# RESEARCH AND MANAGEMENT OF EUTROPHICATION IN COASTAL ECOSYSTEMS

EDITED BY : Jesper H. Andersen, Jacob Carstensen, Marianne Holmer, Dorte Krause-Jensen and Katherine Richardson PUBLISHED IN : Frontiers in Marine Science

#### Frontiers eBook Copyright Statement

The copyright in the text of individual articles in this eBook is the property of their respective authors or their respective institutions or funders. The copyright in graphics and images within each article may be subject to copyright of other parties. In both cases this is subject to a license granted to Frontiers. The compilation of articles constituting this eBook is the property of Frontiers.

Each article within this eBook, and the eBook itself, are published under the most recent version of the Creative Commons CC-BY licence. The version current at the date of publication of this eBook is CC-BY 4.0. If the CC-BY licence is updated, the licence granted by Frontiers is automatically updated to the new version.

When exercising any right under the CC-BY licence, Frontiers must be attributed as the original publisher of the article or eBook, as applicable.

Authors have the responsibility of ensuring that any graphics or other materials which are the property of others may be included in the CC-BY licence, but this should be checked before relying on the CC-BY licence to reproduce those materials. Any copyright notices relating to those materials must be complied with.

Copyright and source acknowledgement notices may not be removed and must be displayed in any copy, derivative work or partial copy which includes the elements in question.

All copyright, and all rights therein, are protected by national and international copyright laws. The above represents a summary only. For further information please read Frontiers' Conditions for Website Use and Copyright Statement, and the applicable CC-BY licence.

ISSN 1664-8714 ISBN 978-2-88963-432-3 DOI 10.3389/978-2-88963-432-3

#### About Frontiers

Frontiers is more than just an open-access publisher of scholarly articles: it is a pioneering approach to the world of academia, radically improving the way scholarly research is managed. The grand vision of Frontiers is a world where all people have an equal opportunity to seek, share and generate knowledge. Frontiers provides immediate and permanent online open access to all its publications, but this alone is not enough to realize our grand goals.

#### Frontiers Journal Series

The Frontiers Journal Series is a multi-tier and interdisciplinary set of open-access, online journals, promising a paradigm shift from the current review, selection and dissemination processes in academic publishing. All Frontiers journals are driven by researchers for researchers; therefore, they constitute a service to the scholarly community. At the same time, the Frontiers Journal Series operates on a revolutionary invention, the tiered publishing system, initially addressing specific communities of scholars, and gradually climbing up to broader public understanding, thus serving the interests of the lay society, too.

#### Dedication to Quality

Each Frontiers article is a landmark of the highest quality, thanks to genuinely collaborative interactions between authors and review editors, who include some of the world's best academicians. Research must be certified by peers before entering a stream of knowledge that may eventually reach the public - and shape society; therefore, Frontiers only applies the most rigorous and unbiased reviews.

Frontiers revolutionizes research publishing by freely delivering the most outstanding research, evaluated with no bias from both the academic and social point of view. By applying the most advanced information technologies, Frontiers is catapulting scholarly publishing into a new generation.

#### What are Frontiers Research Topics?

Frontiers Research Topics are very popular trademarks of the Frontiers Journals Series: they are collections of at least ten articles, all centered on a particular subject. With their unique mix of varied contributions from Original Research to Review Articles, Frontiers Research Topics unify the most influential researchers, the latest key findings and historical advances in a hot research area! Find out more on how to host your own Frontiers Research Topic or contribute to one as an author by contacting the Frontiers Editorial Office: researchtopics@frontiersin.org

# RESEARCH AND MANAGEMENT OF EUTROPHICATION IN COASTAL ECOSYSTEMS

Topic Editors:

Jesper H. Andersen, NIVA Denmark Water Research, Denmark Jacob Carstensen, Aarhus University, Denmark Marianne Holmer, University of Southern Denmark, Denmark Dorte Krause-Jensen, Aarhus University, Denmark Katherine Richardson, University of Copenhagen, Denmark

Wordle based on the keywords from the papers published in this Research Topic on Research and Management of Eutrophication in Coastal Ecosystems. See www.wordle.net for details.

Citation: Andersen, J. H., Carstensen, J., Holmer, M., Krause-Jensen, D., Richardson, K., eds. (2020). Research and Management of Eutrophication in Coastal Ecosystems. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88963-432-3

# Table of Contents

*05 Editorial: Research and Management of Eutrophication in Coastal Ecosystems*

Jesper H. Andersen, Jacob Carstensen, Marianne Holmer, Dorte Krause-Jensen and Katherine Richardson


Carlos M. Duarte and Dorte Krause-Jensen

*86 Disentangling Environmental Drivers of Phytoplankton Biomass off Western Iberia*

A. Ferreira, P. Garrido-Amador and Ana C. Brito

*103 Land Uses Simplified Index (LUSI): Determining Land Pressures and Their Link With Coastal Eutrophication*

Eva Flo, Esther Garcés and Jordi Camp


Luz María García-García, Dave Sivyer, Michelle Devlin, Suzanne Painting, Kate Collingridge and Johan van der Molen


E. Therese Harvey, Jakob Walve, Agneta Andersson, Bengt Karlson and Susanne Kratzer

*200 Sediment Stocks of Carbon, Nitrogen, and Phosphorus in Danish Eelgrass Meadows*

Theodor Kindeberg, Sarah B. Ørberg, Maria Emilia Röhr, Marianne Holmer and Dorte Krause-Jensen


Joseph V. McGovern, Stephen Nash and Michael Hartnett


Fuensanta Salas Herrero, Heliana Teixeira and Sandra Poikane


Richard Tian, Jennifer Keisman and Emily M. Trentacoste

# Editorial: Research and Management of Eutrophication in Coastal Ecosystems

Jesper H. Andersen<sup>1</sup> \*, Jacob Carstensen<sup>2</sup> , Marianne Holmer <sup>3</sup> , Dorte Krause-Jensen<sup>4</sup> and Katherine Richardson<sup>5</sup>

*<sup>1</sup> NIVA Denmark Water Research, Copenhagen, Denmark, <sup>2</sup> Department of Bioscience, Aarhus University, Roskilde, Denmark, <sup>3</sup> Department of Biology, University of Southern Denmark, Odense, Denmark, <sup>4</sup> Department of Bioscience, Aarhus University, Silkeborg, Denmark, <sup>5</sup> Center for Macroecology, Evolution and Climate, University of Copenhagen, Copenhagen, Denmark*

Keywords: nutrients, loads, eutrophication, oligotrophication, recovery, ecosystem-based management

**Editorial on the Research Topic**

#### **Research and Management of Eutrophication in Coastal Ecosystems**

This Research Topic brings together 22 papers on research and management of eutrophication in coastal ecosystems. We, the editors of the Research Topic, hope the readers will find the papers as interesting as we do. We are delighted with the breadth and diversity of the papers. The Research Topic includes 18 original research papers, two reviews, a mini review, and this editorial, which span the entire palette of eutrophication themes, from inputs of nutrients and organic matter, direct effects, indirect effects, to climate change, management strategies etc. across a wide geographical range. The papers of this Research Topic are anchored in EUTRO 2018, the "Fourth International Symposium on Research and Management of Eutrophication in Coastal Ecosystems," which follows up on three earlier symposia: EUTRO 1993, EUTRO 2006, and EUTRO 2010. The broad span of topics related to eutrophication are also reflected in the keywords from the 22 papers.

The roots of EUTRO 2018, as well as EUTRO 1993, EUTRO 2006, and EUTRO 2010, extend from the first Danish Action Plan on the Aquatic Environment, which was adopted by the Danish Parliament in 1987 (Andersen, 2018). In addition to a number of measures aimed at reducing inputs of nutrients to the aquatic environment, the action plan also included a marine research programme known as "Havforskningsprogram 90" (Christensen et al., 1998; Andersen, 2012). The action plan targeted agriculture, industry and urban waste-water treatment, and the successful implementation of nutrient reduction measures led to significant reductions in the inputs of both nitrogen (**Graph 1**) and phosphorus to Danish coastal waters.

"Havforskningsprogram 90" was carried out during the years 1990–1994 and focused on inputs, turn-over, direct and indirect effects of nutrient enrichment in Danish coastal and marine waters (Christensen et al., 1998). The results of the research program, together with eutrophication research results from other parts of the world, were presented at Elsinore, Denmark, at EUTRO 1993, organized by the Danish Environmental Protection Agency in collaboration with the Commission of the European Communities, Directorate-General for Science, Research and Development. Although the concepts of marine eutrophication were discussed before EUTRO 1993, this symposium clearly contributed to its definition (e.g., Nixon, 1995) and highlighted the drivers of nutrient enrichment and eutrophication as well as the biological responses to different nutrient regimes (e.g., Duarte, 1995). EUTRO 1993 reported the experiences from the first generation of nutrient management strategies, i.e., the Danish Action Plan on the Aquatic Environments and the 50% reduction targets adopted by HELCOM and OSPAR (Andersen, 2012). The Symposium Proceedings were

#### Edited and reviewed by:

*Angel Borja, Technological Center Expert in Marine and Food Innovation (AZTI), Spain*

> \*Correspondence: *Jesper H. Andersen jha@niva-dk.dk*

#### Specialty section:

*This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science*

Received: *25 November 2019* Accepted: *28 November 2019* Published: *20 December 2019*

#### Citation:

*Andersen JH, Carstensen J, Holmer M, Krause-Jensen D and Richardson K (2019) Editorial: Research and Management of Eutrophication in Coastal Ecosystems. Front. Mar. Sci. 6:768. doi: 10.3389/fmars.2019.00768*

published in Ophelia—International Journal of Marine Biology as Volumes 41 and 42 and included several seminal papers (e.g., Aksnes et al., 1995; Duarte, 1995; Heip, 1995; Legendre and Rassoulzadegan, 1995; Nixon, 1995; Paerl, 1995; Richardson and Heilman, 1995).

EUTRO 2006, or in full "Research and Management of Eutrophication in Costal Ecosystems—An International Symposium," took place 20–23 June 2006 in Nyborg, Denmark and was organized by the Danish Environmental Protection Agency, the Swedish Environmental Protection Agency, Fyn County and DHI Water & Environment. EUTRO 2006 followed the adoption and implementation of the EU Water Framework Directive (WFD; European Commission, 2000) and the US assessment of eutrophication in the nation's estuaries (Bricker et al., 1999). Hence, EUTRO 2006 focused not only on drivers of eutrophication and biological responses to changes in nutrient input but also on different approaches for assessing eutrophication status of coastal ecosystems. At this symposium, Nixon (2009) presented the definition of oligotrophication. The symposium also focused on different modeling tools and Decision Support Systems, e.g., the Baltic Nest system (Savchuk, 2007). EUTRO 2006 Symposium Proceedings were published in Hydrobiologia (Andersen and Conley, 2009b) and included the seminal papers Conley et al. (2009a,b) and Nixon (2009), as well as a synthesis paper by Duarte (2009).

EUTRO 2010, the "Third International Symposium on Research and Management of Eutrophication in Coastal Ecosystems" took place 15–18 June 2010 and was organized by the International Council for Exploration of the Sea (ICES), the US National Oceanographic and Atmospheric Agency (NOAA) and DHI. A synthesis paper with the title "The Eutrophication Commandments" was published by Fulweiler et al. (2012).

EUTRO 2018 "International Symposium on Nutrient Dynamics in Coastal and Estuarine Environments" took place 18–20 June 2018 in Nyborg, Denmark and marked the 25th anniversary of the first symposium. The EUTRO 2018 presentations underlined that, today, more than 25 years after EUTRO 1993, we have a conceptual understanding of both eutrophication and oligotrophication as well as documentation of management strategies having the potential to transform eutrophication trajectories toward ecosystem recovery. However, several presentations highlighted that systemic time lags may slow down recovery and that it may take decades, if not centuries, before the results of management actions are fully seen (e.g., Murray et al.).

EUTRO 2018 was comprised of seven keynote presentations, two thematic workshops, 49 oral presentations and a symposium summary. The themes for the oral presentations were:


Themes for the thematic workshops were "From monitoring data to integrated assessments" and "Steps toward a harmonized assessment of eutrophication in Europe's seas." For detailed information about EUTRO 2018, please refer to the EUTRO 2018 programme and book of abstracts (Andersen, 2018).

Theme 1 included three original research papers on phytoplankton and harmful algae blooms, one from Kuwait Bay and the Northern Persian Gulf, one from the Western Iberian and one from Ireland:

• Phytoplankton data (2007–2016) from Kuwait Bay and the Northern Persian Gulf document that seasonal and interannual dynamics of plankton communities are linked to land-based inputs of nutrients as well as climate and salinity changes. The combination of nutrient inputs and a warming climate may, moreover, have long-term consequences on the environmental conditions (Devlin et al.).


Four papers were presented under theme 2 on assessment and management tools:


Theme 3 on benthic communities contained four papers:

• The starting point for Christie et al. is the suggestion that the large-scale replacement of sugar kelp by turf algae (ephemeral, filamentous algae) in southern Norway around the year 2000 represented a possible irreversible regime shift. Based on a very large spatio-temporal dataset, the study documents that the seabed state has flipped between sugar kelp and turf algae in several areas and on temporal scales spanning from seasons to years. The paper highlights a complex spatial and temporal distribution pattern between sugar kelp and turf algae and discusses prerequisites and drivers for an irreversible regime shift or a continuation of natural fluctuations, as well as possible mitigation actions (improved coastal water quality, restoration).


Three papers related to theme 4 on land-use and nutrients:


reductions. This, in turn, may indirectly increase the removal of N from less reduced sources. Accordingly, reductions in remote sources in areas not suffering from eutrophication can have a positive effect on areas where eutrophication is a problem.

• The establishment of nutrient criteria (i.e., assessment criteria for nutrient concentrations) for European marine, coastal, and transitional waters can be supported by the statistical approach developed by Salas Herrero et al.. The approach proved sufficient for coastal lagoons but cannot stand alone when developing nutrient criteria for a variety of coastal water types.

Two papers are included from theme 5 on mitigation, oligotrophication, and recovery:


Both papers highlight specific effects of reduced nutrient inputs and document that the two systems have started to recover.

Two papers are related to theme 6 on monitoring, remote sensing, and modeling:


The review papers by Boesch and Deininger and Frigstad cover two different aspects in relation to eutrophication and oligotrophication:


A mini review by Duarte and Krause-Jensen advocates for a broader, more comprehensive approach to reduce eutrophication that considers all major pathways of nutrient budgets of coastal ecosystems, i.e., nutrient inputs, where intervention is most commonly deployed, nutrient export, sequestration in sediments, and nitrogen emissions to the atmosphere as N<sup>2</sup> gas (denitrification). The proposed supplementary management levels involve local-scale hydrological engineering to increase flushing and nutrient export from (semi)enclosed coastal systems, ecological engineering such as sustainable aquaculture of seaweeds and mussels to enhance nutrient export and restoration of benthic habitats to increase sequestration and denitrification in sediments. These ecosystem-scale interventions should be complemented with policy actions to protect benthic ecosystem components.

The papers included in this Research Topic not only take stock of progress regarding our understanding of eutrophication and oligotrophication of coastal marine waters but also represent a perspective on the future. From the results presented at EUTRO 2018, it is obvious that eutrophication trends have been reversed in some coastal systems, e.g., Chesapeake Bay (Zhang et al., 2018), the Danish coastal waters (Riemann et al., 2016), the Baltic Sea (Murray et al.), the Wadden Sea (van Beusekom et al.), and that these systems now display an oligotrophication phase, with various degrees of recovery and restoration toward a more natural ecosystem structure and functioning. We expect a growing number of scientific publications focusing on ecosystem-based management strategies, reduction of loads, oligotrophication and recovery in the coming years.

Today, we understand that individual coastal ecosystems respond in idiosyncratic ways to changes in nutrient inputs and that these systems are controlled by multiple stressors, including, but not limited to nutrient inputs and climate change. Thus, eutrophication cannot be abated only through reduction in nutrient loading. Additional management strategies targeting other controlling factors and stressors and considering effects of climate change are also required. It is also important to recognize that recovery processes may span several decades.

In 2023, we plan to follow up on EUTRO 2018 and this Research Topic by organizing EUTRO 2023, the "Fifth International Symposium on Research and Management of Eutrophication in Coastal Ecosystems." The EUTRO 2023 symposium is in the planning stage and will take place in Nyborg, Denmark in June 2023. EUTRO 2023 will be organized by institutions representing scientists from universities and research organizations, as well as practitioners from competent authorities and stakeholder organizations.

### AUTHOR CONTRIBUTIONS

All authors contributed to designing and writing this editorial based on a draft prepared by JA.

#### REFERENCES


Ecosystems\_Programme\_and\_book\_of\_abstracts


#### FUNDING

This editorial and Research Topic is anchored in EUTRO 2018, the Fourth International Symposium on Research and Management of Eutrophication in Coastal Ecosystems. We thank Danish Centre for Environment and Energy (DCE), the Danish Environmental Protection Agency, NIVA Denmark Water Research, and the Swedish Water and Marine Agency (SwAM) for their dedicated and long-term support of the EUTRO Symposia.

### ACKNOWLEDGMENTS

We thank the participants in EUTRO 2018 and all the contributing authors and reviewers of this Research Topic. We also thank the European Environment Agency, University of Southern Denmark (SDU) and NYBORG STRAND Hotel and Conference Centre for support of EUTRO 2018.


**Conflict of Interest:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Andersen, Carstensen, Holmer, Krause-Jensen and Richardson. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Barriers and Bridges in Abating Coastal Eutrophication

Donald F. Boesch\*

University of Maryland Center for Environmental Science, Cambridge, MD, United States

Over the past 30 years concerted campaigns have been undertaken to reverse nutrient-driven eutrophication in coastal waters in Europe, North America, Asia, and Australia. Typically, eutrophication abatement has proven a more recalcitrant challenge than anticipated, with ecosystem improvements only recently beginning to emerge or falling short of goals. Reduction in nutrient loads has come mainly from advanced treatment of wastewaters and has lagged targets set for diffuse agricultural sources. Synthesis of the major campaigns—varying in terms of physical settings, ecosystem characteristics, nutrient sources, socio-economic drivers, and governance—identified barriers inhibiting eutrophication abatement and potential bridges to overcome them. Actionable science can be advanced by: application of the well-established and emerging knowledge and experience around the globe, client-responsive strategic research, and timely and conclusive adjudication of scientific controversies. More accountable governance requires: enduring engagement of high-level officials of the responsible governments; effective communication of the causes, risks and benefits to the public and stakeholders; quantitative and accountable allocation of responsibility for nutrient load reductions; and binding requirements, as opposed to simply voluntary actions. Effective reduction in nutrient loads requires: reduction strategies for both nitrogen and phosphorus; inclusion of actions that reduce atmospheric emissions of nitrogen in addition to direct inputs to waterways; efficacious regulations; public subsidies based on performance; limitations on biofuel production that increases nutrient loads; and enhancing the sinks and losses for legacy nutrients retained in soils and groundwater. Outcomes must be measured and strategies appropriately adjusted through: sustained monitoring of essential indicators and processes, the use of multiple models, truly adaptive management, and cautious interventions within the coastal ecosystem. The changing climate must be taken into account by reassessing achievable future conditions and seeking alternatives for mitigating and adapting to climate change that also reduce nutrient loads.

Keywords: eutrophication – applied issues, coastal ecosystems, nutrients (nitrogen and phosphorus), hypoxia/reoxygenation, agriculture, Baltic, Chesapeake Bay United States, Gulf of Mexico

## INTRODUCTION

Cultural eutrophication has resulted in consequential changes in coastal ecosystems around the world. Early concerns focused on organic inputs from sewage and industrial wastes. As treatment of these discharges improved, it became clear that inputs of nutrient elements, particularly forms of nitrogen and phosphorus from treated waste and diffuse agricultural and atmospheric

#### Edited by:

Dorte Krause-Jensen, Aarhus University, Denmark

#### Reviewed by:

Robinson W. (Wally) Fulweiler, Boston University, United States Jonathan Lefcheck, Smithsonian Institution, United States

> \*Correspondence: Donald F. Boesch boesch@umces.edu

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 28 September 2018 Accepted: 28 February 2019 Published: 18 March 2019

#### Citation:

Boesch DF (2019) Barriers and Bridges in Abating Coastal Eutrophication. Front. Mar. Sci. 6:123. doi: 10.3389/fmars.2019.00123

**10**

sources, had more expansive consequences detrimental to human uses and ecosystem health. These consequences include diminished water clarity, harmful and nuisance algal blooms, oxygen deficient or hypoxic zones, and degradation of habitats important for living resources.

Extensive scientific research has focused on the progression, causes and effects of eutrophication of coastal ecosystems around the world that became evident in the latter 20th century. A decade ago, Nixon (2009) noted that while many efforts were underway to reduce nutrient loads to degraded coastal ecosystems, there was sparse documentation of the responses to this "oligotrophication" or trend reversal. A burgeoning literature has now emerged on the abatement of nutrient loading and the attendant responses of coastal ecosystems and their catchments (Carstensen et al., 2011).

Abatement of eutrophication does not necessarily lead to restoration of the coastal ecosystems to some previous state (Duarte et al., 2009), but could nonetheless lead to rehabilitation of important ecosystem functions (Choi, 2007) or reconciliation with human uses (Rosensweig, 2003). Effectiveness of abatement is generally evaluated with respect to predetermined objectives such as reduction in nutrient loads, improved water clarity, reduced hypoxia or incidence of harmful algal blooms, or recovery of submersed vegetation or other biotic components of the ecosystem. Measured against such metrics, responses have variously been effective, ineffective, recalcitrant and sometimes surprising.

The goal of this review is less an evaluation of ecosystem responses, but rather the extent to which organized campaigns to abate eutrophication around the world have achieved their management objectives. I explore how, based on this experience, both scientists and society can better navigate obstacles and find effective solutions.

This review is derived from my presentation at the Fourth International Symposium on Research and Management of Eutrophication in Coastal Ecosystems in Nyborg, Denmark, in June 2018. As a framework for assessment of the global experience I chose the metaphors of barriers and bridges used in a book entitled Barriers and Bridges to the Renewal of Ecosystems and Institutions (Gunderson et al., 1995). The book's editors, early thought leaders in ecosystem resilience theory and adaptive management, assembled six case studies to assess how can barriers be broken and bridges built for effective regional ecosystem management. Two of those case studies focused on the Chesapeake Bay (Costanza and Greer, 1995) and the Baltic Sea, large ecosystems where abatement of coastal eutrophication has been a central and concerted focus for both research and management. In the Baltic chapter, Jansson and Velner (1995) reflected on the recent contentious decision to build the Öresund Bridge between Sweden and Denmark. Building bridges seemed to me to be an apt metaphor for a symposium held at the foot of the Belt Sea Bridge in a nation long at the forefront of abating eutrophication and now interconnected by monumental bridges.

Here I review the major campaigns to abate coastal eutrophication in Europe, North American, Asia, and Australia (**Figure 1**). The socio-environmental systems explored vary considerably in terms of physical settings, ecosystem characteristics, nutrient sources, socio-economic drivers, and governance, offering a rich case studies for instructive synthesis. I then offer my perspectives on barriers and bridges to future approaches for achieving eutrophication abatement objectives (**Table 1**). The relevant literature has already become vast; thus, I cite more recent references rather than provide comprehensive documentation.

## EUROPE

## European Union Directives

The abatement of eutrophication in the coastal seas of Europe is a central focus of several regional sea conventions among littoral and riparian nations. The actions taken by the European Unionmember parties to these conventions are generally in conjunction with the implementation of several EU directives. A directive, in contrast to a regulation, is an act that requires member states to achieve a particular result without dictating the means of achieving that result. The Urban Waste Water Treatment Directive, adopted in 1991, has the objective to protect the environment from adverse effects of urban and certain industrial wastewater discharges, encompassing collection, treatment and discharge. The Nitrates Directive (NiD), adopted in 1991, aims to protect water quality by preventing agricultural sources from polluting ground and surface waters through promoting good farming practices. The Water Framework Directive (WFD), adopted in 2000, committed EU member states to achieve good environmental status by 2015 for all water bodies, including marine waters up to one nautical mile from shore. The Marine Strategy Framework Directive (MSFD), adopted in 2008, aims at achieving and maintaining good environmental status in European seas by 2020. The European Commission has produced a set of detailed criteria and methodological standards to help member states implement the MSFD, including abating eutrophication. The MSFD distinguishes four European marine regions located within the geographical boundaries of existing regional sea conventions: the Baltic Sea, the North-East Atlantic Ocean, the Mediterranean Sea and the Black Sea.

### Baltic Sea

Abatement of eutrophication in the Baltic Sea has received more concerted effort and sustained research than any other coastal region in the world. The Baltic Sea: (a) has a wet and largely boreal climate; (b) is an expansive (415,000 km<sup>2</sup> ) and relatively shallow, shelf-sea (mean depth 58 m), but with deeper (>110 m) basins and a shallow restricted connection with the North Sea; and (c) has a vast catchment (1.64 million km<sup>2</sup> ) inhabited by over 85 million people (Reusch et al., 2018). These characteristics combine to create a brackish, tideless sea with year-round thermal stratification that is highly susceptible to eutrophication. Stratification isolates the deeper basins, making them vulnerable to year-round hypoxia that developed during the 20th century and is unprecedented during at least the prior 1,500 years (Jokinen et al., 2018). The hypoxic basins are influenced not only by nutrient-driven productivity, but also periodic intrusions of more saline waters from the North Sea.


In addition, eutrophication has diminished water clarity, caused toxic or noxious algal blooms, resulted in localized hypoxia, and reduced or altered macrovegetation in nearshore environments, fjords, lagoons and archipelagos around the Baltic.

Faced with mounting evidence of environmental degradation, in 1974 Baltic Sea nations joined in the Convention on the Protection of the Marine Environment of the Baltic Sea, better known as the Helsinki Convention. The Helsinki Commission (HELCOM) implements the Convention's plans and directives through its ten contracting parties: Denmark, Sweden, Finland, Germany, Poland, Lithuania, Latvia, Estonia, the Russian Federation, and the European Union. All sovereign parties except Russia are, since 2004, members of the EU. The Heads of Delegation of the contracting parties, typically

nations' environmental ministers, are responsible for HELCOM decisions and national implementation. Commitments made are considered binding, although the consequences of noncompliance have not been tested.

A HELCOM Ministerial Declaration stipulated in 1988 that by 1995 emissions of nitrogen and phosphorus to the Baltic Sea should be reduced by 50% of the 1985 emissions level. These ambitious targets were not met and in 2007 HELCOM launched the Baltic Sea Action Plan (BSAP) that set new targets requiring a decrease in nitrogen and phosphorus loads by 16 and 70%, respectively, from 1997–2003 loads by 2015 (Elofsson and von Brömssen, 2017). Determination of Maximum Allowable Inputs (MAI) of nutrients relied on models covering the entire sea and catchment that are used to allocate the reduction burden among sub-basins and countries. Allocations are refined on an ongoing basis based on new data and model refinements. Wulff et al. (2014) estimated that if control measures are optimized the annual costs to meet the BSAP basin targets would be €4.7 billion.

Substantial reductions in nutrient loads from their peaks in the 1980s have been achieved for both nitrogen and phosphorus, about 24 and 50%, respectively (Reusch et al., 2018). Most of these reductions have come from improvements in wastewater treatment plants, driven largely by the EU's Urban Waste Water Treatment Directive. Further load reductions of about 16 and 38% from present estimated loads, for nitrogen and phosphorus, respectively, are required to meet the BSAP targets. Inputs from diffuse sources, largely agriculture, have been the most difficult to reduce. The difficulty in achieving reductions is compounded by the legacy storage of nutrients in soils and waterways, for phosphorus, and groundwater, for nitrogen. McCrackin et al. (2018) estimated the residence time for phosphorus in Baltic catchments of about 30 years, but concluded that the system has shifted from accumulation to depletion of legacy storage. While it remains difficult to achieve near-term goals for load reductions to the sea, addressing pathways of rapid transport such as overland flow, and mobile stores, such as cropland with large soil-phosphorus reserves, would accelerate load reductions.

Phosphorus also has a long residence time within the Baltic Sea and its coastal waters that delays alleviation of eutrophication symptoms even after anthropogenic loadings are reduced. While efforts were being made to reduce loadings from land, the phosphorus pools in Baltic Sea basins were increasing (Savchuk, 2018). As flow restrictions at the sea's entrance greatly limit phosphorus export, the only loss is to sediment deposition. However, a "vicious circle" of nutrient limitation coupled with hypoxia compounds and prolongs the effect of the accumulated phosphorous in the open Baltic (Vahtera et al., 2007). Hypoxic bottom waters release phosphate that supports the growth of nitrogen-fixing cyanobacteria once surface-water nitrogen supplies are depleted. Enhanced primary production intensifies bottom-water hypoxia that, in turn, fuels more intensive cyanobacteria blooms, particularly in summers such as in 2018 when surface waters are warm and winds weak. The manifestations of eutrophication in the open Baltic, hypoxia and cyanobacterial blooms, may not be abated for many decades. Still, improvements in a multi-component indicator of eutrophication have been in evidence in several Baltic Sea basins beginning in the 1990s. Modeling suggests that if the BSAP load reduction targets are met "good ecological status" would be achieved in the 2030s in the Arkona Basin and in the 2060s in the Kattegat, Bornholm Basin and Gulf of Finland (Murray et al., 2019).

Some businesses, politicians and scientists have advocated quicker alternatives, such as mechanical oxygenation or chemical stabilization of the phosphorus (Conley, 2012). Debates among Baltic scientists over geoengineering, as well as over the efficacy of reductions of nitrogen loads (Schindler et al., 2008; Conley et al., 2009), have sometimes been as vicious as these biogeochemical cycles.

Eutrophication in nearshore waters is directly problematic for citizens and resource users as it results in reduced water clarity, consequent loss of attached plants such as Zostera and Fucus, fish kills, and nuisance accumulation of macroalgae. Hypoxia has also been increasing at many coastal zone sites (Conley et al., 2011). Eutrophication abatement in these environments has usually been addressed by national level actions, but also within the framework of the EU's WFD and MSFD and the HELCOM's BSAP. The need for reduction in nitrogen loads to coastal waters, despite the aforementioned controversies, has now been clearly demonstrated through improvements in water quality and biological conditions following reductions in nitrogen loading, without the confounding effects of nitrogen fixation, in Danish coastal waters (Riemann et al., 2016; Staehr et al., 2017) and Swedish fjords (Savage et al., 2010; Walve et al., 2018).

Experiences in Danish estuaries and coastal waters (including along the North Sea coast) are particularly instructive. Growing environmental deterioration during the 1970s was brought to a head by a fish and shellfish-killing episode of hypoxia in the Kattegat in 1981. This motivated the enactment in 1987 of the Danish National Action Plan on the Aquatic Environment, aimed at reducing nitrogen and phosphorus discharges by 50 and 80%, respectively. A succession of additional agricultural measures was imposed, including limiting nitrogen application to 10% (later 15% below) below the economic optimum rate and obligatory planting of catch crops (Hansen et al., 2017). Sustained and coupled monitoring and assessment programs were established in 1988 to follow the outcomes, including: implementation of agricultural practices; changes in nutrient concentrations in nutrients in groundwaters, surface waters and marine waters; and biological responses. By around 2003, declines in nitrogen leaving the land were evident as follows: discharges from point sources by 74%, the nitrogen surplus in agriculture (the amount of nitrogen applied less the nitrogen removed in crops) by 31%, leaching from the root zone in agricultural land by 33%, and concentrations in streams by 29–32% (Kronvang et al., 2008). After another decade of observations, an assessment concluded that nutrient inputs from land to Danish coastal ecosystems were reduced by about 50% for nitrogen and 76% for phosphorus since 1990 (Riemann et al., 2016). This was accompanied by concomitant shifts in nutrient concentrations in receiving waters, reduced phytoplankton biomass, increased macroalgae coverage and expansion of eelgrass meadows. However, bottom water oxygen conditions in deeper water did

not improve, ostensibly because of more frequent stratification and warmer temperatures.

Ecosystem responses to the abatement of eutrophication can be recalcitrant even when nutrient pollution reduction goals are met. There are myriad reasons for falling short in the "return to Neverland" as Duarte et al. (2009) frame this dilemma. These include both lags in nutrient delivery from the catchment and continued release of accumulated internal loads, as well as "state changes" due to changing climatic conditions, food chain shifts, invasive species and the dislocation of ecosystem engineers. As an example of the last, along the Swedish Kattegat coast, 60% of the eelgrass meadows were lost since the 1980s, but failed to recover despite a significant reduction of nutrient loads and in nitrogen concentrations (Moksnes et al., 2018). Wind-driven resuspension of sediments destabilized by the loss of eelgrass beds limited light penetration and, thus, eelgrass recolonization of deeper bottoms. Drifting algal mats—attributed to overfishing of predators—also cover eelgrass meadows. These feedbacks spread the loss of eelgrass into neighboring areas. Recovery from hypoxia in basins in the Archipelago Sea of Finland is made difficult both by postglacial uplift, that further isolates basins and results in remobilization of organic matter as the wave base shifts and the reduced ice cover prolongs the period experiencing surface waves, as well as by warming surface waters (Jokinen et al., 2018). Regionally specific processes affect recovery, which can be idiosyncratic.

The effects of climate change on rehabilitation of eutrophication have received particular attention for the Baltic Sea, where the climate has already rapidly changed. Water temperatures will continue to rise and river discharges are projected to increase at these latitudes. Changes in wind velocity and direction are also thought to have effects. The general consensus of models and assessments is that climate change will make it more difficult to achieve the goals of the EU directives and the BSAP to achieve and maintain good environmental status that is benchmarked on some earlier condition (Neumann et al., 2012). Nonetheless, models indicate that, while the warming climate might amplify the effects of eutrophication, nutrient load reductions following the BSAP will lead to improved environmental conditions even under future climate changes (Saraiva et al., 2018).

Reusch et al. (2018) make the case for the Baltic Sea as a "time machine," or harbinger of future interactions between climate change and other human impacts in coastal ocean environments more generally. As we move to consider experiences in abating eutrophication in other parts of the world there are several things to keep in mind from the Baltic experience: (a) Baltic nations have sustained a ministerial-level commitment though HELCOM for nearly 45 years and have allocated nutrient load reduction targets that are considered binding. (b) The scientific capacity, activity and engagement in this area surpass by far that of any other region in the world. In particular, strategic transdisciplinary research that is internationally collaborative has been further enhanced by the BONUS Programme: Science for a Better Future of the Baltic Sea (Snoeijs-Leijonmalm et al., 2017). This program has harmonized national research funding and heavily leveraged it with European Commission support. (c) To varying degrees there has been a commitment to environmental monitoring of outcomes and periodic assessments of eutrophication status (e.g., Andersen et al., 2015). (d) Substantial reductions in nutrient loads have been achieved, but reductions from diffuse sources, particularly agriculture, have lagged. Except for the more aggressive approaches pursued by Denmark, efforts under the Nitrates Directive and other EU directives have been insufficient to achieve agricultural source reduction targets. Furthermore, delayed responses as a result of soil, groundwater, and watershed storage and legacy internal loads further compound this recalcitrance.

#### North Sea

The North Sea is a relative shallow shelf-sea (mean depth of 90 m) that, unlike the Baltic Sea, has dynamic exchange with the Atlantic Ocean (Emeis et al., 2015) and consequently residence times of less than 1 month for the coastal North Sea as opposed to 4.5 years for the Baltic Proper (Artioli et al., 2008). While several British estuaries are eutrophied, the designated problem areas for eutrophication in the sea itself lie mainly along its southeastern margins along the continental coast (Claussen et al., 2009). This shallow (<50 m) region is well mixed and receives the discharge of several rivers (Meuse, Scheldt, Rhine, Ijssel, Ems, Weser, and Elbe) that drain catchments totaling 428,500 km<sup>2</sup> , populated by 140 million inhabitants (Emeis et al., 2015).

The North Sea is also the subject of a regional seas convention, OSPAR, effectively the merger in 1992 of the previous Oslo Convention of 1972, covering the dumping of wastes at sea, and the Paris Convention of 1974, covering land-based sources of marine pollution and the offshore industry. The OSPAR Commission has jurisdiction for the broader European North-East Atlantic Ocean, thus has contracting parties additional to those counties that border the North Sea (Belgium, Netherlands, Germany, Denmark, Norway, and the United Kingdom). As recounted in detail by de Jong (2006), eutrophication began to command attention after the occurrence of hypoxic events in the German Bight of the North Sea in 1981 and 1982 that followed from the rapid growth of phosphorus and, particularly, nitrogen discharges from the continental rivers during the 1960s and 1970s. The environmental ministers of the OSPAR contracting parties agreed in 1987 to reduce the river loads of dissolved inorganic phosphorus and nitrogen to 50% of the levels in 1985 by the year 1995.

The phosphorus goal was actually exceeded due to improvements in wastewater treatment and the replacement of phosphates in detergents (Emeis et al., 2015). However, by 1995 dissolved inorganic nitrogen loads had only declined by about 20%. Nutrient loads have continued to decline, reaching about −81% for phosphorus and −45% for nitrogen by 2010. Any further reductions of nitrogen loads from the rivers will have to come from more effective controls on diffuse sources. Also notable are the reductions experienced in atmospheric deposition of nitrogen on the North Sea, which was of the same magnitude as the river loads in 1985. Deposition was almost halved by 2010 due to reductions in emissions of oxidized nitrogen from land-based sources. At the same time, nitrogen deposition from

ship exhausts, responsible for about 17% of nitrogen deposition in the North Sea, is increasing.

Driven by the currents exiting the English Channel, the freshened and enriched surface plume off the river mouths flows along the coast to the north. It stimulates primary production not only in coastal waters, but also in the Wadden Sea, the expansive system of intertidal lagoons running from the Netherlands into Denmark. As loads of phosphorus were reduced more quickly and substantially than for nitrogen, the N/P-ratios in river discharges increased between 1980 and 1992, from 23 to 62 for the Rhine River and 75 to 124 for the Elbe River (Emeis et al., 2015). Primary production in waters near the river mouths was light or phosphorus-limited, thus reducing phosphorus loads allowed more excess nitrogen to be exported up the coast, expanding the reach of eutrophication until nitrogen loads were also reduced.

The environmental conditions in the Wadden Sea have reflected these changes (van Beusekom, 2005). Primary production increased threefold while nutrient loading from the rivers grew. Algal mats replaced seagrasses (Zostera spp.). A three-fold decline in the biomass of benthic infauna was observed in the portions of the Dutch Wadden Sea most affected by the Rhine discharge when phosphorus loading was reduced. In the more distal German Wadden Sea, a reduction of phytoplankton biomass was not seen until after nitrogen loads were also reduced. Despite the apparent abatement of eutrophication the ecosystem had not, however, returned to its previous state largely because of the changes in this intertidal ecosystem related to sea-level rise and invasions of non-indigenous species, including the marsh grass, Spartina anglica, and the introduced oyster Crassostrea gigas, which has supplanted mussel beds (Schumacher et al., 2014).

#### North-East Atlantic Ocean

Eutrophication of coastal waters along the Atlantic coast of Europe from the English Channel through the Iberian Peninsula is manifest in intense phytoplankton blooms, including some that produce nuisance foam events or toxins; local hypoxia; and changes in coastal communities. This eutrophication is caused primarily by the nitrogen and phosphorus enrichment of river discharges and to a lesser extent by atmospheric deposition of nitrogen (Desmit et al., 2018). The region also falls within the scope of the OSPAR Convention, which calls for reducing both nitrogen and phosphorus inputs from rivers by 50% compared with 1985 levels. As all the littoral nations are members of the European Union, they also are responsible for complying with the relevant EU directives (WFD, MSFD, UWWTD, and NiD).

Except for the Duoro River, all of the major rivers experienced substantial declines in total phosphorus discharge between 1991– 1995 and 2001–2005, ranging from 30% for the Loire and Tagus rivers to 55% for the Seine (Romero et al., 2013). Except for the Garonne River, loadings of total nitrogen generally increased during this period, by 27 and 22% for the two largest rivers, the Seine and the Loire. In the bay off the mouth of the Seine River, the decrease in phosphorus loads matched a general decrease in phytoplankton biomass in the summer, but the sustained high loads of nitrogen matched an increase in the abundance of dinoflagellates, including Phaeocystisresponsible for foam events.

Declines in phosphorus discharges were largely due to the ban of phosphates in household detergents and phosphorus removal from wastewater discharges. In the well-monitored Seine River below Paris, treatment of point source discharges led to remarkable improvements in water quality, with sharp reductions in ammonium and phosphate and progressive increase in dissolved oxygen (Romero et al., 2016). Summer anoxia in the Seine estuary has nearly disappeared, aided, in part, by the recent introduction of denitrification of wastewaters. Measures to control nutrients from diffuse sources, particularly agriculture, have, however, been relatively ineffective. Nitrate concentrations increased by 150% between the 1980s and 2010, after which it showed some reversal of this trend. This is due to the ninefold increase in the use of nitrogen fertilizers after the 1950s, compounded by the promotion of artificial drainage of wetlands that reduced nitrogen retention within the watershed. Only a 2– 3% decrease in nitrate was observed in the Seine after 2010 from what was reported in 1998, despite the fact that control measures to reduce the over-dosage of fertilizers have existed for over two decades. The reported nitrogen surplus for agricultural soils has decreased, so the lack of decline in the river might be due to insufficient reductions or to groundwater lag times that could extend several decades.

Ménesguen et al. (2018) modeled the amount and locations of the nutrient source reductions needed to achieve good ecological status within coastal waters. The ineffectiveness or inertia of present measures for reducing agricultural nutrient losses has also prompted use of these models to explore whether the eutrophication abatement goals of the EU's North-East Atlantic Ocean goals are achievable. Linking marine ecosystem models with outputs from a watershed model, Desmit et al. (2018) concluded that decreases in nitrogen fluxes from land sufficient to prevent eutrophication symptoms could not be achieved by implementing wastewater treatment and conventional good agricultural practices, alone. Achieving effective nutrient load reductions would likely entail substantial social, economic and agricultural changes that reshape connections between crop production and livestock farming, and between agricultural and local human food consumption. These changes include less waste production and a shift toward human diets where half of the animal products consumed are replaced by vegetal proteins, known as a demitarian diet.

#### Mediterranean Sea

While the Mediterranean Sea, like the Baltic, has limited exchange with the Atlantic Ocean, it is vastly larger and deeper. Moreover, evaporation from its surface exceeds freshwater inputs via precipitation and rivers. Consequently, its open waters have elevated salinities and are exceptionally oligotrophic. Eutrophication is in evidence in small coastal lagoons and bays affected by intense human activity (Karydis and Kitsiou, 2012). In more open waters, nutrients supplied by rivers are responsible for a large fraction of the pelagic primary production and for bottom hypoxic area only in the Adriatic and Aegean basins (Macias et al., 2017).

Largely based on concerns about pollution, the Mediterranean nations and the European Economic Community adopted

the Mediterranean Action Plan, also known as the Barcelona Convention, in 1976. The Plan was amended in 1995 as the Convention for the Protection of the Marine Environment and the Coastal Region of the Mediterranean, which entered into force in 2004. Member states committed to adopt: measures against land-based pollution, protection of biological diversity, and pollution monitoring. In addition, all of the littoral nations along the northern coast except for Bosnia-Herzegovina, Montenegro, and Albania, are members of the European Union and bound to address the relevant EU directives discussed earlier. The MSFD recognizes a specific Mediterranean ecoregion and calls for achieving good environmental status in the marine environment by 2020, including specifically that human-induced eutrophication is minimized. In response to the MSFD, a Trophic Index (analogous to the HEAT index used for the Baltic) has been widely used to address eutrophication status in northern Mediterranean waters (Pavlidou et al., 2015). Some other non-EU nations on the Mediterranean have developed similar directives paralleling those of the EU.

The northern Adriatic Sea, the most extensive shallow area in the Mediterranean, has received the most attention regarding the effects and abatement of coastal eutrophication. Located between Italy and the Balkans and delimited by the 100 m isobath, its average depth is only 35 m. Eutrophication was manifest in seasonal hypoxia in deep waters during the 1970s and 1980s that was severe enough to cause fish kills and regional elimination of benthic species. From the late 1980s there have been incidents of extensive mucilage production when algal blooms confront severe phosphorous limitation (Degobbis et al., 2005). Discharges from the Po River account for about 65% of the freshwater, nitrogen and phosphorus loads. The effects of associated eutrophication are strongly influenced by variations in circulation and climate. In the lagoons within the Po delta, macroalgal blooms also displaced seagrasses.

Nutrient loads from the Po declined beginning in the late 1980s, but particularly after 2000, and this was accompanied by reductions in soluble reactive phosphorus (SRP) concentrations, chlorophyll a, and plankton biomass in the northern Adriatic (Viaroli et al., 2018). This trend can be reversed during years of high river flow. Net anthropogenic phosphorus inputs to the Po catchment declined by 35% by around 2000 and the SRP loads from the river declined even more. The decline resulted not only from wastewater treatment but also retention of phosphorus in soils and sediments held within streams and rivers. Over the same time, net anthropogenic nitrogen inputs declined by about 18%, but the dissolved inorganic nitrogen discharges from the river have remained unchanged. Consequently, there is currently a large excess in available N relative to P in the northern Adriatic, where the phytoplankton was already phosphorus-limited, causing shifts in micro- and macroalgae and mismatches between grazers and phytoplankton in food webs. The challenge, then, is to reduce loadings of nitrogen by more efficient agricultural practices and by restoring biogeochemical processes in irrigation and drainage ditches, streams and the river that promote denitrification. Abating eutrophication of the coastal waters of the northern Adriatic will require harmonizing policies over large spatial scales and restoring the Po river basin.

### Black Sea

The Black Sea is even more confined than the Mediterranean, but is brackish because of large freshwater inputs. Permanently stratified, it has a large and deep anoxic basin. Concerns about anthropogenic eutrophication are limited to the sea's margins, in particular on the expansive northwestern continental shelf that receives the discharge of the largest rivers flowing into the Black Sea, the Dniester, Dnieper, and particularly the Danube. Because of low sedimentation rates and clear waters, the northwestern shelf harbored a rich benthic biota, characterized by an expansive field of the red macroalga Phyllophora and large populations of the filter feeding mussel, Mytilus galloprovincialis (Mee, 2006). Soviet scientists documented the development and expansion of seasonal hypoxia in bottom waters that grew from 3,500 km<sup>2</sup> and around 30 m deep in 1973 to 30,000 km<sup>2</sup> from 15 to 45 m deep by 1978 (Capet et al., 2013). The Phyllophora field and mussel beds were largely destroyed in the affected area, either due to the lack of oxygen or reduced light availability. Over-fishing and the proliferation of an invasive ctenophore that altered foodwebs combined to cause a collapse and loss in resilience of this productive ecosystem (Mee et al., 2005).

Not surprisingly, the development of hypoxia followed the rapid growth in nutrient loading from the rivers discharging to the northwestern shelf between 1960 and 1990 (Oguz, 2008; Oguz and Velicova, 2010). For phosphorus, growing sources were from urban wastewater, as well as fertilizer application and animal production. For nitrogen it was due mainly to heavy fertilizer application under the centrally planned, Communist economies of Eastern Europe. When Communist states began to collapse in 1989, economies were disrupted and agricultural subsidies disappeared. By 1991, consumption of fertilizers within these watersheds declined by 70% for phosphorus and 50% for nitrogen. Nutrient loads discharged by the Danube River dropped over the next few years, particularly for phosphate. It appeared that the extent of hypoxia on the shelf shrank substantially within only 6 years, although it has taken a longer time for the benthic communities to recover (Mee, 2006).

This story of Black Sea ecosystem recovery has been frequently cited as demonstration of the fairly rapid coastal response to a significant decrease in nutrient inputs to a large watershed. More recent analysis employing models of hypoxia formation and persistence matched with available observations (Capet et al., 2013) suggests that recovery is more prolonged than projected by Mee (2006), who had only limited observations during the recovery period. The models suggest that hypoxia has not been reduced as much as suggested. Although hypoxia was less extensive in 2009 than in 1987, it expands with warmer sea surface temperatures. The accumulation of organic matter in sediments introduces important inertia in the recovery process under declining enrichment, having an effect for as long as 9 years. The inertia in the benthic system may be more complex than just organic matter storage and entails recovery of the functions of biological communities, making resilience in response to disturbances difficult to recover once lost.

In 1992 the six Black Sea countries ratified the Convention on the Protection of the Black Sea Against Pollution (Bucharest Commission) and in 1996 adopted the Strategic Action Plan

for the Rehabilitation and Protection of the Black Sea, which was updated in 2009. The Bucharest Commission developed common strategic goals with the International Commission for the Protection of the Danube River, including avoiding exceeding the nutrient loads experienced during the mid-1990s. Two of the littoral nations, Bulgaria and Romania, and most of the nations in the Danbue watershed are members of the European Union and thus are responsible for implementing the EU directives (O'Higgins et al., 2014).

Achieving the nutrient load exceedance goal, much less additional load reductions, is highly problematic for several reasons. Monitoring of nutrient concentrations and loads is limited and inconsistent, even the benchmark 1990s load levels are poorly defined (Strokal and Kroeze, 2013). Furthermore, there is no formal allocation of loading limits among the catchments and nations on which to target abatement actions. The Danube catchment supplies 70% of the nutrient loads to the northwestern shelf. There 86% of the nitrogen emissions and 71% of phosphorus emissions now come from diffuse sources, but expenditures to reduce eutrophication go overwhelmingly to improve wastewater treatment (O'Higgins et al., 2014). While the EU member nations within the catchment are expected to comply with E.U. environmental directives, they also receive agricultural subsidies that increase fertilizer use and intensify animal production. The EU directives have not demonstrated the ability to improve agricultural practices at the spatial scales required to address diffuse pollution, leaving it to each member state to define and address unclear requirements. Furthermore, Black Sea nations have very different economic conditions, languages, culture and traditions all within a region of great geopolitical tension, further exacerbating the scale mismatch in environmental management.

### NORTH AMERICA

Many estuaries, bays and shallow continental shelf environments in the United States experience eutrophication, particularly along its east and Gulf coasts (Bricker et al., 2008). Here I examine the eutrophication abatement efforts in only three of those ecosystems because of their large scale, the concerted management efforts specifically directed at abating eutrophication, and illustrative range in the dominance of point versus non-point sources. Many bays and estuaries are included within the US Environmental Protection Administration's (USEPA) National Estuary Program (NEP). There are now 28 so designated, notably excluding the Chesapeake Bay that has its own federal authorization. Each NEP location has developed a Comprehensive Conservation & Management Plan that, in most cases, includes a strategy for abating eutrophication. In general, execution of these strategies relies on collaborative rather than regulatory approaches (Lubell, 2004). Few have specific nutrient reduction targets and timelines and governance and accountability mechanisms to track progress.

Even coastal waters such as San Francisco Bay, Puget Sound, and Delaware Bay that were thought not susceptible to eutrophication because of vigorous tidal exchange and mixing are getting a second look. Similarly, Canada's St. Lawrence River estuary, one of the largest in the world, has experienced increased hypoxia, with controversies as to whether this is caused by physical controls (Lefort et al., 2012; Bourgault and Cyr, 2015). However, in no regions of North America has abatement of eutrophication receive more policy and scientific attention that in Tampa Bay, Chesapeake Bay, and the northern continental shelf of the Gulf of Mexico.

#### Tampa Bay

Tampa Bay, located on the Gulf of Mexico coast of the Florida peninsula, is an exemplar for successful abatement of coastal eutrophication. Science-based nutrient reduction goals were achieved and desired ecosystem responses occurred (Greening et al., 2014). The bay is a moderately large (1,036 km<sup>2</sup> ) and shallow (mean depth 4 m) embayment. Its catchment of 5,700 km<sup>2</sup> is 43% urbanized and experienced dramatic population growth from 0.5 to 1.5 million people between 1950 and 1980. Population growth increased nutrient loads such that by the late 1970s residents and tourists observed the effects of eutrophic decline, including diminished water clarity, accumulations of macroalgae, noxious phytoplankton blooms, and loss of about 50% of the bay's seagrass meadows. Occasional hypoxia further threatened living resources.

The bay was already phosphorus-enriched because it receives drainage from inland areas mined for phosphorus ore. The ecosystem was and remains strongly nitrogen-limited as indicated by N:P ratios and bioassays. Therefore, remediation focused on nitrogen removal from point discharges of sewage and industrial wastes that, in the mid-1970s, comprised 60% of the total nitrogen load. Political responses at the state and local levels led the way, with the enactment in 1978 of a Florida statute that required advanced treatment for all wastewater treatment plants discharging to Tampa Bay. Additional nutrient limits were required for stormwater discharges beginning in 1985.

Tampa Bay is included in the NEP and the Tampa Bay Estuary Program (TBEP) has played a critical role in developing an action plan, completed in 1996 and updated in 2006, which has guided the nutrient load reduction efforts. In 1996 the Tampa Bay Nitrogen Management Consortium was created among governmental and industrial organizations to collaboratively allocate additional load reductions needed to meet the targets and offset the effects of continued regional growth. This led to actions to reduce emissions of nitrogen oxides from power plant exhausts, urban fertilizer use restrictions, and other nonpoint source controls of agricultural and mining activities. US federal regulations did not come into play until 1998, when the USEPA recognized a Total Maximum Daily Load (TMDL) based on TBEP-determined targets and Consortiumdetermined allocations.

Under the TBEP, goals and targets were set in a multistep process: (a) a minimum seagrass coverage goal of 15,380 ha, 95% of that estimated present in the 1950s, was set; (b) light requirements for an important seagrass species, Thalassia testudinum, were determined; (c) water clarity levels needed to assure those light requirements were estimated; (d) maximum chlorophyll a concentrations consistent with the water clarity

levels were determined; and (e) maximum nutrient loadings that allow achievement of the chlorophyll a concentration targets were computed. Further, there was a commitment to monitor and assess outcomes and goals annually through an adaptive management processes.

More than US\$500 million has been spent in efforts to reduce nitrogen loadings to Tampa Bay. Loads of total nitrogen during 2000–2011 were 61% less than in the mid-1970s, predominantly attributable to a 91% reduction in point source loads. Fertilizer handling losses, although less important, were also significantly reduced and atmospheric nitrogen deposition began to decline after 2000. These load reductions were achieved despite continued rapid growth of the human population to over 2.5 million. Per-capita total nitrogen loads to the bay fell from 9.6 kg/person/yr in the mid-1970s to 1.3 kg/person/yr for 2000–2011.

TN and TP concentrations in the most affected sections of the bay lagged the load reductions by a few years, ostensibly because of storage of nutrients in sediments, but eventually became two to three times lower than in the early 1980s (**Figure 2**). Chlorophyll a concentrations declined in a similar manner by about a factor of two, as Secchi depth increased concomitantly. Seagrass expansion followed except after weather anomalies increased nitrogen concentrations in 1998 and setback seagrass expansion. Seagrass expansion continued and exceeded the 15,378 ha goal in 2014 (Sherwood et al., 2016). Within several years of wastewater nutrient load abatement, the ecosystem has returned from a turbid phytoplankton-based state to the clear-water seagrassbased system present in the 1950s. Greening et al. (2014) credited the improvements to development of numeric water quality targets, citizen involvement, collaborative actions, and state and federal regulatory programs. Additionally, Sherwood et al. (2016) emphasized the importance of long-term monitoring for measuring progress and thus sustaining political and societal will to continue actions and investments.

Recovery of seagrass beds also occurred following reductions in nitrogen loading to other smaller bays nearby along the Florida Gulf (Tomasko et al., 2018). Seagrass coverage within

the region increased by 12,171 ha between the 1980s and 2016, even with heavy human development along the shorelines and within the catchments. Abatement of eutrophication in other coastal waters of Florida has proven much more recalcitrant. In particular, worsening eutrophication is evident in bays and estuaries south of the Tampa Bay region that receive drainage from the greater Everglades ecosystem. In order to manage flooding in heavily populated areas in southeastern Florida, large quantities of water that naturally flowed through the Everglades from Lake Okeechobee, situated in the center of the Florida peninsula, are diverted to both coasts into the St. Lucie Estuary on the Atlantic coast (Kramer et al., 2018) and the Caloosahatchee River estuary on the Gulf of Mexico coast (Heil et al., 2014). These flows deliver large loads of nutrients from agricultural lands north of the lake or from internal phosphorus loads within the lake. Diversion of excess water through the St. Lucie estuary stimulates blooms of cyanobacteria and has reversed water quality improvements and seagrass recovery in the adjacent Indian River Lagoon. For the Caloosahatchee estuary, nutrient diversions may be intensifying or prolonging red tides, blooms of the toxin-producing dinoflagellate Karenia brevis, in bays and along the coast. Nutrient source controls, even if effective, would likely not produce results in the estuaries for decades. Increasing the capacity of stormwater treatment areas within the agricultural regions south of the lake faces practical limitations and high costs (Wetzel et al., 2017). Moreover, this would not address the reduction of nutrient losses at their source. Meanwhile, scientific debates about the role of nutrient inputs from lower Everglades as causes of blooms of cyanobacteria and algae in Florida Bay continue (Shangguan et al., 2017).

#### Chesapeake Bay

With tidal waters extending over 11,600 km<sup>2</sup> , the Chesapeake Bay is 11 times larger than Tampa Bay, but 32 times smaller than the Baltic Sea. Its 166,000 km<sup>2</sup> catchment is, on the other hand, only 10 times smaller than the Baltic and 29 times larger than Tampa Bay, illustrating why, with it < 1 m tidal range and 6-month mean residence time, it is highly susceptible to changes in delivery of nutrients from diverse land-based sources. The Chesapeake Bay has long been the focus of research related to eutrophication (Kemp et al., 2005). In the 1950s efforts focused on reducing organic waste discharges from Washington, D.C., at the head of the tidal Potomac River, one of the major tributaries discharging to the bay. When noxious algal blooms continued after the completion of secondary treatment, additional phosphorous removal from wastewater discharges was begun in 1974 and fully implemented by 1986.

In 1972 Tropical Storm Agnes resulted in record flows in virtually all the major rivers discharging to the Chesapeake Bay. Recovery from the effects of the massive freshet was slow with a lingering reduction in water clarity, more pervasive hypoxia in bottom waters, and seemingly permanent loss of submersed vascular plants over a large extent of the bay and its tidal tributaries. A 5-year study was commissioned that suggested large-scale eutrophication was the underlying culprit. Still, there were often passionate scientific debates as to whether, for example: hypoxia was primarily driven by stratification during high river discharge or by anthropogenic nutrient inputs; submersed vegetation was lost because of the expanded use of agricultural herbicides or eutrophication; and nitrogen loads played any role or if controlling phosphorous alone was sufficient (Kemp et al., 2005). Results of the 5-year study prompted the first Chesapeake Bay Agreement in 1983 among three states (Pennsylvania, Maryland and Virginia), Washington, D.C., and the federal government, thus establishing the Chesapeake Bay Program (CBP). The agreement simply committed the parties to collaborate to improve and protect the water quality and living resources. It was not until 1987 that the parties committed to develop and implement a strategy to achieve at least a 40% reduction of nitrogen and phosphorous entering the mainstem of the Chesapeake by the year 2000. The scope and timing of the committed nutrient load abatement was remarkably similar and coincident with the HELCOM and OSPAR plans for the Baltic and North seas.

Setting the initial 40% nutrient reduction was based on simple models available at the time and general estimates of what would be effective and achievable (Boesch et al., 2001). In developing the strategy it was realized that nutrients emanate from states in the catchment not then party to the agreement and from atmospheric sources not subject to water pollution controls. Limited to controllable inputs, the total load reductions sought were actually about 24% of the nitrogen and 35% of the phosphorus total loads. The agreement specified that load reductions would be "equitably" allocated among the jurisdictions and these allocations were made in 1988. Each state voluntarily set out to achieve their assigned nitrogen and phosphorus load reductions.

As 2000 approached it was clear that the programs put in place fell woefully short of what was needed to achieve the committed load reductions. A broader and bolder agreement was adopted that required a technically more sophisticated determination and allocation of nutrient load reductions sufficient to achieve water quality standards for designated used in various parts and depth zones in the bay. The new approached linked "airshed" (for atmospheric deposition), watershed, and estuarine ecohydrodynamic models (Linker et al., 2013). The new targets called for reduction in loads of nitrogen and phosphorus of about 43% from 1985 baseline levels (actually a 5-year hydrological average) by the year 2010.

The voluntary approach was continued, but recognizing that it might again fall short, the Chesapeake agreement parties, which included the three additional states that partially lie in the bay's catchment, subsequently agreed that should they fall short they would develop and implement a Total Maximum Daily Load or TMDL plan, which they would be legally bound to meet. Somewhat analogous to the European Union's Water Framework Directive in intent if not process, section 303(b) of the US Clean Water Act requires that for "impaired waters" in which technology-based regulations are not stringent enough to meet the water quality standards set by the states, a TMDL must be calculated as the maximum amount of a pollutant that can be put in a water body and still meet water quality standards. States are then legally responsible through voluntary or regulatory programs to reduce the pollutant loading to the level of the TMDL. While the Clean Water Act prohibits the USEPA from

regulating agricultural runoff other than from combined animal feeding operations, the states could put in place statutes and regulations to control such non-point source pollution.

When it was apparent that they would fall well short of the 2010 target, the parties began to formally develop the TMDL plan, issued by the USEPA in December 2010. The plan refined allocations of load reductions to be met by 2025 and requires reporting and evaluation at 2-year milestones. USEPA's options if a state falls short of its legal obligations are fairly weak, including revocation of delegated authority for permitting point-source discharges and loss of federal grant support. Within weeks of its issuance, the American Farm Bureau Federation, joined by other agricultural business organizations, filed suit in federal court to halt the implementation of the Chesapeake TMDL on the grounds that it was not lawful in providing detailed allocations of load reductions, requiring reasonable assurances, and mandating states to address water quality impairment not just in that state but also in downstream states. The suit also alleged that the TMDL relied on overextended scientific models and flawed data. It was clear from the plaintiff's arguments that their primary goal was to avoid a precedent that could be applied to the much larger agricultural enterprise in the Mississippi River basin. The federal court upheld the legality of the Chesapeake TMDL and the Farm Bureau's appeal of the court's decision was denied.

Estimated annual nutrient loads to the bay are depicted by source category and period in **Figure 3**. Loads at the end of 2016, the mid-way point in the implementation of the binding TMDL, are compared with the starting baseline ca. 1985, prior to the implementation of the TMDL in 2009, and reaching the TMDL by 2025. While direct measurements of wastewater loads are included in these estimates, other loads are estimated by watershed model simulation for hydrological conditions consistent with the base condition around 1985. For any given year, the loads could be greater or lower than those loads depending on the annual freshwater discharge that can vary by a factor of three. Furthermore, in order to guide implementation, source reductions assumed for diffuse source abatement practices are immediately credited as load reductions in the watershed model. This adds uncertainty associated with assurance of actual implementation and presumed efficiency of the practice, as well as lag times between implementation and realized load reductions to the bay.

On the basis of the modeled estimates, the CBP concluded that by 2017 actions had been taken that would reduce phosphorus loads nearly to the TMDL target, largely because wastewater treatment had already exceeded the load reductions allocated for that sector. Agricultural loads of phosphorus had been reduced by 69% of the TMDL planning target, but little reduction in loads from urban runoff had yet been achieved, mainly because land development had continued to expand. The watershed model may have underestimated phosphorous loads from soils with a large phosphorus surplus due to repeated application of manures (Kleinman et al., 2011).

For nitrogen, loads from wastewater have been reduced below the 2025 allocation because of substantial investments in enhanced nitrogen removal, largely funded by user fees. Controls on nitrogen oxide emissions from power plants and vehicles through implementation of the U.S. Clean Air Act substantially reduced atmospheric deposition such that this could account for most of the observed decline of nitrate fluxes in the Potomac River catchment (Eshleman and Sabo, 2016). On the other hand, simulated nitrogen load reductions from agriculture, urban runoff and septic disposal of domestic wastes are significantly behind schedule. Practices have been implemented to meet only 18% of the load reductions allocated to agriculture under the TMDL plan, despite substantial financial support, technical assistance and state regulations for nutrient management. While the watershed model estimates that the practices put in place since the 1985 should reduce nitrogen loads from agriculture by 25%, loadings delivered to the bay from its major rivers have not declined comparably and those draining heavily agricultural areas not at all (Zhang Q. et al., 2015). This indicates either that delivered load reductions lag for a decade or more due to groundwater storage or that management actions have been less effective than credited, or both. Even so, at the end of 2016 estimated nitrogen load reductions from agriculture were behind schedule in each state, but particularly in Pennsylvania, which occupies 35% of the bay's total catchment but has no bay shoreline.

In addition to new Watershed Implementation Plans for each jurisdiction designed to close the gap to the 2025 targets, the CBP is developing strategies to offset the additional nutrient load that has resulted from the sediment infilling of pools behind hydroelectric dams along the lower Susquehanna River (Zhang et al., 2016), the large tributary river discharging to the head of the Chesapeake Bay. The Program is also estimating what additional nutrient load reductions would be required to maintain dissolved oxygen standards under the changing climatic changes (principally warmer estuarine temperatures and increased runoff; Irby et al., 2018) that will have occurred between 1985 and 2025.

Beyond the CBP watershed model estimates (**Figure 3**), monitoring of wastewater discharges and rivers draining to the bay demonstrates overall abatement of nutrient loading, much of it evident prior to the imposition of a TMDL in 2010. Yet, until recently there were few signs of the anticipated responses of the estuarine ecosystem. Murphy et al. (2011) were the first to document the reduction in late summer hypoxic volume beginning in the 2000s. Gurbisz and Kemp (2014) chronicled the rapid reestablishment in the early 2000s of a large submersed plant bed near the head of the estuary that had virtually disappeared in 1972. More recently, Lefcheck et al. (2018) showed that coverage by submersed vascular plants throughout the bay increased 17,000 ha since 1985, reaching a level not seen in a half century. They attributed this to a 23% decline in dissolved nitrogen concentrations. Testa et al. (2018) reported a decline in NH<sup>4</sup> and an increase in NO2+<sup>3</sup> concentrations in channel bottom waters during late summer as dissolved oxygen concentrations have increased, suggesting a vicious biogeochemical cycle may be breaking (Kemp et al., 2005). Zhang et al. (2018) demonstrated progressive improvement in a multimetric index of water quality attainment that integrates dissolved oxygen, water clarity, submersed vascular vegetation and chlorophyll a. This began in

https://www.chesapeakeprogress.com/clean-water/watershed-implementation-plans; accessed on 12 February 2019.

the late 1990s following wastewater treatment plant upgrades and reductions in atmospheric deposition of nitrogen.

While the intended rehabilitation of the Chesapeake Bay is far from completed, high-level, intergovernmental commitment has been sustained for more than 35 years and demonstrable nutrient load reductions have been accomplished. More reductions in nutrient loads will be required and it remains to be seen if they will be accomplished by the third deadline in 2025 and, if not, what will be the legal and political repercussions. Even then, it may take decades for the full rehabilitation of the ecosystem to be realized. Changes in climate and in socioeconomic drivers will also influence outcomes. But, it is too soon to confine the rehabilitation goals for the Chesapeake to Neverland status (sensu Duarte et al., 2009).

#### Northern Gulf of Mexico

Bottom waters over the inner continental shelf of the northern Gulf of Mexico near the Mississippi River delta experience recurring seasonal hypoxia, the first large area of coastal hypoxia to which the term "Dead Zone" was applied. Until the mid-1980s hypoxia was sparingly observed and thought to be isolated, ephemeral and natural (Rabalais et al., 2002). Compelling

evidence grew that expansive and seasonally persistent hypoxia had occurred after the 1960s coincident with a three-fold increase in the nitrate loads from the Mississippi and Atchafalaya rivers, the two principal distributaries from the 3.2 million km<sup>2</sup> river basin (Rabalais et al., 2007). Based on mid-summer, quasi-synoptic surveys conducted since 1985, bottom hypoxia (dissolved oxygen concentrations < 2 mg/L) has extended over an area averaging 13,700 km<sup>2</sup> and as much as 23,000 km<sup>2</sup> , typically in water depths of 10–40 m. The extent depends on the amount and timing of river discharge, prevailing winds and the occurrence of early tropical storms that mix the water column. Hypoxic conditions usually extend west from the Mississippi river mouth along Louisiana to the upper Texas coast. Annual surveys underestimate the extent of bottom hypoxia as it can form, dissipate and move throughout the summer.

Seven years after the first suggestions that the extensive shelf hypoxia likely resulted from increased nitrate loading from the Mississippi River (Turner and Rabalais, 1991), the US Congress enacted the Harmful Algal Bloom and Hypoxia Research and Control Act of 1998 (HABHRCA). The Act required the submission of an integrated assessment of northern Gulf hypoxia that examined its distribution, dynamics and causes; ecological and economic consequences; sources of nutrient loads; and methods, costs and effects of reducing these loads. That assessment was published in 2000 and an Action Plan was endorsed the next year by the interstatefederal Mississippi River/Gulf of Mexico Watershed Nutrient Task Force (Rabalais et al., 2002). The assessment concluded that hypoxia in the northern Gulf of Mexico is caused primarily by increased nitrogen delivered from the Mississippi-Atchafalaya River Basin in combination with the natural stratification of Gulf waters. Nitrate flux nearly tripled between the 1955– 1970 and 1980–1996, while organic carbon and phosphorus fluxes probably decreased over the latter 20th century. Ninety percent of the nitrate load came from diffuse sources, principally agricultural lands in Iowa, Illinois, Indiana, southern Minnesota, and Ohio, 1,800 km or more upstream of the Gulf of Mexico. An approximately 40% reduction in total nitrogen flux would be necessary to return to loads comparable to those during 1955–1970, requiring limiting losses from fields and enhancing nitrogen retention and denitrification within the basin.

The goal of the 2001 Action Plan was to reduce the 5-year running average extent of the Gulf hypoxic zone to less that 5,000 km<sup>2</sup> by 2015 through implementation of specific, practical, and cost-effective voluntary actions to reduce the discharge of nitrogen. It did not commit to a specific reduction in nitrogen discharge, but recognized that a 30% nitrogen load reduction was probably needed. The Plan proved controversial to upstream agricultural interests, so in 2006, as part of a planned 5-year reassessment, a new panel was convened under the auspices of USEPA's Science Advisory Board to evaluate more recent science and options for reducing the size of the hypoxic zone. The USEPA took the unusual step of inviting public comment on the scientists being considered for the panel and then deciding not to appoint any of those who had been involved in the 2000 integrated assessment. Nonetheless, the panel reaffirmed the principal conclusions of the integrated assessment (EPA Science Advisory Board, 2008). It found the 5,000 km<sup>2</sup> goal a reasonable endpoint, but noted that it would not likely be achieved by 2015 because of limited progress in implementing policies, programs and strategies to reduce nutrient loads and lag times in ecological system response. The panel recommended the reduction in both nitrogen and phosphorus fluxes by at least 45%.

The Task Force has continued to meet regularly over the 18 years since developing the Action Plan. It is presently comprised of representatives of five federal agencies and 12 states within the Mississippi River Basin. Its members are federal officials at the assistant secretary level or below and department heads or below from the states. Notably, officials from the departments of agriculture represent most of the states responsible for large agricultural nutrient loads—Iowa, Indiana, Illinois, and Ohio—rather than environmental protection or natural resources department heads.

After the reassessment, the Task Force produced the 2008 Gulf Hypoxia Action Plan that identified actions to accelerate the reduction of nitrogen and phosphorus losses and advance the science, track progress and raise awareness. Following another reassessment in 2013, the Task Force decided to maintain the goal of reducing the hypoxic zone to less than 5,000 km<sup>2</sup> , but extended the target date 20 years to 2035. The Task Force set an interim 20% reduction target for both nutrients by 2025 as a milestone for immediate planning and implementation actions. Once again, it allocated no specific nutrient load reductions to states or sub-basins. States were expected to complete and implement comprehensive nitrogen and phosphorus reduction strategies. These were completed between 2013 and 2016, but vary in designation of specific nutrient load reduction targets and geographical targeting of technical and financial assistance for reducing agricultural nutrient losses. Both the states and the USEPA have contested lawsuits seeking enforceable limits on nitrogen and phosphorus pollution through the preparation and implementation of a TMDL and resisted the development of numeric water quality criteria for nitrogen (Sigford, 2016). Even the state of Louisiana, the most affected by Gulf hypoxia, has resisted designating seasonally hypoxic shelf waters as "impaired," as this might initiate a TMDL process.

Resistance to allocating and targeting nutrient loss abatement has limited progress despite the results of effective semiempirical modeling tools that locate and quantify the sources of nitrogen and phosphorus loads delivered to the Gulf of Mexico (Robertson and Saad, 2014). Delivered nitrogen sources are highest from the Corn Belt centered over Iowa and Indiana, but the highest phosphorus yields are associated more with animal agriculture and wastewater treatment discharges. Models have even been used to target the specific subwatersheds where existing cropland conservation practices augmented by better fertilizer management could achieve the hypoxia reduction goal most cost-effectively (Rabotyagov et al., 2014).

More recent estimates of the effects of nutrient load reductions on the extent of shelf hypoxia have ranged from relatively linear production-respiration models to more complex threedimensional hydrodynamic models (Fennel et al., 2016; Scavia et al., 2017). Some models suggest that nitrogen loads would

have to be reduced as much as 60% to accomplish the 5,000 km<sup>2</sup> maximum hypoxia areal extent goal. If both nitrogen and phosphorus loads were reduced simultaneously, a 48% reduction in each would be required. Benchmarking the goal by hypoxic volume rather than aerial extent, as is done for the Chesapeake Bay, may be more relevant to living resource impacts. A more modest 25% nitrogen load reduction would substantially reduce the thickness of the hypoxic layer to a relatively thin layer near the bottom in much of the affected area (Scavia et al., 2019).

Because it its complexity and the economic stakes, science underpinning the abatement of Gulf hypoxia has attracted its critics. During the initial integrated assessment, scientists aligned with agricultural interests or supported by the fertilizer industry challenged the basic conclusion that shelf hypoxia had worsened as a result of increased nitrogen loading from the Mississippi-Atchafalaya Basin (Rabalais et al., 2002). Some oceanographers argued that shelf hypoxia was caused by upwelling, stratification from increased river flow, or increased loading of terrestrial organic matter (e.g., Rowe and Chapman, 2002). A USEPA analyst blamed hypoxia on phosphorous pollution from industry and cities and asserted little would be accomplished by reducing nitrogen loads (Ferber, 2004). A member of the USEPA Science Advisory Board panel, joined by other regional scientists, criticized what they termed the "nutrient-centric view" of hypoxia abatement, stating that "mandating a specific nutrient reduction target level is "difficult to defend" and "doomed to fail" (Bianchi et al., 2008, 2010). They asserted that the goal of reducing hypoxia to 5,000 km<sup>2</sup> was scientifically unjustified and expressed doubt that reducing nutrient loads would contract hypoxia occurring west of the Atchafalaya River. Further, these authors suggested worsening hypoxia could have been caused by outflowing organic matter from deteriorating coastal wetlands.

Scientists experienced with eutrophication and hypoxia in the Gulf and elsewhere countered these criticisms (Boesch, 2003; Boesch et al., 2009). Furthermore, subsequent research using mass balances (Das et al., 2011) and stable isotopes (Wang et al., 2018) demonstrated that inputs of wetland organic carbon could have only a trivial effect on shelf hypoxia. Recent models have also indicated that the hypoxic volume on shelf west of the Atchafalaya would be substantially ameliorated by reductions of riverine nutrient loads (Scavia et al., 2019). While legitimate scientific debates are appropriate and often helpful, the aforementioned criticisms of reliance on nitrogen source abatement cast doubts that slowed Task Force actions at critical intervals. Criticisms also inspired an attempt in the U.S. Congress in 2014 to amend HABHRCA to prevent any further work on the Action Plan until another assessment was conducted to address the issues raised by Bianchi et al. (2008). The paralyzing effects of such controversies illustrate the need for more effective mechanisms to resolve scientific questions concerning important matters of public policy in a timely fashion.

Despite hundreds of meetings, state reduction plans, and substantial federal state and private expenditures for improvements in nutrient management practices over the past 17 years of the Action Plan, there is no evidence that nutrient loads to the Gulf or hypoxia have yet been reduced. The extent of hypoxic bottom waters mapped in the 2017 was 22,770 km<sup>2</sup> , the largest ever measured. The 5-year running average was essentially the same as at the starting point for the Action Plan. Nutrient loading varies widely from year-to-year as a function of river flow, so that even 5-year averages of loads can be misleading. Once normalized for variations in flow by statistical adjustment for time, discharge and season, neither riverine nitrate concentrations nor loads showed any decline between 1985 and 2010 (Sprague et al., 2011; Murphy et al., 2013). Flow-adjusted nitrate loads have been essentially stable since 1995 (**Figure 4**), in contrast with declining flow-adjusted loads to the Chesapeake Bay (Oelsner and Stets, 2019).

The lack of a significant decline in nitrate loading does not necessarily mean that actions taken for more efficient use and retention of agricultural nutrients have been ineffective. Flowweighted nitrate concentrations in the Illinois River catchment declined significantly since 1990 due to increasing nitrogen use efficiency in largely tile-drained agriculture and the depletion of legacy nitrogen, resulting in a 10% load reduction (McIsaac et al., 2016). However, loads did not decline significantly in other watersheds in the Basin and increased in the Upper Mississippi and Missouri rivers (Murphy et al., 2013). Nitrate loads from the Iowa, contributing about 29% of loads to the Gulf, have been increasing since 1999 (Jones et al., 2018). Nitrogen use efficiency in the agricultural heartland was already high by global standards and has shown only small decline (Swaney et al., 2018), but the area under cultivation has expanded, largely as a result of the dramatic increase in crops grown not for human food or animal feed, but for biofuels. Federal requirements for the inclusion of ethanol and biodiesel in motor fuels have driven the expansion and focus of cropping, with 40% of the maize crop currently produced to refine ethanol. Increased production of maize would likely increase nitrate export to the Gulf by 14% between 2002 and 2022, while improved crop nitrogenrecovery efficiency (NRE) and reduction in atmospheric nitrogen deposition could result in only a 9% reduction (McCrackin et al., 2017). Significant nitrogen load reductions will require aggressive actions to improve basin-level NRE and remove nitrogen from wastewaters. Practices that reduce nitrogen losses from tile-drained cropland, responsible for up to half of nitrate export, would also have substantial benefit. Greater reliance on perennial grasses rather than maize for bioenergy development could also substantially reduce nitrogen and phosphorus loads, particularly if these crops were strategically placed as buffers between streams and intensively fertilized food crops (Ha et al., 2018).

It should be expected that reduction of nutrient fluxes from the Mississippi and Atchafalaya rivers lags the actions taken to reduce their losses from crop production because of temporary storage in soils, groundwater and streams. These lags could delay delivered load reductions for years to decades. Employing a model that accounts for such memory effects of past nutrient use, Van Meter et al. (2018) estimated that even if agricultural nitrogen use became 100% efficient, it would take decades to achieve a 60% reduction. More effective management of tile drainage and restoration and reconnection of wetlands and bottomland forests could speed up the process by enhancing denitrification and nutrient retention.

## ASIA

With its large and dense populations and rapidly increasing demand for animal protein, the coastal waters of Asia experience widespread and generally increasing eutrophication. While efforts were made to abate eutrophication in Japan and Hong Kong decades ago, scientific documentation of eutrophication in much of the rest of Asia has emerged just during the last two decades as there yet are few concerted efforts to abate coastal eutrophication.

accessed on 12 February 2019. Loading estimates produced by Lori Sprague and Casey Lee, U.S. Geological Survey.

In Japan, coastal eutrophication has been manifest mainly in bays such as Tokyo and Ise Bays and in portions of the Seto Inland Sea that are surrounded by large human populations and consequently received heavy loadings of organic matter and nutrients from municipal and industrial wastewaters. In the 1950s extensive red tides in the Seto Inland Sea and fish kills around Tokyo Bay raised concerns about degrading coastal water quality. The Water Pollution Control Law enacted in 1970 required development of effluent and water environmental standards. By 1980 Total Pollutant Load Control Systems

(TPLCS) were put into practice for Tokyo Bay, Ise Bay and the Seto Inland Sea (Yanagi, 2015). Under this system the Minister of the Environment prepares a Basic Policy for Areawide Total Pollutant Load Control and the Prefecture governors are responsible for formulating and implementing pollutant load control plans. While these plans initially focused on organic pollutant loads, some reductions on nutrient loads were also achieved in the process of abating organic loading. In 2001 total nitrogen and total phosphorus was added as specific reduction targets.

For the Seto Inland Sea, phosphorus and nitrogen loads were reduced by 55 and 45%, respectively, between 1979 and 2009, almost all of that attributable to treatment of industrial and domestic wastewaters. Hypoxia has been reduced in heavily impacted waterbodies such as Dokai and Osaka bays, but their rehabilitation has been hindered as a result of large internal sediment loads of phosphorus and deterioration of the physical habitat. Physical restoration and bioremediation are being actively pursued. Red tides in the Seto Inland Sea became less extensive and severe and phytoplankton biomass declined on broader scales. In the Suo Nada, the westernmost portion of the Seto Inland Sea between Honshu and Kyushu, chlorophyll a levels declined substantially, particularly in shallow waters, between 1984 and 2004, closely following reductions in loads of phosphorus and nitrogen (Nishijima et al., 2016). The introduction of advanced treatment for some wastewater discharges has also yielded reductions in nutrient concentrations and chlorophyll a in Tokyo Bay (Kubo et al., 2019). For at least for some Japanese coastal waters, reduced primary production has led to concerns by some fishers, politicians and even scientists that abatement of anthropogenic nutrient loads has been too hasty and has led to severe phosphorus limitation and declines in the secondary production of exploited resources (Yamamoto, 2003).

Based on its experiences in abating eutrophication, Japan has been assisting other East Asian countries in applying TPLCS. Korean coastal waters also experience eutrophication associated with diffuse source runoff and sewage discharges, particularly evident in bays along South Korea's southern coast. This is manifest in seasonal hypoxia (Lee et al., 2018) and harmful algal blooms and red tides (Lee et al., 2013). To abate eutrophication the South Korean Ministry of the Environment has begun to implement the Japanese TPLCS within several catchments.

Efforts have also been undertaken to control the deleterious effects of eutrophication, particularly red tides, in the coastal waters around Hong Kong, a mega-city of 7.4 million people. In the 1980s sewage discharges were redirected from the shallow and poorly flushed Tolo Harbour to the deeper and wellflushed Victoria Harbour (Xu et al., 2010). In 2002 wastewaters discharged to Victoria Harbor were diverted to a new treatment plant that provides chemically enhanced primary treatment and discharges outside of the harbor (Lee et al., 2006). The frequency of red tides, that cause fish kills by depletion of dissolved oxygen rather than by toxins, has been reduced but not eliminated. Controlling excessive nutrient loading presents a complex set of challenges for Hong Kong, which borders on the west to the Zhujiang (Pearl) River estuary that is experiencing emerging hypoxia (Qian et al., 2018). As with the Mississippi, discharges from this large river into the estuary are greatly enriched with nitrogen—nitrate concentrations increased two to three-fold since the mid-1970s, while phosphate concentrations remained constant. Consequently, primary production in the lower estuary is strongly phosphorus limited. During high river flows and when winds drive surface waters to the east, nitrogen-enriched water is driven through Hong Kong's islands to its eastern waters, where production is nitrogen-limited, at least during the summer. This illustrates the complex challenges to management where nutrients are delivered from continental sources interact with those derived from local wastewaters. Fortunately, Hong Kong has sustained a program since 1982 that has monitored water quality and algal blooms and aids researchers in understanding complex dynamics and detecting long-term changes.

Eutrophication is pervasive and worsening along many parts of the coast of China (Jiang et al., 2018), prompting public and media attention, expanded research, and incipient environmental policies. Reminiscent of the Mississippi River and the northern Gulf of Mexico, a zone of seasonally hypoxic bottom water in the East China Sea associated with the plume of the China's largest river, the Chang Jiang, has developed seasonally since at least 1993 (Zhu et al., 2011). This hypoxia is fueled by in situ primary production supported by increased nutrient loads from the river (Wang et al., 2016) that has shifted landward as sediment loads were reduced due to trapping by the Three Gorges Dam (Chen et al., 2017) and extensive tidal flat construction and channelization affected flows from the estuary (Wu H. et al., 2018). So-called green tide blooms of drifting macroalgae (Ulva prolifera) have proliferated along the Yellow Sea coast (Liu et al., 2013), gaining global notoriety when they interfered with sailing events during the 2008 Olympic Games. While expanding seaweed aquaculture within the region apparently seeds these macroalgal blooms, the growth of the drifting green algae is stimulated by anthropogenic nutrient inputs.

In addition to nutrient inputs from populous coastal cities and ubiquitous coastal agriculture and aquaculture, large rivers, including the Huang He (Yellow), Chang Jiang (Yangtze) and Zhujiang (Pearl), drain much of interior China into the Bohai Gulf, Yellow Sea, and East and South China Seas. Between 1970 and 2006 modeled inputs of dissolved nitrogen and phosphorus to these seas increased by a factor of two to five (Strokal et al., 2014). Inputs of particulate nitrogen and phosphorus and dissolved silica declined due to damming of rivers, notably the construction of the Three Gorges Dam on the Chang Jiang. As opposed to the recently declining or stable nutrient loads from large rivers in Europe and North America, nitrogen and phosphorus discharges monitored from the eight main rivers flowing into the coastal waters of China increased by 40 and 75%, respectively, from 2006 to 2012 (Tong et al., 2015). While nearly half of the dissolved nitrogen and virtually all of the dissolved phosphorus loads to the Bohai Gulf in 2000 originated from human sewage, almost all of the dissolved nitrogen and nearly half of the dissolved phosphorus inputs to seas to the south emanate from agricultural sources or, in the case of nitrogen, atmospheric deposition (Strokal et al., 2014). Development scenarios for 2030 to 2050 suggest that it will be difficult to

hold the line on, much less reduce, nitrogen and phosphorus loads. Improved sewage treatment is required to reduce risks of harmful algal blooms in the Bohai Gulf, but improved agricultural management is required elsewhere. Given the demands for agricultural production in China, achieving more efficient use of fertilizers will require fundamental restructuring of Chinese agriculture by increasing farm size (Wu Y. et al., 2018).

While long experiencing eutrophication of restricted coastal waters, the Indian subcontinent, with its dense and growing populations and developing economies, is also increasingly susceptible to regional-scale eutrophication. Models similar to those discussed for China estimate that dissolved nitrogen and phosphorus loads discharged to the Bay of Bengal doubled between 1970 and 2000 (Sattar et al., 2014). While the load increases emanated primarily from fertilizers and manure, the contribution of nitrogen and phosphorus from human wastes, much of which is not collected or treated, is likely underestimated (Amin et al., 2017). With economic and social development projected to increase, constraining the increase in nutrient loads will require more efficient use of fertilizers and manure and recycling sewage wastes in agriculture (Sattar et al., 2014).

## AUSTRALIA

Eutrophication in coastal waters of the more sparsely populated and mostly arid Australia has been evident principally in a few bays and estuaries (Davis and Koop, 2006). Cyanobacterial blooms created nuisances in the Peel-Harvey estuary, south of Perth, beginning in the 1960s. Creating a second channel entrance from the restricted estuary to the ocean in 1994 alleviated the nuisance blooms. Concomitant reductions in phosphorous loading may have also played a role.

Moreton Bay, near Brisbane, is impacted by runoff and deposition of fine-grained sediments from rivers and nutrients discharged by wastewater treatment plants. In the late 1990s local governments began to implement advanced wastewater treatment, specifically including nitrogen removal (Abal et al., 2001). By 2011 nitrogen and phosphorus discharged in wastewater declined by 65 and 46%, respectively (Gibbes et al., 2014). By 2012 seagrasses that had largely disappeared from the mainland shore of the bay had returned to some extent. Nonetheless, the scientific justification for nitrogen removal—isotopic indicators indicating biological uptake of sewage-derived nitrogen and nutrient addition bioassays (Glibert et al., 2006)—was questioned by Wulff et al. (2011). They argued that there is ample benthic nitrogen fixation such that phosphorus is effectively the limiting nutrient in Moreton Bay. They suggested that reducing nitrogen loads exacerbated blooms of the nitrogen-fixing harmful cyanobacterium Lyngbya majuscule, although others pointed to bioavailable iron in humic runoff as the principal driver of these blooms (Albert et al., 2005). Local authorities have not relaxed advanced nutrient removal wastewaters.

In an arm of the Hawkesbury River estuary north of Sydney, Larsson et al. (2017) followed the response of phytoplankton after the implementation of nitrogen removal in a wastewater effluent using a long-term monitoring database. The concentrations of oxidized nitrogen and ammonia declined in the estuary and chlorophyll a concentrations also declined during the summer months, when the risk of harmful algal blooms is greatest.

While these examples from Australian coastal waters are from relatively enclosed waters mainly affected by wastewater discharges, a much larger challenge confronts the abatement of land-based nutrient pollution of the Great Barrier Reef (GBR) and the broad lagoon separating it from the mainland (Kroon et al., 2016). The area at risk is vast—the GBR Marine Park alone covers 344,400 km2—and the excess nutrients emanate predominantly from agriculture in areas lightly populated by humans. Changes in this coastal ecosystem began with extensive land clearing of catchments after 1850, but have become more profound with the intensification of agriculture, including widespread cattle grazing and more concentrated dryland cropping and irrigated production of sugar cane and fruit. As elsewhere, use of fertilizer nitrogen increased dramatically after the 1950s, and loads of nitrogen and phosphorus to the coastal marine system increased many-fold. Increased nitrogen availability resulted in higher phytoplankton biomass, which is associated with outbreaks of the coral-eating crown-or-thorns seastar. In addition, macroalgae proliferated on inshore reefs, crowding out corals, and fine sediments and nutrients led to reduced water clarity and loss of seagrass beds. To be sure, eutrophication is not the cause of the massive die-off of corals on the northern Great Barrier Reef due to recent bleaching induced by extremely high temperature (Hughes et al., 2018), but it contributes additional stress, particularly to inner shelf reefs.

In 2003 the Australian and State of Queensland governments jointly began to implement a Reef Water Quality Protection Plan that relies predominantly on the use of voluntary best management practices for agricultural land uses. Nearly AUS\$1 billion will have been spent through 2022 to implement the plan. However, other government policies allowing further development of water resources in order to increase agricultural production actually work against the adoption and effectiveness of these best practices. Models developed as part an integrated monitoring, modeling and reporting program suggest some progress was made toward the meeting the 2013 nutrient load reduction goals. Still, the estimated load reductions fell short because disincentives and perceived risks limited the sustained application of best management practices. Even if these practices had been fully and successfully implemented they would be insufficient to achieve nutrient load reduction goals.

As with the controversy concerning limiting nutrients in Moreton Bay, assessments of eutrophication status and solutions for the GBR have been the subject of "vicious" debates within the scientific community. In an ongoing feud, Larcombe and Ridd (2018) argue that systemic failings occur in the quality control of environmental science, especially in "agenda-driven science" used to inform government policy. They cited the science related to eutrophication of the GBR as a case in point, concluding that it overstates declines in ecosystem conditions and the need for the prescribed management actions. Leading scientists involved in Reef Plan assessments offered point-by-point rebuttals, bluntly arguing: "that the critiques demonstrate biases, misinterpretation, selective use of data and oversimplification, and also ignore previous responses to their already published claims" (Schaffelke et al., 2018).

## BARRIERS AND BRIDGES

fmars-06-00123 March 14, 2019 Time: 16:28 # 18

Based on experiences drawn from the globally distributed campaigns I have categorized the barriers to abating coastal eutrophication under five overarching themes (**Table 1**). I briefly summarize these particular barriers and consider bridges that can help overcome them.

## Advancing Actionable Science

Limited knowledge of the causes and consequences of changes observed in coastal ecosystems during the 20th century initially slowed action to abate eutrophication, even in those regions where there are now substantial campaigns to accomplish this objective. Are the symptoms just natural phenomena? Do they actually matter to the health of the ecosystem and living resources on which humans depend? Are they stimulated by organic matter or nutrients, and which nutrients? What are the principal nutrient sources and can they be reduced? What level of rehabilitation can be achieved? We now have a wealth of understanding that can assist other regions of the world abate eutrophication more expeditiously. The global scientific community can assist by applying their knowledge and experience to accelerate diagnosis and prognosis.

Even in regions that are comparatively well studied, understanding is often fragmentary across the biophysical and social sciences and among agricultural, hydrological and marine sciences. Fragmentation results not only from the cultural segregation among fields of science and engineering, but also because of the narrow focus of research sponsors. In the Chesapeake Bay region the states have sustained the infrastructure of research institutions and the salaries of core scientists, but these scientists have to compete for research support at a national level, where priorities are not well aligned with the requirements of Chesapeake Bay rehabilitation. Scientists and agency analysts are then left to weld together the fragments of understanding. To bridge not only disciplinary barriers but also international ones, research funding agencies in the EU countries bordering the Baltic Sea pooled their resources and attracted additional support from the European Commission to sponsor the strategic research program BONUS: Science for a Better Future of the Baltic Sea Region. The program has been very successful in advancing actionable research and building transdisciplinary and international science (Snoeijs-Leijonmalm et al., 2017). Its expansion to include the North Sea, with additional national funding partners, is under active development.

Often-public controversies among scientists regarding causes and solutions have sometimes hindered or confused purposeful and efficient abatement of coastal eutrophication. Examples were provided from Baltic and Black seas, Gulf of Mexico, Chesapeake Bay andAustralia.Often, criticswere oceanographers, limnologists, agricultural scientists or economists unfamiliar with the science underpinning efforts to abate coastal eutrophication or without a broader perspective on how the planetary boundaries for biogeochemical flows of nitrogen and phosphorus have been breached (Steffen et al., 2015). Armed with such a global perspective, it would be surprising if extensive hypoxia had not developed in the northern Gulf of Mexico as nitrate loading from the Mississippi River tripled. In some cases, controversies were driven by feelings of disciplinary exclusion and even personal animus. Traditional institutional mechanisms such as advisory boards and review committees are too often inconclusive in resolving such controversies, falling in the trap of focusing on the unknowns rather than what is known with some confidence. More effective bridging of this barrier requires more responsive and conclusive adjudication of key issues and controversies.

## Providing Accountable Governance

Governance of campaigns to abate eutrophication are commonly handicapped by the lack of authority and responsibility by the guiding intergovernmental council. Contrast, for example, the Mississippi River/Gulf of Mexico Watershed Nutrient Task Force, which has resisted even allocating voluntary nutrient load reduction targets, with the HELCOM Heads of Delegation or the Chesapeake Bay Program Governing Council. The former are mid-level officials, some of whom regard their role as preventing aggressive abatement actions, as opposed to officers who are elected or otherwise directly accountable to the public for delivering on commitments. Engagement of responsible parties at the highest practicable level is an important bridge to achieving reliable, and even binding, commitments and account for outcomes.

Public awareness and support is essential for the success of campaigns to abate eutrophication because such campaigns require sustained efforts and incur costs to the public treasury, ratepayers or responsible parties vested in the status quo. Where the source of the nutrient loads is remote from the coastal waters experiencing the effects of eutrophication, e.g., in the Corn Belt from the Gulf of Mexico, the Danube Basin to the Black Sea, and the farms of Pennsylvania for the Chesapeake Bay, this can be problematic. The affected fishers along the coast are usually not organized around the issue of eutrophication and have muted voices compared to agriculture, for example. Non-governmental organizations can be highly effective in advocating for environmental protection and rehabilitation as demonstrated in the Baltic Sea and Chesapeake Bay regions. The news media can also sustain public awareness, but many are now financially challenged to retain experienced environmental journalists. However, new opportunities are provided via online news and social media. Scientists enjoy a trusted position in society and can be effective in communicating understanding concerning the causes and risks of eutrophication and the benefits of its abatement by honing skills in working through NGOs and the news media.

Over-generalized commitments, such as reducing the extent of hypoxia in the northern Gulf of Mexico or avoiding exceeding nutrient loads to the Black Sea, are unlikely to yield intended results. At a minimum, the nutrient load reductions needed to achieve good environmental status (EU) or a TMDL (US) must be estimated and then proportionally allocated to the

responsible jurisdictions and sectors. Even though politically determined allocations may not be the most hydrogeochemically or economically efficient (Iho et al., 2015), they provide the framework for actions and accountability.

Even then, voluntary, non-binding commitments to reduce nutrient loads have proven ineffective in achieving abatement targets. Technological improvements in wastewater treatment required by directives or regulations resulted in substantial load reductions in the Baltic and North seas, the North-East Atlantic, and Tampa and Chesapeake Bay. However, general frameworks such as the Water Framework Directive or TMDL have been less successful in achieving load reductions neededfrom diffuse sources (Voulvoulis et al., 2017). This is particularly true for agriculture, which in the U. S. is exempt from regulation under the Clean Water Act and in Europe is weakly and inconsistently regulated underNitratesDirective. In June 2018, the EU Court of Justice ruled that Germany had breached the Nitrate Directive by allowing an excessive use of manure as fertilizer. A bridge tomore effective reduction of diffuse agricultural nutrient losses is exemplified by the Danish action plans, progressively implemented since 1985, which included obligatory limitations on fertilizer application and requirements for manure management and planting of catch crops (Hansen et al., 2017). These were eventually rolled back in 2016 because of political pressures.

#### Reducing Nutrient Loads

Narrow debates over whether reducing loads of nitrogen or loads of phosphorous was required to abate coastal eutrophication confounded most of the campaigns at one time or another. These remained unresolved not only because of dogmatic claims that only one nutrient can be limiting, but also on overly simplistic interpretations of N:P concentrations compared to the Redfield ratio or bioassays performed in immediate receiving waters. The case studies reveal the complexity of nutrient limitation, as either nutrient can be limiting seasonally or with distance from the source. Within the open Baltic Sea reducing the availability of phosphorus is essential to breaking the vicious circle of hypoxia and diazotrphic nitrogen fixation (Vahtera et al., 2007; Savchuk, 2018), while reduction of nitrogen loads is also required to reverse eutrophication in coastal regions. Reducing phosphorus loads to waters with highly enriched concentrations of nitrogen risks exporting unassimilated nitrogen farther afield or changing the nutrient balance in ways that promote harmful algal blooms (Paerl et al., 2018). The bridge across this barrier is a strategy of dual-nutrient control that takes into account the specific conditions and processes along full environmental transitions and appropriately phases nitrogen and phosphorus load reductions (Conley et al., 2009; Paerl et al., 2016).

The atmospheric deposition of nitrogen emanating from emissions of nitrogen oxides from fossil fuel combustion or from releases of ammonia from animal wastes is often not taken into account in eutrophication abatement strategies. Whether falling on forested or agricultural landscapes, urban surfaces or directly onto coastal waters, the atmosphere can be a significant source of nitrogen inputs, but one that is regarded beyond control of water pollution agencies. Significant reductions in atmospheric deposition of nitrogen within the Chesapeake Bay catchment and North Sea were realized through regulations developed under the air quality directives and laws designed primarily to reduce human health risks. This has had less effect on the open Baltic Sea, where nitrogen fixation has compensated for reduced atmospheric deposition. Nonetheless, projections of future reductions in deposition are now included in the sustainable implementation plans for nitrogen loads to the Chesapeake and Tampa bays and even greater gains are possible as we rely more on renewable energy.

Once required by directive, national law or regulation, the collection and advanced treatment of wastewaters resulted in substantial reductions in loadings of phosphorus, and then nitrogen, leading to significant abatement of eutrophication. Except for Denmark, reductions in diffuse sources of nutrient loads, primarily from agriculture, have generally failed to achieve load reduction goals, even taking into account lags between implementation of practices and water quality outcomes. For economic and social reasons, policy makers have been reluctant to impose regulatory requirements on agriculture. Rather than following a "polluter pays" principle, policies have often taken a "pay the polluter" approach in the form of subsidies and technical assistance for implementation of pollution reduction practices (Iho et al., 2015). This has produced only modest nutrient load reductions, particularly when contravened by other agricultural policies, including commodity subsidies that incentivize high yield and biofuel mandates.

Bridges to overcome barriers to more effective abatement of diffuse agricultural sources differ between phosphorus and nitrogen because of their origins and hydrogeochemical behaviors. For nitrogen, strategies should be fundamentally be founded on minimizing the nitrogen surplus (the difference between the nitrogen applied and the nitrogen removed in crops), a concept also captured in achieving a nitrogen balance (McLellan et al., 2018) or maximizing nitrogen use efficiency (Zhang X. et al., 2015). Many studies (e.g., Osmond et al., 2015) have demonstrated that farmers commonly exceed agronomic prescriptions of application rates, which already err on the side of achieving maximum yields. The nitrogen surplus is a robust determinant of nitrogen losses and can be calculated from readily available farm data. Additional components to the strategy should include targeting of mitigation measures to portions of catchments that, because of soil and other factors, disproportionately contribute to nutrient loads (e.g., Hashemi et al., 2018), controlling sub-surface drainage systems to promote nitrogen assimilation in crops and denitrification in soils (Poole et al., 2018), and construction of wetlands that receive farm runoff (Tournebize et al., 2017). Ribaudo et al. (2016) estimated that if it were required that nitrogen applications not exceed the agronomic needs of crops in order to be eligible for US federal conservation and commodity program payments (including crop insurance), excess nitrogen application in the Mississippi-Atchafalaya River Basin could be reduced by 60%. Such compliance requirements are already used for soil conservation and wetland protection. If strict regulation of agricultural practices is achievable, such compliance requirements could be extended to constraining the nitrogen surplus and additional pollution-reduction practices in targeted areas.

At the same time that efforts are underway to stem the loss of nutrients from agricultural food production, producing crops for biofuels has grown rapidly, displacing food crops, intensifying the cultivation of maize, and expanding cultivation of marginal lands. Planned expansion in production of maize-based ethanol would increase the flux of dissolved inorganic nitrogen down the Mississippi-Atchafalaya River system sufficient to offset the 22% interim reduction goal of the Hypoxia Action Plan (Donner and Kucharik, 2008). Bridges over this barrier to eutrophication abatement include requiring more stringent interventions to control nutrient pollution for fuel-production agriculture and accelerating the transition to cellulose rather than starch-based biofuels. Perennial grasses grown for biofuels require much less fertilization, sequester carbon, and have dramatically less nitrate leaching and nitrous oxide production (Smith et al., 2013; Robertson et al., 2017; Ha et al., 2018).

Campaigns to abate eutrophication should be neither oblivious nor resigned to the long lag times between actions to reduce nitrogen and phosphorus losses and reductions in delivered loads. Accounting for progress through models that assume reductions in delivered loads are immediate, such as in the CBP, belies the reality that it may take years for these reductions to be fully realized. It is critical to ascertain the influence of lag times in the catchment through more diagnostic groundwater monitoring and modeling (e.g., Hansen et al., 2017), not only to avoid imposing changes to intervention measures due to impatience (Vero et al., 2018), but also to guide when and how to make adjustments if they are not effective. Progress can be accelerated by addressing rapid transport pathways, such as surface runoff of phosphorus and the large soil phosphorus reserves in croplands (McCrackin et al., 2018) and large nitrogen losses from subsurface drainage systems. About half of the agricultural nitrate load in the Mississippi River catchment comes from such tile drainage systems (McCrackin et al., 2017) and can be substantially reduced by use of controlled drainage, especially if the drainage then flows through wood-chip bioreactors or constructed wetlands (Christianson et al., 2018). Rehabilitation of existing wetlands and riparian zones is another fast-acting measure for reducing nutrient loads.

## Assessing Outcomes and Adapting Strategies

Assessing outcomes requires monitoring of changing inputs and ecosystem responses, but monitoring is often inadequate for this task (e.g., hypoxia monitoring in the Gulf of Mexico). Even where there is adequate monitoring it may be not be sustained (e.g., the exemplary Danish monitoring program) or monitoring data produced are not regularly interpreted (e.g., until recently in the CBP). Bridging the monitoring barrier requires a truly sustained commitment to monitoring of nutrient inputs and essential indicators and processes of coastal ecosystem responses. Periodic integrated assessments of outcomes based on monitoring results (Andersen et al., 2015; Zhang et al., 2018) should be institutionalized.

Models are also required to connect nutrient sources, delivery mechanisms, fate and responses (Ménesguen and Lacroix, 2018). They should only be as complex as necessary. Even in this era of high-speed computing, overly complex and highly resolved models may convey a sense of false precision and constrain their use to explore abatement options. Use and intercomparison of multiple models, with various levels of complexity and agility, and the reporting of confidence limits rather than single deterministic projections, should be encouraged.

Over-reliance on models not reconciled with observations is dangerous. While useful for tracking abatement actions, use of model estimates of load reductions, as the CBP, risks false confidence in the degree of implementation and effectiveness of these actions and belies the reality of lagged responses. Truly adaptive management approaches offer a bridge over this pitfall. Adaptive management (Allen and Garmestani, 2015) requires reconciling model estimates, including pertinent lags, with observed outcomes on a regular basis and a management commitment to make adjustments to strategies based on these assessments.

Because of compounding pressures that accompanied the eutrophication of a coastal ecosystem, abatement of nutrient loading to the targeted level might not achieve the intended rehabilitation of the ecosystem. Intervention within the coastal system may be an option, including enhancing denitrification in wetlands, seagrass beds and restored oyster reefs; enhancing sediment sequestration; hydrological engineering to increase flushing and nutrient export from relatively enclosed water bodies; and seaweed or shellfish aquaculture (Duarte and Krause-Jensen, 2018). Of course, precaution should be used to avoid unintended or undesirable consequences.

### Addressing Climate Change

Campaigns to abate coastal eutrophication are now taking place at the same time that ecosystems are being altered by global climate change. The discharges and nutrient loads from rivers will be affected by increased temperature and changes in the amount and distribution of precipitation. For example, Sinha et al. (2017) estimated that precipitation changes—both total and extreme precipitation—could increase riverine total nitrogen loading from the continental US by 19% (median estimate) by the end of the century. In addition, coastal waters are influenced by increased temperature, sea-level rise and changes in cloudiness and winds in ways that affect the formation of hypoxia and other effects of eutrophication. Models project that these various forces will exacerbate hypoxia in the Baltic Sea (Saraiva et al., 2018), northern Gulf of Mexico (Laurent et al., 2018) and Chesapeake Bay (Irby et al., 2018), counteracting improvements made by reducing nutrient loading and requiring more aggressive load reductions to achieve rehabilitation goals. The CBP is currently estimating how this will affect the Total Maximum Daily Load for nitrogen and phosphorus even within 2025 time frame for achievement. Scientific analysts will have to continually reassess future conditions that are achievable and the climate-smart strategies capable of reaching evolving goals.

The higher greenhouse gas concentrations rise, the more difficult it will be to abate coastal eutrophication and the accompanying complications from acidification. Yet, waterquality and climate change strategies are almost always decoupled. Analysts run models using high emissions pathways,

counseling managers to plan for the worst. Yet, if human society were to meet the goals of the Paris Climate Agreement, global temperature would be stabilized below a 2◦C increase. As nations work to reduce emissions, there will be many opportunities to reduce nutrient losses to the sea also, including the near-elimination of atmospheric deposition of nitrogen, carbon sequestration, and adjustments of agricultural practices responsible for large greenhouse gas emissions. The development of mitigation and adaptation strategies that limit climate change and its impacts is the super-bridge that must be built during the mid-21st century. Done smartly, it could present significant co-benefits for abating coastal eutrophication.

#### CONCLUSION

While often more challenging than anticipated, campaigns to abate 20th century eutrophication of coastal ecosystems have achieved rehabilitation targets or are on pathways toward

#### REFERENCES


those targets. This collective experience reveals many common barriers to rehabilitation and bridges to overcome them for science, governance, achievement of nutrient load reductions, assessing outcomes and adapting strategies, and addressing the complications of 21st century climate change.

#### AUTHOR CONTRIBUTIONS

DB conceived, researched, and wrote the paper.

### FUNDING

This work was supported by the University of Maryland Center for Environmental Science and the Keith Campbell Foundation for the Environment, which also provided open access publication fees.

a decade of change. Harmful Algae 8, 21–32. doi: 10.1016/j.hal.2008. 08.028


reduction. J. Environ. Manage. 218, 451–464. doi: 10.1016/j.jenvman.2018.04. 078


fmars-06-00123 March 14, 2019 Time: 16:28 # 22

on Cochlodinium polykrikoides. Harmful Algae 30, S3–S14. doi: 10.1016/j.hal. 2013.10.002


Estuary. Environ. Sci. Technol. 50, 2255–2263. doi: 10.1021/acs.est5b0 6211


**Conflict of Interest Statement:** The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Boesch. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Shifts Between Sugar Kelp and Turf Algae in Norway: Regime Shifts or Fluctuations Between Different Opportunistic Seaweed Species?

Hartvig Christie\*, Guri S. Andersen, Trine Bekkby, Camilla W. Fagerli, Janne K. Gitmark, Hege Gundersen and Eli Rinde

Norwegian Institute for Water Research, Oslo, Norway

#### Edited by:

Marianne Holmer, University of Southern Denmark, Denmark

#### Reviewed by:

Perumal Karthick, Sea6 Energy Pvt Ltd., India Mads Solgaard Thomsen, University of Canterbury, New Zealand

> \*Correspondence: Hartvig Christie hartvig.christie@niva.no

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 September 2018 Accepted: 07 February 2019 Published: 22 February 2019

#### Citation:

Christie H, Andersen GS, Bekkby T, Fagerli CW, Gitmark JK, Gundersen H and Rinde E (2019) Shifts Between Sugar Kelp and Turf Algae in Norway: Regime Shifts or Fluctuations Between Different Opportunistic Seaweed Species? Front. Mar. Sci. 6:72. doi: 10.3389/fmars.2019.00072 Around year 2000, sugar kelp (Saccharina latissima) forests were observed to disappear in southern parts of Norway, being replaced by mats of turf algae (i.e., filamentous ephemeral algae) loaded with sediments. Among more than 600 stations covering 35 000 km of coastline, about 80% on the Skagerrak coast and about 40% on the North Sea coast were dominated by turf. Various types of turf algae replaced S. latissima in a discontinuous pattern. This large spatial scale event was reported as a possible irrevocable regime shift, not caused by a single factor but related to multiple stressors, where eutrophication and ocean warming were proposed to be the most important. Recent observations have however, revealed that the seabed state has flipped back and forth between sugar kelp and turf algae in several areas and on temporal scales spanning from seasons to years. The relative abundance of S. latissima at monitoring sites at the Norwegian southern coast has fluctuated dramatically during the last 12 years, varying from sparse to common at several of these sites. In 2016, sugar kelp abundance had increased in more than half of the sites, compared to earlier years. Our monitoring data as well as other field observations and field experiments question the regime shift paradigm. Although traditionally considered as a perennial macrophyte, several of our studies indicate that sugar kelp possesses many of the characteristic traits of an opportunistic species, such as high dispersal potential and colonization rate, which enables the species to rapidly colonize available substrate. However, where turf algae persist, space for recolonization of sugar kelp will most likely be minor. In this paper we explore the spatial and temporal shift dynamic between sugar kelp and turf algae based on monitoring data and other studies. Based on a synthesis of mapped fluctuations between the two states, and studies on sugar kelps recolonization abilities, we discuss prerequisites and drivers for an irrevocable regime shift or a continuation of natural fluctuations, as well as possible mitigation actions.

Keywords: sugar kelp, turf algae, regime shift, flips back, opportunistic algae, eutrophication

## INTRODUCTION

fmars-06-00072 February 20, 2019 Time: 17:19 # 2

An increasing global concern related to regime shifts from perennial foundation kelp species to turf algae (i.e., filamentous, ephemeral algae) have recently initiated "dramatic" headlines in the scientific literature, such as "Turf wars" (O'Brien and Scheibling, 2018) and "Rise of turfs: a new battlefront. . .. . ." (Filbee-Dexter and Wernberg, 2018). These papers (and recent references therein) relate large-scale shifts, earlier described as eutrophication effects (Duarte, 1995; Nixon, 1995; Valiela et al., 1997; Schramm, 1999; Cloern, 2001), to be the result of the combined influence of a complex multifactorial set of direct and indirect stressors (see Filbee-Dexter and Wernberg, 2018), where climate change may play an important role. Norwegian kelp forests are among the many global examples of ecosystems that have been known to experience such shifts. The first shift from sugar kelp (Saccharina latissima) to turf algae was reported in the early 2000's (see Andersen et al., 2011; Moy and Christie, 2012). The concern for the declining distribution of sugar kelp initiated monitoring and scientific studies in southern Norway with focus on this species. The comprehensive sugar kelp project 2005–2008 covering a high number (>600) of sites has been followed by a more site-specific monitoring program, which is still running. Data from more than 15 scientific reports (in Norwegian) was synthesized in the final report from "the sugar kelp project" (Moy et al., 2009). The report concluded that the declines of S. latissima were discontinuous within both small and large spatial and temporal scales, but an overall evaluation concluded that the loss of kelp was considerable both at the southeast coast and the southwest coast. Moy and Christie (2012) classified the ecological state of sugar kelp in five quality classes, indicating how the degree of kelp loss varied both between and within regions. Although extensive dominance by turf algae, S. latissima were observed to occur sparsely or more abundant (between poor and god classification) at about 60% of the sites.

Andersen et al. (2011), Andersen et al. (in press), and Moy and Christie (2012) concluded that the heavy growth of epiphytic algae on S. latissima kelp lamina reduced light penetration below critical levels and thus caused kelp death. Turf algae replaced kelp on the seafloor, trapped inorganic and organic sediments, and prevented recruitment and recovery of the kelp (Moy and Christie, 2012). The dominant turf algae in the Skagerrak coast (southeast Norway), with short dominating turf algae (such as Cladophora spp, Heterosiphonia japonica, Trailliella intricata), differed from of the larger turf algae species within the North Sea coast (south-west Norway, e.g., Spermatochnus paradoxus), dominating the ephemeral vegetation during the summer season. The reported shifts were discontinuous in space and time (Moy et al., 2009; Moy and Christie, 2012) and the causes to the observed patterns were difficult to identify. Areas with high water movement, caused by waves and currents, were in these studies suggested to be unfavorable for the turf algae. This relationship was reported earlier by Pihl et al. (1999) from the Swedish west coast, and confirmed by Bekkby and Moy (2011), who reanalyzed the Norwegian sugar kelp monitoring data and developed spatial distribution models. After visiting 605 stations in southern Norway (2005–2008), Moy and Christie (2012) estimated that a shift from kelp to turf dominance had occurred at approximately 80% of the Skagerrak stations and at approximately 40% at the North Sea stations. Bekkby and Moy (2011) modeled that approximately 50% of the sugar kelp forests areas within the Skagerrak area was lost. More frequent monitoring of 10 stations between 2005 and 2008 (Moy and Christie, 2012), showed persistence of turf algae dominance at many sites, but also documented recovery of S. latissima and temporarily (seasonal) recovery in spring, often followed by increased epiphytic load and turf algae dominance throughout summer. The large spatial and temporal variation in kelp and turf algae abundance along the Norwegian coast, initiates the following hypothesis about their possible development; the turf algae dominated sites may persist for several years, the turf algae sites might recover to sugar kelp dominance, or the sites might fluctuate between years with dominance of either one or the other of the two groups (**Figure 1**).

On a global scale, the distribution of perennial macrophytes (seagrasses and larger seaweeds) shows decreasing trends (Waycott et al., 2009; Araujo et al., 2016; Krumhansl et al., 2016). The reported shifts from sugar kelp to turf algae in southern Norway has contributed to this general trend. Sugar kelp has traditionally been considered as a stable perennial kelp species (e.g., Bekkby and Moy, 2011; Araujo et al., 2016) with a life span of about 3 years and with spore production each winter. In their review, Filbee-Dexter and Wernberg (2018) stated that: «Shifts from kelp forests to turfs have not shown recovery, but reefs have remained in a degraded turf state». This is certainly a valid statement so far, with exception for the annual kelp Undaria pinnatifida (South et al., 2017). But looking with "new eyes" on the data from Moy and Christie (2012) and relating these also to more recent relevant studies from Norway (not easily accessible reports in Norwegian) has made us question this statement and the idea of sugar kelp as a stable, perennial species. While flips in kelp systems mainly have moved from kelp to another less favorable state (Steneck et al., 2013; Filbee-Dexter et al., 2016; Krumhansl et al., 2016; Filbee-Dexter and Wernberg, 2018) representing persistent regime shifts, several observations from southern Norway indicate that flips back to kelp may occur. The aim of this paper is (1) to explore the shift

dynamic pattern at monitored stations, to assess the extent of recovery of sugar kelp at turf dominated stations, and to test any temporal trends in the fluctuations between the two groups, (2) explore, through data from previous studies, life history traits of sugar kelp important for recovery; and (3) to discuss prerequisites and drivers for potential irrevocable regime shifts to turf algae communities or the existence of natural fluctuations between sugar kelp and turf algae, considering both as opportunistic species. Understanding such dynamics will have implication for evaluation of mitigation actions. This paper will not try to highlight pressures behind the shift from kelp to turf algae, as this probably involve several physical, chemical, and biological interactions that even might work synergistically. A complete understanding of the drivers and the dynamics of these shifts is not possible based on analysis of monitoring data only, and demand complex further investigations.

## DATA COMPILATION AND ANALYSIS

This study is based on available data from earlier published material (included data reports in Norwegian) from the 1990's and up to 2017. Most of the data were sampled within the projects "the sugar kelp project": 2005–2008; "the sugar kelp monitoring program": 2009–2012; and the monitoring program "ØKOKYST": 2013–2017. Data were available from several scientific reports (e.g., Moy et al., 2006, 2007, 2008, 2009), from a synthesis paper (Moy and Christie, 2012), and from recent monitoring reports (Fagerli et al., 2017; Naustvoll et al., 2018).

The main aim of the sugar kelp monitoring was to map the state of the sugar kelp forests (sugar kelp abundance versus turf algae abundance) and to detect possible changes in ecosystem state over seasons and years. Hence the data can be used to identify any continuous or discontinuous shift of the two states (kelp and turf) over space and time during the monitoring period. During the surveys in 2005–2008, more than 600 stations were recorded along the southern Norwegian coast (a complex coastline of 35 000 km) and classified as described below. If transects (mainly 0–20 m depth transects) were visited or different depths were recorded on some stations, the data from 5 to 6 m depth were used for the overall comparisons. All stations were in moderately wave exposed, or wave sheltered areas (SWM > 100 000 in Isaeus, 2004, see Gundersen et al., 2011) rocky bottoms (bedrock, boulders, stones), which is where sugar kelp is expected to grow. Most of the stations had not been visited before, so expected presence of sugar kelp was estimated from 30 recordings in the 1980's and 1990's at sites where sugar kelp dominated from about 1 m and down to about 15 m depth (see Moy and Christie, 2012). Both the understory species and the species dominating the turf community varies along the depth gradient and between regions (south vs west, see Moy and Christie, 2012). Turf algae consisted mainly of ephemeral algae, with high abundance during summer and reduced abundance in the winter season.

In the early period of monitoring (2005–2008) the sugar kelp and turf algae abundance were determined by use of drop-camera (with depth sensor, operated from a boat), and only in a few cases by SCUBA diving. Diving was mainly done at selected sites that were revisited during years and seasons. The ecological status of each site was classified after a semi-quantitative abundance scale of sugar kelp (0: absent, 1: single specimen, 2: scattered, 3: common, and 4: dominating), combined with occurrence of turf algae (cf Moy and Christie, 2012). In subsequent programs (2009–2016) annual monitoring was continued at 10 stations in Skagerrak and along the North Sea coast (in the West). Dropcamera was replaced by dive surveys, where abundance of all macroalgal species (or taxa) was recorded semi-quantitatively by the identical 5-step scale as presented above. All visible species were recorded along fixed transects, approximately 0.5 m on each side of the diver's position, i.e., 1 m<sup>2</sup> at each depth). Observations were made for every meter from 1 to 4 m below surface and for every second meter from 4 to maximum 30 m depth. The long-term monitoring of fixed sites provides an opportunity to document any ecosystem shifts, or fluctuations between sugar kelp and turf dominated communities. National reports from the monitoring programs document annual fluctuations in the cover of sugar kelp and indicates that the variation is negatively associated to the abundance of turf (Moy et al., 2009; Fagerli et al., 2017). Based on the annual monitoring data we aimed to assess the extent of kelp recovery from turf algae dominance, and to test if the abundance of turf influences the density of sugar kelp.

The spatio-temporal variability of S. latissima cover was analyzed with a linear mixed effect model. Data from 6 m depth, from 11 m monitoring stations, was selected for the analysis. Algal cover data from 74 species/taxa were accumulated and grouped together in one generic "turf " group based on their morpho-functional traits. Three factors were included in the model: cover of turf (fixed with one level), cover of the kelp L. hyperborea (fixed with 1 level) and year (fixed with 7 levels), and station (random with 11 levels). L. hyperborea was included in the model since they are often found at the same locations and may affect the abundance of sugar kelp. All possible interactions were included in the full model and Akaike's information criterion (AIC) was used for model selection. The linear mixed effect model "nlme" (Pinheiro et al., 2012) was applied for the analysis. We also performed an ANOVA of the cover values with the two fixed factors; group (i.e., turf or kelp, where kelp included both kelp species, L. hyperborea and S. latissima) and time (i.e., year), using station as a random factor. To further explore the relationship between kelp and turf algae, we also calculated Pearson's correlation coefficient between the recorded cover of all species, grouped as turf or kelp.

The basis for evaluation of sugar kelp dispersal traits, are data from former studies of recruitment and regrowth of sugar kelp on areas far from any sugar kelp spore sources have been used. Sugar kelp colonization and recovery rate has been recorded after removal of sea urchins inside a large barren ground area (Leinaas and Christie, 1996), where bottom substrates have become available after sea urchin mortality (Rinde et al., 2014; Christie et al., 2019) and on artificial reefs (Christie, 2011).

As a basis for understanding the recruitment and spreading potential of sugar kelp, we have used the experiments performed by Andersen (2013). This study recorded S. latissima recruitment

and how recruitment rate relates to the development of fertile tissue (sori) on adult kelp throughout a reproductive period. The seasonal differences in the extent of sori, recruitment and the time-related pattern (minutes to hours) of settlement and recruitment immediately following spore release was investigated combining both field and laboratory work.

## RESULTS AND DISCUSSION

fmars-06-00072 February 20, 2019 Time: 17:19 # 5

## Spatial and Temporal Variation in Sugar Kelp Abundance

The data from the survey period 2005–2008 indicate a complex spatial distribution pattern of kelp and turf along this long coastline (**Figure 2**), where red turf algae loaded with sediments dominated on the Skagerrak coast and longer brown filamentous algae dominated on the North Sea coast (Moy et al., 2009; Moy and Christie, 2012). This discontinuous distribution was to some extent explained by the degree of exposure to waves (Bekkby and Moy, 2011; Moy and Christie, 2012), but this relationship was not consistent. Temporal changes were found all along the investigated coastline but was more pronounced at the North Sea coast (Moy and Christie, 2012). The observations from the repeated samplings formed the basis for a conceptual model of the change in macroalgal composition (see Figure 4 in Moy and Christie, 2012) illustrating the decline of sugar kelp and the seasonal fluctuation of turf from dominant in summer and reduced in winter after 2002 in Skagerrak. There, a recovery of sugar kelp occurred during 2007–2008. The decline of kelp and shift to dominance of turf in the North Sea were recorded in 2006. In 2008 sugar kelp recovered to high abundance at the expense of turf algae. The recovery of sugar kelp was minor at Skagerrak due to persistence of turf algae loaded with sediments (Moy and Christie, 2012). Moy and Christie (2012) also showed the frequent coverage of perennial understory macroalgae that may serve as a substrate for sugar kelp spores, but that do not function as proper substrate when the sporophyte grows to larger size during summer leading to dislodgement (see O'Brien and Scheibling, 2018). This was assumed to cause a seasonal loss and variation in S. latissima abundance. The bad ecological status of sugar kelp in the Hardangerfjord in the early 2000's (Moy et al., 2007) contrasts to the conditions reported later (Husa et al., 2014; Sjøtun et al., 2015), indicating a later recovery of sugar kelp in this area, in line with the conclusions of Moy and Christie (2012) from other parts of the west coast.

## Shifts in S. latissima Abundance in Skagerrak and at the North Sea Coast 2005–2017

More recent monitoring of 10 stations at the Skagerrak (southeast) coast and two stations at the North Sea (southwest) coast (Fagerli et al., 2017; Moy et al., 2017; Naustvoll et al., 2018) shows inconsistent changes in abundance of S. latissima between years and sites (**Table 1**, **Figure 3**). At the Skagerrak coast half of the stations showed improved growth and sugar kelp recovery compared to the status reported in the previous monitoring period (Moy and Christie, 2012, red dots in **Figure 1**). At most of the stations the abundance of sugar kelp has fluctuated between rare, frequent and common throughout the period of monitoring (**Table 1**), and even fluctuated between absent and to dominant at one site. There are large differences between the stations in development of sugar kelp and turf cover in the period 2009–2016 (**Figure 3**). Some stations have several alternations between absence and scattered occurrences of sugar

 are shown to the right, white means no data available.

abundance/cover

kelp (for example Brattholm). At these stations, and at Eigebrekk and partly Gleodden, turf algae and sugar kelp show a reverse pattern over time. These observations clearly document how sugar kelp can recover at earlier turf dominated sites, and that the fluctuations may occur frequently. The inconsistency between stations is too large to be explained only by temperature or other environmental factors (waves and nutrients, as was suggested by Bekkby and Moy, 2011). The shift to good condition at several stations in 2015 and 2016 should however, have been investigated closer. Taking advantage of such shifts in experimental studies of drivers are crucial to increase the understanding of the shift dynamics and the underlying mechanisms.

Also at the two stations in the North Sea (Tingsholmen and Rossholmen) the cover of sugar kelp and turf varied between the two stations and between years (**Figure 3** and **Table 1**). Although the stations are situated within short distance, the temporal variation was different: the kelp increased from "scattered" to "common" at Rossholmen in 2015, but kelp was absent at Tingsholmen in 2014, despite scattered abundance in 2013 and 2015. **Figure 3** and **Table 1** shows that the abundance of kelp and turf shifts between years in an unpredictable pattern, indicating stochastic factors driving the abundance of this presumably opportunistic species in both ecoregions (see later). Five of the 12 stations had recovery from 0 coverage of kelp (combined with high abundance of turf) in the period 2009–2016 (i.e., Robbersvik, Brattholm, Gleodden, Eigebrekk, and Tingsholmen).

The linear mixed effect model identified significant effects of both the coverage of L. hyperborea (p < 0.001) and the abundance of turf algae (p < 0.0001) on the coverage of sugar kelp. A significant negative correlation (Pearson's correlation coefficient) were found between sugar kelp and turf algae cover (-0.67, p < 0.0001), also indicating a causal negative impact of turf

Christie et al. Flips Between Kelps and Turf

on sugar kelp. The ANOVA analysis showed a significant effect of time (p < 0.0001) and of the interaction between time and group (p = 0.003), but not for the group factor alone (p = 0.07).

#### S. latissima, an Opportunistic Species?

Kelps have complex life histories where the large, sporophytes alternate with microscopic gametophytes via flagellated spores (planktonic dispersal stages). The production of spores in S. latissima is large, and kelp spores may disperse over great distances (Schiel and Foster, 2006; Cie and Edwards, 2011). Although most settle near the mother plants (Graham, 2003; Gaylord et al., 2006), large-scale oceanographic processes may serve as key drivers of connectivity between kelp populations. High reproduction, high dispersal rates, and high growth rate are typical traits of opportunistic species, as well as the short lifetime of S. latissima (Bartsch et al., 2008; Andersen et al., 2011).

Data from Leinaas and Christie (1996), presented in **Figure 4**, document a rapid recolonization of S. latissima to a small isolated island after removing sea urchins (Strongylocentrotus droebachiensis). As sugar kelp normally release spores in the winter season (Andersen, 2013) the small sporophytes start to grow in the spring and may be observed in early summer. The average density of small sporophytes was more than 470 per m<sup>2</sup> (**Figure 4**, see also Leinaas and Christie, 1996). This finding is supported by more recent recordings of sugar kelp beds in areas where sea urchin densities are decreasing in northern Norway (Rinde et al., 2014; Christie et al., 2019).

A similar pattern of rapid recolonization of sugar kelp to available substrate was recorded at 12 artificial reefs deployed in an area dominated by sea urchins, with no kelp observed in the area. In July, 2006, these large artificial reefs made of concrete and plastic tubes were launched at about 10 m depth at Hammerfest (Barents Sea Norway, Christie, 2011). The recolonization pattern was recorded by diving and photo once or twice a year, for 4 years. After 3 months (October, 2006), mainly small filamentous algae and tubeworms had settled and could be identified at the reef structures. After the following winter and the recruitment season of sugar kelp, the first small kelp sporophytes were observed in April 2007. In July 2007 larger sugar kelps dominated, and the density of sugar kelp was approximately 30 individuals per 2.5 m of the plastic (PVC) tubes of the reefs (roughly about 60 per m<sup>2</sup> ). The outer surface of the reefs was densely colonized with sugar kelp for almost 3 years (from autumn 2007 and until the summer of 2010) before sea urchins invaded the reefs and overgrazed the kelps.

Rapid recruitment of kelp on artificial substrate excludes the existence of a dormant spore banks on the substrate (Hoffmann and Santelices, 1991) as source of the spores. The rapid recruitment and high colonization rate of sugar kelp on the artificial reef structures clearly document that sugar kelp has a great ability to disperse, colonize, and recover kelp forests on available substrate if the conditions are suitable, even when the substrate is far away from a spore source population.

The study by Andersen (2013) revealed synchronous development of fertile tissue, high concentrations of viable spores, consistent settlement patterns and a relatively steady in situ recruitment on clean substrate throughout the winter

months. Connectivity between kelp populations is reinforced by reproductive synchrony because higher densities of spores in the currents increase the probability of long-distance dispersal (Reed et al., 1997). The seasonal development and demise of visible sori in S. latissima are processes that largely overlap along the south coast of Norway (Andersen et al., 2011; Andersen, 2013). The tight link between the timing of recruitment and these patterns shown by Andersen (2013) support the notion that the potential for connectivity between sugar kelp populations in Norway is high. This may enable forest regeneration by natural recruitment from distant remnant source populations. In fact, kelp recolonization of barren grounds and colonization of artificial reefs far from source populations is consistent with long-distance dispersal of S. latissima.

### Regime Shifts or Flips Back and Forth?

It is now 16 years since the first report on turf algae replacing sugar kelp S. latissima (Moy et al., 2009), then indicating a large spatial scale regime shift along the Norwegian coast. Frigstad et al. (2013) suggested the period close to the millennium shift to be a period of regime shift also in the pelagic ecosystems, which coincides with the shift from sugar kelp to turf. Although there have been considerable reductions in the abundance and spatial distribution of sugar kelp in southern Norway and these forests have been classified as endangered (Skagerrak) and vulnerable (North Sea) on the Norwegian red list for ecosystems and habitat types (Lindgaard and Henriksen, 2011), the species still occur and even recover in large areas along these coastlines.

The data presented here show that sugar kelp may quickly colonize and recover in areas taken over and dominated by turf algae. The sugar kelp, with its high recruitment potential and efficient dispersal of spores during winter (when ephemeral algae are reduced) have a large opportunity to seed new sporophytes and to grow dense populations on available substrate each spring. This may become a new long-lasting sugar kelp bed, or a bed of short duration (months, few years) depending on the amount of epiphytic growth, kelp mortality, and turf algae formation. The kelp recovery may not only vary on a temporal scale, but can also be discontinuous on a spatial scale (see **Table 1**). Our data

show that seafloor areas covered by turf and loaded with sediments may persist, and local regime shifts may occur, but also that the positive feedback mechanisms of turf (see Filbee-Dexter and Wernberg, 2018) may be challenged. It is not clear which disturbance factors that occur mainly during the winter season and that make the substrate available for new kelp spore settlement. Many turf /ephemeral algae die and disappear during the autumn/winter season (e.g., Moy and Christie, 2012), and rough winter storms and whiplash effects of remaining kelps may sweep away both remaining turf and sediments (Russell, 2007).

In a scenario of further ocean warming, increasing eutrophication, and water darkening (Aksnes et al., 2009; Moy et al., 2009; Frigstad et al., 2013), the conditions will likely, gradually and additionally favor the turf at the expense of S. latissima. How increased ocean temperatures will work together with acidification and high levels of nutrient to impact macroalgae (Connell et al., 2008; Gorman et al., 2009; Falkenberg et al., 2013) is not fully understood. Even if temperature is more favorable for the sugar kelps in deeper parts of their depth distribution, Andersen et al. (in press) described a scenario with decreased light and increased respiration to squeeze the kelps' vertical distribution to shallow areas, leaving reduced seafloor areas as suitable. So far exposure to critical surface temperatures, reduced light, and increased competition from epiphytic and understory turf growth, sugar kelps remains and have been able to recover. There are still healthy sugar kelps close to the surface in Oslofjord at 10. August, 2018 (own observations) although 2018 has been the "warmest summer ever" in southern Norway, with more than 2 months of surface water temperature at or exceeding the critical level of this species (Luning, 1984; Müller et al., 2009) (temperatures higher than 20oC from end of May to early August, and even longer periods at 22–23oC, shown by regular temperature measurements at NIVA's research station).

### CONCLUSION

As sugar kelp have a potential of wide distribution along large parts of the coastline in southern Norway, as well as covering a depth range of 0–25 m, small scale mitigation actions will likely have limited effect. The chance of restored kelps to survive will depend on the growth condition of turfs and epiphytes. On the other hand, the chance of natural restoration of kelps will also depend on available substrate. In years with good conditions for kelp dispersal and restoration, as indicated from **Table 1**, the natural recolonization of sugar kelp may be much more efficient than any local mitigation action. Kraufvelin et al. (2006) showed restoration of perennial algae when nutrient supplies were reduced, and Lefcheck et al. (2018) presented how long-term nutrient reductions improve large coastal regions. Improving

#### REFERENCES

Aksnes, D. L., Dupont, N., Staby, A., Fiksen, Ø, Kaartvedt, S., and Aure, J. (2009). Coastal water darkening and implications for mesopelagic regime shifts in Norwegian fjords. Mar. Ecol. Prog. Ser. 387, 39–49. doi: 10.3354/meps 08120

coastal water quality (eutrophication, browning) will probably be the most important mitigation action. However, if larger areas are totally depleted and a regime shift to turf seems irrevocable, adult sugar kelps may be transplanted to ensure a spore source in the area to enable kelp recovery if conditions seems satisfactory. This paper highlights a complex spatial and temporal distribution pattern between sugar kelp and turf algae, and do not speculate on physical, chemical and biological factors that contribute to create these patterns. Further multifaceted research projects are needed to reveal the causes to the complex patters of kelp-turf distribution presented here.

Both in northern Norway, at the west coasts of South and North America, and at the west African coast kelp beds are persistent and no turf are reported to disturb this persistence. When it comes to NE America, Australia, and Europe, regime shifts from kelp to turf have been reported, also in areas where S. latissima is the dominating kelp (Filbee-Dexter et al., 2016; Filbee-Dexter and Wernberg, 2018). In Norway the sugar kelp has been reported to be far more efficient when it comes to dispersal and colonization than Laminaria spp (Leinaas and Christie, 1996), and may by its opportunistic traits be more able to quickly take advantage of any space available. Thus, our data from the S. latissima areas of the south coast of Norway differ from the systems where persistent regime shifts from kelps to turf occur, although turf seems to persist at some areas also in southern Norway.

### AUTHOR CONTRIBUTIONS

All authors discussed the scope, agreed on the hypothesis and aims of the paper, and contributed to writing the manuscript. HG, TB, CF, and JG produced the most recent data, **Figure 3**, and **Table 1**. HC, GA, CF, JG, and ER have many years of experience in the field of collecting data used in this manuscript.

### FUNDING

Most studies were funded by the Norwegian Environment Agency and the data can be used without conflict of interest. Some of the data were produced with funding from the Research Council of Norway, with extra support from NIVA.

#### ACKNOWLEDGMENTS

We are grateful for the effort of Frithjof Moy (former NIVA, now Institute of Marine Research) and Lise Ann Tveiten (NIVA) during the early years of the Sugar kelp project.

Andersen, G. S. (2013). Patterns of Saccharina latissima recruitment. PLoS One 8:e81092. doi: 10.1371/journal.pone.0081092

Andersen, G. S., Steen, H., Moy, F., Christie, H., and Fredriksen, S. (2011). Seasonal patterns of sporophyte growth, fertility, fouling and mortality of Saccharina latissima in Skagerrak, Norway – implications for re-forestation. J. Mar. Biol. 2011, 1–8. doi: 10.1155/2011/690375



and ecosystem consequences. Limnol. Oceanogr. 42, 1105–1118. doi: 10.4319/lo.1997.42.5\_part\_2.1105

Waycott, M., Duarte, C. M., Carruthers, T. J. B., Orth, R. J., Dennison, W. C., Olyarnik, S., et al. (2009). Accelerating loss of seagrass across the globe threatens coastal ecosystems. Proc. Natl. Acad. Sci. U.S.A. 106, 12377–12381. doi: 10.1073/ pnas.0905620106

**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Christie, Andersen, Bekkby, Fagerli, Gitmark, Gundersen and Rinde. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Reevaluating the Role of Organic Matter Sources for Coastal Eutrophication, Oligotrophication, and Ecosystem Health

Anne Deininger1,2,3 \* and Helene Frigstad2,3

<sup>1</sup> Department of Natural Sciences, University of Agder, Kristiansand, Norway, <sup>2</sup> Norwegian Institute for Water Research, Grimstad, Norway, <sup>3</sup> Centre for Coastal Research, University of Agder, Kristiansand, Norway

Organic matter (OM) in aquatic systems is either produced internally (autochthonous OM) or delivered from the terrestrial environment (ter-OM). For eutrophication (or the reverse – oligotrophication), the amount of autochthonous OM plays a key role for coastal ecosystem health. However, the influence of ter-OM on eutrophication or oligotrophication processes of coastal ecosystems is largely unclear. Therefore, ter-OM, or ter-OM proxies are currently not included in most policies or monitoring programs on eutrophication. Nevertheless, ter-OM is increasingly recognized as a strong driver of aquatic productivity: By influencing underwater light conditions and nutrient- and carbon availability, increased ter-OM input may shift systems from autotrophic toward heterotrophic production, but also alter the interactions between benthic, and pelagic habitats. Thus, by changing baseline conditions in coastal zones, ongoing, and predicted changes in inputs of ter-OM due to climate change (e.g., in precipitation) and anthropogenic activities (e.g., reduced sulfate deposition, damming, and coastal erosion) may strongly modify eutrophication symptoms within affected ecosystems, but also hinder recovery from eutrophication following a reduction in nutrient loadings (i.e., oligotrophication). In this review, we aim to shed light upon the role of ter-OM for coastal eutrophication and oligotrophication processes and ecosystem health. Specifically, we (1) discuss the theoretical interactions between ter-OM and eutrophication and oligotrophication processes in coastal waters, (2) present global case studies where altered ter-OM supply to coastal ecosystems has shifted baseline conditions, with implications for eutrophication and oligotrophication processes, and (3) provide an outlook and recommendations for the future management of coastal zones given changes in ter-OM input. We conclude that it is essential to include and target all OM sources (i.e., also ter-OM) in monitoring programs to better understand the consequences of both eutrophication and oligotrophication processes on coastal ecosystems. Our review strongly urges to include ter-OM, or ter-OM proxies in eutrophication monitoring, and policies to safeguard coastal ecosystem health also under changing climatic conditions and globally increasing anthropogenic perturbations of coastal ecosystems.

Keywords: browning, coastal darkening, dissolved organic carbon, eutrophication, nutrients, organic carbon, terrestrial organic matter

Edited by:

Marianne Holmer, University of Southern Denmark, Denmark

#### Reviewed by:

Rodrigo Riera, Catholic University of the Most Holy Conception, Chile Cesar De Castro Martins, Federal University of Paraná, Brazil

> \*Correspondence: Anne Deininger anne.deininger@niva.no

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 October 2018 Accepted: 03 April 2019 Published: 25 April 2019

#### Citation:

Deininger A and Frigstad H (2019) Reevaluating the Role of Organic Matter Sources for Coastal Eutrophication, Oligotrophication, and Ecosystem Health. Front. Mar. Sci. 6:210. doi: 10.3389/fmars.2019.00210

## INTRODUCTION

fmars-06-00210 April 23, 2019 Time: 19:29 # 2

Eutrophication has remained a major threat to coastal ecosystems globally throughout the past decades and is today acknowledged as a complex process caused by increased inputs of various nutrients (Nixon, 1995; Boesch, 2002; Seitzinger et al., 2010). Under natural conditions, nutrient runoff from land plays a crucial role in fueling the productivity of coastal, and estuarine systems and by providing key ecosystem services (Nixon, 1988; Barbier et al., 2011; Cloern and Jassby, 2012; Cloern et al., 2014). However, excess supply of nutrients may overstimulate internal production of organic matter (autochthonous OM), resulting in reduced water quality and ecosystem health by causing symptoms such as increased growth of macrophytes and harmful algae, as well as a depletion in bottom water oxygen, and periodic fish kills (Nixon, 1995; Schindler, 2006; Diaz and Rosenberg, 2008; Conley et al., 2009).

Historically, the main nutrients within the focus of eutrophication research and policy have been nitrogen (N) and phosphorus (P) (Schindler, 2006). The increased input of these elements, especially their inorganic forms, can be linked directly to anthropogenic activities within the catchment of a waterbody (e.g., agriculture, fertilization, and wastewater) or from deposition of long-range transported air pollutants (e.g., industrial or agricultural emissions) (Nixon, 1995; Diaz and Rosenberg, 2008; Conley et al., 2009). Therefore, reducing inorganic nutrient loads to coastal zones has been the main strategy in national and international policies to reduce eutrophication (Boesch, 2002; Carstensen et al., 2006; Duarte et al., 2009). The resulting reduction in nutrient loading to coastal systems has been termed "oligotrophication," indicating a recovery from eutrophication (Nixon, 2009).

Besides inorganic nutrients, terrestrial organic matter (ter-OM) is an important component of riverine runoff that has been gaining increasing attention in recent years (Solomon et al., 2015; Ward et al., 2017; Bianchi et al., 2018). Historically, ter-OM was assumed to be stable and relatively resistant to transformations within coastal and estuarine systems (Carlson and Hansell, 2015). However, there was an expressed discrepancy between the high amounts of carbon being transported from land to coasts, compared to the relatively small amounts of terrestrial carbon detectable in oceanic waters and sediments (Hedges et al., 1997; Bianchi, 2011). Today, there is strong evidence that significant transformations of both particulate and dissolved ter-OM fractions occur in coastal and estuarine waters, as well as along their transport downstream (Coble, 2007; Osburn and Bianchi, 2016; Massicotte et al., 2017). These transformational processes are photodegradation, biodegradation and flocculation (causing increased sedimentation), and the relative importance of these processes will vary along the salinity gradient (Weyhenmeyer et al., 2012; Riedel et al., 2016; Massicotte et al., 2017; Bianchi et al., 2018).

Terrestrial organic matter export to coastal ecosystems has been increasing in both boreal (**Figures 1A,B**; Monteith et al., 2007; de Wit et al., 2016), and arctic regions (**Figures 1C,D**; Dickson et al., 2000; Peterson et al., 2002; Carmack et al., 2015). Climate drivers such as precipitation changes and increased

FIGURE 1 | Examples of altered ter-OM loads to coastal ecosystems in (A,B) the boreal zone, due to increased precipitation with potential effects on kelp forests (Skagerrak, Norway), (C,D) the arctic zone, due to thawing permafrost and melting glaciers (Longyearbyen, Spitzbergen), (E,F) tropical zones, due to episodic extreme events with potential effects on e.g., coral reefs (Australia), and (G,H) global coastal zones due to direct human impacts on land-ocean interactions, such as structural modifications of shorelines (Pescadores Islands, Taiwan), or dam construction (Gordon Dam, Australia). Photo courtesy: (A,C,D,F,G) A. Deininger, (B) J. Gitmark, (E) J. Schmaltz (NASA), and (H) L. Lagrainge.

temperatures are believed to be important for both regions, however, with decreasing sulfate depositions (i.e., acid rain) as an additional driver of ter-OM input in boreal regions (resulting in the often termed "browning" of boreal freshwaters). In addition, structural changes to hydrological cycles (dams, land use), coastal erosion, sea level rise and urbanization (**Figures 1G,H**), are causing altered (increasing or decreasing) inputs of ter-OM globally, with impacts on ecosystems already under significant stress, such as tropical coral reefs or seagrass meadows (**Figures 1E,F**; Nyström et al., 2000; Butler et al., 2015; Selvaraj et al., 2015; Regier and Jaffé, 2016).

How and to what extent changed input of ter-OM might influence coastal zones simultaneously receiving excess supply

of inorganic nutrients (i.e., eutrophication) is currently an unanswered question (Nixon, 2009; Ferreira et al., 2011; Newton et al., 2012), especially since ter-OM or ter-OM proxies are not considered in most monitoring programs or policies aimed to address eutrophication. However, there are indications that increased ter-OM may lead to symptoms typically associated with eutrophication, such as net heterotrophy, anoxic bottom water zones or a decline in macrophyte cover (Andersson et al., 2013). This overlap in symptoms might be problematic, since it may lead to a misdiagnosis of the causes based on the symptoms, especially since information regarding changes in ter-OM input in monitoring programs is often lacking. Duarte et al. (2009) describe how several coastal ecosystems have not returned to original conditions or reference status following nutrient reductions to alleviate eutrophication effects, which the authors attribute to changed baseline conditions of the ecosystems over time. Increased ter-OM inputs to coastal ecosystems will affect baseline conditions, together with other natural, and anthropogenic pressures. It is therefore possible that altered runoff of ter-OM might prevent improvement in ecosystem health of systems undergoing oligotrophication (i.e., decreased inorganic nutrient inputs).

With this review, we aim to reevaluate the role of ter-OM for coastal eutrophication, oligotrophication and ecosystem health. Firstly, we will discuss potential theoretical consequences of altered ter-OM input on coastal ecosystems undergoing eutrophication or oligotrophication, by combining existing knowledge collected within the fields of ter-OM and eutrophication research. Secondly, we will provide case studies where interactions of ter-OM with eutrophication or oligotrophication are already visible, and expected to increase with future human perturbations or climate impacts on these ecosystems. In sum, our paper seeks to improve the knowledge on the interactions between eutrophication and altered ter-OM runoff to coastal ecosystems, and to strengthen the collaboration between these two fields of research. We advice current eutrophication programs and policies to include all OM sources (i.e., also ter-OM) in order to improve our understanding of the interactive effects of human pertubations to coastal ecosystems and to secure coastal ecosystem health also in the face of accelerating human pressures and climate change.

## THEORETICAL INTERACTIONS OF ter-OM WITH EUTROPHICATION AND OLIGOTROPHICATION

Terrestrial organic matter may interact with its environment in different ways, depending largely on its chemical and optical properties, but also abiotic (e.g., temperature, salinity), and biotic conditions (e.g., bacterial community composition) (Hedges et al., 1997; Bianchi et al., 2018). When moving from land to sea, ter-OM compounds may undergo substantial modifications along the whole gradient from the upstream catchment to the downstream coastal region (Massicotte et al., 2017). Understanding the different processes involved, as well as the various consequences for ecosystem functioning is currently a vivid field of research (Massicotte et al., 2017; Bianchi et al., 2018; Painter et al., 2018).

Generally, the different ter-OM compounds may be transformed via three main types of reactions: (1) biodegradation, (2) photochemical oxidation (i.e., photodegradation), and (3) flocculation (causing increased sedimentation) (Weyhenmeyer et al., 2012; Riedel et al., 2016; Massicotte et al., 2017; Bianchi et al., 2018). Below, we discuss the various physical and chemical characteristics of the carbon, nutrient and micronutrient fractions.

The chromophoric fraction (CDOM) of the dissolved component of ter-OM, or CDOM, is a major driver of the underwater light climate and photochemistry (**Figure 2A**; Coble, 2007; Kirk, 2010; Osburn et al., 2016). By absorbing light, CDOM may either cause protection of underwater organisms against harmful ultraviolet (UV) radiation, but also induce lightlimitation for aquatic primary producers, causing a decrease in autotrophic production after a certain threshold (**Figure 2B**; Falkowski and Laroche, 1991; Wikner and Andersson, 2012; Thrane et al., 2014). Additionally, the coastal light environment can also be strongly impacted by larger sized particulate-OM, as well as inorganic particles entering with the terrestrial runoff. These particles can not only cause light absorption, but also shading and scattering, and could potentially have an even stronger impact on the coastal light environment than CDOM, depending on respective concentrations, size and chemical properties (Sholkovitz, 1976; Kirk, 2010; Bainbridge et al., 2018; Margvelashvili et al., 2018). Through its UV-absorption potential, ter-OM may also reduce the amount of light reaching

benthic environments and may, depending on depth, result in a shift from benthic to pelagic primary production (Jones, 1992; Fabricius, 2005; Karlsson et al., 2009). Further, shading caused by ter-OM might inhibit growth-stimulating nutrient effects on autotrophs, and lead to an additional advantage for heterotrophs though reduced niche competition (Wikner and Andersson, 2012; Andersson et al., 2013). This resulting release from nutrient competition with autotrophs, may enhance nutrient availability for bacterial production and heterotrophic processes in both pelagic and benthic habitats (**Figure 2C**; Jones, 1992; Jansson et al., 2003). Decreased oxygen production by autotrophs and increased oxygen demand by heterotrophs (i.e., respiration) might shift coastal ecosystems toward net heterotrophy, with increased release of greenhouse gases (e.g., CO2, N2O) and a potential increase in coastal dead zones (**Figure 2D**; Wikner and Andersson, 2012; Andersson et al., 2013; Lapierre et al., 2013). Additionally, changes in CDOM might induce shifts in benthic and pelagic community composition toward mixo- and heterotrophic species due to altered baseline light conditions in these environments (Fabricius, 2005; Moy and Christie, 2012; Deininger et al., 2017). Another consequence of a shift from autotrophic to heterotrophic production might be a reduction in food quantity, quality, and food web efficiency to higher trophic levels: bacerial-based food chains are typically lower in essential polyunsaturated fatty acids (PUFAs) important for consumers, as well as longer and therefore less efficient in energy transfer than phytoplankton-based food chains (Nixon, 1988; Müller-Navarra et al., 2000; Berglund et al., 2007). Lastly, reduced light levels caused by higher ter-OM conditions may hinder re-establishment of macrophyte cover, coral survival, or original phytoplankton communities following oligotrophication of coastal ecosystems (Fabricius, 2005; Karlsson et al., 2009; Filbee-Dexter and Wernberg, 2018).

The carbon compounds of ter-OM (i.e., dissolved organic carbon, DOC) may serve as a major energy source for hetero- and mixotrophic producers and stimulate heterotrophic production, until other factors will further limit growth (e.g., temperature, predation) (**Figures 2A,C**; Stepanauskas et al., 2002; Berglund et al., 2007). However, the bioavailability of DOC has been shown to strongly depend on molecule size and molecule structure of the carbon component, as well as the present bacterial community composition (Kamjunke et al., 2017; Creed et al., 2018). Nevertheless, it is likely that heterotrophic production in general will be stimulated by ter-OM inputs, supporting higher levels of production at high nutrient inputs (i.e., **Figure 2C**, eutrophic scenario), than at relatively lower nutrient inputs (**Figure 2C**, oligotrophic scenario). Further, this response will depend on the initial trophic state of the system and the duration of the nutrient pulses (hours, days, weeks), with different functional groups dominating depending on their adaptation strategies (Berglund et al., 2007; Deininger et al., 2016, 2017).

Besides the carbon compounds, ter-OM also contains complexed elements, such as macronutrients, micronutrients, and contaminants. The macronutrient compounds of ter-OM consist mainly of N and P (i.e., organic N, P; **Figure 2A**). Also for these ter-DOM compounds, bioavailability may be highly variable irrespective of total concentrations (i.e., TN, TP), making measurements of TN and TP potentially poor indicators for predicting productivity responses (Bergström, 2010; Berggren et al., 2015; Soares et al., 2017). However, in contrast to carbon compounds, macronutrients can support auto-, mixoand heterotrophic processes, depending on bioavailability, competition between species, as well as other abiotic conditions such as light availability (**Figures 2B–D**; Berggren et al., 2015; Soares et al., 2017). Due to these multiple interactions, it is difficult to predict, which productivity processes organic N and P components might ultimately influence. Nutrient effects might strongly depend on initial baseline nutrient conditions: In oligotrophic and ter-OM poor systems, it has been hypothesized that increasing ter-OM concentrations might initially increase autotrophic production until systems eventually become increasingly light-limited (e.g., ∼5 mg L−<sup>1</sup> in boreal and arctic lakes) (**Figure 2B**, oligotrophic scenario) (Seekell et al., 2015). A similar response might not be expected in eutrophic systems, where ter-OM-attributed nutrient effects might comparably be diluted in the overall high-nutrient baseline conditions. Additionally, self-shading effects might occur through increased autochthonous OM production (i.e., phytoplankton blooms), which can substantially decrease light availability (**Figure 2B**, eutrophic scenario), with increasing habitat suitability for heterotrophs (**Figures 2C,D**). Also micronutrient compounds, and especially iron (Fe), may strongly stimulate coastal basal production, especially heterotrophicand N-fixing bacteria, which have been shown to be more efficient in accessing complexed Fe than autotrophs (Hyenstrand et al., 2000; Sorichetti et al., 2014, 2016). In coastal zones, the concentration of Fe is comparably higher than in open oceans due to the large shelf sources (Johnson et al., 1997, 1999). Fe-DOM complexes are photoreactive, however given its strong positive charge, Fe is typically strongly bound, and precipitates fast when dissociated from DOM.

Lastly, ter-OM may also affect contaminant dynamics in coastal zones, as many contaminants are also positively charged, such as mercury (Hg), either to its inorganic [Hg(II)] or its organic form (methyl-Hg) (Haitzer et al., 2003; Ravichandran, 2004). The latter is a highly potent neurotoxin, and biomagnifies in aquatic food webs (Morel et al., 1998; Ullrich et al., 2001; Schartup et al., 2017). Thus, increased inflow of ter-OM might ultimately result in increased inflow of attributed toxic contaminants, such as methyl-Hg, threatening both ecosystem health and water quality. Interaction effects with eutro- or oligotrophication are currently unclear but might also depend largely on the ter-OM-induced shifts from autotrophic toward heterotrophic production, which might increase the direct or indirect consumption of ter-OM by aquatic consumers, and therefore the attributed methyl-Hg uptake (Karlsson et al., 2012; Berggren et al., 2015; Schartup et al., 2017).

In summary, the effect of changing inorganic nutrient supply (i.e., eutro-, oligotrophication) on coastal ecosystems might systematically be altered by shifted baseline conditions caused by changes in ter-OM input. Ter-OM may alter light quantity and quality for autotrophic producers through its chromophoric compounds (**Figures 2A,B**), stimulate growth of different producers and consumers via its carbon, macronutrient (N, P)

(**Figures 2A–D**) and micronutrient (mainly iron, Fe) fractions, and/or induce toxicity through its contaminant compounds (e.g., mercury) (Stepanauskas et al., 2002; Fabricius, 2005; Creed et al., 2018). Importantly, to understand the interaction of ter-OM with its surrounding environment (and eutro-, oligotrophication), both chemical and optical properties, but also the reactivity of the material has to be considered (Hedges et al., 1997; Osburn and Bianchi, 2016; Bianchi et al., 2018).

## CASE STUDIES OF ALTERED ter-OM LOADS TO COASTAL ECOSYSTEMS

To date, most mechanistic understanding related to ter-OM effects on aquatic ecosystems comes from systems where aquatic ter-OM concentrations are amongst the highest globally, namely boreal freshwaters and the Baltic Sea (Sobek et al., 2007; Kirk, 2010). Here, ter-OM has been recognized for decades as a fundamental driver of aquatic ecosystem productivity, ecosystem structure, functioning, and greenhouse gas emission through its light-absorbing, nutrient- and carbon-providing properties (i.e., **Figure 2**; Solomon et al., 2015; Andersson et al., 2018; Creed et al., 2018). However, there is increasing evidence that there are similar processes involved for ter-OM processing and cycling when entering systems with comparably lower concentrations than above, such as the Amazon or the Mississippi river to ocean continuum, the Great Barrier reef, and the Arctic Ocean (**Figure 1**) (Fabricius, 2005; Seidel et al., 2015; Duan et al., 2017; Tanski et al., 2017). Below, we will present case studies from ecosystems where increased ter-OM has already changed biotic and abiotic conditions (e.g., Baltic Sea), to ecosystems where ter-OM has recently started to increase (boreal, arctic, tropical, and global patterns), to cases where ter-OM input is abrupt, extreme, short-term and predicted to increase in frequency (globally, episodic events).

### Pilot Case: Baltic Sea

The Baltic sea is the world's largest semi-enclosed sea. This setting might be the reason why eutrophication has drastically affected its ecosystem health, making the Baltic a pilot study for both research and implementing policies on eutrophication (Larsson et al., 1985; Heisler et al., 2008; Andersen et al., 2011; Paasche et al., 2015). In addition, the nature of the catchment (i.e., high amounts of peatland), as well as long water residence times might be additional contributing factors to why runoff of ter-OM has already led to comparably high ter-OM concentrations and ecosystem effects compared to other coastal ecosystems with higher water mass exchange (Søndergaard and Thomas, 2004; Andersson et al., 2013). Recently, interactions between eutrophication and ter-OM in the Baltic have received increasing attention (Wikner and Andersson, 2012; Andersson et al., 2013). In a mesocosm experiment, Andersson et al. (2013) showed that additions of inorganic nutrients alone resulted in typical eutrophication symptoms, whereas additional supply of humic carbon resulted in increased bacterial production and increased net heterotrophy (**Figures 2C,D**). This agrees with an analysis of long-term monitoring data of the northern Baltic Sea, showing that increased riverine discharge over a 40-year period correlated with increased bacterial production and net heterotrophy (Wikner and Andersson, 2012). However, the trend in heterotrophy was not only driven by increased bacterial production, but largely also (71%) by induced light limitation, and a reduction in phytoplankton productivity (**Figure 2B**). These findings indicate that ter-OM plays an important role as a regulator of coastal productivity, but more studies are needed to confirm these observed patterns both at high- (e.g., northern Baltic) and lower ter-OM (southern Baltic) concentrations.

#### Other Boreal Cases

As a result of the increased ter-OM input and consequent browning of boreal freshwaters, most cases of increased downstream ter-OM runoff to coastal zones are reported for the boreal zone. Besides the Baltic, other European cases of potentially increased inputs of ter-OM have been reported for the Norwegian coast, e.g., the "coastal darkening" of the Skagerrak and the North Sea (**Figure 1A**; Aksnes et al., 2009; Dupont and Aksnes, 2013; Frigstad et al., 2013). Here, increases in ter-OM are hypothesized to be connected both to increased discharge of rivers flowing into the Skagerrak (i.e., increasing the loads of OM) (Skarbøvik et al., 2017), but also an increase in the concentrations of the various components of OM per se (de Wit et al., 2016). For example, Frigstad et al. (2013) found a decrease in the inorganic nutrient fraction for the coastal Skagerrak area (southern Norway), while there was an increase in the dissolved and particulate organic fractions of both carbon and nutrients, believed to be related to increased loading of ter-OM (**Figure 2A**). In the same area, a largescale shift from sugar kelp (Saccharina latissima, **Figure 1B**) to ephemeral algae occurred around 2002 (Moy and Christie, 2012), connected with high summer temperatures and siltation from river runoff. Over the same time-period, the soft-bottom fauna in Skagerrak has shown improved status believed to be related to an improvement in eutrophication status (i.e., oligotrophication) (Trannum et al., 2018). However, to be able to disentangle the effects of eutrophication, oligotrophication and increased ter-OM in this region, it is important to have consistent ecosystem monitoring that also includes measurements relevant for detecting changes in ter-OM, such as DOC, CDOM, and light.

## Arctic- and Subarctic Cases

Increases in ter-OM have been reported for many coastal shelf areas of the Arctic (Dickson et al., 2000; Peterson et al., 2002; ACIA, 2004; Kaiser et al., 2017). In this region, increased precipitation driven by climate change, and thawing of permafrost and glaciers have already resulted in increased riverine runoff of ter-OM to arctic and subarctic coasts across Scandinavia, Russia, and North America (**Figures 1C,D**; Dickson et al., 2000; Peterson et al., 2002; ACIA, 2004; Kaiser et al., 2017). These trends are likely to continue, given the predicted increases in both precipitation and temperature, further inducing permafrost thawing, and ter-OM runoff (Froese et al., 2008; Schuur et al., 2008; Carmack et al., 2016). Upstream arctic and subarctic catchments harbor up to 50% of the worlds soil carbon, wherefore increased runoff will potentially have large

consequences for carbon stocks and cycling (Gorham, 1991; Waelbroeck et al., 1997; Kaiser et al., 2017). Thus, runoff from both permafrost and glaciers can be expected to have large consequences for downstream coastal ecosystems and ecosystem baseline conditions (**Figures 2B–D**; Dunton et al., 2012; Lydersen et al., 2014; Kaiser et al., 2017). Although eutrophication is not a major problem in the Arctic, including ter-OM measurements in current monitoring programs and research expeditions will greatly benefit basic research around the interactions effects of ter-OM and eutrophication, since arctic shelf seas may serve as unique reference systems and provide important insights into current and potential future drivers of primary production (Mann et al., 2016). Additionally, studying ter-OM in the Arctic will greatly improve current understanding behind the triggers of ter-OM runoff, but also interactions with e.g., coastal freshening and the importance of ter-OM quality and source (e.g., boreal soils vs. glacier runoff) (Blair and Aller, 2012).

### Tropical Cases

Human activities have been causing large changes to landocean interactions in tropical regions by affecting upstream catchments properties (details see section "Global Cases Caused by Structural Modifications"), altering floodplain structures and inundation patterns (e.g., for rice paddies, shrimp farms), as well as by potentially affecting weather patterns leading to an increasing frequency of extreme precipitation events (i.e., hurricanes and cyclones causing episodic input of ter-OM) (Nyström et al., 2000; Butler et al., 2015; Selvaraj et al., 2015; Regier and Jaffé, 2016). Generally, tropical landocean transects are important hot-spots for biogeochemical ter-OM cycling due to several reasons: their comparably high temperatures, their seasonal inundations over large areas, and their diversity in catchment characteristics, ranging from tropical forests to dryland and from pristine areas to heavily human perturbed areas (dams, river withdraw, and channel alterations) (Burns et al., 2008; Hamilton, 2010; Seidel et al., 2015). Much is still unclear about the ecological equilibrium within these diverse systems, especially within benthic environments, such as coral reefs or macrophyte meadows which are currently decreasing globally (Nyström et al., 2000; Comte and Pendleton, 2018; Filbee-Dexter and Wernberg, 2018). Findings management strategies for these threatened ecosystems is under strong investigation (Pittman and Armitage, 2016; Regier and Jaffé, 2016; Comte and Pendleton, 2018). For example, recent studies from Australia have indicated that increased input of ter-OM following episodic rain events and attributed increases in nutrient concentrations, turbidity, and sedimentation may strongly change baseline conditions important for the growth, survival, but also reproduction and recruitment of corals, and organisms living within coral associated habitats (**Figures 1E,F**; Nyström et al., 2000; Fabricius, 2005; Regier and Jaffé, 2016). However, the current state of knowledge does not allow to identify the specific drivers influencing tropical coastal ecosystems, largely due to missing monitoring data of comparable parameters along the landocean continuum, but also the complexity of processes along the gradient (Burns et al., 2008; Hamilton, 2010; Seidel et al., 2015; Regier and Jaffé, 2016). Similarly, it is difficult to predict the effects of increased temperature due to global warming on these highly dynamic ecosystem processes (Hamilton, 2010). However, including ter-OM or ter-OM proxies, in addition to eutrophication proxies, in the monitoring of tropical coastal waters would aid in understanding these hotspots, but also hot moments (i.e., episodic events) of ter-OM input and cycling in these rapidly changing ecosystems (Hamilton, 2010; Seidel et al., 2015).

## Global Cases Caused by Structural Modifications

Terrestrial organic matter input may also be systematically altered directly through human activities. For example, decreased river runoff may in some cases be directly attributed to anthropogenic activities within the river catchment. These activities may be the building of dams and reservoirs, but also land use changes (agriculture, forestry), urbanization, structural changes to shoreline or the extraction of water for irrigation purposes (**Figures 1G,H**) (Milliman et al., 2008; García-Ruiz et al., 2011; Su et al., 2018). Within midlatitude rivers for example, river runoff has decreased by ∼60% during the last half of the 20th century (Milliman et al., 2008). As discussed above, changing upstream runoff has large consequences for downstream ecosystems, by changing the transport, transformation, and delivery of material (both ter-OM and inorganic nutrients) into coastal zones (Seitzinger et al., 2010). Given the ongoing and predicted demographic pressures in many water-scarce regions of the world (Africa, Asia, Australia, Mediterranean, southern Mexico, and northern Brazil), the decrease in river runoff is likely to accelerate and deserve careful monitoring of downstream ecosystems (Milliman et al., 2008; García-Ruiz et al., 2011; Donnelly et al., 2017).

Furthermore, coastal erosion is increasing globally, and is currently impacting at least 75% of the world's shorelines (Rangel-Buitrago et al., 2018b). Generally, erosion from landmasses (both ter-OM and inorganic material) into the ocean is a natural phenomenon that can occur both as a slow pervasive process, but also rapidly during extreme events (Larson and Kraus, 1995; Rangel-Buitrago et al., 2018b). However, the ongoing and predicted human urbanization of coastal zones, as well as the various anthropogenic effects on the climate (temperature, precipitation, and storm events) increases the rates of coastal erosion, with large ecological, economic, and societal impacts (Pranzini, 2018; Rangel-Buitrago et al., 2018a; Williams et al., 2018). The ongoing increase in sea level is also a significant driver of coastal erosion (Toimil et al., 2017; Rangel-Buitrago et al., 2018a).

Currently, none of the above-mentioned human impacts on ter-OM transport to coastal zones have been discussed in relation to coastal eutrophication or oligotrophication, especially not along the aquatic continuum. However, filling these knowledge gaps is essential for understanding and predicting the future state and baseline conditions of global coastal zones (Newton et al., 2012; Hyndes et al., 2014; Ward et al., 2017).

## Episodic Weather Events

fmars-06-00210 April 23, 2019 Time: 19:29 # 7

Lastly, climate-driven changes in precipitation can strongly influence runoff patterns from land to sea (Milliman et al., 2008; Donnelly et al., 2017; Su et al., 2018). According to climate change scenarios, the frequency and intensity of extreme, episodic weather events is going to increase in various regions around the globe (Milly et al., 2002; Sánchez et al., 2004; Huntington, 2006; Kysel*ı* et al., 2012). Increased storm runoff will potentially have similar effects on coastal ecosystems as the cases discussed above and will strongly depend on ter-OM quality and quantity. Thus, while the carbon fraction may serve as an additional energy source for bacteria, nutrients will, depending on light and trophic conditions, additionally support primary producers (**Figures 2B–D**) (Guadayol et al., 2009; Bec et al., 2011; Pecqueur et al., 2011; Liess et al., 2016). Since storm runoff events occur globally, including ter-OM in coastal eutrophication monitoring is not only crucial for the boreal, arctic or tropical zones, but in general.

## SUMMARY, OUTLOOK, AND RECOMMENDATIONS FOR MANAGEMENT

The theoretical interactions and case studies described above illustrate that there are considerable complexities in the interactions between altered ter-OM inputs and inorganic nutrient loadings, making it currently challenging to understand and predict overall ecological responses, and effects of these stressors on ecosystem health. In **Figure 2**, we suggest a simplified conceptual framework to summarize how increasing ter-OM may affect coastal systems under high inorganic nutrient, and low inorganic nutrient conditions (i.e., eutrophic and oligotrophic scenarios).

In general, ter-OM may change light quality and quantity through its CDOM and shading/scattering caused by particles and inorganic material (**Figure 2A**, gray line), while ter-OM will also increase the carbon- and nutrient availability in receiving waters, with the ecological response depending on the bioavailability of the various ter-OM components, but also ecosystem community compositions (**Figure 2A**, black line). We predict an overall negative effect on autotrophic production of increasing ter-OM (**Figure 2B**). However, in oligotrophic systems, the nutrient fractions within ter-OM would initially act to increase production up to a certain threshold value, after which light-limitation would lead to decreasing autotrophic production. This threshold value would be dependent on the composition of the ter-OM, and specifically the light attenuation of the CDOM fraction. With increasing eutrophication of the system, we hypothesize that this nutrient effect of ter-OM would decrease, and eventually vanish, with light limitation being the main driver for the reduction in autotrophic productivity. With the reduction in light quantity and quality along the gradient of increasing ter-OM, a decrease in benthic: pelagic productivity can be hypothesized. For heterotrophs (**Figure 2C**), we predict an overall positive effect of increasing ter-OM on productivity for both nutrient scenarios (i.e., oligo- and eutrophic scenarios), until a saturation threshold is reached. At nutrient saturation, any additional nutrients and carbon will not lead to higher productivity and instead other factors will be limiting growth of heterotrophs (e.g., temperature, predation etc.). In sum, the overall effect on the metabolic balance (i.e., autotrophic: heterotrophic production) of increasing ter-OM would be a shift toward net heterotrophy, by causing a decrease in autotrophic processes and a simultaneous increase in heterotrophic respiration. By its influence on the metabolic balance of coastal ecosystems (i.e., net heterotrophy), changes in ter-OM input could have implications for the global carbon budget of coastal ecosystems, causing increased release of greenhouse gases (e.g., CO2, CH4) and a potential increase in coastal dead zones (Wikner and Andersson, 2012; Andersson et al., 2013; Lapierre et al., 2013), which deserves increased attention.

This developed conceptual framework, describing interactions between ter-OM and eutrophication or oligotrophication processes, remains to be challenged in future studies by conducting e.g., full-factorial experiments, using long-term monitoring data and performing meta-analysis of global case studies in different global coastal ecosystems, but also along the land-ocean continuum and by changing abiotic conditions (e.g., salinity, temperature). Such studies will greatly improve our understanding of baseline shifts within coastal ecosystems and help to determine thresholds where shifts could be expected, for example for light-limitation of autotrophs or transitions from net autotrophy to heterotrophy in different coastal systems.

In addition, there is a need for methodological harmonization between the different research areas and along salinity-gradients (i.e., catchment, freshwater, coastal, and marine). Although the case studies described above all deal with changes in ter-OM and nutrient input, they have rarely been put into context with each other. This might be due to the different drivers between systems (e.g., precipitation, reduced sulfate deposition, and structural changes in hydrological cycles), but also the situation in different regions (arctic, boreal, and tropical), and the different methodologies used. Future harmonization and improvements in methodology will strongly improve and aid the realization of comparative studies. To date, ter-OM is operationally classified by its dissolved organic carbon (DOC) content, where DOC is defined by its size (i.e., liquid that passes through a 0.45 µm filter in most, but also 0.2 or 0.7 µm filter in some cases) (Creed et al., 2018). However, there is currently a rapid development in methods for improving the characterization and determination of ter-OM, and CDOM, which should aid comparisons between regions and studies in the future (Bravo et al., 2018; Wünsch et al., 2018).

Our review shows that despite the growing recognition of the importance of ter-OM for aquatic productivity (Fabricius, 2005; Solomon et al., 2015; Creed et al., 2018), there is still significant improvements to be made in incorporating this knowledge into both research and policy efforts on eutrophication. Even though alternate sources of OM have been discussed as contributing to eutrophication for decades (Nixon, 1995; Nixon, 2009), the focus of most policies and monitoring programs is still connected to the classical "nutrient-enrichment" understanding of the eutrophication problem (e.g., in the European water

framework directive, WFD). We argue that the similarity in effects between eutrophication and ter-OM (reduced light availability, net heterotrophy, anoxic bottom water zones, and decrease in macrophyte cover) make it challenging for both researchers and policy makers to properly determine the correct cause-effects relationships. A potential misdiagnosis hinders our understanding of the complex responses observed in coastal ecosystems, and more worryingly could be interpreted as that current environmental policies to reduce inorganic nutrient loadings are inefficient in improving eutrophication status and ultimately ecosystem health.

Nixon (2009) argued for the need of "macroscopes" when understanding and managing eutrophication, including effects of climate change, and changes in the total supply of organic matter (i.e., both autochthonous and ter-OM) to the system. Andersen et al. (2006) argued that measurements of primary production should be mandatory when monitoring eutrophication. We echo these recommendations, and specifically also advice to include measurements of ter-OM or ter-OM proxies into eutrophication monitoring programs, such as DOC, CDOM, and light (including spectral composition). We believe this is timely, given both the number of cases showing a failure of reducing inorganic nutrient loadings for improving environmental conditions and ecosystem health in many coastal areas (Duarte et al., 2009), and the improving methodology for determining the quality and quantity of ter-OM (Mann et al., 2016; Osburn and Bianchi, 2016; Bianchi

#### REFERENCES


et al., 2018; Creed et al., 2018). Adding measurements of ter-OM in coastal monitoring programs would reduce uncertainty in predicting future states and baseline shifts of coastal ecosystems in response to both eutrophication and oligotrophication. This is important, since ignoring interactions between ter-OM and eutrophication and oligotrophication might slow down future efforts to secure coastal ecosystem health.

#### AUTHOR CONTRIBUTIONS

Both authors wrote the manuscript, contributed to revision, and approved the submitted version.

#### ACKNOWLEDGMENTS

We thank the strategic group on Land Ocean Interactions at the Norwegian Institute of Water Research (NIVA) for providing a fruitful context for us and specifically Ø. Kaste, A. Poste, and A. L. King, as well as the reviewer for their valuable comments on the earlier version of this manuscript. Further, we are grateful to the marine section at NIVA that supported us to attend the 4th International Symposium on Research and Management of Eutrophication in Coastal Ecosystems (EUTRO 2018) that has resulted in this work.


Hansell and C. A. Carlson (Amsterdam: Elsevier), 65–126. doi: 10.1016/b978- 0-12-405940-5.00003-0



matter export to the deep sea during tropical storms. Mar. Geol. 364, 32–42. doi: 10.1016/j.margeo.2015.03.005


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Deininger and Frigstad. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Seasonal and Temporal Drivers Influencing Phytoplankton Community in Kuwait Marine Waters: Documenting a Changing Landscape in the Gulf

Michelle J. Devlin<sup>1</sup> \*, Mark Breckels<sup>1</sup> , Carolyn A. Graves<sup>1</sup> , Jon Barry<sup>1</sup> , Elisa Capuzzo<sup>2</sup> , Francisco P. Huerta<sup>3</sup> , Fahad Al Ajmi<sup>3</sup> , Mona M. Al-Hussain<sup>3</sup> , William J. F. LeQuesne<sup>1</sup> and Brett P. Lyons<sup>2</sup>

#### Edited by:

Jesper H. Andersen, NIVA Denmark Water Research, Denmark

#### Reviewed by:

Hans H. Jakobsen, Aarhus University, Denmark E. Therese Harvey, NIVA Denmark Water Research, Denmark

> \*Correspondence: Michelle J. Devlin michelle.devlin@cefas.co.uk

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 04 October 2018 Accepted: 06 March 2019 Published: 04 April 2019

#### Citation:

Devlin MJ, Breckels M, Graves CA, Barry J, Capuzzo E, Huerta FP, Al Ajmi F, Al-Hussain MM, LeQuesne WJF and Lyons BP (2019) Seasonal and Temporal Drivers Influencing Phytoplankton Community in Kuwait Marine Waters: Documenting a Changing Landscape in the Gulf. Front. Mar. Sci. 6:141. doi: 10.3389/fmars.2019.00141 <sup>1</sup> Centre for Environment, Fisheries and Aquaculture Science, Lowestoft, United Kingdom, <sup>2</sup> Centre for Environment, Fisheries and Aquaculture Science, Weymouth, United Kingdom, <sup>3</sup> Kuwait Environment Public Authority, Safat, Kuwait

This assessment of environmental drivers and phytoplankton community in Kuwait (Arabian Gulf) describes changes in environmental conditions linked to the rapid urbanization over recent decades. To describe these changes, we have analyzed a longterm water quality dataset (1984–2017) and explored potential changes in a sub-set of phytoplankton community data by analyzing 10 years of phytoplankton data available (2007–2016) for Kuwait Bay and the Northern Gulf waters. The longer-term water quality data show that dissolved nutrient concentrations, with the exception of a recent fall in SiO4, have been increasing over 30 years, but with a high degree of variability reflecting the changing rate of inputs from river outflows and point sewage sources. The correlative analysis between the environmental parameters and phytoplankton in the period from 2007 to 2016 shows that the seasonal variability of the phytoplankton are influenced by several co-stressors including higher temperatures, coastal sewage runoff and changing salinity. While the extended nutrient enrichment has changed the trophic state of Kuwait waters, the seasonal and temporal correlations highlight that recent changes in phytoplankton community seem to be responding to cumulative pressures of eutrophication, climate, and salinity changes. The seasonal and temporal changes in the coastal phytoplankton community, responding to co-stressors, present challenges for managers to consider the simultaneous management of local, regional and external pressures. Continued declines in water quality within a system that is influenced by a warming climate can potentially have long-term consequences on the resilience of the Northern Gulf environment.

Keywords: sewage, phytoplankton, nutrients, Kuwait, eutrophication, assessments

## INTRODUCTION

fmars-06-00141 April 3, 2019 Time: 16:23 # 2

The Arabian Gulf (also known as the Persian Gulf and hereafter referred to as the Gulf), a shallow (35–40 m average depth), semi-enclosed marginal sea, is characterized by extreme natural environmental conditions and severe anthropogenic pressures (Sheppard et al., 2010). It is connected to the Gulf of Oman and the Arabian Sea through a narrow bottleneck of the Strait of Hormuz. Sheppard (2016) and references within identified local and global impacts as the main causes of environmental damage on the marine Gulf ecosystems. The global impacts are those attributed to climate change (Price et al., 1993; Price, 1998; Van Lavieren and Klaus, 2013) particularly temperature extremes in an already high temperature environment. The local impacts are experienced at a smaller spatial scale but are ubiquitous across the Gulf and typically associated with dredging, urban expansion and pollution discharges, particularly nutrient enrichment from sewage inputs. Water quality has been recognized as one of the key drivers impacting unique coastal ecosystems and human health in the Gulf (Sheppard, 1993, 2016; Al-Yamani et al., 2007; Sheppard et al., 2012; Devlin et al., 2015b; Al-Said et al., 2017). This study focuses on Kuwait, a Northern Gulf country that faces those same environmental challenges related to changing climate and increasing pollution. Kuwait experiences a wide range of environmental conditions, with temperature and salinity being at the extremes of known conditions (Riegl, 2003; Riegl and Purkis, 2012; Devlin et al., 2015b), atmospheric fallouts carried by North Western dust storms and particulate matter derived by river transport from the Shatt Al-Arab and increasing coastal pollution (Sheppard, 1993; Sheppard et al., 2012; Al-Sarawi et al., 2015). Coastal water quality issues are significant for Kuwait, with many sewage outfalls discharging directly into the marine environment that are either not properly treated or illegally bypassing the sewage treatment facilities (Lyons et al., 2015; Saeed et al., 2015).

The impacts of eutrophication and coastal sewage issues have been identified in many Gulf countries including long-term coastal pollution in Kuwait (Devlin et al., 2015a; Lyons et al., 2015; Smith et al., 2015), changes in macrobenthic community structure in Bahrain (Naser, 2011), eutrophication related blooms in the Red Sea (Mohamed and Mesaad, 2007), eutrophication impacts on coral reefs in Oman (Al-Jufaili et al., 1999) and longterm degradation from coastal pollution reported across the Gulf marine systems (Price et al., 1993; Sheppard, 1995, 2016; Subba Rao and Al-Yamani, 2000; Sheppard et al., 2010, 2012). The recent observed global increase in severity and frequency of harmful algal blooms (HAB) events has also been linked to eutrophication (Gilbert et al., 2002; Gilbert, 2006; Heisler et al., 2008) although not all HAB events can be directly linked to anthropogenic influences. The Gulf countries have seen substantive coastal expansion with deteriorating water quality conditions arising from unregulated sewage outfalls. These environmental changes have been related to several incidents of algal blooms, HABs, and fish-kills (Heil et al., 2001; Gilbert et al., 2002; Al-Yamani et al., 2004; Sheppard et al., 2010; Al Shehhi et al., 2014). The primary form of nitrogen can differ depending on the pressures and can influence the plankton species (Howard et al., 2007) with growth rate and proliferation of HAB species varying with the nitrogen source (Thessen et al., 2009; Caron et al., 2017; Glibert, 2017; Seubert et al., 2017). Depending on the species and scale, phytoplankton blooms can have significant impacts locally or over wider areas. Examples of negative impacts caused by eutrophication and phytoplankton blooms in the Gulf countries include mass mortality of marine animals such as fish kills (Glibert et al., 2002) shellfish (Richlen et al., 2010), and alternation of marine habitat and community structure (Sheppard et al., 2010, 2012; Sheppard, 2016).

Assessment of the plankton community and possible changes that could be related to nutrient enrichment are an important part of understanding the complexity of eutrophication (Devlin et al., 2009; Foden et al., 2010). Studies on phytoplankton in the Gulf have been varied, with some focusing on the species lists and distribution, particularly in relation to extreme conditions of salinity and temperature and describing the north to south gradient of plankton community structure in the Gulf (Subba Rao and Al-Yamani, 1998; Al-Handal, 2009). The analysis of the temporal and spatial phytoplankton variability in Polikarpov et al. (2008) shows that Kuwait's northern waters differed from areas further south in terms of phytoplankton structure with the environmental heterogeneity mainly attributed to the influence of the Shatt al-Arab system. The Shatt Al-Arab estuarine waters in the north are generally characterized by low species diversity, high biomass and higher production in contrast to Kuwait waters which have higher species diversity but lower biomass and production than the northern Gulf waters. More than 1,480 eukaryotic and prokaryotic primary producers have been described for the Gulf region (Jacob and Al-Muzaini, 1995) with Subba Rao and Al-Yamani (1998) identifying a dominance of epiphytic, tycho-pelagic diatoms to the north and a mix of diatoms and dinoflagellates to the south. Several papers report on these Gulf differences (Dorgham and Moftah, 1989; Al-Yamani et al., 2006; Nezlin et al., 2010; Quigg et al., 2013; Polikarpov et al., 2016) with a general agreement that there are two main regions in the Gulf along a diagonal northwest–southeast axis, with diatomdominated phytoplankton assemblage off the south and along the Iranian coast but with flagellate-dominated phytoplankton in the north and along the Arabian coast (Polikarpov et al., 2016). Changes in the Gulf phytoplankton community have been attributed to nutrient increases (El-Gindy and Dorgham, 1992; Al-Azri et al., 2014; Devlin et al., 2015a), and salinityrelated changes, particularly in the Northern Gulf region (Al-Said et al., 2017; Ben-Hasan et al., 2018). The only northern source of freshwater inflow into the Gulf is from the Shatt Al-Arab river and this flow from this river has decreased significantly over the past decades due to dam building and water storage in upper catchments of the river (Abdullah et al., 2015; Abdullah, 2017). These changes in long-term salinity have the potential to impact significantly on the phytoplankton community with a corresponding impact on the food web functioning and fish populations in the Gulf (Polikarpov et al., 2008; Al-Said et al., 2017). Changes in reduced productivity (Ben-Hasan et al., 2018) and decreases in fish catch landings in Kuwait (Bishop et al., 2011) have been partly attributed to increases in long-term salinity changes (Al-Said et al., 2017; Al-Yamani et al., 2017; Alosairi and Pokavanich, 2017).

Coastal pollution has been identified as one of the drivers of environmental impacts in Kuwait waters, including well documented fish kills in 1999 and 2001 (Gilbert et al., 2002) and ongoing chronic sewage issues (Saeed et al., 2012a, 2015; Lyons et al., 2015). In the face of population growth, inadequate sewage infrastructure, and increasing environmental degradation, increasing impacts are being observed across Kuwait's marine systems (Al-Ghadban et al., 2001; Sheppard et al., 2012; Burt, 2013; Al-Sarawi et al., 2015, 2018; Devlin et al., 2015b; Sheppard, 2016). Long-term changes in water quality in Kuwait have been associated with coastal sewage discharges (Devlin et al., 2015a) and modifications to the Shatt Al-Arab river flow (Al-Said et al., 2017; Al-Yamani et al., 2017; Alosairi and Pokavanich, 2017) but limited analysis exists on the impacts on the coastal phytoplankton from cumulative stressors of coastal pollution, salinity changes and a changing climate. We review the historical trends in environmental variables, including dissolved nutrient concentrations, salinity, temperature, dissolved oxygen, turbidity, and chlorophyll from 1983 to 2016 to assess the scale of change, the persistence of eutrophic conditions within Kuwait Bay and the Gulf and how the local and regional drivers influence the phytoplankton community. We explore the shifts in the phytoplankton community in Kuwait marine waters through linked environmental variables from 2007 to 2016 and explored these correlations in context of the long-term changes of the key environmental drivers. This study integrates long-term environmental data with recent phytoplankton data with the aim of understanding seasonal and temporal variability and the main drivers that influence the variability in the phytoplankton communities.

## MATERIALS AND METHODS

#### Field Sampling

The State of Kuwait is situated at the north-western corner of the Arabian Gulf and has a coastline of approximately 500 km. Kuwait Bay is one of the main features of the Kuwait marine environment alongside the Northern Gulf waters that run adjacent to the eastern side of Kuwait (**Figure 1**).

Kuwait Environment Public Authority (KEPA) has been collecting water and sediment parameters at sites throughout Kuwait waters from 1983 to present time (**Figure 1**). Samples are collected at each station at monthly intervals across 13 locations. The seawater stations, of which six are located within Kuwait Bay (hereafter known as "Bay" sites), and seven distributed along the Kuwait coastline (hereafter known as "Gulf " sites), are typically monitored for environmental and biological data under the Kuwait Marine Monitoring Program<sup>1</sup> .

A series of physical and chemical variables were measured either in situ or in seawater samples collected throughout the year from these 13 nearshore stations. In-situ sampling includes temperature, transparency (Secchi Disk), pH, salinity (S) and dissolved oxygen (DO) collected via Hydrolab DS5 water quality multiprobe, attached to the Hydrolab Surveyor 4 recorder. Water samples were also collected for dissolved inorganic nutrients (DIN), total suspended solids (TSS), and chlorophyll-a (Chl-a). The nutrient parameters analyzed under the water quality monitoring program are nitrate + nitrite (TOxN), ammonia (NH<sup>4</sup> <sup>+</sup>), dissolved inorganic phosphate (DIP) and silicate (SiO4).

The biological program measures phytoplankton abundance and diversity. From 2007, samples for phytoplankton analysis at each station were collected with two techniques: (a) a vertical net haul (with mesh size of 20 µm) for qualitative analysis only (presence/absence of taxa); and (b) water samples from Niskin bottles or sampling hose for quantitative analysis (identification and enumeration). Quantitative samples from Niskin or hose were collected at surface and at depth intervals with sampling depth of 2–5 ms depending on local conditions. Phytoplankton data collected in the two different areas are identified as Bay and Gulf communities.

#### Laboratory Methods

Within 6 h of sampling, water samples collected for the analysis of chlorophyll concentration were filtered (on GFC filters), and pigment concentration was determined by fluorometric technique, following maceration of algal cells and pigment extraction in acetone (Parsons, 2013). A Turner 10-005R fluorometer was used for the analysis and was periodically calibrated against diluted chlorophyll extracts prepared from log-phase diatom cultures (Jeffrey and Humphrey, 1975). After filtration of known volume of water samples, concentrations of TSS were determined gravimetrically from the difference between loaded and unloaded membrane filter weights after drying filters at 60◦ for a minimum of 12 h.

Water samples were analyzed for concentrations of dissolved inorganic nitrogen (DIN) including NH<sup>4</sup> <sup>+</sup> (ammonia), TOxN, which is the sum of nitrate (NO3) plus nitrite (NO2) and DIN which is the sum of TOxN and NH<sup>4</sup> <sup>+</sup>. Other dissolved nutrients included DIP and SiO4. All nutrients were analyzed by standard procedures (Ryle and Wellington, 1982) and conducted on a HACH analyzer (prior to 2009) and then on a Skalar San++ continuous flow analyzer (Skalar Analytical, Breda, Netherlands) after 2009. Detection limits are 0.006 mg/L for NH<sup>4</sup> <sup>+</sup>, 0.01 mg/L for NO2+NO<sup>3</sup> and 0.01 µg/L for DIP. As a component of instrumental quality control, artificial seawater was used to establish baseline characteristics. Analyses of the total dissolved nutrients (total dissolved nitrogen and total dissolved phosphate) were carried out using persulfate digestion of the water samples (Valderrama, 1981) and were then analyzed for inorganic nutrients, as above.

Phytoplankton samples were fixed with Lugol iodine solution, immediately after collection and analyzed with a compound microscope. Karlson et al. (2010) suggested that phytoplankton net samples should not be used for quantitative analysis and as the focus of this study is on quantitative investigation of changes in the phytoplankton community, only samples from water samples collected from Niskin or hose were used in the analysis and discussed further.

Phytoplankton samples from Niskin bottles were homogenized by inverting the samples multiple times, and 1 ml sub-samples were transferred to a Sedgewick-Rafter

<sup>1</sup>http://www.emisk.org/emisk/

chamber using a pipette. All taxa observed were identified and enumerated. The procedure was repeated three times for each sample and the results obtained from replicates were averaged; average abundances in 1 ml were then scaled up to determine cells L−<sup>1</sup> . This counting technique is considered accurate when samples contain sufficiently high numbers of phytoplankton cells, as it has a detection limit of 1,000 cell L−<sup>1</sup> (LeGresley and McDermott, 2010).

Biomass (as µgC L−<sup>1</sup> ) of the 30 most abundant phytoplankton taxa was estimated multiplying the abundance of a given taxon (as cell L−<sup>1</sup> ) by its cell carbon content (as pgC cell−<sup>1</sup> ) and divided by 10<sup>6</sup> (see **Supplementary Table S1**). Due to lack of size measurements of the phytoplankton taxa in the samples, the cell volumes were obtained from global coastal datasets (Harrison et al., 2015), from size measurements in Tomas (1997) and Al-Kandari et al. (2009), using geometric formulas given by Edler (1979) (**Supplementary Table S1**). Ceratium furca represented an exception as the biovolume of this dinoflagellate was calculated based on equation by Thomsen (1992). Biovolumes were then converted to carbon using the equations of Menden-Deuer and Lessard (2000).

#### Scope of Data Analysis

fmars-06-00141 April 3, 2019 Time: 16:23 # 5

Full details of collection, sampling, and analysis of the water quality data are given in Devlin et al. (2015a). Analysis of the long-term water quality changes, building on the water quality data set examined in Devlin et al. (2015a) and Lyons et al. (2015) will be presented to illustrate the impacts of the longterm enrichment of Kuwait marine waters and the impacts on the baseline condition. The integrated analysis will focus primarily on the water quality and quantitative phytoplankton data collected between 2007 and 2016. The full scope of the spatial and temporal analysis applied to the available datasets over the different time periods is summarized in **Figure 2**.

## Long-Term Changes in Water Quality Parameters (1983–2016)

#### Independence of Replicates Between Stations Within Sites

One-way analysis of variance models with normal errors were fitted to see if there was evidence that variation between stations was greater than variation within stations for data collected during 1983 and 2015. This tells us (a) whether there might be any pseudo-replication in our data set, and (b) whether there is evidence of differences between stations within the Kuwait Bay and Arabian Gulf sites. This was repeated for all the 10 variables (DIN, NH<sup>4</sup> <sup>+</sup>, TOxN, DIP, SiO4, salinity, temperature, chlorophyll, DO, TSS). The years 1983 and 2015 were chosen as these were at the beginning and toward the end of the time series, and both contained relatively large numbers of observations. Our analysis of variance model showed that variation between stations was not greater than variation within stations. Of the 20 models we fitted, only one was statistically significant at the 5% level. This was Ln TSS for the Arabian Gulf which had a p-value of p = 0.046. Given that from 20 tests, random chance would be expected to result in one statistically significant result at the 5% level even if there were no differences, it seems reasonable to assume that there is little evidence for differences between the stations. Therefore, considering the stations as replicates within each region is a highly plausible assumption. The six Kuwait Bay stations (Z01, Z02, Z03, Z04, Z05, Z06) and the seven Arabian Gulf stations (Z00, Z07, Z08, Z09, Z10, Z11, Z12) were then aggregated into two separate sites, known as the "Bay" and the "Gulf " sites.

#### Trends in the Water Quality Variables

The raw environmental data were plotted against year for Arabian Gulf and Kuwait Bay sites and investigated for skewness. Skewed data was log transformed for easier interpretation of the long-term data. To investigate the long-term trends, we used Generalized Additive Models (GAMs) to model each water quality (WQ) parameter as a function of time to the transformed and non-transformed data (S and DO). The function gam in the R package (R Core Team, 2017) was used for this modeling. Temporal GAM models are reported for each of the water quality parameters (LnChl-a, LnDIP, DO, LnNH<sup>4</sup> <sup>+</sup>, S, LnSi, LnT, LnTSS, LnTOxN, and LnDIN) for the time period 1994 to 2016. Thin plate regression splines were used for the smoothing with the degree of smoothing determined by generalized cross validation. No restrictions were placed on the degrees of freedom for the smoothed terms. Smooth trend plots were computed for each WQ variable across the whole of the time series (1983–2016). Models were constructed for data aggregated across sites within Kuwait Bay (Z01–Z06) and in the Gulf (Z00, Z07–Z12).

## Seasonal (Within Year) Trends in Water Quality Parameters

Within year variations of the WQ data were also explored across stations within the two sites (Kuwait Bay and the Gulf). To visually represent seasonal variations, the monthly means of the WQ parameters were calculated from data which had the long-term trend subtracted (the long-term mean are the trend lines shown in **Figures 3C**). Effectively, these means are from the residuals of the GAM models applied for the long-term trend analysis. This approach ensures that seasonal variation is not confounded with the long-term trends. Approximate 95% confidence intervals for the means were calculated using the percentile bootstrapping method (Manly, 2006).

## Trends in Phytoplankton Data (2007–2016)

#### Seasonal Changes in Phytoplankton Community

For an analysis of simple diversity indices over the year, the phytoplankton data were averaged by month over the time period to express the seasonal diversity indices. The number of taxa (S), abundance of taxa (N), Simpsons (D), evenness, Margalef (d), and Shannon index (H') was averaged for each month. Counts of diatoms and dinoflagellates were averaged for each month over the same time-period and presented against the average water temperature value.

To identify the most common and abundant phytoplankton species sampled from Kuwait Bay and the Gulf, the annual abundance of each species was summed and ranked for every year. The annual ranks were then averaged over the decade and the top 30 species were compiled based on the rank of the decade averages. The monthly abundance of the top 30 species across the time series (as well as the average monthly climatology) is displayed by means of a shade plot, created using the software R, with package ggplot2 (R Core Team, 2017).

To investigate seasonal changes in the phytoplankton community and the environmental drivers, non-parametric multivariate analyses were carried out in PRIMER v7 (Clarke and Warwick, 1994). Due to the requirement to match biotic data with environmental data the 10-year data set was reduced to only include months where both phytoplankton data

and environmental data were available. This resulted in a total of 63 monthly data points for the Gulf and 72 monthly data points for Kuwait Bay. Species abundance data was averaged for each month over the period 2007–2016 for Kuwait Bay and the Gulf and ln(x + 1) transformed. Resemblance matrices were compiled based on Bray–Curtis similarity. To aid interpretation, monthly averaged data over the time period was calculated for the Kuwait Bay and the Gulf. Analysis of similarities (ANOSIM) was computed to determine if the within site differences significantly contrasted with the between site differences from the Kuwait Bay and Arabian Gulf sites. Non-metric Multi-Dimensional Scaling (n-MDS) methods were applied to ordinate the seasonality of the phytoplankton communities for the two sites. A simple cyclical model was applied to determine if there was a statistical relationship between the monthly averaged community data and the seasonal cycle using the RELATE function for the Kuwait Bay and the Gulf. The seasonal model relates the annual community shifts to a cyclical model of the year, i.e., January and December are 1 step apart, while June is 6 steps apart.

The environmental data NH<sup>4</sup> <sup>+</sup>, DO, DIP, salinity, TSS, TOxN, and temperature was averaged for each month and all variables other than DO and temperature were ln(x + 1) transformed and normalized. Principal components analysis (PCA) was computed to ordinate the environmental variables with season for Kuwait Bay and Arabian Gulf. PCA analysis is a means of projecting the Euclidean distances of normalized environmental variables to describe the relationship between samples (or months in this case). PCA reduces the dimensionality of a data matrix to a low-dimensional summary and provides an interpretation of the axes (PC scores). Similar samples are represented by points located close to each other in the projection. The community data for all available months (using years as a factor) was matched to the environmental data to determine the combination of environmental variables that best explain the patterns in the community composition, using Spearman rank correlations in the BEST (Bio-Env) function. The BEST procedure determines the selection of environmental variable subsets that maximize the rank correlation between the similarity matrices of the community and environmental data by checking all combinations of variables (Clarke and Warwick, 1994; Clarke and Gorley, 2015). Prior to running BEST analyses environmental variables was checked for intercorrelation and DO was removed due to the high intercorrelation with temperature (−0.67 in the Gulf and −0.71 in Kuwait Bay). A permutation test with 999 iterations was completed to test for the agreement in the multivariate matrices. BEST analyses were completed on the full similarity matrices using year as a factor. However, for illustrative purposes monthly average community (n-MDS) and environmental data (PCA) are provided.

#### Temporal Changes in Phytoplankton Communities

Temporal variability of total abundance and biomass of the top 30 species, by month and by year, was calculated by summing abundance (or biomass) of all the taxa in a given sample. The resulting total abundance (or biomass) was then plotted using the software R (package ggplot2), fitting a local polynomial regression to the data points. Monthly long-term abundance (LnA) and biomass (LnBio) changes were plotted for Kuwait Bay and Gulf communities for the top 30 occurring species.

To test the data further we looked for seasonality and interannual variation over the time period (2007–2016). This was done by grouping the data into diatom and dinoflagellate lifeform groups. Lifeforms are aggregations of taxa that group together individual taxa with a similar functional role which are less likely to experience the extreme seasonal fluctuations of single species indicators and their use increases spatial intercomparability (McQuatters-Gollop et al., 2019). Shifts in community structure based on lifeforms were identified by applying the Plankton Index (PI) approach, the details of which are described elsewhere (Tett et al., 2007, 2013; Gowen et al., 2015). Briefly, the abundances

degrees Celsius, S (salinity) has units of ppt, Chl-a has units of µg/L. Logged raw data are presented for Kuwait Bay (A) and Gulf (B) with smooth trend plots computed for each water quality variable across the whole of the time series (1983–2016). The GAM model is presented independently of the data (C) note the smaller range on the x axis. The GAM model output for Kuwait Bay is shown by the black line and the red dotted line for the Gulf.

of two independent lifeforms (log-transformed and corrected for a zero value) during a chosen "starting period" are plotted against one another to define a "reference envelope in statespace. A sample's position at any point in time is defined in state space" by orthogonal axes of (log-transformed) lifeformpair abundance.

In order to capture normal seasonal and inter-annual variability in phytoplankton community structure, at least a 3-year reference interval is recommended, with full seasonal coverage. The reference envelope is defined such that it contains 90% of the reference data which is defined as monthly samples for the period 2007–2009 analyzed independently for both the Bay

and the Gulf. The PI value for any year or set of years outside the reference period is defined as the ratio of observed lifeform pair data which fall inside the reference envelope to the total number of data points being compared to the reference period. For example, if comparing a full year of data to the reference period (sampled monthly) and all 12 of the data points data fall within the reference envelope the PI is 1, if only 3 months fall within the reference envelope the PI is 3/12, or 0.25, and if all of the data fall outside the reference envelope the PI is zero, representing a complete change. The statistical significance of the PI value is given by an associated binomial p representing the probability of finding that number of points inside the envelope given change variation and an expectation of 0.9 (Tett et al., 2008). 'Significance' is assigned when p < 0.01, corresponding to PI values <≈ 0.7. While a useful guide, statistical significance does not necessarily indicate ecological significance.

## RESULTS

#### Long-Term Data (1983–2016)

#### Trends in Water Quality Variables

The raw environmental data were plotted against year for both the Arabian Gulf and Kuwait Bay (see **Supplementary Figures S1A,B**). All nutrient, TSS and Chl-a data showed a substantial degree of skewness due to the large increases in concentrations (nutrients) or episodic peaks (TSS and Chl-a) and were ln-transformed to reduce skewness. Salinity and DO were excluded from the ln-transformation.

Log transformed data shows increasing values for NH<sup>4</sup> +, TOxN, DIP, DIN for both Bay (**Figure 3A**) and Gulf (**Figure 3B**) sites. Temporal GAM models are reported for each of the water quality parameters, as per the methods described in Devlin et al. (2015a), with an additional 6 years added to the longterm analysis of environmental parameters reported in Devlin et al. (2015a) and Lyons et al. (2015) (**Figure 3**). The smoothed trends over time help estimate the long-term trends in the data. Strong temporal trends in data show significant increases in nearly all nutrient variables, including all dissolved nitrogen components of DIN (NO2, NO3, NH<sup>4</sup> <sup>+</sup>) and DIP (**Figure 3C**). The increase in NH<sup>4</sup> <sup>+</sup> is most likely related to the increased industrial discharges emanating from the sewage and industrial outfalls, many of them illegal discharges. The rapid increase in NH<sup>4</sup> <sup>+</sup> from 2009 to 2012 corresponds to the period of time that a pump malfunction resulted in raw sewage discharging into the Gulf. The temporal patterns observed for TOxN and SiO<sup>4</sup> demonstrate more variability, with increases over the early 1990s, and decreasing concentrations from the early 2000s. The simultaneous increases in TOxN, DIP, and NH<sup>4</sup> <sup>+</sup> in Kuwait Bay during the late 2000s are likely to be associated with industrial and sewage discharges which enter directly into Kuwait Bay from the Suibikhat industrial area (Al-Omran, 1998; Al-Ghadban et al., 2001; Saeed et al., 2015). All water quality parameters, with exception of SiO4, have been reducing over the recent years, indicating that the remediation of the more chronic sewage issues in the Gulf in 2012 through improvements in sewage infrastructure (Saeed et al., 2012a, 2015) have had positive effects, including reduction in NH<sup>4</sup> <sup>+</sup>, reductions in TSS, DIP and a slight increase in DO concentrations. The change in water quality parameters for the Bay and the Gulf are similar for the nutrients but show small differences in TSS and Chl-a, likely related to the reduced river flow from the Shatt Al-Arab river, driving longer retention times within Kuwait Bay (Alosairi and Pokavanich, 2017), increasing turbidity and reducing phytoplankton growth in the Bay.

#### Within Year Trends (Seasonality)

The extent of seasonal variation within years is shown in **Figure 4**. Positive increases are seen for TOxN and DIP in the period for October to December, aligning with periods of higher rainfall and potentially influenced by flow from Shatt Al-Arab River (Al-Yamani et al., 2007; Al-Yamani, 2008). In contrast, particularly for the Gulf area, increases in NH<sup>4</sup> <sup>+</sup> are variable over the year, with the highest deviation occurring in March, April, May, and November indicating that the source of NH<sup>4</sup> <sup>+</sup> is likely to be related to man-made industrial inputs and not affected by weather conditions or river inputs.

#### Seasonal Changes in Phytoplankton Community

The phytoplankton dataset (2007–2016) comprises 214 species divided among eight genera. The set of diversity indices averaged over the calendar year show similar seasonal patterns for the Bay and the Gulf (**Figure 5**). The number of species ranged between 25 and 43, with slightly more species in the spring to summer months. The abundance values change over the year, with higher abundances experienced in March, May, and September, likely influenced by the warming waters from March. Simpsons, evenness and Shannon indices are relatively constant, with dips in the values, particularly for the Gulf, in the March to May period, indicating that the high abundances in this period is driven by a reduced number of dominant species. The average monthly counts (logx) of the diatoms and dinoflagellates show the abundance increasing in February for the Bay community and in March for the Gulf community. High counts of diatoms are also measured in June and September in the Bay and in May, June and September in the Gulf. The phytoplankton community is dominated by diatoms, with higher counts of dinoflagellates measured in February and March and higher numbers of dinoflagellates occurring in Kuwait Bay compared to the Gulf community.

The seasonality of the top 30 species is presented in ranked species abundance shade plots (**Figure 6**). Highest counts in Kuwait Bay were measured during February, June, and September, dominated by diatoms, but with higher dinoflagellate counts in February and March (**Figure 6A**). The Gulf community responded differently to the seasonal cycle, with higher abundances and biomass persisting from March to June, and an autumn bloom in September (**Figure 6B**). January was characterized by low abundances and biomass for both the Bay and Gulf communities. The top 30 most abundant species in the Bay and Gulf (2007–2016) are dominated by diatom species. Leptocylindrus minimus, Leptocylindrus danicus, Eucampia zodiacus, and several species of Chaetoceros sp. (e.g., Chaetoceros pseudocurvisetus) dominate in terms of

abundance in the Bay with C. pseudocurvisetus thriving in the summer conditions, compared with some of the other species. Leptocylindrus minimus is the most abundant in the Gulf, with high counts of Chaetoceros sp., but with higher abundances of Thalassionema fraunfeldii, Thalassionema nitzschioides, and Asterionellopsis glacialis. There are only three dinoflagellates species in the top 30 for both Kuwait Bay and the Gulf: Karenia brevis, Karenia selliformis, and Prorocentrum micans in the Bay, and Karenia brevis, Prorocentrum micans, Ceratium furca in the Gulf. Karenia brevis is part of the Karenia (dinoflagellate) genus, a marine dinoflagellate that is responsible for red tides and associated mortality of seabirds and fish.

Seasonal patterns in community composition and the underlying environmental drivers were explored using matched environmental and phytoplankton data averaged for each month over the decade. Significant differences in community composition occurred between the Gulf and the Bay (ANOSIM R = 0.35, p = 0.001, test between sites using months as samples). Seasonal cycling in community composition is evident, both in the Bay (**Figure 7A**), and to a lesser extent in the Gulf (**Figure 7B**) and significantly corresponded to the yearly cyclicity model (Gulf: ρ = 0.129, p = 0.04, Bay ρ = 0.454, p = 0.001). The MDS shows the average monthly phytoplankton community, where months in proximity in the ordination space are most similar and months separated are most dissimilar. Consecutive months are linked with a trajectory line. The low ρ-values for the Gulf are in part due to the low resemblance of the February communities (**Figure 7B**) reducing the fit with the model in the Gulf. It should be noted that only 3 years of matched monthly data was available during February in the Gulf, and diversity indices with the full data set indicate the result may be driven by a lack of data rather than a significant shift in community composition in February. The PCA of the environmental data shows a transition between the warmer summer months (May– October), typified by lower nutrients and TSS, and winter months (November–April) characterized by lower temperature and higher DO (**Figures 7C,D**). Seasonal cycling of the average environmental parameters was evident in the Bay (ρ = 0.501, p = 0.001) and in the Gulf (ρ = 0.530, p = 0.001).

Environmental variables (excluding DO due to the intercorrelation with temperature) were matched to the community composition using all the monthly data from 2007 to 2016 for the Bay and the Gulf. In Kuwait Bay the within year environmental variables with the BEST correlation coefficient (0.415) to the seasonal community composition were NH<sup>4</sup> and temperature (p < 0.05) (**Table 1**). Individually, temperature and NH<sup>4</sup> <sup>+</sup> (0.225 and 0.224, respectively) are the most highly correlated factors to the community composition. In the Gulf the BEST suite of environmental variables correlating to 0.409 of variation in the community composition are salinity and temperature (p < 0.01). Temperature is the single factor most highly correlated with the changes in community composition in the Gulf (0.371). The environmental factors with the greatest bearing on the seasonal community composition have been overlaid as vectors in the average seasonal MDS ordination (**Figure 7**).

#### Temporal Changes in Phytoplankton Community

Monthly long-term abundance (LnA) and biomass (LnBio) changes were plotted for Kuwait Bay (**Figure 8A**) and the Gulf (**Figure 8B**) communities for the top 30 occurring species identified in **Figure 6**. The top 30 species varied between the Bay and the Gulf in terms of abundance and biomass ranking. Coscinodiscus spp., Guinardia flaccida, G. striata, L. annulata

Kuwait Bay, (B) counts of diatoms (lower blue segment of bars) and dinoflagellates (upper red segment of bars), and temperature (gray open circles and line) in the Gulf, (C) number of species (S), (D) abundance (N), (E) Simpsons index (D), (F) Evenness (J), (G) Margalef and (H) Shannon index. In (C–H) black filled circles = Bay, gray open circles = Gulf.

and species of the genus Rhizosolenia, accounted for the highest biomass across the time series, both in the Bay and Gulf, indicating that larger (less abundant) taxa were more important in terms of carbon than smaller but more abundant taxa (e.g., Chaetoceros spp.). These same three species also had high abundances, but seasonal abundance was noted in many species, with higher counts, particularly in 2010, in the Bay communities.

The two dinoflagellates species in the top 30 for the Bay (Karenia brevis, Karenia selliformis) have high abundances in 2010 and 2013 in the Bay, with peaks in biomass in 2010 and 2013. Dinoflagellates in the Gulf communities (Karenia brevis, Prorocentrum micans, Ceratium furca) have high abundances in most years but only small contributions to the biomass in the Gulf.

We looked further at the HABs species to see if there was any further evidence of HABS increasing in frequency and the environmental drivers (using the reduced data set). A total of 39 species listed as being potential HABs based on a review of the HABs of Kuwait (Al-Yamani et al., 2012) and the Intergovernmental Oceanographic Commission UNESCO HAB database<sup>2</sup> were identified from the species list. The frequency of these HABs species, the maximum counts and number of calendar months in the year they are present in the phytoplankton for the Bay and the Gulf communities is provided in **Table 2**. Overall, no clear trends in abundance of HABs over the period 2007–2016 were observed, however some species showed differences over the time period. Ceratium fusus increases in the Gulf but decreases in the Bay. Neurotoxin producing Karenia brevis significantly increases in the Bay, particularly during February and March.

The temporal changes between 2007 and 2016 in two major plankton lifeforms (diatoms and dinoflagellates) was investigated further through analyzing the community change over time. The temporal analysis of two lifeform indicators, based on a functional group approach, have proved relevant for the description of plankton community structure and biodiversity (Gallego et al., 2012; Garmendia et al., 2013) and have been used to assess community response to pressures such as sewage pollution (Tett et al., 2008, 2013). The annually calculated PI values with respect to the chosen 2007–2009 "reference period" indicate change in community since the initial observations with the current method (**Figure 9**). The overall PI for 2010–2016 is 0.53 for the Bay and 0.49 in the Gulf, indicating that about 50% of the observations from the most recent 7 years fall outside the reference period defined by the first 3 years of observations. In the Bay, the change appears to be driven by lower counts of diatoms which occur during all seasons, and both higher and lower counts of dinoflagellates during all seasons. In the Gulf, there is little change in the diatom community, but an increase in dinoflagellates. Looking at each year since 2009 individually, while there is no strong trend, the most recent years are the most different (lowest PI) which suggests that the community

<sup>2</sup>http://www.marinespecies.org/hab/index.php



BEST analysis was run on monthly community resemblance data from 2007 to 2016 using year as a factor. Individual parameters with a correlation coefficient > 0.2, and the highest correlation coefficient of two or more variables are provided (in bold).

structure is continuing to shift over time, with the rate of change increasing in recent years. While a change in community does not, in itself, signify a negative change, the results of the lifeform analysis suggests the Kuwait phytoplankton community, in terms of the two major lifeforms, are experiencing ongoing alterations in community structure and composition.

Although the Bay and Gulf sites presented similar phytoplankton community structure (with diatoms dominating both in terms of abundance and biomass), the trends of the total abundance and biomass at the two areas (calculated as sum of all abundances or biomass in a given sample) (**Figure 10**) were opposite, particularly in recent years. In fact, in the Bay, total phytoplankton abundance and biomass showed a decline, while in the Gulf both phytoplankton abundances and biomass showed an increasing trend.

#### Drivers of Change in Phytoplankton Community

The analysis on the long-term environmental data shows some significant changes occurring in many of the environmental parameters in both the Bay and Gulf coastal waters. There are many cumulative impacts that are driving these changes and the different statistical approaches presented in this paper provide some useful insights into the main drivers of environmental and community variability. The environmental correlations were explored in relation to seasonality and temporal changes. Potential drivers of this seasonal and temporal variability are also briefly discussed (**Table 3**).

### DISCUSSION

The environmental data collected over a 33-year period by KEPA at various coastal and marine sites is a crucial baseline dataset, offering a unique long-term perspective of a changing desert coastal system, where a low nutrient system, influenced mainly by seasonal river flow from the Shatt Al-Arab river and the larger hydrodynamic movement of the Gulf waters has shifted to a high saline, turbid, nutrient-saturated system with many of the changes being recorded in the long-term data (Devlin et al., 2015a; Al-Said et al., 2017). Nutrient enrichment of Kuwait Bay and Arabian Gulf has been ongoing for over two decades, with dissolved nutrient concentrations increasing from the early 1990s coincident with changing river flow and rapid urbanization of Kuwait. Infrastructure around sewage treatment is progressing, with state of the art treatment plants in place, but with many illegal sewage outfalls and drains still directing untreated sewage into the coastal environment (Al-Sarawi et al., 2015; Lyons et al., 2015). The NH<sup>4</sup> <sup>+</sup> signal increased rapidly from the early 2000s, and most likely related to direct sewage inputs, including the untreated sewage discharging into the Arabian Gulf (Sites Z07, Z08, Z09, **Figure 1**) for 3 years due to a pump malfunction (Al-Mutairi et al., 2014a,b; Saeed et al., 2015). The nutrient signal, both from NH<sup>4</sup> <sup>+</sup> and TOxN are showing signs of reduction over the past 4 years, but with concentrations still highly elevated (mean NH<sup>4</sup> <sup>+</sup>: 47 µg/L) from the baseline concentrations measured in the early 1980s (mean NH<sup>4</sup> <sup>+</sup>: 0.7–4.0 µg/L).

However, the change in nutrient concentrations relates both to coastal sewage inputs and Gulf wide changes influencing the time scale of nutrient enrichment. The variability in the long-term trend analysis for TOxN and NH<sup>4</sup> <sup>+</sup> indicates the changing sources of dissolved nutrients, with the strong TOxN signal related to diffuse sources from the Shatt Al-Arab river and remineralization of the NH<sup>4</sup> <sup>+</sup>. Reductions in precipitation and freshwater flow have had a dramatic impact on the hydroenvironment and salinity regime of the northern Arabian Gulf (Abdullah et al., 2015; Alosairi and Pokavanich, 2017). Any analysis of phytoplankton community change needs to consider the impacts of multiple drivers influencing the phytoplankton composition and abundance (**Table 3**). The long-term changes in environmental parameters demonstrate that the trends, while showing substantial increases over time, have had variable dips and peaks over the three decades of monitoring. This variable movement highlights the changing pressures and inputs with early years of nutrient enrichment driven by coastal urbanization and regional riverine inputs from the Shatt Al-Arab river. The increase in TOxN is likely related to diffuse inputs from the river loads as well as local coastal inputs. Kuwait's population has grown from just under 300,000 in 1960 to over four million in 2018 which has driven a coincident increase in sewage issues and coastal pollution. The recent years have seen the nutrient concentrations fall, particularly for TOxN which could relate to infrastructure improvements, such as sewage capture and treatment, but also the reduction in nutrient delivery from river flow. The changing hydrodynamics of the Shatt Al-Arab river can be seen in the variable salinity for both the Bay and the Gulf (**Figure 3**).

Changes in chlorophyll concentrations can provide a useful insight into the total phytoplankton variability (Devlin et al., 2007) due to the (generally) good agreement between planktonic primary production and algal biomass (Boyer et al., 2009). Chlorophyll concentrations for the Gulf and the Bay varied through time, with GAM analysis showing lower concentrations in the Bay compare to the Gulf and both areas showing a slight increase in recent years. Seasonally, log-chlorophyll concentration peaked during the winter period, likely as result of the higher nutrient concentrations and lower temperatures

at this time of year; while the opposite was observed in June (lowest chlorophyll concentration and maximum temperature). Furthermore, seasonality drivers strongly influence the changes in phytoplankton community over the annual cycle, with higher abundances corresponding to increasing temperatures from March.

The discrepancy between the chlorophyll trend and phytoplankton abundance trend could be the result of different factors. Firstly, it is important to note that the phytoplankton community data does not take into account the smallest fractions of the phytoplankton population (i.e., picophytoplankton and nanophytoplankton < 5 µm), while the chlorophyll measurements account for the entire phytoplankton population > 1 µm. The smaller phytoplankton taxa could therefore be playing an important role where chlorophyll and/or biomass are stored in the smaller size spectrum of the phytoplankton community. In this context, Brown et al. (1999) demonstrated that picophytoplankton (particularly picoeukaryotes) accounted for up to 35% of the total phytoplankton community biomass at coastal stations of the Arabian Sea (off Oman), and even a higher proportion (up to 56%) of the primary production. Another potential reason for the discrepancy between chlorophyll and biomass is that chlorophyll is a component of the phytoplankton cell biomass, which, as seen in **Figure 10**, can show different temporal and species distributions compared to phytoplankton abundance. The cellular content of chlorophyll, in respect to the carbon content, varies with different factors, including the type of phytoplankton groups dominating the community, and the local environmental conditions, e.g., light, nutrient availability, and temperature (Geider, 1987; Cloern et al., 1995; Geider et al., 1997; Sathyendranath et al., 2009; Jakobsen and Markager, 2016). So, for example, in nutrient-sufficient phytoplankton, the carbon to chlorophyll ratio increases with increasing light level (at constant temperature) and decrease with increasing temperature (at constant light levels; Geider, 1987). Nutrient (nitrogen) availability is also of significant importance in affecting the carbon to chlorophyll content, as seen for example in stratified waters (Taylor et al., 1997) and in coastal waters (Jakobsen and Markager, 2016). However, N-availability is unlikely to be an issue in the nutrient-rich waters of Kuwait Bay. Finally, the discrepancy between biomass and chlorophyll trends could also be related to the uncertainty associated to the biomass calculations, particularly for the taxa lacking size measurements and for which the global cell volume estimates by Harrison et al. (2015) were used.

Diversity indices are similar in the Bay and the Gulf, with reductions in diversity and evenness in May and June. In Kuwait Bay, the colonial diatom Chaetoceros pseudocurvisetus, along with Eucampia zodiacus and Leptocylindrus minimus dominated in May, with Chaetoceros persisting in high numbers throughout the summer months. Taxa abundances are generally higher in the Gulf, with several species increasing in the spring months, including Karenia brevis. Chaetoceros pseudocurvisetus still dominates through the summer months in the Gulf. A potential explanation for the success of C. pseudocurvisetus during the summer months (when TOxN reaches the lowest values) is the ability of producing both resting spores and resting cells in response to nutrient depletion (Kuwata et al., 1993). The two resting phases maintain or reduce metabolic processes (i.e., low respiration and low photosynthetic activity) until the environmental conditions return to a favorable state. Transformation of spores or resting cell is dependent on the available SiO<sup>4</sup> concentration, as spore cell walls are heavily silicified, while resting cell maintains an appearance similar to

TABLE 2 | Maximum values of HAB species counted in the phytoplankton community over the period 2007–2016 using matched data (63 months in the Gulf, 72 months in the Bay).


Only HABS with maximum counts of greater than 1,000 cells/L were included. HAB codes, 1 (Al-Yamani et al., 2012), 2 (IOC HABS database), 3 (both Al-Yamani et al., 2012 and IOC HAB sources). Bold > 100,000 cells/L maximum abundance.

the vegetative state (Kuwata et al., 1993). Diversity descriptors (**Figures 5A,B**) show that the phytoplankton community in February, as the weather warms, differs between the Gulf and Bay with April and May having similar diversity indices between the two sites. The Kuwait Bay community is more variable in the summer to autumn months, but winter months (November to December) are characterized by distinctive populations in the Bay.

The outcomes of the n-MDS and the PCA (**Figure 7**) analysis show a seasonal cycle, noting that the seasonal cycle represented is based on the 10-year averages in the phytoplankton community and environmental parameters. The BEST analysis (**Table 1**) identifies temperature as an important environmental variable in driving seasonal community composition, explaining 0.371 in the Gulf and 0.224 in the Bay. The enclosed waters of Kuwait Bay experience slightly warmer temperatures and the combination

of temperature and NH<sup>4</sup> <sup>+</sup> is correlated to 0.415 of the seasonal community composition. Salinity and TSS are likely to play a role in the seasonal cycle in the Bay due to the higher turbidity conditions, which are influenced by the longer retention times of the Bay given the decline in Shatt Al-Arab river flow and reduced capacity for the flushing. However, the correlation of TSS and salinity to the community data was not significant (**Table 1**). In the Gulf salinity and temperature combine as the primary environmental variables correlated to the community composition (0.409; **Table 1**).

The amount and form of nitrogen available in the water (particularly whether in the reduced, NH<sup>4</sup> <sup>+</sup>, or oxidized form, e.g., NO<sup>3</sup> <sup>−</sup>) have important implications for the productivity and for taxonomic composition of the phytoplankton community [see review by Glibert et al. (2016)]. It is generally accepted that NH<sup>4</sup> <sup>+</sup> is the preferred form of nitrogen taken-up by phytoplankton (see Raven et al., 1992 and references therein). This is the result of two distinct processes: (1) NH<sup>4</sup> <sup>+</sup> requires less energy to be transported by the microalgae across the cell membrane; and (2) NH<sup>4</sup> <sup>+</sup> shows an inhibitory effect on NO<sup>3</sup> <sup>−</sup> uptake and assimilation (see Dortch, 1990 and references therein). Particularly, the inhibitory effect of NH<sup>4</sup> <sup>+</sup> can manifest at different concentrations; for example, at concentrations lower than 50 nmol L−<sup>1</sup> in the oligotrophic waters of the Atlantic Ocean (L'helguen et al., 2008), or at concentrations > 4 µmol L−<sup>1</sup> in San Francisco Bay (Dugdale et al., 2007). The inhibitory effect of NH<sup>4</sup> <sup>+</sup> on NO<sup>3</sup> <sup>−</sup> uptake (in conjunction with changes in the phytoplankton community associated to high NH<sup>4</sup> <sup>+</sup>), can ultimately result in reduced productivity of phytoplankton community (Dugdale et al., 2007; Glibert et al., 2014). An example is presented by the San Francisco Bay Delta where increased NH<sup>4</sup> <sup>+</sup> concentration in the estuary due to sewage effluents from wastewater treatment plants, in combination with turbid conditions, has been shown to modulate occurrence of spring blooms and chlorophyll concentration, with consequences for the higher trophic levels (Dugdale et al., 2007; Glibert et al., 2014).

The TOxN concentration in the Bay is likely influenced by industrial and sewage discharges and reduced flow from the Shatt Al Arab river driving longer retention times for remineralization. Ongoing sewage failures have kept the NH<sup>4</sup> <sup>+</sup> concentration high, though there is a small decrease in NH<sup>4</sup> <sup>+</sup> in past 3 years which matches the timing of improvements in sewage infrastructure (Saeed et al., 2012b, 2015; Lyons et al., 2015). It could be expected


TABLE 3 | Summary of the variable trends in the environmental parameters (dissolved nutrients, TSS, salinity and temperature and phytoplankton community descriptors (Chl-a and community measures) are described over seasonality and temporal scales.

#### Seasonal and temporal drivers influencing phytoplankton composition in Kuwait marine waters

Chlorophyll-a Log-chlorophyll concentrations peaked during the winter period. This relates to the higher nutrient concentrations and lower temperatures at this time of year; while the opposite was observed in June which had the lowest chlorophyll concentration and maximum temperatures.

Chlorophyll concentrations for the Gulf and the Bay varied through time, with trend analysis showing lower concentrations in the Bay, but with slight increases in recent years, and higher concentrations in the Gulf, increasing from the late 1990s.

Low correlation between chlorophyll and abundance, particularly in the Gulf, could mean the smallest fraction of phytoplankton population (1–5 µm) is influencing the biomass. The trend (in the Gulf) suggests that there may be a transition to smaller flagellates. This could impact on the nutritional quality of the phytoplankton, but this needs to be tested through further analysis on the full plankton community.

to nutrients but it is difficult to separate this from the interaction with

temperature.

#### TABLE 3 | Continued

fmars-06-00141 April 3, 2019 Time: 16:23 # 18


Potential environmental drivers are described for each parameter.

that the high NH<sup>4</sup> <sup>+</sup> concentration in Kuwait Bay and Gulf (and its inhibitory effect on NO<sup>3</sup> <sup>−</sup> uptake), particularly in the 2000s, would be affecting phytoplankton composition and productivity. As support of this hypothesis, the decrease in NH<sup>4</sup> <sup>+</sup> (occurring in the last 4 years of the time series) was concomitant with a small increase in phytoplankton abundance and biomass in the Gulf communities, and with an increase in chlorophyll concentration in both Gulf and Bay (**Figure 3**).

One of the effects of the presence of NH<sup>4</sup> <sup>+</sup> as dominant N form may be the change in the structure of the phytoplankton community. With NH<sup>4</sup> <sup>+</sup> as the dominant nitrogen form in warm waters phytoplankton groups such as cryptophytes, dinoflagellates, cyanobacteria and chlorophytes may proliferate, with smaller size classes (<5 µm) driving production (Glibert et al., 2014; Glibert et al., 2016). Contrarily, with NO<sup>3</sup> <sup>−</sup> as main N-form, particularly under cooler water conditions, diatoms (cell size > 5 µm) are more likely to dominate the phytoplankton community and production (Glibert et al., 2014; Glibert et al., 2016). We could speculate that the NH<sup>4</sup> <sup>+</sup>-enrichment that occurred in the 2000s in Kuwait coastal waters may be linked to the decline in cumulative abundance and biomass of the top 30 species in the Bay, which are almost all diatoms. Interestingly, Al-Said et al. (2017) reported a change in the phytoplankton community of the Gulf of Kuwait between 2000 and 2013, with a reduction in diatom species and an increase in flagellates such as cryptophytes, and Phaeocystis globosa. However, the authors associated the change to variation in salinity rather than nutrients.

The initial analysis of HABs shows that the phytoplankton species which have been identified as HAB species are occurring in Kuwait marine waters and, for several species, in high abundances. The role of sewage effluent in stimulating or initiating HAB events in Kuwait marine waters is a key question for understanding the complex interactions between drivers and response in the phytoplankton community (Riegl et al., 2012).

This outcomes of this multi-faceted analysis show the difficulties in attributing changes in phytoplankton community to any single stressor and highlights the need to explore the cumulative and synergistic impacts of a combination of stresses including urban sewage discharge and alteration of discharge regime of Shatt Al-Arab River (Devlin et al., 2015a; Alosairi and Pokavanich, 2017).

While the multi-decadal nutrient enrichment is evident in the environmental data (1983–2016), the 10-year phytoplankton community data appears to be responding to several anthropogenic influences, all of which continue to be present in Kuwait marine waters. In an extreme environment, such as in the Bay and the Gulf marine waters, it would be expected that the temperature conditions would be the driving force in the timing of the phytoplankton community dynamics, however, temperature is now just one of several drivers with other anthropogenic factors such

as salinity and dissolved nutrients influencing the seasonal phytoplankton community. Recent work (Al-Said et al., 2017; Ben-Hasan et al., 2018) has identified significant changes in phytoplankton community organization on a decadal scale and suggests that salinity-related environmental changes have resulted in a coincidental decrease in species diversity and significant changes in phytoplankton community between the years 2000 and 2013 off Kuwait. The shorter temporal trend in the phytoplankton data (2007–2016) through the analysis of lifeform changes (**Figure 9**) and temporal changes (**Figure 10**) shows a variable phytoplankton population, with a small but significant change in abundance and biomass (**Figure 10**). The data suggest that the phytoplankton community in the Bay has decreasing abundance and biomass, in comparison to the Gulf community which shows an increasing abundance and biomass. However, the phytoplankton community data that is available represents a period where nutrient inputs are already high and makes it difficult to resolve how the community would have responded to the long-term nutrient enrichment described in the GAM water quality plots (**Figure 3**). The long-term chlorophyll data does offer some perspective on what may have occurred in the phytoplankton community data, with decreasing phytoplankton biomass from the late 1990s to 2010 (Devlin et al., 2015a), attributable to higher number of dinoflagellates.

## CONCLUSION

The work presented here demonstrates the importance of longterm datasets which provide details of the historic baseline in environmental drivers and their subsequent rate and nature of change through time. As with all regional case study approaches, there are many factors that need to be accounted for before the interactions of drivers and community impacts can be fully resolved. This study provides a first step towards this goal by highlighting the different stressors that can impact on a system already characterized by extreme conditions in temperature and salinity, such as those found in the Northern Gulf (Sheppard et al., 2010; Riegl and Purkis, 2012). In our changing world, knowledge of how these interactions with localized stressors can reduce ecosystem resilience to warming waters is critical to how managers respond when prioritising actions to reduce multiple pressures. Recent papers exploring the impacts of cumulative pressures in tropical ecosystems identify that management actions need to address local pressures as well as working on national and global scales to combat climate change (Hughes et al., 2008; Ortiz et al., 2018; Wolff et al., 2018).

For Kuwait, continued support for the long-term monitoring program and ongoing analysis of the phytoplankton community in the context of the changing environmental drivers is required to fully elucidate the contribution of individual and

### REFERENCES

Abdullah, A. D. (2017). Modelling Approaches to Understand Salinity Variations in a Highly Dynamic Tidal River: The Case of the Shatt al-Arab River. Boca Raton, FL: CRC Press. doi: 10.1201/9781315115948

cumulative anthropogenic stressors to the negative changes we observe in the phytoplankton community. This information will be essential to guide management action being implemented through local, national and regional programs. Local actions towards remediating the impacts of coastal sewage outfalls, regional actions towards restoring river flows, and national actions on climate change will offer the best outcomes for the health of Kuwait marine waters and the interdependent phytoplankton communities.

## DATA AVAILABILITY

The KEPA monitoring program data, (on which the data used in this study was extracted) and full details of the Kuwait monitoring program can be found on the eMISK website (http://www.emisk.org/emisk/). Data used in this paper is available on request following approval from the Kuwait Environment Public Authority (fahad.alajmi@epa.org.kw). The phytoplankton community data are available on request from CG (Caroline.Graves@cefas.co.uk). The water quality data are available on request from MD (michelle.devlin@cefas.co.uk).

## AUTHOR CONTRIBUTIONS

MD, MB, CG, JB, BL, WL, and EC contributed to the manuscript. FA, FH, and MA-H contributed to the data and editorial comments.

## FUNDING

Funding for this work came from the eMisk Programme funded by the Kuwait Environment Public Authority (KEPA).

## ACKNOWLEDGMENTS

Thank you to Dieter Tracey for assistance on the maps. We would particularly like to thank all the people who have worked in the Water Quality Laboratories and on the KEPA vessels over the 30 year time period and have been responsible for the collection, analysis, and reporting of the long-term monitoring data. Your input into this work has been invaluable.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00141/full#supplementary-material

Abdullah, A. D., Masih, I., van der Zaag, P., Karim, U. F., Popescu, I., and Al Suhail, Q. (2015). Shatt al Arab River system under escalating pressure: a preliminary exploration of the issues and options for mitigation. Int. J. River Basin Manage. 13, 215–227. doi: 10.1080/15715124.2015.10 07870


recruitment changes in the Northwestern Arabian Gulf? Mar. Pollut. Bull. 129, 1–7. doi: 10.1016/j.marpolbul.2018.02.012


present state and challenges within the European directives. Mar. Pollut. Bull. 66, 7–16. doi: 10.1016/j.marpolbul.2012.10.005


dinoflagellate Cochlodinium polykrikoides. Harmful Algae 9, 163–172. doi: 10. 1016/j.hal.2009.08.013


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Devlin, Breckels, Graves, Barry, Capuzzo, Huerta, Al Ajmi, Al-Hussain, LeQuesne and Lyons. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Intervention Options to Accelerate Ecosystem Recovery From Coastal Eutrophication

#### Carlos M. Duarte1,3 \* † and Dorte Krause-Jensen2,3†

<sup>1</sup> Red Sea Research Center (RSRC), King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia, <sup>2</sup> Department of Bioscience, Aarhus University, Aarhus, Denmark, <sup>3</sup> Arctic Research Centre, Aarhus, Denmark

Three decades following the onset of efforts to revert widespread eutrophication of coastal ecosystems, evidence of improvement of ecosystem status is growing. However, cumulative pressures have developed in parallel to eutrophication, including those associated with climate change, such as warming, deoxygenation, ocean acidification and increased runoff. These additional pressures risk countering efforts to mitigate eutrophication and arrest coastal ecosystems in a state of eutrophication despite the efforts and significant resources already invested to revert coastal eutrophication. Here we argue that the time has arrived for a broader, more comprehensive approach to intervening to control eutrophication. Options for interventions include multiple levers controlling major pathways of nutrient budgets of coastal ecosystems, i.e., nutrient inputs, which is the intervention most commonly deployed, nutrient export, sequestration in sediments, and emissions of nitrogen to the atmosphere as N<sup>2</sup> gas (denitrification). The levers involve local-scale hydrological engineering to increase flushing and nutrient export from (semi)enclosed coastal systems, ecological engineering such as sustainable aquaculture of seaweeds and mussels to enhance nutrient export and restoration of benthic habitats to increase sequestration in sediments as well as denitrification, and geo-engineering approaches including, with much precaution, aluminum injections in sediments. These proposed supplementary management levers to reduce eutrophication involve ecosystem-scale intervention and should be complemented with policy actions to protect benthic ecosystem components.

#### Keywords: coastal, eutrophication, recovery, intervention, management

## INTRODUCTION

The rise of coastal eutrophication as a global problem, first addressed in the 1970's (Ryther and Dunstan, 1971; Nixon, 1995), led to effort to reduce nutrient inputs into the ecosystems, spreading from pioneering efforts in northern Europe and North America to the world. Three decades after these efforts were initiated, nutrient inputs have been reduced by 25% (e.g., Nitrogen inputs to Chesapeake Bay, Lefcheck et al., 2018) to 50% (e.g., Danish coastal water, Riemann et al., 2016), and in a few cases reverted back to nutrient inputs comparable to those before the onset of eutrophication (e.g., Tampa Bay, Sherwood et al., 2017).

#### Edited by:

Alice Newton, University of Algarve, Portugal

#### Reviewed by:

Donald F. Boesch, University of Maryland Center for Environmental Science (UMCES), United States Angel Pérez-Ruzafa, University of Murcia, Spain

\*Correspondence:

Carlos M. Duarte carlos.duarte@kaust.edu.sa

†These authors have contributed equally to this work

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 21 August 2018 Accepted: 23 November 2018 Published: 11 December 2018

#### Citation:

Duarte CM and Krause-Jensen D (2018) Intervention Options to Accelerate Ecosystem Recovery From Coastal Eutrophication. Front. Mar. Sci. 5:470. doi: 10.3389/fmars.2018.00470

**78**

However, the expectation that phytoplankton primary production in coastal ecosystems is proportional to nutrient inputs (Nixon, 1995) and, therefore, should be reversible upon reducing nutrient inputs, through a parallel "oligotrophication" process (Nixon, 2009), has failed to be confirmed (e.g., Duarte et al., 2009; Carstensen et al., 2011). Evidence that eutrophication and oligotrophication are not simple complementary processes driven by a single lever, nutrient inputs, revealed a number of complexities in the responses of coastal ecosystems to eutrophication (Cloern, 2001). These include the operation of feed-back mechanisms, such as internal phosphorus loading through vertically segregated primary production in upper layers and remineralization in anoxic bottom layers, and shifting baselines due to concurrent changes and cumulative human pressures acting upon coastal ecosystems (Cloern, 2001; Duarte et al., 2009; Carstensen et al., 2011).

Forcing of biogeochemical processes in coastal ecosystems is now driven by both disruptions in the watershed, such as increased nutrient inputs from agricultural practices, and global processes linked to elevated anthropogenic emissions of CO<sup>2</sup> to the atmosphere, leading to warming, ocean acidification, deoxygenation, changes in precipitation and accelerated sea level rise (Doney, 2010). Warming affects the oxygen budget of coastal ecosystems through increased stratification and elevated respiratory demands (Steckbauer et al., 2011), and compounds with global trends toward ocean de-oxygenation to increase the likelihood of coastal hypoxia (Breitburg et al., 2018). Likewise, vertically segregated primary production in upper layers and remineralization in bottom layers of stratified eutrophic coastal waters lead to elevated CO<sup>2</sup> in the bottom layer, which compounds with the trend toward elevated CO<sup>2</sup> resulting from ocean uptake of anthropogenic CO<sup>2</sup> released to the atmosphere (Borges and Gypens, 2010; Cai et al., 2011; Wallace et al., 2014). These combined trends render eutrophic coastal ecosystems more vulnerable to acidification. Warming of coastal waters is expected to favor the occurrence of harmful algal blooms (Gobler et al., 2017). Such blooms, including green tide blooms of Ulva sp., may also be favored by elevated CO<sup>2</sup> (Gao et al., 2017). Warming also enhances eutrophication-impacts on ecosystem components such as seagrasses, with recent mass-mortality of seagrass with heat waves reported in a seagrass meadows in the Mediterranean Sea (Marba and Duarte, 2010), Florida Bay (United States, Carlson et al., 2018), and Shark Bay (Australia, Arias-Ortiz et al., 2018), and may, therefore, counter the recovery trend expected from nutrient reduction efforts. Lastly, increased precipitation with climate change is projected to enhance eutrophication in some coastal waters because it increases nutrient delivery to and stratification of coastal waters (Sinha et al., 2017).

Early evidence for the effectiveness of reducing nutrient inputs was derived from the observation that the dramatic reduction in fertilizer use in areas impacted economically by the collapse of the Soviet Union led to improvement in water quality, including that of the Black Sea (McQuatters-Gollop et al., 2008; Oguz and Velikova, 2010) and Cuban coastal waters, where fertilizer applications declined fivefold at this time (Baisre, 2006). Following three decades of sustained efforts to reduce nutrient inputs, coastal ecosystems around the world now start to show encouraging evidence of recovery (McCrackin et al., 2017), as recent evidence for Tampa Bay (Sherwood et al., 2017), Danish coastal waters (Riemann et al., 2016), Chesapeake Bay (Lefcheck et al., 2018), the Wadden Sea (van Beusekom, 2010; Dolch et al., 2013), and areas within the Baltic Sea (Andersen et al., 2017), among others, demonstrate.

However, increasing pressures from climate change threaten to shift baselines and arrest coastal ecosystems in a eutrophic state despite reduced nutrient inputs, thereby jeopardizing some of the expected benefit from nutrient load reductions. For instance, Boesch (2018) indicated that for the Chesapeake Bay it could require an additional expenditure of US\$ 1 billion over 5 years to offset the model-estimated effects of climatic changes on the severity of hypoxia, even if existing nutrient-load reduction targets were reached by 2025. There is, therefore, a pressing need to accelerate efforts to recover coastal ecosystems from eutrophication.

Here, we argue that diversifying the strategy to consider additional intervention options is needed to accelerate efforts to recover coastal ecosystems from eutrophication, and provide an overview of the intervention options to revert coastal eutrophication and their effectiveness in catalyzing the oligotrophication process. We address interventions at the catchment level and focus, particularly, on interventions in the coastal marine ecosystem.

## NUTRIENT MASS BALANCES IN COASTAL ECOSYSTEMS AND INTERVENTION OPTIONS

Conventional interventions to mitigate against coastal eutrophication focus on efforts to reduce nutrient inputs by improving waste-water facilities and regulate agricultural practices, such as fertilizer application and manure management schemes (Nixon, 1995; Boesch, 2002). These actions have played a fundamental contribution to mitigate eutrophication and drive coastal ecosystems toward recovery, and must be maintained and enhanced. Further interventions on land include creating freshwater wetlands to enhance nutrient and organic matter removal (Vymazal, 2007) and using winter catch-crops to further remove nutrients from nutrient over-enriched soils (Andersen et al., 2014). However, nutrient inputs from the watershed is only one of four main processes affecting nutrient mass balances in coastal ecosystems, i.e., nutrient inputs, nutrient export, nutrient burial, and nutrient emission (**Figure 1**), all of which should be considered in intervention options.

Nutrients entering coastal ecosystems may be exported with hydrological flows, in either dissolved or particulate form, buried in sediments, or, in the case of nitrogen, emitted to the atmosphere (**Figure 1**). The resulting nutrient concentrations are governed by the balance between these processes. Hence, inputs are modulated by exports, governed by hydrological exchanges and the residence time of the

FIGURE 1 | Schematic representation of nutrient fluxes in coastal sediments, including, along with each main pathway, the main intervention options available.

TABLE 1 | Examples of reported cases of hydrological engineering (ecoengineering with ecohydraulics) to mitigate eutrophication in coastal lagoons and estuaries.


coastal ecosystem, and burial in sediments, as formulated half a century ago by Vollenweider (1970). Each of the four main processes is amenable of intervention, thereby providing a broader slate of levers to complement efforts to reduce nutrient inputs on land to revert eutrophication. In addition, efficient nutrient cycling, a component of ecosystem homeostasis, depends on the integrity of food webs in coastal ecosystems (e.g., Cloern, 2001; Pérez-Ruzafa et al., 2011). Therefore, interventions at the ecosystem scale should ensure that the heterogeneity supporting efficient nutrient cycling is not disrupted.

## INTERVENTIONS TO ENHANCE NUTRIENT EXPORT

There are two main avenues for interventions to enhance nutrient export, one involving hydrological engineering of the coastal system to reduce residence time and increase flushing and dilution with coastal waters, and the other by harvesting ecosystem components able to remove nutrients from the water column, such as seaweed and filter-feeders (**Figure 1**). Hydrological engineering of ecosystems (also termed ecoengineering with ecohydrology), the manipulation of flows to enhance ecosystem status in coastal ecosystems, has a relatively long tradition (Elliott et al., 2016), and its effectiveness has been demonstrated in dozens of cases across the world, including ecosystems in Europe, North America, Oceania and others. This intervention is only possible, however, for semi-enclosed estuaries and lagoons and, particularly, intermittently closed and open lakes and lagoons (Schallenberg et al., 2010), where the flushing can be regulated by changing sluice practices or broadening the channels connecting the lagoon to the open coastal waters. For instance, such practices have been applied to increase the flushing of Chilika lagoon (India, Ghosh et al., 2006), Lake Veere (Netherlands, Wijnhoven et al., 2010), Waituna

Lagoon (New Zealand, Schallenberg et al., 2010), Ringkøbing Fjord (Denmark, Petersen et al., 2008), and the Peel-Harvey estuary (Australia, Humphries and Robinson, 1995), in all cases leading to improved water quality (**Table 1**).

In addition to enhancing nutrient export and dilution, hydrological engineering of coastal lagoons and semi-enclosed estuaries to increase the influx of coastal waters has often been reported to facilitate the development of populations of benthic marine organisms, such as bivalves, seagrasses, mangroves and salt-marshes, which further contribute to removing nutrients through burial and nitrogen emissions to the atmosphere (see below). For instance, changes in sluice practices to increase coastal water input into Ringkøbing Fjord not only enhanced dilution, but allowed the establishment of a population of the filter feeding bivalve Mya arenaria, which contributed to remove plankton and improve water quality, thereby catalyzing a regime shift (Petersen et al., 2008). However, ecohydrological interventions need be considered with care, using models to predict possible responses, to avoid negative experiences due to ill planned interventions (Schallenberg et al., 2010), such as those when the coastal waters flushing the lagoon are also eutrophied and, therefore, little or no dilution is achieved.

Ecological engineering may further contribute to nutrient export (**Figure 1**). In particular, seaweed aquaculture is a very effective tool to remove nutrients (Xiao et al., 2017). An assessment of the role of seaweed aquaculture in nutrient removal from highly eutrophied Chinese coastal waters showed that 1 ha of seaweed aquaculture removes nutrients equivalent to the nitrogen inputs to 17.8 ha and phosphorus inputs to 126.7 ha of Chinese coastal waters (Xiao et al., 2017). Chinese seaweed aquaculture annually removes approximately 75,000 t nitrogen and 9,500 t phosphorus, and at current growth rate of seaweed aquaculture, this industry will remove 100% of the current phosphorus inputs to Chinese coastal waters by 2026 (Xiao et al., 2017). Seaweed farms bring about multiple additional environmental benefits, including climate change mitigation (Duarte et al., 2017), along with economic benefits. Yet, their effective role in alleviating eutrophication has not been sufficiently acknowledged, nor generated compensation to the farmers (Xiao et al., 2017). Similarly, harvesting of filterfeeders from established aquaculture farms and reefs contribute to nutrient removal (Lindahl et al., 2005; Petersen et al., 2016). The only full-scale experiment on the use of mussel farms for eutrophication abatement we are aware of was conducted in Skive Fjord (Denmark), where mussel farms were calculated to remove 0.6–0.9 ton N ha−<sup>1</sup> year−<sup>1</sup> and 0.03–0.05 t P ha−<sup>1</sup> year−<sup>1</sup> (Petersen et al., 2016). Harvesting of natural seaweed and bivalve populations would also remove nutrients but may involve damage on seagrass beds and benthic processes (e.g., Erftemeijer and Lewis, 2006) and may also damage the ecosystem services of natural macroalgal communities e.g., as a habitat for fish (e.g., Teagle et al., 2017). Harvesting of natural populations should, therefore, either be avoided or carefully regulated to remain sustainable. Indeed policies, such as blue carbon strategies (Duarte et al., 2013), to protect and restore benthic habitats are an important lever to maximize the benefits of ecosystem engineering on overall ecosystem status. Likewise, whereas fisheries can remove nutrients from marine ecosystems, doing so sustainably is a challenge, as most fish stocks are already overexploited. Moreover, fisheries can also impact of food-web structure and disrupt nutrient cycling, therefore rendering coastal ecosystems less efficient in processing nutrients (Cloern, 2001; Pérez-Ruzafa et al., 2011). Indeed, observations and experiences from freshwater ecosystems show that addition of piscivorous fish or depletion of herbivorous fishsuffice to cause a regime shift arresting ecosystems in eutrophied states (Smith and Schindler, 2009). Our understanding of these responses is, however, insufficient at this stage to propose them as an intervention to accelerate recovery from coastal eutrophication. Similarly, the use of seaweed aquaculture to remove nutrients needs to adopt best practices to remove potential negative-side effects, such as oxygen depletion from organic inputs to the sediments, the introduction of exotic species (Naylor et al., 2001), or seeding green tides (Liu et al., 2009, 2010).

## INTERVENTIONS TO ENHANCE NUTRIENT SEQUESTRATION

Coastal ecosystems also support important processes leading to nutrient sequestration in sediments and biomass, which may help remove nutrients and assist in the oligotrophication process (**Figure 1**). Indeed, internal loading from nutrients recycled from sediments, particularly important in anoxic bottom layers (Vahtera et al., 2007), is a major buffer arresting coastal ecosystems in eutrophic states despite reduced nutrient inputs, as documented for some Baltic areas (e.g., Pitkänen et al., 2001).

Interventions to enhance nutrient sequestration may involve efforts to establish, protect and restore ecosystem components with a demonstrated capacity to support high nutrient burial rates (**Figure 1**). These include seagrass meadows (e.g., Gacia et al., 2002; Kennedy et al., 2010; Fourqurean et al., 2012), oyster reefs (e.g., Newell, 2004; Pollack et al., 2013; Kellogg et al., 2014), mangroves and salt-marshes (e.g., White and Howes, 1994; Valiela and Cole, 2002; Sanders et al., 2016), which remove nutrients from the water column and accumulate part of these nutrients in the sediments. Conserving and assisting the recovery of seagrass, oyster reefs, mangroves and salt-marshes through restoration do not only contribute to enhance nutrient burial, but provide, in parallel, benefits from the multiple ecosystem services these habitats provide. In addition, these habitats also lock nutrients into their biomass, but this only contributes to mitigate eutrophication in as long this is maintained, enhanced or, in the case of oysters, harvested, and also involves a strong seasonal component with maximum nutrient retention in the growth season of the organisms.

Intervention options also include geoengineering efforts to lock nutrients, particularly phosphorus, in the sediments (**Figure 1**). For instance, a recent successful experiment injected dissolved polyaluminum chloride into sediments at 6 m depth in Björnöfjärden in the Baltic Sea to bind phosphorus and prevent its recycling, responsible for sustained cyanobacterial blooms in this area (Rydin et al., 2017). Aluminum injections were highly successful, achieving a remarkable reduction in dissolved

inorganic phosphorus concentration, reduced chlorophyll a concentration, and enhanced submarine light penetration allowing deeper submerged aquatic vegetation penetration and much increased abundance of benthic fauna (Rydin et al., 2017). However, such interventions need be considered with care, as aluminum hydroxide has very low solubility at circumneutral pH, but dissolves at alkaline and acid pH, and might then be toxic to aquatic organisms. Additional options to remove phosphorus may include iron injection.

## INTERVENTIONS TO ENHANCE NITROGEN EMISSIONS AS N<sup>2</sup>

Conversion of reactive forms of nitrogen to N<sup>2</sup> gas is a major component of the natural nitrogen cycle, removing about 50%, or 24 Tg N yr−<sup>1</sup> , of the reactive nitrogen entering estuaries and coastal ecosystems (Galloway et al., 2004). This pathway is achieved through denitrification and anammox, processes that operate in hypoxic to anoxic environments (Canfield et al., 2010). Recent results have pointed at the same ecosystem components mediating high nutrient burial rates in coastal ecosystems, seagrass meadows, oyster reefs, mangrove and saltmarshes (e.g., Valiela and Cole, 2002; Sousa et al., 2012), as important vaults for removal of reactive nitrogen as N<sup>2</sup> gas emitted to the atmosphere. Denitrification rates in seagrass meadows are reported to be about fivefold greater than those in unvegetated sediments (Reynolds et al., 2016; Eyre et al., 2016; Zarnoch et al., 2017), while oyster reefs also support much higher denitrification rates than adjacent sediments free of such components (Kellogg et al., 2014; Caffrey et al., 2016; Smyth et al., 2016). Anammox rates in seagrass meadows remain poorly studied, but the only report available points at very high rates, albeit lower than denitrification rates (Salk et al., 2017). In oyster reefs, however, anammox rates have not yet been reported, to the best of our knowledge. Mangroves (Cao et al., 2017; Reis et al., 2017) and salt-marshes (Sousa et al., 2012) have also been reported to support very high anammox and denitrification rates, particularly when receiving high nitrate inputs (Koop-Jakobsen and Giblin, 2010; Reis et al., 2017). Hence, a high capacity for removal of reactive nitrogen, through both burial and, particularly, removal of reactive nitrogen as N<sup>2</sup> gas, is a key, but hitherto insufficiently realized ecosystem service of seagrass meadows (Reynolds et al., 2016; Zarnoch et al., 2017), oyster reefs (Cerco and Noel, 2007; Kellogg et al., 2014; Caffrey et al., 2016; Smyth et al., 2016), mangroves (Valiela and Cole, 2002; Cao et al., 2017; Reis et al., 2017) and salt-marshes (Valiela and Cole, 2002; Koop-Jakobsen and Giblin, 2010; Sousa et al., 2012).

The loss of seagrass meadows with eutrophication (Duarte, 1995), the major loss factor for seagrass meadows globally (Orth et al., 2006), thus, shifts key functional processes that lead to feed-back mechanisms preventing seagrass from re-establishing (Duarte, 1995; Maxwell et al., 2017; Moksnes et al., 2018). However, most importantly in this context, the loss of seagrass with coastal eutrophication leads to the loss of a major pathway for nitrogen removal from the ecosystem and helps explain the difficulties to revert eutrophication by only reducing nutrient inputs. Seagrass and oyster reef restoration, are well-established ecoengineering solutions that can, therefore, help remove nitrogen and, hence, accelerate oligotrophication. For instance, Reynolds et al. (2016) demonstrated how Zostera marina restoration by seed broadcasting in coastal waters of Virginia, by seed broadcasting, led to accelerated rates of nitrogen loss compared to those expected under unassisted recolonization.

## CONCLUSION AND PERSPECTIVES

Efforts to reduce nutrient inputs to eutrophied coastal ecosystems are starting to show clear signs of recovery three decades after these efforts were initiated. However, additional pressures from growing anthropogenic climate change, including warming, elevated CO2, increased precipitation and oxygen loss, and alterations to food-web or ecosystem structure risk jeopardizing these efforts and arrest coastal ecosystems in eutrophied states. There is, therefore, a need to accelerate efforts to achieve the desired oligotrophication of coastal ecosystems to further assist the very significant scientific, monetary and political capitals already allocated to revert eutrophication.

Efforts to accelerate oligotrophication requires a more comprehensive strategy to reduce nutrients in coastal waters than the simple approach taken to date, consisting in reducing nutrient inputs alone. As nutrient reduction efforts continue, the low-hanging fruits to meet this goal, involving addressing point source emissions, are already largely addressed and reducing diffuse sources, particularly atmospheric inputs, remain challenging and expensive, but can be achieved. For instance, reduction of NO<sup>x</sup> emissions in compliance with the Clean Air Act Amendment of the United States has been shown to be a main driver of nitrogen inputs to Chesapeake Bay (Eshleman and Sabo, 2016). Indeed, reducing fertilizer application and NO<sup>x</sup> emissions to the atmosphere brings about additional benefits in terms of air quality and reduction of greenhouse gasses, which need be considered in assessment of cost-effectiveness of this measures (Birch et al., 2011). The same is true for the enhancement of vegetated habitats, which not solely serves in eutrophication mitigation but also brings co-benefits e.g., in terms of climate change mitigation and adaptation in addition to providing several other ecosystem services (Duarte et al., 2013).

Adopting a nutrient mass balance approach to consider the regulation of coastal nutrient budgets further confirms that nutrient inputs are not the sole lever upon which interventions to revert eutrophication can act (**Figure 1**). Intervention options also include 1) hydrological engineering to increase nutrient export through enhanced flushing and dilution, 2) ecological engineering in terms of restoration of seagrass meadows and oyster reefs, which sequester nutrients and accelerate the loss of reactive nitrogen as N<sup>2</sup> gas and in terms of nutrient removal with seaweed and filter-feeder aquaculture, and 3) geological engineering to enhance nutrient burial through interventions to

lock phosphorus in sediments. While efforts to reduce excess nutrient inputs to coastal waters are essential and must continue, their effectiveness will be enhanced by considering additional intervention options. Cost-benefit analyses to consider the most effective combination of intervention options must also consider the additional benefits derived from some of these options. In particular, interventions creating habitats, such as seagrass and oyster reefs and seaweed farms, bring about multiple benefits that can greatly enhance the returns on investment of eutrophication abatement programs.

## REFERENCES


## AUTHOR CONTRIBUTIONS

Both authors conceived the research and wrote the paper.

## FUNDING

This study was supported by the Ministry of Environment and Food of Denmark (contract 33010-NIFA-16651) and by the Danish Centre for Environment and Energy (DCE).


of toxic algal blooms in the North Atlantic and North Pacific oceans. Proc. Nat. Acad. Sci. U.S.A. 114, 4975–4980. doi: 10.1073/pnas.1619575114


Sea. Ambio 36, 186–194. doi: 10.1579/0044-7447(2007)36[186:IEFENC]2. 0.CO;2


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Duarte and Krause-Jensen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Disentangling Environmental Drivers of Phytoplankton Biomass off Western Iberia

#### A. Ferreira<sup>1</sup> , P. Garrido-Amador<sup>1</sup> and Ana C. Brito1,2 \*

<sup>1</sup> MARE – Marine and Environmental Sciences Centre, Faculdade de Ciências, Universidade de Lisboa, Lisbon, Portugal, <sup>2</sup> Departamento de Biologia Vegetal, Faculdade de Ciências, Universidade de Lisboa, Lisbon, Portugal

Phytoplankton are the main primary producers in marine ecosystems, supporting important food webs. They are recognized as important indicators of environmental changes in oceans and coastal waters. Ocean color remote sensing has been extensively used to study phytoplankton (i.e., chlorophyll a – CHL – as a proxy of phytoplankton biomass) throughout the world, yet there is still much to understand in terms of what influences phytoplankton communities at regional scales. The main aim of this study was to investigate the drivers of CHL variability in the Western Iberian Coast (WIC), for the period 1998–2016. Satellite CHL data were acquired from the Copernicus Marine Environment Monitoring Service. A positive annual trend of CHL was observed at the Northern coastal WIC, near Galicia, and a negative trend was observed in Southern areas, near the Gulf of Cádiz. An empirical orthogonal function analysis was implemented to identify regions with similar patterns of CHL variability. Six regions were obtained. A set of climate indices, satellite, and model variables were then used as environmental predictors in generalized additive models that explained between 22.8 and 52.8% of the total variance of CHL anomalies calculated from a detrended and deseasoned dataset. In the Northern oceanic region, positive anomalies were linked to high North Atlantic Oscillation values and negative anomalies to high mixed layer depths. In the Southern oceanic region, positive CHL anomalies were found to be associated with high concentrations of nitrogen, that may indicate nitrogen limitation. CHL in coastal areas were found to respond to basin-wide (e.g., Atlantic Multidecadal Oscillation) and coastal processes (e.g., upwelling and continental runoff), yielding positive anomalies with low salinity (SAL) values. In coastal areas off major rivers, such as Douro and Guadalquivir, the positive response of CHL to increased nitrogen concentrations and decreased SAL was evident. Considering the changes in climate expected for this region, related to the decrease in precipitation and increase in summer temperatures, as well as some apparent weakening of upwelling, possible significant impacts on the phytoplankton community can be anticipated. These results are therefore also relevant for environmental management, especially in the context of the European Marine Strategy Framework Directive.

Keywords: chlorophyll a, Western Iberian Coast (WIC), environmental drivers, ocean color remote sensing (OCRS), empirical orthogonal functions (EOFs)

#### Edited by:

Jacob Carstensen, Aarhus University, Denmark

#### Reviewed by:

Fabrizio D'Ortenzio, Centre National de la Recherche Scientifique (CNRS), France E. Therese Harvey, NIVA Denmark Water Research, Denmark

> \*Correspondence: Ana C. Brito acbrito@fc.ul.pt

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 27 September 2018 Accepted: 28 January 2019 Published: 28 February 2019

#### Citation:

Ferreira A, Garrido-Amador P and Brito AC (2019) Disentangling Environmental Drivers of Phytoplankton Biomass off Western Iberia. Front. Mar. Sci. 6:44. doi: 10.3389/fmars.2019.00044

## INTRODUCTION

fmars-06-00044 February 27, 2019 Time: 11:11 # 2

Phytoplankton have long been used as a bioindicator of environment changes in oceanic and coastal ecosystems. Their advantages as a bioindicator include their role as the main primary producers in the global ocean, their rapid life cycles, and their sensibility to changes in its surrounding conditions (e.g., in nutrient availability or light). In the past decades, ocean color remote sensing (OCRS) has become the most costeffective method for studying phytoplankton, using chlorophyll a (CHL) as a proxy of phytoplankton biomass, providing data with high geographical coverage, spatial resolution, and sampling frequency (Racault et al., 2014). Consequently, remote-sensingderived CHL is now considered an essential climate variable by the Global Climate Observing System (GCOS; Bojinski et al., 2014) and a large number of studies on the response of phytoplankton community to environmental changes have been performed, particularly at a global scale (e.g., Behrenfeld et al., 2006; Rost et al., 2008; Beardall et al., 2009; Paerl and Huisman, 2009; Boyce et al., 2010; Hallegraeff, 2010; Litchman et al., 2012). Additionally, coastal studies have already shown responses of phytoplankton biomass and community structure to changes in nutrients, temperature, and salinity (SAL) (e.g., Mendes et al., 2018). However, the phytoplankton communities at a regional scale have complex variations, fluctuating from region to region and there is still much to unravel on community dynamics (e.g., Brotas et al., 2013; Brito et al., 2015). Additionally, these communities' responses are typically intertwined with local and regional coastal events and processes (Jones et al., 2013; Carstensen et al., 2015; Dorado et al., 2015; Howard et al., 2017). Determining how environmental changes in coastal ecosystems drive phytoplankton is essential toward understanding coastal ecosystems, particularly under the everpresent threat of anthropogenic climate change. Key questions that should be addressed at the regional scale include: (1) what are the most important drivers of CHL anomalies? (2) How do these drivers affect CHL variability? (3) What are the implications of the pressure–response relationships to the marine policies in this region?

The present study focuses on the Western Iberian Coast (WIC; 36◦ to 45◦N, 6◦ to 12◦W), a complex regional ecosystem located amidst the transition between North–East Atlantic subtropical and temperate waters. One of the main features of WIC is its location on the Northernmost section of the Canary Current Upwelling System, one of the four major Eastern boundary upwelling systems of the global ocean (Wang et al., 2015). The oceanography of the region is known to be largely dominated by mesoscale structures such as jets, meanders, eddies, upwelling filaments, and countercurrents, superimposed on seasonal cycles (Relvas et al., 2007). Moreover, it is a region expected to sustain severe climate change impacts (Philippart et al., 2011). The last report from the Intergovernmental Panel on Climate Change (IPCC) predicted a decrease in precipitation, increased summer temperatures, as well as in the frequency and intensity of heatwaves (Kovats et al., 2014). In fact, during early August 2018, the Western Iberia region experienced a severe heatwave with maximum temperatures higher than 40◦C for several days, reaching up to ∼47◦C in some locations. Despite being still controversial (see Varela et al., 2015), recent studies have reported an apparent weakening of upwelling events (e.g., Lemos and Pires, 2004). This associated with sea surface warming in Western Iberia Sea (e.g., Goela et al., 2014) are likely to cause a change in phytoplankton community, potentially leading to a shift from diatom dominance to small flagellates (and potentially harmful species), with direct effects on food webs.

During recent years, phytoplankton communities have been investigated in WIC, with several in situ and remote-sensing studies performed. These studies have contributed to a more indepth knowledge of local phytoplankton biomass variability and phenology (e.g., Navarro and Ruiz, 2006; Silva et al., 2009; Krug et al., 2018), community composition and structure (e.g., Lorenzo et al., 2005; Mendes et al., 2011; Goela et al., 2014), as well as its relationship with specific environmental processes, such as riverine discharges (e.g., Moita et al., 2003; Prieto et al., 2009; Guerreiro et al., 2013; Vaz et al., 2015) and coastal upwelling (e.g., Cravo et al., 2010; Pérez et al., 2010; Guerreiro et al., 2013; Vidal et al., 2017). Some investigations (e.g., Navarro and Ruiz, 2006; Krug et al., 2017) were also able to regionalize areas in CHLcoherent regions (i.e., areas with similar CHL variability patterns) to study the environmental drivers of phytoplankton. However, most of these studies have been focused on specific sections of the coastal zone and an overall view of the WIC is needed.

This study aims to contribute to bridging this knowledge gap by considering WIC as a whole, including its coastal (coastal waters are here considered as marine waters where continental freshwater discharges and other sea–land processes have a strong influence) and oceanic domains, and using long-term, datasets with high spatio-temporal resolution. The primary objective was to investigate the drivers of CHL variability in the WIC using satellite and modeled data. This objective was subdivided into three specific goals: (i) analyze CHL anomalies over a nearly 20 years data time series (1998–2016); (ii) identify the main drivers of the CHL anomalies along WIC; and (iii) assess how these drivers influence CHL variability. WIC is also an area of interest under environmental management, representing a large extension of marine waters and making up most of the EU Marine Strategy Framework Directive (MSFD; 2008/56/EC) subregion "Bay of Biscay and the Iberian Coast." Under the MSFD, achieving these objectives would deliver key information toward the evaluation of MFSD descriptors 1 (Biodiversity), 4 (Food Webs), and 5 (Eutrophication), all of which consider phytoplankton as a major component.

## MATERIALS AND METHODS

#### Study Area

The WIC (36◦ to 45◦N, 6◦ to 12◦W; **Figure 1**) is a highly complex and heterogeneous area, being under the influence of several large-, meso-, and small-scale oceanographical phenomena. WIC is located on the Eastern boundary of the North Atlantic basin. It is influenced by major climatic patterns of the North Atlantic, such as the North Atlantic Oscillation (NAO), the

Atlantic Multidecadal Oscillation (AMO), or the Eastern Atlantic (EA) pattern.

Western Iberian Coast is inserted in the Northern Canary Current System, one of the four major Eastern boundary upwelling systems worldwide (Wang et al., 2015). Upwelling off the WIC is highly seasonal, with higher intensity during the boreal summer (Wooster et al., 1976; Fiúza et al., 1982; Lemos and Pires, 2004), particularly off upwelling centers (e.g., off Galiza and Sagres). The influence of the nutrient input driven by upwelling on the marine phytoplankton off the WIC has been thoroughly studied for several regions within the WIC (e.g., Tilstone et al., 2003; Ribeiro et al., 2005; Rossi et al., 2013; Goela et al., 2014). Other significant agents in this region include several large river basins along the coast (e.g., Tagus, Douro, Guadiana, Guadalquivir), which deliver valuable nutrient inputs for coastal phytoplankton communities. Moreover, WIC's proximity to the mouth of the Mediterranean Sea has been seen to shape nearby biological communities (González-García et al., 2018).

#### Biological Data: Satellite Chlorophyll a

Weekly surface CHL remote-sensing data (L4 with 1 km spatial resolution) for the WIC were acquired from the North Atlantic CHL concentration from satellite observation (daily average) Reprocessed L4 (ESA-CCI); available at http://marine.copernicus.eu/ for the period 1998–2016. This product is derived from merging Remote Sensing Reflectance (RRS) data from three different sensors: Sea-viewing Wide Field-of-view Sensor (SeaWiFS), Moderate Resolution Imaging Spectroradiometer-Aqua (MODIS-Aqua), and Medium Resolution Imaging Spectrometer (MERIS). Subsequently, CHL is estimated by applying the OC5ci algorithm, which combines algorithms OC5 (Gohin et al., 2002), for coastal waters, and OCI (Hu et al., 2012) for the open ocean. This dataset is already optimized for the NE Atlantic, displaying a good agreement with the in situ measurements (R <sup>2</sup> = 0.79; RMSE = 0.26) and spatial coverage of about 95%. Moreover, an intercomparison effort of ESA-CCI derived-CHL (albeit an earlier version of the product) with in situ data from WIC was done by Sá et al. (2015). They reported values of unbiased root mean square error, standard deviation, and correlation coefficient that were comparable to other products, such as MODIS OC3M chlorophyll algorithm. Coastal pixels less than 4 km from the shore were excluded to avoid difficulties associated with nearshore Case II waters, such as CHL overestimation in the presence of colored dissolved organic matter (CDOM) and total suspended matter (TSM).

Mean and seasonal (i.e., mean for each annual season) CHL climatologies for 1998–2016 were calculated. CHL anomalies were also determined by detrending (i.e., removing the leastsquares trend) the CHL time series and subsequently by deseasonalizing (i.e., removing the mean associated with each month) it. The linear trend was obtained by calculating the slope of the least square regression model. Thus, the resulting anomalies are also independent of the seasonal variability and are genuinely representative of anomalous CHL variability.

## Physical and Biogeochemical Data

Daily sea surface temperature data (SST; 4 km spatial resolution) for the period 1998–2016 were extracted from the Advanced Very High-Resolution Radiometer (AVHRR) Pathfinder SST product (PFV53; available at www.nodc.noaa.gov; Casey et al., 2010; Saha et al., 2018). This product is produced by the National Oceanic and Atmospheric Administration (NOAA) National Centers for Environmental Information (NCEI) and uses data from the AVHHR instruments aboard NOAA's series of polar satellites.

This dataset is considered highly reliable particularly due to its corresponding global multi-year in situ validation matchup dataset and has been utilized in several oceanographical studies off WIC (Santos et al., 2012; Alves and Miranda, 2013). Plus, only pixels flagged with quality level flags 6 and 7 (i.e., having passed all or almost all quality tests) were chosen to ensure the highest data quality (Kilpatrick et al., 2001).

1998–2016 ∼8 km spatial resolution weekly model-derived data on mixed layer depth (MLD; m), SAL (psu), sea surface height (SSH; m), and the zonal (U) and meridional (V) components of surface water flow (m s−<sup>1</sup> ) were also acquired, from CMEM's Atlantic – Iberian Biscay Irish (IBI) – Ocean Physics Reanalysis Product (IBI\_Reanalysis\_PHYS\_005\_002; available at http://marine.copernicus.eu/). The water flow velocity (WVel; m s−<sup>1</sup> ) was also calculated from the U and V components. This product is based on the NEMO v3.6 ocean general circulation model (Madec, 2008) and forces the model with regional atmospheric fields available at the European Centre for Medium-Range Weather Forecasts (ECMWF)<sup>1</sup> to ensure its optimization to the IBI region. This model has the advantage of assimilating in situ temperature and SAL data, as well as satellite SST and altimeter data. It has recently been validated for the IBI region using several surface and mooring buoys along WIC (for more details, see Levier et al., 2014, 2015; Sotillo et al., 2015).

Dissolved inorganic nitrogen (DIN; µM), calculated as the sum of nitrate (NH4) and ammonium (NO<sup>−</sup> 3 ), phosphate (PO3<sup>−</sup> 4 ; µM), and silicate (Si) (µM) for the period 2002–2014 were acquired, with a weekly periodicity and 8 km spatial resolution, from CMEM's Atlantic – IBI – Ocean Biogeochemistry Non-Assimilative Hindcast Product (IBI\_Reanalysis\_BIO\_005\_003; available at http://marine.copernicus.eu/). This product uses both the ocean physical IBI reanalysis product mentioned above and the biogeochemical model PISCES (Levier et al., 2014) to derive a 3D high-resolution product with several biogeochemical variables optimized to the IBI region. In situ validation efforts have delivered good results, particularly for nitrates and phosphates (R <sup>2</sup> > 0.75; RMSE = 2.04 and 0.1 for nitrates and phosphates, respectively; Bowyer et al., 2018). L3 weekly Photosynthetic Active Radiation (PAR; Einstein m−<sup>2</sup> day−<sup>1</sup> ) with a spatial resolution of 4 km was acquired, for the period 1998–2016, from the GlobColour project website (http://www. globcolour.info/; Frouin et al., 2003).

#### Climate Indices

Due to their relevance for the understanding of global and regional ecosystems in the North Atlantic, the following climate indices were obtained for 1998–2016: the NAO index, the East Atlantic (EA) pattern, the AMO, the Multivariate El-Niño/Southern Oscillation Index (MEI), and the Western Mediterranean Oscillation (WeMO). NAO, EA, AMO, and MEI were acquired from NOAA, while WeMO was acquired from the Climatology Group of the University of Barcelona.

All of the chosen climate indices have been seen to influence the climate over the NE Atlantic region (e.g., de Castro et al., 2008; Hurrell et al., 2013; Krug et al., 2017; Jiménez-Esteve and Domeisen, 2018). NAO reflects differences in atmospheric surface sea-level pressure in the North Atlantic Ocean between the subtropical high and subpolar low. Its positive phases are usually linked with high SSH and pressure and low precipitation and temperatures over Southern Europe (Hurrell et al., 2013). EA is similar to NAO in its calculation and is considered the second main variability pattern in the North Atlantic. EA, in its positive phase, is associated with higher temperatures and low precipitation across Southern Europe (Iglesias et al., 2014). AMO is a multidecadal SST-based index known for its importance to decadal-scale events, such as rainfall patterns and hurricane activity. While AMO has been under a constant positive phase since the 1990s, recent studies have found that AMO has become slightly negative the last few years due to cold anomalous SST values in the subpolar region and is expected to maintain its negativity (Frajka-Williams et al., 2017). MEI is closely related to ENSO, considered one of the most critical interannual climatic phenomena worldwide. Despite its intrinsic relation to the Eastern Pacific Ocean, ENSO has been observed to impact North Atlantic climate, such as interacting with NAO (Jiménez-Esteve and Domeisen, 2018). WeMO is a somewhat new pattern identified for the Western Mediterranean basin and its vicinities (Martin-Vide and Lopez-Bustins, 2006) and it is correlated primarily with rainfall patterns over the Iberian Peninsula (Martin-Vide and Lopez-Bustins, 2006).

## Data Analyses

Prior to data analyses, the dataset was transformed to match the required temporal and spatial resolution, as needed. Linear relationships between CHL anomalies and environmental variables were assessed using Pearson correlations. Additionally, empirical orthogonal function (EOF) analysis (Lorenz, 1956) was used to identify and study spatio-temporal patterns of CHL variability along the WIC. Generalized additive models (GAMs) were used to identify and evaluate the effect of several environmental predictors on CHL variability.

#### Empirical Orthogonal Function (EOF) Analysis

Empirical orthogonal function analyses are often used in climate studies and can be very useful to understand the variability of a given variable, as it breaks a complex signal by delivering mathematically independent spatio-temporal modes and how these modes vary. However, despite its usefulness, EOF modes are not necessarily related to climate patterns and should be interpreted carefully. In this study, conventional EOFs (i.e., unrotated) were used. Rotated EOF analyses (e.g., varimax rotation; Kaiser, 1958) were also tested but were seen to worsen the interpretability of the modes and were thus discarded. To run the EOF analysis, a complete matrix corresponding to the response variable must be provided. Thus, CHL was averaged monthly, and the remaining missing data were filled following Racault et al. (2014) procedure, i.e., a moving mean with 3-pixelsize window was sequentially run longitudinally, latitudinally, and over time. After this procedure, the remaining missing data were residual (<0.01%). A final moving mean with a 3 × 3 pixel-size window was run to fill the remaining pixels

<sup>1</sup>https://www.ecmwf.int/

as in Krug et al. (2017). Afterward, CHL was detrended and deseasonalized before applying the EOF analysis.

The EOF analysis allowed the extraction of variability modes from the CHL dataset into a spatial and temporal component. The significance of the resulting EOF modes was evaluated according to the method described by North et al. (1982). According to this method, an EOF mode is considered significant if its sampling error is smaller than the difference between the mode's explained variance and the next mode [see North et al. (1982) for more details]. The resulting modes were used to identify regions with coherent CHL variability along the WIC using a similar methodology to Krug et al. (2017). The spatial component partitions WIC according to the spatial patterns of the mode, where negative and positive areas correspond to deviations regarding the mean of the mode over the dataset period, as further described in the section "EOF Analysis." These negative/positive areas are considered as CHL-coherent regions, in terms of their own variability patterns and are considered to be out-of-phase in relation to each other.

#### Generalized Additive Models (GAMs)

Generalized additive model techniques were applied to evaluate the environmental drivers of CHL variability. GAMs are an extension of generalized linear models (GLMs), whereas the relationship between the response variable and the predictors can be non-linear. Thus, GAMs are more suitable for complex situations where nonlinear patterns may arise which would be missed or wrongly interpreted in a GLM. As such, it has been used extensively to study CHL drivers (e.g., Jayaram et al., 2013; Krug et al., 2017; Liu et al., 2017). In the present study, GAMs were applied to CHL anomalies within each region of coherent CHL variability patterns identified through the EOF analyses. GAM results were checked for residual autocorrelation and, since no significant evidence of autocorrelated was found, no autocorrelation structure was added to the GAM function. Simple linear correlations between each environmental variable were used as a screening procedure to exclude variables with high co-correlation (R <sup>2</sup> > 0.7). GAM analyses were performed using the package mgcv 1.8-24 (Wood, 2006, 2017), in R (R Development Core Team, 2018).

## RESULTS

#### CHL and Environmental Climatologies

From the average CHL concentration during 1998–2016 (**Figure 2**), several coastal CHL hotspots can be identified along the WIC: off the Galician Rias (∼42.5◦N), off most Northern and Central Portuguese coast (∼41.5◦N–40◦N), off the mouth of the Tagus estuary (∼38.5◦N) and in the Cádiz Gulf (37◦N, −8 ◦E to −6 ◦E). Several CHL plumes extending over several kilometers off the coast can also be found facing major estuaries or over the Estremadura Spur (∼39◦N) and Aveiro region (∼39.5−40.5◦N), where the continental shelf is wider. Overall, the average CHL distribution off WIC (**Figure 2E**) can be divided into two gradients: (i) an offshore south–north gradient, with lower mean CHL concentrations (<0.2 mg m−<sup>3</sup> ) below 37◦N and increasingly higher concentrations toward the Northern latitudes of WIC (e.g., ∼0.4 mg m−<sup>3</sup> circa 45◦N); and (ii) an offshore– inshore gradient, with mean CHL over 1 mg m−<sup>3</sup> next to the coast. Moreover, results suggest that CHL in the WIC displays a marked seasonality, typical of temperate waters (**Figures 2A–D**). Higher CHL concentrations in oceanic waters off the WIC were typically detected during the timing of the spring bloom. During summer, however, CHL is distinctively higher in the coastal area, when compared with other seasons, due to the influence of upwelling. An exception would be the Gulf of Cadiz region, where CHL is higher nearshore during spring as opposed to summer. Trend analysis revealed a general weak positive linear trend over the WIC (**Figure 2F**). CHL trends over 0.02 mg m−<sup>3</sup> year−<sup>1</sup> were found off the northWestern coast. Nonetheless, a negative trend was observed on the Southern coast, particularly in the Cadiz Gulf, where values under −0.06 mg m−<sup>3</sup> year−<sup>1</sup> were found. While these can be considered weak trends, it is essential to think of these values as yearly trends that, at specific locations, may accumulate over a 19-year dataset.

During 1998–2016, WIC presented itself as a highly complex region regarding its environmental forcing. Nutrient-wise, while higher concentrations were expectedly observed closer to the coast, each nutrient (DIN, PO3<sup>−</sup> 4 , Si, and Fe) shows distinct patterns over the WIC. High mean DIN concentrations (>5 µM; **Figure 3A**) were found in the coastal zone of the WIC, mainly between 38◦N and 43◦N. Nevertheless, a clear offshore latitudinal gradient can be seen, as DIN increases toward higher latitudes. A similar pattern can be observed for phosphate (PO3<sup>−</sup> 4 ; **Figure 3B**). However, relatively high mean phosphate concentrations (>0.3 µM) were also observed on the Eastern Gulf of Cádiz (off the Guadalquivir estuary mouth and Cádiz). Si are mainly concentrated on the coastal zone of the WIC (38◦N–43◦N) and off the Eastern Gulf of Cádiz, where it reached mean concentrations over ∼7 µM (**Figure 3C**). Mean iron (Fe) concentrations were mainly higher along the coast (>3 × 10−<sup>3</sup> µM), while also displaying a north–south decreasing gradient (**Figure 3D**). Mean SST ranged from ∼15 to 19◦C along the WIC (**Figure 3E**). Overall, SST was seen to be higher toward Southern latitudes. However, a cold coastal southward plume caused by coastal upwelling and coastal currents can be seen along the Western coast with temperatures below 17◦C. MLD also varied considerably over the study area, ranging from <15 m on coastal areas to >40 m offshore (**Figure 3I**). The highest MLDs were detected north of 43◦N (>50 m). PAR analysis revealed almost a linear north–south gradient toward the Southern region WIC, ranging between 29 and 39 einstein m−<sup>2</sup> day−<sup>1</sup> (**Figure 3F**). Mean SAL across WIC was generally above 35.5 psu, except at some coastal areas that are strongly influenced by riverine inputs (e.g., off Northern Portugal) where salinities as low as ∼34 psu were detected (**Figure 3G**). Mean SSH, from 1998 to 2016, ranged between −0.3 and −0.38 m. Lower heights were typically detected on coastal areas around 43–45◦N, while higher values were seen in oceanic waters below 39◦N (**Figure 3H**). Mean water flow components and speed (**Figures 3J–L**) show that water along the WIC typically flows southward, while along the Southern coast it flows eastward near the coast and westward in the Southern offshore areas. From 39◦N to 42◦N,

water can be seen flowing slightly toward offshore, a sign of upwelling. Several regions where surface water flow speed averaged over 0.15 m s−<sup>1</sup> were detected: off Peniche (∼39.3◦N), off Sagres (∼37◦N), and the Gulf of Cádiz, flowing into the Gibraltar Strait (36–37◦N), where maximums over 0.3 m s−<sup>1</sup> were measured.

North Atlantic Oscillation index (**Figure 4A**), from 1998 to 2016, exhibits high variability, typically varying interannually between positive and negative phases. Nonetheless, three major positive (1999–2001, 2012, and 2013–16) and two negative (2009–2011 and 2013) periods may be detected. The AMO has remained mostly positive since 1998 (**Figure 4B**), fluctuating between 0 and 0.5. Otherwise, only a few minor negative episodes were recorded over the studied period (e.g., 2009). The East Atlantic (EA) pattern index (**Figure 4C**) was mostly positive, with one long positive period between 2000 and 2004 followed by two short positive instances centered in 2007 and 2010. Since 2012, EA has been increasingly positive with rare negative episodes. The El-Niño/Southern Oscillation Index (MEI) fluctuations over the years indicate several El-Niño events between 1998 and 2016 (**Figure 4D**). Strong El-Niño events occurred in 1998 and 2014– 16, while weaker El-Niño followed by strong La-Niña events occurred in 2006–7 and 2009–10. Regarding the WeMO index, apart from a small period between 2000 and 2002, was seen to be on a steady negative phase (**Figure 4E**). Negative peaks were detected in 2006 and 2011.

FIGURE 4 | Temporal variability of climate indices, from 1998 to 2016: (A) North Atlantic Oscillation (NAO), (B) Atlantic Multi-decadal Oscillation (AMO); (C) East Atlantic (EA) pattern; (D) Multivariate El Niño-Southern Oscillation (MEI); and (E) Western Mediterranean Oscillation (WeMO). Black lines represent 3-month moving means.

## Correlation Analysis

Linear relationships between CHL anomalies and environmental variables included in this study are presented for the entire period of the datasets (1998–2016; **Figure 5**). Several relevant significant linear relationships were identified for some areas off WIC, a summary of the most relevant is described in **Table 1**. Nitrogen (DIN) and phosphates (PO3<sup>−</sup> 4 ) yielded a significant positive correlation with CHL anomalies, especially along coastal areas, as well as on the offshore area located under 40◦N. These positive relationships suggest a direct driver–response relationship at those specific areas. Interestingly, higher negative correlations between SST and CHL anomalies were seen around Sagres, a known upwelling center in WIC.

Salinity exhibited an inverse relationship with CHL anomalies along the Western and Southern coast of the Iberian Peninsula, most likely as a result of nutrient-rich river discharges. Correlations were particularly strong in the Gulf of Cadiz, suggesting river discharges might have a big role in CHL variability here. However, SAL did correlate positively with CHL anomalies over a patch north of Galicia, also suggesting differences in CHL anomalies patterns. The surface water flow indices (U and V) exhibited mostly negative correlations off the West coast and several positive correlated patches above 43◦N (**Figures 5J,K**). Their main distinction occurred off the Southern coast and Gulf of Cadiz, where positive coefficients were identified between CHL anomalies and U, while negative coefficients were seen for V. The mean WVel showed positive correlations with CHL anomalies off the Western and Southern coast. The difference in linear correlations between U and V over WIC is a testament to the importance of wind-driven coastal processes such as upwelling, in this area. Negative correlations between U and V and CHL anomalies along the Western coast of the Iberian Peninsula indicate SW-oriented water transport, highly associated with coastal upwelling in WIC, contributes to higher CHL. Overall, this heterogeneity suggests different patterns of CHL variability over WIC, with distinct environmental drivers being responsible, underlining the importance of the regionalization process utilized in this study.

## EOF Analysis

Empirical orthogonal function analysis returned six statistically significant modes. Overall, these six modes explained a total of 57.6% of the total extant variance. Despite these six significant modes, only the first four modes (∼51.7% of the total variance) were further contemplated and analyzed to avoid overcomplexity. **Figures 6A–D** display the first four modes produced by the EOF analysis, including their spatial mode and the temporal variability for each mode, from 1998 to 2016. Mode 1 explained 28.75% of the total variance and its spatial component is mostly positive over WIC, meaning most WIC displays a similar variability pattern according to this mode. Moreover, offshore values were typically higher, except in a few negative zones, located on the Southern coast, particularly off Cádiz (**Figure 6A**), where the variability pattern was different. Regarding its temporal variability, it is clear that there are consistent negative peaks during early spring, while positive peaks showed no apparent

(× 10−<sup>3</sup> µM); (E) SST (◦C); (F) PAR (einstein m−<sup>2</sup> day−<sup>1</sup> ); (G) salinity (psu); (H) SSH (m); (I) MLD (m); (J) zonal component (U) of water flow (m s−<sup>1</sup> ); (K) meridional component (V) of the water flow (m s−<sup>1</sup> ); and (L) mean water flow velocity (WVel; m s−<sup>1</sup> ).

TABLE 1 | Minimum, average, and maximum correlation coefficients observed between the environmental variables and CHL anomalies.


Percentage (%) of non-land pixels identified with significative correlations (pvalue < 0.05) are also presented. Please note that only relationships with maximum (minimum) correlation coefficients higher (lower) than 0.3 (−0.3) are presented.

seasonality, occurring irregularly during summer, autumn, and winter (**Figure 6A**). However, their magnitude is inconstant, particularly of the positive phases, as major peaks were identified in 2005 and 2009. Interestingly, 2016 was the only year where mode 1 did not change its signal, remaining negative throughout the year. Overall, this mode appears to be capturing, on its negative phase, CHL anomalies during the timing of the typical Northeast Atlantic temperate spring bloom. Mode 2 explained circa 12% of the total variance and displayed a very distinct spatial component than mode 1 (**Figure 6B**). According to mode 2, two main contrasting areas can be found: a positive area north of 41◦N, excluding a coastal strip off Western Galicia,

and a negative area occupying most of the area south of 41◦N. Nevertheless, small areas varying positively with mode 2 were also identifying off the Douro and Guadalquivir estuaries. Mode 2 temporal component is complex and does not suggest any apparent seasonality. Regarding the temporal component, it is highlighted by two strong peaks: a negative peak during late

winter of 2009 and a positive one during mid-spring of 2011. Mode 3 (**Figure 6C**) is responsible for explaining 6.36% of the total variance and, similarly to mode 2, also defines two main large areas over WIC with contrasting signals. On the one hand, a positive area encompasses most of the Iberian coastline and a sizeable oceanic patch between 39◦N and 43◦N. On the other hand, the offshore waters above 43◦N and under 39◦N may be observed to vary negatively with mode 3. Two main negative periods, with peaks during the spring of 2004 and 2012, highlight the temporal component of this mode. Intriguingly, the two main positive peaks of mode 3 appear to occur immediately after these negative periods, also during spring (2005 and 2013). Lastly, mode 4 (**Figure 6D**) explains 4.52% and has the most complex spatial and temporal components of the four considered EOF modes. Spatially, it appears to divide WIC longitudinally, resulting in a positive oceanic area east of −10◦E and a broad negative more coastal area, with lower negative values alongshore. This pattern suggesting its variability is related to coastal processes. Two main positive peaks were identified in 2009 and 2015, while, at least, three major negative peaks occurred in 2001, 2009, and 2016.

Using the spatial components from EOF modes 1–4, six areas with coherent CHL variability over WIC were defined (**Figure 7** and **Table 2**). First, the positive/negative areas for the first four modes were combined, as shown in **Table 2**. As a result, 16 (all) possible combinations were produced, 5 of which returned no pixels. Subsequently, linear correlations were performed on the CHL anomalies of the remaining 11 areas to avoid redundancy. Areas geographically next to each other with high correlation (R <sup>2</sup> > 0.85) were merged, resulting in six regions with distinct chl a variability patterns (**Table 2**). Half of the regions were oceanic, while the other half were mainly coastal, revealing sharp differences between coastal and oceanic CHL variability. Overall, each designated region appears to be associated with distinct CHL

for the Western Iberia Coast (WIC), derived by combining the positive and negative signals of the EOF spatial components used (EOF1, EOF2, EOF3, and EOF4), according to the scheme presented in Table 2.

TABLE 2 | Summary of the creation process leading to the regionalization of WIC in six CHL coherent regions using the positive and negative signal of the modes resulting from the EOF analyses.


The percentage of the area covered by each area regarding WIC is also displayed. Due to rounding, the sum of the total area covered is not exactly 100%.

variability patterns and, as seen by the subsequent GAM analyses, associated with distinct environmental drivers.

Region A is located in the Northern region of WIC, ranging from 43◦N to 45◦N, coinciding almost entirely with the negative Northern area observed in the spatial component of mode 3. This area is mainly oceanic. Region B, the largest region defined in this study, is also mostly oceanic, located in the Southern part of the WIC region, but it also includes small portions of the Southern coast. Region C covers most of the coastline, except in the Northern part of Galicia. Region D is mainly composed of the coastal zone in the northWestern part of Galicia, yet it also includes several small zones off the mouths of the Douro and Guadalquivir estuaries (corresponding to the positive areas defined by mode 2; **Figure 6B**). Region E occupies a sizeable oceanic area off central WIC, extending roughly from 39◦N to 42◦N. Region F is the smallest region created (0.5% of the study region), coinciding with the areas with negative values in the spatial component of Mode 1.

#### GAM Analyses

Results of the GAM analyses, which identified the most important environmental drivers, are presented in **Table 3**. Each significant model predictor is identified, as well as its degree of significance. The model's adjusted R 2 and explanatory power (i.e., variance explained) are also provided. Plus, the partial effects of the main detected predictors on CHL anomalies are exhibited in **Figure 8** for each region, using a schematic display. An additional figure with all the results is provided as **Supplementary Material**.

Overall, AMO, MLD, and V were the most common environmental predictors identified by the GAM analyses, being present in most of the models. Other common predictors were

TABLE 3 | Summary of the best performing generalized additive models (GAM) performed on the effects of environmental predictors on CHL anomalies of each of the specific regions derived by combining the signal of the EOF spatial component.


The model explanatory power (MEP) or percentage of variance explained, the adjusted coefficient of determination (R<sup>2</sup> ), and the statistically significant model predictors are displayed. Note that ∗∗∗ , ∗∗, and <sup>∗</sup> indicate p-values <0.001, <0.01, and <0.05, respectively.


FIGURE 8 | Partial effects of individual environmental predictors, derived from the best performing GAM, on CHL anomalies of each region (A–F). Only the main significant predictors are shown (p-value < 0.001). Arrows display the change on CHL anomalies associated with very low (VL), low (L), high (H), and very high (VH) values of the predictor. M corresponds to the midpoint in the range of observed values. Arrows facing up (down) indicate positive (negative) CHL anomalies, while arrows facing right indicate near-zero anomalies. Gray arrows indicate a small-to-medium increase in CHL anomalies, while black arrows indicate a large increase (see also Figure 9).

DIN and NAO. These results show that despite WIC spatial complexity, there is a set of main environmental drivers that contribute to the CHL variability in most regions (**Figures 8**,

predictor are represented on the x-axis. This plot illustrates how the schematic representation in Figure 9 was derived. Arrows facing up (down) in Figure 8 indicate that the predictor is associated to CHL anomalies above (below) 0.1 (–0.1) mg m−<sup>3</sup> (represented by the gray dotted line). If the change (increase or decrease) is higher than 0.2 mg m−<sup>3</sup> , the arrow is black, otherwise it is considered a small-to-medium change (gray arrow).

**9**). Nonetheless, each model indicated a unique combination of significant environmental predictors. The variability of CHL anomalies in region A was seen to be explained by NAO, MLD, PAR, SSH, EA, V, and PO3<sup>−</sup> 4 (40.6% of the total variance explained). Results suggest very high values of the NAO index are associated with positive CHL anomalies. In addition, high MLDs were seen to be paired with negative anomalies. Such results are expected since a shallow or warmer mixed layer contribute to bloom initiation in the Northern Atlantic (Cole et al., 2015). PAR, while being a significant predictor of CHL in this region, did not appear to have a clear association with either positive or negative CHL anomalies. GAM model for region B yielded a wide suite of environmental predictors responsible for CHL variability: DIN, MLD, AMO, SSH, V, and WeMO. Among these, DIN, MLD, AMO, and SSH were considered the most significant. DIN, in particular, is suggested to have a large effect on CHL on region B, as relatively high concentrations were linked to strong positive CHL anomalies. High MLDs exhibited the opposite effect on CHL, i.e., causing negative anomalies. AMO and SSH revealed no clear trend toward either positive or negative anomalies. Concerning region C, GAM analysis identified seven significant predictors: AMO, MLD, SST, V, EA, SAL, and U, with AMO and MLD being the more relevant. Very high negative AMO index values (i.e., colder temperatures) were linked to strong positive CHL anomalies. This analysis also suggested negative CHL anomalies associated with both low and high values of MLD. For region D, the GAM model explained up to 39.9% of its total variance using AMO, EA, V, SSH, NAO, U, and SST as predictors. Similar to region C, very high negative AMO values are also strongly linked to strong positive CHL anomalies. Both strong negative and positive phases of the EA pattern were seen to cause negative CHL anomalies

on region D, particularly during strong positive phases. No clear results were found for V. Region E was the region with the weakest model, only explaining circa 22.8% of its total variance. Unlike other regions, only three environmental variables were identified as drivers of CHL anomalies: AMO, MLD, and U. As with regions C and D, AMO was also fundamental, revealing a common pattern between these three adjacent regions. Region F was the region with the most powerful GAM (52.8% of the total variance explained). SAL, V, DIN, MLD, SST, and NAO were the predictors identified by the model, with SAL, V, and DIN being the most significant. SAL is clearly a relevant predictor in region F, located at Cádiz Gulf, as low-to-average values of SAL in this area are clearly associated with high positive anomalies. Positive CHL anomalies were also found to be related to high DIN concentrations. Moreover, very low and low (Negative) V was linked to CHL positive anomalies, indicating stronger northto-south water transport, while positive values of V suggested a negative effect on CHL anomalies.

## DISCUSSION

## CHL Variability of WIC

Chlorophyll a was seen to be highly heterogeneous over the WIC region, both spatial and temporally. Four CHL coastal hotspots were identified: (1) off the Galician Rias, (2) off central-Northern Portugal, (3) off Tagus estuary, and (4) Gulf of Cádiz (**Figure 2**). On the first two, CHL is promoted by both local riverine nutrient input from local rivers with mixed average annual discharges (Hurrell et al., 2013) and coastal upwelling. Since these nutrient input sources are temporally independent (riverine input peaks during winter–spring, while upwelling occurs during summer– autumn), there is a sustained flow of nutrients almost all-year round. However, a recently identified declining trend in the intensity of coastal upwelling off northWestern Iberia (Pérez et al., 2010) may lead to a future decline in phytoplankton biomass and impact local ecosystems. Off the Tagus estuary, the situation is slightly different, as the major contribution to phytoplankton growth is likely to come from the Tagus riverine discharges. Nonetheless, this region is also influenced by coastal upwelling, which, due to its sheltered location, can lead to harmful algal blooms during relaxation (Moita et al., 2003). Finally, the Gulf of Cádiz is a complex basin mainly fueled by local freshwater discharges and its semi-enclosure, thus explaining why this is the only considered CHL hotspot where maximum concentrations are found on spring (as opposed to summer on the remaining regions).

Overall, the trend analysis suggests there are two regions where CHL appears to be heading separate ways. On the one hand, a positive trend was found in certain areas off Galicia and central-Northern Portugal. As seen above, this region is typically influenced by upwelling (Cravo et al., 2010; Pérez et al., 2010; Guerreiro et al., 2013; Wang et al., 2015; Vidal et al., 2017). Thus, it would be expected that such a positive trend would be linked to coastal upwelling. Several multidecadal studies have noticed upwelling off northWestern Iberia has become less intense the past 40 years (e.g., Pérez et al., 2010; Álvarez et al., 2011; Santos et al., 2011). However, trends in upwelling intensity may be contradictory as it is extremely reliant on factors such as the length of the considered time series, area, and season. Moreover, upwelling intensity is theoretically expected to increase according to future climate scenarios (Bakun et al., 2010; Wang et al., 2015) and recent studies have shown that upwelling off the WIC is no exception as local upwelling favorable winds are expected to increase during the XXI century (Casabella et al., 2014; Álvarez et al., 2017; Sousa et al., 2017). Since this study analyzed a recent period (1998–2016), the positive trend found for CHL over northWestern Iberia corroborates these studies. Overall implications of a positive CHL trend off northWestern Iberia would include a possible increase in productivity of exploitable species (e.g., small pelagic fish). On the other hand, a negative trend was detected off Southern Iberia, particularly on the Eastern Gulf of Cádiz. One reason for this could be the decrease in runoff of the main rivers located on the Gulf of Cádiz, such as Guadiana and Guadalquivir, which is only expected to worsen in the future (Estrela et al., 2012). Since the riverine nutrient input is the primary nutrient source of phytoplankton in the Gulf of Cádiz (Caballero et al., 2014), it would be expected for CHL concentration to diminish in response to lower discharges (Briceño and Boyer, 2010). Such a decline in CHL concentration could lead to impact local ecosystem processes and function (Winder and Sommer, 2012).

## Disentangling CHL Variability: EOF Analyses as a Tool

Each one of the EOF modes exhibited different spatiotemporal patterns which help disentangle CHL variability along WIC during 1998–2016. Mode 1 captured the North Atlantic temperate spring bloom on the more oceanic areas, as these areas are less influenced by other factors that would mask this effect (e.g., upwelling, river discharges, anthropogenic pressure). This demonstrates the importance of the temperate spring bloom for the overall phytoplankton biomass variability over WIC. It also draws attention to the possible consequences of climate change on the magnitude and timing of the Northeast Atlantic spring bloom. For instance, several studies suggest that peak biomass, mean cell size, and diatoms relative abundance should decrease, leading to impacts on zooplankton feeding and, ultimately, to the efficiency of the food web (Sommer and Lengfellner, 2008; Sommer and Lewandowska, 2011), while others suggest a possible mismatch and decoupling between phytoplankton and zooplankton phenology (Edwards and Richardson, 2004).

Overall, mode 2 may be capturing phytoplankton response to the existent biogeochemical and physical meridional gradient between SW and NW Europe. WIC is considered as a transition area due to its climate. According to the Köppen–Geiger climate classification (Peel et al., 2007), the most widely used climate classification scheme, WIC can be divided into three temperate subclimates: temperate with a dry hot summer (36– 40◦N), temperate with a dry warm summer (40–42◦N), and temperate without a dry season and a warm summer (42– 44◦N). While this classification is based on temperature and precipitation for terrestrial regions, this main change from a

temperate climate with dry summer to one without a dry season is reflected on several marine properties. This is supported by the sharp differences across the 42–44◦N strip observed for some environmental drivers in this study (e.g., SST, DIN, SAL, SSH, MLD, PO4, and PAR; **Figure 3**). Among the climate indices considered, NAO has been seen to be linked to differences in wind, temperature, and precipitation between SW and NW Europe (Hurrell et al., 2013). Thus, it is not surprising to see major NAO events (**Figure 4**) roughly reflected in the temporal variability component of mode 2. For instance, 2009–2010 winter NAO index was one of the most negative in recorded history, resulting in unusually high negative SST anomalies and temperatures over Europe (Osborn, 2011). Other recent negative NAO events can also be seen to generally correspond to negative peaks in the temporal component of mode 2 (e.g., 2012–13 and 2016). Also, there is a large positive peak mid-2011 that could correspond to a strong positive NAO phase in 2011–12. All in all, this mode underlines the importance of the influence of climate, mainly captured by NAO. Modes 3 and 4 correspond to much less variance explained, and, thus should be interpreted more carefully. One strong possibility for explaining the spatial cross-shelf positive–negative pattern seen for Mode 4 is that coastal nutrient input is driving this variability mode. Other studies have identified similar modes with cross-shelf patterns related to nutrient input (e.g., Tran et al., 2015; Krug et al., 2017; Liu et al., 2017), supporting this hypothesis.

All in all, most of the considered modes identified by the EOF analyses could be, at least, partly linked to known oceanographical processes. As EOF modes start to explain lower percentages of the total variance, their spatial and temporal components become more complex, resulting in a possible diminishing of the ecological meaning of modes (Navarra and Simoncini, 2010). Such may have happened with mode 3, resulting in a mode with unclear spatial and temporal components.

## Environmental Forcing of CHL Anomalies in Selected Areas

Overall, coherent CHL-variability regions and its drivers were successfully identified for each region, providing relevant results for further understanding of CHL variability over WIC. GAM results suggest that physical processes and climate are the main factors that drive phytoplankton variability in Northern WIC, while in the Southern WIC, phytoplankton is more dependent on nutrient concentration. While nutrient availability is indeed linked to physical processes and climate, DIN, Si, and PO3<sup>−</sup> 4 concentrations do not appear to be limiting phytoplankton in Northern WIC. Region E, interestingly, may be acting as a transition region between regions A and B, thus capturing characteristics from each region. AMO, the main significant predictor identified by the GAM model in region C, is known to be associated with changes in the Atlantic Meridional Overturning Circulation (AMOC; Wang and Zhang, 2013). AMOC influences the heat transport along the North Atlantic, and changes may lead to stronger heat transport and reduced precipitation over WIC (Pohlmann et al., 2004, 2009) that is likely to influence areas C, D, and E. Regarding regions C and D, CHL variability seems to be highly influenced by a combination of basin-wide (e.g., NAO, EA, AMO) and coastal processes (e.g., continental runoff, upwelling–downwelling). While they shared several significant predictors, unique predictors were also identified for each region. MLD and SAL are particularly relevant for region C, the second due to its nature as a proxy of riverine input. Regarding region D, CHL was seen to be linked with V (northward transport) and both very high negative and positive values of EA. Interestingly, since region D overlaps with the NW Iberian shelf, known for its upwelling, it would be expected for positive CHL anomalies to be associated with negative values of V, thus indicating upwelling-favorable winds, which was not observed. One possibility is that northward transport (positive V) may be associated with the Portuguese Coastal Counter-Current (PCCC). PCCC is a well-defined poleward surface current which may occur from September–October to February–May, typically associated with downwelling-favorable winds in the NW (Ambar and Fiúza, 1994; Álvarez-Salgado et al., 2003). During this period, when continental runoff is high, the subsequent plumes from the Galician Rías Baixas and Northern Portuguese rivers (e.g., Minho, Douro) contribute to the PCCC flow (Álvarez-Salgado et al., 2003; Ferreira Cordeiro et al., 2018). Consequently, this leads to shelf-wide mesotrophic conditions off NW Iberia due to the inflow of nutrient-rich waters (Álvarez-Salgado et al., 2003; Otero et al., 2008). Since region D width roughly matches NW Iberia's shelf, the PCCC could be influencing GAM results as CHL positive anomalies during Autumn–Winter could be standing out. Finally, region F is a purely coastal region, and the results reflect that condition. SAL was the main driver, underlying the importance of riverine discharges in this region. This is underlined by the strong association found between high DIN concentrations and CHL anomalies. Negative V was also seen to drive positive CHL anomalies in this region, alluding to the coastal upwelling already seen in this region.

Overall, results suggest CHL variability over WIC is spatially heterogeneous and driven by a multitude of environmental agents. In more oceanic waters, CHL anomalies were seen to be associated with a suite of drivers (e.g., MLD, AMO, and NAO) possibly linked with basin-wide processes, such as the North Atlantic temperate spring bloom or the NAO. Meanwhile, for more coastal regions, upwelling and river discharges were seen to be important factors, with drivers such as SAL and the V and U components. Thus, WIC seems a highly complex area where CHL variability was seen to be highly heterogeneous with oceanic and coastal regions being associated with distinct drivers.

### Final Remarks

Determining how environmental changes drive phytoplankton is essential for understanding oceanic and coastal ecosystems. This study sought to contribute to this matter by focusing on CHL anomalies over the WIC, assessing the most critical environmental drivers and the underlying patterns between them and CHL anomalies. For the first time, this was successfully

achieved at the WIC scale, as this study identified not only the main modes of CHL variability over WIC but also regions with similar variability. This study, most of all, points out that complex areas such as WIC must be handled carefully under the context of environmental management and water quality, and that remote-sensing and modeled data are essential toward unraveling such complexity. The results of this and similar studies are essential for evaluating WIC under the EU MSFD subregion "Bay of Biscay and Iberian Coast," contributing to the knowledge of CHL variability patterns for regions where in situ data are spatially and temporally scarce.

Nevertheless, some limitations should be taken into account: (i) the 19-year dataset here used might not be enough to analyze certain aspects of these complex processes (e.g., how AMO impacts CHL variability due to its multidecadal nature) and (ii) this study was based purely on remote sensing and modeling data, as in situ is scarce and temporally patchy over WIC. Future studies should pursue the use of complementary methodologies, integrating long-term in situ datasets with satellite and modeling data. This would be key to further understand how the phytoplankton community is affected by environmental drivers, especially in terms of: (i) bloom phenology (i.e., bloom timing, including the initiation, duration, and maximum levels); as well as (ii) the community composition and structure. A more in-depth comprehension of processes affecting the phytoplankton community is required to evaluate their implications on local food webs and to

### REFERENCES


contribute to regional environmental marine policies and management strategies.

### AUTHOR CONTRIBUTIONS

The conceptual idea of the manuscript was developed by AB. AB also contributed to data analysis, decision making and manuscript preparation. AF and PG-A both worked on data acquisition and analysis. AF wrote part of the manuscript as well.

## FUNDING

AB was funded by Fundação para a Ciência e a Tecnologia (FCT) Investigador Programme (IF/00331/2013). PG-A and AF also received support from Fundação para a Ciência e a Tecnologia (FCT) Investigador Programme (IF/00331/2013). This study also received further support from Fundação para a Ciência e a Tecnologia, through the strategic project (UID/MAR/04292/2013) granted to MARE.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00044/full#supplementary-material




**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Ferreira, Garrido-Amador and Brito. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Land Uses Simplified Index (LUSI): Determining Land Pressures and Their Link With Coastal Eutrophication

#### Eva Flo\*, Esther Garcés and Jordi Camp

Departament de Biologia Marina i Oceanografia, Institut de Ciències del Mar, CSIC, Barcelona, Spain

Human activities on land result in the high-level production of nutrients. When these nutrients reach coastal waters, they could drive the eutrophication process. Here we present the Land Uses Simplified Index (LUSI), an easy-to-use tool for assessing continental pressures on coastal waters. This assessment is done by indirectly estimating continental nutrient loads and concentrations, and their influence on coastal waters. LUSI is based on systematic information describing both the land uses that influence coastal waters by providing nutrient-rich freshwater inflows (urban, industrial, agricultural, and riverine) and the coastline morphology, which can modify this influence, as it determines the degree of coastal water confinement and therefore the likelihood that these inflows will be diluted. A low LUSI value indicates that coastal waters are not or only slightly influenced by continental pressures and/or that these pressures are diluted. On the contrary, a high LUSI value indicates that coastal waters are strongly influenced by continental pressures and/or that these pressures are not diluted. LUSI fulfills a methodological gap, as a simple method to assess coastal pressures when there is a lack of information. Furthermore, it fulfills the requirement of the Water Framework Directive for a true pressure assessment (i.e., not confounded with impact), which for coastal waters imply using pressure data from land. An additional and important feature of LUSI is that it allows the establishment of pressure-impact relationships with impact indicators, such as those related to the Biological Quality Elements of the above Directive. For example, a relationship based on LUSI, as a proxy of pressure, and on the chlorophyll-a concentration, as a proxy of phytoplanktonic biomass and, thus, of the eutrophication impact. By providing insights into the land uses that trigger eutrophication in coastal waters, LUSI aids in the design of measures aimed at remediating anthropogenic damage caused to the environment.

Keywords: land uses, pressure assessment, coastal waters, pressure-impact relationship, chlorophyll-a, dissolved inorganic nutrients, water framework directive

## INTRODUCTION

In marine ecosystems, coastal areas are of major environmental, economic, and social importance. They are among the most diverse and biologically productive ecosystems on Earth, with the net primary production of phytoplankton estimated to be 50 Pg C y−<sup>1</sup> (Malone et al., 2017). Coastal areas also provide other important ecological functions, including filtering terrestrial inflows, and

#### Edited by:

Jesper H. Andersen, NIVA Denmark Water Research, Denmark

#### Reviewed by:

Joanne I. Ellis, King Abdullah University of Science and Technology, Saudi Arabia Ana C. Brito, Universidade de Lisboa, Portugal

> \*Correspondence: Eva Flo evaflo@icm.csic.es

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 16 August 2018 Accepted: 15 January 2019 Published: 19 February 2019

#### Citation:

Flo E, Garcés E and Camp J (2019) Land Uses Simplified Index (LUSI): Determining Land Pressures and Their Link With Coastal Eutrophication. Front. Mar. Sci. 6:18. doi: 10.3389/fmars.2019.00018 Flo et al. Land Uses Simplified Index

offering food and shelter for a wide variety of organisms (Creel, 2003). For humans, they are "golden areas" of marine socioeconomic development (Zhang, 2012), sustaining major activities related to agriculture, fisheries, industry, urbanization, transport, and tourism (Springer et al., 1996; Cole and McGlade, 1998; San Vicente, 1999; Creel, 2003; Lopez y Royo et al., 2009). Coastal areas comprise 20% of the Earth's surface yet contain over 45% of the entire human population (Schäfer et al., 2010). The population in low-elevation coastal zones (below 10 m of elevation) in the year 2000 was 625.2 million people and is expected to increase to 938.9 million people by the year 2030 and 1,318.3 million people by the year 2060 (Neumann et al., 2015). Therefore, coastal areas are strongly affected by anthropogenic pressure, in the form of high population densities and intense human activities (Lopez y Royo et al., 2009). Coastal development has the potential to produce an excess of nutrients that are delivered to coastal waters via several routes, such as rivers, run-off, and submarine groundwater discharges. Moreover, human activities modify water flows, such as by the construction of dams, increased water extraction, and the deviation of rivers, resulting in profoundly altered nutrient loads, and concentrations (Scialabba, 1998) arriving at coastal waters. Consequently, the risk of eutrophication is higher in coastal waters than in other marine ecosystems.

Several policies have been enacted throughout the world with the aim of restoring and protecting coastal waters, including the Clean Water Act (United States Congress., 1972), The Water Act 2007 (Australia Office of Parliamentary Counsel., 2007), the Water Framework Directive (WFD) (European Commission, 2000), and the Marine Strategy Framework Directive (European Commission, 2008). For all of them, one of the main objectives is to combat eutrophication, which implies an ability to assess both its impact and the pressures that give rise to it, and therefore devise remediation measures. In the case of the WFD, it mandates assessments of true coastal pressures (i.e., not confounded with impact) and the establishment of pressure-impact relationships.

Pressure and impact related to coastal eutrophication have been broadly examined. Among the methods used to assess the degree of eutrophication of coastal waters, chlorophyll-a values are the most widely accepted impact indicator (Bricker et al., 2007; Ferreira et al., 2007; Nixon, 2009; Borja et al., 2012). Although coastal pressures are not easy to quantify and proof of their impact is hard to obtain, several assessment methods are available. Quantitative methods are typically based on readily measurable variables, such as wastewater discharges or human population size (Bonamano et al., 2016; Cai et al., 2016). Others evaluate land use from land cover maps (Comeleo et al., 1996; Jordan et al., 1997; Nedwell et al., 2002; Giupponi and Vladimirova, 2006; Rodriguez et al., 2007; Xian et al., 2007; Lopez y Royo et al., 2009, 2010; Tran et al., 2010) or satellite images (Barale and Folving, 1996; Stefanov et al., 2001; Lopez y Royo and Casazza, 2007; Xiao and Weng, 2007). Furthermore, Aighewi et al. (2013) proposed incorporating land use and land cover information into pressure models to obtain a more holistic assessment of coastal pressures. Semi-quantitative methods have also been proposed, including by Aubry and Elliott (2006); Lopez y Royo et al. (2009); European Commission (2011a); Neto and Juanes (2014), and Batista et al. (2014). These methods yield an approximation of a quantity, usually by a gradient of categories; for instances low, moderate or high population density. Other authors, including Andrulewicz and Witek (2002); Borja et al. (2011), and Weisberg et al. (2008), have advocated pressure assessments based on expert judgment. However, most of the various proposed methodologies to assess coastal pressures are complex or require large amounts of data.

Here we present the Land Uses Simplified Index (LUSI), an easy-to-use tool to assess continental pressures on coastal waters, and a case study to highlight its utility, from the Catalan coast (NW Mediterranean). The first aim of that study was to use LUSI to assess continental pressures on coastal waters and the second to establish a pressure-impact relationship based on LUSI and chlorophyll-a concentrations, as a proxy of phytoplanktonic biomass and, thus, of the eutrophication impact. Additional aims were to validate LUSI against dissolved inorganic nutrient concentrations measured in coastal waters and to determine whether land cover map selection influences the results of LUSI. Furthermore, LUSI utility to assess continental pressures and to establish pressure-impact relationships with the Biological Quality Elements of the WFD in other coastal areas were explored bibliographically. The results demonstrate the strengths of LUSI, its ability to improve scientific knowledge regarding coastal eutrophication, and its applications, especially in the development of remediation measures.

## LAND USES SIMPLIFIED INDEX (LUSI)

The main objective of LUSI is to assess coastal pressures related to eutrophication. LUSI serves as a proxy enabling the indirect assessment of continental nutrient loads and concentrations, and their dilution in coastal waters. Therefore, it estimates the eutrophication risk of coastal waters. It is based on systematic information describing both the anthropogenic land uses that influence coastal waters (urban, industrial, agricultural, and riverine) and coastline morphology. The latter determines the degree of coastal water confinement and therefore the likelihood that continental freshwater inflows and the nutrients they contain will be diluted. LUSI not only fulfills the methodological gap, by providing a simple method to assess coastal pressures when there is a lack of information, but also the requirements of the WFD, by yielding a true pressure assessment (i.e., not confounded with impact) and allowing the establishment of pressure-impact relationships with impact indicators, such as those related to the Biological Quality Elements of the WFD.

### Rationale

Ketchum (1972) defined the coastal area as the band of dry land and adjacent ocean space (water and submerged land) in which terrestrial processes and land uses directly affect oceanic processes and uses, and vice versa. One of the most important coastal process boosted by continental pressures is eutrophication. Eutrophication is driven by nutrients and it has been greatly enhanced by human activities on land, which result in the high-level production of nutrients that reach coastal waters (Chislock et al., 2013). Accordingly, the rationale for LUSI is based on the following assumptions:


#### Requirements

LUSI requires information on the anthropogenic land uses that influence coastal waters and on coastline morphology. The pressures taken into account by LUSI include agricultural (irrigated land only), industrial, and urban land uses as well as riverine effects. Information on land uses and coastline morphology is available from different sources, including governmental sources, such as census data, satellite maps, such as those from Landsat or Google Earth, airplane, and drone survey images or combinations of them, as suggested by Lautenbach et al. (2011). However, land cover maps are the most useful for the calculation of LUSI, as they provide information on land use, the area occupied by the various types of land use, and the morphology of the coastline. There are several publicly available land cover maps with different degrees of coverage (continent, country, and region). Their appropriateness with respect to LUSI depends on the area of interest or whether distinct areas will be compared. For example, for Catalonia, in the NW Mediterranean, three maps are available: the Coordination of Information of the Environment land cover map [CORINE land cover map or CLCM; European Enviromental Agency (2012)], which covers Europe; the Sistema de Información sobre Ocupación del Suelo de España [SIOSE; Instituto Geográfico Nacional. (2011)], which covers Spain; and the Mapa de Cobertes del Sòl de Catalunya [MCSC; Centre de Recerca Ecològica i Aplicacions Forestals (2009)], which covers Catalonia. To determine riverine influences, the mean salinity value of the coastal water area of interest must be obtained. A truly representative value implies the need for a raw dataset acquired by a sampling frequency sufficient to capture the variability in the salinity. Such information is sometimes available from water management or environmental agencies. Ideally, for the calculation of LUSI, the land cover map and salinity dataset should cover the same time period.

## Protocol

To calculate LUSI for a coastal water area, its quantitative information on pressures is classified into categories and assigned a score; then, all the scores are summed and multiplied by a correction factor related to coastline morphology. The protocol to calculate the LUSI is as follows:


TABLE 1 | Constituents of storm waters from areas differing in their predominant land use.


Biological oxygen demand (BOD5), total suspended solids (TSS), ammonium (NH4), total Kjeldahl nitrogen (TKN), total nitrogen (TN), phosphate (PO4), and total phosphorus (TP) concentrations (C) in mg/L and loads (L) in kg/ha/year are shown. Modified from Kadlec and Wallace (2009).

maximum salinity was established at 38.4. Riverine pressure is divided into three categories and each one is associated with a score (**Table 2**). These three categories were established following those of the Water Framework Directive Intercalibration Process for the Mediterranean Sea regarding the specific typology for Biological Quality Element Phytoplankton (Camp et al., 2016).


#### **LUSI** = (**urban score** + **agricultural score** + **industrial score** + **riverine score**)× **coastline correction factor**

LUSI values provide a semi-quantitative assessment of continental pressures on the coastal waters of the studied area. They have no units and range from 0.75 to 8.75. A low TABLE 2 | Pressures categories and their scores used to calculate LUSI.


Land use pressures [urban, agricultural (irrigated), and industrial] are classified based on the percentage of land coverage (%LC) used for the respective activities, and riverine pressure based on the mean salinity of coastal water.



LUSI value indicates that coastal waters are not or only slightly influenced by continental pressures and/or that these pressures are diluted. On the contrary, a high LUSI value indicates that coastal waters are strongly influenced by continental pressures and/or that these pressures are not diluted.

LUSI is to be applied where there is no previous information on continental pressures on coastal waters. However, if information from the area of study is available, such as punctual or diffuse continental nutrients loads reaching coastal waters or coastal waters residence times, the LUSI can be modified to include it. For example, for concave coastal areas with long residence time it can be assumed that the continental inflows would be diluted at a lower rate than in concave areas with short residence time. In the former case, the confinement correction factor for concave coastal morphology can be increased accordingly to the residence time to obtain a more accurate value of the influence of land pressures on the studied coastal area.

### CASE STUDY: CATALAN COAST (NW MEDITERRANEAN)

#### Materials and Methods Study Area

The Catalan coast (**Figure 1**) is located between 3◦ 19′ 59.94′′ E and 42◦ 29′ 0.09′′ N and between 0◦ 9 ′ 41.69′′ E and 40◦ 31′ 27.56′′ N. It occupies 7257.50 km<sup>2</sup> (Agència Catalana de l'Aigua, 2005a) and delimits the Mediterranean Sea over 870.0 km (Institut d'Estadística de Catalunya., 2010). The geography, demographics, and socioeconomic development of the Catalan coast are representative of the NW Mediterranean coast (Flo et al., 2011b).

The continental topography ranges from rocky and steep headlands to sandy and flat bays. There are also deltaic areas, the most important of which is the Ebre delta (Serra and Canals, 1992). This landscape encompasses several watersheds, which consist of ephemeral streams, nine medium to small rivers, and the Ebre River in the south, all of which open directly into the Mediterranean Sea. The Ebre River drains a watershed of 84,230 km², with a mean water discharge at the river's mouth of 416 m<sup>3</sup> /s (Ludwig et al., 2009). Other major rivers in Catalonia drain an area of 13,400 km² and have a mean water discharge of 0.3–16.3 m<sup>3</sup> /s (Liquete et al., 2009). Land use differs along the river basins, with agriculture accounting for 0.4–49.2%, forests for 17.2–77.2%, and urban areas for 5.5– 81.7% (Institut d'Estadística de Catalunya, 2015). Agricultural land use is relatively important in southern river basins and urbanization in central river basins. In terms of surface area, 13.8% of the coastal zone is urbanized (Biblioteca del consorci el Far., 2010; Institut d'Estadística de Catalunya, 2015). The population in the coastal zone is 4,942,044 inhabitants, which represents 66% of the total population in Catalonia (Institut d'Estadística de Catalunya., 2016). However, the population density is highly variable along the coast, ranging from 33 inhabitants/km<sup>2</sup> in the coastal area of the Ebre basin (Ferré, 2007) to 15,319.6 habitants/km<sup>2</sup> in the Barcelona metropolitan area (Institut d'Estadística de Catalunya., 2016).

The underwater topography is complex: the continental shelf is usually narrow [average = 15–20 km; Maldonado (1995)] but wide along its border with the Ebre delta (maximum = 54.5 km), and almost non-existent (minimum = 1.6 km) in front of several canyons located along the coast (Platónov, 2002). The general surface circulation is due to the Liguro-Provençal current, which moves from the NE toward the SW; close to the shore, local superficial currents move toward the NE or also toward the SW (Agència Catalana de l'Aigua, 2005b). The tidal range is small and the sea weather, according to the Douglas scale (Harbord, 1897), is typically described as slight and occasionally as rough or very rough, especially during autumn (Agència Catalana de l'Aigua, 2005b).

Following the implementation of the WFD, the Catalan coast was divided into 36 water bodies (C01 to C35, from NE to SW), including the two bays situated in the north (T01) and south (T03) of the mouth of the Ebre River (**Figure 1**; **Supplementary Table 4**).

#### Data Collection

Pressure information to calculate LUSI was obtained from two sources. First, the land cover map of Catalonia, the MCSC, was used to determine the three land use pressures [urban, agricultural (irrigated), and industrial] and to establish the coastal morphology of each water body. The most recently available version of the map is from 2009. Second, the riverine pressure was determined by using the salinity data from the National Catalan Coastal Water Monitoring Program, conducted by the Catalan Water Agency (ACA) in collaboration with the Marine Science Institute (ICM-CSIC). As the implementation of the WFD follows 6-year cycles, the salinity data covered the period between 2007 and 2012 and matched the year represented by the land cover map.

Dissolved inorganic nutrient concentrations, which are used to validate LUSI, were also gathered for the same period and program.

To test whether the choice of the land cover map influences the results of LUSI, LUSI was also calculated using the land cover map of Europe, the CLCM, from 2006. The characteristics of the two maps are detailed in **Table 4**.



Chlorophyll-a concentrations to establish a pressure-impact relationship together with LUSI values were also obtained for the same period and Program.

Data on salinity, chlorophyll-a and dissolved inorganic nutrient concentrations were collected from 80 sampling stations in Catalan coastal inshore waters (Flo, 2014) (**Supplementary Figure 1**). These stations are located between 0 and 200 m from the shore, depending on the water depth (<2 m depth), and are distributed along all water bodies, with 1–7 stations per water body, depending on the coastal length. The surface water at the stations was sampled 4 or 12 times a year, depending on the variability of the measurements, previously established. Salinity was directly measured using a WTW probe (model 315). Total chlorophyll-a was quantified in 60-mL subsamples filtered through 25-mm Whatman GF/F glass-fiber filters that were stored frozen. The filters were subsequently extracted in 8 mL of 90% acetone for 48 h, and concentrations of chlorophyll-a (µg/L) were measured using a Turner Designs fluorometer, following the method of Yentsch and Menzel (1963). Dissolved inorganic nutrients were determined in 50-mL subsamples that had been frozen upon their arrival in the laboratory. Nitrate (NO3), nitrite (NO2), ammonium (NH4), phosphate (PO4), and silicate (SiO4) concentrations (µM) were measured using an autoanalyzer (Evolution II, from Alliance Instruments, and AA3 HR Bran+Luebbe, from Seal Analytical) and the colorimetric techniques of Grasshoff et al. (1983).

#### Data Processing and Statistics

Information on land use pressures was processed from CLCM and MCSC as follows. The 44 categories of the thirdlevel legend of the CLCM and the 241 categories of the fifth-level legend of the MCSC were assigned either to the three land use pressure types considered within LUSI [urban, agricultural (irrigated), or industrial] or to the category "other" (**Supplementary Tables 1**, **2**). The percentage of land coverage related to these pressures was established using Miramon software (Pons, 1994) for each map and water body, after which pressure scores were assigned accordingly. The low urban scores of two water bodies located in the Barcelona Metropolitan Area (C21-Llobregat and C22-El Prat de Llobregat-Castelldefels) were later changed by expert judgment to the maximum urban score. This modification was performed because the maps greatly underestimated the urban score in the area of land neighboring these water bodies. For example, several areas of land within the airport premises were classified as meadows and grasslands according to level 5 of the MCSC, but they should have been classified as airport. The same changes were made for the CLCM and MCSC.

Information on riverine pressure (salinity) and on dissolved inorganic nutrient and chlorophyll-a concentrations was processed as follows. During the study period, 3,562 samples were collected, with 23–300 samples per water body, depending on its size. The mean salinity of each water body was calculated and a riverine pressure score assigned accordingly. The 90th percentiles of the chlorophyll-a and dissolved inorganic nutrient concentrations were determined for each water body using six different methodologies [methodologies 4–9 from Hyndman and Fan (1996)]. This statistical parameter was chosen following the WFD intercalibration process agreements for chlorophylla. Since chlorophyll-a and dissolved inorganic nutrient concentrations are not normally distributed (as determined from histograms and Shapiro-Wilk tests), the obtained values were log-transformed following Equation (2):

$$\mathbf{v}' = \log \mathbf{l} \mathbf{0} (\mathbf{v} + 1)$$

as is commonly done for environmental data (Cassie, 1962; Legendre and Legendre, 1979; Heyman et al., 1984; Zar, 1984). The mean and standard deviation of the dissolved inorganic nutrient and chlorophyll-a concentrations were calculated for each water body with their six log-transformed 90th percentiles.

The continental pressures on coastal waters were then assessed by calculating LUSI for each water body based on the CLCM and MCSC maps, using the previously assigned pressure scores and taking into account the morphology of the coast.

To validate LUSI against dissolved inorganic nutrient concentrations and to establish a pressure-impact relationship based on LUSI and chlorophyll-a concentrations, first, the corresponding Spearman correlations (ρ) were determined together with the respective p-values. Linear models between LUSI values and dissolved inorganic nutrient concentrations and between LUSI values and chlorophyll-a concentrations were then adjusted, their p-values were calculated, the goodness of fit (R 2 ) determined and the diagnostics checked. The means of the six transformed 90th percentiles of dissolved inorganic nutrient and chlorophyll-a concentrations were used to adjust the models; the standard deviations of the same variables were used to weight the models; and for each model the 90% confidence interval was calculated.

To test whether the choice of land cover map affects the results of LUSI, the LUSI values calculated with MCSC and CLCM were compared using the Wilcoxon matched-pairs signed-rank test. A regression line of the LUSI values calculated using MCSC and CLCM was then compared to an identity regression using the analyses of covariance (ANCOVAs). A first analysis included the interaction of the independent variables to determine whether the slopes were significantly different; in a second analysis, the interaction was omitted to determine whether the intercepts were significantly different.

The level of statistical significance for all performed tests was set at 0.01.

Statistical tests were performed using R (R Development Core Team, 2008) and STATISTICA software (StatSoft, 2003). Linear models were plotted using R software and the maps were drawn using Miramon software.

#### Results and Discussion

#### Assessment of Continental Pressures on Coastal Waters Using LUSI

**Figure 1** (**Supplementary Figure 2**) shows that nearly the entire Catalan coast is influenced by continental pressures to some extent, with LUSI values ranging between 0.75 and 6.25. Lower LUSI values were obtained in areas where natural land use accounts for a high percentage of land coverage; as is the case for water body C05 (0.75), adjacent to the Cap de Creus marine and terrestrial natural park located in northeast Catalonia. Coastal areas that receive important fluvial inflows, such as water bodies located around the Ebre delta (C33 and C34; 5), the Ebre bays (T01 and T03; 6.25), or water bodies receiving the waters of the Muga (C07; 6.25), Fluvià (C08; 6.25), and Ter (C11; 6.25) rivers, located on the northeast Catalan coast, had the highest LUSI values. High LUSI values were also determined for coastal zones with the highest percentage of urban land coverage, primarily the Barcelona Metropolitan Area, in the center of the Catalan coast (C21; 5). However, the values in the metropolitan area were not as high as those of areas influenced by Catalonia's main rivers.

This assessment reveals the Catalan areas where continental pressures are able to fuel the eutrophication of coastal waters.

#### **Validation of LUSI using dissolved inorganic nutrient concentrations**

A validation implies a demonstration of a significant relationship between two independent datasets of indicators for the same parameter, comprising one of the datasets the results of the method to be tested. Accordingly, LUSI was validated by comparing its results with dissolved inorganic nutrient concentrations from coastal waters, given that land pressures, indirectly evaluated by LUSI, are mainly related to these concentrations.

The correlation and adjusted linear model based on the LUSI values and the dissolved inorganic nutrient concentrations were significant (**Table 5**; **Supplementary Figure 3**). The lowest correlation was 0.6 and the highest 0.87. All linear models showed positive relationships, with slope values ranging between 0.05 and 0.21.

These results validated LUSI as a proxy for indirectly assessing continental nutrient loads and concentrations and therefore continental pressures on coastal waters.

#### **Comparison of LUSI calculated with different land cover maps**

LUSI values calculated using two different land cover maps, with regional and European coverage, were compared to test whether the choice of the land cover map influences the results of LUSI.

European (CLCM) and regional (MCSC) land cover maps are based on different reference images, as CLCM uses images captured by SPOT5 and MSCS those from LANDSAT (**Table 4**). In addition, the images were captured during different times:



Spearman correlations (ρ), and their corresponding p-values together with the equations of the linear models (LM), their R<sup>2</sup> values, and their p-values are shown. See the text for details on the calculation.

the CLCM images in 2005 and those of MSCS in 2009. The two maps also differ in their legend levels, as level 3 of CLCM has 44 categories and level 5 of MSCS 241 categories. Nonetheless, 95% of the Catalan water bodies had similar LUSI values when calculated with MCSC and CLCM (1.25 was the maximum difference), and 67% of the water bodies had the same LUSI values. According to the Wilcoxon matched-pairs signed-rank test, there were no statistically significant differences between the two sets of LUSI values (p = 0.045). A comparison of a regression line of the LUSI values calculated using MCSC and CLCM and an identity regression also failed to reveal statistically significant differences between the slopes (p = 0.043) and the intercepts (p = 0.024) (**Figure 2**).

These results showed that the choice of land cover map does not significantly affect the calculation of LUSI, as similar continental pressure information is obtained.

#### Pressure-Impact Relationships

The WFD requires that pressure indicators are unambiguously linked with biological impact indicators, as this will allow an elucidation of the involved mechanisms and increase the probability that management actions will be effective. For coastal waters, these relationships should be established for the following Biological Quality Elements: phytoplankton, macrophytes, angiosperms, and macrofauna. Regarding coastal eutrophication, a pressure-impact relationship can be established between LUSI and chlorophyll-a, as a proxy of phytoplanktonic biomass and, thus, of the eutrophication impact. The mechanism underlying this relationship is related to nutrients: nutrient-rich freshwater and riverine inflows reach coastal waters, where nutrients boost phytoplanktonic growth, thus, enhancing eutrophication.

#### **Establishment of a pressure-impact relationship based on LUSI and chlorophyll-a for the catalan coast**

For the Catalan coast, the link between LUSI and chlorophylla is depicted in **Figure 3**: coastal waters adjacent to areas of land dominated by human activities or coastal waters receiving water from the Catalonia's main rivers had higher chlorophylla concentrations than areas with less human pressure or without rivers. Moreover, the effect of coastal morphology on the confinement of coastal waters could also be visualized: chlorophyll-a concentrations were higher in bays, where water is confined, than in the waters around headlands, where continental freshwater inflows become more diluted. The correlation and adjusted linear model (i.e., pressure-impact relationship) based on the LUSI values and chlorophyll-a concentrations were significant (**Table 5**; **Figure 4**; **Supplementary Figure 4**; **Supplementary Table 3**). The correlation (0.88) confirmed the relationship between LUSI and chlorophyll-a, while the positive slope coefficient (0.12) of the pressure-impact relationship implied that chlorophyll-a is positively related to LUSI. Therefore, water bodies receiving no or minimal pressure from the continent are either without signs of impact or the impact is minimal (low LUSI and chlorophyll-a values). Conversely, water bodies subject to continental pressures comprise more highly impacted waters (higher LUSI values are related to higher chlorophyll-a values). These results show that continental pressures are linked with eutrophication along the Catalan coast. However, it should be clarified that in the Mediterranean Sea high chlorophyll-a concentrations do not necessarily imply changes in the balance of organisms or in the development of hypoxia or anoxia, as is the case in other seas (Diaz and Rosenberg, 1995; Breitburg, 2002; Gray et al., 2002). Moreover, the obtained pressure-impact relationship between LUSI and chlorophyll-a fulfills the requirements of the WFD.

The theoretical background concentration of chlorophylla in the absence of continental pressures, calculated using the linear model, was compared with the natural background concentration of chlorophyll-a measured in coastal waters to validate the established pressure-impact relationship. The theoretical background concentration of chlorophyll-a was calculated using the minimum LUSI value of 0.75; the result was a minimum chlorophyll-a concentration of 1.04 µg/L. This value is similar to the 1.01 µg chlorophyll-a/L measured in the water body of Cap Norfeu (C05), located within the Cap de Creus marine and terrestrial natural park, in northeast Catalonia. Its LUSI value was 0.75, indicating that it is subject to minimum pressure from the continent. The similarity between the theoretical and measured minimum chlorophylla values supports the validity of the previously established

pressure-impact relationship. Furthermore, 1.04 µg chlorophylla/L can be established as the reference condition for the Biological Quality Element Phytoplankton in the Catalan coast.

The above pressure-impact regression and the reference condition for the Biological Quality Element Phytoplankton are applicable for the entire Catalan coast, as all the water bodies of this area were considered in their establishment. However, this kind of exercise can be performed with subsets of data, considering the desired factor: land use pressure, riverine pressure, or coastline morphology. For example, to establish the reference conditions in the Catalan coast for the Biological Quality Element Phytoplankton within the Water Framework Directive Intercalibration Process for the Mediterranean Sea, three regressions were performed considering the three types of the specific typology for this element, which are based on salinity values and are the same of those categories established for riverine pressure; as a result, three different reference conditions were established (Camp et al., 2016).

#### **Comparison of pressure-impact relationships based on LUSI and biological quality elements of the WFD from several coastal areas**

Within the context of the WFD implementation, LUSI has been used together with several impact indicators related to all the Biological Quality Elements linked to coastal waters to establish pressure-impact relationships in the Mediterranean and Black Seas and in the Atlantic Ocean. The broad and successful use of LUSI was made possible by the open distribution of its rationale and its description within the Milestone 5 report regarding Intercalibration Phase 2 of the WFD (Flo et al., 2011a). An extended bibliographic exploration, including scientific articles but also oral presentations and technical reports, resulted in 48 relationships found, performed with independent datasets (**Table 6**). Most of the cases reinforced the well-performing of LUSI in assessing continental pressures and in establishing these relationships. Furthermore, they demonstrated the utility of LUSI beyond the Catalan coast, to other coastal areas of the Mediterranean Sea but also of the Black Sea and the Atlantic Ocean, and in relation to other Biological Quality Elements, in addition to phytoplankton.

The comparison of the pressure-impact relationships based on LUSI values and biological impact indicators shows a very broad range of goodness of fit values, from 8 to 93% (**Table 6**). In some cases, well-defined, linear pressure-impact relationships could not be established, as indicated by the low goodness of fit values. In other cases, however, interactions between pressures and impacts were simpler, more readily identifiable and better known, resulting in liner relationships with a high goodness of fit. Relationships related to angiosperms showed acceptable goodness of fit (minimum value of 61%) while those related to macrofauna were generally poor fitted (maximum value of 31%). Macrophytes and phytoplankton relationships showed high and low goodness of fit. Nonetheless, in all cases the results should be carefully interpreted, as for example a high goodness of fit value does not always imply causality.

The disparity among goodness of fit values of the compared pressure-impact relationships exemplify the complexity of these relationships. For the marine environment and its functioning, the calculated pressure-impact relationships are a simplification. Generally, they are theoretical and linear, indicative of a unique and direct mechanism linking an identified pressure with a specific impact. These relationships, while a very useful tool in

WFD implementation, especially for management purposes, are nonetheless rare in the environment, where multiple pressures, both natural and anthropogenic, interact in multiple ways to affect a wide range of organisms, resulting in numerous impacts throughout the ecosystem. In nature, interactions among pressure, and impact indicators create complex networks and involve physical, chemical, and biological mechanisms that coexist in time and space (Garcés and Camp, 2012). The influence of these factors on the area of study can modify the goodness of fit of the pressure-impact relationship.

An example of the influence of other factors in the fitting of a pressure-impact relationship is that of the distance to the coast of the sampling stations when establishing this relationship between LUSI and the Biological Quality Element phytoplankton. The relationship between continental nutrients and chlorophylla is more direct at the coastline than in outermost waters, where nutrient availability decreases and other complex oceanographic mechanisms that affect phytoplankton growth come into play (Basterretxea et al., 2018). For Mediterranean Spanish coastal waters, the goodness of fit of pressure-impact relationships between LUSI and chlorophyll-a concentration was higher when chlorophyll-a was sampled at the coastline (62%) than at coastal outermost waters (51%) (**Table 6**). Consequently, current sampling strategies for chlorophyll-a should be reconsidered: there should be greater focus on those sites where continental pressures are stronger and the risk of eutrophication is higher, thus where clearer pressure-impact relationships are more likely. These sites are near the coastline, not at the offshore coastal boundary, where the relationships are more complex and the risk of eutrophication is lower. This change will not only cut the expenses of sampling strategies, as those areas, at least in the Mediterranean Sea, are only reachable on food from the coastline and thus it is not necessary to use a boat, but also it will provide better information on the areas that are at higher risk of eutrophication (i.e., the coastal inshore waters).

Another example of factors that can influence the fitting of a pressure-impact relationship is LUSI modifications (**Table 6**). Several authors have adapted LUSI to a specific coastal area (Derolez, 2011; European Commission, 2011b; Cicero, 2012; Alcoverro, 2013; Thomas-Bourgneuf, 2013; Marin et al., 2015; Revilla, 2015) in order to include already know information of land pressures. These adaptations necessitated modifying some of LUSI's characteristics, such as adding more categories for a specific pressure, considering other kinds of pressures, changing the confinement correction factor, or taking into account a different continental area. Some modifications do

not significantly affect the calculation of LUSI, such as take into account a different but similar continental area. However, the more closely adapted the modified index is to the specific characteristics of the coastal area of interest, the less useful it is for other coastal areas with different characteristics and, importantly, for subsequent comparisons between them. Accordingly, to compare pressures assessments of different coastal waters preformed with the original LUSI and with a modified LUSI an intercalibration exercise should be previously done. LUSI modifications could imply a loss of simplicity and a lack of comparability, however, they are a suitable option to obtain more accurate pressure assessments. In all cases when a modified LUSI is used, it is advisable to carefully interpret the results.

The comparison of several pressure-impact relationships (**Table 6**) has allowed to establish some evidences. First, the sea or ocean of study seems not to influence the performance of the relationships, as for example for the Mediterranean Sea there are well-defined relationships and others that are not acceptable. Second, for pressure-impact relationships based on LUSI and chlorophyll-a, the distance to the coast of the sampling stations is key, as LUSI has been developed to be confronted with impact data from the coastal inshore waters, where the pressures from land are more evident and pose a risk of eutrophication. Third, Biological Quality Elements phytoplankton, angiosperms, and macrophytes can provide well-defined relationships, while the examples found in the literature did not show any acceptable example for macrofauna. Fourth, LUSI modifications provide well-defined relationships when applied in the area for which it was developed. For example, LUSI-VAL applied in Valencia together with chlorophyll-a provided a goodness of fit of 87%. However, their performance it is not assured when applied to other areas. For example, LUSI-VAL together with chlorophylla provided an acceptable relationship in the Marmara Sea (54% of goodness of fit) but not in the South-eastern Black Sea (37% of goodness of fit). Even though these evidences, more information is needed to establish the differences and the commonalities regarding the factors that intervene in the calculation of LUSI and in the posterior establishment of pressure-impact relationships.

#### Guidance on Management Actions Provided by LUSI Regarding Eutrophication

The restoration and protection of coastal waters against eutrophication requires information on the pressures that strengthen its impact. With this knowledge, decision-makers can decide on where management actions are needed and the types of action that will be most effective. LUSI provides three kinds of information to guide this decision-making process.

First, the assessment of continental pressures using LUSI yields information on coastal areas vulnerable to eutrophication and thus at high risk (high LUSI values; **Figure 1**). These areas will require the implementation of a management plan if eutrophication is indeed detected. However, in the absence of evident eutrophication, the implementation of a monitoring plan as an early warning system will suffice.

Second, the pressure-impact relationship established between LUSI and chlorophyll-a concentrations identifies the coastal areas characterized by a mismatch between these values (**Figure 4**). For example, in coastal areas with a lower or higher than expected chlorophyll-a concentration, given their LUSI value, other mechanisms related to eutrophication are probably TABLE 6 | Pressure-impact relationships between LUSI, as pressure (P), and biological indicators related to the Biological Quality Elements of the Water Framework Directive for coastal waters, as impact (I).


(Continued)

#### TABLE 6 | Continued


For each relationship, its equation, R<sup>2</sup> value, p-value, the coastal area for which it was calculated, and its reference are shown. Relationships are sorted by Biological Quality Element and then by its R<sup>2</sup> value.

<sup>a</sup>Calculated from correlation.

<sup>b</sup>Calculated from data.

in play; for example, continental nutrients may be diluted at a different rate than expected or transport mechanisms are exporting or importing nutrients. In these cases, further studies aimed at a better understanding of the complexity of the area are warranted before the design of a management plan. Water body C19, located in front of Barcelona, is an example of a water body with a lower than expected chlorophyll-a concentration (2.43 vs. 4.01 µg/L) given its LUSI value (4). In this case, the city's wastewater systems, which include wastewater treatment plants, submarine outfalls, pumping stations, and upstream collectors (Agència Catalana de l'Aigua, 2011), are able to partially neutralize the continental pressures. This also demonstrates that disagreements between measured and expected chlorophyll-a values can be used to test the efficiency of wastewater treatment systems. Conversely, C22, located at the south of the Barcelona Metropolitan Area, is an example of a water body with a higher than expected chlorophyll-a concentration (6.06 vs. 4.01 µg/L) given its LUSI value (4). This water body receives nutrients that reached coastal waters in the northern adjacent water body C21. Continental nutrients from the Barcelona Metropolitan Area and from the Llobregat river reach water body C21 and then are transported by the general surface circulation dictated by the Liguro-Provençal current that moves from NE to SW to the water body C22. Thus, management of water body C22 should take into account the transfer of nutrients from adjacent areas, with subsequent actions directed at water body C21, where the nutrients reach coastal waters.

Third, the LUSI scores provides insights into the land uses that trigger the eutrophication of coastal waters and therefore guide the choice of management actions if an impact of eutrophication is detected. For the Catalan coast and the studied period, the assessment of the impact based on chlorophyll-a concentrations and on the methodology established by Camp et al. (2016) resulted in the identification of four water bodies impacted by eutrophication (**Table 7**; **Supplementary Table 4**). These water bodies require management plans aimed at remediating the negative effects of continental nutrient arrival. The LUSI scores of these coastal areas provides information on the human activities occurring on land that could be responsible for the eutrophication detected in coastal waters, and therefore on the type of management actions that should be included in these plans. In the case of the Catalan coast, the management plans for the four water bodies should include actions related to urban land uses, as all four were characterized by high scores for this source of pressure. Regarding other land-use pressures, the management actions for water bodies C07 and C22 should target agricultural land uses, and those for water body C27 industrial land uses. Only water body C07 show a high score related to salinity, and thus to riverine pressure, reflecting inflows of the Muga river. Therefore, the management plan for this coastal area should include actions related to land uses ongoing throughout the watershed. A study of land uses in the Muga watershed revealed agriculture and urban land uses as the main human activities. Accordingly, the management plan of C07 should include actions related to urban and agricultural land uses, with their implementation not only in the vicinity of the coastal waters of this water body but also across the whole watershed.

General coastal management recommendations can be provided in addition and thus expand the scope of this study. First, actions should be implemented on land, where the nutrients that trigger coastal eutrophication are generated. Second, they should be implemented at a local scale, to respond to specific issues. Third, until integrated management plans are fully implemented, both precautions (Rice, 2003) and common sense are strongly recommended.

### FURTHER CONSIDERATIONS REGARDING LUSI

LUSI is a method to assess continental pressures, manifested as the risk of eutrophication, on coastal waters and it fulfills most of the criteria proposed by Lopez y Royo et al. (2009) for this type of tool: simplicity, time and cost-effectiveness, repeatability, reliability and broad applicability. Other strength of LUSI is that it is based on publicly available and periodically updated data. LUSI provides a synthetic pressure assessment that integrates all potential sources of impact to yield a single indicator value that is comparable among coastal areas.

LUSI is neither quantitative nor accurate, but it is a simple and reliable proxy of continental pressures, as shown herein. However, according to Lopez y Royo et al. (2009), the use of simple pressure assessment methods implies an oversimplification of the information. The question then arises whether continental nutrient fluxes (load and concentration) must be evaluated quantitatively and precisely to obtain an accurate characterization of continental pressures, as recommended by Fedorko et al. (2005), Giupponi and Vladimirova (2006), and Rodriguez et al. (2007). The approaches of those authors have two main drawbacks. First, although information on continental pressures from an identified point source, such as river water and wastewaters, is available and updated, for continental diffuse pressures, such as those caused by runoff, data are scarce or still lacking. Second, an accurate characterization of continental pressures implies considerable investments of time, economic and human resources, and, if models are to be constructed, computational effort; thus, an accurate characterization is likely to be feasible only at local scales and nearly impossible at regional or larger scales. Despite these drawbacks, continental pressures assessed using more accurate approaches have been used to establish pressureimpact relationships. The choice between a simple vs. an accurate method to assess continental pressures may depend on the objectives of the study, specifically, on the spatial scale, and the required degree of detail. However, the goodness of fit values of pressure-impact relationships obtained with more accurate assessments of continental pressure (Böhmer et al., 2014; Gerakaris et al., 2017) were similar to those established by LUSI. Thus, in establishing pressure-impact relationships, simplicity is not necessarily at odds with accuracy, as our results showed that LUSI is both a reliable and an appropriate approach.


Impact has been assessed based on chlorophyll-a concentrations measured from 2007 to 2012 and following the methodology established by Camp et al. (2016). Pressure has been assessed by LUSI and taking into account the 2009 Mapa de Cobertes del Sòl de Catalunya, for which all its parameters are shown.

LUSI offers a broad range of applications in the fields of environmental science, environmental law and socioeconomics. Direct applications include those related to environmental policies based on the drivers-pressuresstate-impact-responses (DPSIR) framework (Organisation for Economic Co-operation Development, 1993), such as described in the WFD and presented herein. Moreover, LUSI has indirect applications related to the benefits of access to pressure information, especially regarding diffuse sources of nutrients. This information could lead to better protection of the environment and therefore human health and wellbeing. For example, in Europe, LUSI already provides useful information for the Nitrate Directive (European Communities Council, 1991b), the Habitats Directive (European Communities Council, 1992), the Bathing Water Quality Directive (European Communities Council, 1976), the Urban Wastewater Treatment (European Communities Council, 1991a), and the New Political Framework for Tourism (European Economic Social Committee, 2010), in addition to supporting Integrated Coastal Zone Management objectives.

## CONCLUSION

LUSI is a method to assess the continental pressures on coastal waters. It serves as a proxy enabling the indirect assessment of continental nutrient loads and concentrations, and their influence on coastal waters. Therefore, it estimates the risk of eutrophication in coastal waters. It is based on systematic information on the land uses that influence coastal waters by providing nutrient-rich freshwater inflows (urban, industrial, agricultural, and riverine) and on coastline morphology, which can modify this influence, as it determines the degree of coastal water confinement and therefore the likelihood that these inflows will be diluted. LUSI was designed as a simple tool to assess coastal pressures when there is a lack of information, thus filling a methodological gap. Moreover, it also meets the requirements of the WFD regarding a true pressure assessment (i.e., not confounded with impact) and the establishment of pressure-impact relationships with impact indicators related to the Biological Quality Elements. The utility of LUSI was highlighted in a case study from the Catalan coast (NW Mediterranean), where continental pressures were assessed using LUSI, a pressure-impact relationship between LUSI and chlorophyll-a concentrations was established, and guidance on management actions regarding eutrophication provided. Furthermore, the well-performance of LUSI in assessing continental pressures and in establishing pressureimpact relationships was bibliographically demonstrated not only for the Biological Quality Element phytoplankton but also for angiosperms and macrophytes and its spatial scope broaden to the Black Sea and the Atlantic Ocean. Our results demonstrate the ability of LUSI to improve scientific knowledge regarding coastal eutrophication and to generate information that is useful to environmental managers in their efforts to restore and protect coastal waters against eutrophication and to achieve an integrated coastal management plan.

## AUTHOR CONTRIBUTIONS

All authors listed have made a substantial, direct and intellectual contribution to the work, and approved it for publication.

### FUNDING

This study used data collected within the National Catalan Water Monitoring Program funded by contract CTN0802809 between the Catalan Water Agency (ACA) and the Marine Science Institute (ICM-CSIC). This research is a contribution to the GRADIENTS project (Estructura fina de los gradientes costeros a lo largo de la costa Mediterranea), funded by the Spanish National Program (CTM2012-39476-C02-02) and the DEVOTES project (Development of innovative tools for understanding marine biodiversity, and assessing good environmental status), funded by the European Union's Seventh Framework Programme for research, technological development, and demonstration under grant agreement number 308392.

#### ACKNOWLEDGMENTS

R. Ventosa and M. Abad are acknowledged for dissolved inorganic nutrients analyses. J. Ballabrera for constructive

#### REFERENCES


discussions of the statistical approach and J. Jones, W. Ran, and two reviewers for helpful comments on earlier drafts of this work.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00018/full#supplementary-material

Baltic Lake Benthic Invertebrate Ecological Assessment Methods (Luxembourg: Publications Office of the European Union).


for ecosystem changes during past and future decades? Progress Oceanogr. 80, 199–217. doi: 10.1016/j.pocean.2009.02.001


based on phytoplankton in coastal waters. Mar. Pollut. Bull. 75, 218–223. doi: 10.1016/j.marpolbul.2013.07.028


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Flo, Garcés and Camp. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Managing Eutrophication in the Szczecin (Oder) Lagoon-Development, Present State and Future Perspectives

René Friedland<sup>1</sup> \*, Gerald Schernewski 1,2, Ulf Gräwe<sup>1</sup> , Inga Greipsland<sup>3</sup> , Dalila Palazzo1,4 and Marianna Pastuszak <sup>5</sup>

<sup>1</sup> Leibniz-Institute for Baltic Sea Research Warnemünde, Rostock, Germany, <sup>2</sup> Klaipeda University Marine Science and Technology Center, Klaipeda, Lithuania, ˙ <sup>3</sup> Norwegian Institute of Bioeconomy Research, Ås, Norway, <sup>4</sup> STA Engineering, Pinerolo, Italy, <sup>5</sup> National Marine Fisheries Research Institute, Gdynia, Poland

#### Edited by:

Marianne Holmer, University of Southern Denmark, Denmark

#### Reviewed by:

Angel Pérez-Ruzafa, University of Murcia, Spain Nafsika Papageorgiou, University of Crete, Greece

\*Correspondence: René Friedland rene.friedland@io-warnemuende.de

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 15 October 2018 Accepted: 21 December 2018 Published: 16 January 2019

#### Citation:

Friedland R, Schernewski G, Gräwe U, Greipsland I, Palazzo D and Pastuszak M (2019) Managing Eutrophication in the Szczecin (Oder) Lagoon-Development, Present State and Future Perspectives. Front. Mar. Sci. 5:521. doi: 10.3389/fmars.2018.00521 High riverine nutrient loads caused poor water quality, low water transparency and an unsatisfactory ecological status in the Szczecin (Oder) Lagoon, a trans-boundary water at the southern shore of the Baltic Sea. Total annual riverine N (P) loads into the lagoon raised at the 20th century from approximately 14,000 t TN (1,000 t TP) to 115,000 t TN (10,500 t TP) in the 1980ties and declined to about 56,750 t TN (2,800 t TP) after 2010. Nutrient concentrations, water transparency (Secchi depth) and chlorophyll-a showed a positive response to the reduced nutrient loads in the Polish eastern lagoon. This was not the case in the German western lagoon, where summer Secchi depth is 0.6 m and mean chlorophyll-a concentration is four times above the threshold for the Good Ecological Status. Measures to improve the water quality focused until now purely on nutrient load reductions, but the nutrient load targets and Maximal Allowable Inputs are contradicting between EU Water Framework Directive and EU Marine Strategy Framework Directive. According to the HELCOM Baltic Sea Action Plan, the thresholds of the annual riverine nutrient inputs to the lagoon would be about 48,850 t N (1,570 t P). Actions in the river basins that would allow meeting these targets are hardly achievable. Even if the load targets would be fully implemented, they are not sufficient to transfer the lagoon into a non-eutrophic state. The implementation of EU Water Framework Directive is further hampered, as consistent water quality thresholds for the two parts of Szczecin Lagoon are missing. An approach to harmonize them is presented, which incorporates the spatial differences. By implementing consistent water quality targets, Szczecin Lagoon could serve as blueprint for other trans-boundary waters. In the western lagoon, nutrient load reductions in the past decades had no effect on the water quality. High water residence times, frequent sediment resuspension and the missing submerged vegetation inhibit the load reduction effects on the water quality. Internal measures in the western lagoon are necessary, which aim at removing nutrients and increasing water transparency to overcome the hysteresis effect and to initiate a recovery of macrophytes. Cultivation of zebra mussels seems the most promising approach.

Keywords: eutrophication, water quality, nutrient loads, WFD, MSFD, integrated ecosystem modeling, internal measures

## 1. INTRODUCTION

With an area of 687 km<sup>2</sup> , Szczecin (Oder-) Lagoon is one of the largest lagoons in Europe. It is located at the border between Germany and Poland (**Figures 1**, **2**) and is divided into Large Lagoon (Polish: Wielki Zalew) and Small Lagoon (German: Kleines Haff). The lagoon is shallow (average depth of 3.8 m), has a salinity between 1 and 3 PSU and is connected with the Baltic Sea via three outlets. Large Lagoon, connecting the harbor of Szczecin with the open Baltic, plays an important role for shipping and transportation of goods, therefore a 12 m deep channel is crossing it (Radziejewska and Schernewski, 2008). With an average water discharge of about 500 m<sup>3</sup> /s and a drainage area of 120,000 km<sup>2</sup> (Pastuszak, 2012a), the Oder (Odra) River is one of the most important rivers in the Baltic region. It significantly controls water and nutrient budgets of the lagoon (Schernewski et al., 2012b). Ongoing high riverine nutrient loads are responsible for the heavily eutrophied status of the lagoon (Bangel et al., 2004; Schernewski et al., 2008, 2011), indicated among other indicators by low Secchi depths (**Figure 2**). Szczecin Lagoon is a part of several European conservation-orientated programs, like Natura2000 (Wolnomiejski and Witek, 2013). Because of its size, state of pollution, economic and ecological importance, Szczecin Lagoon has received a lot of scientific attention. It plays an important role as converter and sink for nutrients and pollutants effecting also water quality of the Baltic Sea (Radziejewska and Schernewski, 2008).

Szczecin Lagoon is managed within EU's Water Framework Directive (WFD; WFD, 60EC), which aims to establish a "Good Ecological Status" (GES) in all surface waters. GES is defined as a 50% deviation from "reference conditions" that describe a "high status with no, or very minor disturbance from human activities" (CIS-COAST, 2003). The WFD takes into account that coastal waters are controlled by nutrient loads from river basins, therefore it recommends comprehensive river basin management plans, linking coastal water objectives and measures in the catchments. In the Baltic Sea, the WFD is complemented by the HELCOM Baltic Sea Action Plan (BSAP), a comprehensive program to restore GES in the Baltic marine environment (HELCOM, 2007, 2013b). HELCOM BSAP can be seen as the regional implementation of the EU Marine Strategy Framework Directive (MSFD, 2008/56/EC). HELCOM BSAP has several operational ecological targets with respect to eutrophication, especially nutrient concentrations close to the natural level, satisfactory water transparency or natural blooms of algae. To achieve the targets, the Baltic countries agreed (beside other obligations) on Maximum Allowable Inputs of nutrients (MAI) and reduction targets in order to reach GES in the Baltic Sea (HELCOM, 2013a,c).

One problem in the Oder River and Szczecin Lagoon system is that coherent GES targets and MAI do not exist (Pastuszak et al., 2003, 2005, 2012a,b; Schernewski et al., 2011; Pastuszak and Witek, 2012a,b). For Small Lagoon, nutrient target threshold concentrations calculated by Brockmann et al. (2005), were far too ambitious to be reached even with an highflying river basin management (Schernewski et al., 2008, 2012b; Nausch et al., 2011). Furthermore, the nutrient targets did not match with chlorophyll-a, although closely linked. Therefore, Schernewski et al. (2015) carried out a full re-calculation of target concentrations for nitrogen, phosphorus and chlorophyll-a in all German coastal waters of the western Baltic Sea, which included a harmonization of water quality targets of waters managed by WFD and MSFD. Then, MAI and target concentrations in rivers for the German Baltic catchments were calculated, so that water quality targets were fulfilled on average. Although, this approach was a major step forward, it has several weaknesses in the case of Szczecin Lagoon, mainly: (a) thresholds are only provided for Small Lagoon (on WFD water type level), whereas for the Polish part harmonized thresholds need to be complemented. Most parameters in the lagoon show strong spatial gradients, therefore, spatially refined analysis and thresholds are required; (b) the German GES thresholds were derived from the state of 1880, by enhancing the estimated pre-industrial concentrations of chlorophyll-a, TN and TP by 50% (Schernewski et al., 2015). This approach allowed to derive consistent thresholds for all coastal and marine waters of the German Baltic Sea. But it stays an open questions, if the targets are reachable in a naturally eutrophic system; (c) water quality improvements via river basin nutrient load reductions usually take decades until they are fully visible (Grimvall et al., 2000; Oenema et al., 2005). Moreover, against the background of ongoing climate change, it is uncertain whether present water quality targets are suitable for future climate (Friedland et al., 2012; Meier et al., 2012a,b, 2014; BACC, 2015; Michalak, 2016), or if an adaptation will be necessary; (d) present targets focus on nutrients, but in naturally eutrophic systems, like Szczecin Lagoon, it is questionable if GES, with respect to biological parameters, can be achieved by a nutrient management alone. At least, more biologically related indicators and thresholds, like water transparency (Secchi depth) or occurrence of submerged macrophytes (Borja et al., 2012), should be considered; (e) there exists no spatial-explicit approach to link MAI with GES-targets on the scale of single water bodies, instead both are integrated over larger regions, e.g., the subbasins used by HELCOM, or the entire south-western Baltic Sea at Schernewski et al. (2015).

The objectives of our study are to: (i) give a comprehensive overview about the development of external nutrient loads into Szczecin Lagoon between 1880 and today; (ii) document resulting changes of major water quality parameters in the lagoon between 1880 and today; (iii) show spatial and interannual variability of parameters within the two parts of the lagoon; (iv) analyze the re-eutrophication process and hysteresis effects; (v) test the hypothesis that management actions in the river basins are sufficient to achieve the GES thresholds; and (vi) present the state of and discuss the consequences for water quality objectives and policy implementations.

## 2. METHODS

### 2.1. Nutrient Loads to Szczecin Lagoon

Szczecin Lagoon's nutrient budget is dominated by waterborne nutrient loads from Oder, Peene, Uecker, and Zarow, with Oder accounting for more than 90% of the total waterborne loads (Wielgat, 2002). Contribution of atmospheric deposition of nitrogen (N) and phosphorus (P) is marginal, as it constitutes 1% (N) and 0.2% (P), respectively to overall loads. Loads from point sources play a minor role, being strongly reduced in the

1990ties (Wielgat, 2002). Although lower, German waterborne nutrient loads impact especially Small Lagoon, where only approximately 20% of the Oder nutrient loads end up (Wielgat, 2002). Since late 1970ties, the German waterborne loads have been regularly monitored, nowadays being a part of the German implementation strategy of the WFD. Nutrient concentrations are usually measured twice a month (Bachor, 2005), likewise the freshwater runoff, which is partly also computed automatically on a daily basis. Monitoring of nutrients in Oder River is conducted in Krajnik Dolny (Kowalkowski et al., 2012; Pastuszak, 2012a,b) at least once a month (Jadczyszyn and Rutkowska, 2012). Riverine data collection started in 1988 and has been elaborated by Pastuszak et al. (2018). Reconstructions prior 1988 were used from Behrendt et al. (2008) and from Baltic Nest Institute (Wulff et al., 2013).

Environmental data, such as riverine nutrient concentrations and loads, often exhibit substantial natural variability caused by weather conditions at, or prior to the sampling occasion. The Partial Mann-Kendall test (Libiseller and Grimvall, 2002) was used to detect long-term changes in solute concentrations and flow-normalized loads. The test has its methodological basis in the seasonal Mann-Kendall-test (Hirsch and Slack, 1984), which has been used to detect temporal trends in environmental data, with the difference that water discharge is included as explanatory variable. The trend assessment (**Table 1**) was performed by comparing the calculated loads with the flow-normalized loads, based on a semi-parametric regression method (Stålnacke and Grimvall, 2001; Hussian et al., 2004). This normalization method aims at removing the natural fluctuations in loads and concentrations caused

TABLE 1 | Flow-normalized loads to Szczecin Lagoon for the reference period (1997–2003) and recent times (2010–2014), values in brackets indicate the loads without flow-normalization.


∗∗Statistically significant downward (p < 0.05)

<sup>∗</sup>Downward but not statistically significant (0.05<p<0.1)

P-values of the long-term trends were computed using the flow-normalized input data, asterisks indicating the degree of statistical significance (no asterisk = no significant change).

by variability in water discharge or other weather-dependent variables, so that the loads at "normal" water discharge are estimated. Such removal or reduction of irrelevant variation in the collected data can help to clarify the impact of human pressures on the environment (here, the impact on riverine loads of nutrients). It has also been shown that semiparametric normalization models were almost invariably better than ordinary regression models (Hussian et al., 2004).

Maximum Allowable Inputs (MAI) by HELCOM BSAP (HELCOM, 2013a,c) were downscaled for the single rivers from the national reduction schemes. Additionally, Poland and Germany have target concentrations defined within the national implementation strategies of EU's WFD (WFD, 60EC). The German WFD target concentrations are 2.6 mg TN/l and 0.1 mg TP/l (BLANO, 2014). Poland uses differentiated target concentrations (Garcia et al., 2012) for large lowland rivers, like Oder, (type 21; 4.0 mg TN/l, 0.29 mg TP/l) and small rivers directly feeding the coastal zone, like the German rivers entering Szczecin Lagoon (type 22; 2.7 mg TN/l, 0.32 mg TP/l).

To estimate the pre-industrial nutrient loads, the GIS (Geographical Information System) based catchment model MONERIS (MOdeling Nutrient Emissions in RIver Systems) was applied (Behrendt and Dannowski, 2005; Venohr et al., 2011). It was adjusted to present observations for Oder (Kowalkowski and Buszewski, 2006; Schernewski et al., 2008; Kowalkowski et al., 2012) and the German catchment area (BLANO, 2014; Ackermann et al., 2016). The simulated pre-industrial loads (referring to 1880) and a summary of underlying assumptions are published for German catchment (Hirt et al., 2014) and for Oder (Gadegast et al., 2012, 2014; Gadegast and Venohr, 2015). Thereby Gadegast and Venohr (2015) gave only the mean TN concentrations (1.6 mg TN/l for Oder; 1.0 mg TN/l for Peene and Uecker; 0.6 mg TN/l for Zarow), which were upscaled to annual loads using the recent runoffs.

## 2.2. Water Quality of Szczecin Lagoon

In Small Lagoon, eutrophication parameters are gathered monthly at three stations (KHJ, KHP, and KHM; **Figure 2**); additionally, observations are carried out near the bottom at station KHM. Until 2006, three more stations (KHO, KHK, and KHQ) were monitored. The same focusing at the most relevant stations took place in Large Lagoon, where data are collected from March to November at stations C, E, and H (**Figure 2**), while stations B, D and F were sampled only until 2006. First observations and macrophyte mappings in Szczecin Lagoon were carried out in the 1880s (Brandt, 1894/96). The regular Secchi depth monitoring started in the 1960ties (Baudler et al., 2012) and was afterwards supplemented by additional parameters covering nearly the last 50 years. A first data set was compiled by Bangel et al. (2004) and extended using public available data from LUNG (State Agency for Environment, Nature Conservation and Geology Mecklenburg-Vorpommern) and annual reports on the nature of German - Polish border waters published at WasserBLIcK<sup>1</sup> .

Threshold determining the "Good Environmental State" (GES) in Small Lagoon are set for summer Secchi depth (1.7 m; Sagert et al., 2008); summer chlorophyll-a (14.3 µg/l; BLANO, 2014; Schernewski et al., 2015) and nutrient concentrations (2.3 µmol/l TP; 38.1 µmol/l TN; BLANO, 2014; Schernewski et al., 2015). The latter were derived by using the Integrated Modeling Approach (Schernewski et al., 2015), which combined recent observations with model simulations to estimate the preindustrial state (1880) serving as reference state. GES threshold concentrations were derived by enhancing the reference concentration with 50% (CIS-COAST, 2003).

## 2.3. Modeling of Water Quality Parameters

Any 3-d model of Szczecin Lagoon must reflect the strong spatial gradients, but also the exchange with the open Baltic

<sup>1</sup>https://www.wasserblick.net/servlet/is/34786/

Sea (Pomeranian Bight), which takes place via three narrow outlets (Swina, Peenestrom, and Dziwna; **Figure 2**). So does the model system ERGOM-MOM (Schernewski et al., 2015), covering the entire Baltic Sea. It consists of the ocean model MOM (Pacanowski and Griffies, 2000) and the pelagic ecosystem model ERGOM that is suitable for applications in the Baltic Sea (Eilola et al., 2011; Friedland et al., 2012). It has a horizontal resolution of 1 nautical mile in the western Baltic Sea and the vertical water column is sub-divided into layers with a thickness of 2 m. In order to adjust the simulated water exchange between Szczecin Lagoon and the Pomeranian Bight on this coarse grid to realistic levels, the depth of Swina had to be reduced to 5 m.

The biogeochemical module ERGOM (Neumann, 2000; Neumann et al., 2002; Neumann and Schernewski, 2008) consists of three dissolved inorganic nutrients (nitrate, ammonium and phosphate), three functional phytoplankton groups (large and small cells, nitrogen fixers), fast-sinking dead organic material and a bulk zooplankton, which is grazing on the phytoplankton. Detritus is partly mineralized back into ammonium and phosphate, while the other portion accumulates at the sea bottom, where it is subsequently buried, mineralized or resuspended. All state variables are linked via advection-diffusion equations to the circulation model. Using stoichiometric ratios, the production and consumption of oxygen is calculated from all biogeochemical processes. Vice versa, the oxygen conditions determine, whether phosphate is bound to iron in the sediment (oxic situation) or is released (during anoxia). To evaluate the quality of the model outputs the method of Edman and Omstedt (2013) and Meier et al. (2018) was applied (**Figure 8**), by combining the coefficient of correlation R and a cost function C, which is the mean model bias normalized by the standard deviation of the observations.

To study the water exchange of Small Lagoon, a highresolution model system was set up, using the coastal ocean model GETM (Burchard and Bolding, 2002; Klingbeil and Burchard, 2013) instead of MOM. The refined model set-up covered Szczecin Lagoon and the surrounding waters with a horizontal resolution of 150 m (approximately 33,000 horizontal grid cells instead of 228 at ERGOM-MOM) and 20 vertical layers that adapt toward stratification (Hofmeister et al., 2010; Gräwe U, 2015). Following previous approaches (Gräwe et al., 2012; Schernewski et al., 2012a; Schippmann et al., 2013a,b), every day ten individual particles were released and followed until they either left Small Lagoon (to Peenestrom or Large Lagoon) or they reached the maximum age of 60 days. The particles represent inert substances driven exclusively by advection (Oliveira and Baptista, 1997). The horizontal diffusivity was taken into account using the random walk method, described by stochastic differential equations of motion (Heemink, 1990).

The model system ERGOM-MOM was further applied to two scenario simulations. To simulate the pre-industrial state (averaged for the period 1880 to 1885, after a spin-up of 25 years with a coarse model and additionally five years with ERGOM-MOM), the needed waterborne loads were taken for the German catchment and Oder river from MONERIS (Gadegast et al., 2012; Hirt et al., 2014). For the remaining catchments, the atmospheric deposition and weather forcing, reconstructions (Gustafsson et al., 2012; Ruoho-Airola et al., 2012; Schenk and Zorita, 2012) were used. For the future scenario, a transient simulation of the 21st century was conducted, assuming a moderate climate change by applying scenario RCP 4.5 (Pachauri et al., 2014) and that the MAI from BSAP (HELCOM, 2013b,c) are Baltic-wide fully implemented. The water temperature increase until the end of 21st century is with approximately 1.5 K quite low compared to most other studies (BACC, 2015). For the later calculations, the model results at the end of 21st century (averaged for 2080–2100) were used.

## 3. RESULTS

## 3.1. Nutrient Loads to Szczecin Lagoon–Development and Targets

Annual estimated riverine loads to Szczecin Lagoon during 1997 to 2003 (reference period of HELCOM BSAP) were approximately 70,400 t TN and 4,600 t TP (flow-normalized: 61,825 t TN and 4,222 t TP; **Figure 3**; **Table 1**). In recent years (average between 2010 and 2014), the loads dropped to approximately 66,400 t TN/a and 3,200 t TP/a (flow-normalized: 56,750 t TN/a and 2,790 t TP/a). The flow-normalized TN loads reached their maximum in the early 1990ties (approximately 85,000 t TN/a), while flow-normalized TP loads peaked at the end of 1980ties (approximately 8,500 t TP/a). The reconstructed loads were highest in 1977 and 1981 with up to 115,000 t TN/a and 10,500 t TP/a, when runoff was 870 m<sup>3</sup> /s (1977) and 841 m<sup>3</sup> /s (1981) substantially above the long term average (504 m<sup>3</sup> /s; **Table 1**). Using the catchment model MONERIS, the estimated pre-industrial TN loads differ substantially (between 14,100 and 26,400 t TN/a; **Table 2**). TP loads increased from approximately 1,000 t TP/a in 1880 (**Table 2**) already strongly until 1960, where they were on the same level as recently (**Figure 3**). The TN loads of 1960 were still near to the upper estimate for the historical levels, before they increased rapidly (**Figure 3**).

Nearly all rivers show a significant decrease of flownormalized TP loads, for TN loads this is only the case for Oder and Uecker (**Table 1**). The German rivers had further a decreasing trend of the freshwater runoff (**Table 1**). The loads showed the strongest decrease in the early 1990ties, while since 1995 the flow-normalized German TP loads stayed on the same level (approximately 100 t TP/a), except 2011. The flow-normalization has thereby only a minor influence on the German TP loads (**Figure 3**), while it smoothed the TN loads, revealing that there was nearly no decrease (**Figure 3**). The flownormalization resulted thereby in a smaller annual load for Oder in 15 of 27 years for TN (17 of 27 for TP) and for the German rivers in 20 of 37 years for TN (19 of 37 for TN).

Applying the flow-normalization led to lower nutrient loads from Oder for the reference period and recent years, while the German flow-normalized loads are higher in both period. But the method worked well to reduce the impact of floods on the loads (**Figure 3**). In 2010, strong rainfalls resulted in an extremely high outflow (up to 1,400 m<sup>3</sup> /s in June) and extraordinary high loads (91,000 t TN/a and 4,100 t TP/a), while this peaks are not visible in the flow-normalized loads. But also some exceptional years

FIGURE 3 | Annual total nitrogen (TN) and total phosphorus (TP) loads discharged by Oder (Upper) and summed for Peene, Uecker and Zarow (Lower) based on observations (blue) and flow-normalization (black); supplemented by reconstructed loads from Baltic Nest Institute (yellow; Wulff et al., 2013) and Behrendt et al., 2008 (orange, only for Oder).

TABLE 2 | Flow-normalized loads to Szczecin Lagoon for the reference period of BSAP (1997–2003) and load targets according to HELCOM Baltic Sea Action Plan (HELCOM, 2013b); loads calculated with adoption of the Polish WFD target (Garcia et al., 2012) and the German WFD-target (BLANO, 2014); estimated pre-industrial loads from catchment model MONERIS.


remained, e.g., in 2014 Oder runoff was extremely low (400 m<sup>3</sup> /s), resulting in an increase of flow-normalized TN loads compared to previous years. Vice versa, the flow-normalized TP loads had a recent peak in 2013, when runoff was very high (626 m<sup>3</sup> /s). In summer 2011 very strong rainfalls occurred, resulting in extremely high runoff combined with high TP concentrations due to the erosion in agricultural areas, especially in the Uecker catchment. As a consequence, monthly flow-normalized TP load was 13.7 t in August 2011, which is one order of magnitude above the long-term August average (1.5 t).

Downscaling the Maximal Allowable Inputs (MAI) from HELCOM BSAP, leads to reduction needs of 8.2% (21.1%) and 22% (64%) for German and Polish waterborne TN (TP) loads compared to the reference period. The MAI of Oder (on basis of the flow-normalized loads; **Table 2**) are then approximately 44,630 t TN/a (1,482 t TP/a) and for the German tributaries 4,230 t TN/a (80 t TP/a; **Figure 3**). After 2010, only Zarow fulfilled the TN target. None of the rivers was below the TP MAI of HELCOM BSAP. Adopting the different WFD threshold concentrations together with the average annual water discharges, MAI arising from the WFD assumptions are substantially differing from each other and also from the HELCOM MAI (**Table 2**). The German WFD target results in an allowable TN load of approximately 2,600 t TN/a for the German rivers, what means an additional TN reduction need of 30% compared to BSAP and 4% compared to the Polish WFD target. The German WFD TP target is substantially below than the Polish one, but less strict than BSAP and was almost equal to the reference loads. Assuming the German WFD target for the Oder, corresponds to an annual TN-MAI of approximately 41,400 t TN (**Table 2**). This would mean an additional reduction need of 7.3% compared to BSAP and would be a substantial tightening (35%) compared to the present Polish WFD target.

## 3.2. Water Quality in Szczecin Lagoon—Development and Targets

Secchi depth (SD), as well as near surface concentrations of chlorophyll-a (Chl-a), total nitrogen (TN) and total phosphorus (TP) are shown for Small Lagoon (represented by station KHM) and Large Lagoon (station C) in **Figure 4**. Throughout the study period, SD was in Large Lagoon always higher than in Small Lagoon, while Chl-a concentrations are substantially lower than in Small Lagoon. Until the middle of the 1990ties, Chla was in the same range for both parts. Then, it decreased by half in Large Lagoon (**Figure 4**). This decrease of Chl-a is significant for the whole time period, but also for the recent years (**Table 3**). It is strongly correlated with the reduced riverine TP loads (R <sup>2</sup> = 0.85; **Figure 5**), but only slightly with the TN loads (R <sup>2</sup>=0.14; **Figure 5**). In Small Lagoon, concentrations of Chl-a dropped in the 1980ties. Afterward, they remained on the same level (**Figure 4**) unaffected from the decreased nutrient loads (**Figure 6**). No significant trend was found, instead the p-value even increased for recent years (**Table 3**).

TN and TP concentrations in Small Lagoon decreased from the maximal concentration in mid-1980s of approximately 300 µmol TN/l (13 µmol TP/l) to 100–150 µmol TN/l (5– 7 µmol TP/l) mid of the 1990ties, resulting in a significant downward trend, which can not be found for recent years (**Table 3**). In Large Lagoon, TP concentration showed a continuous declining trend since its peak around 1990, resulting in significant downward trend (**Table 3**). This decreasing trend is strongly correlated to the riverine TP loads (R <sup>2</sup>=0.66). TN concentrations declined between 1985 and 1990, remained quite constant afterwards and started to increase after 2005 (**Figure 4**), unaffected of the decreasing TN loads, so that no downward trend was found (**Table 3**).

In Small Lagoon, the GES thresholds are most strongly violated in the case of Chl-a. The present state is five times the target value (**Figures 5**, **7**). But TN and TP are also

TABLE 3 | P-values of the long-term decreasing trend for water quality parameters observed at Station C (Large Lagoon) and KHM (Small Lagoon) computed for either the whole time period (see Figure 4) or since 1995.


∗∗Statistically significant (p<0.05)

<sup>∗</sup>Downward (upward for Secchi Depth) but not statistically significant (0,05<p<0,1)

Asterisks indicate the degree of statistical significance (no asterisk = no significant change).

more than twice the threshold concentrations. Transferring the target setting approach from Small Lagoon to the Polish part allows to derive harmonized targets, which include the spatial characteristics. Therefore, the present observed state is transferred by the ratio between simulated pre-industrial and present state to its reference state. Target concentrations are derived by enhancing the reference state with 50%. With 17.3 µg Chl-a/l; 42.5 µmol TN/l; 2.4 µmol TP/l the derived targets would be slightly above the ones for Small Lagoon (14.3 µg Chl-a/l; 38.1 µmol TN/l; 2.3 µmol TP/l). The deviation between present state and targets for TN and TP concentrations are quite the same as in Small Lagoon, while Chl-a in Large Lagoon would be twice the target (**Figure 7**). Extending the target setting approach also to SD<sup>2</sup> , would result in target values of 2.87 m (Large Lagoon) and 1.97 m (Small Lagoon). This target thresholds would be approximately three times the present state.

#### 3.3. Modeling of Water Quality Parameters

Using observations from all available stations in Szczecin Lagoon and from the Southern Pomeranian Bight (**Figure 2**), the model skill of ERGOM-MOM was calculated for surface Chla and dissolved nutrients (DIN, DIP), as well as for the near bottom values of salinity and oxygen (**Figure 8**). Following the classification of Omstedt et al. (2012), simulated surface DIN, bottom oxygen and salinity are good or acceptable for nearly all stations. Modeled surface Chl-a and DIP suffer from the low correlation coefficients, although the normalized biases are in the same range like for the other quantities. The single stations of the different regions cluster strongly. The strongest deviation at Chl-a in Small Lagoon occurred thereby in autumn and winter, so that the derived water quality targets for summer Chl-a are not impacted (**Figure 7**). But their too low concentration has a strong impact on the DIP concentrations, as phytoplankton is missing to bind DIP. On the other hand, the increase of DIP concentrations in late summer after the depletion in spring and early summer

2 Schernewski et al. (2015) did the target setting only for Chl-a, TN and TP.

FIGURE 5 | Observed summer Chl-a at station C (Large Lagoon) as function of the annual TN and TP inputs to the entire lagoon (blue cycles—annual values; orange diamonds—averaged over 5 years). Additionally are shown: (i) MAI of HELCOM BSAP and proposed GES-thresholds (black lines); (ii) pre-industrial simulated water quality and nutrient loads (black square); (iii) simulated water quality using MAI (green, including a moderate climate change).

was excluded.

starts approximately 1 month earlier in the observations then in the model simulation, worsening the correlation.

While the simulation covering the pre-industrial period was used to derive the water quality thresholds, the transient simulation of the 21st century was run to test the Baltic-wide implementation of MAI from HELCOM BSAP. In both parts of Szczecin Lagoon, none of the GES thresholds is met by the scenario simulation (**Figure 7**). Especially, the simulated Chl-a and TN concentrations are strongly above the thresholds, while TP targets are almost met. In both simulations, the simulated Chl-a concentrations are quite the same for both parts of Szczecin Lagoon, dissipating the spatial differences for the present state seen in the observations and the model.

To estimate the water exchange of Small Lagoon during summer month, in the GETM-setup particles were released daily. For two-monthly periods, the residence times were calculated then (**Figure 9**), resulting in a quite consistent spatial pattern. For nearly the whole Small Lagoon the particle age was equal to 60 days, what was the maximal time period, for which the particle were followed. This means, the particles did not left Small Lagoon, so that the water exchange is very limited.

### 4. DISCUSSION

Szczecin Lagoon is strongly utilized by human activities and a crucial supplier of ecosystem services (Inácio et al., 2018; Schernewski et al., 2018). Anthropogenic pressures, especially high nutrient loads, have led to a strongly eutrophied state, while at once central European policies, like the Water Framework Directive (WFD, WFD, 60EC), claim an improved water quality, up to reaching the "Good Ecological State" (GES). Due to its spatial heterogeneity, the inconsistent driving external policies and the opposite reaction to the changed pressures, Szczecin Lagoon can serve as case study for the ongoing re-eutrophication process in many other coastal waters.

Szczecin Lagoon is characterized by high concentrations of chlorophyll-a (Chl-a) and a low Secchi depth (SD; **Figure 1**),

resulting in a low water transparency. Numerous measures

threshold for Large Lagoon.

were undertaken to reduce the extensive nutrient supply (Kowalkowski et al., 2012; Pastuszak, 2012a,b; Pastuszak and Igras, 2012; Pastuszak and Witek, 2012a), resulting in load reductions of 30% (TN) and 70% (TP) compared to the peak values three decades ago (**Figure 3**). The two parts of Szczecin Lagoon (Small and Large Lagoon; **Figure 1**) reacted differently to the lowered loads with mostly significant improving trends at Large Lagoon and no trends at all in Small Lagoon (**Table 3**). Following Duarte et al. (2009) to combine water quality and nutrient load data, revealed that the declining supply of nutrients in the last two decades resulted only in Large Lagoon in a diminishing intensity of phytoplankton blooms (**Figure 5**) and an increase of SD (**Figure 4**). The water quality in Small Lagoon remained unaffected (**Figure 5**). In June 2012, SD reached even 2.55 m in the central part of Large Lagoon, which was the highest value since the continuous water quality monitoring started. This is already in the range of historical state (2–2.5 m; Brandt, 1894/96). Starting from 2012, meadows formed by Potamogeton perfoliatus and Myriophyllum spicatum have been regularly observed at the depth of 2–2.2 m in the eastern part of Large Lagoon (A. Wozniczka, NMFRI, pers. ´ comment). These appearances are practically identical to those reported for the end of the 19th century (Brandt, 1894/96). Together with the re-occurrence of Characeae (Brzeska et al., 2015), this is pointing to an improving ecological state due to the increased water transparency in Large Lagoon. This improvement was accompanied by nearly a halving of Chla concentrations (**Figure 4**), what is strongly correlated to the reduced TP loads (**Figure 5**). Hence, further load reductions are the most important measure to improve the water quality enduringly in the eastern part of Szczecin Lagoon.

Contradictory, the western part of Szczecin Lagoon (Small Lagoon) stayed unaffected from the decreased loads. At the end of the 1960s, when the first regular monitoring started, SD was higher (up to 0.9 m) than today (around 0.6 m), but already strongly below the historical values and also the GES target (1.7 m; Sagert et al., 2008). The historical range (2–2.5 m; Brandt, 1894/96) was not differentiated between Small and Large Lagoon, but it seems likely that SD in Small Lagoon was always less than in Large Lagoon due to the hydrographic background conditions. This difference is reflected by our suggested SD targets (1.97 m for Small Lagoon and 2.87 m for Large Lagoon). Central problem of Small Lagoon hampering any water quality improvement is thereby the low water exchange and high water residence times (**Figure 9**). This leads to the long-lasting internal cycling of nutrients, enhanced by the permanent resuspension of organic-rich sediments (Radziejewska and Schernewski, 2008). The low light availability hampers the growth of submerged macrophytes and resulted in a regime shift from a macrophyte dominated system to a phytoplankton dominated one (Duarte et al., 2009). This is further enhanced by the hysteresis effect (Scheffer, 2009), as macrophytes are missing to stabilize the sediments (Karstens et al., 2015) and to reduce the resuspension, resulting in a further lowered water transparency. Berthold et al. (2018) combined nutrient concentrations and phytoplankton biomass in several coastal waters of the southern Baltic Sea and emphasized the importance of hysteresis and resilience factors when reversing eutrophication, concluding that the pure reduction of nutrient loads will not be sufficient to reach the GES thresholds. Duarte (2009) showed for coastal waters, which received high external loads in the 1970ties and 1980ties that returning to oligotrophic conditions is extremely difficult, long time taking, and might in practice be impossible. The recovery of a degraded ecosystem is complex (Duarte et al., 2015), may last decades to centuries (McCrackin et al., 2017) and is often only possible, when external pressures are reversed and additional restoration efforts undertaken (Duarte et al., 2015). Carstensen et al. (2011) concluded from the development between 1970 and 2010 of coastal waters in Europe and USA that the reduction of nutrient loads had no striking impact on surface Chl-a. Riemann et al. (2016) extended the analysis to all coastal waters in Denmark, which experienced since 1990 a bisection of nutrients loads, while SD, eelgrass coverage and benthic gross primary production (Krause-Jensen et al., 2012) increased only slightly. In Small Lagoon, the same occurred, as a lessening of nutrient loads has not led to lower values of eutrophication parameters, so

FIGURE 9 | Water residence times in Small Lagoon in April, June and August 2008 (Left) and their standard deviations (Right) calculated using ERGOM-GETM. Small Lagoon is always characterized by residence times near to the maximal value (the particle were only followed for 60 days).

that reaching the desired GES is not possible only by reducing the nutrient loads. Further, Petkuviene et al. (2016) and Zilius et al. (2014) reported for muddy sediments in Curonian Lagoon (which are also dominating in Small Lagoon) a release of reactive P, boosting phytoplankton blooms. Boström and Pettersson (1982) and Jensen and Andersen (1992) observed a release of P in shallow lakes also under oxic conditions, concluding that not only the pure redox conditions at the sediment-water interface are crucial for the P-release. Søndergaard et al. (1996) showed the strong influence of resuspension on pelagic nutrient concentrations in shallow lakes and explained this by a high ratio of sediment surface to water column. This holds also for Small Lagoon, where average water depth is below 4 m. To improve water quality in Small Lagoon in a suitable manner, it is therefore not only important to further reduce the external nutrient supply, but to tackle the resuspension and the hysteresis effect. Submerged vegetation is missing to stabilize the sediments, to reduce the resuspension of particles and to bind excess nutrients (especially P).

Hence, an extension of measures (in addition to limiting the nutrient loads) within Small Lagoon seems necessary or authorities have to acknowledge that GES cannot be reached. Possible internal measures cannot replace nutrient input reductions, but they are adequate supplements if they affect the central problems, which are the missing submerged vegetation and the frequent resuspension of sediments. Berthold et al. (2018) discussed therefore bio-manipulation approaches, like the planting of submerged macrophytes or the transition of the food web from planktivorous to piscivorous fish. More promising is to increase the biomass of filter feeders by supplying them appropriate material to grow on (Stybel et al., 2009) as suitable hard substrate is missing in Small Lagoon. Filter feeders reduce the phytoplankton density and improve thereby the underwater light climate. Petersen et al. (2008) showed that in an eutrophied coastal water in Denmark (Ringkobing Fjord) a sudden regime shift towards a clear water state could be initialized by supporting the growth of already occurring suspension feeders. This led to an increasing coverage of benthic vegetation and an ongoing improving trend of water transparency, even when the suspension feeder biomass did not further increase. To ensure at once the active removal of nutrients, mussel farm approaches can be adapted, allowing to harvest the mussels, when they are big enough (Schernewski et al., 2018). Schernewski et al. (2012b) estimated that in Small Lagoon every year 1,000 t N and 70 t P could be taken out by mussel farms utilizing Dreissena polymorpha. SD could be increased by up to 30 cm, so that light availability and overall growing conditions of submerged macrophytes would improve. The efficiency of mussel farms to locally improve SD and water quality was carried out along Baltic coastal waters in Sweden (Lindahl et al., 2005), Denmark (Nielsen et al., 2016) or Germany (Schröder et al., 2014). With a spatial explicit mussel farm simulation model, Friedland et al. (2018) showed that SD could locally be increased even up to 60 cm, but at an increasing risk of anoxia, resulting in a further P-liberation from the sediment. Nevertheless, even in the maximal scenario of Friedland et al. (2018), the potential water quality improvement was not enough to reach the GES thresholds of Chl-a. Instead of a maximal mussel farm approach with unpredictable implications, most promising is to establish spatially and temporally restricted farms to support the (re-)occurrence of submerged macrophytes at wisely chosen spots (Schernewski et al., 2018)—presuming that they will grow, if the underwater light conditions get good enough due to the clearance effect of the mussels. After the initial recovery of the submerged vegetation, the mussel farm could be dismantled and moved to a new spot to prevent a disturbance of the newly grown vegetation.

Yet, water quality management measures focused exclusively on reducing nutrient loads, which are the most important anthropogenic driver resulting in the present eutrophied state. But the management is hampered by non-harmonized nutrient load targets from the driving legislations (**Table 2**). The allowable nutrient inputs from the two central eutrophication related European policies WFD and MSFD (Marine Strategy Framework Directive; MSFD, 2008/56/EC) are not synchronized, reflected by not harmonized load thresholds (**Table 2**). Central program to reach the MSFD targets in the Baltic Sea is the HELCOM Baltic Sea Action Plan (BSAP; HELCOM, 2013b), which includes binding nutrient input thresholds (Maximal Allowable Inputs, MAI) for all Baltic states. Although aiming on the right issues mainly to fight the eutrophication in the Baltic Sea—HELCOM BSAP (HELCOM, 2013b) is not suitable for a coastal water like Szczecin Lagoon (Håkanson and Bryhn, 2008; Håkanson et al., 2010). Szczecin Lagoon is not an explicit part of the ecosystem model NEST (Savchuk et al., 2012), which was used to determine the MAI. Instead, it is combined with the entire Pomeranian Bight and parts of the Bornholm Basin to one subbasin. This means that the nutrient retention within Szczecin Lagoon is not included in NEST, although at least 20% of Oder's nutrient loads are retained in Szczecin Lagoon according to our simulations. (Pastuszak et al., 2005) reported retention rates even up to 45% for N and 37% for P loads. The reduction demands by the updated version of BSAP (HELCOM, 2013b) resulted in unreachable low target, which is nearly the same as in the preindustrial situation (**Table 2**) and 16.3% more ambitious than the German WFD target (0.1 mg TP/l; **Table 2**; BLANO, 2014). Håkanson and Bryhn (2008) and Håkanson et al. (2010) grouped eutrophication indicators according to the trophic state of the different parts of the Baltic Sea and concluded that only Gulf of Finland, Gulf of Riga, Kaliningrad proximity, Vistula, and Szczecin Lagoon are eutrophic and need attention. Therefore, all actions related to nutrient load reduction should be focused on these regions but not on the entire Baltic Sea. Håkanson et al. (2010) emphasized further that the proposed level of load reductions may need updating due to further reorganization of the Baltic ecosystem as a result of regime shifts (Möllmann et al., 2005, 2009; Tomczak et al., 2016; Zettler et al., 2017) and due to the impact of climate change on water quality (Friedland et al., 2012; Meier et al., 2012a,b, 2014). For the central Baltic Sea, not the nutrient load to Szczecin Lagoon is crucial, but the nutrient export from it to the open Baltic. Hence, strengthening the nutrient removal within Szczecin Lagoon, e.g., by establishing mussel farms also in the eastern part, could therefore be a support to reach the targets of the BSAP.

If mussel farms are proven to be a suitable supporting measure to enable a sustainable water quality improvement, transferring them to other coastal systems seems possible. Applying the lagoon categorization of Kjerfve (1986), Kjerfve and Magill (1989), and Umgiesser et al. (2014) to Szczecin Lagoon as whole water body is not possible due to its high spatial heterogeneity. While Large Lagoon is a restricted lagoon, Small Lagoon is choked, comparable to lagoons like Mar Menor (Mediterranean Sea), where low freshwater runoff and a limited exchange with the open sea take place (Umgiesser et al., 2014). But, Mar Menor reacted opposite to the increased nutrient loads over the past decades with high water transparency (up to 5 m Secchi Depth; Pérez-Ruzafa et al., 2008, 2018) and only a low phytoplankton density, resulting in a classification as oligotrophic water (Marín et al., 2015). On the other hand, for Curonian Lagoon (Baltic Sea) a decline of submerged macrophytes due to eutrophication is also reported (Sinkevicien ˇ e et al., ˙ 2017). The zonation of Curonian is thereby quite comparable to Szczecin Lagoon with a slow water renewal in the southern part (Umgiesser et al., 2016), while the northern part is flushed out as fast as Large Lagoon (76.5 days; Umgiesser et al., 2016). For Curonian Lagoon the potential of mussel farms is already discussed (Bagdanavici ˇ ut ¯ e˙ et al., 2018). Other coastal waters like Darss-Zingst-Bodden-Chain (Schubert et al., 2010) or Rügensche Binnenbodden (both German Baltic Sea; Selig et al., 2007) experienced a comparable strong reduction of submerged vegetation and water transparency, and are not reacting to decreased nutrients loads (Berthold et al., 2018).

Friedland et al. Managing Eutrophication in Szczecin Lagoon

Many other trans-boundary waters (Newton et al., 2014), e.g., Vistula Lagoon (Chubarenko et al., 2014), Curonian Lagoon (Povilanskas et al., 2014) or Flensborg Fjord, are lacking like Szczecin Lagoon from a consistent and joint water quality target setting across political borders. Applying the Integrated Modeling approach of Schernewski et al. (2015) to both parts of Szczecin Lagoon allowed to derive harmonized target values (**Figure 7**), accounting the differences in the present state of both parts, but also the differing reactions to nutrient load changes. It can therefore be a blueprint for other transboundary waters. The method of Schernewski et al. (2015) to define GES targets was an important step forward, as it allowed to derive consistent and harmonized targets for all German coastal waters of the Baltic Sea. Although it included the water body specific present states as well as simulated reactions to changed nutrient loads, the estimated pre-industrial state (referring to 1880) might be unrealistic low for some waters after a century of excess nutrient loads, e.g., the simulated Chl-a concentration in Small Lagoon for 1880 is approximately 14% of the recent one. Beside the uncertainty of the model simulations, this deviation raises the question if 1880 is suitable as state referring to a very good water quality or if water quality targets in Szczecin Lagoon should base on another reference state, e.g., before 1960. On the other hand, the pre-industrial nutrient loads can only be derived from models. Although using both times the catchment model MONERIS, estimated pre-industrial TN loads by Gadegast and Venohr (2015) were double of Gadegast et al. (2012) (**Table 2**). Following the model results for a very optimistic future scenario (**Figures 5**, **6**), where all nutrient load reductions were in charge, GES was still not achieved. The mismatch between water quality targets and nutrient load thresholds stresses out the importance of complementing water quality targets of coastal waters with reduction demands in the catchment area, as introduced by Arheimer et al. (2015) or within the updated Danish Water Action Plan (Maar et al., 2016). Although differing in details, these approaches have in common that they are tailor-made and include the specific characteristics of the water bodies and catchment areas.

The used ecosystem model (ERGOM-MOM) was well able to reproduce the mean eutrophication parameters (**Figure 7**), although the model skill (**Figure 8**) revealed some weaknesses. Future improvements may result in more reliable estimates of the pre-industrial state and derived GES thresholds. ERGOM was in its initial phase (Neumann, 2000; Neumann et al., 2002) adjusted to the open Baltic Sea, dominated by phytoplankton communities not being exposed to a steady supply of nutrients, like in Szczecin Lagoon. Further, the interaction with higher trophic levels and macrophytes is missing. 45 trophic-functional components are identified of Large Lagoon's food web (Wolnomiejski and Witek, 2013), while ERGOM only includes three phytoplankton groups and one bulk zooplankton, in which not only all zooplankton species but also all their individual life stages are merged. Nevertheless, with respect to the nutrient dynamics, ERGOM is mostly well adjusted. With denitrification, the most important N retention process is included, which contributes up to 70% of the N retention within Szczecin Lagoon (Pastuszak et al., 2005). In our simulations, denitrification results in an average loss of 4.7 mmol N/m<sup>2</sup> /d or 16,500 t N per year (integrated over the whole lagoon; Allin et al., 2017). This fits with reported values, e.g., 2-4 mmol N/m<sup>2</sup> /d (Radziejewska and Schernewski, 2008) or 15,800 t N/a (based on model data; Wielgat and Witek, 2004), while (Dahlke et al., 1998) reported an annual loss of only approximately 500 t N for Small Lagoon (based on observations). On the other hand, the liberation of P under anoxic conditions is involved, but might not be enough, if P is released from the sediments, although in the bottom waters oxygen is still available. Here the description of the early diagenesis with one static layer that can be either oxic or anoxic is not sufficient, but the qualitative and quantitative description will be improved in the future (Radtke et al., 2018).

## 5. CONCLUSION


## AUTHOR CONTRIBUTIONS

RF was responsible for overall structure of the manuscript, compiling the data, running the simulations and for writing the manuscript. MP collected nutrient loads of Oder and contributed to the writing. GS contributed to the writing. UG and DP were responsible for the simulation of the water exchange. IG conducted the flow-normalization of loads.

### FUNDING

The work was supported by BONUS BALTCOAST (03F0717A) and partly by KÜNO project MOSSCO-Synthese (03F0740B). BONUS BALTCOAST project has received funding from BONUS (Art. 185), funded jointly by the EU and Baltic Sea national funding institutions.

### REFERENCES


#### ACKNOWLEDGMENTS

We thank A. Hiller (IOW) for supporting the map compiling (**Figure 2**) and P. Stålnacke (nibio) for supporting the flownormalization. Supercomputing power was provided by HLRN (North-German Supercomputing Alliance). We thank LUNG and WIOS Szczecin for providing the observational data.


NLWKN. 39 Seiten, Available online at: http://www.nlwkn.niedersachsen.de/ download/98787


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Friedland, Schernewski, Gräwe, Greipsland, Palazzo and Pastuszak. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Optimizing Monitoring Programs: A Case Study Based on the OSPAR Eutrophication Assessment for UK Waters

Luz María García-García<sup>1</sup> \*, Dave Sivyer <sup>1</sup> , Michelle Devlin<sup>1</sup> , Suzanne Painting<sup>1</sup> , Kate Collingridge<sup>1</sup> and Johan van der Molen<sup>2</sup>

<sup>1</sup> Lowestoft Laboratory, Centre for Environment, Fisheries and Aquaculture Science, Lowestoft, United Kingdom, <sup>2</sup> Department of Coastal Systems, NIOZ Royal Netherlands Institute for Sea Research, Utrecht University, Den Burg, Netherlands

The data and results of the UK second application of the OSPAR Common Procedure (COMP) for eutrophication were used as a case study to develop a generic system (i) to evaluate an observational network from a multi-variable point of view, (ii) to introduce additional datasets in the assessment, and (iii) to propose an optimized monitoring program to help reduce monitoring costs. The method consisted of tools to analyse, by means of simple statistical techniques, if any reduction of the available datasets could provide results comparable with the published assessments, and support a reduced monitoring program (and limited loss in confidence). The data reduction scenarios included the removal of an existing dataset or the inclusion of freely available third-party data (FerryBox, satellite observations) with existing datasets. Merging different datasets was problematic due to the heterogeneity of the techniques, sensors and scales, and a cross validation was carried out to assess possible biases between the different datasets. The results showed that there was little margin to remove any of the available datasets and that the use of extensive datasets, such as satellite data, has an important effect, often leading to a change in assessment results with respect to the thresholds, generally moving from threshold exceedance to non-exceedance. This suggested that the results of the original assessment might be biased toward sampling location and time and emphasized the importance of monitoring programmes providing better coverage over large spatial and temporal scales, and the opportunity to improve assessments by combining observations, satellite data, and model results.

Keywords: nutrients, chlorophyll, eutrophication, assessment, OSPAR, optimization, monitoring

## INTRODUCTION

Marine monitoring is an essential element of reporting and assessment of the marine environment and provides insight into coastal and ocean processes, as well as scientific support for management. Sustained, reliable and good quality in situ observations are needed for model and satellite calibration, validation, forecasting, environmental and ecological assessments, but they can come with significant economic costs. Optimization of the monitoring systems and improvement of their cost-effectiveness have become a priority and a subject of international concern in the recent years,

#### Edited by:

Jesper H. Andersen, NIVA Denmark Water Research, Denmark

#### Reviewed by:

Philip George Axe, Swedish Agency for Marine and Water Management, Sweden Lech Kotwicki, Institute of Oceanology (PAN), Poland

> \*Correspondence: Luz María García-García luz.garcia@cefas.co.uk

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 October 2018 Accepted: 14 December 2018 Published: 14 January 2019

#### Citation:

García-García LM, Sivyer D, Devlin M, Painting S, Collingridge K and van der Molen J (2019) Optimizing Monitoring Programs: A Case Study Based on the OSPAR Eutrophication Assessment for UK Waters. Front. Mar. Sci. 5:503. doi: 10.3389/fmars.2018.00503

**138**

as demonstrated by the numerous projects that have dealt with this topic. Some examples only in Europe are the projects ODON (Optimal Design of Observational Networks, 2003–2006), OPEC (Operational Ecology, 2012–2014), JERICO (Toward a Joint European Research Infrastructure network for Coastal Observation, 2011–2015), and its continuation JERICO-NEXT (2015–2018) and JMP-EUNOSAT (Joint Monitoring programme of the Eutrophication of the North Sea with Satellite data, 2017–2019). An interesting summary on the assessment and optimal design of ocean observing networks in Europe together with the short, mid-term and long-term objectives can be found in She et al. (2016). The design of an observational network requires an existing knowledge of the system, which generally rests on the existence of a good dataset (Fu et al., 2011) which provides an optimal number of observations over space and time to answer a specific purpose. In this sense, a first step for the design of an optimal monitoring programme is the assessment of the existing ones, for which two different methods are frequently applied: statistical and dynamic methods (see She et al., 2006; Fu et al., 2011). The statistical methods are generally based on a multi-indicator approach (She et al., 2006): they include the system error, sampling error (North and Nakamoto, 1989; She and Nakamoto, 1996), noise-to-signal ratio (Meyers et al., 1991; Smith and Meyers, 1996; Guinehut et al., 2002), effective coverage (She et al., 2007), explained variance (Fu et al., 2011) or the field reconstruction error, which is mainly based on the optimal interpolation method (see, for instance, She and Nakamoto, 1996).

The dynamic methods use models and data assimilation techniques in the design of optimal observing systems. The most commonly used tools are Observing System Experiments (OSEs), in which the actual observations are assimilated to produce nowcasts and forecasts, and the Observing System Simulation Experiments (OSSEs), that use models to simulate future observing systems before being deployed, and analyze the impact on the forecasts by assimilating or not these virtual observations (see, for instance, Oke and O'Kane, 2011).

In this paper we focus on the optimization of the monitoring systems for the eutrophication assessments in the United Kingdom. Eutrophication is defined as: "the enrichment of water by nutrients causing an accelerated growth of algae and higher forms of plant life to produce an undesirable disturbance to the balance of organisms present in the water and to the quality of the water concerned" (from the Urban Waste Water Treatment Directive (UWWTD [(EC, 1991a)]; Borja et al., 2010; Foden et al., 2011). Diverse approaches to monitoring are employed depending on the regulatory requirements: the Nitrates Directive (EC, 1991b), the Water Framework Directive (WFD, EU, 2000), the Oslo Paris Convention (OSPAR) or the Marine Strategy Framework Directive (MSFD, EU, 2008), which is the key policy driver for future assessments of eutrophication status (see Borja et al., 2010; Painting et al., in preparation). Assessments are generally based on the same key indicators (e.g., nutrients, chlorophyll, dissolved oxygen) and principles, but may differ in terms of the assessment area. For instance, in the UK the WFD is applied to estuaries (typically with a salinity <30) and coastal water bodies within 1 to 3 nm of the coastal baseline and MSFD and OSPAR assessments focus on coastal and offshore waters that extend beyond WFD areas and generally have salinities >30. Assessments may also differ in terms of time periods assessed or the focus on trends or impacts of nutrient enrichment in terms of exceeding assessment levels. A comparative review of the details of the different regulations applied in UK waters, the procedures used for evaluating the eutrophic status of the different water bodies and the employed thresholds on each of the assessments can be found in **Tables 3**–**5**, respectively, of Devlin et al. (2011; see also Borja et al., 2010; UK National Report, 2017; Painting et al., in preparation).

We used data and results from the second application of the OSPAR Common Procedure for the assessment of eutrophication (OSPAR COMP2, hereafter) in the coastal and offshore waters of the UK portion of the southern North Sea. The OSPAR COMP2 covered the years 2001–2005 and was selected as a case study as the final data set and assessment results have been published by OSPAR (OSPAR, 2008) and Foden et al. (2011). The overall aim was to evaluate the monitoring system employed for this assessment and how it could be improved.

A heuristic approach to the evaluation of the monitoring system has been considered in this case, mainly consisting of analyzing scenarios of different dataset aggregations (including or excluding certain datasets, etc.) and their impact on the OSPAR COMP2 results.

The specific aim was to evaluate whether we could obtain similar results (keeping the quality in terms of confidence and representativeness) to the COMP2 assessment by using less data (or, in other words, by identifying and removing redundant data) or third party data (such as FerryBox or satellite chlorophyll), thus reducing the costs of monitoring for future assessments.

The methods proposed here do not fall into the category of the dynamic methods above, since we did not use models or data assimilation techniques. Nor do our methods fall strictly into the category of statistical methods, although the employed techniques provide some indirect information on the effective coverage. Indeed, analyzing whether the current monitoring systems cover the relevant spatio-temporal scales, and considering how to avoid issues related to autocorrelation and combining datasets in order to provide a robust and unbiased assessment for eutrophication assessment is beyond the scope of this paper and will be addressed in Collingridge et al. (in preparation).

This paper addresses the following questions applied to eutrophication assessments:


#### MATERIALS AND METHODS

#### Indicators

The primary indicators of eutrophication status analyzed in the OSPAR COMP2 are the concentration of nutrients, chlorophyll, and dissolved oxygen. Further details on the OSPAR criteria used

to assess eutrophication can be found in Foden et al. (2011). In this paper, we focus on two of these indicators: winter nutrient concentrations [dissolved inorganic nitrogen (DIN), which is the sum of nitrate, nitrite, and ammonium] and growing season chlorophyll concentrations. Dissolved oxygen was excluded from the analysis because it constituted a much smaller dataset.

### Assessment Areas

Thirteen marine areas were assessed for eutrophication under OSPAR in England and Wales (see **Figure 1**), although the full OSPAR COMP2 was only applied to eight areas (marked with a red dot in **Figure 1**). The remaining five areas were not assessed further after the application of a screening procedure that identified them as non-problem areas.

For this study, we selected one of the assessed areas—East Anglia—where an extensive dataset was available. Only the data within this geographical region were analyzed here.

## Datasets for the OSPAR COMP2 Assessment

The description of all the datasets used for the OSPAR COMP2 in East Anglia, together with the assessment itself, is given in OSPAR (2008). The available datasets for DIN and chlorophyll were taken from across the UK estuarine, coastal, and offshore monitoring programs.

#### Datasets for Winter DIN

Three different datasets were available for winter DIN:


**Figure 2A** shows the spatial distribution of all data used for the DIN assessment in East Anglia. Estuarine waters (salinity <30) and inshore coastal waters are mainly covered by the wfd dataset, whereas sap and sbu mainly cover the coastal (salinities ≥30 and <34.5) and offshore water bodies (salinity ≥34.5). **Figure 2B** represents the distribution in time (along the assessment years) of the available datasets. The distribution of data was not homogeneous over time, with 2002 and 2004 better represented than the rest of the years, and 2005 being especially poor in terms of data availability. Focusing on the different datasets, the number of observations from wfd was highest for 2001, 2002, and 2004, but it is very variable along time and very scarce in 2005. The sbu dataset is quite homogeneous in time, and it is the most important dataset in 2003 and 2005. The sap dataset is the smallest of the three.

#### Datasets for Growing Season Chlorophyll

WFD and Warp SmartBuoy data were available for assessments of growing season chlorophyll. The spatial coverage of these datasets did not extend to the full assessment area, with data mainly confined to the coast (**Figure 2C**). The temporal distribution of the data was not homogeneously distributed between the assessment years (**Figure 2D**): the number of observations was highest in 2004 and lowest in 2005. The wfd dataset showed an increase in the number of observations between 2001 and 2004, and dropped to zero in 2005. The sbu dataset was quite homogeneous along time.

### Additional Datasets

For the assessment of chlorophyll, two additional datasets were explored:

• The Cuxhaven-Harwich FerryBox chlorophyll data <sup>1</sup> . This ferry line was operative in the period 2002–2005. It was the first Ship of Opportunity in which a FerryBox system was installed (see Petersen et al., 2011). Continuous data (temperature, salinity, chlorophyll, oxygen saturation, and pH, nutrients, etc.) were recorded en-route at a temporal resolution of ∼20 s, which corresponds to a data point every 550 m, on average. In order to reduce the number of observations and the spatial and

<sup>1</sup>http://www.ferrybox.com/

temporal autocorrelation, the data were averaged considering a time interval of 10 min, which reduced the spatial resolution to a data point every 3.6 km on average, with a total number of 3,278 observations in the assessment period. Sensitivity analysis were carried out to investigate the effect of the averaging period (1 min, 10 min, 30 min, and 60 min) on the assessment results (only for the aggregation scenario wfd+fbx described below). We will use the notation fbx to refer to this dataset.

• The MODIS daily chlorophyll satellite images. Daily composites of chlorophyll data captured by the MODIS-aqua sensor and processed at IFREMER with the OC5 algorithm (Gohin et al., 2002) were available at a horizontal resolution of 1.1 km during the period 2002–2005. In order to assign a value of salinity to each satellite chlorophyll observation, the results of a model that covered the assessment area were considered. In this way, the modeled salinities were interpolated at the positions of the satellite observations on a daily basis. Details of the model setup and its validation can be found in van Leeuwen et al. (2015) or Ford et al. (2017). The satellite chlorophyll data are called sch throughout this paper, and comprise a total number of 419151 observations for the assessment period, which is more than two orders of magnitude higher than the number of observation used for the OSPAR COMP2 assessment of chlorophyll (see **Figure 2D**).

## Statistical Techniques for the OSPAR COMP2 Assessment

#### Statistical Techniques for the Assessment of DIN and Chlorophyll

**Table 1** summarizes the assessment statistics together with the thresholds applied to the two indicators considered in this paper. The mean winter DIN was normalized to salinity to account for the gradients caused by river inputs that affect the nutrient concentrations. In the OSPAR COMP2, mixing diagrams were used to assess the winter concentrations of DIN (with winter defined as the months of January, February, November, and December of the same year) during the period 2001–2005 (see Foden et al., 2011). The mixing diagrams are used to plot concentrations of DIN against salinity each year, and to calculate linear regressions (**Figure 3**). All ranges of salinities (from estuarine to offshore) are considered to construct the mixing diagrams. From the linear regression equations, a normalized mean nutrient concentration is calculated for reference salinities for coastal and offshore waters of 32 and 34.5, respectively. Finally, the normalized mean values are compared against the defined salinity-normalized nutrient thresholds.

Chlorophyll is assessed using the 90th percentile value during the growing season owing to the distribution of chlorophyll data. The 90th percentile value of the data accounts for the variability and skewness of data associated with episodic high bloom periods and sampling frequency (see Devlin et al., 2007; OSPAR, 2008; Foden et al., 2011).

These statistics were only applied when more than 5 observations were available in any time period being assessed (per year, or for the whole assessment period), as was done in Foden et al. (2011) and OSPAR (2008).

#### Estimating Confidence

Three measures of confidence were used/applied in this paper:


All three measures combined give information on the confidence in the assessment. Most of the methods that will be described hereafter are different from those employed in the OSPAR COMP2, and are mainly based on the guidelines published in Annex 8 of OSPAR (2013). It is not the purpose of this paper to compare the confidence obtained using these methods with those published in the OSPAR COMP2, but to have a measure that allows for a comparison of the effect on a modern assessment of the different optimization scenarios that were applied to the data (see section Aggregation scenarios).

#### Confidence in the Representativeness of the Data

The confidence in the representativeness of the data was analyzed in terms of temporal representativeness, spatial representativeness and number of data points.

The representativeness of the available data in time over the assessment period (2001–2005) was calculated taking into account the methodology described in the guidelines published in Annex 8 of OSPAR (2013) and Brockmann and Topcu (2014). The method does not only account for the temporal coverage of the data, but also for the resolution at which large, fast changes (gradients) are sampled. The idea is that if the gradient is flat, not many measurements are necessary to sample the variability and a gap in the measurements would not result in a significant loss of representativeness. However, if the gradient is steep, we would need a higher frequency of sampling to be able to capture the variability, and a gap would have more weight in reducing the representativeness.

The Brockmann and Topcu (2014) method consists of dividing time and/or space into regular intervals/cells and checking whether all of the intervals have been sampled. If an interval has been sampled, it gets the full confidence of 100/N, with N the number of intervals/cells in which the time/space has been divided. Thus, if all the intervals/cells have been sampled, the final representativeness is 100% (P<sup>N</sup> 1 100/N).

If an interval is not sampled, it gets a reduced score that depends on the difference in gradient between the next sampled cells (calculated as a percentage of the overall gradient) and the number of connected empty cells. In general, the representativeness of an empty interval is given by:

$$R = OR - G \* n \* \frac{OR}{100} \tag{1}$$

with R the representativeness of the empty interval (%), OR the full representativeness of the interval (%), n the number of empty intervals, and G the maximum difference between minmax values of the nearest sampled cells divided by the overall difference in min-max (in %). This is a slight modification of G with respect to Brockmann and Topcu (2014) and follows Annex 8 of the guidance (sections B1 and B2, OSPAR, 2013). If R is negative, it is assigned a score of 0, since it is not contributing to the overall representativeness. For this study, the width of the temporal intervals for the calculation of the temporal representativeness was chosen to be 1 month. Notice that, since we evaluated the temporal representativeness only for the winter months for DIN (January, February, November, and December) and the growing season for chlorophyll (March to September), the empty intervals for which the closest available data were located more than 6 months appart were assigned a score of zero to avoid calculating gradients with data corresponding to different years.

The spatial representativeness was assessed by dividing the assessment area into 1 × 1 km grid cells and counting the number of cells that were occupied by observations from the different dataset combinations and dividing this result by the total number of grid cells in the polygon corresponding to the assessment area for the whole assessment period. The results are also given as a percentage. In this case, the gradient steepness was not considered in the calculation of the spatial representativeness. Note that the results are expected to depend on the selected temporal/spatial discretization.

#### Confidence in the Metrics

We can provide a confidence rating of the statistics/metrics by calculating the uncertainty associated with the metrics used in the assessments (averages, percentiles, etc.). In general, the uncertainty of the metrics will increase with the variability of the observations and decrease with an increasing number of observations. In the present paper, each metric has an associated 95% confidence interval, which was considered as a proxy for the uncertainty.

Differences in the confidence in the metrics of the aggregation scenarios with respect to the actual OSPAR COMP2 assessment, which will be called the reference assessment from now on, were calculated by considering the change in the width of the 95% confidence intervals.

TABLE 1 | Statistics applied to DIN and Chlorophyll and assessment thresholds.

## Confidence Relative to the Threshold

In order to limit the risk of mis-classification as non-problem area, a metric is provided to estimate the confidence level with respect to the threshold. Sections A5 and A6 of Annex 8 in OSPAR (2013) provide different methods to calculate the confidence in the classification depending on whether assessments are based on means or on percentiles.

For assessments based on means (i.e., DIN), two methods are considered:


level that would lead to the conclusion that the test values are below the classification limit." In other words, in this case we provide the width of the confidence interval (which is the difference between the threshold value and the mean) and we need to calculate the confidence level, which is done by means of the survival function considering a one-sided t-distribution.

In the case of assessments based on percentiles (i.e., chlorophyll assessment), the confidence is calculated as the cumulative probability of the binomial distribution:

$$\text{Cumulative probability} : P\left(\mathbf{x} < k\right) = \sum\_{\mathbf{x}=0}^{k-1} \left(\frac{n}{\mathbf{x}}\right) \left(\frac{p}{100}\right)^{\mathbf{x}} \left(1 - \frac{p}{100}\right)^{n-\mathbf{x}}, \text{(2)}\right)$$

where n is the total number of observations, k are the observations below the threshold, which is defined by the p percentile (90th percentile in our case).

This cumulative probability is the confidence level for the conclusion that the p percentile is less than value number k. Consequently, if k of n observations are below the classification limit, this confidence level also applies to the conclusion that the p percentile is less than the classification limit (OSPAR, 2013).

## The Optimization Approach

As an initial step toward the optimization of the monitoring systems, we applied heuristic techniques to assess the observational system that was used for OSPAR COMP2, although they can be easily extended to other assessments. The techniques consisted of analyzing several scenarios of aggregation of the available datasets (see sections Datasets for winter DIN and Datasets for growing season chlorophyll) for each of the studied variables (DIN and Chlorophyll). In addition, alternative datasets provided by different observational platforms were available for chlorophyll (FerryBox and satellite data) for the period 2001–2005, allowing for an analysis of the impact on the assessment of using higher spatial- and temporal- resolution datasets. Using additional datasets like FerryBox or satellite chlorophyll involves merging data from different sensors that are not necessarily cross-validated. For this reason, we have carried out a cross-validation exercise among the different datasets to give a quantitative evaluation of the existing mismatch. Finally, all the aggregation scenarios were compared to the reference assessment.

## Aggregation Scenarios

#### **For DIN**

Of the three datasets considered for the DIN assessment (see section Datasets for winter DIN), two are operated by Cefas: sap and sbu. Therefore, we focused on evaluating the importance of these two datasets. The scenarios were:


#### **For chlorophyll**

Only two datasets were available for the OSPAR COMP2 assessment of chlorophyll in East Anglia: wfd and sbu. We used two additional datasets to analyze their effect on the assessment: the Cuxhaven-Harwich Ferrybox chlorophyll data and MODIS daily chlorophyll satellite images (see section Additional datasets for more details).

The aggregation scenarios that were studied for chlorophyll were:


#### Cross-Validation Between Datasets

Merging datasets obtained from different sensors requires an analysis of the similarity between the available measurements. In this section we introduce a cross-validation tool quantifying the degree of mismatch between the different datasets compared one by one. The cross-validation tool was applied to all the DIN and chlorophyll datasets and consisted of the following steps: (a) spatial gridding of the study area considering a 1 × 1 km grid, (b) finding the points belonging to two different datasets that coincide in the same grid at the same time (considering a daily resolution), (c) calculation of the number of crossing points, correlation, root-mean square error (RMSE), and standard deviation to determine the matching between datasets.

#### Metrics for Comparing the Aggregation Scenarios

To analyze the importance of a certain dataset for the eutrophication assessment and, ultimately, to answer the questions posed in the Introduction (see section Introduction), we needed to compare the results of the different aggregation scenarios with a reference which, in this case, was the original OSPAR COMP2 eutrophication assessment. We used the following criterion: an aggregation scenario will be considered "similar to" the reference if the assessment result for the analyzed variable lies within the 95% confidence interval of the reference assessment.

## RESULTS

### Cross-Validation of the Different Datasets

For DIN, all the space/time matchings between wfd and sap, wfd and sbu, and sap and sbu occurred in the Thames estuary adjacent to the Warp SmartBuoy in accordance with the distribution of DIN observations in **Figure 2A**.

All chlorophyll data locations, including the additional datasets, are given in **Figure 4A**. The positions of the space/time matchings for chlorophyll between wfd and sbu, wfd and sch, sbu and sch, and fbx and sch, are shown in **Figures 4B–E**, respectively. No space/time matching between wfd and fbx and sbu and fbx were found.

The statistics for all these matchups were represented by means of normalized Taylor diagrams (Taylor, 2001, see **Figure 5**). The number of matchups for each dataset combination is also included.

For DIN, all dataset combinations were positively correlated (>50%, see the circles in **Figures 5A,B**), but with a low number of matching points and, especially in the case of the comparison

FIGURE 4 | (A) Datasets for the chlorophyll assessment of East Anglia (COMP2) together with the two additional datasets: FerryBox (fbx) and satellite chlorophyll (sch) explored in the scenario testing (see section Aggregation scenarios). (B–E) Space and time matching between the different datasets.

between sbu and sap, different variability, and high RMSE. In the case of chlorophyll, the number of matchups was reasonably high for all the cross-comparisons, except for wfd and sbu, which had only 19. The worst statistics were obtained for the cross validation of wfd and sch (see **Figure 5A**). The number of matchups in space was quite large with respect to the total number (56 crossing points), meaning that very few data points matched at different times at the same location. The opposite happened for the cross-validation of sch with sbu, which represented a unique point in space for which a large number of temporal matches occurr. In this case, better matching statistics were obtained compared to wfd and sch (compare the purple square in **Figure 5B** with the same in **Figure 5A**). The comparison of fbx and sch (**Figure 5C**), with the largest number of matching points, resulted in a slightly higher RMSE than for the cross validation of sbu with sch, but the representation of the variability was better, with a similar correlation.

#### Reduced Sampling Scenarios and Use of Additional Datasets Winter DIN

The results of the OSPAR COMP2 assessment for winter DIN (wfd+sap+sbu), i.e., the reference scenario, and the different aggregation scenarios are given in **Table 2** for each of the assessment years and the whole assessment period for the coastal and offshore regions in East Anglia. The confidence in the metrics and confidence relative to the thresholds were reported by the 95% confidence interval and the confidence in the threshold column, respectively, for each of the water types (coastal/offshore) and each of the aggregation scenarios.

A representation of the assessment results and the confidence in the metrics is shown in **Figure 6A** (coastal waters) and **Figure 6B** (offshore waters), where the impact of aggregating the different datasets is clearly seen.

The confidence in the representativeness of the data used to produce the assessment results is summarized in **Table 3**, with


Results show the mean and 95% confidence interval for each year and the confidence with respect to the threshold (Conf. Thres.) both for the coastal and offshore waters. The results for the whole assessment period are also shown ("All"). In the Reference column, numbers in red highlight results above the threshold and numbers in green show results below the threshold. For the rest of the columns, numbers in black bold indicate that the results are similar to the reference (see section Metrics for comparing the aggregation scenarios) and there is no change in the assessment result; plain numbers show that the results are not similar to the reference, but there is no change in the assessment result and orange numbers show those cases for which there is a change in the assessment results. n/a indicate no data or insufficient data. The underlined numbers highlight the situations for which the results are similar to the reference but there is a change in the assessment results.

FIGURE 6 | Bar plot showing the results of the OSPAR COMP2 assessment for DIN in coastal waters (A) and offshore waters (B) together with the 95th confidence intervals. Each bar represents one of the analyzed data aggregations given in section Aggregation scenarios. The red line shows the assessment threshold.

TABLE 3 | Temporal and spatial representativeness of the winter DIN reference dataset and all the aggregation scenarios.


**Figure 7** showing the number of available data for the reference and the aggregation scenarios in the selected temporal intervals (see **Figure 7A**). This figure also shows an illustration of the monthly averaged, minimum and maximum time series for the reference and the wfd aggregation scenarios (see **Figure 7B**), which gives an idea of the steepness of the gradients. The shaded areas correspond to the gaps in the wfd dataset. The scores (in percentage) for each of the time intervals of the wfd dataset following Brockmann and Topcu (2014) (see section Estimating confidence) are plotted in **Figure 7C**. Notice that the gaps for November and December 2003 are assigned a score zero because the closest available observations correspond to February 2003. In this case, the calculation of a gradient between February 2003 and January 2004 to produce a reduced score would not be meaningful. The same argument was applied to assign a score zero to December 2014. The total temporal representativeness for wfd in **Table 3** is the result of summing the percentages for all the temporal intervals in **Figure 7C**.

A summary of the changes of the different scenarios with respect to the reference assessment is given in **Table 6**.

The reference assessment for DIN—wfd+sap+sbu—was characterized by good temporal representativeness (100%) and not so good spatial representativeness (6.06%, see **Figures 2A**, **7A**). However, the number of observations was not homogeneous over time (see **Figure 7A**, blue bars), with good coverage in 2002 and few data in 2005, or space (see **Figure 2A**), with higher coverage closer to the coast and in the Thames Estuary.

The results of the assessment for the coastal region are shown in **Table 2** and **Figure 6A**. The assessment threshold was exceeded in all years and over the whole assessment period. In offshore waters (**Table 2** and **Figure 6B**), the mean did not exceed the threshold in 2003. The assessment for the whole period also indicates that the threshold was exceeded. For all the years and the overall assessment, the confidence in the metrics was relatively high, as indicated by the small confidence intervals.

All the considered reduction scenarios led to overall assessment results for DIN that were aligned with the reference assessment for both the coastal and offshore waters (see **Table 6**). Only in the case of the aggregation scenario wfd+sbu were overall results within the 95% confidence interval of the reference assessment (see **Tables 2**, **6** and **Figure 6**). In other words, only wfd+sbu led to overall results similar to the reference in the sense of section Metrics for comparing the aggregation scenarios. On a year-to-year basis, wfd+sbu also provided similar results to the reference for all the years except 2005 in the coastal waters (the year for which less data were available, see **Figure 7A**), and for all years except 2003 and 2005 in the offshore waters. Notice that, although the results are similar to the reference in 2004 for offshore waters in the sense of section Metrics for comparing the aggregation scenarios, there is a change with respect to the assessment results, leading to non-exceedance. This aggregation scenario (wfd + sbu) slightly reduced the confidence in the metrics (<10% for coastal waters and <13% in offshore waters) but the spatial and temporal representativeness remained almost unchanged.

For the rest of the aggregation scenarios there was an important reduction (>28%) in the data representativeness (either in time or space) and/or in the confidence in the metrics, implying that some of the relevant spatio/temporal scales have



Metrics for comparing the aggregation

orange numbers show those cases for which there is a change in the assessment

there is a change in the assessment

 scenarios) and there is no change in the assessment

 results.

 result; plain numbers are used to show that the results are not similar to the reference, but there is no change in the assessment

 results. n/a indicate no data or insufficient data. The underlined numbers highlight the situations for which the results are similar to the reference but

 result and

TABLE 5 | Temporal and spatial representativeness of the chlorophyll reference dataset and all the aggregation scenarios.


TABLE 6 | Summary of the changes with respect the reference assessement of the different aggregation scenarios.


Arrows pointing upwards (downwards) indicate an increase (decrease) with respect to the reference, and the color scale indicate the percentage of change (blue, <10%, green, between 10 and 25%, and red, more than 25%).

been lost. However, there were still periods of time when certain datasets presented coverage in time and space similar to the reference assessment (see **Figures 2A**, **7A**), and hence similar results coincident with these periods of high data coverage. These were 2002 and 2004 for the coastal waters in scenario wfd, 2003 for the coastal and offshore waters with sap and 2002, 2003, and 2004 for the coastal waters, and 2003 for the offshore waters with wfd+sap.

#### Chlorophyll

**Table 4** gives the results of the OSPAR COMP2 assessment for chlorophyll (wfd+sbu) and for all the aggregation scenarios. It is interesting to notice that, given that the assessment for chlorophyll is based on a percentile (90th percentile), the 95% confidence intervals are not symmetrical, so we provided the width of the lower and upper confidence intervals. **Figures 8A,B** depict the results of the assessment for the coastal and offshore waters, respectively. The confidence in the representativeness of the data used for the assessment is summarized in **Table 5** (plots not shown).

The reference dataset—wfd+sbu—showed good representativity in time (97.63%), although the representativeness in space was <10% (see **Table 5**). The sbu dataset was spread quite homogeneously in time, and it was the only available dataset in 2005. On the other hand, wfd presented high variability in terms of number of data points,

with 2004 the year for which more data were available (e.g., see **Figure 2D**). The results of the assessment for the coastal regions are shown in **Table 4** and **Figure 8A**. The assessment threshold was exceeded for all the years in the combined period, except for 2005, with very low confidence relative to the threshold (<16%). An exceptionally high value was obtained in 2004, due to intensive sampling in June 2004 coinciding in time and space with a massive phytoplankton bloom at the northern part of the East Anglia region (see **Figure 9**). For offshore waters, the results never exceeded the threshold, during the whole assessment period, with very high confidences relative to the threshold.

For chlorophyll, the reduction scenarios consisted of analyzing the individual datasets that comprised the reference scenario (wfd and sbu). In the case of wfd, the overall results of the assessment were aligned with the reference, although not providing similar results in the sense of section Metrics for comparing the aggregation scenarios (see **Table 6**). Removing the sbu dataset led to a decrease in the temporal representativeness (30%), and reduced the variability of the dataset, causing an increase in the confidence in the metrics (65%). Wfd alone produced similar results to the reference in years 2002, 2003, and 2004 in coastal waters, and in 2002 and 2003 in offshore waters (the only 2 years for which the number of data was enough to carry out the assessment), but led to a result different to the reference in years 2001 and 2003 in coastal waters (no threshold exceedance vs. threshold exceedance in the reference assessment, see **Table 4** and **Figure 8**).

Using the sbu dataset alone resulted in a change in the assessment results for coastal waters (the threshold was not exceeded), but not for offshore waters. In the latter case, the results were not similar to the reference according to section Metrics for comparing the aggregation scenarios (see **Table 4**). It was not surprising that removing the wfd dataset had more impact in the coastal waters, because most of the samples were collected in this water body (see **Figure 2**). The associated loss in spatial representatitivity was high (see **Table 5**). Sbu data alone could produce results similar to the reference assessment for years 2003 and 2005 in coastal waters (we recall that for 2005 it was the only available dataset), and 2003, 2004, and 2005 in offshore waters, but the results were different for 2002 and 2004 in coastal waters (see **Table 4** and **Figure 8**). Notice that the years for which the sbu results were different to the reference assessment are the same as those for which wfd alone produced similar results, meaning that for these years, only wfd was covering the relevant spatio/temporal scales.

The rest of the studied scenarios for chlorophyll consisted of the utilization of additional datasets. For all these scenarios the results of the overall assessment in coastal waters were opposite to the reference assessment, always resulting in no threshold exceedance. On the contrary, the offshore waters were aligned with the results of the reference assessment, although never within the 95% confidence interval (see **Tables 4**, **6** and **Figure 8**). In all the cases, the inclusion of additional datasets led to an increase in the spatial representativity (>25%), in the confidence in the metrics (>65%, mostly in coastal waters) and in the confidence relative to the threshold (>99%, see **Table 6**), and only a slight change in the temporal representativeness.

All scenarios that included sch (wfd+sch, wfd+sbu+sch, and wfd+sbu+fbx+sch) were biased to this dataset for being the biggest one, and all of them produced opposite results to the reference assessment for years 2002, 2003, and 2004 in coastal waters, and only similar results to the reference assessment for year 2004 in offshore waters (see **Table 4**).

When sbu was replaced by fbx (scenario wfd+fbx), the results were similar to the original assessment in 2004 for offshore waters. If fbx was combined with the original dataset (wfd+sbu+fbx), years 2004 and 2005 became similar to the reference in the offshore waters, although 2002 became different to the reference (no threshold exceedance) in the coastal waters. It is important to notice that no fbx data were available for years 2001 and 2003 and that the results in **Table 4** were obtained considering an averaging interval of 10 min for the fbx data. The results of a sensitivity test for aggregation scenario wfd+fbx showed that different averaging intervals led to different assessment results. In general terms, the shorter the averaging interval, the lower the 90% percentile values. For example, different averaging intervals for coastal waters resulted in 90th percentiles for the chlorophyll assessment of 6.2 (1 min average interval), 13.2 (10 min average interval), 21.9 (30 min average interval), and 30.4 (60 min average interval), compared with the 33.76 value of the reference assessment (see **Table 4**). It was beyond the scope of this paper to investigate the most appropriate averaging interval that guarantees that all temporal and spatial autocorrelations are removed, and should be the subject of further research.

#### DISCUSSION

The consideration of different dataset aggregation scenarios can test if different scales of eutrophication monitoring effort can deliver similar results without significantly affecting confidence and representativeness in assessments. These reduced scenarios can provide cost efficiencies but need to be considered in terms of the adequacy of the reduced datasets. Here we discuss the various options using the questions posed in the Introduction.

#### What Impact Does Each Dataset Have on the Results? i.e., Would We Obtain the Same Assessment Results if We Excluded the SmartBuoys (sbu) or the Ship-Based Sampling (sap)? Winter DIN

The outcomes of the winter DIN assessment showed that none of the individual datasets (wfd, sbu or sap) can individually reproduce the results of the OSPAR COMP2 assessment for either single years or the whole period, thus showing that none of them are redundant in the calculation of the assessment statistic (normalized mean). Moreover, either sap or sbu data would be necessary for the assessment of the offshore waters, which is not possible with the wfd data alone, since they cover only the coastal waters.

The availability of estuarine data (covered by the wfd dataset) was crucial for the results of the assessment given the way the normalized means are calculated (see **Figure 3**). For instance, in 2001 and 2002 (see **Figures 3A,B**), wfd included low salinity/lower nitrate data, which were less available in 2003 (see **Figure 3C**). This led to steeper slopes in the mixing diagram in 2003 and, hence, higher values of DIN that reflected a lack of observations, and not necessarily a situation of nutrient enrichment. Similarly, in 2004 and 2005, when no estuarine data were available (**Figures 3D,E**). The mixing diagrams for

years 2001 and 2002 suggest that the salinity/DIN gradients from the estuaries to the offshore waters are strongly spatially variable within the East Anglia region. Therefore, splitting this region into more meaningful areas in terms of river plume dynamics, hydrodynamics, etc. would probably have led to a more realistic eutrophication assessment. Assessment areas delineated based on ecologically relevant typologies (salinity, extent of the river plume, ecohydrodynamic characteristics) have been explored in a detailed case study for the Thames and Liverpool Bay area (Greenwood et al., submitted) . Also, the use of smaller assessment areas in inshore coastal waters has been found to provide better information for managers and policy makers (Elliott, 2013). In the most recent UK OSPAR assessment (OSPAR COMP3, UK National Report, 2017), estuarine data were not considered for the calculation of the normalized means, resulting in a more consistent comparison of results from year to year.

#### Chlorophyll

For chlorophyll only two datasets were available for the OSPAR COMP2: wfd and sbu. As for DIN, neither of them could, on their own, produce similar results to the reference for all years, meaning that both datasets were providing important information to the assessment.

The nature of the variability in chlorophyll maxima in time and space makes chlorophyll difficult to sample, even with continuous sampling devices such as the SmartBuoy. The way the chlorophyll assessment is designed can lead to false nonexceedance results for different reasons. For example:


#### Wider Implications

If the SmartBuoy was removed, although we could get results similar to the original assessment in terms of threshold exceedance, they would not lie in the 95% confidence interval of the reference because of the loss of temporal representativeness. Removing the SmartBouy (data) would save approximately £70 k per year, but would result in an increase in the uncertainty due to the decreased temporal representativeness. This higher uncertainty would result in lower confidence in the assessment outcomes. As an aside, the SmartBuoy programme as a whole contributes to increased scientific knowledge in the area by providing long-term time series on environmental changes, data for satellite, and model calibration and validation, etc. Since 2002, more than 60 peer-reviewed papers have been published using SmartBuoy data.

The removal of the ship-based sampling did not seem to significantly affect the results of the assessment for DIN, except for 2005 in coastal waters, and 2003 offshore (these are water body/year combinations for which data are particularly scarce). A substantial amount of these data is collected during SmartBuoy turn-around cruises, and used to calibrate SmartBuoy observations. Hence, reducing SmartBuoy deployments would also reduce the volume of available ship-based data. In addition to SmartBuoy calibration, this dataset is used for the validation of satellite and FerryBox data.

The conclussions with respect to the relevance of each dataset presented here are only valid for the East Anglia Regional Sea. We cannot anticipate if some datasets would be redundant or not in other assessment areas because of the different dynamics and hence, characteristic spatio-temporal scales. However, the employed methodology is easily and quickly applied to other regions and assessments.

The statistical techniques proposed in this paper allow for a preliminary assessment of the monitoring system based on simple methods. In particular, we are able to evaluate if a dataset is redundant or not. A dataset is redundant if it covers spatio-temporal scales that have already been covered by other available datasets. This information is relevant and constitutes an important step toward the optimization of the monitoring system, but with this methodology we still do not quantify to which extent the relevant spatio-temporal scales have been covered by the available datasets. The calculation of the temporal and spatial representativeness presented in this paper gives a partial idea of the data coverage (notice however that lower values would be expected if they were combined in a 3D matrix: longitude x latitude x time), but not of its effectiveness in covering the relevant scales. More sophisticated statistical techniques, such as the effective coverage and the explained variance (see She et al., 2007; Fu et al., 2011) or assimilative model-based methods (OSEs and OSSEs, see She et al., 2007; Oke and Sakov, 2012; Turpin et al., 2016) could be used for this purpose, although this was beyond the scope of this paper.

#### Does the Addition of New Platforms (Not Included in the Assessment, Like FerryBox or Satellite Data) Significantly Change the Conclusions of the Assessment? Added Value

According to the results in section Chlorophyll, when the high frequency platforms are considered in the assessments, either by replacing the SmartBuoys (aggregation scenarios wfd+fbx and wfd+sch) or by combining them with the existing datasets (aggregation scenarios wfd+sbu+fbx, wfd+sbu+sch, and wfd+sbu+fbx+sch), the conclusion of the assessment changes significantly, especially in coastal waters. Indeed, in coastal waters a change in the comparison with the thresholds occurs, leading to non-exceedance results when the assessment reported exceedance. The results are less dramatic in offshore waters, for which the assessments using fbx or sch are similar in some years and, when they are not, at least they do not demonstrate a change in the comparison with the threshold.

The high frequency platforms are providing information at scales that are not covered by the available monitoring. But we need to be able to explain the differences with the reference assessments by identifying the issues with the aggregation of the different data sources and the weaknesses of the assessment tools that are being used for chlorophyll currently.

#### Data Quality and Quantity Considerations

FerryBox and satellite chlorophyll observations are gathered using different sensors and methodologies from in-situ observations and this is the reason why we cross-validated the different datasets to estimate possible biases that could be affecting the final solution. The cross-validation exercise carried out in section Cross-validation of the different datasets could not give information about the degree of comparability between FerryBox and the in-situ datasets as there were no common points. However, fbx could be compared with sch, resulting in statistics like those for the comparison of the in-situ datasets (see, for instance, the statistics for wfd+sbu vs. fbx+sch in **Figures 5A,C**). However, according to the information in the FerryBox website<sup>2</sup> , the sensors to obtain the chlorophyll concentrations in the FerryBox (fluorometers) need major improvements to account for the dependence of the measurements on the physiological needs of phytoplankton and on the prior illumination, and this would increase the uncertainty of this dataset. FerryBox data have been collected in recent years by the Research Vessel "Cefas Endeavor," and calibrated against in-situ observations. These high quality data were not available for the OSPAR COMP2, but they constitute an additional and reliable dataset that can be used for future assessments.

In the case of the satellite chlorophyll, the cross validation exercise showed that its temporal variability is comparable with that of sbu (**Figure 5C**). However, when we compare the sch dataset with wfd we get high bias and RMSE (see **Figure 5A**). Most of the matchups between these two datasets (see **Figure 4C**) are influenced by the Thames river plume, and to a lesser extent, other rivers (Orwell, Stour, Colne, and Blackwater), therefore these results are not surprising, since satellite chlorophyll products tend to perform less well in turbid waters.

In the case of the FerryBox, data are gathered as the ship moves, with a temporal resolution of 20 s. This implies a huge amount of information that would be biasing the assessments to the observations on the FerryBox routes. For our particular study, we decided to average the data using a time interval of 10 min, which considerably reduced the number of data points, and increased their distance. However, this averaging procedure did not guarantee that all the temporal and spatial correlations were removed from the dataset, for which specific investigations would be required.

Satellite chlorophyll observations constituted the largest dataset in terms of combined temporal and spatial coverage. In this paper, we have considered daily products gridded at 1 km resolution. Several problems occur with the merging of satellite observations with other datasets. Firstly, retrieval of Level-2 products in coastal waters, where suspended sediment and CDOM co-occur with phytoplankton, is inherently complicated by the optical complexities of these waters (see Qin et al., 2007; Petus et al., 2010; Prieur and Sathyendranath, 2018). However, advances are being made toward the development of reliable satellite products generated with the appropriate algorithms for the different water types (i.e., the EU funded JMP EUNOTSAT project), which will reduce the current associated uncertainty. Secondly, the huge amount of data can dilute any information from other datasets, which may be more reliable. This raises questions about the accuracy of the classical assessment, and on the influence and interpretation of the statistics used. In this sense, the incorporation of high frequency observations into the assessment might require the consideration of smaller assessment areas or revisiting the actual thresholds, which are based on much less observations. Finally, clouds and other artifacts reduce coverage of the relevant areas, which reduces the availability of data, and may introduce bias toward conditions associated with clear weather.

### CONCLUSIONS

We conclude that all the in-situ datasets used in the OSPAR COMP2 assessment were relevant to replicate the results of the initial assessment, with almost no margin to reduce costs without increasing the uncertainty in the eutrophication assessments and impacting on the ability of the data to deliver the OSPAR assessment requirements. The only case in which a reduction was acceptable was aggregation scenario wfd+sbu for the DIN assessment, since the removal of the sap dataset had almost no impact on the confidence representativeness of the data and the confidence in the metrics.

The spatial and temporal coverage and the methods used in the eutrophication assessment were biased toward certain times and locations where the sampling was more intensive. This was evident in the different annual outcomes, such as the DIN assessment in 2002 resulting in a lower assessment value than the consecutive years as more estuarine sites were sampled, or the chlorophyll assessment, that resulted in a higher value in 2004 than all other reporting years due to the sampling of a bloom.

In order to avoid these biases, an in-situ sampling programme which is more homogeneous in time and space would be required, but is not feasible due to cost. The incorporation of remote sensing and model data, which are currently not used in the eutrophication assessments, could provide the required resolution but need to be integrated with the appropriate methods. The aggregation of in-situ, satellite and modeling data offers the appropriate integration of available datasets to ensure cost efficient monitoring programs collecting data at the appropriate frequency. There would be a need for each dataset to account for its own uncertainty.

In this paper we have made an initial merging test between in-situ and satellite chlorophyll data that resulted in big changes in the assessment results. This was caused by the fact that the satellite chlorophyll was a massive dataset that contained many more low chlorophyll values (outside blooms) than high chlorophyll values, which lowered the 90th percentile. This might be an indication that the classical assessment methods should be revisited, but it first requires a more in depth study focused on

<sup>2</sup>https://www.ferrybox.com/about/sensors/index.php.en

the best way to aggregate in-situ, remote sensing and model data, which is in preparation (Collingridge et al., in preparation).

#### DATA AVAILABILITY STATEMENT

The raw data supporting the conclusions of this manuscript will be made available upon request. The datasets will be published on the Cefas Data Hub (https://www.cefas.co.uk/cefas-data-hub/).

### AUTHOR CONTRIBUTIONS

LG-G wrote the manuscript and analyzed the data. JvdM and DS had the original idea and were crucial in the design of the methodology. SP and KC provided the data and helped with

#### REFERENCES


the details of the OSPAR COMP assessment and MD provided guidance throughout the process.

### FUNDING

This study was funded by Cefas Seedcorn Project "Optimizing monitoring programmes using model results" (DP381) and part funded through a Service Level Agreement (Defra-funded).

## ACKNOWLEDGMENTS

The authors would like to thank our colleague Jon Barry for his review and comments on the manuscript and for the interesting discussions.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Crown Copyright © 2019 Authors: García-García, Sivyer, Devlin, Painting, Collingridge and van der Molen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Utilizing Eutrophication Assessment Directives From Transitional to Marine Systems in the Thames Estuary and Liverpool Bay, UK

Naomi Greenwood1,2 \*, Michelle J. Devlin1,2, Mike Best <sup>3</sup> , Lenka Fronkova<sup>1</sup> , Carolyn A. Graves <sup>1</sup> , Alex Milligan<sup>1</sup> , Jon Barry <sup>1</sup> and Sonja M. van Leeuwen<sup>4</sup>

*<sup>1</sup> Centre for Environment, Fisheries and Aquaculture Science (Cefas), Lowestoft, United Kingdom, <sup>2</sup> School of Environmental Sciences, University of East Anglia, Norwich, United Kingdom, <sup>3</sup> Environment Agency, Kingfisher House, Peterborough, United Kingdom, <sup>4</sup> COS Division, NIOZ Royal Netherlands Institute for Sea Research, Utrecht University, Texel, Netherlands*

#### Edited by:

*Dorte Krause-Jensen, Aarhus University, Denmark*

#### Reviewed by:

*Theo C. Prins, Deltares, Netherlands Sai Elangovan S, National Institute of Oceanography (CSIR), India*

\*Correspondence: *Naomi Greenwood naomi.greenwood@cefas.co.uk*

#### Specialty section:

*This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science*

Received: *31 October 2018* Accepted: *25 February 2019* Published: *16 April 2019*

#### Citation:

*Greenwood N, Devlin MJ, Best M, Fronkova L, Graves CA, Milligan A, Barry J and van Leeuwen SM (2019) Utilizing Eutrophication Assessment Directives From Transitional to Marine Systems in the Thames Estuary and Liverpool Bay, UK. Front. Mar. Sci. 6:116. doi: 10.3389/fmars.2019.00116* The assessment of eutrophic conditions is a formal requirement of several European Directives. Typically, these eutrophication assessments use a set of primary indicators which include dissolved inorganic nutrients, chlorophyll, dissolved oxygen and secondary information such as phytoplankton community data. Each directive is characterized by a different geographical or political boundary which defines the area under assessment. Several disparate sources of data from the Thames estuary and Liverpool Bay in the United Kingdom collected from different monitoring programs were combined to generate a fully integrated dataset. Data sources included remote sensing, ecosystem models, moorings, freshwater inputs and traditional ship surveys. Different methods were explored for assigning ecologically relevant assessment areas including delineation of the assessment area based on salinity, extent of the river plume influence and ecohydrodynamic characteristics in addition to the traditional geographically defined typologies associated with the different directives. Individual eutrophication indicators were tested across these revised typologies for the period 2006–2015, and outcomes of the different metrics were compared across the river to marine continuum for the two UK areas. There have been statistically significant decreasing trends in the loads of ammonium, nitrite and dissolved inorganic phosphorous between 1994 and 2016 in both the Thames estuary and Liverpool Bay study areas but no statistically significant trends in loads of nitrate or dissolved inorganic nitrogen. There have been statistically significant increases in riverine nitrogen:phosphorous between 1994 and 2016. Nutrient concentrations exceeded assessment thresholds across nearly all areas other than the large offshore assessment areas, and outcomes of the chlorophyll metric were often below assessment thresholds in the estuarine-based areas and the offshore areas, but exceedances of thresholds occurred in the near coastal areas. However, trait-based indicators of phytoplankton community using functional groups show changes in plankton community structure over the assessment period, indicating that additional metrics that quantify community shifts could be a useful measurement to include in future eutrophication assessments.

Keywords: eutrophication, assessment, marine, freshwater, nutrients

## INTRODUCTION

Coastal marine ecosystems have been impacted globally by a range of anthropogenic activities including elevated inputs of nutrients (Jickells, 1998; Conley et al., 2009; Rabalais et al., 2009). Inputs of nutrients from direct discharge of waste water and diffuse sources including agricultural runoff and atmospheric deposition have led to many regions experiencing eutrophication, which includes undesirable changes in the marine ecosystem including increased primary production, accumulation of organic matter and an associated decrease in oxygen concentration (Nixon, 1995, 2009; Cloern, 2001; Cloern and Jassby, 2010; Breitburg et al., 2018). Coastal eutrophication is likely to continue into the future due to the increasing use of fertilizers, discharge of human waste and hydrologic modifications, with these impacts exacerbated by the warming climate (Rabalais et al., 2009; Seitzinger et al., 2010; Paerl et al., 2014).

Humans have significantly altered the balance of the nitrogen (N) and phosphorous (P) cycles which has led to documented changes in riverine and coastal N:P ratios with potential consequences for marine phytoplankton communities, including altered species composition and reduced biodiversity (Turner et al., 2003; Philippart et al., 2007; Grizzetti et al., 2012; Paerl et al., 2014; Burson et al., 2016). Nutrient inputs to riverine and coastal systems may come from diffuse sources (e.g., agricultural run-off and atmospheric deposition) and point sources (e.g., sewage treatment and industrial discharge). Measures to reduce nitrogen and phosphorous inputs are frequently more successful at reducing a single source of nutrients via targeted policies rather than all nutrients and it is recognized that parallel reductions in both nitrogen and phosphorous inputs are required to reduce coastal eutrophication and the impacts associated with a changing nutrient regime (Conley et al., 2009).

In Europe, environmental directives exist for assessing the status of freshwater, coastal and marine environments. These include the Nitrates Directive (EC, 1991), Water Framework Directive (WFD, see EU, 2000) and more recently, the Marine Strategy Framework Directive (MSFD, see EU, 2008). Agreements on eutrophication assessments have also been made under the Oslo and Paris Conventions for the Protection of the Marine Environment of the North-East Atlantic (OSPAR, see OSPAR, 2013) and the Baltic Marine Environment Protection Commission-Helsinki Commission (HELCOM, see Andersen et al., 2011). These environmental frameworks are used for making harmonized assessments of marine eutrophication through the assessment of several key criteria which detail the cause and impact of increased nutrient delivery. Whilst each directive or process is different in terms of scope, assessment area and integration rules, there is a set of common indicators across all the directives. These common or "primary" indicators include dissolved inorganic nutrients, chlorophyll-a, phytoplankton abundance and dissolved oxygen to assess conditions relative to thresholds identifying an accepted ecological state (Devlin et al., 2007a, 2011; Foden et al., 2011; Tett et al., 2013). Each directive typically relates to the assessment of criteria within a set of geographical or political boundaries which do not always represent an ecological boundary. These geographical constraints can mean that assessments are not always coordinated across a river to coast continuum, making it difficult to align with a program of measures that can be developed through the river basin management plans (UK Gov, 2016).

Within the United Kingdom (UK), there are several national agencies with responsibility for collecting data and reporting on the assessment of coastal and marine environmental status under the requirements of the WFD, OSPAR Common Procedure (OSPAR CP) and the MSFD. To date in the UK, assessments under the WFD and OSPAR have been made separately based on geographically defined regions (Foden et al., 2011; UK National Report, 2017), which does not necessarily provide a coordinated assessment across a river basin. WFD assessments are carried out at a waterbody level, where waterbodies are differentiated by estuarine and coastal typologies. Specific reference conditions have been developed for each type of system where waterbody type is defined by characteristics including tidal range, mixing, exposure and salinity (Devlin et al., 2011). Waterbodies are typically estuarine or coastal areas and range from between 0.05 and 1,200 km<sup>2</sup> . The OSPAR CP screens for problem, potential problem and non-problem areas. The full assessment under OSPAR CP is only applied to problem or potential problem areas. All WFD waterbodies are designated based on the WFD assessment criteria and processes. For the OSPAR CP reporting, a WFD waterbody that has been designated as moderate, poor or bad under the WFD assessment is designated as a problem area under the OSPAR CP. The WFD data used to make the initial WFD assessment is not included in the OSPAR assessment of the regional seas areas.

The aim of this work was to improve our understanding of the eutrophication status across the salinity gradient in two study areas that receive significant anthropogenic freshwater nutrients: the Thames estuary and Liverpool Bay in England, by reporting individual criteria that are common to both WFD and OSPAR, and by exploring the outcomes of these criteria across several different assessment areas. The Thames and Liverpool Bay marine area catchments differ in terms of their geology, agriculture, population and landuse, resulting in different patterns of nutrient discharge. Both have historically received significant anthropogenic nutrient inputs. The primary indicators of eutrophication were applied to alternative ecologically relevant typologies in addition to the WFD- and OSPAR-defined typologies to examine whether changing the geographical boundaries of the assessment areas alters the outcomes of the primary indicators. Metrics from both WFD and OSPAR were applied to the different assessment areas and tested over a 10-year period (2006–2015). The objective was not to repeat or test the recent WFD or OSPAR outcomes but to investigate the stability of the metrics as the assessment areas are shifted into more ecologically relevant areas. Additional reporting of state was also tested by the inclusion of communitybased indicators through the assessment of the long-term change in functional phytoplankton life forms. The outcomes of the metrics were also used to improve understanding of the effects of anthropogenic nutrient loading across the river to marine continuum.

## MATERIALS AND METHODS

#### Study Areas

Two study areas were selected: Thames in the south east of England and Liverpool Bay in the north west of England (**Figure 1**). In the Thames catchment, the total farmed area is 1,398 × 10<sup>3</sup> hectares, of which 12% is permanent pasture and 78% is arable land. In contrast, in the catchment area of Liverpool Bay, the total farmed area is 940 × 10<sup>3</sup> hectares, of which 62% is permanent pasture and 21% is arable land (Defra, 2018). The remainder of farmed land in both catchments is used for dairy herds, pigs and poultry. This difference in agricultural land use contributes to different nutrient loadings between the two study areas. Both study areas also have large centers of population within their catchments. The population in the Thames River Basin District is over 15 million people (Environment Agency, 2018a) and it is nearly 7 million people in the North West River Basin District, which encompasses the Liverpool Bay study area (Environment Agency, 2018b).

## Datasets

Datasets for salinity, dissolved inorganic nutrients, chlorophylla, suspended particulate matter (SPM), dissolved oxygen (DO) and phytoplankton community in the period from 2006 to 2015, which covered a range of spatial and temporal scales, were compiled from different sources (see **Supplementary Table 1** for a summary) described in detail below. Trends in freshwater nutrient inputs were assessed between 1994 and 2016 as an indicator of long-term pressure on the two study areas.

#### Freshwater Flow and Inorganic Nutrient Loads

Riverine inputs of freshwater and inorganic nutrients for 1994– 2016 were processed from raw data provided by the Environment Agency and the National River Flow Archive, with permission from the Welsh government to use the Welsh data from these datasets. Data for each river were calculated as the sum of river only loads plus direct sources (sewage plus industrial discharges) downstream of the last tidal gauge point. Loads were catchment corrected to account for ungauged areas. Further details are given in Lenhart et al. (2010) and van Leeuwen et al. (2015). Data for annual loads of nitrate, nitrite, ammonium, dissolved inorganic nitrogen (DIN) and dissolved inorganic phosphorus (DIP) were calculated. Dissolved inorganic nitrogen (DIN) was calculated by summing the concentrations of nitrate, nitrite and ammonium. Annual molar ratios of DIN:DIP were also calculated. Data from the following rivers were summed for each study area. Some rivers include late-joining tributaries (where the river is already tidal) or separate rivers that join in the estuary. Some rivers have data for flow only and not nutrient loads. This detail is included in brackets as necessary after the river name.

Thames: Thames (Beam, Beverley Brook, Brent, Crane, Ingrebourne, Lee, Mardyke, Ravensbourne, Roding, Thames, and Wandle), Chelmer (Chelmer and Blackwater), Colne, Medway, Stour, Deben (flow only), Alde (Alde and Ore, flow only).

Liverpool Bay: Clwyd (Clwyd and Elwy), Alt, Conwy, Dee, Derwent, Douglas, Ethen (flow only), Ellen (flow only), Esk (Esk and Lyne), Kent (Kent and Beela), Leven, Lune, Mersey, Ribble (Ribble and Darwen), Weaver, Wyre.

The yearly trends in the nutrients from both Liverpool Bay and Thames were assessed using Generalized Additive Models (GAMs, Wood, 2006), fitted by restricted maximum likelihood. This approach was adopted because the trends are not all linear and so a method that could fit more flexible trends was needed. The fit of the GAMs is shown in **Supplementary Figure 1** for Thames estuary and **Supplementary Figure 2** for Liverpool Bay. The statistical significance of the trend was assessed using a likelihood ratio test against the null model of no trend. The residuals from the GAMs were checked for temporal autocorrelation by plotting the empirical autocorrelation function for various lags and comparing this with the values expected if the series was uncorrelated. The residuals from these models did not exhibit any autocorrelation and so uncorrelated GAM models were used.

#### Ship Based Sampling—Estuarine to Coastal

Samples were usually collected at 2 m depth using a CTD. The sites are well mixed and samples from 2 m depth are considered to be representative of the surface water. Samples were filtered for inorganic nutrient analysis (nitrate, nitrite, ammonium, phosphate and silicate), chlorophyll-a and SPM. Salinity and dissolved oxygen were determined using YSI handheld conductivity and oxygen meters. Details of sample collection and analysis are given in Kennington et al. (1999).

#### Ship Based Sampling—Coastal to Offshore

Samples were collected using Niskin bottles mounted on a CTD rosette and from the non-toxic pumped supply on research vessels for the years 2006–2015. For Liverpool Bay, research cruises occurred, on average, eight times per year between 2006 and 2011 and on average six times per year from 2012 onwards. For Thames, research cruises occurred, on average, eight times per year between 2006 and 2013 and on average four times per year from 2014 onwards. In addition, samples were collected on other ad hoc research cruises in both study areas. Details of sample analysis and quality assurance for dissolved inorganic nutrients are given in Gowen et al. (2002). Methods for the analysis of dissolved oxygen, suspended particulate matter, chlorophyll-a and salinity are given in Greenwood et al. (2010).

#### SmartBuoy

Instrumented moorings ("SmartBuoy") have been deployed in the Thames region and Liverpool Bay since 2000 and 2003 respectively (see **Supplementary Table 2** for details). The instrumentation, parameters measured and methods are described elsewhere (Weston et al., 2008; Greenwood et al., 2010). Daily average values for salinity, SPM, chlorophyll-a and DO were calculated from 10-min burst measurements made every half an hour over a 24-h period. Total oxidized nitrogen (TOxN) was measured between every 2 h and every 4 days depending on the instruments deployed. Data for total oxidized nitrogen (TOxN), chlorophyll-a, DO and phytoplankton species composition and abundance were used in this study. Seven-day average values were calculated as SmartBuoy data is temporally

correlated with a range of approximately 7 days (Capuzzo et al., 2015). Ammonium (NH4) is not determined on SmartBuoy, therefore DIN was set equal to TOxN. Historical data from CTD and underway samples at these locations for November to February (the assessment period for nutrients) show that TOxN accounted for 97% of DIN at the Thames SmartBuoy and 86% of DIN at the Liverpool Bay SmartBuoy between 2006 and 2015.

#### Phytoplankton

Water samples (150 ml) were collected on SmartBuoys using automated water samplers into pre-spiked bags containing acidified Lugol's iodine. Samples were collected weekly and returned for analysis at Cefas when moorings were recovered every 1–3 months. Samples were decanted from the sample bags into 175 ml glass jars and topped up with 1 ml of acidified Lugol's iodine. A minimum of one sample per month was selected for analysis from each deployment location where sample availability allowed. Samples were analyzed at Cefas using the Utermöhl method (Utermöhl, 1958) under inverted Olympus IX71 microscopes within 1 year of collection. Species were identified and enumerated in the samples and counts recorded in cells per liter.

Since the inception of WFD monitoring from late 2006, Environment Agency (EA) phytoplankton samples have been collected from sites in WFD waterbodies from a combination of Coastal Survey Vessels, rigid hulled inflatable boats (RIBs), and rarely from jetties or bridges in estuaries (Devlin et al., 2012). The frequency of sampling in WFD transitional and coastal waters is typically one sample per calendar month from 3 to 5 sites per water body. Ideally, samples should be 28–31 days apart throughout the year. There must be at least a 14-day interval between sampling occasions at each site. The phytoplankton samples are collected from the mixed surface layer usually between 1 and 2 m below the water surface using a standard NIO/Niskin-style water sampler, avoiding the surface film and without disturbing bottom sediments. In coastal waters or nonturbid waters >5 m depth, where the diurnal vertical migration of phytoplankton with light availability must be accounted for, samples for phytoplankton were mainly collected during daylight hours. However, for some samples, the use of integrated depth sampling using a Lund-type tube system negated the need to constrain the sampling window to daylight hours.

Samples of 250 ml volume were collected in clear PET bottles filled to approximately 90%, leaving sufficient headspace to allow for preservation and homogenization. Samples were preserved with acidified Lugol's iodine, stored in the dark, ideally at a temperature of 3◦C ± 2 ◦C for no longer than 6 months. The samples were analyzed using the Utermöhl method (Utermöhl, 1958) under inverted microscopes. Analysis was conducted at Cefas until 2013 then at both Cefas and an external laboratory from 2013 onwards. One in every 30 samples was analyzed by multiple analysts as a check for quality assurance and interanalyst comparability.

#### Satellite Suspended Particulate Matter

Spatially gridded (1.1 km resolution) monthly averages (from daily images) of non-algal Suspended Particulate Matter (SPM) were downloaded from the Cefas Data Hub (doi: 10.14466/ CefasDataHub.31). This is an interpolated and merged dataset from SeaWiFS, MODIS, MERIS and VIIRS satellite sensors. SPM was derived using the Ifremer OC5 algorithm (Gohin, 2011) and full processing details are given in Cefas (2016).

#### Biogeochemical Model (Thames Only)

Data for nitrate, phosphate, chlorophyll-a, salinity and DO were extracted from the Cefas 50-year hindcast of the North Sea (using the GETM-ERSEM-BFM model) covering 1958–2008 as described in van der Molen et al. (2014) and van Leeuwen et al. (2013, 2015) with an extension of the hindcast run to 2010.

#### Assessment Areas

The assessment areas tested in this work come from several sources including the WFD typologies for transitional and coastal waterbodies defined under (EC, 2003) and the coastal and offshore areas described in Foden et al. (2011) for the 2009 OSPAR comprehensive assessment (COMP2). The WFD waterbodies are split into coastal and transitional waters with transitional waters defined as 'bodies of surface water in the vicinity of river mouths which are partly saline in character as a consequence of their proximity to coastal waters, but which are substantially influenced by freshwater flows' (EC, 2003). WFD Coastal waters are defined as mean 'surface water on the landward side of a line, every point of which is at a distance of one nautical mile on the seaward side from the nearest point of the baseline' (EC, 2003). These waterbodies represent the classification and management unit under the WFD assessment approach.

In addition to these typologies, an additional three typologies defined using ecological and bio-physical characteristics were investigated, including (i) eco-hydrodynamic areas, (ii) salinityderived areas and (iii) river plume-influenced areas. The typologies were created using ArcGIS 10.5 and exported as shapefiles. The open-source language R and environment were used to assign typologies to the input datasets (R Core Team, 2017). The resulting assessment areas are shown in **Figure 1**. Note that the offshore areas used in COMP2 (Southern North Sea in Thames and Northeast Irish Sea in Liverpool Bay) extend beyond the areas considered in this study (Foden et al., 2011).

#### Ecohydrodynamic Areas

Ecohydrodynamic regions were identified as five distinct hydrodynamic regimes, based on stratification characteristics from the results of a 51-year simulation (Thames) and a 15-year simulation (Liverpool Bay) using the coupled hydrobiogeochemical model GETM-ERSEM-BFM. The five regimes are as follows: permanently stratified, seasonally stratified (not applicable for the areas of interest in this study), intermittently stratified, permanently mixed and region of freshwater influence, or ROFI (van Leeuwen et al., 2015; **Figure 1**).

#### Salinity-Derived Areas

Salinity was used to define additional assessment areas given its importance in determining physical conditions, biological processes and a useful proxy to define riverine influence (Foden et al., 2008). A combination of Cefas and EA ship data selected between 2006–2016 and 0.5–35.5 salinity were used to interpolate a salinity surface. Since the combined salinity dataset was autocorrelated but not normally distributed and salinity was expected to be similar locally, the Inversed Distance Weighted (IDW) interpolation was selected (Stachelek and Madden, 2015). The sensitivity analysis on the IDW input parameters was conducted using a cross-validation dataset (10% of the input dataset), and the surface with the lowest root-mean-square error and absolute error was chosen. The interpolated salinity surface was divided into three regions. The area for the region of freshwater influence (ROFI) and a salinity range between 0.5 and 15 was 2.3 km<sup>2</sup> and 0 km<sup>2</sup> for Liverpool Bay and Thames, respectively. The area for transitional waters was selected based on a salinity range between 15 and 25 and covered 108 km<sup>2</sup> for Liverpool Bay and 34 km<sup>2</sup> for the Thames. The area for coastal waters was selected based on a salinity range between 25 and 34.5 and covered 10,820 km<sup>2</sup> for Liverpool Bay and 4,809 km<sup>2</sup> for the Thames.

#### River Plume-Influenced Areas

The riverine plume assessment areas for Liverpool Bay and Thames were defined by the extent of the riverine-influenced turbidity. Whole river plumes were defined based on the mean satellite SPM between 2006 and 2015. Transects at 10 km intervals running from the coast toward the sea were plotted to identify the steepest gradient in SPM with distance from the coast as an estimate of the extent of the river plume. The value of SPM at the point of steepest gradient was extracted (25 mg l−<sup>1</sup> for Thames and 10 mg l−<sup>1</sup> for Liverpool Bay). This threshold was used to map out the influence of the Thames and Liverpool river flow from the mean satellite SPM. The areas designated as plume using this method are 5,724 km<sup>2</sup> for Thames and 1,792 km<sup>2</sup> for Liverpool Bay (**Figure 1**).

#### Assessment Methods

The primary indicators for eutrophication, DIN, DIP, chlorophyll-a and DO and the secondary indicator of phytoplankton abundances (counts) were applied over each assessment area, where there were sufficient data for each parameter. Data limitations meant that not all indicators could be run for all assessment areas. The indicators followed the OSPAR harmonized criteria (Malcolm et al., 2002; Foden et al., 2011; UK National Report, 2017) and the WFD nutrient (Devlin et al., 2007b) and marine plant assessment tools (Best et al., 2007; Devlin et al., 2007c, 2012). Data were divided into three salinity ranges; transitional (<30), coastal (Thames 30–34.5, Liverpool Bay 30–34) and offshore (Thames ≥ 34.5, Liverpool Bay ≥ 34) according to Foden et al. (2011). All available data from all sources were pooled in calculations. We recognize that in some cases this may bias results toward more temporally or spatially rich data sources, an important issue that relates to data aggregation which is the focus of future research.

#### Degree of Nutrient Enrichment

A flowchart showing the processing steps for nutrients is given in **Supplementary Figure 3**. The assessment procedures followed UK National Report (2017) for OSPAR and Devlin et al. (2007b) for WFD. DIN and DIP data were selected for the winter assessment period from November to February, inclusive using data for the whole water column. Observations with corresponding salinity < 5 were excluded based on the WFD method. Data for DIN in the transitional range were normalized based on their salinity classification to salinity 25 (transitional), salinity 32 (coastal) and salinity 34.5 (Thames offshore) or 34 (Liverpool Bay offshore). This process involved fitting a linear regression to the DIN vs. salinity relationship for all data within each region and using that regression to determine the expected DIN at the normalization salinity. The OSPAR assessment of DIN was calculated as the normalized mean of all data for each assessment year. The mean was compared with the appropriate threshold; transitional waters (30 µmol l−<sup>1</sup> ), coastal waters (18 µmol l−<sup>1</sup> ) and offshore waters (15 µmol l−<sup>1</sup> ).

Under WFD, the assessment of DIN for transitional and coastal waters is modified where suspended particulate matter (SPM) was >10 mg l−<sup>1</sup> . If the threshold was exceeded, the 99th percentile of salinity-normalized DIN was calculated. Waterbodies were classified by SPM as intermediate (10<SPM<70 mg l−<sup>1</sup> ), turbid (70 < SPM < 300 mg l−<sup>1</sup> ) or very turbid (>300 mg l−<sup>1</sup> ), dependent on the mean annual SPM. The 99th percentile threshold value was calculated from a sliding scale between these three conditions of SPM and the maxima (99th percentile) value of annual DIN; 70, 180, and 270 µmol l−<sup>1</sup> for intermediate, turbid and very turbid waters, respectively.

To provide the OSPAR assessment, the annual value of mean salinity-normalized DIN for each assessment year is reported as above (+) or below (–) the threshold or identified as insufficient (?) where there was not enough data to calculate a mean. The overall assessment for the 10-year period was made by summing the result for each individual year and assigning a color indicating predominantly above or below thresholds in the results table (**Table 1**). The WFD assessment of DIN was made in the same way as for OSPAR but data for all 10 years were pooled to calculate a single mean salinity-normalized DIN. This was compared to the same thresholds as for the OSPAR assessment of DIN. The thresholds used for the WFD assessment of DIN are the boundaries between good and moderate status (Devlin et al., 2007b).

The OSPAR assessment of DIN:DIP was made by calculating the molar ratio of all co-observed DIN and DIP data between November and February inclusive for each assessment year. The ratio was compared to the thresholds of >8 and <24 for all salinity ranges. The individual year assessments were combined as for the assessment of DIN.

#### Phytoplankton (Chlorophyll-a and Phytoplankton Counts)

The assessment procedures followed UK National Report (2017) for OSPAR and Devlin et al. (2007c) for WFD. Flowcharts showing the processing steps for chlorophyll-a and phytoplankton are given in **Supplementary Figures 4, 5**. Assessment of coastal and offshore waters was applied in the growing season from March to October, inclusive. In contrast, assessment of transitional waters was applied to the full year. For the OSPAR assessment of phytoplankton, only chlorophylla was considered. The chlorophyll-a assessment for coastal and offshore waters is the calculation of the 90th percentile of all growing season chlorophyll-a data and reported for each assessment year. This annual value is compared with the appropriate threshold; coastal (15 µg l−<sup>1</sup> ) and offshore (10 µg l −1 ). Scoring for each year was the same as described for DIN. For the WFD (coastal) assessment of phytoplankton in coastal and offshore waters, the 90th percentile of all data between March and October was calculated for the entire 10-year period, with data first averaged monthly. This single value was compared to the assessment threshold of 15 µg l−<sup>1</sup> . The chlorophyll metric was then combined with the WFD phytoplankton outcomes.

For the WFD coastal metric, the percent of all taxa counts >10<sup>6</sup> cells l−<sup>1</sup> and the percent of chlorophyll-a observations >10 µg l−<sup>1</sup> were calculated. The mean of these two values was calculated and a re-scaled metric from 0 to 1 was calculated based on the percent exceedance (Devlin et al., 2007a). The percentages of diatoms and dinoflagellates below a certain monthly reference were determined. The mean of these two values was calculated and a re-scaled metric from 0 to 1 was calculated based on the percent exceedance. The 90th percentile chlorophyll-a concentration was re-scaled to give a value between 0 and 1. The mean of these three metrics was calculated and compared to the WFD good-moderate boundary of 0.6.

The WFD transitional assessment is similarly composed of chlorophyll-a and phytoplankton counts components. A chlorophyll-a multi-metric for different statistical measures was determined (Devlin et al., 2007c) by splitting the data into two salinity ranges (5–25 and 25–30) and then calculating the following 5 metrics, with corresponding thresholds given in brackets: mean (<15 µg l−<sup>1</sup> ) and median (<12 µg l−<sup>1</sup> ) chlorophyll-a concentrations, and the percent of observations <10 µg l−<sup>1</sup> (>70%), <20 µg l−<sup>1</sup> (>85%) and >50 µg l−<sup>1</sup> (<5%). These calculations used monthly averaged values. For each of the 5 thresholds that was not exceeded in each salinity class with observations from at least 10 months of the year, a score of 1 was assigned, with the final score being the average over the 5 or 10 metric components depending on if both high and low salinity data are available. The resulting 0–1 metric was not compared with a threshold but was combined with the phytoplankton assessment.

The WFD transitional phytoplankton counts metric is composed of two components, the percent of counts for all taxa (combined) exceeding 10<sup>6</sup> cells l−<sup>1</sup> and the percent of counts for any one taxa exceeding 500,000. Both metrics were assessed against a threshold of <20%. A re-scaled metric from 0–1 was calculated based on the percent exceedance (Devlin et al., 2007c). The mean of this phytoplankton metric and the chlorophyll-a metric above were calculated and compared to the WFD goodmoderate boundary of 0.6.

#### Dissolved Oxygen

A flowchart showing the processing steps for DO is given in **Supplementary Figure 6**. The OSPAR assessment for DO was made for coastal and offshore waters. Data were filtered for the


TABLE 1 | Summary of outcomes for nutrient assessment

 tools for Thames estuary.

*below the threshold* 

 *years* 

 *black indicates insufficient data to make an assessment*

 *years* 

 *T, transitional;* 

 *coastal;* 

 *offshore.*

stratification assessment period (July to October inclusive) and only data from within 10 m of the seabed were used. The mean of the lowest quartile of DO was calculated for each assessment year and compared with the threshold >6 mg l−<sup>1</sup> . Scoring for the 10-year period was carried out as for DIN.

The WFD assessment for DO was made for transitional and coastal waters. Data for the whole year from within 10 m of the surface were used. The 5th percentile DO was calculated for the whole 10-year period. For coastal waters this was compared with the threshold >4 mg l−<sup>1</sup> . For transitional waters, this was compared with a variable threshold > 0.0286<sup>∗</sup> salinity + 5 mg l −1 . The threshold used for the WFD assessment of DO is the boundary between good and moderate status (Best et al., 2007).

A comparison of the OSPAR and WFD assessment methods for each indicator is provided in **Supplementary Table 3**.

#### Representativeness of the Data

Confidence in assessment results was investigated, in part, by the representativeness of the data for each parameter as used in the assessment calculations over the 10-year assessment period and the spatial extent of each typology. The calculation of spatial and temporal representativeness followed García-García et al. (2019) which is a modification of the method described by Brockmann and Topcu (2014). The spatial representativeness was assessed by dividing the assessment area into 0.12◦ (latitude and longitude) grid cells (approximately 10 x 10 km grid) as described in the UK National Report (2017). For some of the smaller WFD regions, this coarse resolution means that they are composed of only one grid cell, which was deemed appropriate for the purposes of comparison to larger areas; a finer scale would have required different-sized grid cells based on the size of the assessment region and in consideration of the monitoring program's sampling density. The width of the temporal intervals for the calculation of the temporal representativeness was set to 1 month following the UK National Report (2017) and which is in line with the WFD monitoring schedule. The representativeness was calculated by determining the number of temporal intervals or spatial grid cells in which observations were available and dividing by the maximum possible temporal intervals within that assessment or spatial grid cells whose centers fell within the assessment area. This method is more conservative than that of Brockmann and Topcu (2014), who assigned a non-zero representativeness to intervals without observations based on the size of the data gap and the steepness of the gradient of nearest available data.

#### Plankton Index Tool

Phytoplankton taxa names were matched to biological trait information (Tett et al., 2007, 2008) and analyzed using the Phytoplankton Index (PI) method (e.g., Whyte et al., 2017) in Matlab (Tett, 2016). Briefly, lifeform data at monthly temporal resolution are corrected for a zero value and log transformed [log (counts + z)] and plotted against one another. For diatoms z is set to 70 cells l−<sup>1</sup> and for dinoflagellates, z is set to 700 cells l −1 , which approximate the limits of detection as determined when analyzing a subsample of the original sample volume. The distribution of data during a "reference period" in lifeformlifeform space is used to define a "reference envelope" containing 90% of observations. The plankton index for any comparison dataset is then calculated as the ratio of the comparison data points which fall within the reference envelope in lifeformlifeform space to the total number of compared points. A PI value of 1–0.9 thus represents no change from the reference period, while a PI of 0 indicates no similarities between the two observation periods. For this study the reference envelope was defined using data from the final 3 years of the assessment period: 2013–2015. PI values for the years 2006–2012 thus report on the degree of change from the selected reference period. The diatomdinoflagellate pairing is considered in this study, referred to as lifeform 1 (LF1).

#### RESULTS

#### Changes in Freshwater Inputs

Freshwater nutrient loads to the Thames marine area are dominated by the rivers Thames (79.3% of total riverine DIN load and 83% of total riverine DIP load) and Medway (13.6% of total riverine DIN load and 10.6% of total riverine DIP load). Together they contribute 93% of total riverine DIN and 94% of total riverine DIP loads to the study area. In Liverpool Bay, the river Mersey contributes the greatest proportion of riverine DIN load (36%) and riverine DIP load (47%) but there are also significant contributions from the rivers Ribble, Dee, Weaver, Douglas and Clwyd. Together these six rivers contribute 82% to riverine DIN load and 89% of riverine DIP load to Liverpool Bay.

There are large variations in annual freshwater discharge and inorganic nutrient loads to the Thames estuary and Liverpool Bay (**Figures 2A**, **3A**). Between 1994 and 2016, mean freshwater discharge to Liverpool Bay was 16.1 ± 3.0 × 10<sup>9</sup> m<sup>3</sup> y −1 and ranged between 10.6 × 10<sup>9</sup> m<sup>3</sup> y −1 and 21.8 × 10<sup>9</sup> m<sup>3</sup> y −1 . This is over three times greater than that to the Thames, which has a mean freshwater discharge of 4.8 ± 1.5 × 10<sup>9</sup> m3 y <sup>−</sup><sup>1</sup> with a range between 2.3 × 10<sup>9</sup> m<sup>3</sup> y −1 and 7.6 × 10<sup>9</sup> m<sup>3</sup> y −1 . However, loads of DIN and DIP are similar between the two study areas. Loads of DIN to the Thames marine area ranged between 23,700 and 60,500 tones (**Figure 2B**) compared with between 35,800 and 58,500 tones to Liverpool Bay (**Figure 3B**). In both study areas, DIN is dominated by nitrate. Loads of DIP to Thames ranged between 2,700 and 8,000 tones (**Figure 2B**) compared with between 3,000 and 6,200 tons to Liverpool Bay (**Figure 3B**). There is strong seasonality in freshwater discharge and nutrient loads to both Thames and Liverpool Bay (**Supplementary Figures 7**, **8**). Discharge and loads of DIN and DIP are greatest in December to February and lowest between June and September.

There have been statistically significant decreasing trends (p < 0.001) in the loads of ammonium, nitrite and DIP between 1994 and 2016 to both the Thames estuary and Liverpool Bay study areas (**Figures 2B**, **3B**). There have been no statistically significant trends in loads of nitrate or DIN. Due to the decrease in ammonium loads, the percent contribution from ammonium to DIN has decreased from 30 to 9% in the Thames (**Figure 2C**) and from 35 to 18% in Liverpool Bay (**Figure 3C**). There has

FIGURE 2 | (A) Annual freshwater discharge, (B) Annual nutrient load, (C) % contribution of nitrate, nitrite and ammonium to DIN and (D) the molar N:P ratio for Thames study area.

therefore been an increase in the percent contribution of nitrate to DIN from 68 to 90% in the Thames (**Figure 2C**) and from 63 to 81% in Liverpool Bay (**Figure 3C**). Nitrite contributes between 1 and 2% to DIN over the time period. The percent contribution to total nitrate from the six main rivers to Liverpool Bay (Mersey, Ribble, Dee, Douglas, Clwyd, Weaver) has changed over time. The annual nitrate load and therefore relative contribution from

the Mersey has increased from 22 to 42% as the annual load and relative contribution from the other rivers has decreased. Nitrate loads show a strong correlation with average freshwater flow in the Thames estuary (**Figure 4A**, R <sup>2</sup> = 0.942, p < 0.001). The correlation between nitrate load and average freshwater flow is weaker in the Liverpool Bay study area (**Figure 4B**, R <sup>2</sup> = 0.642, p < 0.001).

Liverpool Bay study area.

There has been a large and statistically significant increase (p < 0.001) in the ratio N:P in riverine loads to both the Thames estuary (**Figure 2D**) and Liverpool Bay (**Figure 3D**) since 1994 due to a larger relative decrease in P compared to N. In the Thames, N:P was a minimum of 9.8:1 in 1997, well below the Redfield ratio of 16:1, increasing to a maximum of 31.7:1 in 2006, well above Redfield ratio. In Liverpool Bay, the N:P ratio was a minimum of 19.2:1 in 1997, just above the Redfield ratio and reached a maximum of 31.1:1 in 2006, well above Redfield ratio. The statistical significance of the trends in nutrient loads and N:P to Thames and Liverpool Bay marine areas are summarized in **Supplementary Table 4**.

#### Application of the Assessment Indicators and Metrics

#### Outcomes Relative to Assessment Thresholds

In the Thames, all transitional assessment areas exceed the threshold for DIN (OSPAR and WFD) and N:P (OSPAR) (**Table 1** and **Figure 5**). Salinity-normalized DIN ranges from 133.0 to 288.8 µmol l−<sup>1</sup> . All coastal assessment areas exceed the DIN threshold under both the OSPAR indicator and WFD metric, with salinity-normalized DIN ranging from 38.1 to 67.3 µmol l −1 . The offshore assessment area 'intermittently stratified' was below the thresholds for DIN (OSPAR and WFD) and N:P (OSPAR), with a mean salinity-normalized DIN of 14.1 µmol l −1 , but there was an insufficient amount of data for making an assessment for OSPAR DIN over this time period. The salinitynormalized DIN for the Southern North Sea assessment area was above the threshold for both OSPAR and WFD (16.3 µmol l −1 ) but below the threshold for the OSPAR N:P indicator. The spatial representativeness is high in most assessment areas, with a few notable exceptions including the salinity-defined transitional area, Hamford Water DIP, salinity-defined coastal area and intermittently stratified eco-hydrodynamic area. The temporal representativeness is low for the intermittently stratified ecohydrodynamic area, variable in the WFD transitional and coastal assessment areas and high for the salinity-defined, turbidity-defined and ecohydrodynamic areas.

In Liverpool Bay, all transitional assessment areas (except two with insufficient data) exceed the OSPAR and WFD threshold for DIN (**Table 2** and **Figure 6**) and all except one exceed the N:P OSPAR threshold. Salinity-normalized DIN ranges from 43.2 to 191.5 µmol l−<sup>1</sup> . Seven of the 12 coastal assessment areas (one with insufficient data) exceed the OSPAR DIN threshold and eight of 12 exceed the WFD threshold. Salinitynormalized DIN ranges from 12.6 to 119.2 µmol l−<sup>1</sup> . The spatial representativeness is high in all assessment areas. The temporal representativeness is more variable with lower values for some of the WFD transitional and coastal assessment areas.

The 90th percentile growing season mean chlorophyll for coastal assessment areas was below the threshold (15 µg l−<sup>1</sup> ) for all areas, ranging between 5.1 and 14.8 µg l−<sup>1</sup> (average 8.8 µg l−<sup>1</sup> , **Table 3** and **Figure 7**). Offshore areas were below the threshold (10 µg l−<sup>1</sup> ), ranging between 1.0 and 5.6 µg l−<sup>1</sup> (average 3.3 µg l−<sup>1</sup> ). One transitional area (Stour in Kent) was below the minimum threshold (0.6) for the combined chlorophyll + phytoplankton metric, with values for all other areas between 0.66 and 1.00 (average 0.87). All coastal and offshore assessment areas were greater than the minimum threshold for the combined chlorophyll + phytoplankton metric, with values for all other areas between 0.72 and 1.00 (average 0.87). The spatial representativeness for chlorophyll is high in all WFD assessment areas whereas the temporal representativeness is more variable. The spatial representativeness for phytoplankton is variable in the WFD assessment areas and higher than the temporal representativeness. The spatial and temporal representativeness for chlorophyll is reasonable in most of the salinity-defined, turbidity-defined and eco-hydrodynamic areas and greater than the spatial and temporal representativeness for phytoplankton.

Where there are sufficient data for PI values to be calculated, PI values for LF1 (diatom-dinoflagellate lifeform pairing) for all assessment areas show a statistically significant change (p < 0.01) due to increasing numbers of dinoflagellates in all assessment areas (**Table 3**). PI values for the four transitional areas assessed were between 0.36 and 0.66 (average of 0.47) and for the eight coastal and offshore areas were between 0.26 and 0.59 (average of 0.40). An example plot for the plume assessment area is shown in **Figure 8**, which is typical for all the assessment areas where there are sufficient data. Except for the transitional assessment area Stour (Essex), there was no significant trend in PI for LF1.

In Liverpool Bay the 90th percentile growing season mean chlorophyll for coastal assessment areas was between 1.7 and 20.8 µg l−<sup>1</sup> (average 10.6 µg l−<sup>1</sup> , **Table 4** and **Figure 9**), with three of the coastal assessment areas (Mersey Mouth, salinity coastal and Liverpool Bay) exceeding both the OSPAR and WFD coastal chlorophyll thresholds (15 µg l−<sup>1</sup> ). In addition, Morecambe Bay exceeds the WFD coastal chlorophyll threshold. The results

against the WFD combined chlorophyll + phytoplankton metric give very nearly the same result; the values for the combined metric for Morecambe Bay, salinity coastal and Liverpool Bay are below the minimum threshold of 0.6, while Mersey Mouth just exceeds the threshold with a value of 0.62. Across all coastal areas, values for the combined chlorophyll + phytoplankton metric ranged between 0.55 and 0.94 (average 0.74). In the transitional assessment areas, values of the combined chlorophyll + phytoplankton metric were between 0.27 and 0.89 (average 0.47), with six of the eight transitional assessment areas below the minimum threshold (0.6). The spatial representativeness for chlorophyll is 100% in all WFD assessment areas whereas the temporal representativeness is more variable. The spatial representativeness for phytoplankton is 100% in the WFD assessment areas except for three WFD coastal water bodies where there are no phytoplankton data. The levels of spatial and temporal representativeness for chlorophyll and phytoplankton are high in most of the salinity-defined, turbidity-defined and eco-hydrodynamic areas.

Where there are sufficient data for PI values to be calculated, PI values for LF1 (diatom-dinoflagellate lifeform pairing) for all assessment areas show a statistically significant change (p<0.01) driven by increasing numbers of dinoflagellates in all typologies (**Table 4**). PI values for the two transitional areas assessed were between 0.03 and 0.05 (average of 0.04) and for the seven coastal areas were between 0.07 and 0.28 (average 0.14). An example for the plume assessment area is shown in **Figure 8**, which is typical of the other assessment areas where there are sufficient data. Lower PI values in Liverpool Bay indicate that changes in Liverpool Bay are greater than in the Thames. Across all assessment areas, the average PI for LF1 for Liverpool Bay is 0.12 compared to 0.42 for Thames. There were no significant trends in LF1.

In the Thames estuary, the 5th percentile DO concentrations (WFD metric) for the Thames (middle) and the transitional salinity assessment areas are below the oxygen minimum threshold (4.5 and 4.6 mg l−<sup>1</sup> respectively, **Table 5**). In all other transitional assessment areas, the 5th percentile DO concentrations are between 5.3 and 7.5 mg l−<sup>1</sup> (average 6.1 mg l −1 ), greater than the minimum oxygen threshold. In coastal and offshore assessment areas, the 5th percentile DO concentrations are above the minimum threshold (4.0 mg l−<sup>1</sup> ), with values ranging between 6.2 and 8.0 mg l−<sup>1</sup> (average 7.0 mg l−<sup>1</sup> ) and the means of the lower 25th percentile (OSPAR indicator) are greater than the minimum oxygen threshold of 6 mg l−<sup>1</sup> . The spatial representativeness for DO is high in nearly all WFD assessment areas whereas the temporal representativeness is more variable. The spatial and temporal representativeness for DO in the


TABLE 2 | Summary of outcomes for nutrient assessment

 tools for Liverpool Bay.

*below the threshold (all years aggregated);*

 *black indicates insufficient data to make an assessment*

 *(all years aggregated).*

 *T, transitional; C, coastal.*

salinity-defined, turbidity-defined and eco-hydrodynamic areas assessment areas is variable, with low values for the Southern North Sea, intermittently stratified and salinity-transitional assessment areas.

In Liverpool Bay, the 5th percentile DO concentrations and means of the lower 25th percentile are greater than the minimum oxygen thresholds in all assessment areas (**Table 6**). Values of 5th percentile DO concentrations were between 5.4 and 9.7 mg l −1 (average 7.2) for transitional areas and between 6.5 and 7.5 mg l−<sup>1</sup> (average 7.0 mg l−<sup>1</sup> ) for coastal areas. The spatial representativeness for DO is 100% in all WFD assessment areas whereas the temporal representativeness is more variable. The spatial and temporal representativeness for DO in the salinity-defined, turbidity-defined and ecohydrodynamic areas assessment areas is mostly high apart from low spatial representativeness in the Northeast Irish Sea assessment area.

#### Comparison of Outcomes Between OSPAR and WFD Methods

Comparison of the WFD nutrient metric (aggregated over a 10 year period) with the reporting of the annual OSPAR primary nutrient indicators gave similar results for both Thames and Liverpool Bay despite the differences in temporal aggregation. The WFD nutrient metric is a salinity-normalized winter mean for the entire 10-year period. The OSPAR nutrient indicator is the same but assessed on 10 individual years and then reported as a final assessment based on the number of (non) exceedances for the 10-year period. In the Thames, there were some assessment areas with missing years where there were insufficient data to make an overall assessment against OSPAR DIN, but a value of WFD salinity-normalized DIN was always calculated for the entire 10-year period (**Table 1**). There were three notable differences in the DIN outcomes for Liverpool Bay where both Mersey Mouth and Liverpool Bay plume areas passed the OSPAR DIN assessment but exceeded the WFD salinity-normalized threshold. In contrast, Morecambe Bay failed the OSPAR DIN assessment but was below the WFD salinitynormalized threshold (**Table 2**).

The Thames coastal and offshore assessment areas have the same outcomes against the OSPAR chlorophyll indicator and the WFD phytoplankton metric (**Table 3**). The only differences were the assessment areas for where there was insufficient data to make an overall assessment against the OSPAR chlorophyll indicator. For Liverpool Bay, the outcomes against the OSPAR chlorophyll indicator, WFD coastal chlorophyll indicator and the WFD combined chlorophyll + phytoplankton metric were the same except for the Morecambe Bay assessment area (**Table 4**).

Outcomes for the coastal assessment areas in Thames and Liverpool Bay against the OSPAR DO indicator and WFD DO


metric were the same, although different thresholds and different statistics were applied in the two assessment processes.

## DISCUSSION

### Freshwater Nutrient Loads

Recent concerns about changing nutrient ratios (Turner et al., 2003; Philippart et al., 2007; Grizzetti et al., 2012; Paerl et al., 2014; Burson et al., 2016) highlight the impact of altered nutrient ratios on phytoplankton communities as a worldwide problem that needs to be considered in the future assessment of eutrophication levels in UK marine waters. Bowes et al. (2018) have shown that concentrations of phosphorous in the upper river Thames have decreased since 1997. These reductions are attributed to improved phosphorous stripping at sewage treatment works leading to reduced phosphorous loads from effluent rather than a reduction in agricultural inputs of phosphorous. A small reduction in nitrate was attributed to a reduction in diffuse sources such as agricultural inputs and not a reduction in nitrate load in sewage effluent. The changes in riverine nutrient loads observed in this study support the outcomes of these studies and highlight that policies targeted at reducing nutrient loads have led to significant decreases in DIP. However, there have been no significant decreases in DIN, particularly nitrate, and effective measures which also target reductions in nitrate are required to counter the significant increase in N:P observed.

The strong correlation between nitrate and river flow in the Thames (**Figure 4A**) suggests that nitrate sources to the Thames are dominated by diffuse sources (e.g., agricultural run-off). The correlation between nitrate and river flow is weaker in Liverpool Bay (**Figure 4B**). This may be because the nutrient load to Liverpool Bay comes from numerous rivers, the relative proportion of which have changed between 1994 and 2016. In addition, a mixed signal from point sources which are invariant with river flow (e.g., sewage treatment and industrial discharge) and diffuse sources such as agricultural run-off may lead to a weaker correlation between nitrate load and river flow.

#### Intra-annual Variability

The seasonal cycle in inorganic nutrients in the Thames displays maximum concentrations in the winter between December to March and a drawdown in April to June coincident with the documented spring bloom (Weston et al., 2008; Blauw et al., 2012). Concentrations of silicate reach a minimum before

compared to the reference envelope (2013–2015) in gray. The time series on the right show log10 (cell counts) for diatoms and dinoflagellates. Small black dots in the time series represent every observation, and the colored points (and line connecting them) match those in the PI diagram with colors indicating season: The reference period data (highlighted in gray in the timeseries) are not shown in the PI diagram. Dark blue: December–February, light blue = March–May, yellow = June–August, red = September–November.

minimum concentrations of ToxN are observed. The PI analysis shows a strong seasonal cycle in the phytoplankton community abundance and composition at both study sites. Abundance is lowest at both sites in the winter months from December to February. The Thames shows a peak in diatom abundance in March and April, which generally decreases during June through November (**Figure 8**). Dinoflagellate abundance is greatest in May and June then decreases throughout the summer. The strong seasonal cycle in chlorophyll and the seasonal succession in phytoplankton community has been previously reported in Weston et al. (2008). Initially dominated by diatoms, the spring bloom then switches to dinoflagellates as silicate becomes limiting. After the spring bloom, diatoms dominate the lower summer phytoplankton biomass and microzooplankton play an important role in controlling phytoplankton growth during the summer (Weston et al., 2008). Blauw et al. (2012; Blauw et al., 2018) demonstrated that in the Thames, horizontal and vertical physical mixing processes driven by the tides are important in controlling phytoplankton concentrations at short time scales. At longer time scales of weeks to months, biological growth and loss processes driven by nutrients and light are important in controlling phytoplankton concentrations.

Previous studies have demonstrated that the seasonal cycle of winter maximum nutrient concentrations in February and drawdown in April/May in Liverpool Bay are recurrent features of this location, with the timing of the drawdown varying by several weeks between years (Foster et al., 1978, 1982a,b, 1983; Gowen et al., 2000; Greenwood et al., 2011, 2012). Concentrations of chlorophyll are low between November and March, peak in April and May and gradually decrease from June onwards. The PI analysis shows that diatom abundance is elevated between March and September with highest dinoflagellate abundances between May and September (**Figure 8**). Previous analysis has shown that at the Liverpool Bay SmartBuoy site the phytoplankton community is dominated by diatoms, with dinoflagellates most abundant between July and October each year (Greenwood et al., 2010, 2011). The variability in the underwater light climate and turbulent mixing are key factors controlling the timing of phytoplankton blooms.

## Comparison of Outcomes From OSPAR and WFD Tools

Using nutrients, chlorophyll and dissolved oxygen is a good baseline for the assessment of eutrophication and, with appropriate thresholds, can provide a useful tool to assess the extent and impact of nutrient enrichment. The results of the assessments show that the current OSPAR and WFD assessment processes for eutrophication that are utilized in UK waters perform well in a cross comparison, showing similar outcomes from the application of the three primary indicators. Eutrophication in UK waters has typically been managed as a coastal issue, constrained to a small number of estuarine and coastal waters (WFD outcomes presented in UK National Report, 2017). Applying the primary indicators through both


TABLE

4


of

outcomes

for

chlorophyll

and

phytoplankton

assessment

tools

for

Liverpool

Bay.

*period; S indicates a significant (p* < *0.01) trend in PI over the assessment*

 *period; NS indicates no significant trend in PI over the assessment*

 *period.*

the WFD and OSPAR assessment approaches shows similar outcomes in the coastal assessment areas tested in this study. This is interesting given that the use of salinity and riverine plume-derived areas has extended the coastal areas beyond the one nautical mile of the coastal WFD typology. The estuarine assessment areas, which include both the WFD transitional waters and the salinity-defined estuarine area fail against the nutrient metrics under both assessment approaches, as would be expected in the heavily modified sub-catchment areas of the Thames estuary (**Figure 5**) and Liverpool Bay (**Figure 6**). The outcomes for the nutrient metrics in the coastal waterbodies are also similar between the WFD and OSPAR process. However, Harwich Approaches, Norfolk East and Hamford Water do have different outcomes between the WFD and OSPAR nutrient assessment, which is more a reflection of the limited data in several of the years than differences in assessment processes, highlighting the sensitivity of the assessments to data frequency. The salinity-defined coastal areas, the turbidity-defined plume, the ecohydrodynamic areas and the COMP2 assessment areas all have similar outcomes in the nutrient assessment, with nearly all the areas failing both the WFD and OSPAR nutrient assessment. The exception is the intermittently stratified areas, which pass the WFD nutrient assessment. The only assessment area that does not fail the nutrient assessment is the larger ecohydrodynamic areas, which reflects the larger offshore, less riverine-influenced area.

The chlorophyll metric could not be compared across the WFD and OSPAR metrics as only the WFD process has a chlorophyll metric for transitional waters. For the transitional assessment areas in the Thames, the chlorophyll metric only fails in one waterbody (Stour), reflecting the high turbidity and light-limiting characteristics of these eastern waterbodies. In contrast, the transitional assessment areas of Liverpool Bay have only two (out of eight) waterbodies (Dee and Wyre) that pass the chlorophyll sub-metric, reflecting the clearer waters with potential for higher phytoplankton growth (Cole and Cloern, 1987; Painting et al., 2007).

The coastal chlorophyll metrics for WFD and OSPAR include the assessment of the 90th percentile value of chlorophyll during the growing season. The OSPAR metric is based solely on the 90th percentile chlorophyll value whereas the WFD process has three sub-metrics: the 90th percentile chlorophyll value, phytoplankton counts and measures


TABLE

5


Summary

of

outcomes

for

oxygen

assessment

tools

for

Thames

estuary.


TABLE 6 | Summary of outcomes for oxygen assessment

 tools for Liverpool Bay.

of seasonal succession of two major phytoplankton groups (diatoms and dinoflagellates). Despite the differences in the chlorophyll process, the comparison between WFD, OSPAR and salinity/turbidity-defined assessment areas shows similar outcomes for the Thames, with the assessment values not exceeding the thresholds in any coastal, salinity, plume defined or offshore areas (**Figure 7**). These non-exceedances represent the higher turbidity waters limiting phytoplankton growth, which can be seen in those lower values of biomass and abundances. These higher turbidity waters seem to extend further offshore into the larger coastal assessment areas. In Liverpool Bay there is also good comparability between results for the OSPAR and WFD chlorophyll assessments, apart from Morecambe Bay, which fails the WFD phytoplankton multimetric but not the OSPAR criteria. Interestingly, the salinity-defined coastal area (**Figure 9**) exceeds both the WFD and OSPAR criteria, suggesting that this larger assessment area would have a different outcome than the smaller WFD coastal areas. This may be due to decreasing turbidity in the salinity-defined coastal area, with increasing distance from the coast that permits greater phytoplankton growth and therefore elevated chlorophyll concentrations. In contrast, turbidity in the turbidity defined plume limits phytoplankton growth, and therefore chlorophyll concentrations are below the assessment threshold (Cole and Cloern, 1987; Painting et al., 2007).

The outcome for the dissolved oxygen assessment in transitional assessment areas shows that most of the transitional waters for Thames and Liverpool Bay do not fail the DO threshold. The exceptions are the Thames Middle transitional water and the salinity-defined transitional assessment area, which are similar areas and therefore have very similar results (**Figure 1**). The DO value for the salinity-defined transitional area in Liverpool Bay is lower than what is calculated for the individual WFD transitional waterbodies. This is because the salinity-defined transitional area is very coastal (**Figure 1**) and does not necessarily encompass the whole WFD transitional waterbodies and therefore only includes the very coastal subset of the WFD transitional data. The outcomes for the dissolved oxygen metrics are similar across the WFD and OSPAR criteria and are also similar across all the different assessment areas. The DO values calculated by the 5th percentile do change, with lower values in the transitional waters, increasing in all the coastal and offshore waterbodies.

To date in the UK, assessments made under WFD and OSPAR have not been integrated, and exceedances under the OSPAR assessment have not fed back into the program of measures under the WFD. However, assessments under the MSFD include WFD coastal waters and use some of the WFD tools for part of the assessment of Good Environmental Status, applied at a broader scale than an individual WFD water body. Therefore, in the future the WFD will provide support for the achievement of Good Environmental Status in marine waters. The thresholds used in the WFD and OSPAR assessments were developed at a time of limited data (prior to 2007). Given the advances in remote sensing, modeling and improved data collection, it is timely to consider testing of these thresholds to ensure they are still appropriate, such as is happening under the project JMP-EUNOSAT (Joint Monitoring Programme of the Eutrophication of the North Sea with Satellite data, 2017–2019). The UK chlorophyll threshold values were derived from the background nutrient concentration, assuming a constant Redfield C:N ratio and set C:Chlorophyll values (Foden et al., 2011). Recent research has shown seasonal variability in the uptake of carbon and phosphorous in shelf seas, which deviates from Redfield (Davis et al., in press; Poulton et al., in press). Therefore, reviewing the current chlorophyll thresholds in light of this recent research would be appropriate.

#### Additional Assessment Areas and Tools

The WFD transitional and coastal waterbodies are useful to flag issues at the fine spatial scale, and the smaller scale of these assessments is required to direct back to programs of measure and river basin management plans. Temporal representativeness is low in some cases, but this is expected given that they represent very small spatial areas. In the Thames estuary, the turbidity-defined plume assessment area is a large reporting area that represents the area of both the salinity-derived transitional and coastal assessment areas, and therefore represents one of the only fully integrated "catchment to coast" assessment areas. In Liverpool Bay, however, the plume assessment area includes the salinity transitional and only a fraction of the salinity coastal assessment areas. The additional process that reports across the salinity gradient, and fully encompasses the riverine influence of these large UK catchments, could provide an important component in understanding the links between river basins, land use information and downstream impacts. The smaller scale WFD waterbodies have been successfully used in identifying direct source impact, with the larger scale COMP2 areas providing a full assessment of any potential offshore issues. Both these processes could be improved by the addition of a more intermediate assessment area, providing information on the direct and diffuse nutrient loads and the full extent of these riverine loads and potential impact. In this paper we used turbidity to define such an intermediate assessment area. The Forel Ule (FU) scale is a color comparator scale used to classify the color of the oceans, regional seas and coastal waters. Remotesensing algorithms have been developed to classify water bodies from satellite imagery (Wernand et al., 2013), and the FU scale could be used as a standardized way of defining the extent of riverine influence. In addition, ecosystem models can be used to track the extent of riverine influence on the marine environment (e.g., Figure 4 in Painting et al., 2013).

The outcomes of the lifeform tool (Tett et al., 2008; McQuatters-Gollop et al., 2019) show that the phytoplankton communities in many of the assessment areas have changed in relation to the diatom and dinoflagellate counts. Whilst many of the assessment areas show a change in the PI value, signifying a change between the reference period (2013–2015) and the assessment period (2006–2012), there is not a statistically significant linear trend over the assessment period. However, the counts of diatoms and dinoflagellates in the assessment areas show an increasing number of dinoflagellates in the last few years (see **Figure 8**). Whilst it is difficult to ascertain if a change in plankton community is a negative one, it does demonstrate that the more established primary indicators that measure biomass and abundances may not be measuring these more complex changes in the phytoplankton community. Future assessment of phytoplankton metrics could include these type of measures to improve our understanding of community changes as well as the more traditional eutrophication indicators.

A protocol is being developed by the wider research community to use PI values to identify key stressors (McQuatters-Gollop et al., 2019). Low PI values observed here, indicating changes in plankton community in terms of the diatom and dinoflagellate lifeforms, appear to be driven in Liverpool Bay by increasing counts of dinoflagellates, especially during winter months, and in the Thames by increasing summertime dinoflagellate counts. These changes appear consistent across the different typologies we explore, with low PI values calculated in all areas with sufficient data. The drivers of community change require further investigation, but likely include changing phytoplankton counts vs. biomass and cumulative pressures including temperature and eutrophication.

### CONCLUSION

The outcomes of this study show that there have been significant reductions in loads to these two large UK catchments for some nutrients, particularly for phosphorus. However, the management of phosphorous has been more successful than management of diffuse nitrogen loads, reflected in the increasing nutrient ratios and a common problem facing many coastal waters in the UK, Europe and internationally. The range of tools or metrics available to assess the impact of these changing nutrient loads are explored in this paper and show a great degree of consistency in different approaches across the WFD and OSPAR process. Many criteria still show similar outcomes despite slight differences in aggregation. The testing of more ecologically appropriate areas, as defined by

### REFERENCES


salinity or riverine plume influence also show a degree of consistency when applying the different assessment metrics. The use of riverine plume and salinity-derived areas do show that the coastal issues of high nutrients and elevated phytoplankton biomass can extend beyond the narrow WFD coastal areas and could be a useful approach for future assessments when looking across the full salinity continuum. Additionally, the use of the phytoplankton lifeform tool, whilst not a typical approach in the eutrophication assessment process, highlights the importance of understanding community change in relation to the long-term nutrient shifts and should also be considered as part of a future eutrophication assessment process.

#### AUTHOR CONTRIBUTIONS

NG, MD, MB, LF, CG, AM, SvL and JB drafted the manuscript. NG, LF, CG, and AM analyzed the data. JB fitted the GAM models, and SvL processed the riverine data from raw archives.

## FUNDING

The work was funded through Defra contract SLA25.

### ACKNOWLEDGMENTS

The authors would like to acknowledge all scientific and technical staff at Cefas, the EA, The National River Flow Archive (CEH) and Afbi for data collection and processing.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00116/full#supplementary-material


online at: https://www.ospar.org/work-areas/hasec/eutrophication/commonprocedure. iv + 201 pp.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Greenwood, Devlin, Best, Fronkova, Graves, Milligan, Barry and van Leeuwen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# The Effect of Optical Properties on Secchi Depth and Implications for Eutrophication Management

#### E. Therese Harvey 1,2 \*, Jakob Walve<sup>1</sup> , Agneta Andersson3,4, Bengt Karlson<sup>5</sup> and Susanne Kratzer <sup>1</sup>

<sup>1</sup> Department of Ecology, Environment and Plant Sciences, Stockholm University, Stockholm, Sweden, <sup>2</sup> NIVA Denmark Water Research, Copenhagen, Denmark, <sup>3</sup> Department of Ecology and Environmental Science, Umeå University, Umeå, Sweden, <sup>4</sup> Umeå Marine Sciences Centre, Umeå University, Umeå, Sweden, <sup>5</sup> Oceanographic Unit, Swedish Hydrological and Meteorological Institute (SMHI), Västra Frölunda, Sweden

#### Edited by:

Katherine Richardson, University of Copenhagen, Denmark

#### Reviewed by:

Kemal Can Bizsel, Dokuz Eylül University, Turkey Matthias Obst, University of Gothenburg, Sweden

> \*Correspondence: E. Therese Harvey Therese.Harvey@niva-dk.dk

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 September 2018 Accepted: 11 December 2018 Published: 10 January 2019

#### Citation:

Harvey ET, Walve J, Andersson A, Karlson B and Kratzer S (2019) The Effect of Optical Properties on Secchi Depth and Implications for Eutrophication Management. Front. Mar. Sci. 5:496. doi: 10.3389/fmars.2018.00496 Successful management of coastal environments requires reliable monitoring methods and indicators. Besides Chlorophyll-a concentration (Chl-a), water transparency measured as Secchi Depth (ZSD) is widely used in Baltic Sea management for water quality assessment as eutrophication indicator. However, in many coastal waters not only phytoplankton but also colored dissolved organic matter (CDOM) and suspended particulate matter (SPM) influence the under-water light field, and therefore the ZSD. In this study all three main optical variables (CDOM, Chl-a, and SPM [organic and inorganic]) as well as ZSD were measured in three Swedish regions: the Bothnian Sea, the Baltic Proper, and the Skagerrak in 2010–2014. Regional multiple regressions with Chl-a, CDOM, and inorganic SPM as predictors explained the variations in ZSD well (R<sup>2</sup> adj = 0.53–0.84). Commonality analyses of the regressions indicated considerable differences between regions regarding the contribution of each factor to the variance, R<sup>2</sup> adj, in ZSD. CDOM explained most of the variance in the Bothnian Sea and the Skagerrak; in general, Chl-a contributed only modestly to the ZSD variance. In the Baltic Proper the largest contribution was from the interaction of all three variables. As expected, the link between Chl-a and ZSD was much weaker in the Bothnian Sea with high CDOM absorption and SPM concentration. When applying the Swedish EU Water Framework Directive threshold for Good/Moderate Chl-a status in the models it was shown that ZSD is neither a sufficient indicator for eutrophication, nor for changes in Chl-a. Natural coastal gradients in CDOM and SPM influence the reference conditions for ZSD and other eutrophication indicators, such as the depth distribution of macro-algae. Hence, setting targets for these indicators based on reference Chl-a concentrations and simple Chl-a to ZSD relationships might in some cases be inappropriate and misleading due to overestimation of water transparency under natural conditions.

Keywords: secchi depth, monitoring, management, eutrophication, CDOM, SPM, Chl-a, EU directives

## INTRODUCTION

Human activities have increased the transport of nutrients and organic matter from land to coastal waters, often resulting in eutrophication (Nixon, 1995). Eutrophication increases phytoplankton biomass and the chlorophyll-a concentration (Chl-a) decreases light availability for benthic vegetation (Nixon, 1995), and—if severe—can deplete oxygen near the sea bottoms with fatal effects on benthic fauna (Diaz and Rosenberg, 2008). In the Baltic Sea, eutrophication is one of the main challenges for good water quality (HELCOM, 2009; Jutterström et al., 2014) and several international agreements and programs are in place in order to mitigate the negative effects. The EU Water Framework Directive (WFD) focuses on the coastal zone, while the EU Marine Strategy Framework Directive (MSFD) (European Commission, 2000, 2008) as well as the Helsinki Commission's Baltic Sea Action Plan (HELCOM, 2007, 2009) focus mainly on targets for the open Baltic Sea basins. The coordinated policy actions taken so far within the Baltic Sea drainage basin have reduced some of the undesirable perturbations of eutrophication, but further improvements are still needed (Riemann et al., 2015; Andersen et al., 2017).

Eutrophication does not only change the nutrient conditions but may also alter the light environment and the light availability for photosynthetic primary production. Sunlight absorbed by phytoplankton is the prime energy source for pelagic food webs (Wozniak and Dera, 2007). Increased phytoplankton abundance increases the attenuation of light (Kd)—i.e., the gradual loss of light with depth—as more light is both absorbed and scattered in the visible wavelengths. However, in optically-complex waters, such as the Baltic Sea and many coastal areas, the light attenuation is also affected by riverine inputs of suspended particulate matter (SPM) (Kratzer and Tett, 2009; Gallegos et al., 2011; Aas et al., 2014; Capuzzo et al., 2015) and colored dissolved organic matter (CDOM) (Kowalczuk et al., 2005; Kratzer and Tett, 2009; Gallegos et al., 2011; Aas et al., 2014). In addition, erosion and resuspension of shallow coastal sediments add SPM to the water mass, to large extent as clay mineral (inorganic) particles (Blomqvist and Larsson, 1994). Depending on the relative proportions of organic and inorganic components, SPM interacts differently with light. Similarly to CDOM, organic SPM absorb light mostly in the blue wavelengths (Morel and Prieur, 1977) whereas inorganic SPM mostly scatters the light (Kirk, 2011).

K<sup>d</sup> is dependent both on absorption and scatter, but it is a nonlinear function of the present optical components (Kirk, 2011). According to Kirk, spectral K<sup>d</sup> (i.e., wavelength dependent Kd) can be estimated from spectral a (absorption) and b (scattering) in the following way:

$$K\_d = \mu\_0^{-1} \left[ \mathbf{a}^2 + (\mathbf{g}\_1 \mathbf{g}\_1 \ast \mu\_0 - \mathbf{g}\_2) \mathbf{a} \ast \mathbf{b} \right]^{0.5} \qquad \text{(Kirk, 2011)} \tag{1}$$

where µ<sup>o</sup> refers to the cosine of the refracted solar beam just below the surface (about 0.86 in the NW Baltic Sea, Alikas et al., 2015). The constants g<sup>1</sup> = 0.425 and g<sup>2</sup> = 0.19 were estimated by Kirk (2011). All combinations of optical in-water components that scatter or absorb light influence Kd, and K<sup>d</sup> is strongly inversely related to water transparency measured as Secchi depth (ZSD) (Jerlov, 1976; Preisendorfer, 1986; Wozniak and Dera, 2007; Siegel and Gerth, 2008; Kirk, 2011). ZSD is measured with a white disc that is lowered down the water column until the depth at which it is not visible anymore; this depth is noted as ZSD (Secchi, 1866; HELCOM, 2017). ZSD is a rough proxy for water transparency as it directly detects changes in the visible underwater light field (Preisendorfer, 1986). In coastal opticallycomplex waters the CDOM, SPM and phytoplankton biomass often co-vary, affecting K<sup>d</sup> (Morel and Prieur, 1977) and the observed ZSD readings simultaneously. Detailed descriptions of the theory for the relationship of ZSD to K<sup>d</sup> and optical properties are given by e.g., Preisendorfer (1986) and Kirk (2011), and recently by Aas et al. (2014) and Lee et al. (2015).

Reduced water transparency is regarded as a decrease in water quality (European Commission, 2000, 2008; HELCOM, 2007) and changes in the underwater light field can reduce both the primary production (Lyngsgaard et al., 2014) and the depth distribution of macrophytes (Orth et al., 2010).

Phytoplankton biomass is often estimated from the Chla concentration (Morel, 1980). Chl-a is therefore one of the commonly measured parameters to trace eutrophication within aquatic monitoring programs (HELCOM, 2007, 2017). ZSD is generally assumed to be inversely related to phytoplankton biomass and is used as an indirect eutrophication indicator (Karydis, 2009; Devlin et al., 2011; Fleming-Lehtinen, 2016) or even as a proxy for Chl-a (Boyce et al., 2010). Strong inverse relationships between Chl-a and ZSD (Lewis et al., 1988; Boyce et al., 2010) and K<sup>d</sup> (Smith and Baker, 1978) have been demonstrated for clear open sea waters.

Time series of ZSD have shown decreased transparency in the North Sea (Dupont and Aksnes, 2013; Capuzzo et al., 2015), the Baltic Sea (Sandén and Håkansson, 1996; Fleming-Lehtinen and Laamanen, 2012; Dupont and Aksnes, 2013), and the North Atlantic (Gallegos et al., 2011). This is mostly explained with increased phytoplankton biomass, but resuspension of sediments and the brownification of natural waters has also been discussed. In optically-complex coastal waters, however, the assumption of a direct inverse relationship between Chl-a and ZSD, i.e., that Chl-a solely determines ZSD, will neglect potential effects of spatial and temporal changes in SPM and CDOM.

There have been several local or regional investigations on how much the different optical components contribute to K<sup>d</sup> in coastal waters (Lund-Hansen, 2004; Kratzer and Tett, 2009; Aas et al., 2014; Murray et al., 2015) or lakes (Thrane et al., 2014; Watanabe et al., 2015). In Himmerfjärden bay in the Baltic Sea, included in this study, inorganic SPM had the strongest effect on the coastal spatial gradient in K<sup>d</sup> (Kratzer and Tett, 2009) although the K<sup>d</sup> (and thus also the ZSD) in the Baltic Sea is generally governed by CDOM absorption due to its dark, humicrich waters (Kowalczuk et al., 2006; Skoog et al., 2011; Gustafsson et al., 2014; Harvey et al., 2015a). However, knowledge about the relative contributions of the different optical parameters to variations in ZSD in different Baltic Sea coastal gradients is still limited. It is important for Baltic Sea management to know what the variations in ZSD depend on locally to enable a better-informed use of ZSD as a water quality parameter. Assumptions of what determines natural ZSD gradients from inner to outer coastal areas will influence the expected reference conditions within the WFD for e.g., the depth distribution of benthic vegetation. Since there are coastal gradients also for CDOM and SPM, simple general Chl-a to ZSD relationships will tend to overestimate the influence of Chl-a on ZSD. Reference values for ZSD for inner coastal areas estimated from such relationships in combination with modeled Chl-a reference values can therefore over-estimate the reference values for ZSD.

This study aims to examine how much Chl-a, inorganic SPM and CDOM contribute to the variations in ZSD in coastal gradients of three different regions with variable concentrations of optical parameters: the Bothnian Sea (BS) in the northern Baltic Sea with very high CDOM absorption, the Baltic Proper (BP) with high CDOM and high SPM near the coast and in the Skagerrak (SK) with lower CDOM absorption. We hypothesized that the direct link between Chl-a and ZSD was weaker in areas of high CDOM absorption and SPM scatter. Furthermore, we investigated how much CDOM and inorganic SPM affect the predictions of changes in ZSD by empirical models, when the Chl-a reference values and the respective level for Good/Moderate (G/M) status is applied.

## DATA SOURCES AND METHODS

#### Areas of Investigation

The Baltic Sea is a semi-enclosed brackish sea, connecting to the North Sea through the Kattegat and the Skagerrak, with several basins separated by sills and shallow areas. In the north, the Bothnian basins have high run-off from land and very low surface salinity (2–6). In the central parts of the BP the salinity is around 7. At the Swedish West coast, the influence of the North Sea is more pronounced, resulting in almost marine conditions in the SK and the Kattegat (salinity 18–26) (Voipio, 1981; Leppäranta and Myrberg, 2009; Deutsch et al., 2012), and a clear influence by tidal action. In the northern basins, there is a strong gradient in CDOM absorption due to the restricted water exchange and a relatively large run-off, discoloring the water distinctly brownish (Jerlov, 1976; Kirk, 2011; Skoog et al., 2011).

We collected samples along several near-shore to off-shore gradients in the BS, the BP and the SK in order to cover representative ranges of values in each sub-area within the regions. The northern-most water samples were collected at 18 locations in a coastal gradient in the Öre Estuary in the western BS, sampled from May to early September 2010 (**Figure 1**, **Table 1**). The Öre Estuary receives a large water inflow from the Öre River, especially during spring (up to 290 m3 s −1 ), transporting nutrients, dissolved, and particulate matter originating from the mountains and surrounding bog areas (Harvey et al., 2015a).

Data from the BP are from seven gradients reaching inner coastal waters to the open sea (**Figure 1**). A total of 28 locations were sampled from June to early September 2010 to 2014 (**Table 1**). The Östhammar gradient (sub-area BP.1) is situated in the archipelago of Uppsala County, north of Stockholm, in southern BS but in proximity to the BP and was still included in the BP data set due to the low freshwater inflow compared to the northern part of the BS. There is only a weak salinity gradient in the BP.1 subarea, and it is the receiver of a Waste Water Treatment Plant (WWTP). The inner coastal area is eutrophicated with relatively high Chl-a concentrations. Himmerfjärden bay (BP.2), situated in the southern Stockholm archipelago, is a large fjord-like bay consisting of several basins with low freshwater input and restricted water exchange. The inner Himmerfjärden bay is a receiver of a regional large WWTP and the area shows symptoms of eutrophication, with increased phytoplankton biomass and decreased ZSD (Engqvist, 1996; Savage et al., 2002). Gälöfjärden (BP.3) is a small, shallow bay southwest of Himmerfjärden bay with relatively high resuspension of sediments. Nyköping bay (BP.4), situated further south of Stockholm and Himmerfjärden bay, is a shallow estuary that receives the freshwater inputs of three rivers (**Figure 1**). The city of Nyköping and its WWTP are located in the inner part of the estuary. The large fresh water inflow and the restricted water exchange in this shallow area, create strong gradients in salinity, nutrients, Chl-a concentration and ZSD. Bråviken bay (BP.5) is deep (ca 30 m) in its central and outer parts, but its inner-most part is rather shallow, with a high freshwater inflow from the river Motala Ström, entering through the city of Norrköping. Slätbaken (BP.6) is a relatively deep bay with sills restricting the water exchange and with a nutrient-rich freshwater inflow that seems to cause the relatively high Chl-a levels. Two locations situated south of Slätbaken, close to one-another, were also included in the study [Kaggebofjärden and Lindödjupet (subarea BP.7)].

In the SK region, 15 locations were sampled in June to August 2012 to 2013 in an elongated fjord system between the islands Orust and Tjörn and the mainland (**Figure 1**, **Table 1**). The innermost sub-area Byfjorden (SK.1) is situated at the head of the fjord system with a large freshwater input from the river Bäveån. The discharge of the WWTP of the city of Uddevalla is located near the river mouth. Byfjorden is eutrophicated, with high levels of nitrogen and Chl-a and low ZSD. Further sub-areas are Havstensfjorden (SK.2), Askeröfjorden (SK.3), and Hakefjorden (SK.4). Hakefjorden receives freshwater from the river Göta Älv, causing (similar as in Byfjorden) a slightly lower salinity (of ∼18) than in the other sub-areas (salinities ∼20). Marstrandsfjorden (SK.5) is directly connected to the open sea and has the lowest Chl-a concentrations.

## In situ Data of Secchi Depth, Chl-a, CDOM, and SPM

At each sampling station all optical water quality parameters (ZSD, Chl-a, CDOM, SPM) were measured or sampled. The water transparency (i.e., ZSD) was measured with a white Secchi disc (25 cm in BS and most BP areas and 30 cm in diameter in SK and for some occasions in BP.2), taken on the shady side of the ship in order to avoid the influence of sun reflection on the viewer's perception of the ZSD. A water telescope was used in the BP for ZSD measurements in order to reduce the sun glint (Werdell, 2010; HELCOM, 2017). Water samples for measuring Chl-a, CDOM, and SPM were collected just below the surface with a


special sampling bucket or a Ruttner sampler. The laboratory analyses were carried out according to established protocols or ISO- standards; Chl-a in BS and SK by HELCOM (2017) and in the BP according to Jeffrey and Vesk (1997) or HELCOM (2017), CDOM by Kirk (2011). In SK and BS samples were extracted with ethanol followed by fluorometry, in the BP extracted with ethanol or acetone followed by spectrophotometry. The analyses were conducted by both accredited monitoring labs (Jeffrey and Vesk, 1997; Werdell, 2010) as well as by a specialized biooptics research group following established ESA MERIS protocol. The slight differences in the methods were shown to have little influence on the results, and extensive tests by the Marine Ecology Laboratory at Stockholm University prior changes from acetone to ethanol in the Chl-a extraction in various monitoring programmes, showed no difference between these methods. Also, a recent evaluation of the CDOM filtration method using glass vs. plastic filtration gear did not show any significant differences. SPM (inorganic fraction, SPIM and organic fraction, SPOM) was determined gravimetrically (Strickland and Parsons, 1972) with one replicate in the BS and the SK, three replicates per station in most cases in the BP. A more detailed description of the methods for the optical data and data collection for this study can be found in Kratzer et al. (2003), Kratzer and Tett (2009), Harvey et al. (2015a,b), and Kari et al. (2017). In Aas et al. (2014) a detailed evaluation of error sources for ZSD measurements are given. It was found that wind-wave effects caused most errors by stirring the surface and increasing the sun glint. The use of a water telescope increased the ZSD with 10–20% (Aas et al., 2014), whilst the wind effect at the surface caused a decrease in the same order (11%) (Sandén and Håkansson, 1996). The difference of

using a disc with 10 rather than 30 cm in diameter reduced the average ZSD by 10–20% (Aas et al., 2014). However, a recent intercomparison between Umeå and Stockholm Universities found no significant difference when comparing the use of Secchi disks of 25 vs. 30 cm diameter. Also, the data from different regions were here analyzed separately so the results are comparable within each respective region. For this study the overall relative error for ZSD was calculated from the coefficient of variation to below 2.6%. The error for the trichromatic Chl-a analysis is within 7– 10% (Kratzer, 2000; Sørensen et al., 2007), and the SPM method has an error of about 10–13% (Kratzer, 2000; Kari et al., 2017), dependent on the range of SPM values. The error for CDOM absorption is within 6% (Harvey et al., 2015a).

Data is provided via Stockholm University, Umeå University, and SMHI upon request. Most of the Chl-a data are available within the national monitoring programme and accessible via the Swedish Oceanographic Data Center at SMHI (the SHARK-database, http://sharkweb.smhi.se), other are hosted by Umeå University or the Marine Ecology Laboratory at the Department of Ecology, Environment and Plant Sciences, Stockholm University. For the CDOM and SPM data some are also available in SHARK. The ranges of values of optical properties for all regions are shown in tables, and these ranges are required for constraining e.g., regional radiative transfer models.

### Data Analysis

Correlations between ZSD and the predicting variables, i.e., CDOM, SPM (inorganic fraction, SPIM) and Chl-a, as well as between the different predictors were tested for possible collinearity for each region. The correlations were derived applying Pearson's correlation on ln-transformed data (where ln stands for natural logarithm).

Multiple general linear models (GLMs) for ZSD with Gaussian distribution family were used for the first data analysis step, treating "region," "season," and "sub-area" as categorical variables to estimate the slopes of possible spatial differences. Generally, only small amounts of SPIM originate from phytoplankton and thus, inorganic SPM (SPIM) is used here as a proxy for landderived and/or resuspended SPM. Organic SPM is assumed to be strongly linked to phytoplankton biomass and therefore already represented by Chl-a. Hence, in the model ZSD was response variable, and Chl-a, SPIM, and CDOM potential explanatory variables. In order to evaluate the performance of the models the modeled ZSD was evaluated against the measured ZSD, the Root Mean Square Error (RMSE; unit in meters), the Normalized Root Mean Square Error (NRMSE, unit in percentage) and the Mean Normalized Bias (MNB; unit in percentage) as well as 95% Confidence Intervals (CI) for the model coefficients. It should be noted that the error metrics for the models here only refer to the fit of each model to the existing regional data sets, but do not indicate how reproducible each model is. For this, an independent data set would be needed for regional model evaluation.

The second step was to apply commonality analysis based on the GLMs to reveal how much each predicting variable uniquely affects the variation in ZSD as well as the common contribution with the other variables (i.e., that cannot be separated due to collinearity) (Kraha et al., 2012; Dormann et al., 2013; Ray-Mukherjee et al., 2014). Commonality analysis is not widely used within ecological research, but is well-established within psychology, social sciences and education. A good description with ecological examples are given in Ray-Mukherjee et al. (2014) and the use of commonality analysis splits the adjusted coefficient of determination (R<sup>2</sup> adj ) into a unique and a common variance of the predicting variables (Chl-a, CDOM and SPIM) to the ZSD and thereby contributes to an improved understanding and interpretation of the different effects on ZSD. The analysis takes the often-neglected collinearity between the predicting variables into account and enables for follow-up analyses of the GLMs.

The GLMs were used to predict the potential increases in summer ZSD at decreased Chl-a levels, i.e., simulating an improved eutrophication status according to EU WFD and MSFD (for some outer locations). For each water body or region, the GLMs were used to calculate ZSD using Chl-a concentrations adjusted to the respective reference (pristine) and G/M boundary Chl-a values. The reference and G/M thresholds values for Chla were calculated according to the specified salinity-dependent equations or the defined thresholds were used according to the Swedish Agency for Marine and Water Management (SwAM, 2012, 2015). The modeled changes in ZSD were recalculated to deviation in % difference from the ZSD G/M threshold for each water body or region. A deviation of 0% equals the G/M ZSD threshold, a negative deviation indicates a lower ZSD, and that the ZSD threshold for G/M status has not been reached. A positive value indicates that the ZSD threshold has been exceeded, i.e., the ZSD is greater than the respective threshold value, thus indicating good water quality. The deviations from the G/M ZSD threshold were then compared to the observed ZSD for each sub-area within each region.

Assumptions of independence, normality and heteroscedasticity were tested, and all data were ln-transformed to achieve normal distribution. Residual analysis and model evaluations were performed for all statistical tests. The number of observations (n) is given for each analysis and the confidence level was set to 5%. For all graphs, statistical and data analyses R 3.0.1 was used (R Core Team, 2013).

## RESULTS

The highest average ZSD was found in the SK area (6.2 ± 1.7 m, mean ± one standard deviation), the lowest in the BS (3.7 ± 0.9 m), and slightly higher in the BP (3.8 ± 2.3 m). **Table 2** shows the ranges, means, medians, standard deviations (Stdev) and standard errors of the mean (SEM) for ZSD, Chl-a, CDOM and SPM (total SPM, and SPIM and SPOM fractions) for the three regions. The same data are presented as boxplots for all sub-areas in the **Figures S1–S3**.

## Data Selection and Empirical Models

A pooled GLM for ZSD for all regions had high predictive power (R<sup>2</sup> adj = 0.85, n = 406). However, this model was not representative for each individual region since stepwise GLMs showed significantly different model parameters between



The table shows the ranges, mean, standard error of the mean (SEM), median and standard deviation (Stdev.) for Secchi depth (ZSD), Chlorophyll-a concentration (Chl-a), colored dissolved organic matter (CDOM), suspended particulate matter (SPM), inorganic suspended particulate matter (SPIM), organic suspended particulate matter (SPOM), Salinity and Number of observations (n).

TABLE 3 | Results from the empirical multiple regression models per region, based on the data selection from the GLM's.


The table shows the coefficient of determination (R<sup>2</sup> adjusted), p-values for the intercepts and the coefficients for each predicting parameter. Non-significant parameters are denoted n.s. The number of observations (n) and the degree of freedom (df) for the models are also presented, as well as the root mean square error (RMSE), the normalized root mean square error (NRMSE) and the Mean Normalized Bias (MNB). The model for all regions, three main models and additional ones for the regions are presented.

regions (**Table 3**). Further analyses per region showed differences between seasons and between gradients. Hence, the selected models for further analyses were focused on the summer data, the assessment period for both Chl-a and ZSD. The BS model had moderately variable input data, except for ZSD (**Figures S1–S3**), giving a relatively low predictive power (R<sup>2</sup> adj = 0.54, n = 131) for ZSD. The SPIM component was not significant but was still included in the analysis (see motivation below). In the BP region the model chosen was based on data from all gradients (R<sup>2</sup> adj = 0.84, n = 85) except for Himmerfjärden (BP.2), as it differed significantly from the other areas (**Table 3**). The SK data were from five basins (sub-areas) along a gradient in TABLE 4 | Results of the commonality analysis, showing the percentage of contribution to the explained variations to the R<sup>2</sup> adj for the main ZSD models in the three regions.


a fjord system (**Figure 1**, **Table 1**) and the GLMs resulted in two models based on differences between sub-areas (**Table 3**). A common model for SK1, 2 and 5 (ModelS1) predicted the ZSD fairly well with a R<sup>2</sup> adj = 0.64 (n = 61), although the SPIM component was not significant. A common model for SK.3 and 4 (ModelS2) had a slightly lower R <sup>2</sup>adj =0.53 (n = 54) but a significant SPIM parameter. ModelS<sup>1</sup> was chosen for comparisons between regions since it had higher predicting power (i.e., R 2 adj) and included three out of the five sub-areas. The main models from each region (**Table 3**), were used for further detailed analysis and comparisons, and are from here on referred to in the text, if not stated differently. Since SPIM has strong scattering properties (Kratzer and Tett, 2009; Kirk, 2011; Kratzer and Moore, 2018), and there was a strong correlation between ZSD and SPIM where we had the largest SPIM gradient, we chose to include the non-significant SPIM parameters (**Tables 3**, **4**). All models had similar intercepts but different parameter coefficients.

#### Commonality Analysis

Commonality analysis showed that contribution of the various optical components to the explained ZSD variation (R<sup>2</sup> adj) was different amongst regions (**Table 4**, **Figure 2**). In the BS a large proportion was explained by CDOM alone (46%), together with the paired interaction of CDOM and SPIM (42%). The SPIM component had no unique effect and Chl-a alone explained only 6%, whilst CDOM and Chl-a combined contributed somewhat more (8%) to the variations in ZSD. The common interaction effect of all three variables in the BS was very low and even negative. In contrast, the common interaction effect of all three variables (∼53%) together with the interaction of Chl-a and SPIM (∼11%) explained most of the variation in ZSD in the BP. In the SK, CDOM alone explained more than two thirds (∼70%) of the variation in ZSD. SPIM contributed uniquely with < 1% and the shared, interactive commonalities were negative, but very close to zero. Moreover, Chl-a had the most pronounced unique effect on ZSD in the SK (9%).

## Correlation and Multiple Regression Analysis

Correlation analysis among the predicting variables and ZSD for each region (**Figure 3**, **Table 5**) were used to evaluate the GLMs and the commonality analyses, and to visualize the data. ZSD was inversely correlated to Chl-a in all regions, with the strongest correlation in the BP, moderate in the SK and the weakest in the BS (**Figures 3A,D,G**, **Table 5**). The relationship between ZSD and CDOM was strong and inverse in all regions (**Figures 3B,E,H**, **Table 5**). The correlation between ZSD and SPIM was inverse and strong in the BP, but weaker in the BS and absent for SK (**Figures 3C,F,I**, **Table 5**). Correlations between the predicting variables (i.e., the optical components) also differed among the regions (**Figure 4**, **Table 5**). There were strong positive correlations between Chl-a and CDOM and between Chl-a and SPIM in the BP, but these were absent or weak in the BS and the SK (**Figures 4A,B,D,E,G,H**, **Table 5**). The relationship between CDOM and SPIM was on the other hand positive and significant in the BS and the BP but not in the SK (**Figures 4C,F,I**, **Table 5**).

## Model Performances in Baltic Sea Regions

The main models predicted ZSD well with R<sup>2</sup> adj between 0.54 and 0.84 and with a rather high precision (RMSE 0.6–1.1 m, equivalent to a NRMSE of only 9–13%) and a very high accuracy with a low bias (MNB only 1.6–3%) (**Table 3**). The relative errors (RMSE) in this study were in the same range as e.g., found in the Oslo fjord in the Skagerrak (Aas et al., 2014). Outliers have a large influence on the error statistics (**Figure 5** and **Figure S4**). In the BS model some ZSD observations deviated from the model predictions both in the lower, <1.5 m and higher, >5 m ZSD range (**Figures 5A,B**). This explains the lower R<sup>2</sup> adj and the NRMSE of 13% of this model. The BP model performed better over the full range of ZSD, except for a few very high values (**Figures 5C,D**). Similarly, the SK model captured the full range of ZSD well, also for the highest values with a slightly higher dispersion in the lower ZSD range (**Figures 5E,F**). The graphs for the additional models are presented in the **Figure S4**, as all models were used for the respective sub-areas in the analysis of Chl-a influence on EU Directive ZSD targets. Overall, the biases for all models were low (slightly negative), indicating a high precision of the estimated ZSD from the models, with an insignificant underestimation of the predicted ZSD, as all MNB were <4% (**Table 3**). The Confidence Intervals (CI) of the model intercepts and coefficients were calculated using a two-sided 95% CI and are presented in **Figure 6** and **Table 6**. The Chl-a coefficients showed the smallest difference among regions and the narrowest CI, indicating a high precision of the estimated coefficients. The SPIM coefficients also had a relatively small range of CI while the range was larger for the CDOM coefficients. The intercepts were all very close to each other, with smaller CI for the main models. Generally, the coefficients and intercepts for SK and BS were closer to each other than to those for BP.

FIGURE 3 | Correlation plots between Secchi depth, ZSD (m) and the predicting parameters; Chl-a [µg l−<sup>1</sup> ] (left), CDOM (m−<sup>1</sup> ) (middle), and SPIM [g m−<sup>3</sup> ] (right). Graphs (A–C) show data from the Bothnian Sea (BS), (D–F) from the Baltic Proper (BP) and (G–I) from the Skagerrak (SK). The Pearson correlation coefficient, r, is shown for each data-set, where \*, \*\*\* denotes significant correlations at 0.01 and 0.0001 levels and no star indicates non-significance. All data have been natural log-transformed.

TABLE 5 | Table showing the Pearson correlation coefficient, r, for each data set, where \*, \*\*\* denotes significant correlations at 0.01 and 0.0001 levels and no star indicates n.s.


All data have been natural log-transformed. The strongest correlations are marked in gray, the modest correlations in light gray and no correlation are presented without marking.

## Chl-a Influence on the EU Directive Secchi Depth Targets

For each sub-area within the regions, the deviations from the G/M ZSD threshold were calculated for observed and for modeled ZSD-values (by the derived empirical GMLs). The modeled changes in ZSD were obtained by adjusting the Chl-a levels in the models to (1) G/M value of Chl-a and (2) to the Chl-a reference value (**Figure 7**). In the BP all observed ZSD values were below the G/M ZSD threshold, except one off-shore station in BP.1. All modeled ZSD values increased but only a few reached or exceeded the ZSD threshold for good environmental status. Even though the Chl-a concentrations were lowered considerably (compared

to the observed values), many of the modeled ZSD values in the sub-areas were still more than 50% below the ZSD threshold. In the SK the measured ZSD were generally above or near the G/M ZSD threshold, indicating good environmental status for ZSD. For simulated reference Chl-a concentrations there was some increase of ZSD for the sub-areas SK.3 & 4 (modelS2), with very few observations below the threshold (**Figure 7**). However, for the sub-areas SK.1–2 & 5 (modelS1), the modeled ZSD for reference and G/M Chl-a decreased to near or even below the ZSD threshold (**Figure 7**). This is explained by the low observed Chl-a values, many already below the Chl-a G/M threshold, and that the G/M Chl-a values used in the models were higher. In the BS most of the observed ZSD values were above the G/M ZSD threshold and the distance increased even more for the ZSD modeled from reference Chl-a values (**Figure 7**). The median of the modeled ZSD values with the G/M threshold for Chl-a was not improved.

### DISCUSSION

Empirical models were built to estimate the ZSD, based on the optical components Chl-a, CDOM and SPIM. The results show that in the studied regions the components contribute differently to the variations in the ZSD. Our results demonstrate that contributions from all optical variables to the variation in ZSD affect the possibility of reaching the current G/M threshold of ZSD as defined by the WFD and the MSFD.

## Different Empirical Models Indicate Different Optical Conditions

The empirical models and commonality analyses reveal a variability and inconsistency in the collinearity between the predicting variables among the studied coastal regions, indicating complex optical conditions caused by different proportions of the three optical components. For example, a large difference in the proportion of inorganic matter, which is strongly scattering, to organic matter, which is highly absorbing, will have a great effect on K<sup>d</sup> as it is a non-linear function of both absorption and scatter (Kirk, 2011) (Equation 1). In the BS the SPIM coefficient in the main model was not significant but the commonality analysis showed that together with CDOM, SPIM still had a strong collinear influence on the variation in ZSD. A strong unique effect was seen for CDOM (∼46%). Riverine CDOM loads are generally much higher in the BS than in the BP (Harvey et al., 2015a) and the SK (**Figure S2**). In the BS gradient there were very high median values of CDOM, even during the summer (1.7 m−<sup>1</sup> ) but the SPIM was relatively low (0.6 g m−<sup>3</sup> ) (**Table 2**). The high background CDOM absorption leads to a relatively low reflectance and makes the water appear rather dark. Thus, a comparatively small increase in SPIM scattering will have a relatively large influence on the reflectance and, thus on the ZSD. The commonality analysis showed a rather interesting result for the SK, here the variation in the ZSD was even more driven by CDOM alone (70%) than in the BS (**Figure 2**, **Table 4**) even though the levels of CDOM were much lower (0.4 on average), and the SPIM values relatively high (about 3.2 g m−<sup>3</sup> on average), which indicates that overall, the light attenuation should be dominated by SPIM scatter (Equation 1). However, the total effect of SPIM (unique and common effects) on ZSD variation was much less in the main model for SK (< 1%) than in the and BP, where SPIM contributed 40 and 85%, respectively. There is an important distinction between which optical component that dominates the light attenuation—e.g., scattering from SPIM in the SK—and which component that drives the variability in light attenuation. The commonality analysis explains the contribution of the optical components to the variability of ZSD and is completely dependent on the covariation and concentrations of the optical components. The difference between regions may partly be due to the effect of tidal action at the west coast, which is hardly detectable in the other two regions. The tidal range on the Swedish SK coast is rather low, ca. 30 cm, compared to many other seas, but higher than in the Baltic Sea, where it is only in the range of a few centimeters. Tidal action may explain a relatively high background of SPIM in the SK and influence the CDOM gradient. During high tides, open sea waters currents

and Intercepts for each region (BS; BP and SK). Triangles indicate the main models and filled circles the additional. The bar lines show the 95% CI of each estimated coefficient.

TABLE 6 | Table showing the 95 % Confidence Interval (CI) of the intercepts and coefficients for the models.


FIGURE 7 | Boxplots of the % differences to the Good/Moderate (G/M) thresholds of Secchi depth, ZSD for the respective water bodies within the WFD or the MSFD. The solid line denotes the G/M level for ZSD; the upper dashed green line 50% above; and the lower dashed red line 50% below the ZSD threshold. The boxes for the different sub-areas represent observed ZSD (Obs), modeled ZSD with Chl-a concentrations set to the respective reference value (ref) and with Chl-a concentration set to the G/M value for Chl-a. The horizontal lines in the boxplots are the median values, horizontal edges the 25th and 75th percentiles, the whiskers indicate the min and max observations within 10th and 90th percentiles, and open circles represent outliers.

move into the fjords at the west coast, markedly decreasing CDOM concentrations (unpublished data from Gullmars fjord, measured by S. Kratzer), whereas during low tides the water runs in the reverse direction, i.e., from the inner fjords to the outer sea, which markedly increases the CDOM concentrations. The results of the commonality analysis can to a large extent be explained by the collinearity between different parameters in the regions, as reflected in the correlation analysis. For example, the correlation between ZSD and SPIM was low for the main SK model, as reflected in the low SPIM coefficient in the model. A weak correlation (r = 0.27) was found between the Chl-a and CDOM, which was also reflected in the shared commonality between the two parameters.

Another explanation for low influence of SPIM is that the commonality analysis is based on multiple regressions (assuming a linear relationship between the parameters) and is thus dependent on the model's accuracy to predict ZSD. The coefficients of determination for the models were moderate both for the SK (R <sup>2</sup> = 0.64) and the BS (R <sup>2</sup> = 0.54) and the commonality analyses only explain ∼50–65% of the variation in ZSD. It should therefore not be used as a tool to explain the (optical) physics behind ZSD variability but is valuable for a rough estimation on which parameters may drive the variability. **Figures 3**, **4**, however, show that in the BP there are very strong correlations between ZSD and all three main optical components. The results of the commonality analysis also showed that SPIM in BP was here the largest single contributor to ZSD variability (**Table 4**). The largest effect on the variation in the BP, however, was from the shared commonality of all three components (∼53%), which was clearly seen in their strong negative correlation with ZSD (**Figure 3**) and the strong positive correlation (indicating strong collinearity) among them (**Figure 4**). This differed from the other regions that had very low values for the three fold-shared commonalities. The largest total effect of Chl-a (∼75%) was found in the BP region, even though the unique effect was slightly stronger in the SK region. However, in the SK there was also a substantial effect shared with CDOM (20%). When evaluating the results of commonality analysis, it is important to stress that the analysis describes the unique separate and combined statistical effects of each variable to the explained variation in the R<sup>2</sup> adj. Consequently, the same percentage value for a unique contributor in a certain region can here mean different actual contribution to the variation in ZSD (Ray-Mukherjee et al., 2014). For example, will the 1 % CDOM contribution from the commonality model in BP correspond to a CDOM contribution of 8.4% of R<sup>2</sup> adj variation CDOM, whereas the 1 % SPIM contribution in the SK model corresponds to 6.4% explanation of R<sup>2</sup> adj.

In summary, based on the commonality analyses the variations in ZSD were overall mostly governed by CDOM, both uniquely and together with SPIM in the BS, by all three parameters jointly and uniquely by SPIM in the BP, and by CDOM, uniquely and together with Chl-a, in the SK. The relative contribution of the inherent optical components to K<sup>d</sup> for a large lake dataset from Norway and Sweden was evaluated in Thrane et al. (2014), where also CDOM was found to be the dominant component, followed by Chl-a and non-algal particles, water excluded. Kratzer and Tett (2009) found that Kd490 was predominantly determined by CDOM absorption, both in the open and coastal Baltic Sea. In coastal areas the effect of SPM scatter also had a strong optical influence (decreasing with distance to the shore), while Chl-a had a relatively small optical influence, both in the open sea and coastal areas.

#### Factors Influencing the Optical Variables

Some of the differences among regions seen in the commonality analyses models may be explained by the differences in the inherent optical properties among CDOM, SPIM, and Chl-a jointly influencing the under-water light conditions and modulated by changes in physical and hydrological schemes. The strong correlations observed between CDOM and SPIM are most likely caused by a strong influence of terrestrial run-off, leading to a large variation in both CDOM and SPIM, even though run-off is lower in summer than in spring. However, the same pattern was not seen in SK, where there on the other hand is a pronounced tidal effect. Dupont and Aksnes (2013) showed that also the distance from shore and the bottom depth are significant factors affecting ZSD, both in the Baltic Sea and the North Sea. In Danish waters, suspended material and Chl-a was found to be negatively correlated to ZSD, with a pronounced effect on the variation in ZSD (Nielsen et al., 2002) similar to our results. Resuspension or erosion of sediments (i.e., increase of SPIM) have been shown to affect water optics in other seas as shown by Otto (1966), Devlin et al. (2008), and Capuzzo et al. (2015), as well as for the BP (Kratzer and Tett, 2009) and thus may also be pronounced in many BS coastal areas less dominated by freshwater run-off than in this study. Even though the SPIM was relatively high in the SK and seemed influenced by tidal dynamics that may lead to a relatively strong resuspension of inorganic sediments it showed an unexpectedly small effect on ZSD variation.

Nutrient input has also been found to be negatively correlated to the ZSD, by increasing Chl-a levels (Nielsen et al., 2002; Fleming-Lehtinen and Laamanen, 2012; Aas et al., 2014). Nutrient inputs are often increased by high run-off that also introduces more allochthonous CDOM and SPM into the ecosystem, as is reflected in the correlations between these variables in this study and in coastal gradients by Kratzer and Tett (2009) and Capuzzo et al. (2015). Phytoplankton can be a substantial fraction of total SPM, which means that Chl-a are often positively correlated with SPOM (e.g., Nielsen et al., 2002). But Chl-a may also correlate with SPIM, e.g., due to the silica in diatom blooms or through run-off, increasing both SPM and nutrients that increase Chl-a. Assuming a phytoplankton carbon to Chl-a ratio of 40 (Sathyendranath et al., 2009), a carbon content of 40% and (a high) ash content of 20% (∼25–40% in diatoms) of dry weight (Whyte, 1987), the range of Chl-a from 1 to 52 µg/l (median 5 µg/l) in the BP should correspond to SPIM concentrations of 0.02 to 1 mg/l (median 0.1), about a magnitude lower than the observed range of SPIM (0.1 to 16 mg/l, median 2.1). The SPIM found in algal detritus should also contribute, possibly in the same concentration range. This means that, overall the cellular contents of inorganics is relatively small in algal cells compared to terrestrial inorganic matter originating from freshwater or coastal erosion.

Kratzer and Moore (2018) investigated the scattering and absorption properties of the Baltic Sea in comparison to other seas and oceans. They found that besides an optical dominance of CDOM absorption, there was also a clear indication of different optical water types—open sea vs. coastal waters. Thus, in optical models different parameterization may have to be sought for these different water types. This should also be considered when investigating the effect of all three main optical components on ZSD. Inner coastal waters are often dominated by SPIM scattering (Kratzer and Tett, 2009; Kari et al., 2018), which will also clearly affect the ZSD. Open sea waters are more dominated by CDOM absorption- and during times of phytoplankton blooms also by phytoplankton absorption and scatter. The proportion of inorganic to total SPM decreases when moving from inner coastal to more open sea waters (Kratzer and Tett, 2009; Kari et al., 2018). It is therefore recommendable that current monitoring programs also include measurements of SPIM in the coastal zone. In the open sea, total SPM or turbidity measurements can be used to indicate phytoplankton or cyanobacteria blooms, as the organic fraction usually falls out very close to the coast (Kratzer and Tett, 2009; Kari et al., 2018). In all regions CDOM correlated quite strongly with ZSD (r ranging between −0.71 and −0.77) (**Table 5**), which makes CDOM a very important optical variable to measure in all areas.

Climatological models for the Baltic Sea predict more land run-off due to an increase in precipitation, especially in the northern Scandinavian regions (Meier et al., 2012a,b). This might lead to a brownification due to an increase in humic matter and suspended organic material to the sea (Larsen et al., 2011; Meier et al., 2012a). Although the relationships between K<sup>d</sup> and ZSD are variable in the Baltic Sea, stronger light absorption due to climate change would affect the light climate and the conditions for primary production by an increase in K<sup>d</sup> and thus a decrease in ZSD (Kowalczuk et al., 2005; Kratzer and Tett, 2009; Harvey et al., 2015a). Hence, the predicted brownification due to an increase in precipitation in northern Scandinavia may have a dominant effect over a possible decrease in Chl-a concentrations due to recent and future reductions in nutrient loads by effective management programs, maybe resulting in unchanged ZSD.

## Secchi Depth and Implications for Eutrophication Management

Our results show that using ZSD as a direct eutrophication indicator may be misleading, as the empirical models, the commonality analyses and the correlations all show that ZSD is more strongly related to CDOM and SPIM than to Chla, or that these variables are inter-correlated in the coastal gradients. Hence, the common assumption that a strong direct inverse relationship with Chl-a makes ZSD a suitable water quality indicator has important limitations in the Baltic Sea. Some anthropogenic measures to reduce nutrient load from runoff, e.g., changes in land use or wetland area, are also likely to affect the input of CDOM and SPM to coastal waters and might mitigate climate change effects. However, natural gradients of CDOM and SPM from the coast to the open sea (Kratzer and Tett, 2009; Harvey et al., 2015a) will likely have a main influence on the response of the ZSD to anthropogenic changes. Even more important is that those relationships vary with region, sub-areas and season, meaning that the same ZSD value may indicate quite different environmental and optical conditions, influencing the management efforts needed for certain ZSD improvements. In different optical water types (e.g., clear ocean vs. coastal waters) the same change in ZSD expressed in meters can imply very different ranges and combinations of the optical components (Kirk, 2011). Due to the logarithmic attenuation of light in the water, a decrease in ZSD from 2 to 1 m is related to much larger changes in concentrations and absorption by optical components than a ZSD decrease from 10 to 9 m (Dupont and Aksnes, 2013), implying very different management efforts and costs.

Long-term decreases in ZSD have been observed in both the open Baltic Sea and the North Sea. In the open Baltic Sea, increased Chl-a have been linked, to some extent, to a decrease in ZSD (Sandén and Håkansson, 1996; Fleming-Lehtinen and Laamanen, 2012; Dupont and Aksnes, 2013), but the potential importance of CDOM and SPIM was not evaluated in these studies. ZSD is commonly used as indicator of eutrophication (HELCOM, 2007; European Commission, 2008, 2010; Karydis, 2009; Fleming-Lehtinen and Laamanen, 2012). In the WFD and the MSFD, targets and thresholds for both ZSD and Chl-a are used to classify if water bodies meet the goals of good eutrophication status (SwAM, 2012, 2015), influencing management action plans. The long-term time-series of ZSD in open sea areas and the Chl-a to ZSD relationships have been used to establish reference chlorophyll values (Hansson and Håkansson, 2006; Larsson et al., 2006). Due to the lack of longterm data for coastal areas, these results have been extrapolated to the coastal zone and neighboring water bodies (Hansson and Håkansson, 2006). Because of the correlations of Chl-a with SPM and CDOM in coastal gradients the relationships may overestimate the response of ZSD to changes in nutrients levels and chlorophyll.

When the coastal reference and G/M Chl-a levels were applied in the different coastal areas to our models, the G/M levels of ZSD were unlikely to be met in the BP and in the SK region, but in the BS, as the latter is less eutrophicated. In the SK the observed ZSD met or were close to the G/M levels and changing the Chla levels mostly had a negative effect on the ZSD. In the BP the target levels of ZSD were met only in a very few cases, even when adjusting the Chl-a levels to the reference conditions. The results clearly indicate that reducing Chl-a alone will not be enough to reach a good status also for ZSD depth, or that the current ZSD targets are not realistic. Riemann et al. (2015) showed that the modest increase in ZSD measured in Danish coastal waters (over 25 years) was in general not only a response to reduced nutrient levels and Chl-a concentrations but was related to changes in other optical components such as lower SPM concentrations, and thus less scattering. An increase in SPIM at constant levels of total dissolved carbon (over 21 years), was found to determine the K<sup>d</sup> in a Danish fjord, despite substantial nutrient reduction leading to a decrease in Chl-a (Carstensen et al., 2013). Therefore, it may be questioned if the G/M levels of ZSD can be reached at all for water bodies with high concentrations of CDOM or SPIM, that are common in coastal waters. In fact, the reference and G/M thresholds for ZSD will be overestimated in such areas. The management (as well as the monitoring programs) could preferably be changed to be adapted more to the background values of CDOM and to some extent SPM.

Another important water quality indicator is the depth distribution of submersed aquatic vegetation (Dennison et al., 1993; Middelboe and Markager, 1997), which is affected by eutrophication because of lowered water transparency (European Commission, 2000, 2010; SwAM, 2015). An appropriate estimation of ZSD target and reference levels taking into account natural gradients in CDOM and SPIM is important for setting correct goals for the depth distribution of submersed aquatic vegetation (Dennison et al., 1993; Gallegos, 2001; Carstensen et al., 2013).

The correlation and commonality analyses showed that the main optical parameters are strongly linked in the Baltic Sea, and changes in Chl-a concentrations generally co-occur with changes in both SPIM and CDOM. Our application of the models when testing the effect on ZSD by changing Chl-a, assumes that we can change one parameter while keeping the other parameters constant. This approach can be somewhat misleading due to the general collinearity between variables. Strictly speaking, all three main optical components tend to change in tandem along a specific gradient, and it is therefore not correct to assume that the model fully captures the response when both CDOM and SPIM values vary while Chl-a is assigned to a fixed value. However, this was done in **Figure 7** for water types with defined fixed reference values and G/Mboundaries for Chl-a (BS and SK) in order to show what an effect a change in Chl-a has in different coastal waters with different optical properties and composition. For the BP, Swedish Chl-a reference values are linked to the salinity gradients so from this point of view, they are more realistic. Anyhow, a certain nutrient reduction, affecting Chl-a levels, should have different effects in water bodies with different levels of CDOM and SPIM. The effect would be largest in the SK, which has stronger unique and common effects from Chl-a concentration than found for the other areas. The other variables are more important, both uniquely and interactively, in the BS and the BP. Therefore, the effects of reduced Chl-a levels on ZSD are not only generally lower than predicted from simple Chl-a to ZSD relationships, they are also very difficult to estimate with a high degree of certainty in these areas.

Monitoring of eutrophication within the WFD and the MSFD does not only take Chl-a and ZSD into account, but they are two of the most important indicators used- together with nutrients, biovolume, oxygen conditions, and phytoplankton composition. Recent assessment of the eutrophication status e.g., within HELCOM and the Holas II assessment (HELCOM, 2018) use an integrated eutrophication assessment tool, considering the joint status of all indicators. A way forward for using ZSD as an indicator for eutrophication is to refine the ZSD reference values and thresholds based on the natural relationship among the optical parameters, preferably by optical modeling also taking historical data into account. Usually optical data sets have not been commonly collected within monitoring programmes but K<sup>d</sup> often is available from measured light profiles. Information about Chl-a and K<sup>d</sup> could then be used to model the historical reference values for ZSD.

## CONCLUSIONS

Secchi depth is a water quality indicator that is easy and relatively inexpensive to measure. However, it is very hard to interpret in the complex optical conditions found in the Baltic Sea. The same ZSD can result from very different combinations of phytoplankton, SPM and CDOM. Changes in ZSD are commonly influenced by changes in all optical constituents in the water column, not only by changes in the Chl-a concentration. With its uniquely long historical record, ZSD is, and will continue to be, an important general water quality indicator, as good water transparency means a lot to laypeople. Hence, as ZSD responds not only to Chl-a and phytoplankton, it is not an appropriate indicator for eutrophication assessment in areas such as the Baltic Sea, especially in coastal areas, with gradients in CDOM and SPIM. This has implication for management of the Baltic Sea, as well as of many other coastal and eutrophic waters. When setting reference and target levels for ZSD for use in management, it is imperative to consider also the contribution and response of the other two main optical components—CDOM and SPIM- besides Chl-a. Knowledge of local conditions and the causes and response in optical changes of a given water body is crucial for developing accurate and attainable goals in the WFD and the MSFD, both for ZSD and for the depth distribution of benthic vegetation. Measurements of both absorption and scattering properties combined with bio-optical modeling may help to identify alternative approaches to using ZSD as indicator for eutrophication. For example, the Chl-specific absorption of phytoplankton could be used as indicator for eutrophication as it is one of the main parameters determining the ability of phytoplankton to absorb light and thus their productivity. As the productive status of a water body may also be light limited, ZSD may still be an important co-factor for describing the eutrophication status of a water body. For routine monitoring programs at least the measurement of SPM (or turbidity), SPIM and CDOM should be feasible. Turbidity can be used as a proxy for SPM scattering and CDOM is measured directly in terms of absorption. Absorption and scattering can also be derived from satellite remote sensing data via the Copernicus program from the European Commission launched by ESA, especially using data from the Ocean Land Color Instrument (OLCI) on Sentinel-3. Using these inherent optical properties may guide a new way forward in eutrophication assessment as the absorption of phytoplankton as well as the K<sup>d</sup> can be derived from remote sensing data and is directly related to the productive status of the water body. These additional optical measurements both from in situ and remote sensing would provide invaluable information to interpret water quality and would improve both water quality assessment and management.

## AUTHOR CONTRIBUTIONS

ETH and SK are responsible for the original research idea, which was further developed together with JW. SK was in charge of the research in Himmerfjärden and the Baltic Proper gradients, together with JW. BK was responsible for the research in the Skagerrak region and AA for the Bothnian Sea region. Data analysis and model evaluation has been conducted by ETH together with primarily SK and JW with inputs from AA and BK. ETH is responsible for writing the article, but all co-authors have contributed significantly and participated in the discussions, especially SK and JW.

## FUNDING

Funding was provided by the FORMAS funded initiative Strategic Marine Environmental Research programs Baltic Sea Adaptive Management (BEAM) based at Stockholm University and Ecosystem dynamics in the Baltic Sea in a Changing climate perspective (EcoChange) based at Umeå University, The Swedish National Space Board (147/12, 110/16, 175/17), and the Swedish Environmental Protection Agency and the Swedish Agency for Marine and Water Management research program Waterbody Assessment Tools for Ecological Reference conditions and status in Sweden (WATERS) [10/179 and 13/33]. Supportive funding was also provided by the Svealand Coastal Water Association (Svealands kustvattenvårdsförbund, SKVVF). The Bothnian Sea data were collected under the EcoChange program. Baltic Proper and Skagerrak data were collected in the WATERS research program. SKVVF contributed with data for areas BP.1-4.

## REFERENCES


## ACKNOWLEDGMENTS

We would like to acknowledge all colleagues who collaborated in the monitoring programs and helped collecting and analyzing water samples (i.e., the Water Quality Association for the Bohus Coast sharing data and aid with sampling, the Marine Ecology Laboratory at the Department of Ecology, Environment and Plant Sciences, Stockholm University, Umeå Marine Sciences Center, and SKVVF). Samplings in the Bothnian Sea were performed by the EcoChange Research program. Sampling in the Skagerrak was coordinated with the BOX-project. The latter was coordinated by Prof. Anders Stigebrandt at the University of Gothenburg. Special thanks to Elina Kari and Madher Adalla for excellent help with the lab analyzes. We would also like to thank Prof. Ragnar Elmgren for valuable comments at an early stage of the manuscript.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2018.00496/full#supplementary-material


HELCOM (2017). Manual for Marine Monitoring in the COMBINE Programme of HELCOM. Helsinki. Available online at: http://www.helcom.fi/Documents/ Action%20areas/Monitoring%20and%20assessment/Manuals%20and %20Guidelines/Manual%20for%20Marine%20Monitoring%20in%20the %20COMBINE%20Programme%20of%20HELCOM.pdf (Accessed on Dec 18, 2018).


Jerlov, N. G. (ed.). (1976). Marine Optics. Amsterdam: Elsevier.


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Harvey, Walve, Andersson, Karlson and Kratzer. This is an openaccess article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Sediment Stocks of Carbon, Nitrogen, and Phosphorus in Danish Eelgrass Meadows

Theodor Kindeberg<sup>1</sup> \*, Sarah B. Ørberg<sup>2</sup> , Maria Emilia Röhr1,3, Marianne Holmer<sup>1</sup> and Dorte Krause-Jensen<sup>2</sup>

<sup>1</sup> Department of Biology, University of Southern Denmark, Odense, Denmark, <sup>2</sup> Department of Bioscience, Aarhus University, Silkeborg, Denmark, <sup>3</sup> Environmental and Marine Biology, Åbo Akademi University, Åbo, Finland

Seagrass ecosystems provide an array of ecosystem services ranging from habitat provision to erosion control. From a climate change and eutrophication mitigation perspective, the ecosystem services include burial and storage of carbon and nutrients in the sediments. Eelgrass (Zostera marina) is the most abundant seagrass species along the Danish coasts, and while its function as a carbon and nutrient sink has been documented in some areas, the spatial variability of these functions, and the drivers behind them, are not well understood. Here we present the first nationwide study on eelgrass sediment stock of carbon (Cstock), nitrogen (Nstock), and phosphorus (Pstock). Stocks were measured in the top 10 cm of eelgrass meadows spanning semi-enclosed estuaries (inner and outer fjords) to open coasts. Further, we assessed environmental factors (level of exposure, sediment properties, level of eutrophication) from each area to evaluate their relative importance as drivers of the spatial pattern in the respective stocks. We found large spatial variability in sediment stocks, representing 155–4413 g C m−<sup>2</sup> , 24–448 g N m−<sup>2</sup> , and 7–34 g P m−<sup>2</sup> . Cstock and Nstock were significantly higher in inner fjords compared to outer fjords and open coasts. Cstock, Nstock, and Pstock showed a significantly positive relationship with the silt-clay content in the sediments. Moreover, Cstock was also significantly higher in more eutrophied areas with high concentrations of nutrients and chlorophyll a (chl a) in the water column. Conversely, siltclay content was not related to nutrients or chl a, suggesting a spatial dependence of the importance of these factors in driving stock sizes and implying that local differences in sediment properties and eutrophication level should be included when evaluating the storage capacity of carbon, nitrogen, and phosphorus in Danish eelgrass meadows. These insights provide guidance to managers in selecting priority areas for carbon and nutrient storage for climate- and eutrophication mitigation initiatives.

Keywords: eutrophication, Zostera marina, blue carbon, sediment storage, ecosystem service

#### INTRODUCTION

The coastal ocean is a highly dynamic component of the Earth's system and features many complex interactions between oceanic, terrestrial, and atmospheric processes. The central role of coastal ecosystems in global and local biogeochemical cycles is in part due to disproportionately high biological activity in conjunction with extensive inputs of carbon and nutrients from rivers, terrestrial runoff, and upwelling (Gattuso et al., 1998; Duarte et al., 2005). Seagrasses are pervasive

#### Edited by:

Christian Grenz, UMR7294 Institut Méditerranéen d'Océanographie (MIO), France

#### Reviewed by:

Jenny R. Hillman, The University of Auckland, New Zealand Theresa O'Meara, Smithsonian Environmental Research Center (SI), United States

> \*Correspondence: Theodor Kindeberg theo.kindeberg@gmail.com

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 31 August 2018 Accepted: 26 November 2018 Published: 10 December 2018

#### Citation:

Kindeberg T, Ørberg SB, Röhr ME, Holmer M and Krause-Jensen D (2018) Sediment Stocks of Carbon, Nitrogen, and Phosphorus in Danish Eelgrass Meadows. Front. Mar. Sci. 5:474. doi: 10.3389/fmars.2018.00474

benthic macrophytes in coastal and estuarine waters spanning all continents except Antarctica (Short et al., 2007). In addition to providing important habitat for a wide range of species, seagrasses also play important roles in biogeochemical cycles including carbon and nutrient cycling (Duarte et al., 2005; McGlathery et al., 2007). Seagrasses, along with other marine vegetated habitats such as mangroves, saltmarshes, and macroalgae, have been recognized as a potential solution to combat anthropogenic perturbations such as excessive emissions of greenhouse gasses and nutrients to the environment (Nordlund et al., 2016; Gattuso et al., 2018; Himes-Cornell et al., 2018). The ability of marine vegetated habitats to sequester and store atmospheric CO<sup>2</sup> was first introduced by Smith (1981), and the term Blue Carbon was defined in the Blue Carbon report by Nellemann et al. (2009). This research field has propelled over the last decade and Blue Carbon has been proposed as a viable tool in both climate mitigation- and adaptation strategies (McLeod et al., 2011; Duarte et al., 2013; Greiner et al., 2013; Luisetti et al., 2013). Similarly, the potential of these habitats to act as nutrient filters has been suggested to alleviate effects of local eutrophication (McGlathery et al., 2007; Piehler and Smyth, 2011).

Due to a combination of high productivity and the ability to trap sestonic particles, seagrasses produce and bury large amounts of organic matter (OM) of which a portion is stored in the often anoxic underlying sediment and effectively sequestered from the ocean-atmosphere carbon pool (Fourqurean et al., 2012a). The sediment stock has been shown to be the predominant contributor to the total pool of organic carbon in seagrass meadows (Kennedy et al., 2010; Fourqurean et al., 2012a; Macreadie et al., 2014). In addition, a fraction of the seagrassderived detritus and dissolved organic matter (DOM) is exported from the meadow and may either be buried in depositional basins and the deep sea, or enter the microbial loop as refractory DOM (Barrón et al., 2014; Duarte and Krause-Jensen, 2017).

Nutrient cycling and retention in seagrass meadows is regulated both directly through the uptake and assimilation in leaves, roots, and rhizomes and indirectly through trapping of sestonic particles containing variable amounts of nutrients (Touchette and Burkholder, 2000; Eyre and Ferguson, 2002; Romero et al., 2006). In addition, elevated rates of denitrification carried out by microbes associated with the rhizosphere have been shown to contribute to additional export of nitrogen at the community and ecosystem scale (Eyre et al., 2016; Reynolds et al., 2016), although nitrogen fixation in seagrass meadows may also be significant (McGlathery et al., 1998; Welsh, 2000). However, the size and variability of the nutrient sink capacity of seagrass meadows, the mechanisms governing the variability, and the temporal aspects of the retention are not well constrained (Risgaard-Petersen et al., 1998; Nielsen et al., 2004; McGlathery et al., 2012).

Several studies have investigated the drivers behind Blue Carbon storage, and it has been shown that they consist of a complex set of physical, chemical, and biological factors that vary considerably between and within systems and are highly dependent on which functional scale is assessed (e.g., spatial, temporal, species, community, ecosystem) (see e.g., Ricart et al., 2015; Röhr et al., 2016; Samper-Villarreal et al., 2016; Belshe et al., 2017a,b; Dahl, 2017; Oreska et al., 2017a; Mazarrasa et al., 2018; Röhr et al., 2018). Sediment properties such as grain size, porosity, degree of sorting, and density are largely controlled by the hydrodynamic regime where, for instance, grain size typically increases with increasing wave- and current exposure, as the higher hydrodynamic energy and short residence times do not allow finer grains to settle (Folk and Ward, 1957; Fonseca and Bell, 1998; van Keulen and Borowitzka, 2003; Mazarrasa et al., 2017). Contrarily, low-energy environments typically observed in sheltered locations, tend to have a larger fraction of fine-grained material such as silt and clay (Jankowska et al., 2016; Röhr et al., 2016). Thus, information on sediment characteristics can serve as a useful proxy for the level of exposure in unvegetated areas, but the relationship may be confounded by water depth and in vegetated sediments by the seagrass canopy and rhizosphere's effect on sedimentation and resuspension (Fonseca and Bell, 1998; Madsen et al., 2001; Yang et al., 2008; Linders et al., 2018). Moreover, sediment grain size is a strong predictor of sediment OM pools and seagrass sediments exhibit increases in both the amount of carbon stored (Dahl et al., 2016; Röhr et al., 2016; Samper−Villarreal et al., 2016; Miyajima et al., 2017) and carbon burial rates (Mazarrasa et al., 2017) with increasing silt-clay content. This correlation may be due to the larger surface area for adsorbing organic molecules at high silt-clay content (Keil et al., 1994; Bergamaschi et al., 1997), and low oxygen availability reducing the remineralization of OM (Dauwe et al., 2001; Serrano et al., 2016), as well as the relatively higher sedimentation of sestonic particles in these low-energy environments (Mazarrasa et al., 2017, 2018). However, plant traits such as shoot density, above- and belowground biomass, and net primary productivity also explain a large part of the variability in carbon stock and sink capacity in local areas (Duarte et al., 2010; Gillis et al., 2017; Oreska et al., 2017a). The relative importance of biological and environmental factors seem to largely depend on which spatial scale is assessed, with biological factors being more important on smaller scales (meadow, patch) and environmental factors more important on larger scales (regional, global) (Dahl, 2017).

In the present study, we made a concerted effort to assess sediment stocks of carbon (Cstock), nitrogen (Nstock), and phosphorus (Pstock) in eelgrass (Zostera marina L.) meadows across Denmark and evaluated environmental properties as explanatory variables for the variability in sediment stocks. Eelgrass is a prevalent seagrass species which, although it has been in drastic decline, creates vast meadows in areas along the Danish coasts (Boström et al., 2014). Despite receiving extensive attention and research efforts in Denmark for over a century, the role of eelgrass in carbon and nutrient storage is not well known. To our knowledge, this study is the first to assess species-specific seagrass sediment storage of carbon, nitrogen, and phosphorus on a nationwide scale ranging from semi-enclosed estuaries to the open coast. Given the large spatial range in sampling areas, we proposed the following research questions: (i) what is the spatial variability in sediment Cstock, Nstock, and Pstock in Danish eelgrass meadows? (ii) To what extent is stock size related to sediment characteristics and/or level of eutrophication? Based on these questions we formulated the following hypotheses: (1) Cstock, Nstock, and Pstock are higher in sheltered areas with less exposure and higher silt-clay content and (2) stocks are higher in areas with higher nutrient and chl a concentrations in the water column.

### MATERIALS AND METHODS

fmars-05-00474 December 7, 2018 Time: 17:37 # 3

#### Assessing C, N, and P Stocks

To assess the sediment stocks of each element (Cstock, Nstock, and Pstock), particulate organic carbon [POC % dry weight (% dw)], total nitrogen (TN % dw), and total phosphorus (TP % dw) concentrations in the top 10 cm sediment were compiled from six studies conducted across Denmark in the period 1999–2014 (**Supplementary Table S1**). In all studies, sediment cores (1–4 cores per site) were collected in eelgrass meadows using similar techniques and subsequently analyzed for elemental composition and sediment properties. Sediment carbon density (g cm−<sup>3</sup> ) was calculated by multiplying dry bulk density [DBD (g cm−<sup>3</sup> ) with POC (% dw)] according to Howard et al. (2014). The sediment density of nitrogen and phosphorus was calculated in the same way but using TN and TP, respectively. In studies where OM [defined as loss on ignition, LOI (%)] was measured instead of POC, a conversion was performed using the empirical linear relationship between POC and OM found in studies measuring both parameters (**Supplementary Figure S1**). For sites where a full depth profile was obtained (0–10 cm), the accumulated amount of C, N, and P was assessed by multiplying the average (n = 1–4 cores) sediment density of carbon, nitrogen, and phosphorus by the length of each core section (cm) and integrating across the 10 cm depth to express as mass per area (g m−<sup>2</sup> ). The total number of sites, where all information needed to assess sediment stock was available, equaled n = 50 for Cstock, followed by n = 47 for Nstock, and n = 36 for Pstock.

Maps of Cstock, Nstock, and Pstock were created using the MATLAB <sup>R</sup> R2017b mapping toolbox (Mathworks, Inc.) and projected using the WGS1984 ellipsoid UTM zone 32N.

#### Environmental Parameters

To assess the level of exposure, a wave energy layer covering the sampling sites was obtained from the EUSeaMap<sup>1</sup> with a resolution of 300 m × 300 m. The wave energy layer is based on mean wind data in the period 2002–2007 (Wijkmark and Isæus, 2010).

The process of eutrophication is defined as "an increase in the rate of supply of organic matter to an ecosystem" (Nixon, 1995). However, several alternative definitions exist that concern the various proxies from which the level of eutrophication can be inferred (Cloern, 2001; Andersen et al., 2006). These include water column concentrations of macronutrients, chl a, and Secchi depth (Andersen et al., 2006). In order to assess relationships between sediment stocks and level of eutrophication we extracted data from the Danish National Monitoring and Assessment Program for the Aquatic and Terrestrial Environments (DNAMAP). Due to the long temporal range of our sediment data (1999–2014), we used annual averages

<sup>1</sup>www.EMODnet.eu

of water column concentrations of TN, dissolved inorganic nitrogen (DIN), TP, dissolved inorganic phosphorus (DIP), chl a, salinity, and Secchi depth from each site for the 5 years prior to the respective sampling events (Wulff et al., 1990; Christiansen and Emelyanov, 1995) (**Supplementary Data Sheet S1**). Based on levels of chl a and Secchi depth at the Danish intercalibration sites of the Water Framework Directive (WFD), all sites in this study can be considered eutrophic (Henriksen, 2009). Thus, we discuss various levels of eutrophication when referring to water column TN, TP, chl a, and Secchi depth.

To obtain information on the effects of sediment fine-grained material on sediment stocks, and to explore the relationship between level of exposure and fine-grained material, we used siltclay content (% dw, 8 < 62.5 µm) reported in each study. We used the average silt-clay content of the top 10 cm at all sites except for 10 sites, where only surface samples (0–2 cm) were available.

#### Statistical Analyses

The sediment cores were sampled in different waterbodies, which we qualitatively categorized into three types (inner fjord, outer fjord, and open coast) based on visual inspection of the sites' respective location (**Figure 1**). Out of the 60 sampling sites, 34 were categorized as inner fjord, 18 as outer fjord, and 8 as open coast. The sites differed in water depth ranging from 0.5 to 12 m, and we performed a one-way ANOVA to assess whether mean water depth differed between waterbodies. Similarly, we performed one-way ANOVAs to assess whether the silt-clay content and exposure differed between waterbodies.

To assess whether sediment stocks in Danish eelgrass meadows depend on waterbody type, we performed oneway ANOVAs. Cstock and Nstock were log-transformed to

categorization.

Frontiers in Marine Science | www.frontiersin.org

meet assumptions of normality and homogeneity of variances, obtained using Shapiro–Wilk's test of normality and Levene's test of homoscedasticity, respectively. Tukey's HSD post hoc tests were performed to test the pairwise difference in sediment stocks between waterbodies.

To examine environmental effects on sediment stock size, we constructed generalized linear models (GLM) for each sediment stock (Cstock, Nstock, and Pstock). Stocks were modeled as a function of silt-clay, TN, TP, chl a, Secchi depth, salinity, and exposure, using the Gaussian family with identity links. The model setup was chosen following evaluation of Akaike Information Criterion (AIC), likelihood-ratio test, and residual deviance. We also assessed the multicollinearity between explanatory variables by using variance inflation factors (VIF). In order to constrain the predictive power of each explanatory variable, we standardized all variables to z-scores by subtracting the mean from each observation and dividing by its standard deviation.

A caveat of using GLMs was that due to missing values at some sites, the model runs were applied to a reduced number of observations and thus did not cover the entire range of sites. Specifically, the GLMs for Cstock, Nstock, and Pstock were only applied to 27, 29, and 24 sites, respectively, and the relative number of waterbodies (inner/outer fjord, open coast) was not always proportional to the full dataset. Consequently, we also assessed the relationships between sediment stocks and environmental variables through simple linear regression analyses of one explanatory variable at a time. Similarly, we assessed the linear correlation between Cstock, Nstock and Pstock. Considering our aim to describe variation across all sites, we focus on these results and primarily use the GLMs to explore the predictive power of environmental factors for a subset of sites.

All statistics were performed with R statistical software (R Core Team, 2017) using the following packages: basic, boot, bbmle, car, and lmtest. A significance level of α < 0.05 was used for all parametric tests.

## RESULTS

## Eelgrass Sediment Stocks of C, N, P in Denmark

Concentrations (% dw) of sediment POC, TN, and TP exhibited significant heterogeneity ranging from 0.005–11.33, <0.001– 1.13, and 0.007–0.17, respectively. Mean (± SE) Cstock, Nstock, and Pstock inferred from the top 10 cm sediment equaled 1013 ± 116, 109 ± 12, and 17 ± 1 g m−<sup>2</sup> , respectively. The spatial variability was large and ranged 155–4413, 24–448, and 7–34 g m−<sup>2</sup> , for Cstock, Nstock, and Pstock, respectively (**Figure 2**). Furthermore, median stocks of C and N were substantially lower than mean values (711 and 68 g m−<sup>2</sup> , respectively), reflecting a lognormal distribution with a majority of lower stocks and a few very high stocks.

We found significant positive linear correlations between Cstock and Nstoc<sup>k</sup> (slope = 6.42, R <sup>2</sup> = 0.59, F1,<sup>30</sup> = 43.22, p < 0.001), between Cstock and Pstock (slope = 44.42, R <sup>2</sup> = 0.13, F1,<sup>29</sup> = 5.41, p = 0.027), and between Nstock and Pstock (slope = 0.04, R <sup>2</sup> = 0.25, F1,<sup>30</sup> = 10.18, p = 0.003).

We identified significant relationships between waterbody (inner or outer fjord, open coast) and sediment stock for Cstock (ANOVA, F2,<sup>47</sup> = 9.39, p < 0.001) and Nstock (ANOVA, F2,<sup>44</sup> = 4.10, p = 0.02) but not for Pstock (ANOVA, F2,<sup>33</sup> = 0.42, p = 0.52). Cstock was significantly higher in inner fjord locations than in outer fjords (post hoc, p < 0.001) whereas Nstock was significantly higher in inner fjords compared to open coast (post hoc, p = 0.03) (**Figure 3** and **Supplementary Table S2**). Neither silt-clay content nor level of exposure differed significantly between waterbodies, but the water depth of the sampling sites differed significantly between waterbodies (ANOVA, F2,<sup>43</sup> = 5.76, p = 0.006) and was on average significantly higher in outer fjords compared to the inner fjords (post hoc, p = 0.006), reflecting that the depth range of eelgrass is deeper in outer fjords (Riemann et al., 2016).

#### Drivers of Variability Sediment Properties

The silt-clay content in the sediment revealed a positive linear relationship with Cstock, Nstock, and Pstock with the highest slope for Cstock and highest R 2 value (0.30) for Nstock (**Figure 4** and **Supplementary Figures S4, S5**).

Silt-clay content was not significantly correlated with exposure (p = 0.29). Similarly, exposure was not significantly related to Cstock (p = 0.37), Nstock (p = 0.41), or Pstock (p = 0.15).

#### Water Column Properties

Water column concentrations of TN, TP, and chl a exhibited a similar pattern as sediment stocks and were all significantly different between waterbodies (ANOVA, F2,<sup>39</sup> = 6.43, p = 0.004; F2,<sup>39</sup> = 8.66, p < 0.001; F2,<sup>42</sup> = 6.56, p = 0.003, respectively) (**Supplementary Figure S2**). TP was significantly higher in inner fjords compared to open coasts (post hoc, p = 0.02), whereas TN, TP and chl a were significantly higher in inner fjords compared to outer fjords (post hoc, p < 0.05) (**Supplementary Figure S2**). For the entire dataset, DIN and DIP exhibited strong linear correlations with TN (TN = 2.6·DIN ++ 260, R <sup>2</sup> = 0.92, p < 0.001) and TP (TP = 1.7·DIP + 18, R <sup>2</sup> = 0.85, p < 0.001; figure not shown), respectively. On average (mean ± SE), DIN accounted for 13.8 ± 1.2% of TN and DIP accounted for 25.8 ± 1.6% of TP.

Salinity showed a different pattern with lowest salinities in inner (22.8 ± 0.6) and outer fjords (22.0 ± 0.8) compared to open coast (18.2 ± 0.2) (ANOVA, F3,<sup>47</sup> = 4.25, p = 0.001) (**Supplementary Figure S3**). Across all sites, salinity ranged from 15.8 to 30.7 and was lowest around Funen and highest in Nissum Bredning, Limfjorden, close to the mouth connecting to the North Sea (**Supplementary Figure S3**).

Cstock displayed a significant positive correlation with both water column nutrient concentrations and chl a, and a significant negative correlation with Secchi depth (**Figure 4**). In contrast, Nstock and Pstock displayed no significant

FIGURE 2 | Maps of average Danish eelgrass sediment stocks of (A) carbon (Cstock), (B) total nitrogen (Nstock), and (C) total phosphorus (Pstock) in the top 10 cm. Empty circles indicate missing data. Note that the colorbar scale differ between maps.

correlations with any water column parameter (**Supplementary Figures S4, S5**). Furthermore, no significant correlation was observed between any of the water column parameters and silt-clay content, except for Secchi depth which exhibited a significant negative correlation with silt-clay content, displaying the highest silt-clay content at low Secchi depths (**Supplementary Table S3**). Elemental ratios in the sediment were not related to water column nutrients, except that sediment N:P ratio was positively related to TP concentration in the water column (p = 0.03). The variation in sediment elemental ratios was highest in C:N, followed by C:P and N:P as indicated by their respective coefficients of variation (CV = standard deviation/mean) of 0.84, 0.58, and 0.57, respectively (**Supplementary Table S4**).

#### Model Predictions

The three GLMs, based on a subset of sites for which all variables were measured, overall supported the patterns of the

FIGURE 4 | Linear regression analyses of average Cstock and (A) sediment silt-clay content, (B) water column salinity, (C) total nitrogen, (D) total phosphorus, (E) chlorophyll a, and (F) Secchi depth. Black solid line indicates the best fit line result of linear regression and shaded area indicates 95% confidence interval. Note that the y-axis limit in panel (A) is different from other panels.

TABLE 1 | Summary table of generalized linear model results.

#### Stock ∼Siltclay + TN + TP + chl a + Secchi + Salinity + Exposure


All explanatory variables were transformed to z-scores prior to model runs. Variance inflation factors (VIF) describe the multicollinearity between explanatory variables in the GLM. Greater value indicates greater multicollinearity. Significant values indicated in bold.

variable-by-variable linear regression analyses. Consistent for all three sediment stocks was that silt-clay content was significantly positively related to stock size (**Table 1**). In contrast to the simple linear regression analyses, however, chl a was the only other variable that was significantly related to Cstock (p = 0.01) and Secchi depth was significantly related to Nstock (p = 0.005). As expected, water column concentrations of TN and TP exhibited the largest multicollinearity (**Table 1**) and were strongly correlated (slope = 15.1, R <sup>2</sup> = 0.82, p < 0.001).

Overall, the GLM predictions performed well and yielded R 2 values of 0.78, 0.66, and 0.71 for Cstock, Nstock, and Pstock, respectively (**Figure 5**).

## DISCUSSION

## C, N, P Stocks in Danish Eelgrass Sediments

We observed large spatial variability in sediment stocks of carbon, nitrogen, and phosphorus where sheltered, inner fjord locations generally comprised larger stocks whereas outer fjords and open coasts exhibited stocks that were sometimes orders of magnitude lower. These findings demonstrate the necessity of considering spatial variability when assessing regional or national sediment stocks and provide guidance into which eelgrass areas may be of particular importance as reservoirs of these elements.

## Sediment Characteristics and Exposure as Explanatory Variables for the Variability in Eelgrass Sediment Stocks

As hypothesized, we observed strong positive relationships between fine-grained material (silt-clay) in the sediment and Cstock, Nstock, and Pstock, which is in line with previous studies on eelgrass sediment carbon content (e.g., Dahl et al., 2016; Röhr et al., 2016; Mazarrasa et al., 2017; Oreska et al., 2017a; Röhr et al., 2018). In general, the grain size composition of marine sediments is controlled by the hydrodynamic regime and is often considered a useful proxy for wave- and current exposure (Yang et al., 2008; Cabaço et al., 2010; Mazarrasa et al., 2017). In absence of in situ current or exposure measurements, we utilized the publicly available EUSeaMap exposure layer. Notably, we did not find a relationship between level of exposure and silt-clay content in the sediment. This could be due to the relatively coarse resolution of the exposure layer used (300 m × 300 m) compared to the point measurements that the sampled cores constitute. The exposure layer also does not address temporal differences in exposure between sites, being based on mean wind conditions for a 5-year period between 2002 and 2007. Moreover, the level of exposure is strongly depth-dependent where the generally shallower sampling depths in inner fjords in this study represent a higher exposure level than the deeper sampling depths in outer fjords and open coast locations. However, we did not observe a significant relationship between water depth and silt-clay content or sediment stocks. The lack of correlation between silt-clay content and exposure may also be due to the attenuating effect of the eelgrass canopy on hydrodynamic energy which allows for fine-grained particles to settle, in addition to the stabilizing effect of the rhizosphere on limiting resuspension (Fonseca et al., 1982; De Boer, 2007; Hansen and Reidenbach, 2012). Lastly, considering that our sites ranged from semi-enclosed estuaries to open coast, variable patterns in sediment discharge from runoff may explain the lacking correlation, where low-exposure sites are often located in areas of higher riverine and/or terrestrial sediment input compared to the open coast and thereby offsetting the role of eelgrass.

## Eutrophication Effects on Eelgrass Sediment Stocks

The literature is equivocal regarding the effects of eutrophication on seagrass carbon storage and burial, reporting both higher storage and burial due to an increased input of sestonic particles (Gacia et al., 2002; Mazarrasa et al., 2017; Samper-Villarreal et al., 2017), and decreased storage and burial due to detrimental effects on seagrass productivity and survival (Macreadie et al., 2012; Jiang et al., 2018). Here, the level of eutrophication appeared to increase Cstock (**Figures 4C–F** and **Table 1**), suggesting that Danish eelgrass meadows have a relatively high reliance on allochthonous carbon (Oreska et al., 2017b). In fact, for 10 of the sites used in this study, the eelgrass net primary production only explained 2.3% of the variation in Cstock indicating a low contribution of autochthonous carbon (Röhr et al., 2016). Moreover, the positive relation between Cstock and silt-clay content and turbidity (Secchi depth) observed in this study suggests that the sources of carbon in the eelgrass sediment are largely of allochthonous origin (Samper−Villarreal et al., 2016). This was also demonstrated for 10 of the sites where analysis of stable isotopes (δ <sup>13</sup>C and δ <sup>15</sup>N) and C:N ratios in sediments and in carbon sources was carried out (**Supplementary Table S1**). An isotope mixing model suggested that the carbon content in eelgrass sediments was largely comprised (up to 78%) of sestonic particles (e.g., phytoplankton and macroalgae), further pointing to the influence of allochthonous carbon on Cstock (Röhr et al., 2016; Kindeberg et al., unpublished).

Unexpectedly, Nstock or Pstock were uncorrelated with water column nutrient concentrations (**Table 1** and **Supplementary Figures S4, S5**). This could be due to the lower sample size for

these stocks compared to Cstock which thus did not cover the entire range of eutrophication levels (TN: 248–1123 µmol L−<sup>1</sup> ; TP: 21–75 µmol L−<sup>1</sup> ; chl a: 2–6 µg L−<sup>1</sup> ). Moreover, it may reflect the different burial efficiencies and remineralization rates for the three elements due to differing nutrient content of the sediment OM, where OM rich in nutrients (low C:N and C:P ratios) is decomposed faster and to a larger extent than nutrient-depleted OM (Kristensen and Hansen, 1995; Banta et al., 2004; McGlathery et al., 2007). It is difficult to ascertain the processes affecting relationships between sediment stocks and C:N:P ratios as elemental ratios of OM (e.g., eelgrass leaves and phytoplankton) are affected both by nutrient supply and light availability, and a confounding effect of eutrophication on increased nutrients and associated decreased light on productivity may also confound the effects of OM decomposition (Zimmerman et al., 1987; Grice et al., 1996). Overall, we observed a quite large variability in C:N:P ratios across the different sites (**Supplementary Table S4**), illustrating the spatial difference in sources and/or age of OM and nutrient availability as well as sediment biogeochemistry (Fourqurean et al., 1997; Kennedy et al., 2010).

Nutrient dynamics in eelgrass sediments are comprised of many complex processes, some of which are directly governed by uptake of dissolved nutrients in the water column and sediment porewaters and others that are facilitated by the rhizosphere's microbial communities (Pedersen et al., 2004). These dynamics may explain the lack of correlation between water column nutrients and Nstock and Pstock. Pedersen and Borum (1993) estimated a total annual uptake rate of nitrogen of 25.3–50 g m−<sup>2</sup> year−<sup>1</sup> in an eelgrass meadow in Øresund, Denmark, of which 38–64% was attributed to the roots' uptake from sediment porewater. If uptake into eelgrass biomass is high and this biomass is subsequently exported out of the system (e.g., during storms), this could account for a significant loss term. Depending on the fraction of the plant detritus that is subsequently deposited in the sediment compared to exported elsewhere (Heck et al., 2008; Duarte and Krause-Jensen, 2017), this could be a factor which may explain the discrepancy between available nutrients in the water column and nutrient stocks in the sediment.

In addition, denitrification and anammox are both loss processes that convert inorganic nitrogen into nitrogen gas (N2, N2O, or NO) which escapes the sediment, and have been shown to be a major contributing process to nitrogen loss in seagrass habitats (Eyre et al., 2016; Reynolds et al., 2016). Rates of eelgrass-associated denitrification differ between studies (3.2–10.1 mg N m−<sup>2</sup> d −1 ) but the net effect on the sediment nitrogen fluxes should also consider nitrogen fixation, which can be significant (1.2–6.5 mg N m−<sup>2</sup> d −1 ) (see e.g., Flindt, 1994; McGlathery et al., 1998; Risgaard-Petersen and Ottosen, 2000; Welsh, 2000; Welsh et al., 2000; Cole and McGlathery, 2012; Russell et al., 2016). Although dependent on several factors and far from straightforward, denitrification rates typically increase and nitrogen fixation rates decrease in seagrass sediments with increasing eutrophication (Seitzinger and Nixon, 1985; Howarth et al., 1988; Welsh, 2000; Seitzinger et al., 2006; Murray et al., 2015; Asmala et al., 2017). This could be a potential explanation as to why Nstock did not mirror the nutrient concentrations in the water column.

Phosphorus cycling in marine sediments is governed by a complex set of biogeochemical processes and is distributed into various pools, where phosphorus loosely adsorbed or bound to iron oxides (Fe-P complexes) are generally the largest pools (Ruttenberg, 1992). In carbonate-rich areas, phosphorus is also often tightly bound to calcium carbonate (CaCO3-P complexes) whereas in siliciclastic muddy sediments, phosphorus is considered relatively more available due to the effects of coupled sulfate- and iron-reduction on releasing ironbound phosphorus (Canfield et al., 1993; Jensen et al., 1995). The effects of eutrophication on increasing sediment organic content may stimulate sulfate reduction thus releasing ironbound phosphorus (Holmer et al., 2006). As opposed to its tropical seagrass counterparts, eelgrass growing in fine-grained, siliciclastic sediments is rarely limited by phosphorus, although contrasting evidence exists (Murray et al., 1992; Holmer et al., 2006). Eelgrass may even excrete some phosphorus into the water column (Brix and Lyngby, 1985), and depending on the rate and magnitude of this process, eelgrass may facilitate an efflux of phosphorus out of the sediment to the water column. Taken together, these two processes may mask the effects of water column nutrient concentrations on Pstock and could, in addition to export of eelgrass biomass, explain the lack of correlation in our dataset.

Notably, neither water column nutrients nor chl a were significantly correlated with silt-clay, suggesting a spatial dependence in the explanatory power of these variables and it is therefore difficult to attribute the variability in stocks solely to either factor. Both eutrophication and too high silt-clay content can have negative effects on eelgrass subsistence and the positive effects on stock size observed here are likely to diminish when reaching a certain threshold (De Boer, 2007; Viaroli et al., 2008). Eelgrass is, as seagrasses in general, sensitive to turbidity and organic-rich sediments due to high light requirements and sensitivity to highly reducing sediments (Goodman et al., 1995; Holmer and Bondgaard, 2001; Kemp et al., 2004; Ochieng et al., 2010). Once a tipping point is exceeded and the eelgrass is lost, the system can enter an alternative stable state where sediments are eroded, and the stored carbon, nitrogen, and phosphorus can re-enter the water column (Pendleton et al., 2012; Arias-Ortiz et al., 2017; Moksnes et al., 2018). Depending on the bioavailability of the eroded nutrients, this efflux can result in net CO<sup>2</sup> emissions and fuel a positive feedback loop on water quality deterioration. Further investigation into this complex balance is warranted and in particular the observed decoupling between the effects of eutrophication and sediment silt-clay content on a larger spatial scale should be explored further.

## Danish Eelgrass Sediment Stocks in a Broader Seagrass Context

Due to the relatively shallow sediment depth (10 cm) used in this study, it is difficult to directly compare our results with those from other studies which usually assess sediment stocks based on deeper cores [e.g., ≥100 cm proposed as the practicable standard for carbon stock assessments (Howard et al., 2014; Emmer et al., 2015)] (Fourqurean et al., 2012a,b;

Kindeberg et al. CNP Stocks in Danish Eelgrass

Miyajima et al., 2015). However, when comparing the mean POC concentration (1.1 ± 0.07%) found in our study, Danish eelgrass sediments fall within the lower end of the global range of 0–48.2% and below the global mean of 2.5 ± 0.1% for seagrasses (Fourqurean et al., 2012a). Nevertheless, the identified POC concentrations are well above the reported estimates for Z. marina in the southern Baltic Sea (0.14%, Jankowska et al., 2016) and similar to the recently reported global average for Z. marina of 1.4 ± 0.4% (Röhr et al., 2018). It should be noted, however, that mean POC concentration alone does not provide information of the total accumulated carbon in the sediment (Cstock), which requires information on sediment density and integration over sediment depth (Howard et al., 2014).

Similar to our study, Jankowska et al. (2016) measured Cstock in the top 10 cm, sampled in eelgrass meadows in a sheltered inner bay, an exposed outer bay, and a shallow open coast area in Puck Bay, Poland. In accordance with our results, the highest Cstock values were found in the sheltered, inner location (228.0 ± 11.6 g C m−<sup>2</sup> ) and lowest in the exposed, outer location (50.2 ± 2.2 g C m−<sup>2</sup> ). However, these values are about an order of magnitude lower than the average Cstock we observed in respective waterbodies in Denmark (Inner fjord: 1357 ± 173 g C m−<sup>2</sup> , Outer fjord: 525 ± 117 g C m−<sup>2</sup> ; **Supplementary Table S2**), which could partly be due to a higher fraction of larger grain sizes (mean ϕ = 1.68–2.39) in Puck Bay.

Postlethwaite et al. (2018) measured Cstock in three eelgrass meadows in Clayoquot Sound, BC, Canada, but integrated over deeper sediment depths (34 cm on average) compared to our study. They found Cstock to range from 820 ± 26 g C m−<sup>2</sup> in areas with sandier sediments closest to the coast and up to 2099 ± 365 g C m−<sup>2</sup> near the river mouth. Considering that these Cstock measurements were obtained from more than three times as deep cores than used in our study, their values are generally lower than presented here, as their carbon depth profiles were relatively stable with depth. This difference could partly be due to the more pristine conditions with low anthropogenic influence, high Secchi depth (4.8–8.8 m), and likely lower sediment discharge as compared to most of the sites in Denmark (Postlethwaite, 2018; Postlethwaite et al., 2018). A similar study was carried out south of British Columbia, in Puget Sound, Washington, DC, United States. Poppe and Rybczyk (2018) found eelgrass Cstock to range from 1180–1900 g C m−<sup>2</sup> , averaging at 1420 ± 110 g C m−<sup>2</sup> , but also here the integrated core depth was three times as deep (30 cm) as in our study. A recent comprehensive assessment of Cstock (integrated over 25 cm sediment depth) across the distributional range of Z. marina further established the spatial variability, which ranged from 318 ± 10 – 26523 ± 667 g C m−<sup>2</sup> , and averaged 2721 ± 989 g C m−<sup>2</sup> (Röhr et al., 2018). In line with the results presented here, silt-clay content was found to be the most important predictor explaining 53% of the variation in Cstock. Among the largest observed stocks were typically found in sheltered locations in the Kattegat-Skagerrak ocean margin in which the sites in this study are located. Based on the datasets in the abovementioned studies, Danish eelgrass meadows seem to hold larger Cstock than other areas of comparable latitudes which highlights the influence of the local environmental setting.

Fewer studies have assessed nitrogen and phosphorus stocks in seagrass sediments and the current study is, to our knowledge, the largest-scale assessment available in the literature. McGlathery et al. (2012) measured the sediment nitrogen stock in the top 5 cm sediment in a 9-year-old restored eelgrass meadow in Virginia, United States and found an average Nstock of 16.2 g N m−<sup>2</sup> . Assuming a stable profile between 5–10 cm, and extrapolating across 10 cm, would put this restored eelgrass meadow (32.4 g m−<sup>2</sup> ) on the lower end of the range of Nstock presented here (24–448 g m−<sup>2</sup> ). While not explicitly assessing sediment stocks, Yang et al. (2018) measured the sediment nitrogen concentration in three eutrophic sites colonized by Z. marina and Z. japonica in northeastern China, where the average nitrogen content in the top 10 cm sediment was almost three times higher (0.35%) than what we found in Denmark. Fourqurean et al. (2012b) assessed nitrogen and phosphorus in the top 100 cm sediment in two seagrass systems (mainly Thalassia testudinum) in Western Australia and Florida, United States. The average nitrogen concentration in the sediment was similar between the two areas (0.15 ± 0.004 and 0.19 ± 0.01%, respectively), which is comparable to 0.12 ± 0.01% in our study. Phosphorus content exhibited much greater difference between their sites and was on average 87.7 ± 4.3 µg g−<sup>1</sup> in Florida and 135.7 ± 3.6 µg g−<sup>1</sup> in Australia, which is lower than what we found (310.1 ± 18.6 µg g −1 ) and was explained by the carbonate-rich sediments in these locations in which phosphorus content in the porewater is known to be low (Short, 1987; Fourqurean et al., 2012b).

## Uncertainty, Limitations, and Future Direction

For best practices assessments of sediment stocks and longterm storage in seagrass meadows, sediment cores should ideally extend at least 100 cm to estimate buried carbon and nutrients below the depth of decomposition (Howard et al., 2014; Macreadie et al., 2014). One important caveat of assessing stocks in the top 10 cm is that this shallow sediment layer is considered biologically (e.g., bioturbation and -irrigation), physically (e.g., wave action, sediment resuspension) and biogeochemically (e.g., high remineralization rates, redox oscillations) active, and the sediment properties (including C, N, and P stocks) can vary over short time-scales (e.g., seasonal, annual) (Middelburg et al., 2004; Burdige, 2007; Johannessen and Macdonald, 2016). Consequently, although the top 10 cm in similar areas have been shown to represent approximately the past 60 years (Jankowska et al., 2016), any assertions regarding long-term carbon and nutrient storage cannot be adequately made (Johannessen and Macdonald, 2016; Nilsson, 2018). Normalizing stocks based on 10 cm cores to 25 or 100 cm requires untested assumptions of invariant change with sediment depth and that the 10 cm profile pattern is representative of deeper depths. However, deeper sediment cores (25 cm) from 10 sites in this study revealed that most depth profiles of carbon exhibited either a stable, or even increasing, pattern with sediment depth between 10–25 cm, especially in sites with large Cstock (**Supplementary Figure S6**), suggesting that the risk of overestimation may be limited in this case. Furthermore, from these sites the vertically extrapolated Cstock (from 10 to 25 cm) was generally lower than the measured

Cstock with a mean offset (± SE) of 895 ± 527 g C m−<sup>2</sup> , corresponding to 64.7 ± 20.8% of the average Cstock. Although uncertainties associated with deeper extrapolation persist, the estimated Cstock values presented here should be considered conservative (Röhr et al., 2016).

A second limitation of this study is the lack of adjacent, unvegetated sites to allow for estimation of buried carbon and nutrients in the absence of eelgrass cover. However, Kennedy et al. (2010) reported almost twice as high median organic carbon content (0.34% compared to 0.19%) and total nitrogen (0.031% compared to 0.024%) in sediments within the seagrass meadow compared to adjacent bare. Similarly, Jankowska et al. (2016) observed 1.5–4.8 times higher organic carbon concentrations and 1.6–3.1 times higher nitrogen concentrations in eelgrass-vegetated sediments compared to unvegetated. Similar differences in OM content between vegetated and unvegetated reference sediments have also been documented elsewhere (see e.g., Miyajima et al., 1998; Ricart et al., 2015; Postlethwaite et al., 2018). This aspect should, however, be a focal point in future studies as the lack of unvegetated reference sediment is prevalent in the seagrass sediment stock literature. Yet, selecting a reference is by no means uncomplicated as both the OM produced by seagrass itself and the sestonic particles it captures have been documented to bury far away from the canopy (Kennedy et al., 2010; Duarte and Krause-Jensen, 2017).

A third important consideration concerning studies on seagrass sediment storage is the ability to scale measurements to basin-wide, regional, or even global level. Large scale extrapolation of sediment stocks warrants considerable precaution, as spatial (and temporal) variability is large not only between meadows as shown here, but can also be significant within meadows (Oreska et al., 2017a). Any upscaling of meadow point measurements therefore requires assumptions of invariable abiotic properties (e.g., hydrodynamic regime, residence time, sediment dynamics, geomorphology, water column chemistry) and biotic properties (e.g., meadow configuration, productivity, shoot density). This aspect was to some extent illustrated in the, albeit limited, discrepancy between the results of the GLM and the variable-by-variable regressions where the number of sites and environmental settings differed. Nevertheless, there is a need to provide stakeholders, managers, and policymakers with incentives to conserve and protect these valuable habitats, which are declining at rapid rates (Orth et al., 2006; Waycott et al., 2009). However, rather than producing first-order estimates of sediment storage on a larger scale, we emphasize the small-scale spatial variability and provide guidance into selecting high priority areas for carbon and nutrient storage on the local scale. Denmark does indeed comprise hotspots for carbon and nutrient storage, which can be targeted as high priority areas in the management of seagrass meadows for mitigation of climate change and eutrophication. By further pinning down the associated environmental factors, this study also indicates where additional hotspots may be located. All in all, this nationwide information on C, N, and P stocks highlights the combined climate change and eutrophication mitigation potentials of eelgrass meadows and thereby stimulates incentives for conservation efforts.

## CONCLUSION

In this study, we have further established the large spatial variability in sediment storage of carbon, nitrogen, and phosphorus in seagrass systems. Our study is the first to assess these three elements in concert for a single species on a nationwide basis covering a variety of waterbodies. We have shown that carbon and nitrogen stocks are higher in sheltered inner fjords compared to more exposed locations. Furthermore, we have shown that environmental factors that are often considered detrimental to eelgrass, such as eutrophication and high sediment mud content (De Boer, 2007), to a certain extent have a positive effect on the sediment stocks and this intricate balance should be explored further. The heterogeneous nature of eelgrass sediment stocks and the varying influence of environmental parameters observed here highlights the importance of considering the environmental setting when assessing carbon and nutrient storage in sediments, constructing regional budgets, and evaluating ecosystem services in eelgrass meadows on a national scale.

## AUTHOR CONTRIBUTIONS

TK, SØ, MH, and DK-J conceived and designed the study. MR collected the data. TK and SØ carried out data analysis. TK drafted the manuscript. All authors revised and approved the manuscript.

## FUNDING

This work was built upon initial work presented (TK) at the 4th International Symposium on Research and Management of Eutrophication in Coastal Ecosystems (EUTRO2018). The study was part of the project Havets skove, funded by the Villum Foundation (Grant Number 18530).

## ACKNOWLEDGMENTS

We would like to thank Prof. Jacob Carstensen for providing water column chemistry data from the DNAMAP database. DK-J also acknowledges support from the Ministry of Environment and Food of Denmark (contract 33010-NIFA-16651) and from the Danish Centre for Environment and Energy. We thank CG, JH, and TO for constructive comments that significantly improved the quality of this paper.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/ 10.3389/fmars.2018.00474/full#supplementary-material

## REFERENCES

fmars-05-00474 December 7, 2018 Time: 17:37 # 11




Delta: on the role of physical and biotic controls. Estuar. Coast. Shelf Sci. 77, 657–671. doi: 10.1016/j.ecss.2007.10.024


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Kindeberg, Ørberg, Röhr, Holmer and Krause-Jensen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Assessing the Effects of WFD Nutrient Reductions Within an OSPAR Frame Using Trans-boundary Nutrient Modeling

#### Hermann-Josef Lenhart <sup>1</sup> \* and Fabian Große1,2

<sup>1</sup> Scientific Computing, Department of Informatics, Universität Hamburg, Hamburg, Germany, <sup>2</sup> Department of Oceanography, Dalhousie University, Halifax, NS, Canada

The reduction of riverine nutrients inputs is considered the means of choice to improve the eutrophication status of the southern North Sea. With the European Union's Water Framework Directive (WFD) reduction measures presently under debate, two questions arise: (1) What changes in eutrophication indicators can be expected? (2) How do the reductions by the individual member states contribute to these? We combine an element tracing method (TBNT) with a biogeochemical model to analyze the effects of WFD-compliant nitrogen reductions proposed by OSPAR's North Sea member states. We first analyze changes in selected OSPAR assessment parameters relative to a reference simulation. Second, we quantify the source-specific contributions to total nitrogen (TN) in different regions. An overall nitrogen load reduction of 14 % is achieved. However, the response shows significant spatial variations due to strong differences between the countries' load reductions. TN and dissolved inorganic nitrogen reductions up to 60 % and 35 % are simulated near the Bay of Seine (France) and in the German Bight, respectively. Along the Dutch coast, reductions are below 10 %, and no changes occur along the British coast. Reductions in chlorophyll-a are generally lower. The TBNT analysis for the German Exclusive Economic Zone shows a TN reduction in the coastal region comparable to the N reductions in the German rivers (~25 %). In the offshore region, TN is reduced by only 6 % due to the strong influence of riverine sources with only low reductions and non-riverine sources. Our analysis reveals that non-linear responses in the biogeochemistry cause a faster removal of N from rivers with strong reductions by benthic denitrification, which enhances indirectly the removal of N from less reduced sources. Consequently, reductions in remote sources in non-problem areas can have a relevant positive effect on problem areas. This demonstrates that the TBNT method is an ideal tool to put in practice the "source-oriented approach" advocated by OSPAR, and to inform stakeholders about the effects of defined reduction strategies. However, an assessment framework is required to efficiently use it in management and for decision making, either by OSPAR, or in the context of WFD or Marine Strategy Framework Directive.

Keywords: North Sea, eutrophication, biogeochemical modeling, nutrient tagging, nitrogen cycle, nutrient reductions, Water Framework Directive (WFD), trans-boundary nutrient transports (TBNT)

#### Edited by:

Jesper H. Andersen, NIVA Denmark Water Research, Denmark

#### Reviewed by:

Xavier Desmit, Royal Belgian Institute of Natural Sciences, Belgium Theo C. Prins, Deltares, Netherlands

\*Correspondence: Hermann-Josef Lenhart hermann.lenhart@uni-hamburg.de

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 31 August 2018 Accepted: 08 November 2018 Published: 27 November 2018

#### Citation:

Lenhart H-J and Große F (2018) Assessing the Effects of WFD Nutrient Reductions Within an OSPAR Frame Using Trans-boundary Nutrient Modeling. Front. Mar. Sci. 5:447. doi: 10.3389/fmars.2018.00447

## 1. INTRODUCTION

Eutrophication, i.e., the "increase in the rate of supply of organic matter to an ecosystem" (Nixon, 1995), has been and still is an ongoing problem in the North Sea ecosystem, primarily driven by excess riverine nutrient loads. Its negative impact on the marine environment was first observed in the 1980s in the form of low oxygen conditions (Rachor and Albrecht, 1983; von Westernhagen and Dethlefsen, 1983), and later linked to high riverine nutrients (Brockmann and Eberlein, 1986; Brockmann et al., 1988; Peeters et al., 1995). As a result of this period of severe eutrophication, the ministers of environment decided on the 2nd International Conference on the Protection of the North Sea (ICNS-2) in 1987 to "reduce the river nutrient loads of phosphorus and nitrogen by 50 %" between 1985 and 1995 to mitigate the negative effects of eutrophication (ICNS-2, 1988). However, this goal has only been reached for phosphorus (P), but not for nitrogen (N) (Claussen et al., 2009; Lenhart et al., 2010), and several recent studies on oxygen in the North Sea provide evidence that eutrophication still is an important issue (Greenwood et al., 2010; Topcu and Brockmann, 2015; Große et al., 2016, 2017; Queste et al., 2016).

Since then, OSPAR, as the regional sea convention for the North-East Atlantic (www.ospar.org), has regularly assessed the eutrophication status of the North Sea applying its Common Procedure (COMP; OSPAR, 2003, 2005, 2013). The COMP assessment distinguishes between different categories of eutrophication indicators that describe (I) the "degree of nutrient enrichment," (II) "direct effects" (e.g., algal blooms) and (III) "indirect effects of nutrient enrichment" (e.g., oxygen deficiency), and (IV) "other effects of nutrient enrichment" (e.g., algal toxins). In practice, the COMP is based on thresholds defined for different key parameters, such as winter concentrations of dissolved inorganic N (DIN; category I indicator) or summer surface chlorophyll-a (Chl-a; category II) (Claussen et al., 2009; OSPAR, 2017). Unfortunately, even the latest COMP report shows only limited improvement over the years (OSPAR, 2017). It identifies large parts of the southern North Sea between the Belgian and Danish coasts as so-called "problem areas" or "potential problem areas," while only small regions along the French and British coasts are characterized as such.

With the aim to better understand what improvements in the eutrophication status could be achieved and in what timeframe, the OSPAR "Intersessional Correspondence Group for Ecosystem Modeling" (ICG-EMO) was established in 2005. Its assigned task is the application of marine ecosystem models to foster the understanding of the North Sea ecosystem dynamics and to assess the measures that are required to improve the eutrophication status of the North Sea.

Following different objectives defined by OSPAR, ICG-EMO conducted a series of model intercomparisons on nutrient reduction targets (Lenhart et al., 2010; OSPAR, 2010, 2013). This approach proved to be very beneficial for both scientists and OSPAR. On the one hand, it led to the first application of the OSPAR assessment parameters to model results of both a present state simulation and a reduction scenario to provide insight in potential future changes under nutrient reductions (Almroth and Skogen, 2010; Lenhart et al., 2010; OSPAR, 2013). On the other hand, the modeling community could identify and report inconsistencies within the nationally oriented OSPAR thresholds for the different assessment parameters (OSPAR, 2013), and it provided quantitative information on the nutrient reductions required to reach a North Sea free of problem areas (Los et al., 2014).

In 2000, the European Commission put into practice the Water Framework Directive (WFD; EU, 2000) with the goal to reach a "Good Ecological Status" (GES) in all water bodies of the member states. Since this directive was mainly focused on catchments, with only small assessment areas along the coast, it was extended toward the marine environment by the Marine Strategy Framework Directive (MSFD) in 2008 (EU, 2008). Hence, the WFD and the MSFD constitute legislative frameworks to combat eutrophication in European seas, including the North Sea, which is in line with the objective of OSPAR.

The latest assignment of OSPAR to ICG-EMO was to assess the impact of the different nutrient reduction measures defined in the individual national WFD programs. It was based on the fact, that most Contracting Parties had management plans in relation to their WFD programs available. However, these management plans are designed to combat eutrophication in the different countries' own Exclusive Economic Zones (EEZ). In order to address these management plans, Contracting Parties were asked to submit their WFD programs of measures to the OSPAR Hazardous Substances and Eutrophication Committee (HASEC). From these national WFD programs the reduction levels for nitrogen loads were extracted for the individual rivers entering the North Sea (OSPAR, 2016), which constitute the basis for the ICG-EMO modeling study. At the current stage, these programs only consider reductions in riverine N loads.

In this study, we aim to fulfill the "source-oriented approach" advocated by OSPAR (1999) in the context of these N reductions. For this purpose, we (1) quantify the changes in selected OSPAR assessment parameters in response to the WFD-compliant N reductions; (2) link these changes to the individual national measures and quantify their impact; and (3) identify changes in the N cycle induced by these measures.

In order to address these objectives, we conduct a WFDcompliant reduction scenario using a physical-biogeochemical model of the North Sea and compare the results to a reference simulation. In addition, we apply an active element tracing method (Ménesguen and Hoch, 1997) to the N dynamics of both simulations. This method—often referred to as "Trans-Boundary Nutrient Transports" (TBNT; Blauw et al., 2006; OSPAR, 2010)–allows for the tracing of elements from individual sources through all physical and biogeochemical processes, and thus provides quantitative information on the influence of these sources on the N dynamics in the different North Sea regions. A comparative analysis of the TBNT results of the WFD reduction scenario vs. the reference simulation allows us to quantify the changes induced by the different N reductions on both OSPAR key parameters and the N cycle in the North Sea.

#### 2. MATERIALS AND METHODS

In this section, we first provide a brief description of the physicalbiogeochemical model HAMSOM-ECOHAM, followed by a brief theoretical background of the TBNT method applied to the model. Finally, we provide a detailed description of the model and TBNT setup including the WFD reduction scenario.

#### 2.1. The HAMSOM-ECOHAM Model

Our study is based on a three-dimensional (3D) model consisting of the physical model HAMSOM (HAMburg Shelf Ocean Model; Backhaus, 1985; Pohlmann, 1991, 1996) and the biogeochemical model ECOHAM (ECOsystem model, HAMburg; Pätsch and Kühn, 2008; Kühn et al., 2010; Lorkowski et al., 2012; Große et al., 2016).

The physical model HAMSOM (Backhaus, 1985) is a baroclinic, primitive equation model using the hydrostatic and Boussinesq approximation (Pohlmann, 1991). The current velocities are calculated using the component-upstream scheme. The horizontal dimensions are discretized on a staggered Arakawa C-grid (Arakawa and Lamb, 1977) and z-coordinates are applied to the vertical. HAMSOM calculates the 3D fields of advective flow, vertical turbulent mixing, temperature and salinity, which are used as forcing for ECOHAM. A detailed description of HAMSOM is provided by Pohlmann (1991, 2006), Chen et al. (2013), and Mathis and Pohlmann (2014).

The biogeochemical model ECOHAM (Pätsch and Kühn, 2008; Lorkowski et al., 2012; Große et al., 2016) represents the pelagic and benthic cycles of carbon (C), nitrogen (N), phosphorus (P), silicon (Si), and oxygen (O2). The model includes all parameter groups of an NPZD-type model (nutrients-phytoplankton-zooplankton-detritus) that cover the lower trophic level dynamics. It describes four inorganic nutrients (nitrate, ammonium, phosphate, and silicate), dissolved inorganic C, two phytoplankton (diatoms and flagellates) and two zooplankton groups (micro- and mesozooplankton), and slowly and fast sinking detritus. The "microbial loop" (Azam et al., 1983) is represented by further including labile and semi-labile dissolved organic matter (DOM) and bacteria. For phytoplankton, zooplankton and bacteria individual but fixed C:N:P ratios are applied. For detritus and labile DOM, the C:N:P ratios can evolve freely. In ECOHAM, Chl-a concentrations are derived from C bound in phytoplankton according to the empirical relationship of Cloern et al. (1995). The self-shading effect of phytoplankton on the light climate is implemented and depends on the Chl-a concentration using an attenuation coefficient of 0.02 m<sup>2</sup> (mg Chl-a)−<sup>1</sup> .

The sediment is described by a simple zero-dimensional module (Pätsch and Kühn, 2008). Benthic remineralization follows a first-order approach inhibiting year-to-year accumulation of organic matter (Große et al., 2016). The released dissolved inorganic matter is returned directly into the deepest pelagic layer. Different remineralization rates are applied to organic C, N, P, and Si (opal), resulting in different delays for the release into the pelagic. Benthic denitrification is linked to benthic O<sup>2</sup> consumption following Seitzinger and Giblin (1996), and reducing the O<sup>2</sup> concentration in the deepest pelagic layer. Explicit benthic nitrification and benthic anammox are not implemented (Pätsch and Kühn, 2008). The ECOHAM version applied for this study is identical to that used by Große et al. (2016, 2017). Lorkowski et al. (2012) provide a full description of the ECOHAM model equations and parameter settings.

#### 2.2. The TBNT Method

The element tracing method applied in this study is based on the work by Ménesguen and Hoch (1997), who describe that any selected property (e.g., the source of a N element brought into an ecosystem) can be traced throughout all physical and biogeochemical processes represented by the applied model. Since then, several modeling studies made use of this method with various research objectives (e.g., Wijsman et al., 2004; Blauw et al., 2006; Ménesguen et al., 2006; Lacroix et al., 2007; Neumann, 2007; Timmermann et al., 2010; Troost et al., 2013; Radtke and Maar, 2016), demonstrating the versatility of this method. In the meantime, the term "Trans-Boundary Nutrient Transports" (TBNT) was established (Blauw et al., 2006; OSPAR, 2010) and is used in particular within the OSPAR frame.

Conceptually, the TBNT method labels all matter, which contains a selected chemical element (e.g., N), according to its source when it enters the ecosystem under consideration. Technically, this implies the introduction of an additional set of model state variables and related processes, as all state variables containing the selected element need to be labeled. The physical and biochemical processes working on the labeled state variables are the same as for the overall state variables, i.e., the total amount of labeled and unlabeled material, however, proportional to their relative contribution to this overall amount. Following Große et al. (2017), the temporal evolution of the concentration of a labeled state variable X i can be calculated as:

$$\frac{dC\_X^i}{dt} = \nabla \cdot \underbrace{\left(\overline{\overline{D}} \nabla C\_X\right) \cdot \frac{C\_X^i}{C\_X}}\_{\text{diffusion}} \underbrace{-\nabla \cdot (C\_X \vec{\nu}) \cdot \frac{C\_X^i}{C\_X}}\_{\text{advection}} + \underbrace{R\_{\text{C}x} \cdot \frac{C\_{X\_{\text{con}}}^i}{C\_{X\_{\text{con}}}}}\_{\text{sources/sinks}}.\tag{1}$$

Here, C i X and C<sup>X</sup> represent the concentrations of the fraction of state variable X originating from the i-th input source and that of the corresponding bulk state variable, respectively. The diffusive transport is calculated according to Fick's first law, with the second-order diffusion tensor D. In the advective transport term, Ev represents the 3D velocity vector. RC<sup>X</sup> represents the change in concentration of X due to the sources and sinks (i.e., biogeochemical processes, input from external sources). The index Xcon in the fraction of this term indicates that the relative contribution of the state variable that is consumed by a biogeochemical process is used.

For n individually labeled input sources (i.e., i = 1, 2, ..., n − 1, n), the concentration of each bulk state variable X at each location and point in time equals the sum of the concentrations of its n contributing fractions:

$$C\_X = \sum\_{i=1}^n C\_X^i. \tag{2}$$

### 2.3. Study Setup

2.3.1. Model Setup and Nitrogen Reduction Scenario The model and TBNT setup for this study is identical to that used by Große et al. (2017), hence, we only describe its main aspects. The HAMSOM-ECOHAM model is set up for a domain encompassing the entire North Sea, large parts of the northwestern European continental shelf and parts of the adjacent Northeast Atlantic (**Figure 1**). The horizontal resolution is 1/5 ◦ with 82 grid points in latitudinal direction and 1/3 ◦ with 88 grid points in longitudinal direction. The vertical dimension with a maximum depth of 4,000 m is resolved by 31 z-layers with a surface layer of 10 m thickness. Between 10 and 50 m depth, the vertical resolution is 5 m. Below 50 m, the layer thicknesses successively increase with depth.

We first run the HAMSOM model for the period 1977– 2014 using 6-hourly information for air temperature, cloud coverage, relative humidity, wind speed and direction derived from NCEP/NCAR reanalysis data (Kalnay et al., 1996; Kistler et al., 2001). Daily freshwater run-off data for 254 rivers are provided by Sonja van Leeuwen (pers. comm.) and represent an updated dataset of that used by Lenhart et al. (2010) covering the entire simulation period. Monthly climatologies of sea temperature and salinity based on the World Ocean Atlas 2001 (Conkright et al., 2002) are used for initialization and at the open boundaries. The HAMSOM simulation is carried out with a 10 min time step, and output is stored on a daily interval.

In order to analyze the effects of the WFD-compliant riverine N reductions on the North Sea, we run two different ECOHAM simulations using the same physical forcing produced by HAMSOM. The first simulation runs for the period 1977– 2014 using realistic forcing (hereafter "reference"). We provide daily nutrient loads based on the same dataset as the freshwater discharge (see Große et al. (2017) for details). Annual average rates of atmospheric deposition of NO<sup>x</sup> and NH<sup>3</sup> are derived from data from the EMEP (Cooperative program for monitoring and evaluation of the long-range transmissions of air pollutants in Europe) model and long-term trends (Schöpp et al., 2003) as described in Große et al. (2016). A daily climatology of suspended particulate matter (Heath et al., 2002) is used to include its influence on the light climate.

For the second simulation [hereafter "WFD (reduction) scenario"], we apply the same forcing as described above, except for the riverine N loads. For the latter, reduction levels are derived based on the responses to a questionnaire sent off to the OSPAR contracting parties and asking how they want to fulfill the WFD requirement described within their national management plans. Some contracting parties reported reduction targets for both N and P, like France stating that "a strong nitrate reduction (about 50 %) should be necessary in many rivers, whereas phosphate should be significantly reduced (about 40 %) only in the Seine River". However, the current OSPAR assignment–as a first step–focuses on N reductions alone in order to allow for a stepwise approach, including feedback between

different markers indicate the different river groups: German (DE), Dutch (NL-1/NL-2), Belgian (BE), French (FR), British (UK-1/UK-2), Norwegian (NO) and other rivers ("Others"). The sub-regions of the German EEZ used for analysis are: "Inner Coastal (IC)," "Outer Coastal (OC)," and "Offshore (OF)." Adapted from Große et al. (2017), with permission of the copyright holders.

Lenhart and Große WFD Reductions and Trans-boundary Transports

the modeling community and OSPAR based on the first model results. This approach is supported by Emeis et al. (2015), who found that the N:P ratio in the Rhine River loads had increased from 23 to 62 between 1980 and 1992 due to the significant reductions in P but not in N (Claussen et al., 2009). This implies that future reduction measures should focus on N. Similarly, Lenhart et al. (2010) found that additional reductions in P are required only in the British rivers in order to achieve a 50 % reduction of P relative to the 1985 river loads (ICNS-2, 1988) by 2002. Therefore, further P reductions are considered a secondary objective within OSPAR. Consequently, this study focuses on the implementation of WFD-compliant riverine N load reductions. For the implementation of these reductions into the ECOHAM model, reductions for all affected N state variables need to be defined.

In its response to the questionnaire, Germany did not provide a percentage reduction but a target concentration of total nitrogen (TN), which implies reductions in both DIN and particulate organic N (PON). For the sake of consistency, this implies that PON reductions need to be applied to the rivers of the other contracting parties as well. Based on the assumption that reductions in DIN in the river basin will also result in reductions in PON, we use the same reduction as for DIN in these cases.

France and Belgium referred to the model results of the EMoSEM project (Desmit et al., 2015a, 2018), which combined hydrological and marine ecosystem models, and adopted DIN reductions of 50 % and 37 %, respectively. The Netherlands adopted a DIN reduction of 5 % based on a report from the Rhine Commission (ICBR, 2015), which takes into account the entire catchment and different types of N sources (e.g., agriculture or waste water) in the Rhine's neighboring countries. The United Kingdom did not adopt any N reduction, as they have only small localized "problem areas" in a few harbors and estuaries (OSPAR, 2017), and the effect of potential targeted reductions that address these sites could not be quantified.

The German "Bund–Länder Messprogramm" (BLMP, 2011) provided a target concentration for TN of 2.8 mg N L−<sup>1</sup> at the limnic-marine boundary for all German rivers entering the North Sea. In order to calculate the individual reductions in DIN and PON, we calculate the average DIN:PON ratios for the individual German rivers during the period 2006–2012, as agreed on with stakeholders from the German Federal Environmental Agency, using the above described daily river dataset. With that, we translate the TN target concentration into target concentrations for DIN and PON and calculate the reduction levels based on their 2006–2012 average concentrations, following Kerimoglu et al. (2018). As DIN:PON ratios vary between the different German rivers and throughout the seasonal cycle, we obtain different reductions levels for these rivers as well as for DIN and PON. The combined N reduction in the German rivers results in 28.5 %.

The resulting DIN and PON reductions for all contracting parties are provided in **Table 1**. It should be noted that these DIN reduction levels were presented on the HASEC meeting 2016 in Cork/Ireland and found the approval to be used as the basis for the ICG-EMO modeling activities related to the WFD measures. TABLE 1 | DIN and PON reduction levels for the WFD reduction scenario.


The PON reduction levels were presented on the HASEC meeting 2017, without approval nor rejection. For consistency among the different contracting parties we use both the DIN and the PON reductions in our scenario.

For the WFD reduction scenario, we only simulate the period 2000–2014, initialized with the results for January 1st, 2000, of the reference simulation. Both simulations run with a time step of 30 min and output is stored on a daily basis.

#### 2.3.2. TBNT Setup

In this study, we use the same TBNT post-processing software and setup as in Große et al. (2017), which showed good agreement with other TBNT studies (OSPAR, 2010; Painting et al., 2013; Troost et al., 2013). The full ECOHAM model domain is shown in **Figure 1**. The N tracing is conducted inside a subdomain, which is limited by the North Atlantic (NA) in the North, the English Channel (EC) in the Southwest and the Baltic Sea (BS) in the East (hereafter referred to as "TBNT domain"). Any N state variable that enters the TBNT domain across one of these boundaries is labeled accordingly during the calculation. This implies that N from rivers outside of the TBNT domain, which subsequently enters this domain, is attributed to the corresponding boundary. For the rivers inside the TBNT domain, we define 8 different source groups according to the standard adopted by the ICG-EMO community (ICG-EMO, 2009). All remaining rivers not included in these river groups are collected in a group of "other rivers". The input locations of the different river groups are indicated by the different markers in **Figure 1**. A detailed list of the individual rivers in each group is provided in Table 1 in Große et al. (2017). In addition, we trace the N inputs by atmospheric deposition into the TBNT domain.

The TBNT analysis is conducted for both the reference simulation and the WFD reduction scenario, using the daily ECOHAM output for the N cycle. For the analysis of the reference simulation, we apply a 7-year spin-up by re-running the year 1999. For each iteration, we use the ECOHAM output for the N cycle in 1999. The first iteration starts from an initial distribution with all mass attributed to the "other" rivers, while the following iterations start from the final distributions of the source-specific relative contributions of the previous iteration. With this procedure we achieve a quasi-steady state representing realistic distributions of the source-specific state variables within the TBNT domain at the beginning of the year 2000 (Große et al., 2017). Since the results at the end of the 7th iteration are qualitatively the same as those at the end of the 6th iteration, it can be concluded that a 6-year spin-up is sufficient to reach a quasi-steady state, independent of the the initial distribution. The resulting final distributions of the reference simulation are also used as initialization for the TBNT analysis of the WFD reduction scenario. For both simulations, we run the TBNT software for the years 2000–2014 and use the years 2006–2014 for our analysis. This guarantees sufficient time for the model system to reach a new quasi-steady state under the reduced riverine N loads. In addition, this analysis period corresponds to the latest OSPAR assessment period (OSPAR, 2017; Brockmann et al., 2018).

## 3. RESULTS

In the following, we first provide an overview of the effects of the WFD-compliant reductions on the actual riverine N loads, and second, how these reduced N loads effect the marine environment with respect to key parameters used within the OSPAR assessment of the eutrophication status. Thereafter, we present the results of the TBNT analyses of the two simulations with a special focus on the changes in the German EEZ, and on how the individual N reductions affect the N dynamics in the North Sea in general.

## 3.1. Changes in Riverine Nitrogen Loads Under a WFD-Compliant Riverine N Reduction

As a result of the strong N reductions in some of the North Sea tributaries (see **Table 1**) a significant overall reduction in riverine N loads into the North Sea can be expected. However, the strong differences between the reductions in the individual countries likely also result in changes in their relative contributions. In order to analyze both, **Figure 2A** presents the time series of annual riverine TN input into the model domain inside (solid lines) and outside of the TBNT domain (dotted line). In addition, we show the relative contributions of the individual river groups to the TN input into the TBNT domain for both the reference simulation and the WFD reduction scenario (**Figures 2B,C**, respectively).

In the reference run, the riverine TN input into the TBNT domain shows values of about 1,200–1,300 kt TN a−<sup>1</sup> from 2000 to 2002, before it drops to values of about 800–1,000 kt TN a−<sup>1</sup> during the period after 2002 (see **Figure 2**). In the WFD reduction scenario, the overall river input into the TBNT domain ranges between 700 and 900 kt TN a−<sup>1</sup> after 2002, which corresponds to an overall reduction of 14 % (excl. loads outside the TBNT domain).

The riverine N input into the region outside the TBNT domain does not change between the two simulations as the corresponding N loads (from UK, Ireland and Norway) are not reduced. The loads outside the TBNT domain range between 20 % and 25 % of the total loads (i.e., sum of loads inside and outside of the TBNT domain). Most of these loads originate from Irish rivers and rivers along the British west coast. Due to the generally northeastward circulation west of the British mainland (e.g., Otto and van Aken, 1996; Xing and Davies, 2001), it can be assumed that most of these N inputs are transported toward the northern boundary of the TBNT domain (NA; see **Figure 1**). Consequently, most of their N is lost via benthic denitrification before reaching this boundary, and thus only small amounts of N from these sources will actually enter the TBNT domain. In addition, this amount is unlikely to change between the reference simulation and the WFD scenario, as they are far away from any riverine sources with strong N reductions. Hence, the effect of their implicit inclusion in the open boundary sources (see section 2.3.2) on the study results is negligible.

In the reference simulation, the relative contributions of the individual sources are quite stable throughout the entire period and show only some variations (see **Figure 2B**), e.g., for the German (DE) and large Dutch rivers (NL-1; incl. Rhine and Meuse Rivers), and the rivers at the British east coast (UK-2). The highest contributions range between 20 % and 25 % with only the above named groups reaching these values. For the DE rivers, the flood events of 2002 (Ulbrich et al., 2003) and 2010 (Kienzler et al., 2015; Philipp et al., 2015) are clearly visible in high relative contributions. Usually, the NL-1 rivers account for the highest contribution, followed by the UK-2, DE and French rivers (FR).

The year-to-year variability in the relative contributions of the individual sources basically does not change under WFD reductions (see **Figure 2C**). However, the overall relative importance has increased for the British and Dutch rivers, due to their zero, respectively, low N reductions, now consistently constituting the highest contributions (NL-1 and UK-2). In other words, the relative contributions of those countries with only small N load reductions are amplified under a WFD reduction.

Although the overall decrease in riverine TN loads is likely to reduce N concentrations, and thus that of phytoplankton and Chl-a in the North Sea, the very different national reduction measures will likely result in regionally different responses to these reductions. Therefore, we now provide an overview of the reductions in TN, DIN, and Chl-a in the North Sea.

## 3.2. Reductions in OSPAR Assessment Parameters in Response to Riverine N Load Reductions

Since the riverine N reductions have the most direct effect on TN in the North Sea, **Figure 3A** shows the simulated TN concentration (i.e., sum of all pelagic N state variables) averaged over 2006–2014 and over the water column for the reference conditions. We choose this period as it corresponds to the latest OSPAR assessment period (OSPAR, 2017; Brockmann et al., 2018). The change in TN concentration simulated by WFD reduction scenario ("WFD") relative to the reference simulation is shown in **Figure 3B** and calculated as the difference between the result of the WFD scenario and that of the reference, divided by the latter. Accordingly, a negative change implies a reduction in the TN concentration. TN is only a voluntary assessment parameter within the OSPAR Common Procedure (OSPAR, 2017). Hence, we also show the analogous results for winter (January–February), water column averaged DIN (**Figures 3C,D**)

domain (dotted line), and the relative contributions from the different sources to the loads inside the TBNT domain for (B) the reference simulation and (C) the WFD reduction scenario.

and growing season averaged ("summer"; March–September), surface Chl-a (**Figures 3E,F**), which are mandatory assessment parameters (OSPAR, 2017). For Chl-a, the term "surface" refers to the uppermost model layer.

The TN distribution for the reference simulation (**Figure 3A**) shows that TN concentrations in major parts of the central and northern North Sea, and in the English Channel are less than 10 mmol N m−<sup>3</sup> . Only in the vicinity of major rivers, like the Rhine and Elbe Rivers, concentrations are significantly elevated and exceed 50 mmol N m−<sup>3</sup> (color scale limited to 50 mmol N m−<sup>3</sup> ) with maximum values of 105 mmol N m−<sup>3</sup> at the Elbe mouth. The high-TN signal of these riverine sources shows a gradual decrease toward the offshore regions of the North Sea, and it follows the general cyclonic circulation. In the Bay of Seine, values up to 45 mmol N m−<sup>3</sup> are simulated which do not extend far into the offshore English Channel, probably due to the strong tidal mixing in that region, which diminishes the signal.

The spatial patterns for winter DIN in the reference simulation (**Figure 3C**) are very similar to those in TN, with the only difference that concentrations in the open North Sea are slightly lower, and that they exceed the TN concentrations near the major rivers (e.g., 54 mmol N m−<sup>3</sup> in the Bay of Seine, 127 mmol N m−<sup>3</sup> at the Elbe River inlet), which relates to the different averaging periods. For the summer, surface Chl-a in the reference simulation (**Figure 3E**), values of above 1 mg Chla m−<sup>3</sup> are only simulated in the southern North Sea and along the British coast. Near the major rivers, values can reach or even exceed 10 mg Chl-a m−<sup>3</sup> . In the entire central and northern North Sea, and in most parts of the English Channel, Chl-a concentrations are less than 1 mg Chl-a m−<sup>3</sup> .

The relative changes between the reference simulation and the WFD reduction scenario exhibit qualitatively the same patterns for all three parameters (**Figures 3B,D,F**). The strongest reductions occur in the vicinity of the major rivers, to which significant N reductions were applied, namely the Belgian, French and German rivers. In their plume regions downstream the cyclonic circulation, further reductions are simulated. In contrast, no reductions occur in the entire western North Sea, due to the zero reductions in the British rivers.

The strongest TN reductions of up to 58 % are simulated in the Bay of Seine, while reductions near the Belgian and German rivers are on the order of 36 % (**Figure 3B**). The

values of "REF" and "WFD".

winter DIN reductions show a very similar response to the N load reductions with highest values of up to 64 % in the Bay of Seine and up to 35 % near the Belgian and German rivers (**Figure 3D**). For both TN and DIN clear reductions of up to 10 % can be seen in the plume regions of the French and German rivers, covering the entire eastern English Channel and wide parts of the southeastern North Sea.

For Chl-a, the reductions are generally lower (up to 29 % in the Bay of Seine) and locally confined to the French/Belgian and German/Danish coasts (**Figure 3F**), in response to the strong riverine N reductions in these regions. The two regions are separated by a region with very small reductions (< 4 %) off the Dutch coast, which is different to the reductions in TN and DIN. This implies that the N reductions in this region, which is strongly affected by the Rhine and Meuse Rivers, are too low to cause N limitation to surpass other limiting factors such as light or P limitation (Billen et al., 2011; Desmit et al., 2015b). Along the British south coast smaller changes do occur despite no reductions in the British rivers, which likely is a result of the strong French reductions.

An interesting feature occurs in the inner German Bight, where a region of only small changes in Chl-a extends northwestward from the Elbe River inlet, although changes in TN and winter DIN are significant. This could relate to the high turbidity in the Elbe River plume, causing light limitation to be the main controlling factor of primary productivity (Kerimoglu et al., 2018). Additionally, the negligence of riverine P load reductions in the WFD scenario could play a role here, as P limitation likely limits spring primary production in the coastal North Sea (Billen et al., 2011; Emeis et al., 2015).

The simulated changes in TN, DIN and Chl-a concentrations provide an overview of the potential changes in the North Sea in response to WFD reductions in riverine N loads. They further indicate that only small or no reductions in some North Sea regions likely result from zero or only small load reductions, e.g., in the Dutch rivers. In other regions (e.g., British south coast), reductions occur despite no reductions in the closest riverine sources, suggesting that N inputs from other rivers, to which reductions were applied, affect these regions. In the following, we therefore present spatial distributions of the relative contributions of selected riverine N sources to TN in the North Sea.

### 3.3. Relative Contributions to TN in the North Sea

In order to provide an overview of individual riverine contributions to TN in the North Sea, and their potential changes under WFD reductions, **Figure 4** shows the mass-weighted average relative contributions to TN of four selected river groups during 2006–2014: (**Figures 4A,B**) the German rivers (DE), (**Figures 4C,D**) the French rivers (FR), (**Figures 4E,F**) the large Dutch Rivers (NL-1; incl. Rhine and Meuse Rivers), and (**Figures 4G,H**) the rivers on the British east coast (UK-2). The panels on the left show the results for the reference simulation, while the right side shows those for the WFD reduction scenario.

For both simulations, the distribution maps show a typical point source characteristic, with very high values up to 100 % at the inlet and a strong decrease within the surrounding regions. The relative TN contributions can drop to 40–60 % within a distance of only about 100 km from the inlet. However, there is a far-field effect on TN concentrations in remote regions, e.g., along the Danish west coast in case of the NL-1 rivers.

The DE contribution is highest along the German and Danish coasts with contributions of more than 60 % up to 55◦N in the reference simulation (**Figure 4A**). In the WFD scenario (**Figure 4B**), the DE plume extends slightly less far north and west as a result of the comparably strong N reductions (see **Table 1**).

A very strong difference between the reference run and the WFD scenario can be seen for the FR contribution. In the former, it shows values of above 40 % in the entire eastern English Channel, and remains above 2.5 % almost until the northwestern tip of Denmark (see **Figure 4C**). In the WFD scenario, this signal is not visible and the contribution decreases from 10 % in the Strait of Dover to 2.5 % in the Southern Bight. In the English Channel, the FR contribution also dropped below 15 % in most regions, except near the inlets of the French rivers. This strong decrease in the FR contribution in the eastern English Channel explains the previously shown decrease in TN, DIN, and Chl-a along the British south coast.

Different to the DE and FR contributions, the changes in the NL-1 and UK-2 contributions between the reference simulation and the WFD scenario are very subtle, which relates to the 5% and zero reductions in their riverine N loads, respectively. In both simulations, the NL-1 contribution influences wide parts of the southeastern North Sea with values of above 5 % up to the northwestern tip of Denmark (see **Figures 4E,F**). The strong influence of the NL-1 rivers on TN also explains the weaker response in TN, DIN, and especially Chl-a directly off the Dutch coast. A slight increase in the NL-1 contribution occurs in the southeastern North Sea in response to the strong decrease in the DE contribution. Minor increases in the NL-1 contribution can further be seen in the Southern Bight and north of the Rhine and Meuse Rivers' inlets, which are induced by the reduction in the French and Belgian Rivers. The UK-2 contribution also extends far into the offshore regions of the North Sea in both simulations due to the cyclonic circulation (see **Figures 4G,H**). Highest values (> 75 %) occur near the inlets of Humber and Wash and the contributions stay above 5 % until the Danish northwest coast. As for the NL-1 rivers, minor increases in contribution relative to the reference simulation can be seen in the southeastern North Sea and in the Southern Bight.

The comparison of the spatial distributions of the relative contributions of the selected riverine sources provides a qualitative overview of the effects of WFD reductions on the individual sources in different North Sea regions. However, it does not allow for a detailed analysis of the simulated changes in the different North Sea regions. Therefore, we now present a quantitative analysis of the source-specific changes in the German EEZ to elucidate how the quite different N reductions levels (see **Table 1**) affect the TN concentrations in this region.

## 3.4. Source-Specific Contributions to TN in the German EEZ

For the quantitative analysis of the changes in TN and its sourcespecific contributions in the German EEZ, we consider three sub-regions according to the OSPAR COMP assessment (see **Figure 1**), based on observed salinity (S) gradients (OSPAR, 2017; Brockmann et al., 2018). The "Inner Coastal (IC)" region is characterized by S < 33, while the "Outer Coastal (OC)" region is defined by 33 ≤ S < 34.5. The "Offshore (OF)" region is the region with S ≥ 34.5. The results for the three regions are given in **Table 2** and are calculated as averages over the entire period 2006–2014 in order to provide a general picture of the changes imposed by the WFD reductions. In addition, we show the results for 2010, the year of a summer flood event in the Elbe River (Kienzler et al., 2015; Philipp et al., 2015), which also exhibited a record low in the winter NAO index (Osborn, 2010), reducing

FIGURE 4 | Maps of average relative contributions to TN by selected river source groups for (left) the reference simulation and (right) the WFD reduction scenario during 2006–2014: (A,B) German Rivers (DE), (C,D) French rivers (FR), (E,F) first group of Dutch rivers (incl. Rhine and Meuse; NL-1), and (G,H) rivers on British east coast (UK-2). Same color scale for all panels.


TABLE 2 | TN concentrations (in mmol N m−<sup>3</sup> ) and source-specific relative contributions (in %) in the different subregions of the German EEZ for the reference simulation ("REF") and the WFD reduction scenario ("WFD"), averaged over the entire period 2006–2014 and over 2010, respectively.

The individual Dutch and British contributions are collected in one group each. The smallest contributions (NO and "other" rivers, and BS) are also grouped together. Percentage sums > 100 % are due to rounding.

the Atlantic inflow into the North Sea (Winther and Johannessen, 2006). This is done to provide insight into the importance of natural variability compared to the changes induced by riverine N reductions. For a better overview, we combined the individual Dutch and British river groups into one group each and grouped together the very minor sources, namely the Norwegian (NO) and "other" rivers, and the Baltic Sea (BS). In addition to the relative contributions of the different source groups, we also show the TN concentrations (in mmol N m−<sup>3</sup> ) in each region. In the following, we first describe the changes over 2006–2014, and subsequently highlight some differences to 2010.

In the reference simulation ("REF"), the average TN concentration exhibits a steady decrease from 15.1 mmol N m−<sup>3</sup> in the IC region to 7.6 mmol N m−<sup>3</sup> in the OC region, and to 6.9 mmol N m−<sup>3</sup> in the OF region. This decline reflects the decreasing riverine influence toward the offshore regions of the North Sea. This is also illustrated by the decrease in the relative contributions from the German rivers from the IC (53.6 %) to the OC (8.5 %) and further to the OF region (1.5 %), as well as the increase in the North Atlantic contribution (from 10.2 % in IC to 51.1 % in OF). This implies that the German rivers dominate the dynamics in the IC region, while the North Atlantic is the main control in the OF region.

All other sources exhibit their highest contributions in the OC region, which means that this region is affected by the widest range of different sources. The Dutch rivers constitute the second largest riverine contribution in all subregions, with a maximum of 20.8 %. British rivers show a maximum contribution of 13.3 %. The Belgian and French rivers have the smallest contributions, with highest values of 1.5 % and 3.8 %, respectively. The atmospheric contribution reveals a comparably stable contribution of 12–16.9 % across the regions, while the English Channel contributes only 3–6.1 %.

The comparison of the results of the WFD scenario with those of the reference simulation shows that the TN concentration is reduced by 24.8 % in the IC region. In the OC and OF regions, the reductions only result in 11.3 % and 4.7 %, respectively, as a result of the generally lower influence of riverine sources. The strong reduction in the IC region is mostly due to the comparably strong reductions in the TN loads in the German rivers (see **Table 1**). The relative contribution of the German rivers to TN decreases from 53.6 % to 45 %, due to the only low or zero reductions in the Dutch and British rivers and the nonriverine N sources. This corresponds to a decrease in the absolute contribution (calculated as the product of the TN concentration and the relative contribution) from the German rivers by 36.8 % (not shown). This decrease in the absolute contribution clearly exceeds the actual combined N load reduction of 28.5 % in the German rivers, indicating additional N loss inside or west of the IC region due to changes in the N cycle. In the OC and OF regions, the reduction in the relative contributions from the German rivers is much weaker than in the IC region, due to their generally lower influence. However, their absolute contributions in both regions decrease by even 40 %. This suggests that changes in the biogeochemical cycling of N from the German rivers in the IC region (i.e., upstream with respect to the North Sea's cyclonic general circulation) have an additional indirect reduction effect on the German contribution in these regions.

The strong reductions in the Belgian and French rivers are also reflected in lower relative contributions to TN. Both contributions show their strongest decreases in the OC region (down to 1.0 % and to 1.6 %, respectively), where their contributions are highest in the reference simulation. Similar to the German rivers, the reductions in both groups' absolute contributions (Belgium: 40.6 % to 43 %, France: 61.8 % to 63.7 %) exceed the applied reductions in N loads notably (Belgium: 37 %, France: 50 %; see **Table 1**), especially for the French rivers.

In contrast, the relative contributions from the Dutch and British rivers consistently show a slight increase in the IC and OC regions due to their 5 % and zero N load reductions, respectively. For both groups, the strongest increases occur in the IC region, with a Dutch contribution of 13.7 % and a British one of 7.5 % in the WFD scenario. In the OF region, the British contribution also slightly increases by 0.4 %, while the Dutch one shows a minor decrease of 0.4 %. Interestingly, the absolute contributions of the Dutch rivers exceed 5 % in all subregions, ranging between 7.8 % in the OC and 11.3 % in the IC regions. The same holds for the British rivers, whose absolute contributions are reduced by up to 7 % in the IC region, despite no actual riverine N load reduction.

Regarding the non-riverine sources, which also did not change between the reference simulation and the WFD scenario, the atmospheric and North Atlantic contributions show consistently higher relative contributions in the WFD scenario. For the IC region, the same effect can be seen for the EC contribution. However, in the OC and OF regions, its relative contributions remain the same or decrease slightly. Similar to the riverine sources, the absolute contributions of these sources decreased in the WFD scenario (except for the North Atlantic in the OF region), with the strongest changes in the IC region (atmosphere and North Atlantic: 5.9 %, English Channel: 14.3 %). This shows that reductions in riverine N loads can have an indirect reduction effect even on non-manageable sources like adjacent seas.

The comparison of the 2010 values with those for 2006– 2014 shows that the flood event in the Elbe river caused a dramatic increase in the TN concentration in the IC region (23.8 % higher than the 2006-2014 average). This is also reflected in the much higher relative contribution of the German Rivers of 64.7 %. In the OC and OF regions, the TN concentration in 2010 is similar or even slightly lower than average. This can be attributed to the reduced North Atlantic contribution in all subregions as a result of the very low winter NAO, which resulted in significantly reduced inflow into the North Sea across its northern boundary. Accordingly, the contributions of most other N sources are higher in the OC and OF regions in 2010. Despite these significant differences in the relative contributions of the individual sources, the overall pattern in the changes of the relative contributions between the reference run and the WFD scenario remains the same as for 2006–2014. In addition, the changes in the absolute contributions of the individual sources between the two simulations (not shown) are almost the same as on average. This emphasizes the high potential of riverine N reductions to reduce TN levels in the German EEZ, independent of inter-annual variations in river load.

The excess reduction in the absolute contributions to TN in all subregions and for all–riverine and non-riverine–sources demonstrates that riverine N reductions cause an additional indirect reduction effect, which could play an important role for the long-term removal of N from the system. In the following we want to further elucidate what causes this excess N removal in the different North Sea regions, and how it affects the downstream regions.

#### 3.5. Changes in Source-Specific Benthic Denitrification

The consistent excess reduction in the source-specific absolute contributions to TN under reduced N loads compared to the reference conditions must result from a disproportionally higher loss of N from the North Sea system relative to the overall N inputs. Since both simulations use the same physical forcing, changes in lateral transport are proportional to the actual N load reductions and cannot explain this excess. As benthic denitrification (DNF) constitutes the only N loss term in the ECOHAM model, only changes in source-specific DNF in response to changes in the riverine N inputs can explain these excess N reductions. In order to analyze these changes, we calculated the 2006–2014 average ratios of source-specific DNF per unit of source-specific N river load (into the TBNT domain; DNF/RL) in different North Sea regions for both simulations and calculated the relative change in DNF/RL between the two. An increase in source-specific DNF/RL in a region implies a faster loss of one unit of N load from that source in that region or in other words a relatively higher loss of N under WFD reductions. As the river loads are lower under WFD reductions (or remained the same for the UK) and DNF itself decreases due to less organic matter availability (indicated by the reductions in Chl-a; see **Figure 3F**), such increase can only result from a weaker reduction in DNF relative to the N load reduction. The decreases in riverine N loads further imply that a decrease in sourcespecific DNF/RL in a region is caused by disproportionally less N from that source reaching that region, and thus a decrease in source-specific DNF.

**Figure 5A** shows the five selected North Sea subregions. As most of the riverine sources (except for a few British rivers) are located south of 57◦N, we only considered this part of the TBNT domain and subdivided the region in relation to the main circulation patterns illustrated by the spatial distributions of the relative contributions in **Figure 4**. The resulting changes in DNF/RL for the individual regions are shown in **Figures 5B–F**. We only display the changes for river groups with DNF/RL ≥ 0.01 (i.e., removal of at least 1 % of the overall riverine N load from that source) in the reference simulation, and black-framed bars indicate DNF/RL ≥ 0.1 (i.e., removal of at least 10 %) in the reference simulation.

In the central North Sea (region 1; **Figure 5B**) all river groups, except the rivers along the British east coast (UK-2), show a clear decrease in DNF/RL, i.e., relatively less N reaches this region in the WFD scenario. The strongest decreases occur for the German (DE; 13.5 %) and French rivers (FR; 25.3 %). For the UK-2 rivers, DNF/RL remains the same as in the reference simulation, as they are the only source group draining large amounts of N directly into that region and are not changed between the two simulations.

In the southwestern (SW) North Sea (region 2; **Figure 5C**), both the Dutch (NL-1/-2) and the British rivers (UK-1/-2) show slight increases in DNF/RL due to their very low and zero reductions, respectively. Hence, the amount of N from these sources is increased relative to the other riverine N sources in the WFD scenario, resulting in enhanced cycling of N and leading to increased N loss by DNF for these river groups. Similar to region 1, the French rivers exhibit a strong reduction in DNF/RL, due to less N from these rivers reaching the region.

In the southeastern (SE) North Sea (region 3; **Figure 5D**), DNF/RL is also clearly reduced for the French rivers, while the German rivers reveal a 14 % increase. The latter implies that under WFD reductions the N from German rivers is removed faster in this region, which explains the excess reductions in the German absolute contributions to TN in the German EEZ. It further explains the strong decrease in DNF/RL for the German rivers in region 1, as less N from these rivers reaches that region. The minor increases for the Dutch rivers relate to the 5 % reduction in their N loads in the WFD scenario and their vicinity to the region.

An interesting change in DNF/RL occurs in the English Channel (region 4; **Figure 5E**). Here, only the French rivers, the rivers along the British south coast (UK-1) and the "other" rivers exhibit ratios of DNF/RL > 0.01 as they are the only rivers draining directly into the region and all other rivers are located downstream with respect to the cyclonic circulation. For the French rivers, DNF/RL increases by 19.7 % as a result of the strong N load reduction and the non-linear response in primary production indicated by the smaller changes in Chl-a relative to TN (see **Figure 3**). Surprisingly, the UK-1 and "other" rivers' DNF/RL also increase by 4.6 % and 3.7 %, respectively. As their N loads have not changed, this must be caused by the strong French reduction and it implies that relatively more N from these sources is removed in that region in the WFD scenario.

The enhanced removal of N from the French rivers in the English Channel explains the reductions in DNF/RL for these rivers in all other regions, including the Southern Bight (region 5; **Figure 5F**), where their DNF/RL is 0.1 in the WFD scenario. Here, DNF/RL increases for all other relevant river groups, either as a result of direct N load reductions for the Belgian (BE) and NL-1 rivers, or as an indirect effect of reductions in these two groups and in the French rivers upstream.

In summary, for all river groups with strong N load reductions, relatively more N is removed by benthic denitrification in the regions, where they drain into the North Sea. This results from the non-linear response in phytoplankton growth to the N reductions near the inlets, e.g., due to additional P limitation (Billen et al., 2011; Emeis et al., 2015) or light limitation (Loebl et al., 2009). Consequently, the reduction in organic matter production is weaker than the actual N load reduction, which leads to a relatively higher N loss by benthic denitrification. As a result, relatively less N from these rivers reaches the downstream regions, which causes a relatively higher removal of N from other sources with higher N availability. This suggests that reductions in riverine N sources can significantly reduce the overall N availability even in remote regions due to a relative increase in N loss through benthic denitrification during the transit from the source to the region of assessment (e.g., reductions in UK-2 rivers would indirectly increase N removal of N from the North Atlantic).

## 4. DISCUSSION

## 4.1. Changes in Nitrogen and Chlorophyll-a in the North Sea in Response to WFD-Compliant Nitrogen Reductions

This study provides the first consistent approach to assess the potential impact of the combined national N reduction measures adopted by the OSPAR Contracting Parties to achieve the GES described under the WFD legislation. To our knowledge, it is also the first application of the TBNT method to a nutrient reduction scenario in order to analyze the impact of the individual reductions on the North Sea.

Although the overall riverine N input into the North Sea is reduced by 14 % under WFD reductions relative to the reference state, the different hydrographical regimes of the North Sea in combination with the wide range of national reduction measures (see **Table 1**) result in regionally very different responses to the riverine N reductions. It should be noted that the large range in the countries' N reductions might be amplified by our approach to apply identical reductions to PON and to DIN for countries that did not provide PON reduction targets (all except Germany). As primary production in river basins is usually P limited (e.g., Hecky and Kilham, 1988), reductions in PON loads are likely overestimated in our study. However, the average PON:TN ratios in the loads of the major rivers included in the model are below 0.15 during the simulation period, and only reach values up to 0.4 during summer when TN loads are generally low (not shown). Consequently, the amplifying effect on annually averaged TN and winter DIN can be considered small. With respect to Chl-a this effect might be stronger as PON accounts for a significant portion of TN loads during parts of the growing season. Though, a better estimation of actual PON reductions in response to reduced DIN would require the application of a catchment model, which represents the cycles of N and P, to the individual river basins. Analogously, such catchment model would be needed to estimate potential indirect reductions in the riverine P loads in response to N reductions in the river basin. However, both is beyond the scope of this study.

Compared to other North Sea modeling studies, which applied identical reductions to all riverine sources (Skogen et al., 2004; Lacroix et al., 2007; Lenhart et al., 2010; Wakelin et al., 2015), the strong differences between the individual N reductions of the member states result in a different model response. These studies predicted a general reduction of Chl-a concentrations or primary production along the continental and British coasts, in the southern and central North Sea, and in the eastern English Channel. In our study, significant reductions in TN, DIN, and Chl-a only occur in the southeastern North Sea, the Southern Bight and the eastern English Channel (see **Figure 3**) in response to strong N reductions in the German, Belgian and French rivers, respectively.

Due to the zero reductions in the British rivers under WFD, no reductions in TN, DIN and Chl-a concentrations are simulated along the British coast and wide parts of the western North Sea, where these rivers constitute the only major riverine source of N (see **Figure 4**; Große et al., 2017). In the Dutch coastal zone, the applied WFD reductions also cause only small reductions in TN and DIN and almost no reductions in Chl-a, due to the high Dutch contribution to N and a Dutch N load reduction of only 5 %, which prevents N limitation from exceeding P limitation (Billen et al., 2011; Desmit et al., 2015b) or light limitation (Loebl et al., 2009). This is supported by other TBNT studies (OSPAR, 2010; Painting et al., 2013), which found similarly high contributions of the British and Dutch tivers in these regions, respectively.

Our model also simulated only small reductions in Chl-a in a region extending northwestward from the Elbe River inlet, despite significant N reductions in the German rivers (28.5 % in annual TN load) and their dominant influence in that region. This suggests that light or P limitation surpass N limitation even under TN and winter DIN concentrations being about 25 % lower in that region in the WFD scenario. This is in agreement with Kerimoglu et al. (2018), who used a high-resolution model of the southern North Sea and also found light limitation to play a major role in this region. Other studies also identified P as the main limiting nutrient in the inner German Bight (Lenhart et al., 2010; Emeis et al., 2015; Wakelin et al., 2015), which suggests that additional P reductions might be required to achieve the "good environmental status" in the German Bight. The same might apply for the Dutch coastal waters, however, it cannot be concluded from this study, due to the only small TN and DIN reductions along the Dutch coast.

Despite the differences between the individual N reductions in this study and in previous studies, and the described regional differences, the ranges of the reduction levels simulated for DIN and Chl-a are comparable to other nutrient reduction studies (Skogen et al., 2004; Lacroix et al., 2007; Wakelin et al., 2015). Hence, we can consider carefully designed WFD reduction measures as a potent means to improve the eutrophication status of the southern North Sea.

In this context, it needs to be pointed out that the assessment of the effect of riverine nutrient load reductions adopted by the member states is only one step toward a comprehensive assessment according to the WFD. This is due to the fact that the success of reduction measures is assessed against so-called "pristine conditions," which describe a North Sea undisturbed by anthropogenic influences like elevated river nutrient loads. For the North Sea, these pristine conditions can be defined differently depending on the sources of information used and definition of the term "pristine". Pre-industrial conditions are often considered pristine due to the comparably small anthropogenic impact. They are usually defined as the status of the mid- or late 19th century (e.g., Serna et al., 2010; Kerimoglu et al., 2018), as reliable external nutrient inputs can only be dated back until then (Schöpp et al., 2003; Hirt et al., 2014). However, Desmit et al. (2018) derived "truly pristine" conditions before any anthropogenic disturbance using a catchment model for western Europe. The representation of the status of the marine environment under such historic conditions is usually achieved by combining these information obtained, e.g., from hydrological models (e.g., Gadegast and Venohr, 2015) with marine biogeochemical models (Desmit et al., 2018; Kerimoglu et al., 2018). Alternatively, N isotopes and sediment cores can be used to estimate nutrient inputs under undisturbed conditions, as demonstrated by Serna et al. (2010) for the German Bight.

However, assessing the effect of the WFD reductions against such a historic state is beyond the scope of this study, as our analyses focus on the assessment of the impact of individual N reductions from the different countries on the North Sea.

## 4.2. TBNT Analysis for the German EEZ and Implications for Eutrophication Management

The regionally very different responses in TN, DIN and Chl-a together with the relative contributions of individual riverine N sources and their changes under WFD reductions illustrate the importance of well-defined N (and possibly P) reductions for the improvement of the eutrophication status in the different North Sea regions. For the OSPAR assessment, the North Sea is subdivided into the different national EEZs, which are further subdivided, e.g., in relation to salinity gradients in the case of the German EEZ, taking into account the different regimes (coastal vs. offshore) in the different subregions.

Although the TBNT method was first published two decades ago (Ménesguen and Hoch, 1997), followed by a series of TBNT studies on the North Sea (e.g., Blauw et al., 2006; Lacroix et al., 2007; Painting et al., 2013; Troost et al., 2013; Dulière et al., 2017; Ménesguen et al., 2018), only few analyzed the contributions of the different N sources in the OSPAR assessment regions (OSPAR, 2010). Brockmann et al. (2018) state that the German Bight is affected by trans-boundary input of inorganic and organic nutrients, however, without quantifying these contributions. By analyzing in detail the individual contributions from the different riverine and nonriverine sources to TN in the German EEZ for the reference simulation, we address this topic and pursue the "source oriented approach" advocated by OSPAR (OSPAR, 1999). The additional analysis of the changes under WFD reductions further expands the work carried out by ICG-EMO (OSPAR, 2010), which was published recently in summarized form (OSPAR, 2017). Große et al. (2017) showed that the here applied setup consisting of the HAMSOM-ECOHAM model and the TBNT post-processing software is in good agreement with other TBNT studies (OSPAR, 2010; Painting et al., 2013; Troost et al., 2013) with respect to both riverine and non-riverine N sources. Hence, we consider the results of both the reference simulation and the WFD reductions as realistic representations of the recent and a potential future state.

Our results suggest that only the coastal zone of the German EEZ is dominated by the German rivers (53.6 % averaged over 2006–2014), while the regions farther offshore are strongly affected by the Dutch rivers, the rivers along the British east coast, and the North Atlantic (see **Table 2**). The French and Belgian rivers are only of minor importance in the entire EEZ. However, in the ICG-EMO study (OSPAR, 2010, 2017), which analyzed only the year 2002, the contribution of the French rivers (8 %) exceeds that of the British rivers (5 %), which questions the usefulness of TBNT analyses based on single years for management purposes. Our results for 2010, the year of a flood event in the Elbe River (Kienzler et al., 2015; Philipp et al., 2015), also show very different relative contributions than for 2006–2014 (e.g., 64.7 % for the German rivers in the coastal zone). Dulière et al. (2017) also demonstrated a similarly high year-to-year variability in the atmospheric contribution in the eastern English Channel and Southern Bight. This emphasizes the importance of considering long-term averages when assessing the impact of the individual N sources in a management context.

Due to the low reductions in the Dutch and British rivers, and the high German contribution to TN only in the coastal region, significant reductions in overall TN concentration (25 %) only occur in this region. For 2010, the reduction was only slightly higher (27 %), which suggests that these reductions in TN in the inner German Bight under WFD reductions are a robust estimate, despite the high year-to-year variability in the relative contributions. The discrepancy between the N load reductions of the individual OSPAR member states further results in a shift to higher relative contributions by the Dutch and British rivers in all subregions of the German EEZ, while those of the German, French and Belgian Rivers are significantly reduced. Accordingly, the contributions of the North Atlantic and the atmosphere increase.

Surprisingly, the absolute contributions to TN of all N sources-riverine and non-riverine-are reduced in the WFD scenario relative to the reference simulation. Moreover, the reductions in the absolute contributions of all riverine sources even exceed the actual N load reductions. Our analysis of the changes in benthic denitrification per riverine N load of the different sources (see **Figure 5**) show that under WFD reductions and relative to the total riverine N, benthic denitrification removes N from rivers with strong N reductions faster from the system than under recent conditions. This causes excess reductions in TN, which can be explained by non-linear responses in primary production to the N load reductions, e.g., due to the generally high light limitation in the river plumes of the coastal North Sea (Loebl et al., 2009) and possibly additional P limitation (Billen et al., 2011; Emeis et al., 2015).

Consequently, much less N from sources with strong riverine N load reductions reaches the downstream regions (with respect to the cyclonic circulation), which implies a relatively higher uptake of N from less reduced sources during primary production in these regions. This in turn enhances the loss of N from these sources via benthic denitrification. As a result even the contributions of N from the North Atlantic and the English Channel to TN are reduced by 5.9 % and 14.3 % in the coastal zone of the German EEZ in the WFD scenario. It should be noted that Große et al. (2017) found that benthic denitrification rates simulated by ECOHAM are up to 4 times higher than those reported by Marchant et al. (2016) in a few near shore locations of the German Bight, due to the simple sediment model used in this study. Consequently, the effect in the very nearshore regions might be overestimated. In addition, the release of legacy N stored in the North Sea sediments, e.g., in the German Bight (Serna et al., 2010), may partly balance the effect of N load reductions on benthic denitrification. Both would also reduce the indirect downstream effect, resulting in lower excess reductions. Therefore, we recommend a study on this effect using a more complex sediment model.

Nevertheless, these results clearly indicate that riverine N load reductions can have both a direct near-field and an indirect farfield effect on the reduction of the TN concentration in the North Sea. This suggests that N load reductions even in rivers distant from eutrophication problem areas can have a significant positive impact on the long-term removal of N from the North Sea and should be discussed in the context of WFD reduction measures.

## 4.3. The TBNT Analysis Within an OSPAR Context

The finding that reductions in riverine N sources could significantly reduce the overall N availability even in remote regions is important also within the OSPAR context, as it indicates that even small changes can have an effect on the entire North Sea system. Consequently, this should lead to a change in the treatment of the so-called "non-problem areas" defined within the OSPAR assessment (OSPAR, 2017).

As a result of the severe North Sea eutrophication in the 1980s, the 2nd International Conference on the Protection of the North Sea (ICNS-2) postulated in 1987 to "take effective national steps in order to reduce nutrient inputs into areas where these inputs are likely, directly or indirectly, to cause pollution" and to "aim to achieve a sustainable reduction (of the order of 50 %) in inputs of phosphorus and nitrogen to these areas between 1985 and 1995" (ICNS-2, 1988). In this statement, the term "these areas" was related to problem areas only, implying that only countries with problem areas had to take measures.

With the report on the "Distance to Target" assessment (OSPAR, 2013), which also includes TBNT components, the ICG-EMO group managed to bring forward a new perspective such that contributions from non-problem areas into problem areas should also be taken into account. In the executive summary, they stated "with respect to Eutrophication Problem Areas, all contributing Transboundary Nutrient Transport (TBNT) areas should be included in future modeling and assessment." However, the basic logic is that each member state still has to prove that these contributions are inspected.

In this context, we can claim that our TBNT study for the German EEZ has quantified the contribution from areas also with non-problem area status, like the United Kingdom. Here, it is worthwhile to note that our analysis revealed a multi-year average relative contribution to TN from the United Kingdom on the order of 6–13 % in comparison to only 5 % from the ICG-EMO study (OSPAR, 2010). Our results suggest that this contribution will further increase under the adopted WFD reduction measures. Hence, future reductions might be required in the British rivers if the "good environmental status" in the German EEZ and other parts of the southern North Sea cannot be reached under these measures. The same might apply to the Dutch rivers, which are the most important riverine N source for large parts of the southeastern North Sea.

## 5. CONCLUSIONS

This is the first representation of a WFD-compliant riverine N reduction scenario for the North Sea, which provides a consistent approach based on the combined national measures from OSPAR Contracting Parties under WFD. It furthermore constitutes the first detailed analysis on how changes in the individual riverine source groups affect the response in N and Chl-a to these reductions in the different North Sea regions.

By quantifying the relative contributions of different N sources to TN in the German EEZ under recent conditions and under WFD reductions, our study demonstrates that the TBNT method is a quantitative tool to put into practice the "source-oriented approach" advocated by OSPAR (OSPAR, 1999). Though, our study shows that in a management context, sufficiently long assessment periods need to be evaluated due to the high natural year-to-year variability in riverine N loads strongly affecting the relative contributions of the different N sources. In addition, multi-model studies are required in order to obtain an even more reliable assessment. Here, the good news is that the TBNT method is available for a number of North Sea ecosystem models. However, for the application in management and decision making an assessment framework is needed, either within OSPAR, or in the context of WFD or MSFD.

Our study also suggests that riverine N reductions have a direct near-field and an indirect far-field effect, caused by non-linear responses in NPP and thus N loss via benthic denitrification, which both result in excess reductions of TN in the marine environment. This indicates that riverine N reductions could be a potent means for the long-term removal of N from the North Sea system, and should be considered not only in countries with eutrophication problem areas. However, further studies with a more complex sediment model are recommended to better estimate these effects.

Besides this, there is still need for a better understanding of the balance between the different sources in the North Sea, e.g., under different environmental conditions. It could be worthwhile to conduct a TBNT analysis of the North Sea under pristine conditions. Such study should be based on a combination of hydrological and marine ecosystem models in order to also account for trans-boundary effects in the watersheds (i.e., across national borders). This would provide detailed insight in the natural balance between the different riverine and non-riverine nutrient sources in the North Sea and could provide a baseline for a future distribution of the relative contributions of the different North Sea tributaries and the natural sources like the North Atlantic. This could provide a valuable expansion of the description of the North Sea state under pristine conditions.

#### AUTHOR CONTRIBUTIONS

H-JL and FG contributed equally to the conception of the manuscript and the interpretation of the results. H-JL took the lead in writing the manuscript. FG developed the TBNT software, conducted all analyses and assisted in the writing.

#### FUNDING

This study received funding from the German Environmental Protection Agency (UBA), in the frame of the project Implementation of Descriptor 5 Eutrophication to the MSFD, SN: 3713225221.

#### ACKNOWLEDGMENTS

We thank Markus Kreus (Federal Waterways Engineering and Research Institute) and Johannes Pätsch (Universität Hamburg) for technical support with the model simulations. We would like to thank Sonja van Leeuwen (Royal Netherlands Institute for Sea Research) for providing freshwater discharge and nutrient load data for the major rivers across Europe. We further thank Jerzy Bartnicki (EMEP) for providing atmospheric N deposition data. We also thank Onur Kerimoglu (Helmholtz-Zentrum Geesthacht) for intensive discussions on the derivation and application of the river load reductions. We thank two reviewers for their constructive criticism, which helped improving the manuscript. We would like to thank Joel Graef and Sebastian Glaschke for technical support with the compilation of the figures. The model simulations and TBNT analyses were conducted as part of the dissertation by FG (Große, 2017) and were run on Mistral, the Atos bullx DLC B700 mainframe at the German Climate Computing Center (DKRZ) in Hamburg. This study benefited from the cmocean toolbox of Thyng et al. (2016). All figures were generated with MATLAB (MathWorks Inc., USA).

#### REFERENCES


ökologischen Zustand der Küstengewässer gemäßWasserrahmenrichtlinie. ARGE Bund Länder Messprogramm.


relevance of phytoplankton stoichiometry. Sci. Total Environ. 639, 1311–1323. doi: 10.1016/j.scitotenv.2018.05.215


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Lenhart and Große. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Decoupling Abundance and Biomass of Phytoplankton Communities Under Different Environmental Controls: A New Multi-Metric Index

Sorcha Ní Longphuirt<sup>1</sup> \*, Georgina McDermott<sup>2</sup> , Shane O'Boyle<sup>3</sup> , Robert Wilkes<sup>2</sup> and Dagmar Brigitte Stengel<sup>4</sup>

<sup>1</sup> Environmental Protection Agency, Cork, Ireland, <sup>2</sup> Environmental Protection Agency, Mayo, Ireland, <sup>3</sup> Environmental Protection Agency, Dublin, Ireland, <sup>4</sup> Botany and Plant Science, School of Natural Sciences, Ryan Institute for Environmental, Marine and Energy Research, National University of Ireland Galway, Galway, Ireland

#### Edited by:

Katherine Richardson, University of Copenhagen, Denmark

#### Reviewed by:

Erik Askov Mousing, Norwegian Institute of Marine Research (IMR), Norway Michelle Jillian Devlin, Centre for Environment, Fisheries and Aquaculture Science (CEFAS), United Kingdom

> \*Correspondence: Sorcha Ní Longphuirt s.nilongphuirt@epa.ie

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 October 2018 Accepted: 27 May 2019 Published: 13 June 2019

#### Citation:

Ní Longphuirt S, McDermott G, O'Boyle S, Wilkes R and Stengel DB (2019) Decoupling Abundance and Biomass of Phytoplankton Communities Under Different Environmental Controls: A New Multi-Metric Index. Front. Mar. Sci. 6:312. doi: 10.3389/fmars.2019.00312 Increased nutrient delivery to estuarine systems results in elevated growth of primary producers. This is evidenced by high chlorophyll concentrations and increased frequency of phytoplankton blooms. However, shifts in nutrient loads to estuarine ecosystems can also cause modifications in the structure of phytoplankton communities which can have adverse impacts right through the food web. Acknowledging these modifications is imperative if response mechanisms are to be fully understood. In this study, Ireland's current water framework directive (WFD) tool for determining the status of phytoplankton communities was built upon to encompass not only biomass and bloom frequency but also community structure (diversity and evenness) and abundance. This method allows for comparison with site- and date-specific environmental data which could give an indication of cause and effect relationships. The newly developed phytoplankton index performed well against current methods to determine ecological status. Furthermore, it had a better agreement with other physico-chemical and biological WFD parameters. Statistical analysis captured the relationship between the phytoplankton index and physicochemical parameters, allowing for a more detailed look at the impact of disturbance on the system. The inclusion of community structure acknowledged the imbalances in the phytoplankton communities of some systems even when frequent blooms are not evident. In bloom conditions, the disparity between the chlorophyll and abundance metrics within the phytoplankton index can be linked to winter dissolved inorganic nitrogen concentration and forms, temperature, and light conditions. Application of the phytoplankton index will allow not only for compliance with WFD requirements, but also a method for understanding and assessing ecosystem health of estuarine phytoplankton communities over spatial and temporal timelines in line with changes in physiochemical parameters.

Keywords: bloom, water framework directive, phytoplankton, estuarine, eutrophication

## INTRODUCTION

In a context of efforts to remediate estuarine and coastal systems impacted by anthropogenic pressures, understanding and quantifying the response of biological communities are essential. Phytoplankton communities are one of the first biological elements to respond to the detrimental impacts of nutrient enrichment in estuarine and coastal zones (Nixon, 1995). High levels of

phytoplankton biomass, due to the formation of blooms, can be detrimental to the health of estuarine ecosystems (Smayda, 2004) through a reduction in water quality and dissolved oxygen. This can create unsuitable conditions for the survival of flora and fauna.

Chlorophyll concentration, a proxy for phytoplankton biomass, is often used as an indicator of enrichment. However, the response of the phytoplankton community to nutrient loadings is not limited to chlorophyll concentration alone, but also includes structural changes related to the composition, abundance, frequency, and intensity of algal blooms. Alterations to any of these constituents can modify the energy supply and food quality that fuels production in food webs (Winder et al., 2017). In turn, this can impact on nutrient and energy fluxes, fisheries, aquaculture, and microbial processes (Houde and Rutherford, 1993; Bacher et al., 1998; Cloern et al., 2014).

In recent decades, anthropogenic activities have increased flows of nitrogen and phosphorus from land to surface waters. In a recent study of 18 Irish catchments 90% of all nitrogen entering estuarine and coastal systems emanated from diffuse (mainly agricultural) sources. Total oxidized nitrogen represents 71% of this load, while ammonia represents 3% (O'Boyle et al., 2016). Sources of phosphorus can vary considerably, with diffuse sources represent from 5 to 92% of all phosphorus entering the system (Ní Longphuirt et al., 2016). The N:P load ratio of nutrient sources and concentration ratio of downstream estuarine systems was directly related to the source of loads, with agricultural catchments having higher ratios. Chlorophyll concentrations in Irish estuarine and near-coastal systems are controlled by these nutrient inputs, but this response can be mitigated by factors such as light and residence time (O'Boyle et al., 2015). Analyses of the impact of these factors on the entire phytoplankton community structure (biomass, composition, and abundance) would deepen our understanding of the pressure–response relationship. The importance of structural changes in phytoplankton communities due to anthropogenic activities has been recognized by the water framework directive (WFD) through the inclusion of biological indicators. The ecological quality status of phytoplankton should be assessed using indicators of biomass, the frequency and duration of blooms, and the abundance and composition of phytoplankton data (see EC, 2000, Annex V, Tables 1.2.3, 1.2.4). The current Irish method for WFD monitoring of estuarine and coastal waters is a two-stage process consisting of an assessment of phytoplankton bloom frequency and biomass (EPA, 2006). This was developed to meet the requirements of the EU WFD (2000/60/EC) and National Regulations implementing the WFD (S.I. No. 722 of 2003) and National Regulations implementing the Nitrates Directive (S.I. No. 788 of 2005).

Changes in chlorophyll level and bloom frequency can represent a direct community response to nutrient enrichment in terms of increased primary productivity. The inclusion of abundance and community structure information in assessment approaches can complement these metrics by conveying additional insight into possible shifts in the composition of the community. In addition, a multi-metric index is often considered more robust than its component metrics (Lacouture et al., 2006). Oscillations in the relative concentrations of nutrients can potentially favor different species and species groups (Carlsson and Granéli, 1999; Bharathi et al., 2018). While preferences for different forms of nitrogen can also cause shifts in biodiversity (Glibert, 2017). Responses to nutrients should also be placed in the context of community responses to physical factors such as light, temperature, and residence time (O'Boyle et al., 2015; Cloern, 2018). The addition of community structure also allows the inclusion of heterotrophic species that are not always represented by chlorophyll measurements (Domingues et al., 2008).

To comply with the WFD, several multi-metric indices have been developed (Tett et al., 2008; Devlin et al., 2009; Giordani et al., 2009; Spatharis and Tsirtsis, 2010; Lugoli et al., 2012; Facca et al., 2014). The merits of each of these metrics are evident from their successful application in European systems. However, the metrics are often developed to encompass available datasets. For example, while some incorporate data on size class (Lugoli et al., 2012), others are predicated on having high frequency datasets (Tett et al., 2008; Devlin et al., 2009). Although it is recognized that high-frequency data are preferable (Ferreira et al., 2007), it is often difficult to reconcile adequate sampling effort in terms of spatial and temporal cover with reasonable costs (Garmendia et al., 2013). In situations where sampling frequencies are lower a more robust tool which can comply with WFD reporting requirements is required. The current study presents such a tool which is comparable with previous reporting tools while at the same time represents a method for better identifying responses to environmental pressures.

The proposed phytoplankton index takes elements of the original EPA blooming tool and the integrated phytoplankton index (IPI) developed by Spatharis and Tsirtsis (2010) to create a tool that will be compatible with current methods of estuarine waterbody classification (i.e., the EPA blooming tool), and with environmental and physical forcings in Irish estuarine and coastal waters.

The objectives of this study were (1) to create a phytoplankton index that encompasses all the structural components of the phytoplankton community and (2) to compare this phytoplankton index with corresponding environmental data to identify the parameters that impact on the phytoplankton community. The results of this study provide the basis for a detailed phytoplankton metric which could be incorporated into the reporting structure for the WFD.

#### MATERIALS AND METHODS

#### Data Sources and Sampling Methodologies

This study incorporated data from the EPA's Irish National Monitoring Programme from 2007 to 2016. Details of estuary types and the location of waterbodies can be found in O'Boyle et al. (2015) or viewed on the EPA geoportal website<sup>1</sup> (**Figure 1**).

Monitoring stations in each waterbody are, in general, sampled three times during the months of May–September and

<sup>1</sup>http://gis.epa.ie/

FIGURE 1 | Map of the WFD status of monitored Irish waterbodies (2007–2012). 1, Castletown Estuary; 2, Dundalk Bay Inner; 3, Dundalk Bay Outer; 4, Boyne Estuary; 5, Boyne Estuary Plume Zone; 7, Irish Sea Dublin; 6, Northwestern Irish Sea; 8, Rogerstown Estuary; 9, Broadmeadow Water; 10, Malahide Bay; 11, Liffey Estuary Upper; 12, Liffey Estuary Lower; 13, Dublin Bay; 14, Southwestern Irish Sea Killiney Bay; 15, Broad Lough; 16, Avoca Estuary; 17, North Slob Channels; 18, Slaney Estuary Lower; 19, Wexford Harbour; 20, Upper Barrow Estuary; 21, Nore Estuary; 22, Barrow Nore Estuary Upper; 23, New Ross Port; 24, Upper Suir Estuary; 25, Middle Suir Estuary; 26, Lower Suir Estuary; 27, Barrow Suir Nore Estuary; 28, Waterford Harbour; 29, Dungarvan Harbour; 30, Blackwater Estuary Lower; 31, Youghal Bay; 32, Owenacurra Estuary; 33, North Channel; 34, Lee Estuary Lower; 35, Lough Mahon; 36, Cork Harbour; 37, Cork Harbour Outer; 38, Bandon Estuary Upper; 39, Bandon Estuary Lower; 40, Kinsale Harbour; 41, Argideen Estuary; 42, Clonakilty Harbour; 43, Clonakilty Bay; 44, Ilen Estuary; 45, Roaring Water Bay; 46, Berehaven; 47, Bantry Bay; 48, Inner Kenmare River; 49, Kilmakilloge Harbour; 50, Outer Kenmare River; 51, Cahersiveen Estuary; 52, Valentia Harbour; 53, Portmagee Channel; 54, Tralee Lee Estuary; 55, Tralee Bay Inner; 56, Feale Estuary Upper; 57, Cashen; 58, Limerick Dock; 59, Fergus Estuary; 60, Upper Shannon Estuary; 61, Lower Shannon Estuary; 62, Deel Estuary; 63, Mouth of Shannon (Has 23;27); 64, Kinvara Bay; 65, Corrib Estuary; 66, Galway Bay North Inner; 67, Loch an tSaile; 68, Loch an aibhinn; 69, Loch Tanai; 70, Camus Bay; 71, Kilkieran Bay; 72, Erriff Estuary; 73, Killary Harbour; 74, Westport Bay; 75, Newport Bay; 76, Clew Bay Inner; 77, Clew Bay; 78, Broadhaven Bay; 79, Moy Estuary; 80, Killala Bay; 81, Garavogue Estuary; 82, Ballysadare Estuary; 83, Sligo Bay; 84, Erne Estuary; 85, Donegal Bay Inner; 86, Donegal Bay; 87, McSwines Bay; 88, Killybegs Harbour; 89, Gweebarra Bay; 90, Gweebarra Estuary; 91, Northwestern Atlantic Seaboard (HAs 37;38); 92, Mulroy Bay Broadwater; 93, Swilly Estuary; 94, Lough Swilly contains information © Ordnance Survey Ireland. All rights reserved. Licence Number EN 0059208.

once in the winter (January or February). Hence the sampling excludes naturally occurring spring and autumn blooms. This allows for a focus on the summer period when any growth exceedances would relate to excess nutrients entering the system. Surface and bottom water samples were collected for dissolved inorganic nitrogen (DIN) as nitrate, nitrite, and ammonia (NH4); molybdate reactive phosphorus (MRP); silicic acid (Si); and chlorophyll at each station. Nutrients were analyzed according to the Standard Methods for the Examination of Water and Waste Water<sup>2</sup> . Pigments were extracted using hot methanol (not corrected for the presence of pheopigments) and was measured using a spectrophotometer (Standing Committee of Analysts, 1980). A Hydrolab DS5X Multiparameter Data Sonde was used to measure salinity, pH, dissolved oxygen, and temperature in depth profiles at each station. Transparency was estimated using a Secchi disk and used to calculate the light attenuation coefficient (Kd) and photic depth (Zp) as follows:

$$\text{Kd} = 1.7/\text{Secchi depth.} \tag{1}$$

$$\text{Zp} = 4.61/\text{Kd.} \tag{2}$$

To determine if there was sufficient light available for phytoplankton growth the ratio of mixing depth (Zm) to photic depth (Zp) was calculated as Zm:Zp. Light limitation occurs when this ratio is greater than 5 or the eutrophic depth is <20% of the mixing depth (Cole and Cloern, 1984; Cloern, 1987). Specific details pertaining to sampling methodologies and calculations can be found in O'Boyle et al. (2015) and Ní Longphuirt et al. (2016).

To record phytoplankton abundance and community structure, the surface and bottom water samples taken for nutrients and chlorophyll were subsampled at each monitoring station. These individual samples were then mixed to give a whole waterbody sample. A subsample of the whole waterbody sample was taken in a 30-ml universal tube and preserved with Lugol's iodine. Cell counts were undertaken in 1 ml of sample on a Sedgewick Rafter Cell using a compound microscope. Cells were recorded to an appropriate taxonomic level and damaged cells were not counted as part of the analysis. The Sedgewick Rafter Cell has a limit of detection of 1,000 cells/l. It has been proven to provide accurate results between 10,000 (ICES 2006) and 100,000 cells/l (McAlice, 1971).

#### Phytoplankton Index Development

The current Irish method for WFD assessment of estuarine and coastal waters is the EPA blooming tool. This tool has been inter-calibrated with other tools developed by North East Atlantic countries (Carletti and Heiskanen, 2009). The tool contains a two-stage process consisting of the determination of phytoplankton bloom frequency and biomass. In the new phytoplankton index, these two metrics were combined with the metrics developed here for abundance and community structure. The relationship between the log-transformed individual metrics and pressures was determined using Spearman's rank coefficient (R platform). Once a relationship with pressures was established reference and class boundaries as per the WFD (high–good, good–moderate, moderate–poor, and poor–bad) for each of the metrics were identified.

#### Bloom Frequency

Bloom frequency was determined over the 6-year WFD cycle (four sampling occasions per year) through the analysis of taxonomic abundance of the dominant taxa (EPA, 2011). A bloom is considered to occur when the frequency of individual taxon exceeds 500,000 cells/l, at salinities of ≤17, or 250,000 cells/L for coastal waters of salinities above 17. Reference conditions are met if blooms are under a threshold of 2 for every 3 years and a high status is applied. A bloom every 2 years will place the waterbody at good status, while a bloom every year (or for 25% of sampled dates) will place the waterbody at moderate status. Ecological quality ratios (EQRs) were then calculated by dividing the reference values by the observed values.

#### Biomass

The median and 90th percentile chlorophyll concentrations were determined for each waterbody over a 6-year period (2007– 2012). The reference conditions and class boundaries are salinity dependent; for example, the reference conditions for fully saline waters are 3.33 mg l−<sup>1</sup> (Carletti and Heiskanen, 2009). This gives an EQR of 1 for any concentrations at or below this value (EPA, 2006). As class boundaries had already been developed for chlorophyll in the EPA blooming tool, these were carried over to the new phytoplankton index.

#### Abundance

Abundance can be considered a proxy for ecological disturbance as phytoplankton community growth is directly correlated with nutrient inputs to a system. A five-point scale was developed for abundance based on the full phytoplankton dataset available for Irish estuarine waters (**Table 1**). The median values for all estuaries that never exhibited a bloom and were also classed as "unpolluted" (no exceedances in nutrients or oxygen levels over a 6-year period) were considered reference sites (Government of Ireland, 2009).

The high–good boundary value was set as the reference value plus 50% of the reference value. The upper third quartile was considered the good–moderate boundary for estuaries with a salinity of more than 17, while the moderate–poor boundary was the boundary for bloom conditions (see above), as determined by the EPA blooming tool. The poor–bad boundary was the upper third quartile of all national datasets plus 1.5-times the interquartile range value [outliers were determined by the method developed by Tukey (1977) and used by Spatharis and Tsirtsis (2010)].

#### Community Structure

To identify structural changes in the phytoplankton community several ecological indices, which determine species richness, diversity, and evenness, were calculated for the dataset (2007–2012). These types of quantitative indices are favorable as they allow structural information about the community to be expressed as a single number (Tsirtsis and Karydis, 1998). Equations for the indices can be found in

<sup>2</sup>www.standardmethods.org

TABLE 1 | Reference and class boundaries developed for the abundance of dominant taxa at salinities above and below 17.


Boundaries are based on the 2007–2012 phytoplankton dataset for Irish Estuarine and Coastal waters.

TABLE 2 | Log regression analysis of the relationship between diversity and evenness indices and abundance of the dominant taxon.


The E2 index and Menhinick's index (MI) had the highest correlation with abundance and are indicated in bold.

Spatharis and Tsirtsis (2010). The results of these calculations were then separated based on salinity (i.e., above and below 17 salinity). After standardization the community indices tested were compared with the abundance of the dominant taxa to determine their monotonicity (consistent increase or decrease) with this metric (**Table 2**). Community evenness and diversity were negatively correlated to abundance, as cell numbers increase and dominant species prevail (Spatharis and Tsirtsis, 2010). The E2 index was the evenness index with the highest correlation with the log of abundance. Similarly, Menhinick's Index (MI) was used to represent species richness due to its strong correlation with the log of abundance (Spatharis and Tsirtsis, 2010; Nincevi ˇ c-Gladan ˇ et al., 2015). These two metrics complement each other as MI indicates the richness of the community, while the relative abundance of each species is considered with the evenness index. Hence, the two metrics combined were considered representative of community structure and were given a 0.5 weighting each for the multi-metric phytoplankton index. Because E2 and MI showed a relationship with the log abundance, the boundaries for these indices were calculated from the limits of the five-point scale developed for abundance using the equations in **Table 1**.

Finally, for each metric, boundary conditions were converted into a normalized EQR by first converting the data to a numerical scale between 0 and 1, where boundaries were not equidistant. These values were then transformed into an equal-width class scale between 0 and 1, where 1 is considered high (or reference) and zero is considered low (or poor) (**Table 1**).

The EQR values of the four date-specific metrics (abundance, chlorophyll, E2, MDI) were calculated for the waterbodies sampled (**Figure 2**). While the bloom frequency was calculated over 6 years, the new multi-metric phytoplankton index was calculated from the chlorophyll, bloom frequency, abundance, and combined evenness and diversity metrics. All five metrics can be used in the calculations if an EQR for a 6-year WFD period is required (**Figure 2**):

6-year multi-metric phytoplankton index

= Average {abundance EQR, chlorophyll EQR,

[average (E2, MDI)], bloom frequency EQR}. (3)

Only four metrics are used if only single waterbody and datespecific points are being considered (**Figure 2**):

Date and waterbody-specific multi-metric phytoplankton index

	- [Average (E2, MDI)]}. (4)

## Calibration of the Phytoplankton Index With the Current EPA Blooming Tool

To test the validity of the phytoplankton index the results were compared with the five assessment classes of the currently used EPA blooming tool (metrics: bloom frequency and chlorophyll) over the 2007–2012 period. The assessment classes produced were also compared with the overall WFD ecological status identified for each waterbody. The determination of overall WFD classification is a one-out-all-out system which includes data on various biological quality elements [i.e., phytoplankton, opportunistic macroalgae, macroalgal species richness, angiosperms (seagrass), benthic invertebrates, and fish], supporting quality elements including general physicochemical parameters (nutrients, biological oxygen demand, dissolved oxygen, temperature, salinity), and specific pollutants (EC, 2018). Finally, hydro-morphological risk is considered [see Government of Ireland (2009) for details of quality standards).

## Statistical Analysis of Driver–Response Relationships for the Phytoplankton Index

The response of the phytoplankton index to drivers was tested using statistical analyses techniques on the R statistical software platform (R Core Team, 2013). Following the guidelines of Feld et al. (2016) the dataset was checked for outliers and square root transformed before analyses was undertaken. Pairwise Pearson correlation coefficients for all variables were undertaken to assess co-linearity (package HMisc, Harrell, 2018). Subsequently, non-linear relationships were accounted for using the variance inflation factor (package usdm, Naimi, 2015). Once collinear varibales, as determined by Pearson correlation coefficients, were removed the relationship between physicochemical parameters [salinity, residence time, Zm:Zp, temperature, MRP, DIN (summer and winter) and N:P, TON:NH4, and Si] and the phytoplankton index was examined using random forest (RF) analysis (Elith et al., 2008). Winter nutrients were added as variables as they can be considered the nutrient concentration before biological uptake during the growth period; hence, they are a proxy for loadings which were unavailable (Desmit et al., 2015). This non-parametric regression method fits several models to bootstrapped data subsets, allowing the results to be tested against the observations not used in the models. The data were then split based on predictor thresholds (Breiman, 2001). Interactions among the explanatory variables were then obtained by ranking the deviance explained by individual predictors in R using gbm.interactions. Generalized linear models were then applied to the highest ranking variables. Selection of the best model, which incorporated the least descriptor variables to fit the data, was undertaken by comparing Akaike Information Criterion (AIC) (Burnham and Anderson, 2002). Model fitness was then tested using the ANOVA function in R.

## RESULTS

## Development and Testing of the Phytoplankton Index

The statistical analysis indicated significant correlations between the metrics chosen for the phytoplankton index and forcing parameters (**Table 3**). DIN, TON:NH4, MRP, and Si concentration gradients were strongly linked to abundance, MI, and chlorophyll. NH4 concentrations appeared to be strongly correlated to all metrics except abundance. Light conditions were linked to E2 suggesting the importance of light on the community dynamic. Residence time was also a factor which correlated strongly with abundance and chlorophyll. The analysis indicated that the metrics themselves were all correlated with each other; the correlation between E2 and abundance being the highest (**Table 3**).

The newly proposed phytoplankton index classified the status of Irish transitional and coastal waters based on the fivepoint WFD classification scheme. The phytoplankton index identified 28 waterbodies with "high" status, 45 waterbodies with "good" status, 14 "moderate," and 7 "poor" waterbodies. When comparing both status and individual EQRs the new phytoplankton index performed well against the current EPA blooming tool used to determine the ecological status of phytoplankton for the WFD (**Figures 3**, **4**). The two tools showed a linear correlation (R <sup>2</sup> = 0.85), while the newly developed phytoplankton index tended to give lower EQR values, particularly for the high-status waterbodies (**Figure 4**).

TABLE 3 | Spearman's rank correlation matric for the metrics chosen for the phytoplankton index (n = 1756).


MI, Menhinick's index; MRP, molybdate reactive phosphorus; DIN, dissolved inorganic nitrogen; Si, silicic acid; N:P, the molar ratio between DIN and MRP; Zm:Zp, the ratio of the mixing depth to the photic depth; RT, residence time; TON:NH4, the ratio of total oxidized nitrogen to ammonia; win, winter. <sup>∗</sup>P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001. Correlations in bold have a P < 0.001.

Considering the status of the 94 waterbodies analyzed with the new phytoplankton index, 45 had the same status assignment, 48 had a lower status, and 1 had a higher status (North Slobs: poor to moderate) than the original EPA Blooming Tool.

The new waterbody status was then compared with the overall WFD classification for the 2007–2012 period to determine if the new phytoplankton index would hypothetically alter the overall status of any of the waterbodies. The WFD status determination is a "one out all out" system; hence, the metric with the lowest classification will determine the overall status. The new phytoplankton index agreed with the overall WFD status in 39 cases, this improves upon the old tool which agreed with the WFD classification in only 25 cases. This suggests that the phytoplankton index agrees in more cases with the lowest biological or chemical metric in more waterbodies. In 45 waterbodies the new phytoplankton index designated a status that was higher than the overall status. Hence in waterbodies where the status was higher or equal to the current WFD status (84) the new phytoplankton index would not have altered the overall WFD classification for the 2007–2012 period.

In 10 cases the new phytoplankton index calculated a lower status for phytoplankton than the overall WFD status currently assigned (**Table 4**). This would indicate that if the new phytoplankton index was adopted these waterbodies would yield a lower status. An examination of the results from these systems showed different scenarios for coastal and transitional waterbodies. Five coastal systems (Broadhaven Bay, Valentia Harbour, Dundalk Bay Outer, McSwines Bay, and Donegal Bay Inner) were all classified as "high" for phytoplankton by the currently used EPA blooming tool and showed little or no disturbance to metrics for nutrients, oxygen, and other biological elements, while chlorophyll and phytoplankton counts and were, in general, low. However, in these systems, overall abundance and, more so, evenness and/or diversity, evidenced much lower EQR values. These lower EQRs resulted in the ecosystem classifications shifting from "high" to "good" in four systems and "high" to "moderate" in one system.

In four transitional waterbodies (Garavogue Estuary, North Channel, Dundalk Bay Inner, and Castletown Estuary) slight to moderate disturbances were registered by the original EPA blooming tool with waterbodies being classed as either "good" or "moderate." In these estuaries, the addition of the community structure metrics resulted in all systems dropping one classification on the scale. Although chlorophyll and/or bloom frequency EQRs did show disturbances much lower EQRs were evident for the community structure metrics. For example, in the Garavogue Estuary out of a total of 48 sampling points, 27 showed a difference between chlorophyll and abundance EQRs of over 0.4; with similar differences being observed between the chlorophyll EQR and diversity and evenness EQRs. Chaetoceros spp. (Hyalochaete), Cryptophyte spp., and Skeletonema spp. were the most prevalent species in this system and when dominant showed differences between the metrics of 0.78 ± 0.17, 0.49 ± 0.18, and 0.66 ± 0.39, respectively.

In three of these systems (North Channel, Dundalk Bay Inner, and Castletown Estuary) the original EQR was close to the boundary and so the additional metrics, while only dropping the EQR by a small amount, lead to a change of classification (**Table 4**). One system, the Middle Suir, showed a low chlorophyll EQR (0.25) but the bloom frequency EQR was higher (0.6), overall this resulted in a moderate status. The community structure metrics reinforced the chlorophyll EQR and resulted in a change in classification from "moderate" to "poor." This system is dominated by Coscinodiscus spp. that have large numbers of small chloroplasts (Hasle and Syvertsen, 1997) leading to higher chlorophyll concentrations relative to cell numbers.

Overall these results indicate that disturbances to the community were not always being picked up by the chlorophyll concentration and bloom frequency. Differences between EQRs for chlorophyll and abundance in most of the 10 systems (excluding the Middle Suir Estuary) highlight how high cell abundances may not always be reflected by the chlorophyll concentration measured.

#### Bloom Formation

The dataset was examined to identify which species were responsible for bloom formation, and concurrently, the consistency between the chlorophyll and abundance EQRs when the phytoplankton community was in bloom (**Figure 5**). A bloom was considered as a cell count of the dominant species of over 500,000 cells in salinities below 17 and over 250,000 in salinities above 17. There were 612 blooms recorded in the summers between 2007 and 2016 with over 56% dominated by five species;

Chaetoceros spp. (Hyalochaete) (20%), Asterionellopsis glacialis (14%), Skeletonema spp. (11%), Cryptophyte spp. (6%), and Cylindrotheca closterium (6%). These five species also dominated in periods when blooms were not recorded. In cases where the bloom was dominated by Pseudo-nitzschia > 5µm (mostly Pseudo-nitzschia delicatissima), Navicula < 10µm, Rhizosolenia setigera, Akashiwo sanguinea, and Karenia spp. (mostly Karenia mikimotoi), differences between the two metrics were consistently over 0.6. In 15 of the bloom forming species, including the dominant Chaetoceros spp. (Hyalochaete), Skeletonema spp., Cryptophyte spp., and A. glacialis, the difference between the abundance EQR and chlorophyll EQR was almost always above 0.4 (**Figure 5**), suggesting that at these times the chlorophyll values did not reflect the abundances present.

#### Pressure–Response Relationships

Four distinctive tests were carried out using R-based statistical analysis to determine firstly, the influence of environmental

drivers on the phytoplankton index of the entire dataset and during bloom periods, and secondly, the influence of drivers on the difference between chlorophyll and abundance EQRs for the entire dataset and during bloom periods. Following tests for collinearity (Pearson's rank coefficients followed by variance inflation factors) and interactions between drivers (RF) a reduced number of parameters were chosen to build generalized linear models (GLMs). Following this AIC identified the best model for each relationship. The final model fit for each relationship identified varying parameters which influenced the results (**Tables 5**, **6**). Hypotheses testing using ANOVAs in R suggested that all models were fit for purpose to a significance of P < 0.001.

The model results indicated that for the entire dataset the phytoplankton index was negatively influenced by temperature, winter DIN, and to a lesser extent by residence time and light availability (Zm:Zp). As these parameters increased the phytoplankton index decreased (**Table 5**). Increased N:P had a positive impact on the phytoplankton index.

Winter DIN concentrations and the TON:NH4 ratio negatively influenced the difference between the chlorophyll and abundance EQRs (**Table 5**). Hence, in waterbodies where these parameters were elevated the difference between the phytoplankton index metrics tended to be reduced. Concurrently, temperature, N:P, and Si were all higher as the disparity between the two metrics increased.

During bloom events, higher concentration of winter DIN negatively impacted the entire community structure, while temperature had a positive influence (**Table 6**). High N:P ratios and salinity were also positively correlated to the phytoplankton index during a bloom, albeit with a weaker significance. The discrepancy between chlorophyll and abundance EQRs during a bloom appeared to be negatively influenced by winter DIN, and to a lesser extent TON:NH4 ratios and light conditions. Increased light limitation (determined by Zm:Zp) reduced the difference between the metrics. Temperature positively impacted on the difference between the metrics, so higher temperatures resulted in greater differences (**Table 6**).

#### DISCUSSION

#### Phytoplankton Index Development

The multi-metric index developed in this study was proposed to provide an assessment of phytoplankton health in transitional


TABLE 4 | Irish transitional and coastal waterbodies with a lower overall WFD status when the new phytoplankton index is applied.

Bloom refers to bloom frequency and Abund refers to Abundance.

and near coastal water systems and, further, investigate the influence of physico-chemical parameters on the phytoplankton community. The Phytoplankton index incorporates chlorophyll concentrations and bloom frequency to allow comparisons with past evaluations and in addition proposed metrics for abundance, diversity, and evenness of the community. The correlation between nutrient pressure and the individual metrics in the index validates their inclusion in the phytoplankton index developed. Correlation between the metrics themselves is also observed, and expected. As abundance and chlorophyll rise, and the number of species drops dramatically, impacting on both on species richness and the evenness of the community (Tsirtsis and Karydis, 1998; Bužanci ˇ c et al., 2016 ´ ).

The results indicated that the inclusion of additional metrics for community structure led to a greater level of agreement between the phytoplankton status and overall WFD status assignment between 2007 and 2012. Increasing the number of metrics in a tool is considered more robust, while allowing more sensitivity to changes in the structure of the community

TABLE 5 | Results of generalized linear models produced from Irish transitional and coastal water data from 2007 to 2016.


The models show the relationship between (1) phytoplankton index and environmental predictors and (2) the difference between the chlorophyll and abundance EQRs and environmental predictors, for the entire dataset. Directional effect of each predictory coefficient (C) relative to the phytoplankton index are shown, along with the standard error (SE), the test (t), and significance (p) thereof. WinDIN, winter DIN; Temp, temperature; N:P, molar ratio of DIN to molybdate reactove phosphorus. RT, residence time; Zm:Zp, the ratio of the mixing depth to the photic depth. TON:NH4, the ratio of total oxidized nitrogen to ammonia; Si, silicic acid. P-value significance: ∗∗∗∗P < 0.001; ∗∗∗P < 0.01; ∗∗P < 0.05, <sup>∗</sup>P < 0.1. TABLE 6 | Results of generalized linear models produced from Irish transitional and coastal water data from 2007 to 2012.


The models show the relationship between (1) phytoplankton index and environmental predictors and (2) the difference between the chlorophyll and abundance EQRs and environmental predictors, during bloom events. Directional effect of each predictory coefficient (C) relative to the phytoplankton index are shown, along with the standard error (SE), the test (t), and significance (P) thereof. WinDIN, winter DIN; Temp, temperature; N:P, the molar ratio of DIN to MRP; Sal, salinity; TON:NH4, the ratio of total organic nitrogen to ammonia; Zm:Zp, the ratio of the mixing depth to the photic depth. P-value significance: ∗∗∗∗P < 0.001; ∗∗∗P < 0.01; ∗∗P < 0.05, <sup>∗</sup>P < 0.1.

(Garmendia et al., 2013). At the same time, quantifying different metrics of a community can help overcome the wide diversity of cells sizes and biochemical compositions which are found in the different taxonomical groups that can comprise a phytoplankton community (Litchman and Klausmeier, 2008). In several waterbodies the inclusion of the additional metrics resulted in a reduction in the EQR and in some cases a lower status assignment for phytoplankton. This lead to greater agreement with the other biological and chemical indicators that are used to determine waterbody status under the WFD and respond to nutrient enrichment, hence reflecting an overall disturbance to the ecosystem.

Lower EQRs for one or all the metrics for abundance, diversity, and evenness suggested structural imbalances in the phytoplankton community. In some coastal systems, this structural imbalance was not reflected in the overall chlorophyll concentrations or the number of blooms recorded (**Figures 3**, **4** and **Table 4**). These waterbodies have low nutrient concentrations and anthropogenic influences are considered low (EPA, unpublished data; Ní Longphuirt et al., 2016). As such they are classed as high or good status under the WFD, but would drop a status class if the new phytoplankton index was applied. The reason for the disparity may come from either (1) extremely low phytoplankton numbers or (2) the suitability of the analyses techniques. Diversity can increase to intermediate productivity and subsequently decrease at higher cell numbers. Hence, when phytoplankton are in very low numbers diversity and evenness EQRs will be low. The data in this study were obtained from a monitoring data which, at its conception, were aimed at recording bloom events and not overall community structure. The assessment procedure uses a Sedgewick Rafter Cell, which only contains 1 ml of a sample. This may be an insufficient volume for lower productivity systems. Future applications in coastal systems will require investigation of the different counting techniques to assess the most suitable method for use with the phytoplankton index.

In several systems, the inclusion of additional metrics resulted in a reduction of the overall EQR and associated status. These results evidenced the disparity between the metrics used to determine the phytoplankton index, specifically abundance and chlorophyll. As the phytoplankton respond to increasing resource availability the community structure will comprise cells of a variety of sizes and biovolumes, with larger cells contributing most (Cloern, 2018). It has been shown that (1) larger microplankton dominate during bloom periods in estuarine systems, (2) the biovolume of these larger cells is highly variable, and (3) that biovolume will increase during a bloom as larger cells accumulate (Irwin et al., 2006; Cloern, 2018). This could explain the variability in the metrics and further the variability in EQR differences during bloom events for each species. Additional information on cell size or biovolume along with cell number would be an important consideration for phytoplankton functional traits and their contribution to the community structure (Acevedo-Trejos et al., 2015).

In Irish estuarine and near coastal systems, diatoms were the taxonomic group that dominated the bloom events, which is similar in other temperate areas (Carstensen et al., 2015). Diatoms are well adapted to varying physical and chemical gradients present in near shore estuarine and coastal systems (Lomas and Glibert, 2000) and are considered of high food value for consumers (Winder et al., 2017).

Cryptophytes also played a role representing over 6% of all blooms, slightly higher than in other temperate areas (Carstensen et al., 2015). While the biovolumes of the dominant species in this study were not measured, a comparative study in the Bay of Brest indicated that Chaetoceros spp. (Hyalochaete), Skeletonema spp., and Cryptophyte spp. have on average biovolumes of 1540, 331, and 265 µm<sup>3</sup> cell−<sup>1</sup> , respectively (Klein et al., unpublished in pers comm review). The biovolume of C. closterium (Lower Slaney Estuary) and A. glacialis (Ballysadare Estuary) has been measured at 325 and 444 µm<sup>3</sup> cell−<sup>1</sup> , respectively. In comparison Coscinodiscus spp., which often dominant in the Suir estuarine system, are considered to have high biovolumes (ca. 8,586– 1,077,020 µm<sup>3</sup> cell−<sup>1</sup> ) (Olenina et al., 2006). The relatively large size of these cells could explain the reversed relationship between the chlorophyll EQR and the lower abundance and bloom EQRs (**Table 4**) in this system.

While a disparity between the different metrics contained in the phytoplankton index is apparent, and in some cases, sizable, their inclusion allows the phytoplankton index to consider multiple facets of the phytoplankton community structure and consolidate them into a single EQR. Furthermore, shifts in the relationship between the different metrics can allow for trait-based diagnostics of the community. For example, a higher chlorophyll EQR relative to the abundance EQR could indicate a shift toward smaller faster growing phytoplankton in a system. This can then be correlated with other biological and physico-chemical elements of the ecosystem to understand response trajectories to climatic, physical, and anthropogenic forcings.

## Phytoplankton Index Relationship With Environmental Factors

Pressure–impact relationships, a pre-requisite for ecologically meaningful indicators (Birk et al., 2012), were calculated at date and waterbody steps allowing an increased understanding of the relationship between anthropogenic pressures and physical constraints on Irish estuarine and coastal systems. The statistical analyses of the datasets indicated that the phytoplankton index responded significantly to several environmental pressures. Winter DIN, in this case considered a proxy for loadings of N to each system, had a negative relationship with the phytoplankton index over the entire dataset and concurrently when bloom events were considered alone. These results show a clear pressure–response relationship and validate the ability of a phytoplankton index, which considers not only enhanced biomass but also community structure, to respond to enrichment pressures. Concurrently, in periods where higher winter DIN concentrations were recorded the difference between the EQRs for abundance and chlorophyll was lower, indicating that the metrics were in greater agreement. As a blooming species continues to increase, its biomass peak is regulated by the nutrient supply (Chisholm, 1992), and the metrics used in

the phytoplankton index will converge as both abundance and chlorophyll surpass the upper boundary levels.

The availability of lower amounts of ammonia relative to total oxidized nitrogen appeared to improve the agreement between the abundance and chlorophyll metrics, but had no influence on the overall phytoplankton index outcomes. The preference for oxidized or chemically reduced forms of N can result in shifts in biodiversity, and while diatoms appear to preferentially use nitrates, cyanobacteria, chlorophytes, and dinoflagellates may be adapted to assimilate ammonia [see Glibert (2017) for review]. While biodiversity shifts at higher relative ammonia concentrations may have occurred, the statistically significant relationship could also relate to larger overall dissolved nitrogen in the system which is mostly made up of inorganic nitrogen forms.

The GLM results suggest an increase in the ratio of nitrogen to phosphorus limits phytoplankton growth. These results reinforce the classic paradigm that reduction of a limiting nutrient (in this case phosphorus) can lead to a decrease in phytoplankton growth (Schindler et al., 2008). However, high N:P ratios are also known to alter phytoplankton biodiversity and species composition due to competition between algae with direct optimal nutrient requirements (Collos et al., 2009; Glibert and Burkholder, 2011). While not tested here, this could explain the greater differences between cell numbers and chlorophyll concentrations at higher N:P stoichiometric ratios, due to a shift in the size structure of the phytoplankton communities. The expression of this difference in metrics suggests that higher ratios will have a deleterious impact on the phytoplankton community which may not be captured using chlorophyll concentration alone.

Increased temperature was related to lower EQRs for the phytoplankton index when the entire dataset was considered. As with the nutrients, this relationship is anticipated, as growth rate, and hence the possibility of bloom occurrence, is intrinsically linked to temperature (Sherman et al., 2016). The influence of temperature on the phytoplankton community is however multifaceted; higher temperatures appeared to result in better EQRs during a bloom and at the same time correlated with greater difference between the abundance and chlorophyll EQRs in all cases. This may relate to the influence of controlling factors such as zooplankton which graze on the primary producers. At higher temperature, the grazing rate of zooplankton can accelerate faster than the growth rate of large cells, thereby curtailing their relative abundance (Nixon et al., 2009; Cloern, 2018). Small cells may thus increase in proportion in higher temperatures, augmenting their influence on the community structure. Theoretically this could result in greater divergence between the abundance and chlorophyll metrics as smaller cells will contain lower amounts of chlorophyll.

Phytoplankton biomass in Irish estuarine systems can be modulated by light and/or residence time (O'Boyle et al., 2015). The results of the current study confirm this with both physical factors impacting on the phytoplankton index. Light was also shown to impact on the difference between metrics during bloom events. As with other components limitation by light can alter the size scaling of metabolic rates, resulting in a decrease in the size-scaling exponent (Finkel et al., 2010), thus leading possible changes in phytoplankton community size structure.

The opposing predictor coefficients of the individual stressors on the phytoplankton index indicate that their interaction is, most likely, antagonistic. For example, short residence times can dampen the impact of nutrient concentrations on phytoplankton biomass (Paerl et al., 2014; Hart et al., 2015) and promote faster growing and inherently smaller phytoplankton groups (Reynolds, 2006; Hart et al., 2015). In an Irish context, O'Boyle et al. (2015) found that chlorophyll median and 90th percentile concentrations did not always correlate with nutrient concentrations due to either short residence time and/or low light conditions. The correlation of the phytoplankton index with high DIN, light, and residence time suggests that a multi-metric index may capture the complex relationship between drivers, which are not necessarily shown if biomass measurements are considered alone.

It has been recognized that a greater understanding and recognition of responses to multiple pressures are required when determining programs of measures (Carvalho et al., 2019). The expression of phytoplankton community response to modulating parameters and pressures can result in changes in the size structure, chlorophyll content, and species dominance. Metrics for chlorophyll, abundance, and community structure will represent these changes in diverse ways, while at the same time providing insight into the response mechanism. Examining the relationship between chlorophyll and abundance becomes a proxy for identifying the response of the phytoplankton community to pressures such as light conditions and ratios of available nutrients. This in turn could help flag potential issues before a more comprehensive analysis of community structure is undertaken.

### CONCLUSION

Assessment tools which encompass both the structure and quantitative biomass response of phytoplankton communities are required to comply with the EU WFD and support management policies. The phytoplankton index proposed in this study appears to correlate well with existing methods and hence allows continuity and comparability with historic status classifications while concurrently improving the level of agreement between the status of Irish waterbodies and the overall WFD classification. Any future application of the phytoplankton index will need to consider the analytical methods used as the results appear to be influenced by the counting methods which may alter the assessment in low pressure species poor areas. The disparity between metrics shows their ability to represent different facets of the phytoplankton community structure and, in their amalgamation, a more holistic representation of the response to pressures can be portrayed.

The meta-analyses applied in this study validated the phytoplankton index through the statistically significant relationships between drivers and modulators of phytoplankton community structure and general community health. In addition, they reinforced the idea that changes in community structure, species composition, and cell size should also be considered

alongside chlorophyll and abundance when considering the impact of anthropogenic forcings.

#### AUTHOR CONTRIBUTIONS

SNL collated the data, undertook statistical analysis, and wrote the manuscript. GM undertook counting of phytoplankton, development of initial tool, and expert advise in phytoplankton species. RW inputted significantly to the text and provided

### REFERENCES


**Figures 1**, **2**. SO'B helped with initial tool conceptualization and reviewed the text. DS provided guidance throughout the project and reviewed the text and provided comment.

## FUNDING

This research was fully funded under the EPA STRIVE Research Call, grant number 2012-W-FS-9. The publication fees will be provided by the EPA.

dinoflagellate in Thau Lagoon, southern France. J. Sea Res. 61, 68–75. doi: 10.1016/j.seares.2008.05.008



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Ní Longphuirt, McDermott, O'Boyle, Wilkes and Stengel. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Interannual Improvement in Sea Lettuce Blooms in an Agricultural Catchment

#### Joseph V. McGovern1,2,3 \*, Stephen Nash2,3,4 and Michael Hartnett2,3,4

<sup>1</sup> Oceanographic Services, Ocean Science and Information Services, Marine Institute, Galway, Ireland, <sup>2</sup> Department of Civil Engineering, National University of Ireland Galway, Galway, Ireland, <sup>3</sup> Ryan Institute for Environmental, Marine and Energy Research, National University of Ireland Galway, Galway, Ireland, <sup>4</sup> MaREI Centre for Marine and Renewable Energy Ireland, National University of Ireland Galway, Galway, Ireland

Riverine nutrient loading from agriculture is one of the most prominent pressures in the second cycle of river basin management planning for the European Union (EU) Water Framework Directive (WFD). Better farmyard nutrient management planning is the measure most likely to reduce agricultural nutrient loading to catchment watercourses and coastal receiving waters. The adjoining Argideen Estuary and Courtmacsherry Bay in the south west of Ireland drain a 150 km<sup>2</sup> catchment comprising mainly agricultural land. The receiving waters were allocated Poor ecological status under the WFD at the most recent appraisal. Sub-hourly water quality monitoring in the Timoleague River has been carried out by the Teagasc Agricultural Catchments Program to track the changes in nutrient loading to the Argideen Estuary in response to improved farming practice. A bio-physical model of the adjoining Argideen Estuary and Courtmacsherry Bay was calibrated subject to the prevailing climatic and nutrient loading regime. Six nutrient load scenarios were simulated to determine their impact upon macroalgae and phytoplankton bloom magnitude. In addition, nutrient flow-load relationships were derived for summers 2010 and 2016 to elucidate the improvements induced by better catchment management practice. Flow-load relationships for dissolved inorganic nitrogen (DIN) and dissolved inorganic phosphorus (DIP) for each year were applied to the flow data for the other year, to query the outcome if (1) there had been no change in farm management practice between 2010 and 2016, or (2) the improved farm management practice in place by 2016 had been implemented by 2010. The difference between expediting and delaying improvement in farm nutrient management practice was a 5% increase in DIN loading and a 233% increase in DIP load. Application of this higher estimated load to the calibrated bio-physical model projected an increase in 2016 Ulva bloom magnitude from 381t to 1,391t. Although phosphorus retention within the catchment has improved in recent years, with an attendant improvement in Ulva bloom magnitudes, flow connectivity in the catchment still facilitates a higher phosphorus transfer during large rainfall events. A high amount of phosphorus is stored within the catchment, while point source pressures continue to contribute to phosphorus transfer to streams during periods of low flow.

#### Edited by:

Marianne Holmer, University of Southern Denmark, Denmark

#### Reviewed by:

Autumn Oczkowski, Environmental Protection Agency (EPA), United States Kapuli Gani Mohamed Thameemul Ansari, Indian Institute of Science Education and Research Kolkata, India

> \*Correspondence: Joseph V. McGovern joe.mcgovern@marine.ie

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 October 2018 Accepted: 05 February 2019 Published: 18 March 2019

#### Citation:

McGovern JV, Nash S and Hartnett M (2019) Interannual Improvement in Sea Lettuce Blooms in an Agricultural Catchment. Front. Mar. Sci. 6:64. doi: 10.3389/fmars.2019.00064

Keywords: macroalgae, biophysical model, agriculture, catchment management, flow-load relationship

## INTRODUCTION

fmars-06-00064 March 19, 2019 Time: 16:20 # 2

Opportunistic macroalgae blooms are a common indicator and response to nutrient enrichment, with effects including displacement of primary producers, increased frequency of eutrophication induced hypoxia and uncoupling of interactions between sediment and the water column (Valiela et al., 2003). In the presence of excessive nitrogen, blooms of macroalgae and phytoplankton can outcompete the very seagrasses that reflect Good ecological status (Nelson and Lee, 2001; Burkholder et al., 2007), due to differing compensation irradiance and rates of nitrogen uptake, growth and grazing (Duarte, 1995). A single macroalgae bloom event increases the likelihood of recurrence of bloom at a later time (Zhang et al., 2010). Varying tolerances to salinity extrema and nitrogen availability differentiate between green macroalgae of the genus Ectocarpus and Ulva (Fong et al., 1996). Ulva and the related species Ulvaria thrive in their respective intertidal and subtidal habitats due to differing tolerance to nitrogen, light availability and grazers (Nelson et al., 2008). Allelopathy, defined as the release of inhibitory or toxic compounds, is both an effect of Ulva upon other species of macroalgae (Van Alstyne et al., 2015) and oyster larvae (Green-Gavrielidis et al., 2018), and a mechanism for Ulva to outcompete Gracilaria (Gao et al., 2014). Ulva spp. also has a higher photosynthetic efficiency than other macroalgae species such as Gracilaria and Hypnea due to a higher ratio of surface area to volume reflected in its sheet-like structure (Whitehouse and Lapointe, 2015). Ulva species have a greater capability than other species to physiologically capitalize on the combination of higher temperatures, CO<sup>2</sup> enrichment and increased nitrogen loads anticipated due to climate change and ocean acidification (Parages et al., 2014; Chen et al., 2015). However, the higher growth rate of Ulva induces nitrogen limitation under exposure to higher CO<sup>2</sup> concentrations, which may lead to increasing competition from Gracilaria (Young and Gobler, 2017).

Across the European Union, diffuse and point source pressures account for 38 and 18% respectively of the significant pressures in the second WFD River Basin Management Plan cycle (European Environment Agency, 2018), with agriculture and urban wastewater accounting for the majority of each respective pressure. In the period from 2000 to 2013, Irish efforts to reduce nutrient transfer from catchments to estuarine and coastal waters in response to European environmental directives has yielded greater reductions in phosphorus than nitrogen (Ni Longphuirt et al., 2016), which may lead to greater transfer of nitrogen to the coastal zone. Intensification of dairy farming in response to removal of milk quotas may have a moderately deleterious effect on trophic status with increases of up to 20% in trophic status parameters (O'Boyle et al., 2017). The most recent analysis of nutrient loading to Irish surface waters reported that 82,000 tones of nitrogen and 2,700 tones of phosphorus is discharged per annum (Mockler et al., 2017). While nitrogen loading nationally is predominantly attributed to agriculture, phosphorus sources are more evenly distributed across the spectrum of land uses; the consequences of phosphorus discharge to watercourses are influenced by the presence or absence of hydrogeological attenuation. The strength of these attenuation processes, measured by hydrological and biogeochemical time lags may obscure the impact of measures implemented to improve water quality, with the delay between implementation of measures and confident detection of change ranging from 4 to 20 years (Melland et al., 2018). Interannual climate oscillations between wet and dry weather patterns may also offset or magnify improvements in water quality in response to better catchment management practice (Dupas et al., 2018).

The current Irish River Basin Management Plan (DHPLG, 2018) sets out to improve water quality through improving agricultural sustainability, recruitment of implementation teams to liaise with stakeholders, improvement in urban and domestic wastewater treatment efficiency and regulation of forestry and peat extraction; of the surface waterbodies identified as being at risk of not meeting their environmental objective, 53% are at risk due to agriculture.

In a selection of Irish estuarine waters, Ulva rigida was found to be the most dominant species (Wan et al., 2017); of the eight waterbodies classified using the WFD opportunistic macroalgal blooming tool (Wells et al., 2009), Courtmacsherry Bay and a neighboring bay were the only two waterbodies assessed as having Poor ecological status. Courtmacsherry Bay was also allocated Poor status in the most recent national assessment of water quality (Fanning et al., 2017). Agriculture is the predominant land use in Courtmacsherry Bay and the adjoining Argideen Estuary. In recent years, the Argideen Estuary has been blighted by the annual blooming of the opportunistic macroalgae Ulva rigida, with a total wet weight of up to 2,164 tones, corresponding to a mean density of 1,680 g m−<sup>2</sup> (Wan et al., 2017).

The advent of integrated catchment management in Ireland (Daly et al., 2016) requires overall nutrient load reduction targets for estuarine and coastal waters, as a starting point for identification of appropriate targeted measures at catchment level. Refinement of programs of measures is also necessary to target persistent nuisance opportunistic macroalgae. Biophysical box models such as the Dynamic Combined Phytoplankton and Macroalgae model (DCPM) (Aldridge et al., 2010) facilitate the identification of nutrient load targets, not least due to their numerical simplicity; short runtimes of the order of minutes for a year-long simulation are a significant benefit, compared to coupled 2D or 3D hydrodynamic, solute transport and water quality models. Reliable high-resolution nutrient loading data and river discharge data is a cornerstone of scenario modeling for identification of nutrient load targets. Monitoring data at hourly resolution has been made available for this study by the Irish Agricultural Catchments Program (ACP) (Shortle and Jordan, 2017) which is operated by Teagasc and funded by the Irish Department of Agriculture, Food and the Marine. Activities of the ACP focus on six catchments throughout Ireland, covering a variety of soil types and agricultural activities. One of the six catchments is the Timoleague, which drains to the Argideen Estuary. The ACP has provided daily rainfall measurements, hourly streamflow measurements and concentrations of dissolved inorganic nitrogen (DIN), total nitrogen, dissolved inorganic phosphorus (DIP) and total phosphorus. In an Irish context, the data has provided a rare opportunity to explore the impact of the timing and magnitude of

riverine nutrient loading to an estuary which has been allocated Poor ecological status in recent assessments, mainly attributed to the abundance of intertidal macroalgae.

Previous work by others in an Irish context has sought to quantify the reduction in nitrogen and phosphorus loading to eighteen estuaries over a 20-year period, and the influence of these reductions upon estuarine water quality (Ni Longphuirt et al., 2016). Individual case studies on the Blackwater and Argideen estuaries have set out to determine the influence of hydrological regime (Ni Longphuirt et al., 2015b) and long-term declining in nutrient loading (Ni Longphuirt et al., 2015a) on estuarine water quality. Projected intensification of the dairy sector has also been subject to a modelling study (O'Boyle et al., 2017). However, to the authors' knowledge, no studies have sought to quantify the outcome if there had been no improvement in nutrient management at farm scale. A sustained reduction in estuarine nutrient loads may be expected leading from measures directed with reducing diffuse pollution from agriculture. However, the fine balance in estuarine trophic cycles must be carefully considered when setting out nutrient reduction targets. Excessive changes to nutrient loads may lead to pollutant swapping resulting in under-assimilated nutrients being transferred to neighboring bays, while neighboring bays may be limited by different physical or biological factors irrespective of proximity (O'Boyle et al., 2015).

This study aims to identify the limiting factor for phytoplankton and macroalgae growth in the Argideen Estuary and Courtmacsherry Bay. A collection of scenarios will be appraised considering the identified limiting factor, to quantify the likely impact of nutrient load reductions upon opportunistic blooms. The European Union (Good Agricultural Practice for the Protection of Waters) Regulations, also known as the 'GAP Regulations' or the 'Nitrates Regulations,' were introduced to the Irish Statute books in 2006 (DOEHLG, 2006) to give effect to the European Union's Nitrates Directive (European Commission, 1991). Under the GAP Regulations, the Nitrates Action Program (NAP) was introduced to set restrictions on certain aspects of farm practice in order to improve water quality. The NAP has been reviewed and updated in the intervening years (DOEHLG, 2009, 2010, 2014). The final aim of this research is to assess whether an improvement in water quality in the Argideen Estuary and Courtmacsherry has been induced by changes in nutrient management practice in agriculture, endorsed under successive NAPs.

#### MATERIALS AND METHODS

#### Model

The Dynamic Combined Phytoplankton and Macroalgae (DCPM) model (Aldridge et al., 2013) is a biophysical box model, which describes the growth of phytoplankton and macroalgae in response to available light and system inputs of inorganic nitrogen and phosphorus. Chlorophyll a concentrations are modeled as a proxy for phytoplankton concentrations. The model solves a generic expression (Equation 1) (Aldridge et al., 2010) for the rate of change of solute (salinity, nitrogen, phosphorus, and chlorophyll a) and macroalgae, utilizing an adaptive, 4th order Runge-Kutta solution scheme with a daily timestep.

$$\frac{dY\_{\rm i}}{dt} = \frac{1}{V\_{\rm i}} (\phi\_{\rm i\dot{j}} + V\_{\rm i}\beta\_{\rm Y} + \Gamma\_{\rm Y}) \tag{1}$$

Here, Y<sup>i</sup> is the concentration of solute in box i, 8ij represents tidal exchange between box i and downstream box j (Equation 2) (Aldridge et al., 2010), β<sup>Y</sup> is the biological source-sink term (Equations 3 and 4) (Aldridge et al., 2010), 0<sup>Y</sup> is the solute loading term, E is the tidal exchange rate (d−<sup>1</sup> ), Y<sup>E</sup> and Y<sup>F</sup> are the tidally exchanged downstream concentration and freshwater concentration of solute Y.

$$
\phi\_{\text{i}\dagger} = E(Y\_{\text{E}} - Y\_{\text{i}}) + Q\_{\text{F}}(Y\_{\text{F}} - Y\_{\text{i}}) \tag{2}
$$

$$
\beta\_\mathrm{Y}[m] = (G\_\mathrm{m} - L\_\mathrm{m})X\_\mathrm{m} \tag{3}
$$

$$\beta\_{\rm Y}[nt] = \sum\_{\rm m} (G\_{\rm m} - e\_{\rm m}^{\rm nt} L\_{\rm m}) X\_{\rm m} \frac{1}{q\_{\rm m}^{\rm nt}} \tag{4}$$

G<sup>m</sup> refers to the net growth rate, L<sup>m</sup> is the net biological loss rate, X<sup>m</sup> is the net biomass concentration or weight, m refers to the algae type – chlorophyll a or macroalgae. e<sup>m</sup> nt is a grazing coefficient which reflects the recycling of decayed autotrophs into nutrients. The term q is the yield of phytoplankton or macroalgae per unit of each nutrient nt; thus the term q<sup>m</sup> nt joins Equation 3 and Equation 4 and therefore relates the abundance of primary producers to nutrient concentrations. The growth rate G<sup>m</sup> is determined as a product of the maximum daily growth rate Gmax (d−<sup>1</sup> ) and the minimum of three Monod-type expressions describing nitrogen, phosphorus and light limitation.

The model is underpinned by a user-tailored pro-forma spreadsheet, with each additional row in the spreadsheet detailing the next adjacent waterbody upstream. Each row contains the following user-defined parameters: latitude (◦N) and longitude ( ◦E), waterbody mean surface area between mean high water and mean low water (km<sup>2</sup> ), intertidal area (percent of mean area), spring tidal range (m), annual average river discharge (m<sup>3</sup> s −1 ), annual total loads for inorganic DIN and DIP (kg yr−<sup>1</sup> ), the ratios between summer and annual discharge RSA for freshwater, DIN and DIP, the winter and summer average offshore DIN and DIP concentrations (µM) and chlorophyll a concentrations (µg l−<sup>1</sup> ), salinity on the adjacent shelf (psu), exchange rate E (d−<sup>1</sup> ) and mean light attenuation K<sup>d</sup> (m−<sup>1</sup> ). Default parameters relating to phytoplankton and macroalgae physiology are iteratively modified as part of the model calibration process. Knowledge of bathymetry and waterbody surface area determines the waterbody volume; the bottom topography for each waterbody is defined in terms of the proportion of the waterbody area falling within twenty-two 1m wide depth bins between 10m above lowest astronomical tide (LAT) to 10 m below LAT. Water renewal throughout the system is represented by the exchange rate E (d−<sup>1</sup> ) for each waterbody, which represents the fraction of the volume exchanged with the adjacent downstream waterbody per day. Riverine discharge and nutrient loading is defined as an annual average with a ratio of summer to annual loading RSA to describe variation throughout the year. Calibration of the model

proceeds from the most coastal waterbody inward; calibration is initially completed for salinity, which is a proxy for mixing and dilution. Thereafter, modification of biological parameters is necessary to complete calibration. The model generates seasonal average concentrations of DIN (µM), DIP (µM), chlorophyll a (µg l−<sup>1</sup> ) and salinity (psu), as well as the following macroalgae parameters for summer: macroalgae density (g dry weight m−<sup>2</sup> ), over the entire waterbody area or intertidal area, and the total summer standing stock in tones (wet weight). Summer is defined as April to September inclusive, while winter is October to March inclusive. A full exposition of DCPM is beyond the scope of this publication; please refer to Aldridge et al. (2008, 2010, 2013) for further reading.

#### Study Site

Courtmacsherry Bay and the adjoining Argideen Estuary are situated in the southwest of Ireland. For the purposes of this research, the domain was divided into seven sub-sections of the Argideen Estuary/Courtmacsherry Bay system: Courtmacsherry Inner Bay, Courtmacsherry Bay North, the Lower Argideen Estuary, Flaxfort Strand, Courtmacsherry Intertidal Macroalgae, the Timoleague Receiving Waters and the Upper Argideen Estuary (**Figure 1**). Five rivers discharge to the Argideen Estuary-Courtmacsherry Bay system: (i) the East Cruary, (ii) the Timoleague, both of which enter the Timoleague Receiving Waters, (iii) the Argideen, (iv) the Flaxfort stream entering at Flaxfort Strand, and (v) the Kilbrittain entering at Courtmacsherry Bay North.

#### Datasets

This research exclusively uses river discharge and nutrient load data provided by the Teagasc Agricultural Catchment Program (ACP) for the Timoleague River for the years 2010 and 2016. Years 2010 and 2016 have been selected primarily due to their hydrological similarity. Summer, winter and annual total and average rainfall for both years are all but identical (**Table 1**), while the highest daily rainfall in each year occurred in July (**Figures 2A,B**).

Due to the similarity between the five river sub-catchments in terms of catchment slope, subsoil, groundwater connectivity and practices, freshwater discharges and their associated DIN and DIP loads for the Rivers Argideen (135 km<sup>2</sup> ), East Cruary (9.7 km<sup>2</sup> ), Flaxfort (5.4 km<sup>2</sup> ) and Kilbrittain (23.2 km<sup>2</sup> ) have been scaled up from area-normalized data derived from the ACP data for the Timoleague catchment (5.2 km<sup>2</sup> ) (**Figure 3**).

While the availability of hourly nutrient and flow data for the Timoleague Estuary may reduce the uncertainty relating to freshwater nutrient loadings, additional uncertainty may pertain to the extrapolation of the nutrient and flow data from the Timoleague catchment to the other four river systems that enter the Argideen Estuary/Courtmacsherry Bay system. However, the catchment descriptors amongst the five catchments are similar (**Supplementary Table S1**). In addition, the 2018 CORINE landuse maps indicate that pasture is the most common land-use in all river sub-catchments (**Supplementary Figure S1**). In addition to diffuse nutrient loading from upstream catchments, nutrient loading from point discharge of wastewater effluent was also included; information from the existing wastewater treatment at Courtmacsherry and conveyancing at Timoleague was collated to reflect the outgoing nutrient loads at both locations.

The Irish Environmental Protection Agency sample estuarine and coastal waters three times during summer and once during winter for salinity, DIN (nitrate, nitrite and ammonia), chlorophyll a, DIP, dissolved oxygen, silicate and Secchi depths. Water quality sampling is carried out on both the ebb and flood tide where possible. Macroalgae surveys are carried out once annually at a selection of waterbodies impacted by opportunistic blooms. In the context of the research discussed here, two patches are surveyed within the model domain, in the Argideen Estuary and at Flaxfort Strand.

The net summer phosphorus mass transfer to the system was greater in 2016 than 2010 (**Table 2** and **Figure 4**). Likewise, the summer discharge Q<sup>50</sup> from the Timoleague River was greater in 2016 than 2010 (**Figure 4**). The summer flow exceedance curves indicate that the entire duration of summer 2010 was dryer than an average year, or indeed the year 2016, but with higher flows at lower percentiles (**Figure 4**). Hence, higher than Q<sup>50</sup> flows in 2016 delivered a sustained higher phosphorus transfer (**Table 2**).

The relationships between total DIN loading per day and total discharge from the Timoleague River for 2010 and 2016 appear relatively stable (**Figure 5A**). There has been a substantial difference in the total DIP loading per day throughout the range of long-term flow percentiles in the Timoleague River for 2010 and 2016 (**Figure 5B**), with the lowest daily orthophosphate loading in 2010 (1.39 kg DIP at Q95) exceeding the largest daily orthophosphate loading in 2016 (1.24 kg DIP at Q05). A change in the relationship between flow in the Timoleague River and the molar N: P ratio of nutrient loading to the Argideen Estuary from the Timoleague River is also apparent (**Figure 5C**); while in 2010, the relationship indicates a gradual increase in the molar N: P ratio with increasing flow from Q<sup>95</sup> to Q05, in 2016, the molar N: P ratio peaked between Q<sup>20</sup> and Q30, with the ratio decreasing at higher flows. The implication of this trend is that, at higher flows, estuarine phosphorus limitation that may be induced by excessive nitrate between Q<sup>95</sup> and Q<sup>30</sup> is offset by increasing DIP loading relative to DIN. The consequence of this sequence of rainfall events would be higher potential growth of primary producers after higher rainfall.

There is a considerable difference between DIN, DIP and chlorophyll a concentrations at the tidal boundary for both summers (**Table 3**). The molar N: P ratio varies from 51 in winter to 11 in summer 2010, while for 2016, the ratio is 19 in winter and 25 during summer. Courtmacsherry Bay and the Argideen Estuary are subject to semidiurnal tides. Given the volume that is exchanged on a daily basis during the tidal ebb and flood, the additional coastal P underlying lower molar N: P ratios has been posited as a major factor in sustaining macroalgae in the Argideen Estuary (Ni Longphuirt et al., 2015b) due to the abundance of N in the coastal system.

## RESULTS

### Model Calibration

The calculated RSA was left unchanged throughout model calibration, with the daily tidal exchange value, E, being the

TABLE 1 | Total and average rainfall, by season and per year, for 2010 and 2016.


sole parameter tuned to calibrate physical exchange. Nutrient and freshwater loads were extrapolated from the Timoleague catchment to the other four rivers assuming the RSA and loadings per unit catchment area transferred from 1 catchment to the others.

Daily exchange rate E is related to residence time τ<sup>r</sup> by Equation 5 which is given by Aldridge et al. (2013), as follows:

$$E = -\log(1 - 1/\tau\_{\rm r})\tag{5}$$

Residence time is an expression of the length of time necessary to entirely flush an initial concentration of conservative tracer from a waterbody or part thereof, subject to the influence of riverine discharge and variation in tidal elevation. Equation 6, as proposed by Hartnett et al. (2011), was utilized to determine an initial estimate of residence time τ'<sup>r</sup> based on the length of a waterbody (L) in km, the width of the mouth of the waterbody (B0) in km, and the dimensionless tidal prism ratio (TPR).

$$\mathbf{r}'\_{\mathbf{r}} = \frac{8.65}{TPR} - 2.45B\_0 + 0.59L - 5.05 \tag{6}$$

The final exchange rates, which were identified from tuning the initial value determined using Equations 1 and 2, fell within a narrow range for each waterbody (**Supplementary Table S2**). The greater variation in values between the 2 years occurred further upstream to reflect greater exchange due to higher flushing because of higher freshwater flows.

Following calibration of the physical exchange between adjacent waterbodies, the model was tuned to match as closely as possible the seasonal average concentrations of DIN, DIP and chlorophyll a.

Site specific light attenuation coefficient values K<sup>d</sup> were determined for each waterbody and year using Secchi depth readings together with separate equations given by Devlin

et al. (2008) for estuarine and coastal waters. Using all sitespecific values for parameters such as exchange rate and seasonal ratios for river discharge and nutrient loading brought all model observations to the right order of magnitude. The main differences between the final calibrated models were the macroalgae turnover rate and the wet weight to dry weight ratio (**Supplementary Table S3**). The only other differences between the two model configurations were the nutrient loads, tidal boundary conditions, the exchange rates (**Supplementary Table S2**) and the river discharge.

There is generally good agreement between the model simulated averages and seasonal average observations for summer 2010 and 2016, with respect to nutrients, phytoplankton and macroalgae (**Table 4**); the same gradient observed in monitoring data throughout the continuum from freshwaters to the coastal zone is replicated in the equivalent results generated by the model. It is important to emphasize that the seasonal average concentrations of DIN, DIP and chlorophyll a generated by the model are based on daily resolution outputs over the full summer from April to September inclusive, while the EPA monitoring data is collected at three points during the summer. The concentration averages determined by the model are based on volumes estimated based on from Admiralty charts. As concentrations are determined as mass per unit volume, the consequence of under or overestimation of volume must be kept in mind when interpreting the percentage difference. A lack of knowledge with respect to how the river reaches and their adjacent flood banks interact under high flows also must be considered. Under or overestimation of waterbody volumes also influences the modeled exchange for each daily timestep. The waterbodies with the greatest discrepancy between modeling and monitoring averages are those with the greatest measured chlorophyll a concentration, possibly a reflection of nutrient limitation. Furthermore, the greatest deficit in terms of DIP averages lies in sections with significant growth of primary producers, which are all phosphorus limited.

For both years, chlorophyll a growth in the system is either light of phosphorus limited (**Supplementary Figure S2a**). In the Upper Argideen Estuary, which receive the greatest riverine inflow, chlorophyll a growth is light limited throughout the year, whereas all other waterbodies in the system switch from light to phosphorus limitation between April and September and revert back to light limitation from July onward. As macroalgae in the Argideen Estuary proliferates in the intertidal zone, and is thus exposed twice daily, macroalgae growth switches from light limitation in spring due to low irradiance, to phosphorus limitation in summer (**Supplementary Figure S2b**); the same observation was made for the system in previous works (Ni Longphuirt et al., 2015b).

TABLE 2 | River discharge and nutrient loading data for (A) 2010 and (B) 2016, and (C) effluent characteristics from the existing Courtmacsherry wastewater treatment which discharges to the lower Argideen Estuary.



FIGURE 5 | (A) Summer DIN flow-load relation, (B) Summer DIP flow-load relation, and (C) Summer molar N: P ratio vs. flow; flow values are long term summer flow percentiles Q05-Q95.

TABLE 3 | Tidal boundary concentrations of DIN, DIP, chlorophyll a and salinity for 2010 and 2016 based on the EPA monitoring data collected at station AR120 at the outer edge of inner Courtmacsherry Bay; values determine the extrema of a standard tidal boundary forcing curve which provides separate tidal boundary concentration values for each day.


TABLE 4 | Summer N, P, chlorophyll a and macroalgae observations in the Argideen Estuary/Courtmacsherry Bay system area compared to DCPM simulated summer averages after calibration for (A) 2010 and (B) 2016.


TABLE 5 | Scenarios considered in nutrient load scenario simulations.


#### Scenario Modeling

The changing relationship between daily DIP load and stream discharge between 2010 and 2016 implies that more DIP is being retained in the catchment, and thus not entering watercourses (**Figure 5B**). Additionally, model results point to a systemwide trend of phosphorus limitation during the peak primary production season (**Supplementary Figure S2**). Therefore, the scenarios which were considered are primarily focused on possible further decreases in DIP load (**Table 5**); scenarios 2 and 3 give an insight into the potential outcome if DIP loads continue to decrease during the summer. Although the majority of the system was phosphorus limited during both years, the outcome of a joint reduction in DIN and DIP has also been considered, hereafter referred to as a coupled reduction; scenarios 4 and 5 apply this case. Coupled reductions and DIP reduction alone have been considered in two steps of 33% of the initial diffuse nutrient loading.

Opportunistic macroalgae growth occurs in the intertidal zone in the Argideen Estuary. Consequently, the impact of a coupled decrease in the concentrations of tidally exchanged coastal shelf DIN and DIP have been considered in scenarios 6 and 7.

As previously noted, the years 2010 and 2016 are similar with respect to summer flow exceedance distributions and averages. For this reason, the 2010 flow-load relationship for DIN and DIP was applied to 2016 flows (scenario 8), and vice versa

(scenario 1), to establish the improvements in water quality due to better catchment management practice between 2010 and 2016. Scenario 1 represents the outcome if no effort had been made to improve nutrient management practice between 2010 and 2016. Scenario 8 considers the outcome if better nutrient management practices had been adopted sooner.

#### 2010 – Scenario 1

Scenario 1 considered the possibility of the 2016 flow-load relationships being valid in 2010. The net effect to the nutrient loads was an increase in the freshwater DIN loading of 3%, and a 30% increase in the freshwater DIP loading. A moderate 5% increase in summer chlorophyll a in the freshwater Upper Argideen Estuary (**Figure 6E**) and Timoleague Receiving Waters (**Figure 6D**) reflects the light limitation of these waters for the early and latter stages of the summer. A 40–50% increase in the DIP concentrations in both reflect a greater availability of DIP. This increase in DIP has a greater impact in the downstream Courtmacsherry Macroalgae zone (**Figure 6C**), where there is an observed increase of approximately 25% in the Ulva tonnage, 5% increase in the summer averaged chlorophyll a and 7–8% increase in the DIP concentration. System wide, there is a reduction in summer DIN concentrations (**Figures 6A–E**), which reflects a greater uptake of available DIN in primary producers; this would be the expected result of an increased supply of DIP to a phosphorus limited system, without a concomitant increase in DIN loading.

#### 2016 – Scenarios 2–7

Application of scenarios 2 and 3 resulted in an increase in DIN concentrations throughout the system of 2–17% (**Figures 7A–F**), with the response increasing in proportion to the DIP load reduction applied. An associated reduction in average

DIP concentrations of 4–66% throughout the system was also induced. Due to the phosphorus limitation of primary producers, summer average chlorophyll a concentrations reduced by 2– 66% in response to freshwater DIP load reduction, with the greatest effect observed in the freshwater reaches of the system. Macroalgae standing stock declined by up to 50% in response to a freshwater DIP load reduction of 66%.

Scenario 4 impacted chlorophyll a concentrations in the freshwater Timoleague receiving waters (**Figure 7D**) and Upper Argideen Estuary (**Figure 7E**) more than scenario 2; this reflects the delicate balance regarding nutrient limitation within the system. Scenario 5, however, had a similar impact throughout the system to scenario 3. No additional benefit with respect to Ulva bloom abatement was derived from scenarios 4 and 5, beyond the benefit observed with scenarios 2 and 3.

Scenarios 6 and 7 brought about a minor reduction in chlorophyll a concentrations, DIN and DIP in the brackish-saline reaches of the system, with the effect diminishing with distance from the tidal boundary.

#### 2016 – Scenario 8

Scenario 8 is a counterpoint scenario 1, which presented the opposite approach of applying the summer 2016 flow-load relationships for DIN and DIP to summer 2010 flows; scenario 8 considers the potential impact had there been no improvement in agricultural land management practice between 2010 and 2016. The implication of scenario 8 upon riverine nutrient loading to

the system is an increase in DIN load of 0.2% and an increase in DIP load of 134%.

One of the most notable observations is an increase of 25–185% in DIP in the Courtmacsherry Macroalgae Zone (**Figure 8C**), Timoleague Receiving Waters (**Figure 8D**) and Upper Argideen Estuary (**Figure 8E**), with excesses increasing with distance upstream; there is a moderate reduction in DIP of 0.5–2% in the remaining waterbodies. Due to the greater availability of DIP during the peak primary production period, there was an increase of 6–65% in summer averaged chlorophyll a concentrations, and an increase of 265% in summer Ulva standing stock. The increase in available DIP reflects an increase of 134% in summertime freshwater DIP loading upon baseline summer 2016 loading. DIN loading remained static in comparison to the baseline summer 2016 DIN loading. Therefore, the reduction in observed summer averaged DIN concentrations throughout the system (**Figure 8**) reflect a greater uptake of DIN by primary producers.

## DISCUSSION

A high-level overview of the summertime hydrological patterns of 2010 and 2016 reveal that each summer differs, although summers 2010 and 2016 are close to the 2010–2016 average discharge and flow exceedance curve. Therefore, the outcomes of the six nutrient load reduction scenarios considered in this research for summer 2016 may be interpreted as the potential outcomes for a standard summer.

DIN loading from the Timoleague catchment was linear with increasing rainfall/river flow for the years 2010 and 2016 (**Figure 5A**). DIP transfer with increasing river discharge across the flow percentiles was curved upward (**Figure 5B**), indicating

greater transfer with increasing precipitation. The curve shape is similar for years 2010 and 2016. The most notable observation in 2016 is the reduction in DIP transfer across the range of flow percentiles (**Figure 5B**). There has been a sustained improvement in the phosphorus balance in the Timoleague catchment since the Nitrates Action Plans (DOEHLG, 2006) were introduced in 2006, evidenced by a gradual reduction in phosphorus surplus and a convergence of soil phosphorus indices toward optimum values (Murphy et al., 2015). However, whilst the predominant source of phosphorus within the Timoleague catchment is storm runoff from agriculture, Shore et al. (2017) suggested that point source pressures such as domestic wastewater from one-off dwellings play a significant role in baseflow DIP concentrations; an increase in concentrations of total reactive phosphorus was observed during very low flow in Timoleague, due to baseflow point source pressures. In the same catchment, Jordan et al. (2012) also remarked upon significant phosphorus transfer from baseflow during summer.

In summer 2010, the molar ratio of nutrient loading throughout the flow percentile range was much closer to the Redfield ratio (Redfield, 1934) than 2016 (**Figure 5C**); on a daily basis in summer 2010, irrespective of the mass of nutrients delivered via watercourses, the nutrients were delivered in a ratio that would provide optimum conditions for growth of phytoplankton and macroalgae. In summer 2016, the equivalent curve indicates increasing phosphorus limitation with increasing flow, up to approximately Q<sup>10</sup> flows. Thereafter, greater rainfall and the resultant runoff would contain DIN and DIP concentrations closer to the optimum N: P Redfield ratio of 16: 1. However, when overall nutrient loads are considered (**Table 2**), the molar N: P ratio of nutrient loading to the system in 2010 and 2016 was 190 and 105 respectively. To unite both perspectives on the molar N: P ratio, one must consider the relative frequency of occurrence of low flows in 2016. Days of low flow in 2016 delivered nutrients in more favorable molar ratios of N: P; the infrequent wet days represented by the upper end of the x axis (**Figure 5C**), would have delivered excess DIN relative to DIP, thus inducing phosphorus limitation. Dupas et al. (2017) deduced from the apparent chemostasis in the Timoleague catchment that there was a high P store in the catchment. The same publication noted the same patterns in nutrient flow load curves, namely an inflection on the flow-nitrate curve above 6mm d−<sup>1</sup> rainfall, and the opposite upward flow-phosphate curve above 6 mm d−<sup>1</sup> rainfall. These patterns were linked to the combined influence of shallow and overland flow which connects regions of low nitrate and high soluble phosphorus, thus explaining a narrowing of the N: P ratio under high flow conditions. The author cited the potential cause of low nitrogen transfer being a better retention of nitrogen in grasslands such as Timoleague due to longer growth phase of grass than other land use activities, whereas the risk of P mobilization increases in grassland (Haygarth et al., 1998).

Changes in the nutrient flow-load relations from 2010 to 2016 imply an improvement in the prevailing conditions in the Timoleague catchment for an average year (2010 versus 2016), potentially due to better agricultural practices. Growth of benthic and pelagic primary producers has been almost consistently phosphorus limited in recent years, suggesting that continued improvements in phosphorus management and better phosphorus retention in the Timoleague catchment may result in a considerable reduction in macroalgae and phytoplankton bloom magnitudes in forthcoming years.

A 66% reduction in DIP loading in 2016 would have resulted in a reduction of 50% in the total macroalgae wet weight to 190 tones (**Figure 7C**). Overland flow connectivity in a catchment such as the Timoleague significantly contributes to DIP transfer to watercourses during storm events. Meanwhile, during drier periods, baseflow DIP is responsible for a disproportionate increase in DIP concentrations in watercourses; reducing the contribution from point sources must therefore be considered in detail. The pressure posed by point sources nutrient sources such as domestic septic tanks has been cited in other agricultural catchments (Bowes et al., 2005; Jarvie et al., 2006). Reduction of this pressure would improve trophic status during and after low flow periods; it would allow for easier attribution of stream flow nutrient concentrations to agriculture and thus improvements in nutrient concentrations would be more readily attributable to improvements in farm nutrient management practice.

Notwithstanding the differing hydrological conditions, the results of applying the 2010 flow-load relationships for DIN and DIP to 2016, and vice versa, suggest that there has been an improvement in phosphorus retention within the Timoleague catchment between 2010 and 2016; the combination of 2010 flows and 2016 flow-load relationships lead to an increase in the simulated Ulva tonnage from 784t to 968t. However, applying the 2010 flow-load relationships to the 2016 flows resulted in an increase from 380t to 1391t. There has been a clear improvement in flow-normalized phosphorus transfer between 2010 and 2016, as evidenced by the flow-DIP load curves in **Figure 5B**. Therefore, the improvements already observed in Ulva bloom magnitudes between 2010 and 2016 may continue with sustained improvement in farm phosphorus management in the catchments discharging to the Argideen Estuary.

Basu et al. (2010) has posited that in well managed catchments with legacy nutrient sources, flow weighted concentrations become temporally invariant, such that nutrient transfer is solely a function of stream discharge following on from rainfall, whereas in less managed catchments without any legacy sources, flow weighted concentrations would be highly variable and limited only by phosphorus application rates. Legacy phosphorus storage within the Timoleague catchment is being utilized, with flow normalized phosphorus transfer throughout the range of flow percentiles reducing between the hydrologically similar summers of 2010 and 2016 (**Figure 5B**); the Timoleague catchment is in an apparent period of transition, between the two situations suggested by Basu et al. (2010); while phosphorus transfer appears to be reducing, it has not yet become time invariant from year to year.

Considerable effort will be required globally and nationally to maintain and build upon the improvements in catchment management practice observed in recent years; phosphorus loss mitigation measures may be negated if they are not

climate-proofed into the future (Schoumans et al., 2015). With respect to other similar agricultural catchments nationally and internationally, Haygarth et al. (2014) proposes that catchments can be characterized with detailed timeseries of nutrient loading inputs and water course discharge outputs to identify the whether it is undergoing phosphorus accumulation or depletion.

Schulte et al. (2010) identified an average lag of 7–15 years from cessation of fertilizer spreading to high phosphorus soils and the reduction in soil phosphorus to a satisfactory value of soil index 3. Therefore, improvements in farm nutrient management may have a long lead in time before they yield results. Collins et al. (2018) has proposed 12 basic farm management measures which would cost £52 Stg. per hectare to implement, and result in reductions in phosphorus load of up to 12%.

Many challenges must be addressed to further reduce phosphorus losses to water in the future (Sharpley et al., 2015) such as identification of phosphorus transport pathways and mitigation measures to stem phosphorus loss, and implementation of these measures. Roberts and Johnston (2015) advocate the 4R approach to efficiently managing phosphorus and increasing phosphorus recovery, by applying the right source of fertilizer at the right rate, time and place.

The 2016 N: P ratio for the summer freshwater loading, at 105: 1, is closer than the 2010 ratio (190: 1) to the Redfield ratio of 16: 1 which would lead to optimum growth of primary producers, the tidal molar N: P ratio in 2016, at 25:1, represents a state of phosphorus limitation. Simulated Ulva tonnages in the Argideen Estuary are sensitive to tidal boundary nutrient forcing during the summer. While the total diffuse freshwater DIN loads for 2010 and 2016 were quite similar, the total diffuse freshwater DIP load in summer 2016 was 64% greater than in summer 2010. Consequently, were the tidal DIP concentration in 2016 (0.08 µM) the same as 2010 (0.25 µM), a greater bloom would most likely have occurred as previously determined in this system (Ni Longphuirt et al., 2015b), the possibility of tidally exchanged coastal shelf phosphorus concentrations downstream fuelling opportunistic blooms must be borne in mind. Indeed, previously modelled reductions in P loadings to the estuary were shown to have little effect on the standing stock of macroalgal biomass due to the influx of seawater with relatively high P concentration (Ni Longphuirt et al., 2015b).

The North Atlantic Oscillation (NAO) affects both nutrient losses from catchments (Mellander et al., 2018), and ocean circulation (Marshall et al., 2001), thus influencing the model forcing data at either end of the estuary, i.e., nutrient losses, freshwater discharges and coastal shelf nutrient concentrations. The results presented here consider two summers in isolation. Therefore, the impact of the NAO on estuaries should be considered in future research to build upon this research and provide a wider context for the 2 years discussed.

### CONCLUSION

The DCPM model was applied to the system comprising the Argideen Estuary and Courtmacsherry Bay for 2010 and 2016. A series of scenarios were simulated for the year 2016 to inform integrated catchment management in the Argideen Estuary, with a view to eradicating the annual macroalgae bloom.

A 33% reduction in DIP loading produced a macroalgae bloom which was approximately 33% smaller, while a 66% reduction in DIP loading produced a macroalgae bloom that was 50% smaller. Therefore, in the Argideen Estuary the impact of DIP load reduction upon macroalgae bloom magnitudes diminishes with as the DIP load is reduced.

DIP load reduction without parallel DIN load reduction resulted in an increase in DIN concentrations of up to 20% throughout the Argideen Estuary-Courtmacsherry Bay system. The implication of this observation is that if the DIP loads delivered throughout the flow percentile range, without any concomitant change in the DIN flow-load relationship, DIN will be transferred to coastal waters.

Flow-load relationships for 2010 and 2016 suggested there was an improvement in nutrient management within the Timoleague catchment. The 2010 flow-nutrient load relationships were applied to 2016 daily flow measurements, and 2016 flow-nutrient load relationships were applied to 2010 daily flow measurements. The total nutrient loadings were applied to the model for the respective years to determine the improvements caused by implementation of better agricultural nutrient management, advocated by Nitrates Action Programs, in the intervening years. Including 2016 flow-load relationships in the 2010 model represented the scenario that better agricultural nutrient management practice had been brought about sooner. Including the 2010 flow-load relationships in the 2016 model represented the scenario if there had been no improvement in agricultural nutrient management between 2010 and 2016.

Whilst the hydrological conditions during summer 2010 and 2016 differ, the overall flows were similar. The main distinction between the summers was a difference with respect to flow duration, with some higher daily flows observed in summer 2010.


The differing flow-nutrient load curves between 2010 and 2016 indicate that there was a reduction in transfer of P from the Timoleague catchment to surface waters. Given the P-limitation of primary production for summer 2010, a reduction in the P transfer rate between the hydrologically similar summers of 2010 and 2016 would ostensibly indicate that a reduction in Ulva magnitudes would be induced. A reduction in the monitored Ulva tonnages in the main macroalgae patch of Courtmacsherry from 2010 to 2016 from 784 to 391 tones would bear out this hypothesis. In an effort to confirm this improvement, application of the flow-load relationships from each year to the other indicates that a dramatic increase in Ulva tonnages would have been caused were the nutrient management regime in 2010 to remain in place in 2016.

## DATA AVAILABILITY

fmars-06-00064 March 19, 2019 Time: 16:20 # 16

Requests for the datasets used in this study will be facilitated by the corresponding author.

## AUTHOR CONTRIBUTIONS

JM conceptualized the paper scope, sourced input data, calibrated the model, and authored the first draft of the paper and the subsequent paper version post peer-review. MH and SN have reviewed all research from inception through to completion and reviewed drafts of this paper.

## FUNDING

This project was funded under the Irish EPA Research Program 2014–2020, project number 2015-W-FS-17. The EPA Research Program is a Government of Ireland initiative funded by the Department of Communications, Climate Action and Environment.

## ACKNOWLEDGMENTS

We would like to thank the Irish Environmental Protection Agency and the Government of Ireland for the provision of research fellow status and funding for the JM. We would like to thank the steering committee for their input and

### REFERENCES


guidance: Sorcha Ni Longphuirt (EPA), Tomasz Dabrowski (Marine Institute), Joe Silke (Marine Institute), Ronan Kane (Irish Water) and Trudy Higgins (Irish Water). We would like to thank Robert Wilkes (EPA) for the provision of macroalgae monitoring survey data. We would like to thank the United Kingdom Environment Agency for providing the DCPM model code under the United Kingdom open government license, the details of which are available at the following webpage: http://www.nationalarchives.gov.uk/doc/ open-government-licence/version/3/. We would like to thank Paul Tett of the Scottish Association for Marine Science (SAMS) and John Aldridge of the United Kingdom Centre for Environment, Fisheries and Aquaculture Science (CEFAS) who co-developed the DCPM code, and Karen Edwards of the United Kingdom Environment Agency, for providing access to the DCPM code and arranging provision of the code under the United Kingdom open government license. We would also like to thank Per-Erik Mellander, Sara Vero and Ger Shortle of the Teagasc Agricultural Catchments Program, Johnstown Castle, Wexford, for providing hourly nutrient loading and streamflow data and daily rainfall data for the Timoleague catchment for 2010 and 2016 for this publication, which has yielded a greater insight into nutrient loading regimes.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00064/full#supplementary-material



composition in green macroalgal blooms. Ecology 89, 1287–1298. doi: 10.1890/ 07-0494.1


Hypnea musciformis, and Gracilaria tikvahiae. J. Exp. Mar. Biol. Ecol. 471, 208–216. doi: 10.1016/j.jembe.2015.06.012


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 McGovern, Nash and Hartnett. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Past, Present and Future Eutrophication Status of the Baltic Sea

Ciarán J. Murray<sup>1</sup> \*, Bärbel Müller-Karulis<sup>2</sup> , Jacob Carstensen3,4, Daniel J. Conley<sup>5</sup> , Bo G. Gustafsson2,6 and Jesper H. Andersen<sup>1</sup>

<sup>1</sup> NIVA Denmark Water Research, Copenhagen, Denmark, <sup>2</sup> Baltic Nest Institute, Baltic Sea Centre, Stockholm University, Stockholm, Sweden, <sup>3</sup> Department of Bioscience, Aarhus University, Roskilde, Denmark, <sup>4</sup> Baltic Nest Institute Denmark, Roskilde, Denmark, <sup>5</sup> Department of Geology, Lund University, Lund, Sweden, <sup>6</sup> Tvärminne Zoological Station, University of Helsinki, Helsinki, Finland

#### Edited by:

Dongyan Liu, State Key Laboratory of Estuarine and Coastal Research (ECNU), China

#### Reviewed by:

Hsiao-Chun Tseng, National Taiwan Ocean University, Taiwan Marco Uttieri, Stazione Zoologica Anton Dohrn, Italy

> \*Correspondence: Ciarán J. Murray cjm@niva-dk.dk

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 October 2018 Accepted: 07 January 2019 Published: 25 January 2019

#### Citation:

Murray CJ, Müller-Karulis B, Carstensen J, Conley DJ, Gustafsson BG and Andersen JH (2019) Past, Present and Future Eutrophication Status of the Baltic Sea. Front. Mar. Sci. 6:2. doi: 10.3389/fmars.2019.00002 We modelled and assessed the past, present and predicted future eutrophication status of the Baltic Sea. The assessment covers a 350-year period from 1850 to 2200 and is based on: (1) modelled concentrations of dissolved inorganic nitrogen (DIN), dissolved inorganic phosphorous (DIP), chlorophyll-a, Secchi depth, and oxygen under four different of nutrient input scenarios and (2) the application of a multi-metric indicatorbased tool for assessment of eutrophication status: HEAT 3.0. This tool was previously applied using historical observations to determine eutrophication status from 1901 to 2012. Here we apply HEAT 3.0 using results of a biogeochemical model to reveal significant changes in eutrophication status from 1850 to 2200. Under two scenarios where Baltic Sea Action Plan (BSAP) nutrient reduction targets are met, we expect future good status will be achieved in most Baltic Sea basins. Under two scenarios where nutrient loads remain at 1997–2003 levels or increase, good status will not be achieved. The change from a healthy state without eutrophication problems in the open waters took place in the late 1950s and early 1960s. Following introduction of the first nutrient abatement measures, recovery began in some basins in the late 1990s, whilst in others it commenced in the beginning of the 21st century. Based on model results, we expect that the first basin to achieve a status without eutrophication will be Arkona, between 2030 and 2040. By 2060–2070, a status without eutrophication is anticipated for the Kattegat, Bornholm Basin and Gulf of Finland, followed by the Danish straits around 2090. For the Baltic Proper and Bothnian Sea, a good status with regard to eutrophication is not expected before 2200. Further, we conclude that two basins are not likely to meet the targets agreed upon and to attain a status unaffected by eutrophication, i.e., the Gulf of Riga and Bothnian Bay. These results, especially the prediction that some basins will not achieve a good status, can be used in support of continuous development and implementation of the regional ecosystem-based nutrient management strategy, the HELCOM Baltic Sea Action Plan.

Keywords: eutrophication, Baltic Sea, nutrient loads, modelling, scenarios, integrated assessment, status classification

## INTRODUCTION

fmars-06-00002 January 25, 2019 Time: 12:9 # 2

The causes, process and effects of nutrient enrichment and eutrophication in the Baltic Sea are well understood and well documented (Larsson et al., 1985; Rönnberg and Bonsdorff, 2004; Vahtera et al., 2007; Andersen et al., 2011; Gustafsson et al., 2012; Fleming-Lehtinen et al., 2015; Savchuk, 2018). There is no commonly agreed definition of eutrophication, but there is a conceptual understanding of what the consequences of nutrient enrichment are (Andersen et al., 2006; HELCOM, 2009). Discharges, losses and inputs of nutrients from upstream catchments, atmosphere, the North Sea and nutrient regeneration from sediment pools lead to elevated concentration of nutrients in seawater. In most parts of the Baltic Sea, the direct consequences of elevated nutrient concentrations are increased primary production and phytoplankton biomass (Richardson and Heilmann, 1995; Wasmund et al., 2008), and in some areas manifested as blue-green algal blooms (Finni et al., 2001). The increased production of organic matter has negative consequences in most parts of the Baltic Sea. The enhanced sedimentation of organic matter has led to significantly reduced oxygen concentrations and hypoxia has become a largescale problem (Conley et al., 2011; Carstensen et al., 2014). Subsequently, reduced oxygen concentrations have affected not only benthic invertebrates (Villnäs and Norkko, 2011) but also the spawning success rate of cod, a commercially important fish species (MacKenzie et al., 2000; Köster et al., 2001). The turnover of phosphorus in seabed sediments increases with expanding hypoxia which further amplifies primary production and consequently oxygen demand. The so-called vicious circle (Vahtera et al., 2007), is an important indirect effect of eutrophication.

Baltic Sea countries have been working for decades to reduce nutrient inputs and improve eutrophication status, primarily under the umbrella of the Baltic Marine Environment Protection Commission – Helsinki Commission (HELCOM). With the adoption of the Baltic Sea Action Plan (BSAP) in 2007 (HELCOM, 2007; Backer et al., 2010), this work entered a new phase with reductions based on numerical target values and model calculations for basin-wise Maximum Allowable Inputs and Country-wise Allocated Reduction Targets.

With the 2013 update of the BSAP's eutrophication segment, the Baltic Sea states not only implement the Ecosystem Approach to management of human activities but also set a new standard for the development of an adaptive and evidence-based nutrient management strategy (HELCOM, 2013b). The environmental objectives of the BSAP are to attain, by 2020, a healthy Baltic Sea unaffected by eutrophication, including (1) concentrations of nutrients close to natural levels, (2) clear water, (3) natural levels of algae blooms, (4) natural distribution and occurrence of plants and animals, and (5) natural oxygen levels. These objectives were quantified into numerical targets that were subsequently used to calculate Maximum Allowable Inputs that, if achieved, will lead to reaching the objectives. Achieving these ambitious objectives by 2020 is unrealistic given the long retention time of water (30 years, Stigebrandt and Gustafsson, 2003) and nutrients (9– 50 years, Gustafsson et al., 2017) in the Baltic Sea. Indeed, the Maximum Allowable Inputs were quantified with the prerequisite that the targets will be met when the Baltic Sea has adjusted to a new steady state. The mismatch between the policy goal of 2020 and the practical implementation is well known, and even acknowledged in the 2013 HELCOM Ministerial Declaration (HELCOM, 2013b). Nevertheless, it is still highly relevant to determine the most likely time frames, on a regional basis, for reaching a Baltic Sea unaffected by eutrophication.

The objectives of this study are: (1) to use results of a biogeochemical model to classify eutrophication status of nine Baltic Sea basins for the period 1850–2200 and (2) to identify the basins, which are likely to see improvement to a status not affected by eutrophication, and those basins which are not expected to achieve this status.

### MATERIALS AND METHODS

This study represents a meeting of two processes: (1) the regular assessment of eutrophication status in the Baltic Sea region using indicator-based eutrophication assessment tools (i.e., the HEAT tool) and (2) the implementation of the BSAP, particularly the expected future reduction in nutrient inputs from land-based sources and the atmosphere.

### Study Area

The Baltic Sea is an inland sea in northern Europe surrounded by Sweden, Finland, Russia, Estonia, Latvia, Lithuania, Poland, Germany and Denmark, covering a surface area of 415,200 km<sup>2</sup> (**Table 1**). The Baltic Sea is usually divided into several basins separated by sills, including a transition zone to the North Sea including the Kattegat and the Danish Straits (**Figure 1** and **Table 1**).

The basins vary substantially regarding ice cover, temperature, salinity, maximum depth and residence times. There is also a wide variation in composition of the benthic biota between basins. More information about Baltic Sea characteristics can be found in Bonsdorff (2006), Johannesson and André (2006), Österblom et al. (2007), and Leppäranta and Myrberg (2009). Nutrient enrichment and eutrophication signals within the study area are very well studied and documented (HELCOM, 2009; Andersen et al., 2011; Carstensen et al., 2014; Fleming-Lehtinen et al., 2015). The root causes, inputs and fluxes of nitrogen and phosphorus are, in general, well understood and documented (Vahtera et al., 2007; HELCOM, 2009).

Actions to improve the ecosystem health of the Baltic Sea, including the currently impaired status regarding eutrophication, are under way as part of the Baltic Sea Action Plan (HELCOM, 2007) and the EU Marine Strategy Framework Directive (Anon, 2008). With the most recent update of the Baltic Sea Action Plan, the countries bordering the Baltic Sea have agreed on a comprehensive ecosystem-based nutrient management strategy (HELCOM, 2013b).

#### Data Sources

The Baltic sea Long-Term large-Scale Eutrophication Model (BALTSEM: Gustafsson et al., 2012; Savchuk et al., 2012) is


TABLE 1 | Key characteristic of the Baltic Sea and the nine assessments units in this study.

Based on Fleming-Lehtinen et al. (2015) and Andersen et al. (2017).

a coupled physical-biogeochemical model of the Baltic Sea. It represents the complex topography through 13 basin-specific 1D models of high vertical resolution that are linked horizontally. For the current study, model results from the 13 BALTSEM basins were aggregated into the nine basins used for regional eutrophication status assessment. The version of the BALTSEM model which was used to generate the results used in this study explicity describes the dynamics of nitrogen, phosphorus and silica in separate pools. The model simulates three groups of phytoplankton: diatoms, cyanobacteria and a third group including dinoflagellates and all other phytoplankton. Nutrients are taken up by phytoplankton for growth and are subsequently regenerated by heterotroph organisms in the water column. The model further simulates the transport of nutrients from the water column to bottom sediment in the form of detritus, where the organic nutrient pools are slowly remineralized. Oxygen consumption is coupled to all mineralization processes. BALTSEM has been validated against field data and other models (Eilola et al., 2011; Gustafsson et al., 2012, 2017; Savchuk et al., 2012; Meier et al., 2018a,b). It has been used to simulate the change in ecological indicators (Meier et al., 2012; Neumann et al., 2012) and was applied to calculate the Maximum Allowable Inputs of nutrients to the Baltic Sea in the revision of the Baltic Sea Action Plan (HELCOM, 2013a).

The past eutrophication status of the Baltic Sea in 1850– 2006 was simulated by forcing the BALTSEM model with reconstructed nutrient inputs and atmospheric conditions as described in Gustafsson et al. (2012). Its future status was then assessed by extending the model runs for another 194 years under different nutrient scenarios, while hydrodynamics were driven by a statistical representation of the present climate. Nutrient load scenarios included continuation of present nutrient inputs, as well as declining and increasing nutrient inputs. "Present" inputs (PLC5.5) correspond to the loads observed in the BSAP reference period 1997–2003 as described in the review of the 5th HELCOM Pollution Load Compilation (HELCOM, 2013c). Load reduction scenarios simulate nutrient inputs according to the 2013 update of the BSAP's eutrophication segment, implemented either instantaneously (BSAP0; **Figure 2**) or with a linear decrease in loads over 30 years (BSAP30). Further, a high nutrient input scenario (BAU30) represents potential increases in nutrient supply associated with future intensified agriculture in the Eastern Baltic States (Meier et al., 2011; Hägg et al., 2014) with a 30-year transition from present inputs (**Figure 2**). Details about the scenarios and modelled trajectories for the parameters used as indicators can be found as **Supplementary Material**.

#### HEAT 3.0

In this study, we apply the recent version of the HELCOM Eutrophication Assessment Tool (HEAT 3.0), which has been used for assessing eutrophication in the Baltic Sea for the periods 2007–2011 (Fleming-Lehtinen et al., 2015) and 1901– 2012 (Andersen et al., 2017). HEAT 3.0 is a multimetric indicatorbased assessment tool which compares the values of several indicator parameters with threshold values, which define the boundary between eutrophic and non-eutrophic status. The ratios of observed and threshold values are averaged within the categories (1) Nutrients, (2) Direct effects, and (3) Indirect effects. The worst (highest) ratio from the three categories determines the overall Eutrophication Ratio (ER). An ER value greater than 1.0 indicates a eutrophic status whilst values less than 1.0 indicate a good status.

For a detailed description of the assessment principles and methods, please confer with the above references including the Supplementary Material to these. Additional information on the development of the tool and earlier versions can be found in Andersen et al. (2010; 2011; 2014) and Fleming-Lehtinen et al. (2015). For convenience, the HEAT3.0 method as described in Andersen et al. (2017) is reproduced in the **Supplementary Material** to this study.

The target values applied in HEAT 3.0, for the indicators DIN, DIP, chlorophyll-a, Secchi depth, oxygen debt, are taken from Fleming-Lehtinen et al. (2015). An overview of these values, which are also identical to those applied in the study of temporal trends in eutrophication status of the Baltic Sea 1901–2012 (Andersen et al., 2017), is given in **Table 2**. This also shows the categories for indicator aggregation as described above. Andersen et al. (2017) also included an indicator for benthic invertebrates

but since this is not modelled by BALTSEM, this indicator is not used in the model-based HEAT calculations.

## RESULTS

Long-term temporal and spatial trends in eutrophication status of the Baltic Sea were obtained by taking data originating from modelling and applying the HEAT tool to these model results. As a first step, we compared the HEAT classifications for the period 1901–2012, which are based on BALTSEM model results with HEAT classifications based on observations for the same period. The rationale was to check the strength of the similarity of the two assessments to assess if model-based HEAT assessments (this study) are comparable with previously published observationbased assessments (from Andersen et al., 2017). Observation and model based HEAT values increased from 0.6 to 0.8 in 1900– 1920 to approximately 1.0 in the 1930s. In the 1960s observed HEAT values reached 1.5, modelled 1.3. HEAT values increased even further and reached 2.0 based on observations and about 1.5 for modelled values in the 1990s. The relation between the two assessments show good agreement (**Figure 3**), therefore we

TABLE 2 | Basin-specific target values.


Indicators are winter mean concentration of total inorganic nitrogen (DIN), winter mean concentration of total inorganic phosphorus concentrations (DIP), summer mean concentration of chlorophyll-a (Chl-a), summer mean Secchi depth corrected for CDOM (Secchi), and oxygen debt (Oxygen). From Andersen et al. (2017).

have carried out integrated assessments of eutrophication status based on four different input scenarios all nine regional basins (**Figures 4A–I**).

Increasing loads (BAU30 scenario) lead to a worsening of the eutrophication status in all basins. In some basins, i.e., the Baltic Proper (**Figure 4E**) and Gulf of Finland (**Figure 4G**), ER values can potentially reach 2.5, indicating a bad status with significant deviations of indicators from target values. Eutrophication status will improve in the PLC5.5 scenario with maintained nutrient inputs, but the target of a Baltic Sea unaffected by eutrophication will not be reached. In this scenario, the only basin likely to meet the BSAP objectives is the Arkona Basin (**Figure 4C**), although eutrophication status in the Kattegat will approach the target (**Figure 4A**).

The two load reduction scenarios (BSAP0 and BSAP30) may potentially result in oligotrophication sensu Nixon (2009) and thus give a better future eutrophication status in most of the Baltic Sea basins. However, even in these best-case scenarios, some basins are still unlikely to attain a good status according to the HEAT classification with ER values below 1.0, i.e., the Gulf of Riga and Bothnian Bay. BSAP0 and BSAP30 scenarios result in a good status in 7 out of 9 basins, whereas attaining ER = 1.0 in the Gulf of Riga and the Bothnian Sea seems unattainable.

Based on the assessment of the individual basins (**Figure 4**), we can identify the year when the targets are met for each basin. Since there is some year-to-year variation in ER, we define that the objective of good status in a basin is met when the moving average of ER over a 10-year period falls below 1.0. Using this criterion, the BSAP0 scenario predicts that Arkona Basin is the first to achieve good status, in 2024, followed by the Kattegat and Bornholm Basin in 2057, then the Gulf of Finland and Danish Straits in 2064 and 2080, respectively. Good status is achieved in the Baltic Proper and Bothnian Sea around 2200, just within the time scale of the model simulations.

The BSAP30 scenario describes a similar recovery pathway with basins achieving good status in the same order as for the BSAP0 scenario, however, as might be expected, with somewhat delayed responses. Good status is achieved 9 years later in the Arkona Basin (in 2033) and 8 years later in the Danish Straits (in 2088). The Baltic Proper and the Bothnian Sea just manage to achieve good status before 2200. As described above, Arkona is the only basin expected to return to a good status in the PLC5.5 scenario in 2079, whereas good status will not be achieved for any basin with the BAU30 scenario.

## DISCUSSION

The current eutrophication status in the Baltic Sea is far from the objectives agreed upon in the BSAP (HELCOM, 2013b). This is well documented as shown here and in Fleming-Lehtinen (2016) and Andersen et al. (2017). However, earlier trends of increasing eutrophication have been reversed and the Baltic Sea

has entered a phase of recovery (Andersen et al., 2017). Examples of oligotrophication and partial recovery have been documented in many coastal waters, e.g., in Denmark (Riemann et al., 2016; Staehr et al., 2017), in Sweden (Walve et al., 2018), in the North Sea (Andersen et al., 2016; OSPAR, 2017; van Beusekom et al., 2018) and in United States (Bricker et al., 2008; Oviatt et al., 2017; Zhang et al., 2018).

Comparison of observation-based HEAT classifications and model-based HEAT classifications for entire Baltic for the period 1902–2012 shows a reasonable relation between the two methods of assessing eutrophication status. A model essentially gives a smoothed representation of the data and cannot describe the micro-variability and measurement errors associated with monitoring data, and henceforth affecting basin average values. Therefore, it is not unexpected that the slope of the regression (0.565) shows that the model-based HEAT results vary less than the observation-based HEAT results. The fact that the HEAT assessment based on observed data also includes a benthic invertebrate quality index could possibly explain some of the differences between model-based and observation-based HEAT results. However, both are in reasonable agreement on where the status changes between eutrophic and good. And since the model-based HEAT results capture 85% of the variation seen in the observation-based HEAT results, we conclude that using the model-based assessments provides an important way forward to assess the potential effects of the BSAP into the future.

Our study predicts how future eutrophication status will improve under different scenarios of nutrient reduction. The time required to achieve good status varies from decades in some basins to centuries in others. The time needed for recovery could be reduced with a faster implementation of nutrient reductions. In most basins, there are substantial delays between implementation of measures leading to reduced loads and ecological responses. Lagged responses of marine ecosystems are well-known (Carstensen et al., 2011) also due to large-scale changes associated with global climate and increasing human stress on coastal ecosystems.

Very few similar studies from other regions assessing the temporal trends in eutrophication status have been published. An example from the North Sea is OSPAR (2017), which applies a harmonized assessment framework for the third time and concludes that the spatial extent of 'problem areas' in terms of eutrophication has decreased from approximately 169.000 km (1990–2001) to 119.000 km<sup>2</sup> (2001–2005) and to 100.000 km<sup>2</sup> for the period 2006–2014. An assessment of the effects of nutrient enrichment in United States estuaries (n = 58) from the early 1990s to the early 2000s (Bricker et al., 2008) concludes that conditions had remained the same over this period in most systems (32), whilst they had worsened in 13 and improved in 13. In the future, conditions were predicted to improve in 19% of the assessed systems. However, in 65% of the estuaries conditions were expected to worsen, due to projected increases in nutrient loads with increasing population density.

There are limited numbers of assessments of eutrophication status in EU Member States and even fewer studies that describe how ecological status improves in eutrophied marine waters. In coastal waters susceptible to high nutrient loads, assessment of 'ecological status' according to the EU Water Framework Directive (WFD) can be considered equivalent to a eutrophication status assessment. Accepting this, Member States'

so-called Initial Assessments can provide an indirect indication of whether conditions are improving in coastal waters. Based on a meta-study of national reporting, a recent pan-European assessment (Kristensen et al., 2018) concludes that the overall ecological status of surface waters has not improved. In some coastal waters, the assessed status is even worsening, despite River Basin Management Plans in place to improve water quality.

To attain a Baltic Sea unaffected by eutrophication, reaching the load reduction targets of the ecosystem-based nutrient management strategy (the BSAP's eutrophication segment, HELCOM, 2007, 2013b) is required. Decision-makers and the wider public should be aware of the current poor situation and should be well informed on the time-scales of Baltic ecosystem recovery. Improving the communication between decision-makers and the scientific society should of course be anchored in scientific studies and literature, but the primary means of communication is not scientific papers and complicated graphs. There is in our opinion a need for simplification, where complex information is synthesized in info-graphics, where messages can be more easily understood. For example, the trend information in multiple graphs in **Figure 4** can be summarized as a single graph (**Figure 5**), where the eutrophication status, and the basin-wise trends are presented using simple colour classes. How the Baltic Sea changed from a system unaffected by eutrophication at the beginning of the 20th century to its present eutrophic state is now expressed in a single graph using an intuitive colour scale. The same graph can further show what we can expect for the future: the likely consequence of the agreed load reductions (BSAP0 and BSAP30 scenarios), once implemented, will cause significant improvements and ultimately a Baltic Sea unaffected by eutrophication in most basins.

Andersen et al. (2017) combined the status classifications from the nine basins into an overall Baltic Sea. In this way, data for several indicators representing different features of the ecosystem are synthesized into a single value. In a similar manner, this study presents the overall Baltic Sea eutrophication status classifications for four future nutrient load scenarios (**Figure 6**). The results for individual basins (**Figure 4**) are

fifth class "High" (ER < 0.5) but none of the BALTSEM assessments returned this result.

thus integrated, giving a result which reflect the overall longterm trends in eutrophication status resulting from differences between nutrient inputs scenarios. This integration supports the interpretation of the classification presented in **Figure 5** and reveal, not surprisingly, that PLC5.5 and BAU30 scenarios do not lead to a Baltic Sea unaffected by eutrophication, whilst BSAP0 and BSAP30 scenarios will both, after a considerable

number of years, meet the overall objective of a healthy Baltic Sea.

eutrophication status. The figure also indicates the ER ranges for colours in Figure 5.

The long-term trends in loads, indicators and eutrophication status follow distinctive trajectories for the four scenarios, though the differences between BSAP0 and BSAP30 scenarios narrow continuously. Communicating the links between human activities, the loads to the Baltic Sea, the responses in selected indicators, the time lags and ultimately the overall implication with respect to eutrophication status, to decision makers is important. One of many ways of synthesizing the results of this Baltic Sea-wide study is to compare the trends for selected indicators and for the worst-case (BAU30 scenario) and bestcase scenario (BSAP0 scenario). By doing so, we illustrate the difference between implementing a state-of-the-art ecosystembased nutrient management strategy (BSAP) and doing nothing at all (**Figure 7**).

An interesting finding from the long-term trends is that the biological responses (chlorophyll-a, Secchi depth and oxygen debt) return to a good status with respect to eutrophication earlier than indicators for DIN and DIP concentrations. In part, different degrees of model bias contribute to the difference in timing. While BALTSEM captures phytoplankton trends, it tends to underestimate biomass and therefore chlorophyll-a concentrations (**Supplementary Figure S8**). Since phytoplankton biomass enters Secchi depth calculations, Secchi depth is slightly overestimated (**Supplementary Figure S9**), whereas nutrient and oxygen concentrations are simulated with little bias (**Supplementary Figures S6, S7, S10**). On the other hand, target values for biological variables, nutrient and oxygen levels have been developed by applying statistical change point detection methods to each time-series individually (HELCOM, 2013a). Thus, they do not necessarily reflect a single year in the eutrophication trajectory of the Baltic Sea. That targets might not be met simultaneously and might even be unachievable in individual basins was also taken into account in the 2013 Baltic Sea Action Plan revision. For example, the winter DIN target in the Gulf of Riga was disregarded (Gustafsson and Mörth, 2013), since increasing phosphorus limitation tended to increase DIN concentrations in the Gulf (see also Müller-Karulis and Aigars, 2011), while other indicators like phytoplankton biomass and oxygen debt reached their ecological targets. Both the Gulf of Riga and the Bothnian Bay are exceptionally phosphorous limited and becoming more so. This explains why the revised Baltic Sea Action Plan does not include nitrogen reductions in these basins even though nitrogen concentration targets are exceeded.

In other words, considering biological response indicators alone, it is likely that something resembling a Baltic Sea unaffected by eutrophication would be achieved earlier than when all indicators are used in the assessment. The responses of nutrient indicators to the load reductions (**Figures 7D,E**) do not appear to be lag behind the responses of biological indicators (**Figures 7F,G**) or oxygen debt (**Figure 7H**). However, they take longer to reach the target and do not reach as close to or as far under the target values.

Considering each of the 39 separate combinations of indicator and basin (See **Supplementary Figures S1–S5**) a similar pattern is seen. For BSAP0 and BSAP30, predicted concentrations of chlorophyll-a in 2200 are less than 50% of the target value in all but one of the nine basins, in some cases far less. For winter DIN in the BSAP0 scenario, the best case is in Arkona Basin where the concentration has fallen under the target and reached approximately 70% of the target value by 2200. For five of the other basins (Kattegat, Danish Straits, Bornholm Basin, Baltic Proper, and Bothnian Sea), DIN concentrations in 2200

lie close to the target values and for the three remaining basins (Gulf of Riga, Gulf of Finland, and Bothnian Bay) they are clearly above target values. Thus, according to the model the target values for nutrients are significantly more stringent than those for chlorophyll-a, Secchi depth and oxygen debt. However, since the biological indicators are dependent on nutrients and not the other way around, this situation can be seen in a positive light, in that it ensures that the nutrient targets are sufficiently ambitious to achieve the desired changes in biological indicators.

### CONCLUSION AND PERSPECTIVES

The BSAP may, according to model predictions, be an efficient driver regarding reduction of nutrient loads. However, this requires commitments from all HELCOM Contracting Parties to meet BSAP load reduction targets. Without a collective and strong commitment, we risk failing to attain a Baltic Sea unaffected by eutrophication.

This study indicates that a good status with respect to eutrophication will be met for most parts of the Baltic Sea, if the BSAP nutrient reductions are fully implemented. This recovery has already started but the ultimate effects will not be visible soon, but in a much longer perspective. An encouraging result of the study is that it concludes that the overall objective of a healthy Baltic Sea is within reach. Patience is required, as well as a continuation of the load reductions achieved so far, such that the load reduction targets set by BSAP are met. An interesting and positive finding is that the indicators representing biological responses do seem to respond faster to load reduction than indicators representing nutrient concentrations. This implies that the visual appearance of the Baltic will reach a good status earlier than the integrated assessment based on the HEAT tool and the full range of indicators.

On a less positive note, we should remember that none of the scenarios take into account climate change, where elevated sea temperatures are of concern. Thus, there seems to be an urgent need to include climate change in future updates of the BSAP and to update the projected development in eutrophication status.

### AUTHOR CONTRIBUTIONS

CM and JA conceived the study. BM-K and BG provided modelled trajectories from the BALTSEM model. JC and BG provided observed indicator trajectories. CM did the integration calculations. CM, JA, BG, and BM-K wrote a preliminary version of the manuscript. All authors discussed and reworked the manuscript.

#### FUNDING

This study is a Baltic Nest Institute activity co-funded by Aarhus University (AU) and Stockholm University (SU). Parts of the analyses were funded by EEA ETC ICM 2016 task 1.6.1.g ('Eutrophication in Europe's seas').

#### ACKNOWLEDGMENTS

fmars-06-00002 January 25, 2019 Time: 12:9 # 11

We thank all those who have participated in development, testing, or application of various versions of the HEAT tool over the past

#### REFERENCES


decade. This study is dedicated to the memory of Prof. Fredrik Wulff.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00002/full#supplementary-material



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Murray, Müller-Karulis, Carstensen, Conley, Gustafsson and Andersen. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# A Novel Approach for Deriving Nutrient Criteria to Support Good Ecological Status: Application to Coastal and Transitional Waters and Indications for Use

#### Fuensanta Salas Herrero<sup>1</sup> \*, Heliana Teixeira<sup>2</sup> and Sandra Poikane<sup>1</sup>

<sup>1</sup> European Commission, Joint Research Centre (JRC), Ispra, Italy, <sup>2</sup> Department of Biology & CESAM, University of Aveiro, Aveiro, Portugal

#### Edited by:

Jesper H. Andersen, NIVA Denmark Water Research, Denmark

#### Reviewed by:

Robinson W. Fulweiler, Boston University, United States Akkur Vasudevan Raman, Andhra University, India Urmas Lips, Tallinn University of Technology, Estonia

\*Correspondence:

Fuensanta Salas Herrero Fuensanta.Salas-Herrero@ ec.europa.eu

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 31 August 2018 Accepted: 29 April 2019 Published: 17 May 2019

#### Citation:

Salas Herrero F, Teixeira H and Poikane S (2019) A Novel Approach for Deriving Nutrient Criteria to Support Good Ecological Status: Application to Coastal and Transitional Waters and Indications for Use. Front. Mar. Sci. 6:255. doi: 10.3389/fmars.2019.00255 A huge variability exists in nutrient concentrations boundaries set for the water (WFD) and the marine strategy framework directives (MSFD), as revealed by a survey to EU member states (MS). Such wide variation poses challenges when checking policy objectives compliance and for setting coherent management goals across European waters. To help MS achieve good ecological status (GES) in surface waters, different statistical approaches have been proposed in a Best Practice Guide (BPG; CIS Nutrients Standards Guidance) for establishing suitable nutrient boundaries. Here we used the intercalibrated results from the WFD for the biological quality element phytoplankton to test the applicability of this BPG for deriving nutrient boundaries in coastal and transitional waters. Overall, the statistical approaches proved adequate for coastal lagoons, but are not always robust to allow deriving nutrient boundaries in other water categories such as estuaries, in transitional waters, or some coastal water types. The datasets available for analysis provided good examples of the most common problems that might be encountered in these water categories. Similar issues have been found in freshwater environments, for which solutions are proposed in the BPG and which are demonstrated here for coastal and transitional waters. The different approaches available and problems identified can be useful for supporting the derivation of nutrient concentrations boundaries both for the Water and the MSFDs implementation.

Keywords: eutrophication, nutrients, phytoplankton, water framework directive, marine strategy framework directive

## INTRODUCTION

European water policy aims to achieve good ecological status (GES) for all rivers, lakes, coastal and transitional water bodies of European Union (European Commission [EC], 2000). The ecological status is assessed based on biological quality elements (BQEs), accompanied by physico-chemical and hydromorphological quality elements (Borja, 2005; Birk et al., 2012). To achieve GES, river basin management plans should be put in place, addressing all relevant pressures (Hering et al., 2010, 2015). Eutrophication is still the major pressure in European coastal and transitional waters, therefore, setting of nutrient criteria to reach GES is of utmost importance (European Environmental Agency [EEA], 2012).

Transitional and coastal waters (CWs) are among the most highly impacted ecosystems in the world presenting inherently high variability over both spatial and temporal scales (Paerl, 2006; Reyjol et al., 2014). In those environments, the greatest impacts of increasing nutrient concentrations have been observed at sites with restricted water exchange, resulting in phytoplankton and macroalgal blooms (Tett et al., 2003; Salas et al., 2008; Carstensen and Henriksen, 2009; Teichberg et al., 2010).

A significant amount of research has been done in developing and intercalibrating biological indicators to assess impact of eutrophication in coastal and transitional waters (Borja et al., 2013; Garmendia et al., 2013; Marbà et al., 2013). The most suitable BQEs for assessing eutrophication are phytoplankton and macrophytes (angiosperms and macroalgae) due to their more established and direct response to nutrient conditions (Smith et al., 1999; Boström et al., 2002; Devlin et al., 2007).

However, less attention has been directed to linking ecological status to management actions and establishing meaningful and consistent nutrient criteria to support GES (Hering et al., 2010). A comparison of nutrient boundaries set for the WFD and the marine strategy framework directive (MSFD) in transitional, coastal and marine waters across European member states (MS) (Dworak et al., 2016) revealed a huge variability in nutrient concentrations boundaries, but also in other relevant aspects such as the nutrient parameters and metrics used, the time of year assessed, the reference conditions established. It also revealed that often MS boundaries do not follow their regional sea convention (RSC) nutrient standards. The possible implications of the wide variations in the nutrient concentration boundaries set by the European MS need to be understood in the context of establishing appropriate nutrient boundaries to achieve GES, as urged by the Working Group on Ecological Status (ECOSTAT), as part of the Common Implementation Strategy for the WFD. A best practice guide (BPG, Phillips et al., 2018) has been elaborated in this context. Its purpose is to help MS achieving GES in surface waters. It complements previous guidance on eutrophication assessment (European Commission [EC], 2009) by providing more targeted advice on how to link nutrient concentrations in surface waters to specific policy objectives. The new guide includes a tool kit to facilitate the application of the different statistical approaches proposed to establish the nutrient targets.

On the other hand, one major achievement of the WFD implementation has been the establishment of a common view of ecological status through an intercalibration exercise (Birk et al., 2013; Poikane et al., 2014, 2015). This has ensured that the concepts of ecological status are transferable between groups of organisms (fish, invertebrates, macrophytes, algae) and between countries within the EU. This, in turn, provides a robust view of GES (and other status classes) that can be used as the starting point for the development of nutrient targets. However, so far, this concept has been applied only to freshwaters (Dolman et al., 2016; Free et al., 2016; Poikane et al., 2019).

The aim of this work is to demonstrate the applicability of the statistical approaches proposed in the BPG, hereafter guidance, in coastal and transitional waters. The work focuses on the pressure-response relationships found between nutrients and the BQE phytoplankton in transitional and CWs. The phytoplankton was selected because it is deemed more responsive to the type of pressure under study, i.e., nutrients (Devlin et al., 2007; Suikkanen et al., 2007). The WFD requirement for assessing the ecological status of the phytoplankton quality element, in transitional and CW bodies, includes taxonomic composition, abundance and biomass of phytoplankton, as well as bloom frequency. However, most of the MS use the concentration of Chlorophyll a (Chl a) as a proxy measure for phytoplankton biomass for the intercalibration, whereas most of the indicators of the other sub-elements (phytoplankton composition and blooms) have not yet been intercalibrated (Garmendia et al., 2013). Case study examples with Chl a metrics from the WFD intercalibration (IC) exercise were used for testing the suitability of the different methodologies proposed in the guidance. In addition, where significant pressure-response relationships have been observed, nutrient boundaries were derived, which can be adopted in the respective water types.

## MATERIALS AND METHODS

## The Toolkit

The guidance document is designed to assist in the determination of the concentrations of phosphorus and nitrogen that are likely to support GES. It can be used to check existing boundary values or to develop new ones. The guidance is supported by a toolkit in the form of an Excel workbook and a series of scripts which can be run using R, an open-source language widely used for statistical analysis and graphical presentation (R Development Core Team, 2016). The toolkit provides the full R code, together with a series of examples which can be used to explore the methods.

This toolkit includes different statistical approaches to derive nutrient boundaries:

Univariate linear regression: assuming a linear relationship between the ecological quality ratio (EQR) and nutrients, three regression types are implemented: two ordinary least squares OLS linear regressions between EQR and log nutrients concentration, where each variable is alternatively treated like the independent variable (because none of our two variables in practice can be considered to be free of error); and a third, type II regression, the ranged major axis (RMA) regression. The predicted range of nutrient threshold values are then determined from the range of results obtained from these regressions' parameters.

Logistic regression: this approach treats ecological status as a categorical variable where a logistic model is fitted between categorical data using a binary response, "biology moderate or worse" = 1 or "biology good or better" = 0 and log of nutrient. Nutrient concentrations are determined where the probability of being moderate or worse was 0.5. In the case that additional pressures, other than nutrients, are suspected, a nutrient concentration value was determined at a probability of 0.75 instead.

Categorical methods: nutrient concentrations associated with a particular ecological status class could also be expressed as a distribution from which an upper quantile might be chosen to indicate a nutrient concentration above which good status was very unlikely to be achieved, or a lower quantile below which

good status was very likely to be achieved (average of upper and lower quartiles of adjacent classes), so long as nutrients are the main driver of status. The average of the median of adjacent classes and the upper 75th percentile distribution are two additional categorical approaches tested.

Minimisation of mismatch of classification: estimates the nutrient threshold value that minimizes the mismatch between status (good or better and moderate or worse) for the ecological and the supporting element.

Linear quantile regression: useful alternative when the nutrient-biology interactions are confounded by other stressors, or environmental factors, leading to wedge-shape, or invertedwedge, type of distributions. In such cases, the quantile regression allows different rates of change in the response variable to be predicted along the upper (in the presence of stressors) or lower (in the presence of mitigating environmental factors) quantiles of the distribution of the data (Cade and Noon, 2003).

Detailed information about the methods included in the toolkit is provided in the Guidance (Phillips et al., 2018).

The boundaries predicted using the different toolkit methods were compared with the national nutrient boundaries established by MS, where available.

#### Datasets

The datasets used in this study represent systems in some of the common IC types across geographic intercalibration groups (GIG) in the water framework directive (WFD). For the Baltic, this work includes data for the single transitional waters (TRW) common type defined and also for two of the nine CWs common types in this GIG. For the Mediterranean, there is data for five of the six CWs common types defined for phytoplankton, and also for one TRW common type. For the North East Atlantic there is data for the single TRW common type defined (**Table 1**).

Data from 13 MS was compiled (**Table 2**), which included essentially nutrient parameters and EQR values for the BQE Phytoplankton. In most cases the parameter indicative of biomass (i.e., Chl a) was intercalibrated, and not the full national classification system. In a few cases, datasets included also the raw Chl a data and supporting environmental data such as: turbidity (Turb), flushing regime, tidal range (T), and distance to shore, salinity (Sal), pH, and dissolved oxygen (DO).

The nutrient parameters available varied across water categories (CW and TRW), and also across regional GIG, but they were common within IC types. To keep the examples comparable, in this study we have focused the analysis in the phytoplankton response to total Nitrogen (TN), total Phosphorus (TP) and dissolved inorganic Nitrogen (DIN) parameters, since these are the most commonly applied across GIG and common types in both water categories (**Table 2**). Nutrient data refers to specific seasons of the year, either Summer (Su) or Winter (Wi), but in some cases all year round data was considered (**Table 2**). In a few cases, for the Mediterranean, the season information was not available. The nutrient units used by the MS were kept for the analyses, in order to facilitate the comparison with their own established nutrient boundaries.

All datasets considered in the present study have associated EQR IC boundaries (European Commission [EC], 2018) that allow their use for deriving nutrient boundaries. Some MS use a full intercalibrated phytoplankton EQR method (e.g., Italy, Greece, and France in TRW); however, the majority of the MS reported a partially intercalibrated EQR method based only on the Chl a parameter. Where different methods have been used within a common type, and with different IC boundaries, in such cases, the datasets were analyzed separately. Only in the case of the common type NEA 11 statistical analyses have been applied on a common data set, and not separately for each country. The NEA11 common dataset was derived after EQR's normalization (nEQR), for which the full set of MS quality classes' boundaries were used where available, otherwise assumed equidistant for the missing classes. The toolkit normalize template was used, which is available with the Guidance documentation (Phillips et al., 2018). As relationships within national datasets are often relatively weak, there is a good case for combining data to produce a single dataset spanning several countries. This was the case for the NEA11 dataset.

Datasets were provided by EU MS and compiled by the Joint Research Centre and are only available for results' checking purposes.

## Data Check and Exploration

The recommended steps in the guidance protocol (Phillips et al., 2018) were largely followed and thus, initially, we checked and explored the datasets in order to identify:


Potential outliers were identified if data points fell above or below the 0.975 percentile of residuals, corresponding to 5% of data points that would fall outside of the orthogonal regression line. However, they were only excluded from the analysis if they were considered isolated points, resulting from either measurement errors or sampling situations responding to unusual factors (e.g., samples representing punctual extreme hydrological events); in which case, they are identified in the results' plots. If the outliers reflected instead, for example,

TABLE 1 | Common types and members states (MS) considered for testing the toolkit in this study.


data points falling outside the linear range or trends in data dispersion, then their exclusion /inclusion was evaluated for each statistical method. The point was to allow comparison of the performance of different methods toward different features commonly found in real datasets. In any case, data truncated are identified in the results and the nutrient range modeled or used to set boundaries by each method is acknowledged. The log10 transformation of nutrient concentrations, applied prior to some of the analysis, is commonly used for reducing right skewness in a variable distribution. This transformation is often appropriate for measured variables, which is the case of this dataset. It is also one of the options to get a linear relationship between the nutrients and phytoplankton EQR. Since we are testing OLS regression models, this transformation allows stabilizing the variance of the given variable when the standard deviation is proportional to the mean (too wide variance), bringing it more close to a normal distribution.

#### Selection of the Method

The Guidance includes a road map for the selection of the most appropriated methods. For the linear methods, the selection is based on the strength of the relationship between the BQE and nutrients. To apply these univariate regressions, the dataset should span ideally four ecological quality classes and show a linear relationship for at least the range of the High (H), Good (G), and Moderate (M) classes. In addition, the strength of the observed relationship should be good enough to allow making predictions, and we suggest that the regression model coefficient of determination is r-squared > 0.36. This correlation value (r > 0.6) follows recommendations by Smith (2009) based on Jolicoeur's (1990) work when using type II linear regression methods, as the RMA considered in our work. If the observed linear relationship is weaker, then categorical methods (including logistic regression) should instead be tested and applied to derive nutrient thresholds. However, for some of the categorical methods, it is required that significant differences between the nutrients distribution across ecological status classes are observed before nutrient boundaries can be derived from such methods (Phillips et al., 2018).

## RESULTS

#### Univariate Regressions

Relationships between nutrients and biological elements that are meaningful enough to allow establishing nutrient boundaries, i.e., significant and with an r-squared near or higher than 0.36, were found in Lithuanian and Polish transitional waters (common type BT1), in French, Italian and Greece polyhaline coastal lagoons (TRW MED common type), and in Estonian CWs (common type BC4). In Estonia, only TN presented a robust relationship (R <sup>2</sup> > 0.36).

For Lithuanian transitional waters referring to oligohaline very sheltered coastal lagoons (BT1), the univariate regression of EQR with TN showed a significant acceptable relation (R <sup>2</sup> = 0.41, n = 25) and predicted a concentration for good/moderate


boundary of 1228 µg L−<sup>1</sup> . For Polish water bodies of the same common type, univariate regression results between TN and the EQR Chl a predicted a slightly lower G/M nutrient concentration of 1042 µg L−<sup>1</sup> (R <sup>2</sup> = 0.78, n = 14). Here it was considered a range of nutrient concentration between 400 to 1640 µg L−<sup>1</sup> where a linear trend was observed. Observations out of the linear range were previously dropped from the analysis. Those 12 excluded samples corresponded to three sites showing consistently strong phosphorus limited patterns (N:P molar ratio > 32), and not responding to TN enrichment as the remaining samples of a different site. In the case of TP, for this BT1 common type in Lithuanian waters, the relation was slightly lower and the G/M boundary was predicted at 91 µg L−<sup>1</sup> nutrient concentration (R <sup>2</sup> = 0.34, n = 27). With the Polish data, in this same type, a stronger relationship with TP was observed (R <sup>2</sup> = 0.46, n = 22), after excluding nitrogen limited samples. A similar G/M TP concentration was predicted (100 µg L−<sup>1</sup> ).

In the Mediterranean, all the French, Italian and Greek water bodies of the common type polyhaline coastal lagoons showed significant (p < 0.001) and strong relationship between nutrients and phytoplankton metrics (**Tables 3**, **4**). For Italian and Greek coastal lagoons, the nutrients' relationships were established with a multimetric phytoplankton index (MPI) score (TN R <sup>2</sup> = 0.61, n = 14; TP R <sup>2</sup> = 0.60, n = 15). Despite a good and highly significant relationship is found for TN, the predicted TN G/M boundary (1749 µg L−<sup>1</sup> ) falls beyond the nutrient range used for model predictions (454–1515 µg L−<sup>1</sup> ), which requires caution. In the case of France, regressions where established for a different Phytoplankton EQR method (only Chl a) with nutrients, but results found were equally good: TN (R <sup>2</sup> = 0.64; n = 13) and TP (R <sup>2</sup> = 0.86, n = 13). The EQR Chl a relationship with TP (log<sup>10</sup> transformed data) shows a clear linear trend (**Figure 1**), as captured by the BPG toolkit graphical output of the three types of univariate regressions considered: ordinary least square of nutrient on EQR and vice-versa, and also type II regression of EQR on log<sup>10</sup> nutrient concentration. Predicted boundaries are indicated in **Tables 3**, **4**.

For the CW types analyzed in this study, only Estonian water bodies of the common type BC4 showed good univariate linear relationships with one of the nutrients tested. The linear regression obtained for EQR∼TN (**Figure 2**) was the only relationship robust enough (R <sup>2</sup> = 0.58, n = 31) for deriving boundaries (TN G/M = 25 µmol L−<sup>1</sup> ). For this analysis, the nonlinear portion of the data was excluded (TN > 33 µmol L−<sup>1</sup> ), as well as a data point with extreme EQR = 2.9. Nevertheless, this dataset presented a high overlap between nutrient concentrations distribution across ecological status classes, in particular for the G/M boundary, both for TN and TP (**Figure 2**).

#### Bivariate Regression Models

For Lithuanian transitional waters included in the common type BT1, the bivariate model improves slightly from the univariate models for TN or TP previously presented (adjusted R <sup>2</sup> = 0.55; p < 0.001, n = 23). The predicted nutrient boundaries for H/G and G/M, for both nutrients, are very similar to those of the univariate approach (**Table 5**) and are presented in **Figure 3**.

fmars-06-00255 May 16, 2019 Time: 16:37 # 5

MS Phytoplankton Models R <sup>2</sup> Nutrient range Most likely boundary Possible p-value used TN µg L−<sup>1</sup> Range n GM TN HG TN µg L−<sup>1</sup> µg L−<sup>1</sup> Pred 25th 75th Pred 25th 75th Italy/ Greece EQR\_Phyt (MPIscore) Boundaries: HG 0.78 GM 0.51 EQR v TN (RMA) 0.609 <0.001 n = 14 454 – 1515 1749 1438 1886 1052 1016 1066 GM no data HG 824–1245 Average adjacent quartiles no data 1095 Average adjacent classes n = 14 no data 1077 Average 75th quartile 454 – 1515 1463<sup>∗</sup> 824 Minimize class difference 1700<sup>∗</sup> 920 France EQR\_Phyt Boundaries: HG 0.71 GM 0.39 EQR v TN (RMA) 0.642 (0.539) <0.001(0.002) n = 13(15) 177 – 1612 586 (579) 582 (570) 594 (598) 271 (287) 216 (215) 304 (329) GM 362–929 (360–952) HG 132–419 (130–479) Average adjacent quartiles 600(549) 360<sup>∗</sup> (365)<sup>∗</sup> Average adjacent classes n = 13 (15) 641(560) 369(373) Average 75th quartile 433(433) 419<sup>∗</sup> (429)<sup>∗</sup> Minimize class difference 545(480) 205(255)

TABLE 3 | Summary of predicted values of nutrient concentration at the Good/Moderate (GM) and High/Good (HG) boundaries of TN (µg L−<sup>1</sup> ), per countries (IT/GR, and FR), within Mediterranean TRW common type polyhaline coastal lagoons (TRWMEDpolyCL), obtained by regression and categorical analyses.

The number of observations n used by each method is indicated; the full dataset observations and the corresponding predicted boundary values are indicated in brackets, for comparing outliers' removal effect where applicable. <sup>∗</sup> Non-significant differences between the nutrient concentrations in adjacent classes (Wilcoxon Rank Sum test), treat quantile based results with caution.

TABLE 4 | Summary of predicted values of nutrient concentration at the Good/Moderate and High/Good boundaries of TP (µg L−<sup>1</sup> ), per countries (IT/GR, and FR), within common type TRWMEDpolyCL, results obtained by regression and categorical analyses.


The number of observations n used by each method is indicated; the full dataset observations and the corresponding predicted boundary values are indicated in brackets, for comparing outliers' removal effect where applicable. <sup>∗</sup> Non-significant differences between the nutrient concentrations in adjacent classes (Wilcoxon Rank Sum test), treat quantile based results with caution.

#### Logistic Regression Models

Binomial logistic regression was applied on the NEA 11 data set, as this statistical analysis is the most reliable and flexible categorical method, included in the toolkit, when linear modeling is not appropriate. A good example is in the presence of weak nutrient-biology relationships as observed for NEA11 common type (R <sup>2</sup> = 0.21). This method is not substantially influenced by the mean of the data set and is only slightly influenced by scatter

in the data (Phillips et al., 2018; Phillips et al., unpublished). The binomial logistic regressions of DIN on biology (for normalized nEQRs, and excluding the French dataset that presented opposite trend), for both the H/G and the G/M range are presented in **Figure 4**. Nutrient boundary estimates are presented for a 50% probability of being in moderate or worse status for the G/M, or in good or worse for the H/G, but nutrient values at lower and higher probability thresholds (25% and 75%) are also presented, which provide precautionary and non-precautionary values.

However, like other methods, boundary estimates may still be unreliable if other pressures are operating. This is often the case in estuaries, where multiple pressures are frequently encountered. The very scattered regression observed for the NEA11 data suggests that other pressures besides DIN are contributing to decrease the EQR, influencing data distribution (i.e., wedgeshaped distribution observed in **Figure 5**) and masking the nutrient relationship with biological data.

#### Quantile Regression Models

An additional alternative approach tested for establishing guiding nutrient boundaries in such widely scattered NEA11 data, suspected to be caused by the influence of unaccounted pressure variables, was a quantile regression model. For this NEA11 available dataset, a higher quantile has been adopted, for coping with the influential role of potential unknown stressors (or environmental features) in the shape of the data, as the modeling is only possible to do it with this portion of the dataset. The univariate model fitted indicated that the maximum DIN levels (using the 0.7 quantile) that could still support High/Good and Good/Moderate Ecological Status correspond to nutrient concentrations of 68 and 212 µM, respectively (**Figure 5**). However, the 95% confidence intervals obtained for the G/M boundary are too wide; indicating that, at such nutrient concentrations, large EQRs variation (ranging from 0.39 to 0.81) could be expected.

## Minimize the Mismatch Between Biological and Supporting Element Classification

Another possibility of the statistical approach to establish the nutrient boundaries is the minimisation of mis-match method, as this is the least sensitive to outliers and non-linear relationships.

This is the case for Mediterranean CWs, as well as for all coastal common types, which showed, in general, weak results both for TN and Phosphorus. In most of the cases a wedgeshape data distribution did not allow the use of linear regression approaches. Therefore, only categorical approaches should be adopted for deriving nutrient boundaries from intercalibrated EQR data. Results obtained for this categorical approach with the toolkit are indicated in **Table 5**.

In the case of the common type NEA 11, using this approach (**Figure 6**) the mean estimated high/good boundary for DIN is 52.5 µM, within a range of 47–59 µM, with a total mismatch classifications rate of 30%, ranging from 28–34%. For the good/moderate the mean estimated boundary is 74.5 µM, which is within the range of 66–83 µg L−<sup>1</sup> reported in **Table 6**. At this point the total mis-match of classifications is 28% and lies within the range of 24–34%.

## Overview and Comparison of the Nutrient Boundaries

Where national Good/Moderate and High/Good nutrient boundaries (Country; **Table 5**) within common IC types were available, those values were compared with the range of nutrient boundary values resultant from the application of the BPG toolkit analyses. When linear regression results were not significant then the results from the categorical approaches (Cat appr) are used and indicated instead (**Table 5**). However, where the differences between the nutrient concentrations in adjacent classes are not significant, quantiles derived methods need to be treated with

boundary range (25th, 75th). Bottom panels show range of nitrogen (TN) and phosphorus (TP) concentrations across ecological quality status (EQS) classes: High,

Good, Moderate, Poor and Bad. extreme caution (e.g., **Table 3**). Overall, the nutrient boundaries predicted by the different statistical approaches included in the

DISCUSSION AND CONCLUSION

similar within the common types.

The WFD (European Commission [EC], 2000) introduced, amongst other requirements, a comprehensive ecological status assessment of all surface waters, based on a number of biological, hydromorphological, chemical and physico-chemical quality elements (cf. Annex V 1.1 and V 1.2). Nutrient concentrations are only used as supporting parameters in the assessment of the ecological status. Coastal and estuarine nutrient concentrations are, however, key parameters for the management of eutrophication, since they can be directly linked

toolkit are broadly similar to those established by the MS, and

to nutrient inputs, which can be addressed by abatement measures (Vollenweider, 1992; Paerl et al., 2011). In this context it is important that EU MS set consistent and comparable nutrient boundaries.

In addition, marine strategy directive (MSFD) considers nutrient concentrations as indicators of equal importance as that of the biological indicators (Ferreira et al., 2011). Within the scope of the MSFD, nutrient levels (nutrient concentrations in the water column and nutrient ratios for nitrogen, phosphorus and silica, where appropriate) are amongst the relevant primary criteria in marine waters under Descriptor 5: "Human-induced eutrophication" (European Commission [EC], 2017). Setting consistent nutrient boundaries for the WFD and MSFD is therefore important for a consistent management approach across the continuum of transitional, coastal and marine waters. The recommendations proposed in the guidance and toolkit can promote such consistency, thus having important implications

TABLE 5 | Comparison of range of nutrient boundary values for TN and TP, obtained with linear regression and/or categorical approaches, with the range of national good/moderate and high/good boundary values for some transitional (TRW) and coastal (CW) waters common IC types across Europe (where data are available).


The symbol "<sup>∗</sup> " indicates higher disagreement is signaled.

in coastal and estuarine management, but we have found some problems during its application.

In many datasets the EQR values presented a significant percentage of the data (nearly 25%) beyond the expected EQR range [0–1], and often with very pronounced deviations. This occurred mainly where Chl a based EQR was being used for intercalibration. In such cases, only extreme EQR values were removed from the analysis, since removing all EQR values > 1 would decrease the amount of data available for analysis and, more importantly, could influence the observed statistical properties of the relationship between the phytoplankton BQE and the nutrient pressure. However, these EQR ranges may indicate a problem in the established reference conditions, in certain types. If the natural ranges of Chl a in the new datasets differ considerably from the ones used for establishing reference conditions and/or used in the intercalibration exercise, then the intercalibrated phytoplankton (Chl a) boundaries, defined within a 0–1 range, may compromise the prediction of robust nutrient boundaries. We suggest that these cases should be further scrutinized, in order to check the influence of this aspect in the predicted nutrient boundaries.

Some datasets have relatively few observations which may compromise their use to apply the regression analyses proposed and also some of the categorical ones, since results might not be robust and representative enough. Many do not have a proper coverage of the full gradient of disturbance, and in particular of the range of interest to derive nutrient boundaries, i.e., from High to Moderate status. Both situations might be partially overcome if datasets within common types are normalized and pulled together for the analyses. This would allow increasing the number of observations and the coverage of the gradient of disturbance. This is particularly relevant for MS lacking either good or bad quality samples/sites/conditions, since the full gradient of disturbance could still be captured at the scale of the common type.

FIGURE 3 | Relationship between TN and TP for Lithuania (common type BT1). Points colored by phytoplankton EQS class (H, blue; G, green; M, yellow; P, orange; B, red), dotted line marks the mean N:P molar ratio, broken orange line ratio of 15:1. Green and blue lines mark contours of the good/moderate and high/good boundaries predicted from the bivariate model. The vertical and horizontal dotted lines indicate the predicted boundaries and respective upper and lower range.

Evidence of potential interaction between nutrient parameters, of factors masking the pressure-response relationship (either positively, e.g., when the pressure is mitigated by other factor(s); or negatively, e.g., when multiple stressors occur simultaneously), and of overdispersion in the data, make good cases for the use of alternative statistical approaches, to univariate linear regressions. Other methods might be necessary to predict nutrient boundaries in such datasets. For example, when dealing with the lack of environmental information for predicting relevant features and potential sub-types across broadly defined types (as e.g., the NEA11 type); or with the likelihood of the presence of other pressures, quantile regression would still allow predicting boundaries. However, it must be noticed that, while boundaries derived from a high quantile may be appropriate (or even the single available option), when unmeasured pressures other than nutrients are downgrading the biological status, such boundaries are not precautionary and pose a high eminent risk of negative effects on the biota at those predicted values. It is often the case, where these pressures might be operating together with environmental factors to control phytoplankton growth dynamics, which makes it more difficult to disentangle the most relevant factors. The nutrient boundaries indicated by such an approach should therefore be taken with caution until additional environmental factors are considered and further guidance on adequate quantile selection for this purpose is developed.

Nevertheless, although data quality was demonstrated to be a frequent obstacle for deriving nutrient boundaries, the results show that for about 71% of the examples considered in the current study there is an agreement between the national nutrient boundaries established by the countries and the expected nutrient range derived from the application of the different toolkit approaches.

Most of the disagreements are on the TP boundaries. This fact could be related to the limiting nutrient effect. It is generally considered that nitrogen is the nutrient that limits primary productivity in most oceans (Tyrrell, 1999). This is not the case for the Mediterranean basin where phosphorus appears as the most important limiting nutrient, although it is closely followed

FIGURE 5 | Relationship of nutrient Dissolved inorganic nitrogen (DIN) concentrations (µM) with intercalibrated normalized nEQRs in the NE Atlantic estuaries (NEA11 common type), including observations from Netherlands (NL), Portugal (PT), Ireland (IE), Spain (ES) and the United Kingdom (UK): (left) scatter plot and linear trend, with sites colored according to ecological status (High, blue; Good, green; Moderate, yellow; Poor, orange; Bad, red) and (right) quantile regression fit (Additive Quantile Regression Smoothing rqss using quantreg R package by Koenker) for nEQR v DIN (µM), where horizontal lines indicate EQR boundaries at H/G and G/M, and vertical lines the nutrient boundaries, respectively for H/G and G/M, at the 70th quantile.

sub-sample of the data set selected at random).

by nitrogen in this limiting role (Krom et al., 1991; Estrada, 1996; Pitta et al., 2005; Thingstad et al., 2005). In estuarine and CWs, nitrogen vs. phosphorus limitation can change both temporally and seasonally, depending on the inputs from rivers, agriculture and sewage drainage (Painting et al., 2005). It is therefore important to consider local dynamics, as demonstrated for the Baltic, where different N:P ratio patterns among sites of a common typology suggest that type-specific nutrient boundaries determination may not be adequate across all sites. The example shown indicates that type-specific TN boundary values might be unnecessarily restrictive for phosphorus limited systems in some Polish waters of the common type TRW BT1. In lagoon ecosystems, Rinaldi et al. (1992) assume that the algal biomass is limited by nitrogen when the N:P weight ratio is lower than 5 and by phosphate for N:P values higher than 10, while the intermediary values of the ratio indicate that both nutrients regulate algal growth. This ratio can fluctuate, as it was demonstrated in the Papas Lagoon, depending on the season. MS should check the N:P ratio of their water bodies (based on previous studies or in the results of long term monitoring programs) before selecting the right nutrient parameter for the status assessment of their water bodies.

Analyzing the disagreements found between predicted and national boundaries, in the case of the North East Atlantic estuaries (NEA 11), the available intercalibration dataset has shown that the difference between Good and Moderate status



Results from regression and categorical methods are presented, those signaled (<sup>∗</sup> ) need to be taken with caution. "Possible range" refers to values derived from the interquartile range of the residuals.

was not significant, which compromises the robustness of the results obtained for most of the tested approaches. For the H/G range, however, boundaries suggested by the categorical approaches and quantile regression may be considered for guidance. The categorical and quantile regression results obtained with the analyzed dataset indicate G/M boundaries in line with United Kingdom boundaries adopted for medium to very turbid waters, but do not seem adequate to protect clearer waters. The boundaries are also not in line with the more stringent H/G or G/M French boundaries in either of their NEA11 national subtypes. These results reinforce the need to account for additional environmental factors when setting common datasets for establishing nutrient boundaries across common types, in order to accommodate within type natural variability, particularly for broadly defined types across Europe. A mixed dataset covering a wide gradient of more to less turbid systems would mask the relationship between nutrients and phytoplankton in clearer waters, as turbidity would control phytoplankton growth allowing for good ecological quality values to be attained at higher nutrient values than would be expected for example in non-turbid systems. This is the reason for the widely scattered data observed, and emphasizes the need to evaluate and interpret all values produced using the toolkit.

In the Mediterranean CWs of the common type III E, the disagreements between national boundaries and the values proposed by the toolkit are also observed when compared to the results included in Pagou et al., 2008. In fact, the TN thresholds (0.62–0.65 µmol L−<sup>1</sup> for CWs in good status) proposed by these authors are more in line with the range predicted by the toolkit statistical approaches tested.

Souchu et al. (2010) analyzed nutrients concentration along an anthropogenic eutrophication gradient in French Mediterranean coastal lagoons. TN and TP values in oligotrophic lagoons were around 220 and 13 µg L−<sup>1</sup> , respectively. Also in this case, and for both nutrients, the values are similar or within the H/G range predicted by the toolkit, but in disagreement with the boundaries proposed by France.

On the other hand, results have also shown that relationships found between nutrients and EQR of BQE phytoplankton have been stronger in coastal lagoons than in CWs. This is due to the fact that in CWs and in large and complex estuarine systems such as the Baltic Sea, the relationship between loads and nutrient concentrations is not as simple as in enclosed systems. Loss mechanisms (sedimentation, denitrification) and retention time play key roles but obscure the cause effect relationships. Therefore, we suggest categorical approaches for the establishment of nutrient standards in CWs.

Finally, considering other BQEs (e.g., macroinvertebrates or fish) would not improve the relationships with nutrients, because their responses are more affected by other pressures (e.g., physical disturbance, hydromorphological changes). The improvement of the relationships between BQE phytoplankton and nutrients could be possible if data are collected within a suitable spatiotemporal framework, with sufficiently frequent sampling over a reasonable period of time (Flo, 2017).

Many methods have been developed in the EU and elsewhere to evaluate and track trends in eutrophication in order to fulfill requirements of legislation designed to monitor and protect CW bodies from degradation. In this sense, many assessment tools (Bricker et al., 1999; OSPAR Commission, 2005; HELCOM Eutrophication Assessment Tool – HEAT) have stressed the importance of establishing a more tight linkage between causative factors (nutrients) and direct and indirect effects of eutrophication (Painting et al., 2005).

We conclude that the toolkit applied in the current work is a valuable tool for establishing the nutrient criteria, but attention has to be paid to the quality of data and to the importance of compiling a comprehensive dataset that covers a wider spectrum of conditions, containing as much as possible a balanced number of observations across several EQS classes (at least until the Moderate status).

## AUTHOR CONTRIBUTIONS

All authors were involved in conceiving the ideas and designing methodology. HT led the data analysis. FS led the writing of the manuscript. All authors contributed critically to discussion and gave final approval for publication.

## FUNDING

HT thanks FCT/MCTES for the financial support to the host institution CESAM (UID/AMB/50017/2019).

## REFERENCES

fmars-06-00255 May 16, 2019 Time: 16:37 # 13


status within the European marine strategy framework directive. Estuarine Coast. Shelf Sci. 93, 117–131. doi: 10.1016/j.ecss.2011.03.014



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Salas Herrero, Teixeira and Poikane. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Habitat Model of Eelgrass in Danish Coastal Waters: Development, Validation and Management Perspectives

Peter A. Staehr\*, Cordula Göke, Andreas M. Holbach, Dorte Krause-Jensen, Karen Timmermann, Sanjina Upadhyay and Sarah B. Ørberg

Department of Bioscience, Aarhus University, Aarhus, Denmark

During the last century, eutrophication significantly reduced the depth distribution and density of the habitat forming eelgrass meadows (Zostera marina) in Danish coastal waters. Despite large reductions in nutrient loadings and improved water quality, Danish eelgrass meadows are currently not as widely distributed as expected from improvements in water clarity alone. This point to the importance of other environmental conditions such as sediment quality, wave exposure, oxygen conditions and water temperature that may limit eelgrass growth and contribute to constraining current distributions. Recently, detailed local models have been set up to evaluate the importance of such regulating factors in selected Danish coastal areas, but nationwide maps of eelgrass distribution and large-scale evaluations of regulating factors are still lacking. To provide such nationwide information, we applied a spatial habitat GIS modeling approach, which combines information on six key eelgrass habitat requirements (light availability, water temperature, salinity, frequency of low oxygen concentration, wave exposure, and sediment type) for which we were able to obtain national coverage. The modeled potential current distribution area of Danish eelgrass meadows was 2204 km<sup>2</sup> compared to historical estimates of around 7000 km<sup>2</sup> , indicating a great potential for further distribution. While validating the modeled eelgrass distribution area in three areas (83–111 km<sup>2</sup> ) that hold large eelgrass meadows, we found an agreement of 67% with in situ monitoring data and 77% for eelgrass areas as identified from summer orthophotos. The GIS model predicted higher coverage especially in shallow waters and near the depth limits. Areas of disagreement between GIS-modeled and observed coverage generally exhibited higher exposure level, mean summer temperature and salinity compared to areas of agreement. A sensitivity analysis showed that the modeled area distribution of eelgrass was highly sensitive to light conditions, with 18–38% increase in coverage following an increase in light availability of 20%. Modeled coverage of eelgrass was also sensitive to wave exposure and temperature conditions while less sensitive to changes in oxygen and salinity conditions. Large regional differences in habitat conditions suggest spatial variation in the factors currently limiting the recovery of eelgrass and, hence, variations in actions required for sustainable management.

Keywords: eelgrass, model, summer orthophotos, monitoring, environmental conditions, management

#### Edited by:

Heliana Teixeira, University of Aveiro, Portugal

#### Reviewed by:

Mogens Rene Flindt, University of Southern Denmark, Denmark Irene Martins, Interdisciplinary Centre of Marine and Environmental Research (CIIMAR), Portugal

> \*Correspondence: Peter A. Staehr pst@bios.au.dk

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 29 August 2018 Accepted: 19 March 2019 Published: 04 April 2019

#### Citation:

Staehr PA, Göke C, Holbach AM, Krause-Jensen D, Timmermann K, Upadhyay S and Ørberg SB (2019) Habitat Model of Eelgrass in Danish Coastal Waters: Development, Validation and Management Perspectives. Front. Mar. Sci. 6:175. doi: 10.3389/fmars.2019.00175

## INTRODUCTION

fmars-06-00175 April 2, 2019 Time: 17:28 # 2

Benthic primary producers such as seagrasses play important ecological roles as hotspots for production, storage and export of organic carbon (Duarte et al., 2013; Duarte and Krause-Jensen, 2017) in addition to efficiently retaining nutrients, stabilizing sediments and stimulating biodiversity in shallow coastal ecosystems (Hemminga and Duarte, 2000). They also form important habitats for epifauna, fish and birds (Bostrom et al., 2014). Reductions in water clarity of shallow coastal waters, mostly due to eutrophication, have caused global losses and reduced depth colonization of seagrass meadows (Short and Wyllie-Echeverria, 1996; Orth et al., 2006). Historically, most of the Danish estuaries were dominated by the seagrass Zostera marina (eelgrass), but following the wasting disease in the 1930s, and partial recovery thereafter, the extent and depth distribution of eelgrass decreased markedly, as eutrophication reduced water clarity (Nielsen et al., 2002; Krause-Jensen et al., 2012; Bostrom et al., 2014). The distribution of eelgrass is central in coastal water management, partly because of the ecosystem functions and services it provides (Orth et al., 2006; McGlathery et al., 2012), but also because of relatively strong relationships between depth distribution and nutrient loading driven mostly by reductions in light availability (Nielsen et al., 2002). Eelgrass is therefore a key indicator species for the assessment of marine water quality in Europe (Guidance, 2009). Besides being sensitive to eutrophication, eelgrass meadows reflect and integrate changes in water quality over longer time periods making them an ideal indicator species (Krause-Jensen et al., 2008). A recent study shows that seagrass recovery is longer in systems with a long history of eutrophication altering several growth conditions (O'Brien et al., 2017). Despite large reductions in nutrient loadings and improved water quality in Danish coastal waters, eelgrass meadows are not as widely distributed as expected (Riemann et al., 2016). This indicating the importance of other environmental conditions delaying the recovery and growth of eelgrass. Besides light availability, eelgrass distribution is also related to oxygen and temperature conditions (Koch, 2001; Pulido and Borum, 2010; Raun and Borum, 2013), salinity and nutrient regime (Krause-Jensen et al., 2000; Carstensen et al., 2013). Wave and current exposure and sediment conditions are also important controlling factors (Short et al., 2002; Rasmussen et al., 2009; Yang et al., 2013; Kuusemäe et al., 2016). In addition, overgrowth by epiphytes stimulated by nutrients (Sand-Jensen, 1977; Borum, 1985) the negative physical impact from drifting macroalgae (Canal-Verges et al., 2014), burial of seeds and seedlings by the bioturbation generated by lugworm (Valdemarsen et al., 2011), substrate competition from pacific oyster (Tallis et al., 2009) and a range of anthropogenic activities including dredging and fisheries (Hilary et al., 2005) may affect eelgrass distribution and performance.

Recently, in some local Danish areas, detailed models have been set up to evaluate the importance of such regulating factors (Canal-Vergés et al., 2016; Flindt et al., 2016; Kuusemäe et al., 2016). However, a nationwide map demonstrating the eelgrass distribution and density in Danish coastal waters is still lacking. To provide such nationwide information, we applied a spatial modeling approach, which combines information from different spatial layers available on the important regulating environmental conditions mentioned above. The aims of this Geographical Information System (GIS) model are to provide a tool that enables a nationwide description of the potential spatial distribution and density of eelgrass. Furthermore, the GIS tool aims to enable and evaluate the importance of key environmental conditions regulating eelgrass distribution, to identify limiting factors for the expansion of current eelgrass populations and thereby guide sustainable management of the meadows. In this study we describe the data layers applied in the nationwide GIS model, and how we combine these to provide a map of the potential distribution of eelgrass. By potential, we mean the most likely distribution, given the status of a number of key environmental conditions known to represent habitat requirements of eelgrass. To evaluate the GIS model output, we compared the output with spatial eelgrass distribution data obtained from summer ortho photos (SOP) and in situ eelgrass monitoring data from the Danish nationwide marine monitoring program (NOVANA). We tested for possible differences in environmental conditions between areas where the GIS model and observations both agree and disagree. Finally, we performed a sensitivity analysis of the available data layers to evaluate their relative importance and discuss future management and optimization perspectives of the model.

#### MATERIALS AND METHODS

#### Study Area

The study area covers all Danish coastal waters including the Kattegat, the Danish straits and the Wadden Sea as well as estuaries, lagoons, bays and open stretches along the coastline. In total, more than 7000 km of coastline of shallow waters (<11 m depth) corresponding to 13125 km<sup>2</sup> seafloor are included in this study (**Supplementary Figure S1**). Although the various water bodies are very different, they are almost all characterized as eutrophic with turbid waters, organic enriched sediments, with some of them exposed to increased risks of anoxia (Riemann et al., 2016).

#### Environmental Data

Environmental parameters were selected a priori based on existing knowledge on eelgrass habitat requirements and constraints on growth and distribution. These variables included bathymetry (depth), bottom water temperature, salinity and oxygen concentrations, light attenuation, sediment characteristics and physical exposure to waves and tides (**Table 1**). While other biotic and abiotic conditions have proven to be relevant (**Figure 1**), we had to simplify our model approach to the availability of data on a national scale. Data for estimating the pelagic variables (temperature, salinity, oxygen, light attenuation, and Secchi depth) were provided by the NOVANA monitoring program for the period 1994–2010. Pelagic variables represent sampling biweekly to monthly throughout the year by local departments of the Danish Nature Agency and the results are

TABLE 1 | Description of GIS data layers used to model the probability of eelgrass with a spatial resolution of 100 m × 100 m.


reported to a national database<sup>1</sup> maintained by the Danish Center for Energy and Environment (DCE) at Aarhus University. Data on temperature, light attenuation and salinity were averaged over the summer season (April to October), except for oxygen where we extracted information on the frequency of low oxygen concentrations to assess the potential impact of hypoxia/anoxia on eelgrass distribution. To provide bottom temperature, bottom salinity and frequency of low oxygen concentrations (<2 mg/L) for the geographic area relevant for eelgrass, we applied a geographically weighted regression tool in ArcMap, with the depths information of the observations as explanatory variable. In the extrapolation analysis we separated the coastal zone into

<sup>1</sup>https://oda.dk/main.aspx

Frontiers in Marine Science | www.frontiersin.org

open waters and fjords. A similar approach was used to estimate light attenuation (KD) at shallow depth. To improve the number of light attenuation observations K<sup>D</sup> was estimated both from light profiles and Secchi depths observations as described in the section on light index. Afterward we combined the data into a single raster layer per parameter.

The GIS data layer on sediment characteristics was provided by the Geological Survey of Denmark and Greenland (GEUS) and consists of seven sediment classes: bedrock, hard bottom complex, gravel, sand, hard clay, muddy sand and mud. The sediment map has a scale of 1:250000. The coverage of the seafloor features varies within the area depending on the resolution of the different surveys (Cameron and Askew, 2011). The bathymetric maps for the Kattegat, inner Danish waters Danish estuaries and coastal zones were provided by the Danish Maritime Safety Administration at a nominal resolution of 50 m. Data on wave exposure and current velocities in the Wadden Sea and Western coastline of Jutland were provided by a hydrodynamic North Sea model (MIKE, DHI), whereas wave and current energy at the seafloor (hereafter named exposure) for the inner Danish waters and estuaries were provided from EUseamaps. Finally, as the spatial layers varied in resolution we aggregated data and exported these as gridded files with a spatial resolution of 100 m × 100 m.

#### Model Approach

Our modeling approach largely follows the recommendation by Hirzel and Le Lay (2008) for habitat suitability models. The model assumes that the coverage of eelgrass to a large extent reflects the suitability of the coastal habitats for eelgrass growth and survival. In addition, we assume that this overall suitability can be evaluated based on spatial information on a number of key environmental conditions for which we used the most suitable data.

The model consists of spatial data layers with environmental parameters (input data), which are either used directly in the eelgrass model or used to generate derived data layers, which are important for eelgrass growth (**Figure 1**). For each of the six environmental parameters we explored the relationship with in situ measured eelgrass coverage observed along transects. For a given spatial layer (e.g., bottom temperature) we extracted data on eelgrass coverage and the environmental value (here bottom temperature) on a pixel (50 × 50 m) level. The eelgrass coverage within the pixel was calculated as a mean value within the range of 0–100% coverage.

Correlations between eelgrass coverage and each of the six environmental conditions were modeled using relations derived from statistical analysis of the present data in combination with information from the literature on important thresholds, if available. The resulting data layers were then transformed into continuous index functions with a value between 0 (i.e., eelgrass growth/survival is impossible) and 1 (i.e., the parameter is not limiting eelgrass). Below, we briefly present the individual index functions. A detailed description of the index functions is available in the **Supplementary Material**.

Although habitat suitability models can be derived from discrete functionalities (Brooks, 1997), continuous non-linear functionalities (or smoothened threshold functions) are commonly used (Hirzel and Le Lay, 2008). Non-linear functions were chosen as these provided better fits and commonly used to estimate threshold values (Andersen et al., 2009).

#### Light Index

To develop a light index we used both light extinction coefficient (KD) derived from light profiles and from Secchi depth converted to K<sup>D</sup> following the method of Murray and Markager (2011). Here we parameterized the importance of the actually available light reaching the seafloor, rather than different indices of water clarity. First, we applied a standard equation (Beer's law) to model bottom light levels (Iz) from measured light attenuation coefficients (KD), light at the Sea surface (Isurface) and water column depth as:

$$\mathbf{I}\_{\mathbf{z}} = \mathbf{I}\_{\text{surface}} \cdot \mathbf{e}^{-\mathbf{K}\_{\text{D}} \cdot \text{depth}} \tag{1}$$

For Isurface, a constant value (384 µmol photons m−<sup>2</sup> s −1 ), calculated as the mean daily light reaching the sea surface during summer (April to October) based on a 10 years data set from a measurement station at Højbakkegaard, Denmark. Rather than an arbitrary percent of surface light, we choose to evaluate the importance of light by comparing I<sup>Z</sup> with the eelgrass coverage (in pct) observed at depth during a period from 1994 to 2010. This resulted in a data set of 7750 observations of pct cover and I<sup>z</sup> (**Supplementary Figure S2A**). Data on light at depth represents cell sizes of 50 × 50 m compared to the mean eelgrass cover within the associated transects.

To deduce a pattern from the scattered cover vs. light data set (**Supplementary Figure S2A**), we calculated percent cover for a range of binned light data. This resulted in a bell shaped curve, which we then normalized to 1 (**Figure 2**). The light index model (red line in **Supplementary Figure S2B**) was parameterized using a combination of a polynomial function shown as a dotted blue line. Also, we included a threshold for the minimum light required for eelgrass plants to have a net positive growth (IC). A first condition was therefore that I<sup>Z</sup> should be >I<sup>C</sup> for eelgrass to occur. The critical light level is known to vary as a function of water temperature and to display a strong seasonal variation, reflecting physiological acclimation of plants to prevailing light, temperature and nutrient conditions (Staehr and Borum, 2011). However, to simplify our light model we choose to apply an I<sup>C</sup> value of 25 µmol photons m−<sup>2</sup> s −1 , which represents light requirements when summer mean water temperature ranges between 11 and 17◦C with a median of 15◦C (Staehr and Borum, 2011). Data show an apparent negative effect of high light levels (**Supplementary Figure S2A**), which we assume is an artifact driven by effects of other factors, such as physical exposure at shallow depth. Therefore, above the peak of the polynomial model (140 µmol photons m−<sup>2</sup> s −1 ) we assigned a light index value of 1. Below the I<sup>C</sup> value of 25 µmol photons m−<sup>2</sup> s −1 , the index value was set to zero. The light index model is described in Equation (2):

$$\text{Light}\_{\text{index}} = \text{IF(I}\_{\text{Z}} > 140 \text{ THEN 1}, \text{ELSEIF(140} > \text{I}\_{\text{Z}} > 25)$$

$$\text{THEN (0.000000047 \* I}\_{\text{Z}} "{}^{3} - 0.0000515 \* I\_{\text{Z}} "{}^{2}$$

$$+ \text{ 0.0135 \* I}\_{\text{Z}}), \text{ELSEIF I}\_{\text{Z}} < 25 \text{ THEN 0}$$

#### Sediment Suitability Index

Sea bottom characteristics such as organic matter content affect the ability for eelgrass to become established, as well as the suitability for anchoring, growth and seed survival (Krause-Jensen et al., 2011; Flindt et al., 2016). In this GIS model, we applied a national wide sediment map with information on seven sediment categories (Cameron and Askew, 2011), which we merged into three substrate groups: (1) Mud, sandy mud, muddy sand and bedrock, (2) Sand (3) Hard bottom complex and gravel and coarse sand. To include information on sediment conditions as a determining factor for eelgrass distribution we assigned the following sediment index values ranging between 0 and 1 (unsuitable to suitable): Mud, sandy mud, muddy sand and bedrock (grp1) = 0.1; Hard bottom complex and gravel and coarse sand (grp3) = 0.5; Sand (grp2) = 1. The eelgrass coverage in these three groups is shown in **Supplementary Figure S3**. The final sediment index model is described in Equation (3)

$$\text{Sediment}\_{\text{index}} = \text{IF} \left( \text{grp} = \text{ 1 THEN 0.1, ELSE} \right) \tag{3}$$

$$\text{(IFGRP = 2 THEN 1, ELSE 0.5)}$$

#### Physical Exposure Index

Waves and tides limit eelgrass distribution especially in shallow waters where bottom exposure is most pronounced. To parameterize the effect of the physical exposure on eelgrass cover, we combined two data sources which both have weaknesses or data gaps in sub regions: (1) a map of the frequency with which the wave base reaches the bottom; produced by DHI covering the Danish waters without Limfjorden and (2) an EUSeaMap wave energy map (Log transformed) covering the North Sea including Limfjorden and the southwestern Baltic Sea but not the waters around Bornholm. Before combining the two layers, they were normalized.

Comparing the modeled exposure levels with eelgrass coverage on a pixel basis (5505 observations), showed that eelgrass coverage declined strongly when the normalized physical exposure exceeded a value around 0.2 (**Supplementary Figure S4A**). To parameterize the effect of physical exposure we applied a function, which represents the red line in **Supplementary Figure S4B**. Below the cut off value of 0.2, exposure levels are considered to be sufficiently low to enable full coverage of eelgrass, providing an index value of 1. In parallel to the reasoning conducted for eelgrass cover as a function of light, we interpreted that the apparent decline in eelgrass cover at low levels of physical exposure is largely due to deeper water where light levels limit distribution. The final exposure index model is described by Equation (4).

$$\text{Exposure}\_{\text{index}} = \text{IF} \left( \text{(exposure} > 0.2 \right) \text{THEN} \right)$$

$$\text{(20\*EXP (}-15\*\text{ exposure})\text{), ELSE I)}\qquad \text{(4)}$$

#### Temperature Index

To develop a temperature index, we made a map of summer mean (April to October) bottom water temperatures by interpolating

data from CTD profiles. These modeled bottom temperature data were then combined with the gridded transect observation values (7006 observations). A comparison of eelgrass cover and bottom temperatures in the gridded cells are shown in **Supplementary Figure S5A**. This resulted in a bell shaped curve somewhat similar in shape to what has previously been shown in experimental studies (Nejrup and Pedersen, 2008; Staehr and Borum, 2011). To deduce a pattern from the scattered cover vs. water temperature plot (**Supplementary Figure S5A**) we calculated percent cover for a range of binned temperature data, which we then normalized to range between 0 and 1 (**Supplementary Figure S5B**). The temperature index model was parameterized using a polynomial function (Equation (5)):

$$\text{Tempindex} = -0.006079 \ast \text{Tw}^4 + 0.3074 \ast \text{Tw}^3$$

$$-5.7604 \ast \text{TW}^2 + 47.587 \ast \text{Tw} - 146.49 \tag{5}$$

where Tw is the mean summer bottom water temperature.

#### Oxygen Index

We compared eelgrass coverage at local sites with corresponding oxygen concentrations, obtained from oxygen depth profile measurements within the national monitoring program. As we did not have high frequency continuous oxygen data available, we used monitoring data to determine the frequency of low oxygen concentrations (<2 mg/L) at the sea floor during the summer growth season (April to October). These frequencies of low oxygen data were then combined with the gridded transect observation values providing 5576 observations (**Supplementary Figure S6A**). We interpret the low oxygen frequency data as an information layer reflecting the sensitivity of eelgrass to low oxygen conditions. According to this, we expect low coverage of eelgrass due to mortality and poor performance in areas with high occurrence/frequency of low oxygen.

The negative impact of low oxygen conditions was included in the model by an oxygen index function (Equation (6)):

Oxygenindex = 1 − ((1/(1 + EXP( − (DOLow ∗ 20) + 6)))) (6)

Where DOLow is the relative frequency (ranging between 0 to 1) of days per summer (May to September) with observed DO at depth below 2 mg/L (**Supplementary Figure S6B**).

#### Salinity Index

An experimental study of the effect of salinity on several eelgrass performance parameters showed that optimal salinities range between 10 and 25 psu, with increased mortality and lowered performance at low salinities and moderate effects at high salinities (Nejrup and Pedersen, 2008). The potential success of eelgrass is therefore expected to be reduced in areas of low salinity, but remains only weakly affected by high salinities. To evaluate this further, we compared eelgrass cover at 7750 sites with summer mean bottom salinities (**Supplementary Figure S7A**). The comparison does suggest a somewhat bell shaped dependency/effect of salinity. Since no experimental evidence exists for the apparent negative effect of high salinities, we expect that this arises from other confounding factors such as higher levels of exposure in high saline waters and lower light levels at larger depths where salinities usually increase during stratification.

The effect of low salinity on plant performance and survival was included in the model by a salinity index function (Equation (7); **Supplementary Figure S7B**):

$$\text{salinity}\_{\text{index}} = \text{IF}\left(\{\text{salt} < 7 \,\text{THEN}\left(\text{EXP} 1\* \{\text{salt} - 7\}\right)\right),$$

$$\text{ELSE I}\{\}\tag{7}$$

Where salt is the bottom water salinities determined from CTD profiles and interpolated to the corresponding eelgrass monitoring sites.

#### Eelgrass Habitat Model

We tested different combinations of the six environmental index maps to produce a national map of the potential habitat occupied by eelgrass plants in Danish coastal waters. In all cases the environmental index maps were combined to produce an eelgrass habitat map with values ranging from 0 to 1, reflecting 0–100% expected presence of eelgrass. As suggested by Hirzel and Le Lay (2008) we tested different combination of the data layers in our GIS model. These varied in the prioritization of the importance of the different data layers, especially light and physical exposure which we expected to be of greater importance. However, a comparison between the measured and modeled cover showed that a simple multiplicative model, which gives equal importance to each data layer (Equation (8)), was best:

$$\text{Eelgrassindex} = \text{Light}\_{\text{index}} \* \text{ExpSource}\_{\text{index}} \* \text{Tempindex}$$

$$\ast \text{Oxygen}\_{\text{index}} \* \text{Sodium}\_{\text{index}}$$

$$\ast \text{ salinity}\_{\text{index}} \tag{8}$$

Conversion of parameter layers to index layer and calculation of the eelgrass model index was done on the 100 m × 100 m resolution raster layers with arcpy in ArcGIS 10.3.

#### Model Performance

To assess the performance of the GIS probability map (Zost GISmodel), we performed a validation using independent datasets of eelgrass cover from recent years (**Table 2**). One data set (Zost Map2012) covered three selected study areas encompassing 292 km<sup>2</sup> within 0–5 m depth. The areas were Nibe-Gjøl Bredning in Limfjorden, Saltholm including the Zealand coast facing Saltholm, and the South Funen Archipelago. The eelgrass coverage in these three areas were derived from a pixel based supervised image analysis of aerial SOP from 2012 performed with linear discriminant analysis of color bands (red, green, blue) involving two classes, 'eelgrass' and 'bare sand' (Ørberg et al., 2018). Indeed, other marine vegetation was most likely classified as eelgrass, but the mapping was performed in areas with eelgrass as the dominant vegetation type. A second data set used to validate the GIS modeled distribution was based on NOVANA monitoring transects from 2012 covering Nibe-Gjøl Bredning, Saltholm, South Funen (Zost Moni2012) and all Danish coastal waters also in 2012 (Tot Zost Moni2012). Finally, we used a high-resolution ground truth data set from Mariager Fjord (Zost Obs2009) (Clausen et al., 2015).

Staehr et al. Eelgrass Habitat Model

TABLE 2 | Data layers used in the validation of the map derived from the GIS model predicting potential distribution of eelgrass in Danish coastal waters.


Prior to validation of the GIS model coverage with SOPs and monitoring data points, all data layers were converted into raster's, matching the bathymetry grid size (50 m × 50 m). Hence, the Zost Map2012 values (presence = 100, absence = 0) were averaged within each grid, resulting in values ranging between 0 and 100. Similarly, the Zost Moni2012 and Tot Zost Moni2012 data points (pct cover) were averaged within each grid and the Zost GISmodel was converted from a 100 m × 100 m grid to a 50 m × 50 m grid. We did not average the Zost Obs2009 data within each grid because there was approximately one data point per pixel.

With equal grid sizes of 50 m × 50 m, we translated the reshaped data layers into presence/absence data by the following rules: Values <10 define the absence of eelgrass, and values ≥10 define the presence of eelgrass. Hereafter, we calculated the accuracy of the Zost GISmodel displayed in confusion matrices. To identify areas of particularly low/high accuracy compared to the Zost Map2012, we created maps displaying the agreement/disagreement between Zost GISmodel and Zost Map2012 for the three selected areas with shallow water. Similar to recent evaluations of eelgrass distributions, we used 10% cover of eelgrass as a threshold for defining areas where eelgrass meadows are present (Bostrom et al., 2014).

#### Importance of Environmental Conditions

We applied two different approaches to evaluate the importance of variations in the applied environmental conditions. These consisted of a statistical analysis and a sensitivity analysis. The statistical analysis investigated the importance of environmental variables for areas showing disagreement between predicted and measured eelgrass cover. Here we defined three groups of data: Agreement (grp 0) between predicted (Zost GISmodel) and observed (Zost Map2012); Prediction of eelgrass presence while Zost Map2012 show no eelgrass (−1) or vice versa (1). We applied a non-parametric Kruskal–Wallis test and a multiple comparison post hoc test (nemenyi from Desctools) in R to test whether environmental parameters differed significantly between the three groups. We only studied parameters for which eelgrass habitat requirements and thereby thresholds are less understood (physical exposure, oxygen, temperature, and salinity) compared to the basic requirements, such as light and sediment type.

To investigate the sensitivity of the eelgrass distribution (km<sup>2</sup> ) calculated by the GIS model, we performed a series of model runs where each input data layer (except the sediment map) was varied separately between −20 and + 20%. To gain information about possible regional differences in the sensitivity of the selected variables, we divided the Danish waters into three regions, covering Limfjorden, the Kattegat and the Eastern Baltic Sea (**Supplementary Figure S8**).

### RESULTS

Based on the eelgrass spatial habitat model, we produced a nationwide map (Zost GISmodel) of the potential distribution area of Danish eelgrass meadows (**Figure 2**). From this map, we calculated the total potential eelgrass distribution area in Danish waters to be 2204 km<sup>2</sup> (probability >10% × pixel size × pixel numbers). Furthermore, we compared the Zost GISmodel with ground truth observations and orthophoto mapping of eelgrass cover from recent years.

## Model Validation With Monitoring Data and SOP

As an initial step, we calculated the accuracy of the Zost GISmodel and the SOP based map (Zost Map2012) compared to in situ observations (Zost Moni2012) on eelgrass cover in three shallow areas (a total of 538 pixels). The GIS modeled distribution showed an accuracy of 67.1% (**Table 3A**). In comparison, the Zost Map2012 displayed a higher accuracy (80.3%) (**Table 3D**), thereby serving as a relevant map to validate Zost GISmodel on a larger spatial scale (**Table 3B**). While the accuracy of Zost Map2012 varied between the three areas (**Figures 3B–D**) the level of agreement with in situ data was overall high and stable across all depth intervals (**Figure 3A**). In comparison, the level of agreement of the Zost GISmodel was lower and tended to increase with depth (**Figure 4A**). The low accuracy, caused by the Zost GISmodel underestimating eelgrass distribution at shallower depths, was most pronounced in Nibe-Gjøl Bredning (**Figure 4B**). In comparison, both the Zost GISmodel and the SOP map overestimated eelgrass distribution compared to in situ observations in the very shallow areas (0–1 m) at Saltholm and South Funen (**Figures 3C,D**, **4C,D**).

A direct comparison of the eelgrass distribution between the Zost GISmodel and the SOP map resulted in an overall agreement of 77.3% (a total of 114938 pixels) in the three case study areas (**Table 3B** and **Figures 5A–D**). Disagreements

TABLE 3 | Confusion matrices comparing classification results of the GIS model (Zost GISmodel) with (A) in situ monitoring data in three smaller areas (Zost Moni2012); (B) 2012 aerial orthophoto image analysis in three smaller areas (ZOST MAP2012); (C) in situ monitoring data in all monitored areas in 2012 (Tot Zost Moni2012). In (D) we compare results from in situ monitoring and orthophotos in 2012 and finally in (E) we evaluate the GIS model performance against high-resolution from Mariager fjord in 2009. For each pixel, data were categorized into presence or absence of eelgrass, while less than 10% cover was considered as absence. Correct classification gives the % of pixels classified correctly to each category and in total.


between the Zost GISmodel and the SOP maps in 2012 were highest in the very shallow areas (38%). However, this decreased with depth, suggesting that the Zost GISmodel mainly underestimated eelgrass presence in the 0–1 m depth interval (**Figure 5A**), particularly around Saltholm (**Figure 5C**). The level of agreement/disagreement between the GIS modeled, SOP mapped and in situ monitored distribution of eelgrass were visualized in maps representing the three case study areas (**Figure 6**). While agreement between GIS and SOP maps dominate all three areas, they all have deeper zones where only in situ monitoring data indicate the presence of eelgrass.

In addition to the three case study areas, we evaluated the nationwide accuracy of the Zost GISmodel against an in situ data set representing all Danish coastal waters (Tot Zost Moni2012). We found an overall accuracy of 48.9% of the modeled eelgrass cover (a total of 2882 pixels) (**Table 3C**). While there was a high agreement (76.2%) in areas where in situ data show presence of eelgrass, the agreement was much lower (28.2%) in areas where in situ data show less than 10% coverage of eelgrass (**Table 3C**), suggesting that the Zost GISmodel most often overestimates the eelgrass cover. The level of disagreement between in situ observations and the Zost GISmodel was consistent throughout all depths (**Figure 7**). A similar comparison with high-resolution in situ observations from Mariager fjord 2009 (Zost Obs2009, **Supplementary Figure S9**) gave an overall accuracy of the Zost GISmodel of 30.4%, which also highlights that the agreement is lowest in areas where in situ data show less than 10% coverage of eelgrass (**Table 3E**).

#### Importance of Environmental Conditions

Comparing monitored eelgrass data with the geographically interpolated environmental data layers, showed that eelgrass coverage was higher in areas characterized by high light levels, shallow depth, high wave exposure, low frequency of oxygen depletion and higher temperatures (**Table 4**). Although the dependency of eelgrass cover was far from linear (see **Supplementary Material**), we expect the GIS model to perform overall well in such areas. As a first approach to investigate the importance of environmental conditions for the accuracy of the GIS model, we calculated the levels of three of the selected environmental parameters (exposure level, mean summer temperature and salinity) in areas with agreement and disagreement with the orthophoto mapped eelgrass areas in 2012. For areas of disagreement, we found that we found that mean exposure level, mean summer temperature and mean salinity were all significantly higher (p < 0.0001) higher compared to areas of agreement between the Zost GISmodel and SOP validation data. In areas of disagreement, exposure was higher by 0.063 exposure units, mean summer temperature was higher by 0.17◦C, and mean salinity was higher by 1.16 psu. As disagreement may also depend on habitat characteristics, we further tested whether all three possible outcomes [i.e., 0 = Agreement, 1 = the orthophoto map (Zost Map2012) predict eelgrass while the GIS model does not, −1 = GIS model predict eelgrass while Zost Map2012 does not] differed in level/intensity of each environmental parameter. The exposure level differed between all groups with significantly higher exposure level in areas where the GIS model underestimated eelgrass coverage (**Supplementary Table S1**). Small, but significant differences in mean summer bottom temperature and salinity between the three groups were also apparent. This suggests that the GIS model is currently less reliable in areas where exposure levels are higher than 0.33 on the exposure scale, where mean

Funen. (A) All areas combined, (B) Nibe Gjøl bredning, (C) Saltholm, and (D) South Funen. Green displays the percentage of all pixels where the models agree. Orange and blue display the disagreement between models. Blue show the percentage of pixels where Zost Map2012 displays the presence of eelgrass when Zost Moni2012 displays the absence of eelgrass. Orange is vice versa.

summer temperatures are higher than 14.7◦C and where salinity is higher than 15.1 psu. The correlation analysis of relationships between the applied environmental variables suggested that the high levels of wave exposure prevailed at shallow depths in both Kattegat, Limfjorden and the Baltic Sea regions (**Table 4**). Moreover, temperatures tend to increase toward shallow waters (except Limfjorden), while salinities increased with depth. Areas associated with higher light levels were, as expected, associated with shallower depth, and higher wave exposure, but lower frequency of low oxygen conditions.

Varying the value of each environmental variable in each pixel between −20% and +20% relative to baseline conditions, provided a simple sensitivity assessment of the GIS model to the applied data layers and their parameterization of eelgrass suitability except for the sediment conditions. Comparing three regions, the Kattegat, Limfjorden and the Baltic Sea (**Figure 8**), we found a strong sensitivity in all areas to variations in light conditions. The modeled eelgrass area was also quite sensitive to changes in wave exposure, low oxygen conditions and low bottom water temperatures, while salinity did not appear to have any influence. Limfjorden was the most sensitive area for changes in light and wave exposure, but less sensitive to low oxygen conditions. All three regions were surprisingly sensitive to changes in temperature. However, while higher temperatures had a negative impact on eelgrass coverage in Limfjorden and the Baltic Sea, the sensitivity analysis suggested that the coverage would increase significantly in the Kattegat area with increasing temperatures (**Figure 8**).

## DISCUSSION

### Model Performance

Applying different independent data sets (in situ monitoring data and SOPs) to validate the GIS model showed overall good agreement with the GIS model. In addition, the GIS model performance seems reasonable when comparing with previous modeling efforts (Krause-Jensen et al., 2003; Bekkby et al., 2008). However, model performance varied greatly between geographical areas, and the model did not perform well in predicting small-scale distribution patterns. It should be noted that the GIS model aims to provide estimates of the potential distribution and cover of eelgrass given a combination of key environmental conditions, for which we have spatial data available at the national scale. The overall (national scale) good agreement between model results and data indicate that the

presence of eelgrass when Zost Moni2012 displays the absence of eelgrass. Orange is vice versa.

model contains the main controlling parameters determining eelgrass distribution, whereas factors not included in the model (e.g., drifting macroalgae, epiphytes and bioturbation) locally may play a significant role for eelgrass growth and distribution.

The GIS model estimated a total area of eelgrass with 10% coverage or more to be 2204 km2, which is close to the range recently estimated by Bostrom et al. (2014) (ca. 1400–2100 km<sup>2</sup> ). In agreement with the empirical data on which the GIS model was parameterized, our model validation provided high eelgrass coverage in areas characterized by shallow sheltered waters such as the South Funen area. This area is also know to host many water birds which depend on eelgrass as a food source (Clausen, 2000).

Under-estimation of eelgrass presence at shallower depth in some areas such as Nibe-Gjøl Bredning suggests that the environmental thresholds for eelgrass presence in the GIS model may be too conservative. However, in other areas such as Saltholm, both the GIS model and the SOP maps indicated higher eelgrass distribution compared to in situ monitoring. While higher estimates of eelgrass cover by the GIS model may indicate potential areas of near future colonization, deviations may also reflect inaccuracies in the in situ data or simply inadequacies in the GIS model. Interestingly, the shallow areas around Saltholm (0–1.5 m) are known to be dominated by other rooted macrophytes such as Ruppia sp. rather than eelgrass (Noer and Petersen, 1993; Krause-Jensen and Christensen, 1999). Concerning inaccuracies in the in situ data, our comparison of the GIS model with a high-resolution data set from Mariager Fjord, showed that the GIS model overestimated the current distribution of eelgrass. However, similar to Saltholm, the vegetation at shallow depths in Mariager Fjord was dominated by Ruppia species (Clausen et al., 2015). Overestimation of eelgrass cover by both the GIS model and the SOP maps at shallow depth suggest that they are not capable of distinguishing between different rooted macrophytes such as Ruppia sp. and eelgrass. Future work should investigate specific habitat requirements and possibilities to distinguish RGB signals of different species. Incorporating biotic parameters, such as interspecific competition and foraging, would likely improve the GIS model. Currently, we must acknowledge that both the SOP maps and the GIS model to some extent estimate coverage of not just eelgrass, but rooted macrophytes in general. While the

the percentage of all pixels where the models agree. Orange and blue display the disagreement between models. Blue show the percentage of pixels where Zost Map2012 displays the presence of eelgrass when Zost GISmodel displays the absence of eelgrass. Orange is vice versa.

GISmodel displays the presence of eelgrass when Tot Zost Moni2012 displays the absence of eelgrass. Orange is vice versa.

Blue shows where Zost Map2012 displays the presence of eelgrass when Zost GISmodel displays the absence of eelgrass. Orange is vice versa. Ground truth data from NOVANA 2012 of eelgrass cover are annotated with points grading from gray to dark purple, where gray = 0% cover and dark purple = 100% cover. Gray line marks 2.5 m depth contour line.

GIS model compares reasonably well with the overall trends in eelgrass cover, the current coarse resolution makes it inadequate to thoroughly investigate conditions determining the distribution and size of eelgrass patches. Such analysis would be possible with the detailed SOP's which gives information at 10–20 cm scales in shallow waters. While this would be interesting, this was outside the scope of this current study, which primarily focuses on developing a model that enable us to understand changes at the landscape scale.

Perhaps a large source of uncertainty and error in our GIS habitat suitability model is the use of extrapolation of pelagic data (oxygen, temperature, salinity, light attenuation) from central stations in deeper waters into the nearshore shallow waters. Unfortunately, there are very limited datasets available for Danish waters for the required parameters. Aggregation of a new nationwide coastal dataset was therefore only possible by extrapolation. One key thing to be noted is that the central stations are not really deep, as Danish waters are quite shallow in general with depths typically of less than 20 m at the central sampling stations. Comparisons with a limited data set from the shallow Roskilde Fjord suggests that measurements in the upper part of the water column represented conditions in the shallow sites reasonably well (data not shown). In addition, the station grid is quite dense and as the calibration of the GIS model was done with the extrapolated values, the actual values have less importance. Furthermore, the morphometry and environmental conditions vary substantially in the Danish coastal zone (Conley et al., 2000), hence any error associated with extrapolation will not be systematic.

#### Importance of Environmental Conditions

In agreement with recent spatial predictive probability models, our GIS model also predicts that the probability of finding eelgrass is highest in shallow and sheltered areas (Bekkby et al., 2008), where light conditions are within the optimal range for the species (Canal-Vergés et al., 2016; Flindt et al., 2016; Kuusemäe et al., 2016). These areas were also highly represented by independent data sets based on in situ monitoring and SOP's. Comparing levels of the environmental variables in areas with agreement and disagreement with the independent validation data sets indicated that disagreement was higher in areas with elevated levels of exposure, as well as temperature


TABLE 4 | Spearman correlation analysis of relationships between environmental variables used in the GIS model to estimate eelgrass coverage.

DOlow is the frequency of dissolved oxygen concentrations below 2 mg/L, Temperature and salinity represents mean summer bottom values. I<sup>z</sup> is an estimate of the mean summer light intensity reaching the seafloor. Depth is the depth (m) at which eelgrass cover (%) was recorded. Values highlighted in bold are significant at p < 0.05. n is the number of observations.

and salinity, although the latter were significantly less important. Underestimation by the GIS model therefore mostly occurred in areas with high exposure levels, suggesting a high sensitivity to this variable. A recent model of eelgrass coverage in two Danish fjords, similarly showed that physical exposure, in terms of waves has a strong negative impact on eelgrass growth and distribution (Kuusemäe et al., 2016). The physical exposure data layer applied in our analysis had to be merged from two normalized data sets to cover the entire Danish area. This disabled us from using a physical unit which would have been preferable. However, as the GIS model was calibrated with the derived data, the actual values have less importance. Our highest concern was to obtain a homogeneous data set. A nationwide detailed map of physical exposure should be developed for future models of eelgrass habitats.

A sensitivity analysis of environmental variables underlines that light is a strong determinant of the depth distribution of eelgrass in the Danish coastal waters. This agrees well with the established statistical relationship between maximum eelgrass colonization depth and water transparency, as measured by K<sup>D</sup> or Secchi depth (Nielsen et al., 2002; Carstensen et al., 2013). The applied light dependency in our GIS model included a minimum light requirement threshold value which is known to exist for eelgrass (Staehr and Borum, 2011). However, our data set did not support the importance of such a clear minimum light threshold for the coverage of eelgrass. Rather, eelgrass coverage decreased gradually toward zero as light approached zero. Different studies have shown that seagrass light requirements depend on the environmental conditions and are higher in turbid waters (Duarte et al., 2007) and in areas with higher sediment organic matter content (Kenworthy et al., 2014) compared to seagrasses growing in clearer waters. The apparent absence of such a minimum threshold limit, indicated by our data, suggests that we have underestimated the in situ light levels at the depth limits. Alternatively, interpolating over a large pixel size (100 × 100 m) caused an overestimation of eelgrass coverage.

Both the correlation and sensitivity analysis suggested high importance of temperature for regulating the cover of eelgrass in Danish waters. Temperature affects eelgrass performance directly via effects on photosynthesis, respiration (Staehr and Borum, 2011), growth and survival (Nejrup and Pedersen, 2008), hence affecting eelgrass distribution through several processes (Bostrom et al., 2014). All enzymatic processes related to plant metabolism are temperature dependent (Drew, 1978), and specific life cycle events, such as flowering and germination, are often strongly dependent on temperature (De Cock, 1981; Phillips et al., 1983; Blok et al., 2018). In addition, biogeochemical processes are also affected by temperature, thereby influencing the interaction between plant, sediment and water column. Furthermore, temperature impacts seagrass performance by lowering water column oxygen content, increasing the oxygen diffusion coefficient, increasing respiration (Borum et al., 2006) and greatly reducing plant tolerance to anoxia (Pulido and Borum, 2010). Effects of temperature on eelgrass performance have previously been described by a bell shaped temperature dependency with optimum temperatures around 20◦C (Nejrup and Pedersen, 2008; Staehr and Borum, 2011). In this study, we also applied a bell shaped curve, but by setting a much lower optimum temperature (15◦C). This is because we used the mean summer temperatures at the local sites, calculated as a mean of ca.10–20 measurements during April to October at each eelgrass site. The fact that our data suggests maximum coverage at 5◦C below the normal temperature optimum, implies that high mean temperatures are related to much higher maximum temperatures, which can extend beyond 20◦C. The sensitivity analysis showed surprisingly high sensitivity to changes in the temperature layer. Lowering the bottom temperatures compared to current conditions were all associated with lower cover, and except for the Kattegat area, elevating the temperatures resulted in substantial declines in eelgrass. Given the ability of eelgrass to grow successfully in waters with significantly lower and higher temperatures (Bostrom et al., 2014), we find it unlikely that the occurring increases in summer water temperatures (Riemann et al., 2016) per se will cause major shifts in eelgrass coverage in Danish waters. While temperature is undoubtedly an important parameter affecting growth and distribution of eelgrass, the high sensitivity documented by our model, indicates that the temperature parameterization should be optimized by including better data on the high temperature conditions experienced by

the plants. We recommend that future spatial modeling of largescale eelgrass coverage applies data on frequency of temperatures above optimum. This should reduce the co-variation of mean summer temperatures with other important regulating factors (depth, light, salinity, exposure).

The effects of salinity on eelgrass performance have received relatively little attention despite its potential relevance particularly in estuarine environments, which typically are strongly affected by variations in freshwater inputs from precipitation, rivers and surface run-off (Conley et al., 2000). Coastal waters and estuaries, in particular, are prone to large and sometimes rapid changes in salinity. Eelgrass is a euryhaline species, and is found in both low saline systems (2–5%) such as rivers mouths, the inner estuaries and in waters of high salinity (35–40%) (Nejrup and Pedersen, 2008; Bostrom et al., 2014). It could therefore be argued that salinity is a relatively "unimportant" factor for the distribution of eelgrass, as also suggested by the sensitivity analysis in this study. However, eelgrass does not thrive equally well at all salinities and previous studies have shown that both survival and growth, as well as reproduction and seed germination are affected at extreme salinity (Phillips et al., 1983; Bostrom et al., 2014).

Low oxygen concentrations are also known to reduce growth and increase eelgrass mortality and have been associated with low coverage of eelgrass (Krause-Jensen et al., 2011; Canal-Vergés et al., 2016). Even short periods (12 h) of exposure to anoxic conditions reduce eelgrass performance whereas 24 h reduces the growth and kills eelgrass leaves (Pulido and Borum, 2010). The effect of anoxia is exacerbated when temperatures reach 25◦C and severe at 30◦C (Pulido and Borum, 2010). As temperature itself affects oxygen levels through changes in solubility of oxygen and anabolic oxygen demand, some covariation between temperature and oxygen can be expected, which we have currently not taken into account in our sensitivity analysis. In our study, we

included information on oxygen sensitivity through a data layer representing the summer mean frequency of oxygen conditions below 2 mg/L. The applied oxygen index allowed us to exclude areas with too frequent anoxic events as previously done in other studies (Canal-Vergés et al., 2016; Flindt et al., 2016). As expected, eelgrass coverage showed a decreasing trend with increasing frequency of low oxygen in all areas. In addition, the sensitivity analysis suggested a significant influence of improved oxygen conditions for areal coverage of eelgrass suggesting that our parameterization was useful.

The long history of eutrophication has led to organically enriched sediments with low critical shear stress, which is easy to resuspend, providing poor anchoring for eelgrass (Krause-Jensen et al., 2011; Canal-Vergés et al., 2016). For our large-scale model, we applied a data layer that originally included seven sediment groups. However, the resolution is rather coarse and the seven categories do not contain specific information for eelgrass suitability, such as organic matter content, and furthermore are not exclusive but include areas dominated by other substrate types. Accordingly, the group defined as mud and bedrock contains significant areas with, e.g., sand and gravel which are suitable for eelgrass plants. Considering these limitations, we reduced the original seven sediment groups into three surrogate sediment groups, which differed in the observed coverage of eelgrass (**Supplementary Figure S3**). While sediment conditions are most likely very important for determining the possible coverage of eelgrass, limitations in the current classification and the coarse resolution of the data restricted our ability to fully evaluate the importance of sediment conditions. Future analysis will undoubtedly benefit from better information on this data layer, including higher spatial resolution and information relevant for eelgrass performance (Canal-Vergés et al., 2016).

#### Management Perspectives

The GIS model has several obvious perspectives as a management tool. One aspect is, as highlighted above, to identify key distribution areas of eelgrass at a national scale, which is of obvious interest with respect to generating awareness of the vast overall distribution as well as to hotspots of eelgrass associated ecosystem functions. Awareness of the presence of the meadows and their functions is a first step toward appreciation of the meadows which inspire incentives for protection and restoration and, hence, sustainable management. In addition, the model provides a tool to identify the conditions which are currently restricting nationwide recovery of eelgrass. Recent experiences from eelgrass restoration projects show that recolonization in Danish waters will be from both sexual and vegetative reproduction. Along the shallow edge of the meadow, recolonization is primarily maintained by vegetative recruitment whereas the deep edge to larger extent relies on sexual recruitment. The intermediate depth zone may act as a buffer zone supporting the maintenance of shallower and deeper eelgrass through seed supply and vegetative expansion, thereby stabilizing the meadow by increasing its resilience toward disturbances and its recovery potential upon disturbances (Olesen et al., 2017). In relation to this, the developed habitat model and the potential eelgrass map can highlight areas where eelgrass restoration efforts are likely to be successful because

FIGURE 9 | Sensitivity of modeled eelgrass coverage to changes in input values. The input value in each pixel was varied between −20% to +20% relative to baseline conditions.

habitat conditions are documented fulfilled while monitoring data show that natural colonization has not yet happened.

Scenarios, such as effects on eelgrass distribution through changes in light availability (**Figure 9**), is a way to quantify the potential effect of actions to further improve water quality and clarity. The model can, thereby, in several ways directly help guide management interventions to protect and restore eelgrass meadows. Absence of eelgrass in validation data sets compared to the GIS model may accordingly indicate areas where there is a potential for establishment of eelgrass, given that environmental conditions remain favorable. Moreover, with some modifications, the GIS model also provides a useful tool to evaluate different climate scenarios by applying maps of high summer temperatures and low oxygen conditions.

The limitations of the current model should, however, be kept in mind. The GIS model applies data with a rather coarse resolution (100 × 100 m), implying that not all subareas are well represented by the model, simply due to absence of some of the key input data. In addition, the quality of the input data layers will largely determine the quality and predictability of the output from the GIS map. These limitations also indicate how the model can be improved in the future. Firstly, there is potential for improving the quality of the different data layers once better data are available at a national level. The physical exposure layer is an obvious candidate here. Also, as data layers representing additional potential stressors become available at a national scale, this will also help to better account for and explain local differences in regulating factors. For example, compared with local modeling in Odense fjord by Kuusemäe et al. (2016), the GIS model displays a good resemblance to scenario zero that exclude stressors such as resuspension and lugworm burial of seeds. We should also take note that the transect data, which the GIS model was fitted by, do not represent all Danish coastal waters equally. This may bias the model toward higher suitability and accuracy in areas with more observations. For example, we see that the visual fit between in situ eelgrass depth limits from 2012 and the GIS model seems better in areas with more transects and worse in areas with fewer transects. Moreover, the algorithms used to combine the GIS data layers can be improved. For example, adjustment of the applied sensitivity to water temperature could probably reduce overestimation of eelgrass presence at depth, particularly in the Kattegat region. Similarly, adjustment of the sensitivity to wave exposure could probably improve predictions in shallow waters such as Limfjorden.

#### REFERENCES


However, since our model aimed at a national scale, such local adjustment has not been undertaken in the current exercise. Adjustment of the individual weights of the applied indices in the combined index model could also be considered, although initial trials did not indicate improved model performance when the light and exposure indices were weighted higher. Finally, reiterating the fact that the outcome of the eelgrass model is strongly dependent on the quality of the GIS data layers used, the GIS model described here will be strengthened as new and better data layers become available. While some of these concern higher spatial resolution (e.g., sediment characteristics), others involve higher temporal resolution capable of discerning the duration of periods unfavorable (e.g., low oxygen and high temperatures) for eelgrass growth.

## CONCLUSION

Despite limitations and precautions, the developed GIS model provides a highly useful and long-needed estimate of the current potential distribution of eelgrass in Danish waters as well as an overview of key factors regulating the national distribution of these important meadows. The model thereby constitutes a very important tool to guide the sustainable management of eelgrass meadows at a national scale.

### AUTHOR CONTRIBUTIONS

PS and KT conceived the study. PS, CG, SØ, and SU analyzed the data. PS was responsible for writing with manuscript but got considerable help from all authors.

### FUNDING

This work was supported by the Velux Foundation through the project, Havets skove (#18530) and Storm impacts (#18538).

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00175/full#supplementary-material




**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Staehr, Göke, Holbach, Krause-Jensen, Timmermann, Upadhyay and Ørberg. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Assessment of Eutrophication Status Based on Sub-Surface Oxygen Conditions in the Gulf of Finland (Baltic Sea)

Stella-Theresa Stoicescu\*, Urmas Lips and Taavi Liblik

Department of Marine Systems, Tallinn University of Technology, Tallinn, Estonia

Sub-halocline oxygen conditions in the deep Baltic Sea basins depend on natural forcing and anthropogenic impact. HELCOM has a long tradition of characterizing the status of the seabed and deep waters by estimating the extent of anoxic and hypoxic bottoms. A eutrophication-related indicator "oxygen debt" has been used in the recent HELCOM assessments and a more sophisticated "oxygen consumption" indicator has been introduced. We describe the oxygen conditions in the Gulf of Finland (GoF) in 2016–2017 based on observations at the Keri profiling station where vertical profiles of temperature, salinity and oxygen were acquired up to 8 times a day. The main aim of the study is to test the applicability of high-frequency data from this fixed automated station and the three adapted oxygen indicators for the eutrophication-related status assessments. The results show that the GoF bottom area affected by hypoxia varied in large ranges from 900 to 7800 km<sup>2</sup> with a seasonal maximum in autumn (>25% of bottoms were hypoxic in autumn 2016). Oxygen debt is the simplest indicator, and the assessment results are less influenced by the wind-induced changes in hydrographic conditions. We suggest that oxygen debt should be assessed just below the halocline and based on data from the stratified season only since, in the GoF, the halocline could be destroyed in winter. For the "oxygen consumption" indicator, a rough oxygen budget, where the contributions of advection and mixing are included, was formulated. Average seasonal consumption values of 0.82 and 0.31 mg·l −1 ·month−<sup>1</sup> were estimated in the 50–60 m water layer of the GoF in 2016 and 2017, respectively. The found large difference in consumption values between 2016 and 2017 could partly be related to the uncertainties of advection estimates. We concluded that all three indicators have their advantages and methodological challenges. To increase the confidence of eutrophication assessments both high-frequency profiling should be implemented in the monitoring programs and more accurate estimates of changes due to physical processes are required.

Keywords: eutrophication, assessment, hypoxia, Baltic Sea, Gulf of Finland, bottom waters, oxygen

## INTRODUCTION

The Baltic Sea is an area where oxygen conditions are influenced by climate change (Kabel et al., 2012) and increased eutrophication (Conley et al., 2009; Gustafsson et al., 2012). Eutrophication is driven by excessive inputs of nutrients from rivers and atmosphere (mostly land-based sources) which lead to increased sedimentation of organic material and oxygen depletion in the bottom

#### Edited by:

Jacob Carstensen, Aarhus University, Denmark

#### Reviewed by:

Qian Zhang, University of Maryland Center for Environmental Science (UMCES), United States Kari Juhani Eilola, Swedish Meteorological and Hydrological Institute, Sweden

#### \*Correspondence:

Stella-Theresa Stoicescu Stella.Stoicescu@taltech.ee

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 01 October 2018 Accepted: 31 January 2019 Published: 19 February 2019

#### Citation:

Stoicescu S-T, Lips U and Liblik T (2019) Assessment of Eutrophication Status Based on Sub-Surface Oxygen Conditions in the Gulf of Finland (Baltic Sea). Front. Mar. Sci. 6:54. doi: 10.3389/fmars.2019.00054

layer and the internal loading of phosphorus (Vahtera et al., 2007). The oxygen conditions in the near-bottom layer of the central deep basins of the Baltic Sea are occasionally improved by the Major Baltic Inflows (MBI) (Matthäus and Franck, 1992; Schinke and Matthäus, 1998). The MBIs also strengthen stratification and therefore potentially increase areas with oxygen depletion by inhibiting ventilation of deep layers (Gerlach, 1994; Conley et al., 2002).

The Gulf of Finland (GoF) is an elongated estuarine basin with the largest single freshwater input and the highest nutrient loading to the Baltic Sea from the Neva River (Alenius et al., 1998). Beside the eutrophication effects, the sub-surface distribution and variability of dissolved oxygen (DO) in the GoF are related to multi-scale physical processes. These processes range from short-term oscillations and mixing events to winddriven alterations of estuarine circulation, seasonal development and decay of stratification and sub-halocline transport of hypoxic/anoxic waters from the Northern Baltic Proper, for instance, associated with the MBIs (Elken et al., 2003; Liblik et al., 2013, 2018; Lips et al., 2017). The lack of MBIs during stagnation periods affects the GoF in a way that stratification and hypoxia are decreased (Conley et al., 2009; Laine et al., 2007). Oxygen in the near-bottom layer is dependent on wind conditions as north-easterly and northerly winds support the estuarine circulation and strong south-westerly wind forcing causes the reversal of estuarine circulation (Lehtoranta et al., 2017). The estuarine circulation is characterized by the upestuary (eastward) flow in the deep layer and the down-estuary (westward) in the upper layer. The reversed circulation bears opposite results – the up-estuary flow in the surface layer and the down-estuary (westward) flow in the deeper layer (Elken et al., 2003). Prevailing of the estuarine circulation leads to strengthening of the halocline while reversals cause weakening of stratification. The wintertime deep-water oxygen conditions are strongly dependent on the spatiotemporal variability of the salt wedge originating from the Northern Baltic Proper (Liblik et al., 2013), which moves eastward and westward at the gulf's bottom depending on wind conditions.

According to the Marine Strategy Framework Directive's (MSFD; Directive 2008/56/EC of the European Parliament and of the Council, 2008) Article 11, all member states have to establish and implement monitoring programs to assess the status of their marine environment. The monitoring programs have to provide data that enable the application of different indicators in order to assess the status of marine waters, including in regard to eutrophication effects. HELCOM (Baltic Marine Environment Protection Commission – Helsinki Commission) has set out a number of core indicators (according to MSFD) for describing the eutrophication status where nutrient levels are based on dissolved inorganic nitrogen and phosphorus as well as total nitrogen and phosphorus. Direct effects of eutrophication are assessed based on chlorophyll a levels and water transparency (Secchi depth). Oxygen conditions are used to describe the indirect effects. The overall assessment is given using a tool called HEAT (HELCOM Eutrophication Assessment Tool) which determines the distance between the measured value of an indicator and good environmental status (GES, a predefined threshold value), aggregates the indicator evaluations and gives the confidence of the assessment (HELCOM, 2014).

While indicators based on nutrient levels and direct effects for assessing eutrophication status are applicable in all Baltic Sea sub-basins, the use of oxygen indicators, describing the indirect effects, is restricted to the deep basins, including the GoF. The approved core indicator evaluates the oxygen debt (the "missing" oxygen relative to a fully saturated water column) below the halocline found based on salinity profiles and discrete oxygen concentrations measured at standard depths (HELCOM, 2013). The threshold values are uniform for most of the basins, including the GoF (8.66 mg/l). The latest results for oxygen debt assessments in the GoF are 10.54 mg/l for 2007–2011 and 10.67 mg/l for 2011–2016 (HELCOM, 2014, 2018a), both indicating that GES has not been achieved. Data used for oxygen debt indicator assessments originate from temporally sparse monitoring (in Estonia, national monitoring is carried out six times a year) which could miss the variability of deep-water oxygen conditions, and therefore, the assessment results could be biased. To get a more confident assessment of oxygen conditions in the deep water the data acquired with better temporal coverage could be used.

The development of a more uniform indicator was initiated in the frames of the HELCOM EUTRO-OPER project in 2015. The idea was to estimate the oxygen consumption during summer in the layer below the productive surface layer, it means in the so-called stagnant layer located between the thermocline and the halocline. Oxygen consumption is calculated based on oxygen depletion, diffusion and advection. Due to small temporal differences in salinity and temperature within the stagnant layer, advection was neglected in the first tests of this indicator (HELCOM, 2015b).

Oxygen conditions can also be described by the spatial extent of hypoxia which could be one of the possible indicators of indirect effects of eutrophication (Conley et al., 2009). Considering the hypoxic area as a eutrophication status indicator, the natural variability of oxygen, e.g., the changes due to hydrographical influences and the anthropogenic component have to be separated in order to seclude human-induced effects. The yearly extent of the hypoxic and anoxic bottom area in the Baltic Sea is assessed and presented on HELCOM Environment fact sheets (Naumann et al., 2018; Viktorsson, 2018). The results differ a bit between these latest two assessments due to the temporal and spatial coverage of data used. The first estimate, produced by the Swedish Meteorological and Hydrological Institute (SMHI), uses data from the autumn period, August to October for the whole Baltic Sea, and the second, produced by the Baltic Sea Research Institute, Warnemünde (IOW), uses data from May and excludes the Bothnian Sea, Danish straits, Gulf of Riga and the GoF.

In this study, we defined hypoxia by a threshold of 2.9 mg·l −1 , which is one of the two most commonly used thresholds (the other being 2.0 mg·l −1 ) in literature, based on an extensive review by Vaquer-Sunyer and Duarte (2008). They concluded that a single threshold could not adequately describe the influence of hypoxia to different benthic marine organisms and argue that the conventionally accepted threshold of 2.0 mg/l is well below

the oxygen thresholds for more sensitive taxa. It is important to define a proper hypoxia level for the Baltic Sea and GoF, but we do not discuss this issue here since the hypoxia area estimate methodology will still be the same and the hypoxia threshold could be addressed together with setting the border between the GES and sub-GES.

The hypoxic area assessments by SMHI show that the areal extent and the volume of anoxia and hypoxia have continuously been elevated since the regime shift in 1999 (Hansson et al., 2011). The latest results (for 2017, mean areal results) show that anoxic conditions affect around 18% and hypoxia around 28% of the bottom areas in the Baltic Sea, including the GoF and the Gulf of Riga (Hansson et al., 2017). In the IOW hypoxic area estimations, the GoF is unfortunately excluded, but it is seen that the anoxic conditions in the Northern Baltic Proper in 2016 have changed into hypoxic conditions in 2017 (Naumann et al., 2018).

All above-described assessments are based on data from spatially distributed monitoring stations with low temporal resolution. An alternative approach could be to apply data from automated profiling stations which also catch temporal variability due to prevailing hydrographic processes. In the present study, we analyze the high-resolution time series of vertical profiles of oxygen in the GoF in 2016–2017. The main aim is to demonstrate how these observations could be applied to assess the status of this stratified estuary in relation to the eutrophication effects using the adapted versions of the three possible indicators.

## MATERIALS AND METHODS

#### Core Data Set

The core data set used in the present study originates from the autonomous bottom-mounted profiler, which is deployed at the depth of 110 m near the Keri Island in the GoF since March 2016 (**Figure 1**). The data acquired in 2016 and 2017 were used. The system was developed by Flydog Solutions Ltd., (Estonia) and includes, as the main measurement device, an OS316plus CTD probe (Idronaut s.r.l., Italy) with Idronaut oxygen sensor and Trilux fluorescence sensor (Chelsea Technologies Group Ltd.). The profiler records temperature, salinity, DO content, chlorophyll-a, phycocyanin and turbidity at a rate of 8–9 Hz while moving up with an average speed of 8–10 cm·s −1 . In May– June 2017, a Seabird Electronics SBE19plus probe with SBE43 oxygen sensor provided by the Finnish Meteorological Institute was used instead of the OS316plus probe. The accuracy of the used Idronaut oxygen sensors is 0.1 mg·l −1 .

The data were pre-processed to exclude spikes and recordings when the probe movement was reversed (e.g., due to waves moving the buoyant probe up and down in the near-surface layer) and to compensate the time lag of sensors. While the Idronaut oxygen sensor has a time constant of about 8 s, the time constant of SBE43 sensor was set to 4 s. After the pre-processing of data, the vertical profiles were stored with a constant step of 0.5 m. The data were collected in the water column between 100 and 4 m.

All used CTD probes and oxygen sensors were calibrated at the factory before the deployment. In addition, the core data set was quality controlled against the quality-assured data from the research vessel based measurements conducted regularly (once a month) close to the profiling station using a CTD probe OS320plus (Idronaut s.r.l., Italy) and laboratory analyses of water samples. See the applied methods and onboard quality assurance procedures in the next sub-section. Since the oxygen sensor had a drift in time, the oxygen profiles were corrected using ship-borne measurement results. For each vessel visit, a linear regression line equation with the intercept set to zero was found

by comparison of the ship-borne oxygen profile and the closest oxygen profile from the bottom-mounted profiler. The oxygen profiles between the two consecutive vessel visits were corrected using the coefficient (the slope of the regression line) assuming its linear trend in time. If such ship-borne measurements were not available from a specific period, all data from that period were excluded from the further analysis.

The temporal continuity of the analyzed data can be described by the months covered and the number of quality controlled profiles available (**Table 1**). In both years, the majority of the productive season is covered. The exclusion of data due to quality problems and the stops in the operation of the bottom-mounted profiler (for instance, in January–April 2017) due to different reasons are seen in **Table 1** as well as the next section as blank areas in the graphs. The higher number of profiles per month in summer 2017 was due to the setup of profiling frequency of 8 profiles per day while it was 4 profiles per day in 2016.

#### Ship-Borne Data Collected in 2014–2017

Data from the research vessel cruises were used for the quality assurance of the measurements by the bottom-mounted profiler and determining the spatial variability in the vertical distribution of DO in the gulf. Also, the relationship between the oxygen and salinity in the sub-surface layers in the vicinity of the profiler was found using the ship-borne observations. The analyzed CTD data were gathered in the frames of national monitoring and GoF Year 2014 programs in 2014–2017 (**Figure 1**) as well as additional monthly visits to the Keri station in 2016–2017. The oxygen sensor attached to the OS320plus was calibrated before each cruise. Oxygen profiles used for the analysis were quality checked against the laboratory analysis of water samples using an OX 400 l DO (WWR International, LCC). The accuracy of the Idronaut oxygen sensor on board the research vessel is 0.1 mg·l −1 while the accuracy of the laboratory DO analyzer is 0.5% of the

TABLE 1 | The number of available quality checked profiles from the Keri autonomous profiler.


The values in bold represent the total number of profiles available in both years.

measured value. Altogether 336 ship-borne vertical profiles of DO were available for the present study.

Also the temperature and salinity data of both, the research vessel CTD and the profiler CTD, were quality checked against water samples analyses using a high-precision salinometer 8410A Portasal (Guildline). The average difference of the salinity values was less than 0.02 g·kg−<sup>1</sup> that indicated no need for the correction of the CTD data.

#### Indicators

#### Extent of Hypoxic Area

Bathymetric data from Andrejev et al. (2010, 2011) with the horizontal resolution of 463 m were used to calculate the hypsographic curve of the GoF, and based on this dataset, the GoF area was set equal to ∼27631 km<sup>2</sup> and volume to ∼1042 km<sup>3</sup> . We selected the threshold for hypoxia of 2.9 mg·l −1 (equal to 2.0 ml l−<sup>1</sup> and 89.3 µmol l−<sup>1</sup> ) as also applied in HELCOM Baltic Environmental Fact Sheets (Naumann et al., 2018; Viktorsson, 2018) and many research papers (e.g., Diaz and Rosenberg, 2008; Conley et al., 2009). The upper border of the hypoxic layer was defined for each measured vertical profile as the minimum depth where DO content was below the defined threshold (see examples in **Figure 2**). In some cases (some profiles from March 2016) hypoxic depth was not possible to define from the profiles, and the hypoxic depth was assumed to be deeper than the deepest measured value. Here the depth of hypoxia was found based on the assumption that DO content decreases linearly when moving deeper. The change in DO content with depth was found comparing DO measured at the last depth and DO measured five meters above the last depth.

$$depth\_{\text{hypozia}} = \, depth\_{\text{max}} + \mathfrak{x},\tag{1}$$

where depthmax is the deepest measured depth value and <sup>x</sup> is the change of depth from last measured depth value to the depth value where DO = 2.9 mg·l −1 ; <sup>x</sup> was defined as:

$$\chi = \frac{5\*(2.9 - DO\_{\text{depth}(\text{max})})}{DO\_{\text{depth}(\text{max})} - DO\_{\text{depth}(\text{max})} - 5},\tag{2}$$

where 5 is depth interval (in meters) between the two DO measurements used to find the linear change in DO; 2.9 is the concentration of DO (in mg·l −1 ) which marks hypoxia; DOdepth(max) is the deepest measured DO value and DOdepth(max)−<sup>5</sup> is the DO value measured 5 m above the deepest value. If <sup>x</sup> < 0 or DOhypoxia > 115 m, then the value DOhypoxia = 115 m was used, which is the maximum depth in the Keri area and it is deeper than the maximum depth in the used bathymetric data.

To determine the average spatial inclination of the upper border of the hypoxic layer along the GoF, data from two cruises in May and August 2014 with spatially well distributed station network (**Figure 1**) were analyzed. First, the start of hypoxia was found (same method as for the indicator) for every measured profile. Then these depths were plotted against station longitude, and the linear regression line, indicating the change of hypoxic depth in the east-west direction, was found.

Based on the found hypoxic depth values and the hypsographic curve, corresponding area and volume of hypoxia in the GoF were calculated by applying a developed script (we used software package MATLAB R2016a). Two estimates of both parameters were found (1) assuming that the border of the hypoxic layer was a horizontal plane and (2) assuming that it had a constant inclination in the east-west direction. The monthly values of the estimates are presented since we suggest that it is a long enough period to filter out the impact of spatial variability at mesoscale (remember that we use observations for a single station).

To compare hypoxic area extent results with prevailing wind conditions we used wind data from the only real open sea automatic weather station (other stations are mainly located near the coast) at Kalbådagrund (59◦ 580N, 25◦ 370E) (Alenius et al., 1998). Wind data was corrected to represent wind speed at 10 m (measured wind speed was multiplied by 0.91) (Launiainen and Laurila, 1984). First, the monthly mean hypoxic area extent was related to the prevailing wind direction and speed calculated for the preceding 30 days. For example, the hypoxic area in June was compared with wind data from the 16th of May to the 15th of June. The number of days used for wind analysis was selected to be 30 because that is the time frame used for the calculation of average hypoxic extent. Secondly, the linear correlation between the 3-week average wind stress component from N-NE (20π) and the hypoxic depth was found. This choice of the wind direction and period was based on a study by Liblik and Lips (2011) where they showed that the 3-week average wind component correlated best with the changes in the vertical thermohaline structure in the GoF in summer. Drag coefficient used when finding wind stress was according to Large and Pond (1981).

#### Oxygen Debt Indicator

In the latest HELCOM eutrophication assessments (HELCOM, 2014, 2018b), an oxygen debt indicator proposed during the HELCOM TARGREV project was used. Because the monitoring data are only available from the standard depths, a special procedure was applied to estimate the sub-halocline DO content. First, the salinity profiles were modeled to identify the halocline. Then, the linear segments of the oxygen profile in the halocline and below it were constructed. Oxygen debt was calculated by subtracting the monitored DO content from the concentration of saturation, taking into account the temperature and salinity values. Finally, the volume specific oxygen debt was found for the sub-halocline layer (see more about the method in HELCOM, 2013).

In the present work, the oxygen debt value just below the halocline and not the volume specific average was used as the oxygen debt indicator. The halocline was determined as the

depth range where the vertical salinity gradient was greater than 0.07 g·kg−<sup>1</sup> ·m−<sup>1</sup> (Liblik and Lips, 2011), whereas the salinity gradient was found based on smoothed profiles (over 2.5 m). Temperature, salinity and DO content values just below the halocline were found for each measured profile, and the corresponding oxygen debt values were calculated. In order to assess the eutrophication status, monthly and seasonal averages of oxygen debt were estimated.

For the HELCOM oxygen debt indicator, a constant target value in the GoF, the Northern Baltic Proper and the Eastern Gotland Basin is defined at 8.66 mg·l −1 as an annual average (HELCOM, 2013). We suggest that the uniform target value should not be applied since the basins have different depths and a linear oxygen change below the halocline was assumed. However, if the oxygen debt value just below the halocline is used, the targets could be similar, though still to be defined, for all mentioned basins. The other reason not to use the volume specific average oxygen debt is that we lack hydrogen sulfide observations at the profiling station and the oxygen debt estimate near the seabed could be biased.

#### Oxygen Consumption Indicator

An alternative oxygen indicator is being developed by the experts working with eutrophication-related issues in the HELCOM community, although it is not fully ready nor applied yet (HELCOM, 2015b). The idea is to base the indicator on estimated oxygen consumption in the summer season (June to September) below the productive layer but above the halocline – in a so-called stagnant layer. In the HELCOM report, the sparse monitoring data were used and the advective processes were not taken into account when estimating oxygen consumption. We test this indicator based on vertical profiles of DO content acquired with a high temporal resolution at a fixed position.

According to HELCOM (2015a), oxygen consumption (CONS) in a water layer between its upper (u) and deeper (d) border is calculated as:

$$\text{CONS}\_{\text{(u,d)}} = DEPL\_{\text{(u,d)}} + DFF\_{\text{(u,d)}} + ADV\_{\text{(u,d)}} \quad \text{(1)}$$

where DEPL is oxygen depletion (decrease in oxygen has a positive value) and DIFF and ADV are the change in oxygen content due to vertical diffusion and advection (both horizontal and vertical), respectively. The layer with the borders u and d is selected for a studied year based on vertical thermohaline structure (see section "Oxygen Consumption").

Changes in DO content due to diffusion are estimated as:

$$\begin{aligned} \text{DIFF}\_{(u,d)} &= \\ &-A(u) \left( \kappa(u) \frac{\hat{\varepsilon} O\_2(u)}{\hat{\varepsilon} z} - \frac{A(d)}{A(u)} \kappa(d) \frac{\hat{\varepsilon} O\_2(d)}{\hat{\varepsilon} z} \right) \end{aligned} \tag{2}$$

where A(u) and A(d) are the horizontal cross-sectional area of the studied layer, <sup>∂</sup>O2 (u) ∂z and <sup>∂</sup>O2 (d) ∂z the vertical gradient of oxygen and κ(u) and κ(d) the vertical diffusivity coefficient at its upper and deeper border, respectively. The latter is calculated as:

$$\kappa\_{\text{(u,d)}} = \frac{\alpha\_{\text{(u,d)}}}{N\_{\text{(u,d)}}} \tag{3}$$

where α is an empirical intensity factor of turbulence. N is the Brunt–Väisälä frequency, defined as:

$$\mathbf{N}^2\_{\mathrm{(u,d)}} = -\frac{\mathbf{g}}{\rho\_0} \frac{\partial \rho(\mathbf{u}, \mathbf{d})}{\partial \mathbf{z}} \tag{4}$$

where g is the acceleration due to gravity, ρ<sup>0</sup> is density of the seawater and <sup>∂</sup>ρ(u,d) ∂z is the vertical gradient of density. In the present study, we assumed that the areas A(u) and A(d) are equal (it is correct for deep enough regions) as well as the empirical intensity factor of turbulence is a constant (α = 1.5·10−<sup>7</sup> m<sup>2</sup> ·s −2 ).

The increase in oxygen content due to advection was calculated based on the estimated salinity advection and an assumption that a linear correlation exists between the changes in salinity and oxygen. Since neither salinity sources nor sinks exist in the sub-surface layer, salinity advection could be found as:

$$ADV\_{\text{(u,d)SA}} = \text{CHANG}\_{\text{SA}} - \text{DIFF}\_{\text{(u,d)SA}} \tag{5}$$

where CHANGE is the salinity change between the two profiles in the selected layer and DIFF is the estimated change in salinity due to vertical diffusion (a similar formula was applied as for oxygen diffusion). Oxygen advection was thus calculated as:

$$ADV\_{\rm (u,d)O\_2} = ADV\_{\rm (u,d)SA} \ast a \tag{6}$$

where a is the oxygen change corresponding to a unit change in salinity. This coefficient was found as the slope value of the linear regression line based on salinity and oxygen data from the monitoring cruise in April of the corresponding year, thus, before the analyzed period. For the regression analysis, a layer from 30 to 70 m was selected which is broader than the stagnant layer where oxygen consumption was estimated in the present study.

Oxygen depletion was calculated based on the monthly mean average oxygen concentrations in the selected layer. For example, to find oxygen depletion between June and May, the average concentration in June was subtracted from the average concentration in May. The total monthly change in DO content due to diffusion was found as the daily average diffusion during 30–31 days (e.g., from mid-May until mid-June) multiplied by the number of days. The monthly changes in both, salinity and oxygen, due to advection were also estimated over the similar monthly time step. Finally, monthly oxygen consumption values were obtained as expressed in Eq. (1). A month was chosen as a minimum time step since the consumption estimates for a shorter period are close to the accuracy of the DO measurements.

#### RESULTS

#### High-Resolution View on Temporal Variability of Dissolved Oxygen Content

The time series of vertical distributions of temperature, salinity and DO concentration from spring until autumn in 2016 and 2017 (**Figure 3**) show both the seasonal course and the short-term variations. The surface layer warming and development of the seasonal thermocline with its sharp downand upward movements between 10 and 40 m depth were the

characteristic features of the temperature distribution time-series. The halocline fluctuated between the depths of 60 to 80 m. If to consider a general temporal development of salinity distribution in the deep layer from May to October, then the halocline penetrated deeper from June to late August and got shallower in late September and October in both studied years.

The seasonal course of the vertical distribution of oxygen followed the mentioned development of the thermohaline structure. Oxygen concentrations decreased in the upper layer from spring to late summer mostly due to the decay of the vernal phytoplankton bloom and the temperature increase in the summer months, as it leads to the decrease in saturation concentration. The boundary of the near-bottom hypoxic layer, defined here as the oxygen concentration of 2.9 mg·l −1 (the yellow line in **Figure 3** lower panel), moved up- and downward together with the halocline. It is also seen that the oxygen concentrations decreased in the layer between the thermocline and halocline. At least partly this decrease in oxygen content could be related to the oxygen consumption that will be analyzed in more detail in the present study.

#### Hypoxic Area

The depth, at which hypoxia starts in the water column, varied throughout the year and between the 2 years (**Figure 4**) with an average from May to October of 64.5 m in 2016 and 64.6 m in 2017. The upward movement of the border of the hypoxic layer from winter to spring occurred in 2016. In June–September of both years, the hypoxia border moved

TABLE 2 | Monthly averages of the bottom area and volume (percent of total area/volume) for the leveled and the inclined border of the hypoxic layer in 2016 and 2017; "incl" stands for the inclined layer, "A" for area and "V" for volume.


Also, the average wind speed (m·s −1 ) and direction at Kalbådagrund during the preceding 30 days for each month is given; e.g., for June from the 16th of May to the 15th of June.

downward in the water column from about 60 to 70 m, and after that, it rose again to 60 m in both years and even shallower in late October 2016. At the same time, as the hypoxia border deepened, a significant salinity decrease at the minimum depth of the hypoxia was observed (p < 0.05) in both years from May to October (**Figure 4**). This result suggests either the local consumption of oxygen, since it corresponds to a decrease in oxygen content at a fixed salinity value, or a change in water properties due to physical processes, such as advection of water masses with a different oxygen-salinity relationship.

Since the depth of the upper border of the hypoxic layer (hypoxic depth) revealed high short-term variability (**Figure 4**), the estimates of the hypoxic area and volume of hypoxic waters were analyzed based on the monthly averages. Both approaches, assuming the leveled and the inclined border of the hypoxic layer were applied (**Table 2**). The linear inclination of the border of the hypoxic layer of 1.9 m per 100 km along the GoF from west to east, found based on the GoF Year 2014 data [see stations network in **Figure 1**; correlation between the start of hypoxia and longitude was significant (R <sup>2</sup> = 0.16, p < 0.05)], was used. As seen in **Table 2**, the estimates of the area and volume of the hypoxic waters differed only by <2 and <0.5%, respectively, if the estimates based on the leveled and the inclined border of the hypoxic layer were compared. This good coincidence of the estimates, although an average inclination of the border of hypoxia exists, is most probably explained by a central location of the Keri station in the GoF.

We compared the monthly average extent of the hypoxic area with the average wind vector from the preceding 30 days (**Table 2**). It is seen that for 2016, when the N-NE winds prevailed, then the hypoxic area was enlarged, while with the W-SW winds, the area was somewhat smaller. However, in 2017, the

wind forcing could not explain the observed changes so well. The hypoxic area was biggest in May–June and in September but the corresponding winds for these months from the sector between east and north were almost absent. We also compared the 3-week average wind component from N-NE (20π) with the hypoxic depth at the Keri station (**Figure 5**). If the early spring data were excluded, then the linear correlation was significant (p < 0.05) for both years with a much stronger relationship for 2016 May–October than for 2017 May–October (**Figure 5**).

Almost identical seasonal course in the development of the hypoxic area was observed in both summers when the area affected by hypoxia decreased from the early summer until August and started to grow again in September–October (**Figure 6**). When comparing the hypoxic extent monthly results obtained using different hypoxia thresholds (DO < = 2.9 mg·l −1 and DO < = 2.0 mg·l −1 ), the dynamics are the same and the averages differ only slightly. In 2016, the May to October mean hypoxic area was 2.2% bigger with the hypoxic threshold

R <sup>2</sup> = 0.60, p < 0.05] and 2017 [(B); R <sup>2</sup> = 0.08, p < 0.05]. Wind data from Kalbådagrund are used (see Figure 1 for location). The color bar shows the time of measurements (in months).

**317**

DO < = 2.9 mg·l −1 compared to the area found using the threshold DO < = 2.0 mg·l −1 , while in 2017, the area was 1.6% bigger.

#### Oxygen Debt

Oxygen debt values just below the halocline were calculated for every vertical profile of DO (**Figure 7**). Halocline, defined using the vertical salinity gradient criterion of 0.07 g·kg−<sup>1</sup> ·m−<sup>1</sup> , was detected for all profiles in both years. The mean depth of the point below the halocline, for which the oxygen debt was estimated, was 74.2 m in 2016 and 74.5 m in 2017. If for 2016 the period from May to October was considered (when the data were available in 2017), then the mean depth of this point was 72.4 m. The temporal variability (May to October) could be characterized as the maximum and minimum values of 96.0 and 54.5 m in 2016 and 97.0 and 60.5 m in 2017, respectively, and the 5th and 95th percentiles of 62.0 and 81.5 m for 2016 and 67.0 and 83.6 m for 2017.

For the period from May to October (when data were available for both years), the mean oxygen debt was 11.3 mg·l −1 in 2016 and 11.6 mg·l −1 in 2017. The minimum, maximum, 5th and 95th percentile values of oxygen debt were 8.5 mg·l −1 , 11.9 mg·l −1 , 10.2 mg·l −1 , and 11.8 mg·l −1 in 2016 and 9.9 mg·l −1 , 12 mg·l −1 , 11 mg·l −1 , and 12 mg·l −1 in 2017. If also early spring period for 2016 was taken into account, then the minimum oxygen debt was estimated as 7.1 mg·l −1 .

In both years, we found no linear trend for the halocline base depth from May to October, and the oxygen debt showed no clear seasonal dynamics. However, there was a significant correlation between the halocline base depth and oxygen debt values in 2017 (R <sup>2</sup> = 0.23, p < 0.05).

If monthly mean values are considered, the halocline base depth and oxygen debt showed similar dynamics in both years. Considering the period from May to October, the mean halocline base depth ranged from 66.0 to 75.1 m in 2016 and from 72.2 to 77.2 m in 2017. A smaller variation in monthly mean oxygen debt values in 2017 is also evident, with the values ranging from 11.5 to 11.8 mg·l −1 , compared to 2016 where the monthly oxygen debt values were between 10.9 and 11.5 mg·l −1 (**Figure 8**).

#### Oxygen Consumption

Oxygen consumption was calculated for the period from May to September for the stagnant layer which was defined based on the monthly average profiles of absolute salinity (SA) and conservative temperature (CT) (**Figure 9**). In both years, it was defined as the layer between 50 and 60 m, and the extended layer for the analysis (needed for the calculation of diffusion) was defined by adding data from additional 3 m to both sides of the stagnant layer; so, a depth range of 47–63 m was used. To estimate the sensitivity of the results to the choice of the stagnant layer, the estimates for diffusion, advection and consumption were also found for the layer 45–55 m.

Changes in the oxygen concentration due to advection were calculated based on estimated salinity advection and the correlation between salinity and oxygen. In both years, vertical profiles in the depth range of 30–70 m acquired in April at the national monitoring stations distributed along the GoF (see **Figure 1**) were used for the correlation analysis. The found relationships were statistically significant – in 2016, DO = −3.39<sup>∗</sup> SA + 34.54, R <sup>2</sup> = 0.84, p < 0.05 and, in 2017, DO = −3.57<sup>∗</sup> SA + 36.43, R <sup>2</sup> = 0.73, p < 0.05.

Positive monthly average oxygen depletion values show that during the period in question, oxygen content in the stagnant layer decreased (**Figure 10**). Negative diffusion values indicate that more oxygen left the layer through the deeper boundary than was brought in through the upper boundary of the layer. Positive advection values show that more oxygenated water was brought into the study area (area of the profiling station).

The results show that the changes in oxygen on the monthly scale were mostly related to advection since the depletion and advection had the largest values and were moving in opposite directions (**Figure 10**; remember that a positive value of depletion means that oxygen is decreasing). Monthly depletion values were in the range of ±2.0 mg·l −1 ·month−<sup>1</sup> while advection values ranged from −1.9 mg·l −1 ·month−<sup>1</sup> to 1.7 mg·l −1 ·month−<sup>1</sup> . On

FIGURE 8 | Monthly average oxygen debt just below the halocline (± standard deviation); (A) 2016, (B) 2017. The halocline base depth at Keri station in 2016 (C) and 2017 (D).

a monthly scale, the diffusion had the lowest contribution with the values varying in the range from 0 to −0.4 mg·l −1 ·month−<sup>1</sup> . The resulting monthly average consumption estimates were in the range of −0.5 and 1.6 mg·l −1 ·month−<sup>1</sup> with the lowest values in June–July and the maximum in July–August. This pattern was evident in both years. Note that the consumption values should be positive since no oxygen production should exist in the analyzed stagnant layer. However, the applied method gave small negative values of consumption for June–July in both years.

The total seasonal consumption from May to September as well as monthly average consumption rates were almost three times larger in 2016 than in 2017 – 0.82 and 0.31 mg·l −1 ·month−<sup>1</sup> , respectively (**Table 3**). It is interesting that the consumption rate estimates did not change much if, instead of the layer 50–60 m, the layer 45–55 m was analyzed although depletion was higher in the layer 45–55 m in both years. According to our estimates, consumption gave the highest contribution to the seasonal oxygen depletion. For the layer 50–60 m in 2017 and 45–55 m in 2016, diffusion and advection almost entirely compensated each other. Diffusion values did not vary between the 2 years much while advection was considerably higher in 2016 than in 2017, especially for the layer 45–55 m. This result is in agreement with the observed concurrent decrease of salinity in these depth ranges (see

TABLE 3 | Average values of depletion (DEPL), diffusion (DIFF), advection (ADV), and consumption (CONS).


**Figure 3**; the decrease was larger in 2016 than in 2017) since the oxygen advection was estimated based on salinity advection and linear correlation between oxygen and salinity. Still, the much larger consumption in 2016 than in 2017 has to be analyzed further.

We also analyzed the sensitivity of the consumption estimates to the following assumptions: neglecting the changes in oxygen content due to changes in solubility; constant empirical intensity factor for turbulent flux estimates; linear regression between oxygen and salinity used to estimate oxygen advection.

We calculated monthly average temperature and salinity in the analyzed layer 50–60 m and found that due to their changes ranging, respectively, from 3.27 to 4.66◦C and from 7.25 to 8.01 g·kg−<sup>1</sup> , the oxygen saturation concentration varied between 11.61 and 12.06 mg·l −1 . Since the monthly average oxygen concentration in this layer ranged between 5.70 and 8.74 mg·l −1 , the changes in solubility could cause less than about 15% of the observed variability.

For diffusion estimates, we used the value of the empirical intensity factor of α = 1.5 × 10−<sup>7</sup> m<sup>2</sup> ·s −2 . It is an average value based on earlier studies, where the values ranging from α = 0.5 × 10−<sup>7</sup> m<sup>2</sup> ·s −2 to α = 2.5 × 10−<sup>7</sup> m<sup>2</sup> ·s <sup>−</sup><sup>2</sup> have been applied (Gargett, 1984; Stigebrandt, 1987; Axell, 1998; Meier, 2001). Using these minimum and maximum values of the factor we found that for the 2016 data, the consumption estimates varied only from 0.80 to 0.85 mg·l −1 ·month−<sup>1</sup> while for the 2017 data, the ranges were larger – from 0.25 to 0.36 mg·l −1 ·month−<sup>1</sup> . Although for the latter case, the difference between the minimum and the maximum was about 30% of the estimate, the choice of the factor changed the result only by 0.1 mg·l −1 ·month−<sup>1</sup> .

For the advection estimates we used the linear correlation between salinity and oxygen based on data from the monitoring stations in the GoF from the Osmussaar Island to the eastern GoF (station F1; see **Figure 1**) This assumption of the linear relationship between salinity and oxygen content is valid only in a limited area (see the correlation estimates above) but not for the entire Baltic Sea; thus, also in the GoF, along-isohaline gradients exist. For instance, based on the data from the monitoring cruises in May/June, July, August and October of both years, the average change in oxygen concentration in the GoF along an isohaline corresponding to the average salinity in the layer 50–60 m was 0.9 mg·l <sup>−</sup><sup>1</sup> per a longitudinal degree. In the case of oscillating up- and down-estuary flow in the studied layer, the oxygen advection estimates are not biased, but a relatively large bias could occur if unidirectional flow prevails for a long period. Taking the estimated horizontal oxygen gradient and assuming a constant flow of 2 cm·s −1 for a month could result in a change of oxygen content by 0.83 mg·l −1 ·month−<sup>1</sup> .

### DISCUSSION

We analyzed high-resolution time series of vertical profiles of oxygen at a central, deep station in the GoF close to the Keri Island in 2016–2017. The main study question was whether these observations could be applied to assess the status of this stratified estuary in relation to the eutrophication effects using

three suggested indicators. The present monitoring programs contain observations of DO content at a few stations a few times a year (in Estonia, 6 times a year, and only the surface and nearbottom layer are sampled). Beside the eutrophication effects, the sub-surface distribution and variability of DO in the GoF are related to multi-scale physical processes. Thus, the differentiation between the natural variability and the effects of eutrophication could require high-resolution monitoring data. Since due to economic reasons it is not possible to arrange measurements with high temporal resolution at many stations, we analyzed the applicability of data from one profiling station to assess the oxygen conditions using the following indicators – the extent of hypoxia, oxygen debt, and oxygen consumption.

Although large-scale surveys of the vertical oxygen distribution revealed an along-gulf inclination of the border of the hypoxic layer of 1.9 m per 100 km, the estimates of the extent of hypoxia based on the hypoxic depth from Keri station with and without inclination were very close to each other. In addition, it was shown by Liblik and Lips (2017) that there exists almost no average cross-gulf inclination of the halocline in the GoF. Thus, we conclude that the Keri station is a representative station to assess the sub-halocline oxygen conditions in the entire GoF on seasonal timescales.

Since wind conditions play a major role defining the areal extent of the near-bottom hypoxia in the GoF, the effect of the prevailing wind has to be taken into account when using the hypoxic area extent as a eutrophication indicator. It is known that with prevailing south-westerly winds the estuarine circulation is reversed which moves the surface layer into the gulf and the deep layer out of the gulf (Elken et al., 2003). North-easterly winds have the opposite effect and the deep water – deoxygenated and phosphate-rich salt wedge moves into the gulf from the Northern Baltic Proper (Liblik and Lips, 2011; Lips et al., 2017).

When comparing the monthly average hypoxic area extent with the local wind conditions using wind data from Kalbådagrund we found that for 2016, when the N-NE winds prevailed, the hypoxic area was enlarged, while with the W-SW winds, the area was smaller (**Table 2**). This result is in accordance with a recent analysis of long-term monitoring data by Lehtoranta et al. (2017) and a study by Väli et al. (2013) where the authors concluded that the increase in bottom oxygen concentrations in the GoF in 1990–1995 could be explained by stronger ventilation of the bottom layers due to increased westerlies. Also, a comparison of the 3-week average wind stress component from N-NE (20π) with the hypoxic depth at the Keri station (**Figure 5**) showed significant correlation between these parameters if early spring data from 2016 were excluded. The very deep hypoxic depth in March 2016 could be explained by weaker stratification and mixing of the water column (and a possible collapse of stratification) as it was observed earlier in winter by Liblik et al. (2013) and Lips et al. (2017). In 2017, the winds from the sector from east to north were almost absent, and thus, the intensification of estuarine circulation leading to the penetration of the near-bottom salt wedge into the GoF did not manifest itself in the same extent as in 2016. Secondly, we suggest that the high variability and large extent of hypoxia in late 2016 could be attributed to the influence of the 2014 MBI, which impact reached the GoF in 2016 (Liblik et al., 2018). The MBI influence constituted itself as a deep-water inflow of former deoxygenated water from the NBP to the GoF, which pushed the existing bottom water in the GoF upward resulting in an increase in volume and areal extent of hypoxic waters.

In conclusion, although the extent of hypoxia is an easily understandable indicator of near-bottom oxygen conditions in the GoF, still some issues have to be solved. An analysis of long-term data should be conducted to differentiate between the eutrophication-related and the wind- or inflow-related impacts (a first attempt is made by Lehtoranta et al., 2017). Also, the reference and target values – what percentage corresponds to the good status, have to be defined. For this purpose, the links of the extent of hypoxia with the nutrient load, nutrient concentrations and productivity have to be shown. Also, it is important to consider the feedback that the extent of hypoxia has on nutrient concentrations via the internal nutrient fluxes from the sediments in poor oxygen conditions (Pitkänen et al., 2001).

Oxygen debt indicator is the other easily understandable characteristic of oxygen conditions which could be understood as the apparent oxygen utilization – the difference between the oxygen saturation concentration and measured concentration. As seen in **Figure 7**, the oxygen debt estimates based on a single profile had very high short-term variability. It could be mostly explained by the observed halocline dynamics in the GoF displayed in **Figure 3** and shown by Liblik and Lips (2017). Internal waves cause up- and downward movement of the halocline and influence the depth range with the vertical salinity gradient larger than 0.07 g·kg−<sup>1</sup> ·m−<sup>1</sup> within the halocline layer. Consequently, the estimates of the oxygen debt based on separate profiles could reveal a relatively high variability, but an average of these estimates over a large number of profiles characterizes the sub-halocline oxygen conditions quite well.

Since we did not have the time-series of oxygen profiles covering the winter months, we were not able to calculate the yearly average oxygen debt as required by the HELCOM oxygen debt indicator description (HELCOM, 2018a). However, we suggest that for the GoF, where the halocline could be weak or occasionally absent in winter (Liblik et al., 2013; Lips et al., 2017), the oxygen debt could be estimated on the basis of seasonal data from May to September (or October). It also means that a new threshold for good environmental status should be suggested since the value 8.66 mg·l −1 , which is applied for most of the deep basins in the Baltic Sea, was defined as a volume specific average and for the yearly average (HELCOM, 2018a). We also suggest that the uniform threshold value for the volume specific average oxygen depth should be reconsidered since the basins have different depths while the DO content decreases with the depth. Instead, the oxygen debt just below the halocline could be used which might have similar threshold values for most of the Baltic Sea deep basins.

An advantage of the oxygen debt indicator compared to the extent of hypoxia is that it seems to be less dependent on the advection of the hypoxic salt wedge into and out from the inner GoF since an increase/decrease in the volume of hypoxic waters due to the horizontal advection results in upward/downward movement of the halocline and not in a decrease in DO

content just below the halocline. For instance, from September to October 2016, the extent of hypoxia almost doubled while the oxygen debt increased only from 11.4 mg·l −1 in September to 11.5 mg·l −1 in October.

Oxygen consumption results found in this study as averages for the whole period from spring to autumn (June–September) are in the range of other published results (**Table 4**). While in the earlier estimates, the monthly consumption rates varied between 0.25 and 1.29 mg l−<sup>1</sup> month−<sup>1</sup> , in the present study, the estimates in the Gulf of Finland were 0.82 mg l−<sup>1</sup> month−<sup>1</sup> for 2016 and 0.31 mg l−<sup>1</sup> month−<sup>1</sup> for 2017. This comparison is not entirely correct since the earlier estimates were mostly for longer (multiyear) periods and our study dealt only with the productive season, as well as we made our analysis for the intermediate layer (50– 60 m) while in the earlier studies deeper (sub-halocline) layers were considered. However, this agreement between different results confirms that the method proposed in our study is applicable to rough consumption estimates. The other question is whether it is accurate enough to be used for eutrophication status assessment. For instance, the negative values of monthly consumption estimates in July of both years indicate that the approach may not be appropriate for shorter periods.

We analyzed the sensitivity of the results regarding different choices and assumptions in the suggested method of consumption estimates. We showed that a slight change of the depth limits for the analyzed stagnant layer would almost not alter the results. The changes in solubility, which were not included in the suggested approach, could cause a bias less than about 15% of the observed variability (see section "Oxygen Consumption"). The choice of the empirical intensity factor of turbulence could change the diffusion estimate and consequently the consumption estimate only by 0.1 mg·l −1 ·month−<sup>1</sup> .

For the advection estimates, we used the linear correlation between salinity and oxygen, which is significant based on the data from the GoF. At the same time, we showed that the average change in oxygen concentration in the GoF along an isohaline in the layer under consideration could be as large as 0.9 mg·l −1 per a longitudinal degree. Thus, a relatively large bias in the consumption estimates up to 0.83 mg·l −1 ·month−<sup>1</sup> could occur if unidirectional flow prevails for a long period. Such on average unidirectional flow in the intermediate layer could exist in the GoF for a few months as shown by Lilover et al. (2017) although the characteristic current velocities there are usually smaller than in the surface and near-bottom layer (Suhhova et al., 2018). This bias estimate has the value close to the consumption estimates, and thus, it has to be addressed in the further analysis. For instance, in 2016, the saline and oxygen-deficient sub-halocline waters of the northern Baltic Proper were pushed by the MBI north-eastward to the GoF near-bottom layer (Liblik et al., 2018). It could cause uplift of old near-bottom water in the central and eastern GoF and the westward flow in the intermediate layer which could lead to an oxygen decrease. The observed decrease in oxygen content from July to September 2016 by 2.59 mg·l −1 was mostly assigned to consumption since the average salinity increase was only 0.18 g·kg−<sup>1</sup> and the corresponding oxygen decrease due to advection should not be as large according to the present calculation method based on a linear regression between salinity and oxygen content. This method, which has a bias in advection estimates when the flow is on average unidirectional in the analyzed layer, could also cause the large difference between the seasonal consumption estimates in 2016 and 2017. As a way forward, simultaneous observations of vertical structure of currents near the Keri station has been initiated in 2018.

The assumption of an equal surface area of the upper and lower border of the analyzed layer and no explicit water-sediment oxygen fluxes means that the downward oxygen flux through the lower boundary is entirely estimated as a turbulent flux in the water column. This approach is correct for an open sea area; however, one could expect that the oxygen flux from water to the sediments at the slopes of the basin (Holtermann et al., 2017) and a horizontal mixing toward the basin interior impact the consumption estimate results. As shown by Koop et al. (1990) the oxygen flux from the water to the sediments under oxic conditions could be as large as 777 µmol·O·m−<sup>2</sup> h −1 corresponding to approximately 1.8 mg·l −1 ·month−<sup>1</sup> if a 10 m thick water layer above the seabed is considered. The maximum oxygen flux through the lower border of the analyzed layer found in the present study could cause a decrease in DO content of 0.7 mg·l −1 ·month−<sup>1</sup> . Thus, an additional export of oxygen from the stagnant layer due to the flux from water to the sediments and horizontal mixing should increase the consumption estimates. When interpreting our results, one has to consider that the consumption values include both the local consumption in the studied open-sea area and a part of consumption in the sediments at the surrounding slopes of the basin. This additional export of oxygen from the stagnant layer should increase the consumption estimates. Thus, it does not explain the negative values found in


our study for June–July 2016 and 2017 which most probably are related to the bias in advection estimates.

A comparison of the three indicators reveals that their average seasonal values almost did not vary between the two analyzed years for the hypoxic area extent and oxygen debt while about 2.7 times higher oxygen consumption was found in 2016 than in 2017. The average hypoxic depth in May–October was 64.5 m in 2016 and 64.6 m in 2017, which result in an almost equal hypoxic area extent for both years. The average oxygen debt in May–October was 11.3 mg·l −1 in 2016 and 11.6 mg·l −1 in 2017. These values differ only by 0.3 mg·l −1 , which is <3% of the indicator result. We suggest that the found large difference in oxygen consumption estimates between 2016 and 2017, which is inconsistent with the other indicator results, could be related to the applied methodology and/or the fact that the consumption estimates were found for the intermediate layer while hypoxic area and oxygen debt for the sub-halocline layer. What could be the main natural factors causing this large difference, e.g., influence of productivity or riverine water, since in 2016 the upper layer was fresher than in 2017, needs future studies. The largest uncertainty is caused by the too simplified estimate of the oxygen advection. Although oxygen consumption would be the best indicator, which should directly be dependent on the productivity of the sea area, it requires further development. Regarding the observational program, simultaneous measurements of current profiles could be beneficial to decrease the uncertainty of advection estimates.

## CONCLUSION

We have described the oxygen conditions in the GoF in 2016– 2017 based on observations mostly at the Keri profiling station where vertical profiles of temperature, salinity and oxygen were acquired up to 8 times a day. The applicability of high-frequency data from this fixed automated station and the three adapted oxygen indicators for the eutrophication status assessments were tested. The main results of the analysis are:


### REFERENCES


We concluded that all three tested indicators have methodological challenges to be solved if they are used for the eutrophication effects assessment. The main issue is related to the differentiation between natural changes and eutrophication-related impacts. To increase the confidence of eutrophication assessments both high-frequency profiling should be implemented in the monitoring programs and more accurate estimates of changes due to physical processes are required.

## DATA AVAILABILITY STATEMENT

Part of the data (temperature, salinity) can be found on EMODnet Physics http://www.emodnet-physics.eu/Map/ DefaultMap.aspx when searching for "KeriCable" station. Oxygen datasets are available on request.

## AUTHOR CONTRIBUTIONS

S-TS was the main responsible person in developing methods, analyzing data, and writing the manuscript. UL contributed to developing methods and writing the manuscript. TL contributed to analyzing data regarding the influence of hydrography.

## FUNDING

The work was financially supported by the Institutional Research Funding IUT (IUT19-6) of the Estonian Ministry of Education and Research and Environmental Investment Center environmental program project KIK17144.

## ACKNOWLEDGMENTS

We would like to thank the Finnish Meteorological Centre, for providing wind data, the crew of r/v Salme and our colleagues.



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 Stoicescu, Lips and Liblik. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Wadden Sea Eutrophication: Long-Term Trends and Regional Differences

Justus E. E. van Beusekom<sup>1</sup> \*, Jacob Carstensen<sup>2</sup> , Tobias Dolch<sup>3</sup> , Annika Grage<sup>4</sup> , Richard Hofmeister<sup>1</sup> , Hermann Lenhart<sup>5</sup> , Onur Kerimoglu<sup>1</sup> , Kerstin Kolbe<sup>6</sup> , Johannes Pätsch<sup>7</sup> , Johannes Rick<sup>3</sup> , Lena Rönn<sup>6</sup> and Hans Ruiter<sup>8</sup>

1 Institut für Küstenforschung, Helmholtz-Zentrum Geesthacht, Geesthacht, Germany, <sup>2</sup> Department of Bioscience, Aarhus University, Roskilde, Denmark, <sup>3</sup> Wadden Sea Station Sylt, Alfred-Wegener-Institut, Helmholtz-Zentrum für Polar-und Meeresforschung, Bremerhaven, Germany, <sup>4</sup> Bundesamt für Seeschifffahrt und Hydrographie, Hamburg, Germany, <sup>5</sup> Department of Informatics, University of Hamburg, Hamburg, Germany, <sup>6</sup> Niedersächsischer Landesbetrieb für Wasserwirtschaft, Küsten- und Naturschutz, Betriebsstelle Brake-Oldenburg, Oldenburg, Germany, <sup>7</sup> Institute of Oceanography, University of Hamburg, Hamburg, Germany, <sup>8</sup> Rijkswaterstaat, Utrecht, Netherlands

#### Edited by:

Alberto Basset, University of Salento, Italy

#### Reviewed by:

Paolo Magni, Italian National Research Council (CNR), Italy Hans Paerl, The University of North Carolina at Chapel Hill, United States Melanie Beck, University of Oldenburg, Germany

\*Correspondence:

Justus E. E. van Beusekom justus.van.beusekom@hzg.de

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 03 October 2018 Accepted: 14 June 2019 Published: 04 July 2019

#### Citation:

van Beusekom JEE, Carstensen J, Dolch T, Grage A, Hofmeister R, Lenhart H, Kerimoglu O, Kolbe K, Pätsch J, Rick J, Rönn L and Ruiter H (2019) Wadden Sea Eutrophication: Long-Term Trends and Regional Differences. Front. Mar. Sci. 6:370. doi: 10.3389/fmars.2019.00370 The Wadden Sea is a shallow intertidal coastal sea, largely protected by barrier islands and fringing the North Sea coasts of Netherlands, Germany, and Denmark. It is subject to influences from both the North Sea and major European rivers. Nutrient enrichment from these rivers since the 1950s has impacted the Wadden Sea ecology including loss of seagrass, increased phytoplankton blooms, and increased green macroalgae blooms. Rivers are the major source of nutrients causing Wadden Sea eutrophication. The nutrient input of the major rivers impacting the Wadden Sea reached a maximum during the 1980s and decreased at an average pace of about 2.5% per year for total Nitrogen (TN) and about 5% per year for total Phosphorus (TP), leading to decreasing nutrient levels but also increasing N/P ratios. During the past decade, the lowest nutrient inputs since 1977 were observed but these declining trends are leveling out for TP. Phytoplankton biomass (measured as chlorophyll a) in the Wadden Sea has decreased since the 1980s and presently reached a comparatively low level. In tidal inlet stations with a long-term monitoring, summer phytoplankton levels correlate with riverine TN and TP loads but stations located closer to the coast behave in a more complex manner. Regional differences are observed, with highest chlorophyll a levels in the southern Wadden Sea and lowest levels in the northern Wadden Sea. Model data support the hypothesis that the higher eutrophication levels in the southern Wadden Sea are linked to a more intense coastward accumulation of organic matter produced in the North Sea.

Keywords: eutrophication indicators, Wadden Sea, North Sea, nutrients, long-term trends, phytoplankton, submersed vegetation, sediments

## INTRODUCTION

Increased nutrient fluxes to the coastal ocean are one of the main drivers for coastal change, having a pronounced impact on phytoplankton biomass and primary production, harmful algae blooms, seagrass, green macroalgae blooms, and water transparency (Cloern, 2001; Boesch, 2002; Orth et al., 2006; Smetacek and Zingone, 2013). Given the continuous increase in anthropogenic nitrogen fixation and increasing needs for nitrogen fertilizers (Battye et al., 2017), a further increase

in the above mentioned eutrophication symptoms can be expected (e.g., Breitburg et al., 2018). First signs of coastal eutrophication like enhanced algae blooms were already observed in the 1950s and in Europe. Decisions to combat eutrophication were taken since the 1970s and 1980s (de Jong, 2007). Decreasing eutrophication trends have been observed for instance in European waters like the North Sea, Danish coastal waters or in the Baltic (e.g., Carstensen et al., 2006; Emeis et al., 2015; Andersen et al., 2017) during the 2000s or in Chesapeake Bay (Harding et al., 2016; Lefcheck et al., 2018) showing that a reversal of eutrophication is possible.

In the present paper, we will describe and compare the eutrophication levels and trends of the international Wadden Sea, a shallow coastal sea along the Dutch, German and Danish North Sea coast, based on long-term observations on nutrients and phytoplankton biomass (chlorophyll a). Since the earliest nutrient measurements in the Wadden Sea during the mid-20th century (Postma, 1954, 1966; Hickel, 1989) a strong increase in nitrogen and phosphorus concentrations has been documented (e.g., de Jonge and Postma, 1974; Hickel, 1989) reaching maximum values during the 1980s and 1990s (Jung et al., 2017) and decreasing since (e.g., van Beusekom et al., 2001, 2009; Cadée and Hegeman, 2002). Among the regularly occurring negative effects associated with the increased nutrient inputs are an increased import of organic matter from the North Sea to the Wadden Sea (de Jonge and Postma, 1974), more intense Phaeocystis-blooms (Cadée and Hegeman, 1986), a decline in seagrass distribution (de Jonge and de Jong, 1992), and increased green macroalgal cover (Reise and Siebert, 1994). Measures were taken during the 1970s and 1980s to address eutrophication (de Jong, 2007) through a reduction of riverine nutrient inputs to the North Sea.

Several studies have described local changes in Wadden Sea nutrient dynamics with a focus on the Dutch Wadden Sea (e.g., de Jonge and Postma, 1974; Cadée and Hegeman, 2002; Philippart et al., 2007) and the Northfrisian Wadden Sea (e.g., Hickel, 1989; van Beusekom et al., 2009). A comparison of available data for the entire Wadden Sea on phytoplankton and nutrients was only carried out within the framework of so-called Wadden Sea Quality Status Reports (e.g., van Beusekom et al., 2017). In these reports, the international scientific Wadden Sea community regularly compiles data on the ecological status of the North Sea (see for instance Wolff et al., 2010), showing among others a general decline in eutrophication levels.

Special attention will be given in the present study to long time series and regional differences in eutrophication levels covering the entire Wadden Sea. We will use two previously developed and tested eutrophication proxies (e.g., van Beusekom et al., 2001, 2009; van Beusekom and de Jonge, 2002) to describe temporal trends and regional differences. The proxies include autumn values of NH<sup>4</sup> + NO<sup>2</sup> as an indicator of organic matter turnover and summer phytoplankton chlorophyll a as an indicator of phytoplankton growth potential.

One of the processes responsible for the high productivity of Wadden Sea ecosystem is import of organic matter from the North Sea (Verwey, 1952; Postma, 1954). van Beusekom et al. (1999) compiled organic matter budgets from the Wadden Sea supporting the heterotrophic nature of the Wadden Sea with a primary production level of around 200–300 g C m−<sup>2</sup> y −1 , remineralization levels of about 300–450 g C m−<sup>2</sup> y −1 , and an annual import of about 100 g C m−<sup>2</sup> y −1 . This import is an important driver of nutrient availability in the Wadden Sea (e.g., de Jonge and Postma, 1974; van Beusekom et al., 1999, 2012; van Beusekom and de Jonge, 2002). The import mechanism is as follows (e.g., van Beusekom et al., 2012): Transport of particles (both inorganic and organic) from the North Sea to the Wadden Sea is induced by residual, baroclinic circulation similar to estuarine circulation (e.g., Burchard et al., 2008; Flöser et al., 2011; Hofmeister et al., 2017) with bottom currents directed on average toward the coast, and surface currents directed toward the sea. This traps sinking particles from a large part of the coastal North Sea, transports them toward the Wadden Sea and prevents particles to escape from the Wadden Sea to the North Sea. By the same token, dissolved nutrients, released after organic matter remineralization and not taken up within in the Wadden Sea, can be transported back to the North Sea (e.g., Grunwald et al., 2010). As light conditions are generally much better outside the Wadden Sea, these nutrients can be taken up by phytoplankton, and are transported back into the Wadden Sea after settling into deeper water layers (e.g., Postma, 1981; Maerz et al., 2016). Hofmeister et al. (2017) demonstrated with a model that the transport of sinking phytoplankton and detritus toward the Wadden Sea is able to keep up a nutrient gradient between the North Sea and the Wadden Sea.

We will use a biogeochemical North Sea model (Kerimoglu et al., 2017) to investigate regional differences in the import of organic matter from the North Sea to the Wadden Sea. Model approaches are needed because available budgets are too coarse to resolve regional differences. Also, no recent carbon budgets are available that demonstrate changes in Wadden Sea eutrophication levels. We will show that with the two proxies -summer chlorophyll a and autumn NH<sup>4</sup> + NO2, a general decrease in eutrophication can be linked to decreasing riverine nutrient loads. We will further show that high organic matter import rates in the southern part of the Wadden Sea lead to higher eutrophication level than in the northern part of the Wadden Sea.

#### MATERIALS AND METHODS

#### Area Description

The Wadden Sea is a shallow intertidal coastal sea along the Dutch, German, and Danish North Sea coast (**Figure 1**). Tidal range is about 1 m at the northern and western end and increases to more than 3 m in the central part. At tidal ranges below 3 m, barrier islands can be formed (e.g., Hayes, 1975; Oost and de Boer, 1994), protecting the southern and northern Wadden Sea, but are absent in the central part. About 50% of the sediments emerge during low tide. Sediments are mostly sandy in the more exposed parts, whereas in the more protected parts muddy sediments prevail. Major nutrient discharges directly into the southern Wadden Sea are the IJsselmeer and the Ems river. Major rivers entering the central Wadden Sea are the Weser and

Elbe. Residual currents along the Wadden Sea follow the anticlockwise circulation of the North Sea. Thus, nutrient loads of the Rhine and Maas are carried northward and westward toward the Wadden Sea especially impacting the southern Wadden Sea. The Weser and Elbe nutrient loads are carried northward especially impacting the northern Wadden Sea (**Figure 1**).

Salinity ranges in most tidal inlets on average between about 27 and 31. A clear seasonality is observed with lowest values in winter and highest values in summer (e.g., van Aken, 2008a; van Beusekom et al., 2008; Grunwald et al., 2010). Deviations from long-term means are related to freshwater discharge (e.g., van Beusekom et al., 2008). Closer to the mainland, and especially near fresh water sources, lower salinities prevail (e.g., van Aken, 2008a; Grunwald et al., 2010).

Temperature ranges on average between minimum values of around 3◦C in February and around 18◦C in July/August (e.g., Diederich et al., 2005; van Aken, 2008b; Grunwald et al., 2010) with an increasing trend (Martens and van Beusekom, 2008).

Phytoplankton shows a clear seasonal cycle, mainly due to light limitation in winter (e.g., Loebl et al., 2009). The spring phytoplankton bloom is dominated by diatoms (e.g., Cadée, 1986; Loebl et al., 2007), mainly occurs during April, but can be earlier during cold winters (van Beusekom et al., 2009). It is often followed by a Phaeocystis-bloom during early summer (e.g., Cadée and Hegeman, 2002; Loebl et al., 2007).

Nutrient seasonal cycles reflect nutrient uptake by primary producers, but the shape depends on the nutrient species. Silica (H4SiO4) and nitrate (NO3) show a simple cycle with highest values during winter and low values during summer due to uptake by phytoplankton (e.g., van Bennekom et al., 1974; van Beusekom et al., 2008; Leote et al., 2016). Ammonium (NH4) and nitrite (NO2) also show high winter and low summer values but with a clear autumn maximum (e.g., Postma, 1966; van Beusekom and de Jonge, 2002; Loebl et al., 2007; Reckhardt et al., 2015). Phosphate (PO4) reaches lowest values during spring and early summer, but concentrations already increase during May/June (e.g., de Jonge and Postma, 1974; Loebl et al., 2007; Grunwald et al., 2010; van Beusekom and de Jonge, 2012). An overview of mean seasonal nutrient cycles in the Wadden Sea until the mid 1990s is given in van Beusekom et al. (2001).

#### Data Compilation Riverine Input Data

Riverine nutrient input data were compiled by Lenhart and Pätsch (2004) and updated to 2017. Total nitrogen (TN) and total phosphorus (TP) concentrations, taken from stations within the freshwater part of the estuary near the river mouth, were linearly interpolated between observations (mostly about every 2 weeks) to achieve estimates of daily concentrations. These nutrient concentrations estimates were then multiplied with daily discharges values taken from the last tidal free gauge station to derive daily loads. Data are available through https://wiki.cen. uni-hamburg.de/ifm/ECOHAM/DATA\_RIVER.

#### Dutch Wadden

For the Dutch Wadden Sea, a coherent observational program with methods applicable to the Wadden Sea is carried out by the Dutch governmental agency Rijkswaterstaat since 1977.

Frequency of the sampling is fortnightly and samples are taken between 2 h before or after high tide. Water samples are filtered and analyzed for dissolved nutrients according to CEN/ISO guidelines and intercalibrated. Chlorophyll is measured by HPLC and is intercalibrated. During the past decades, the number of stations that are regularly sampled has decreased and at present five stations are continuously monitored (**Figure 1**). All Dutch monitoring data can be downloaded from the internet<sup>1</sup> .

#### Lower Saxonian Wadden Sea (Norderney, Germany)

Nutrients and phytoplankton are monitored at the island of Norderney since 1987. The exact position of the sampling station has changed during the time series within a radius of 2 km around Norderney harbor. Sampling was carried out at high tide, but includes a period with additional sampling at low tide from 2005 to 2014. Differences in chlorophyll a and NH<sup>4</sup> + NO<sup>2</sup> between these samplings is small (∼1 µg chla/l or ∼6% and ∼2– 3 µmol/l or ∼22% for NH<sup>4</sup> + NO2; compare van Beusekom et al., 2017). In this study, we have combined the data. Nutrients were filtered (0.4 µm), originally analyzed after Grasshoff et al. (1976, 1983) and later using continuous flow analysis (CFA) using German standard (DIN) methods. Labs are quality assured through interlaboratory comparisons: LÜRV (nationally) and Quasimeme (internationally). Data cannot be retrieved from a public data base but all data can be requested from NLWKN.

#### Northfrisian Wadden Sea (Sylt, Germany)

Regular sampling in the List tidal basin (Northfrisian Wadden Sea) was started in 1984 by the Biologische Anstalt Helgoland, now part of the Alfred-Wegener-Institut, Helmholtz-Zentrum für Polar- und Meeresforschung (AWI). Two stations (one in the main channel and one in the entrance to the Königshafen bay) were sampled twice weekly irrespective of the tidal phase. Until 1998, unfiltered nutrient samples were analyzed, afterward water samples were filtered through 0.4 Nuclepore polycarbonate filters and stored frozen until analysis. The impact of filtration on nutrient concentrations was not investigated systematically, but except for phosphate no clear shifts were observed in the time series. Nutrients were analyzed manually (NO<sup>3</sup> with CFA) after Grasshoff et al. (1983) and intercalibrated (Quasimeme). Chlorophyll was analyzed spectrophotometrically after Jeffrey and Humphrey (1975) and intercalibrated. Data are deposited in the Pangaea database (www.pangaea.de). Details can be found in van Beusekom et al. (2009).

#### Danish Wadden Sea

The Danish Wadden Sea has been monitored since 1989. All measurements are according to Danish standards<sup>2</sup> . In short, chlorophyll was extracted with ethanol and the absorbance is measured spectrophotometrically. Nutrients were filtered (GF/F or Advantec GF75) and analyzed using standard wet chemical techniques. In contrast to the other time series, NO<sup>2</sup> is not assessed separately. Labs must participate in quality assurance

<sup>1</sup>https://waterinfo.rws.nl

<sup>2</sup>http://bios.au.dk/raadgivning/fagdatacentre/fdcmarintny/ gaeldendetekniskeanvisninger/#c236812

programs (Quasimeme) to assure comparability of results. Data are accessible<sup>3</sup> .

#### Model Data

We use simulation data from Kerimoglu et al. (2017) to discuss the mechanisms controlling the cross-shore gradients between the North Sea and the Wadden Sea. The three-dimensional coupled model describes the biogeochemical processes within the southeastern North Sea, coupled to the General Estuarine Transport Model (Burchard and Bolding, 2002), defined on a curvilinear grid of 1.5–4 km and terrain-following 20 vertical layers. For the period 2000–2010, his model showed high skill in reproducing the spatial and temporal distribution of salinity, nutrients and chlorophyll; and in particular, capturing the steep cross-shore gradients between the North Sea and the Wadden Sea (Kerimoglu et al., 2017). We extracted transects of particulate organic nitrogen (PON), salinity and residual, cross-shore currents between the ∼40 m depth contour and the Wadden Sea for the summer (May – September) of 2010 near the long-term stations, and present here the temporal average results.

#### Data Handling

In order to have an equal distribution of data over time, we binned data into monthly means. Handling and statistics were done with R (R Core Team, 2016), using the graphical package ggplot2 (Wickham, 2016). Processing and plotting of the model data was performed with nco<sup>4</sup> and Python 2.7.

#### Eutrophication Proxies and Drivers

Two eutrophication proxies were used: summer chlorophyll a and autumn levels of the NH<sup>4</sup> + NO2. Summer chlorophyll a levels were the mean of the monthly means for May – September. We did not include spring values, as the intensity of the spring bloom at least in the northern Wadden Sea is closely related to temperature, with highest biomasses after cold winters (van Beusekom et al., 2009) probably due to low grazer activity. Autumn NH<sup>4</sup> + NO<sup>2</sup> levels were the mean of the monthly means for September to November. As for the Danish time series, no NO<sup>2</sup> values were reported and we estimated the NO<sup>2</sup> contribution as 25% of the NH<sup>4</sup> levels based on the nearby AWI time series from the Northfrisian Wadden Sea. The proxies are based on earlier work (e.g., van Beusekom et al., 2001; van Beusekom and de Jonge, 2002). Support for the usefulness of these proxies was given by van Beusekom and de Jonge (2012), who showed good correlations between summer dissolved organic phosphorus concentrations and the above eutrophication proxies.

As a drivers of the Wadden Sea eutrophication we used riverine TN loads and TP loads as suggested by van Beusekom et al. (2001). Instead of annual loads we used the loads until August as later riverine input cannot impact the production of organic matter during Summer. In case of the rivers Rhine and Maas, we extended the time series to include the December loads of the previous year as suggested by van Beusekom et al. (2001),

<sup>3</sup>https://arealinformation.miljoeportal.dk/html5/index.html?viewer=distribution <sup>4</sup>https://github.com/nco/nco

among others based on the fact that the first high winter discharges are in December and on the longer distance between river mouth and Wadden Sea as compared to the Elbe and Weser. Riverine discharges directly into the Wadden Sea like the river Ems and lake IJsselmeer can have strong local effects (e.g., de Jonge and Postma, 1974; de Jonge, 1990; Jung et al., 2017), but less on the adjacent tidal basins. We therefore used only riverine TN and TP loads by the rivers Rhine/Maas as a common driver for the entire southern Wadden Sea. These two rivers dominate the riverine nutrient concentrations in the North Sea off the southern Wadden Sea. The rivers Rhine and Maas account for around 80% of the nutrient loads, with IJsselmeer, Ems and Noordzeekanaal representing 20%. Moreover, Rhine + Maas loads correlate with the combined loads of Rhine, Maas, IJsselmeer, Ems, and Noordzeekanaal with an r <sup>2</sup> > 0.98.

#### RESULTS

#### Riverine Input

We used the TN and TP loads by the rivers Rhine and Maas as the main driver of the southern Wadden Sea nutrient dynamics, and the TN and TP loads by the rivers Weser and Elbe as the main driver of the northern Wadden Sea nutrient dynamics. Discharge (**Figure 2**) has a major impact on the riverine loads showing a high interannual variability with highest discharges in the 1980s, and at least for the Rhine/Maas rivers a decreasing tendency of peak discharge. Riverine TN and TP loads reflect the annual discharge (**Figure 3**) but in addition show a clear decreasing trend in both river systems. As discharge during the 1980s was higher than present, riverine TN loads were up to three times higher and the riverine TP loads were up to five times higher than present levels. TP decreased faster than TN. In the Rhine/Maas, molar N/P ratios continuously increased from around 25 in the 1980s to about 80 during recent years. In the rivers Weser/Elbe, N/P molar ratios increased from 40 in the 1980s, to about 60–70 in the 1990s, decreasing to recent levels of about 45 (**Figure 4**). Flow-weighed TN concentrations (annual nutrient load/annual discharge, **Figure 5**) reached a maximum around 1985, regularly decreasing at a rate of around 2.5% per year until 2017. Flowweighed TP was decreasing faster, with a rate of about 8% until 1990 in both system. Whereas in the Rhine/Maas this trend continues, the decrease in TP loads (2.5%) of the Weser/Elbe is much slower after ∼1990. Overall, we observed a decrease in flow-weighed concentrations of about 50% (TN) and 70–85% (TP), respectively.

### Chlorophyll a

Mean summer chlorophyll a concentrations (Chl) in the southern part of the Wadden Sea ranged between 2.6 and 48 µg Chl /l (**Figure 6**) with a median of 11.8 µg Chl /l. Most Dutch time series showed a significant decrease over time (**Figure 6** and **Table 1**). The clearest trends were observed in the tidal inlet stations "Marsdiep Noord" (NL), Vliestroom (NL), and "Huibertgat oost" (NL). At the station "Doove Balg west" (NL) situated more closer to the coast, the trend was significant but

with a higher variability. The time series at Norderney (D) also showed a significant decrease. The combined time series of the four southern Wadden Sea tidal inlet stations with a decreasing trend (**Figure 7**) reveal a general pattern with initially increasing chlorophyll a levels peaking in 1987, decreasing afterward.

In contrast to the southern Wadden Sea, summer chlorophyll a levels were much lower in the northern Wadden Sea ranging, between 2.0 and 16.9 µg Chl /l and with median of 6.9 µg Chl /l. A significant decreasing trend was only observed for the longer Sylt (D) time series and no trends were found for the Danish Wadden Sea time series. The contrast between the Sylt and the southern tidal inlet stations of Netherlands and Lower Saxony is clearly shown in **Figure 7**. It is noteworthy, that with decreasing chlorophyll a levels during recent years, regional differences were becoming increasingly smaller.

### Autumn Concentrations of NH<sup>4</sup> and NO<sup>2</sup>

Seasonal dynamics of nitrogen-containing nutrients (NH4, NO2, and NO3) in the Wadden Sea showed decreasing values during spring, low values in summer and increasing concentrations from September onward (e.g., van Beusekom and de Jonge, 2002; van Beusekom et al., 2009), due to decreasing light availability and subsequent decreasing nutrient uptake rates (see section "Area Description"). We used the summed autumn

concentrations of NH<sup>4</sup> and NO<sup>2</sup> as a proxy of organic matter turnover, as these compounds are mainly produced locally by degradation processes. We did use NO3, although NO<sup>3</sup> is also produced locally by degradation processes and nitrification. However, in contrast to NH<sup>4</sup> and NO2, NO<sup>3</sup> presently is the main form of nitrogen in rivers. Thus, both riverine input and local degradation processes may impact Wadden Sea NO<sup>3</sup> concentrations (van Beusekom and de Jonge, 2002). Autumn concentrations of NH<sup>4</sup> + NO<sup>2</sup> have decreased significantly at all southern Wadden Sea stations, whereas in the northern Wadden Sea a significant decrease was only observed at the Danish Wadden Sea station Grådyb (**Figure 8** and **Table 1**). As for chlorophyll a, the autumn NH<sup>4</sup> + NO<sup>2</sup> concentrations in the southern Wadden Sea (median: 11.9 µmol/l) were substantially higher than in the northern Wadden Sea (median: 6.0 µmol/l). It is important to note that at the Dutch Station Dantziggat, a significant decrease in autumn NH<sup>4</sup> + NO<sup>2</sup> was observed.



## Summer Chlorophyll a and Autumn NH<sup>4</sup> + NO2: Two Independent Eutrophication Proxies

Most summer chlorophyll a time series showed significant correlations with riverine nutrient loads (**Figure 9**). It made no clear difference whether TN or TP was used as causative factor as evident by very similar r 2 values (**Table 2** for TN and **Table 3** for TP). As for the temporal trends, correlations between riverine loads and chlorophyll a are clearest for the tidal inlet stations of the southern Wadden Sea [Marsdiep noord (NL), Vliestroom (NL), Huibertgat oost (NL), and Norderney(D)], weak for Doove Balg west (NL), and absent for Dantziggat (NL). The northern Wadden Sea stations all show a weak but significant correlation with lower r 2 values. In most cases, riverine TN loads provided slightly better predictions than riverine TP loads. We also observed strong correlation between autumn NH<sup>4</sup> + NO<sup>2</sup> and riverine nutrient loads in the southern Wadden Sea (**Figure 10**), with no clear differences between TP, and TN as predictors. In the northern Wadden Sea, correlations between autumn NH<sup>4</sup> + NO<sup>2</sup> and riverine nutrient loads were not significant, except for TP in Grådyb (DK). Both proxies, summer chlorophyll a and autumn NH<sup>4</sup> + NO2, were significantly correlated (p << 0.001, r <sup>2</sup> = 0.26), and both identify the southern Wadden Sea as having the highest eutrophication levels.

#### Model Data

Model results (see Kerimoglu et al., 2017) show, that within the regions where a thermohaline stratification occurs, PON concentrations at the bottom layers are slightly higher than at the surface layers within the deep (>20 m) regions (**Figure 11**). This, in combination with an estuarine-like circulation mentioned above, where the residual surface currents are directed toward off-shore and the bottom currents directed toward the shore, results in a net PON transport toward the shore, thus contributing to the eutrophication of the Wadden Sea. The efficiency of this transport mechanism varies between different regions: The intensity of the bottom transport is much higher along the southern Wadden Sea with average values of about 2 cm/s compared to the northern Wadden Sea (1 cm/s or less).

Also, the PON concentrations are in general higher along the southern Wadden Sea than along the northern Wadden Sea.

#### DISCUSSION

#### The Role of Sediments in Wadden Sea Nutrient Cycling

The present results show that riverine nutrients are a major driver of the long term phytoplankton and nitrogen dynamics in the Wadden Sea. An essential link is the primary production of organic matter in the North Sea, import of primary produced organic matter from the North Sea into the Wadden Sea and remineralization within the Wadden Sea (van Beusekom and de Jonge, 2002). Sediments are a major site of organic matter remineralization and constitute a significant sources of nutrients (e.g., Beck and Brumsack, 2012). In shallow coastal settings, benthic, and pelagic contribute about equally to the total organic matter turn-over (Heip et al., 1995). Wadden Sea carbon budgets support this (van Beusekom et al., 1999). Grunwald et al. (2010) estimated nutrient fluxes from the Lower Saxonian Wadden Sea to the North Sea based on high resolution water column nutrient measurement from an automatic measurement station. They concluded, that particulate nutrients have to be imported from the North Sea to balance the export of inorganic nutrients. Grunwald et al. (2010) further identified pore water discharge as a major nutrient source. The release of nutrients from organic matter remineralization in sediments may be fast in permeable sediments, when surface sediments are rapidly flushed (e.g., Huettel and Rusch, 2000; Ehrenhauss et al., 2004; de Beer et al., 2005). Sediments may also trap nutrients on longer time scales, if transport of pore water to deeper sediment layers is involved (e.g., Beck et al., 2008; Røy et al., 2008). This internal storage of nutrients may induce lag effects in the response of the Wadden Sea to changes in eutrophication levels. Jung et al. (2017) constructed a nutrient budget for the Western Dutch Wadden Sea based on long-term nutrient measurements. They explicitly accounted for sediment import from the North Sea as this part of the Wadden Sea has a sediment deficit due to the construction of a dike (Afsluitdijk) between the Wadden Sea and the former Zuiderzee (now IJsselmeer, see **Figure 1**). Jung et al. (2017) concluded that during the initial eutrophication phase (until 1981) P and N were imported. A net export of N was observed since 1981 and for P since 1992.

Despite the potential role of sediments to buffer nutrients, a rapid response to changing riverine N loads can be deduced from work by van Beusekom and de Jonge (2002). They showed that during wet years substantially more NH<sup>4</sup> + NO<sup>2</sup> is released in autumn than during dry years. Also phytoplankton dynamics respond quite fast to changes in riverine nutrient loads (compare **Figures 2**, **7**): During the 1980s, high riverine TN and TP

concentrations and high discharges caused peak loads and high chlorophyll a levels during the late 1980s. This was followed by a period (1989–1993) of very low discharges, strongly reduced

TABLE 2 | Summary statistics of the correlation between riverine Total Nitrogen loads and summer chlorophyll (May – September) and autumn (NH<sup>4</sup> and NO2).


For the southern Wadden Sea Stations Marsdiep noord (NL), Vliestroom (NL), Doove Balg west (NL), Dantziggat (NL), Huibertgat oost (NL), and Norderney (D) the combined loads of the Rhine and Maas were takten (months 12 of the previous year until month 8 of a given year). For the northern Wadden Sea stations Sylt (D), Knude Dyb outer (DK), and Grådyb inner (DK) the combined loads of the Weser and Elbe were taken (month 1 – month 8).

loads, and rapidly decreasing chlorophyll a levels. This suggests a rapid response of the phytoplankton to prevailing nutrient availability despite a potential nutrient buffering by sediments. Also, the present analysis does not suggests a major role of N and P legacies in sediments on the long term eutrophication levels, as the river loads (reflecting interannual dynamics) are a better predictor of the eutrophication level than a simple linear (smoothed) temporal trend. This is in line with the conclusions of Jung et al. (2017).

Nutrient ratios in rivers impacting the Wadden Sea have increased toward high N/P ratios clearly above the Redfield ratio, with potential consequences for the coastal food webs (see below). Wadden Sea sediments play an important role in lowering (and thus improving) these ratios in the Wadden Sea and adjacent North Sea: On the one hand, denitrification rates are high (e.g., Gao et al., 2009; Deek et al., 2012) removing significant amounts of N, on the other hand, sediments have a high capacity to bind and buffer PO<sup>4</sup> (e.g., Leote and Epping, 2015).

## Limiting Factors of Phytoplankton Growth in the Wadden Sea

Phytoplankton dynamics in the Wadden Sea are governed by a multitude of factors including zooplankton grazing (e.g., Loebl and van Beusekom, 2008), macrobenthos filter feeder activity (e.g., Cadée and Hegeman, 1974; Asmus and Asmus, 1993), light



For the southern Wadden Sea Stations Marsdiep noord (NL), Vliestroom (NL), Doove Balg west (NL), Dantziggat (NL), Huibertgat oost (NL), and Norderney (D) the combined loads of the Rhine and Maas were taken (months 12 of the previous year until month 8 of a given year). For the northern Wadden Sea stations Sylt (D), Knude Dyb outer (DK), and Grådyb inner (DK) the combined loads of the Weser and Elbe were taken (month 1 – month 8).

availability (Colijn, 1982), nutrient limitation (Loebl et al., 2008), or interactive effects of light limitation and nutrients (Colijn and Cadée, 2003; Loebl et al., 2009). We will discuss, whether trends in these three factors -grazing, light limitation and nutrient limitation- may have impacted the overall decreasing trends observed in the Wadden Sea during the last 3 decades.

Grazing by benthic filter feeders or (micro)zooplankton may have a large impact on phytoplankton dynamics. However, the observed recent decline in phytoplankton biomass is most probably not due to an increase in grazing pressure: For Wadden Sea zooplankton, only one time series is available since 1984 as part of the AWI-Sylt time series. Long term trends show a longer copepod season, an increase in Acartia sp. in April-May but no clear overall trend in copepod abundance (Martens and van Beusekom, 2008). This is in line with observations from the Helgoland road time series, even showing an overall decline in zooplankton densities (Boersma et al., 2015). Macrozoobenthos densities in the Wadden Sea have been relatively stable during the past decades (Drent et al., 2017) or even decreased in the adjacent coastal zone (Meyer et al., 2018). Philippart et al. (2007) did not observe a clear decrease in macrobenthos biomass in the western Dutch Wadden Sea but did observe a decrease in filter capacity by the macrobenthos. Taken together, no convincing evidence exists to support that recent long-term phytoplankton decline is driven by an increased top-down control. Whereas the long term trend in summer phytoplankton biomass is not driven by top-down control, short term dynamics, and interannual difference in summer phytoplankton biomass can be related to grazer-phytoplankton interactions. The tight coupling between phytoplankton growth rates and microzooplankton grazing rates in summer was demonstrated by Loebl and van Beusekom (2008). In the inner German Bight (Helgoland), Wiltshire et al. (2015) concluded that zooplankton grazing may control phytoplankton dynamics.

There has been a debate whether light, N or P limit phytoplankton dynamics in the Wadden Sea (e.g., de Jonge, 1990; Colijn and Cadée, 2003; Philippart et al., 2007). Colijn and Cadée (2003) used a method by Cloern (1999) to derive from time series, whether phytoplankton growth is limited by nitrogen or light and concluded that light is the dominating limiting factor, especially at Wadden Sea stations close to the coast like in the Dollard (inner part of the Ems estuary), with some N-limited periods during summer at Marsdiep, Norderney, Ems, and Büsum Mole (52.12◦N, 8.86◦E). Loebl et al. (2009) extended this analysis to also cover Si and P. They also concluded that during most of the year light is limiting, followed by Si during most of the growth season for diatoms, P limitation during the summer, and N limitation toward the end of the growth season.

The decreasing trend in phytoplankton biomass could potentially be caused by an decrease in light availability. Available evidence, however, suggests increasing light availability in the Wadden Sea and North Sea. Philippart et al. (2013) analyzed 4 decades of Secchi depth reading and concluded no overall change. Long-term trends (their **Figure 7**) suggest a slight decrease in light availability in the western Dutch Wadden Sea between the mid-1970s until the mid-1990s followed by an increase. This is in line with de Jonge and de Jong (2002) describing a general increase in annual suspended matter concentrations in the Western Dutch Wadden Sea from the 1950s to the mid 1980s, then decreasing until the 2000s. de Jonge and de Jong (2002) related changes in suspended matter concentrations to dredging activities in the Rhine Estuary and Rhine discharge. In the German Bight, Gebuhr et al. (2009) described a general increase in Secchi depth from the mid 1970s. In general, light conditions in the coastal North Sea and Wadden Sea apparently improved rather than deteriorated, suggesting that light is not the dominant factor in the long-term phytoplankton decrease observed in most of the Wadden Sea. Thus, the observed correlation between summer chlorophyll a levels and riverine nutrient loads supports that nutrients are a dominant factor in the long-term decline of phytoplankton. The low correlation coefficients do indicate, however, that many other factors and their interactions modulate the long-term phytoplankton dynamics. For instance, better light conditions would also increase microphytobenthos activity, possibly leading to reduced nutrient fluxes from the sediment (e.g., Sundbäck et al., 2000).

As in most of the Wadden Sea phytoplankton biomass recently decreased, the increase at the station Danziggat in the Dutch Wadden Sea is striking. The increase can be due to an increased accumulation of phytoplankton, an increased growth potential or reduced loss factors in this part of the Wadden Sea. The decreasing trend in autumn NH<sup>4</sup> + NO<sup>2</sup> contrasts with the chlorophyll time series and suggests that increased nutrient availability was not responsible for the observed phytoplankton, but hints at improved growth condition e.g., due to increasing light conditions or decreasing grazing rates or filter rates. Suspended matter concentrations at Danziggat are higher than at the other western Dutch Wadden Sea stations (de Jonge and de Jong, 2002), suggesting stronger light limitation here than at the other Dutch long-term stations. Possibly, strong light limitation prevailed also during summer at this station during

the early 1980s. We hypothesize that improved light conditions since the 1980s (see section "Discussion") enabled enhanced phytoplankton production in this part of the Wadden Sea leading to increasing phytoplankton biomass (compare **Figure 6**).

Both riverine N and P loads have been used to explain longterm trends in phytoplankton biomass and primary production (e.g., de Jonge, 1990; van Beusekom et al., 2001). Our analyses show good correlations between summer chlorophyll a with both N and P. This indicates that N and P availability ultimately determine the eutrophication levels of the Wadden Sea. However, they do not provide conclusive indications whether N or P are ultimately limiting the phytoplankton dynamics.

Riverine P loads decreased faster than the N loads leading to high N/P ratios. Although no large differences were observed in the response of the phytoplankton biomass to decreasing loads of either riverine N or P, the changes of the N/P ratio may have implications on coastal communities. Malzahn et al. (2007) investigated food chain effects of N limited and P limited phytoplankton (Rhodomonas sp.) on Copepods (Acartia sp.) and fishlarvae (Clupea harengus). The P-limited food chain resulted in larval fish with a significantly poorer condition than N-limited or nutrient-sufficient food chains. Meunier et al. (2018) showed that long-term shifts in dissolved PO<sup>4</sup> concentration in the North Sea were closely linked to biomass trends of heterotrophic dinoflagellates and corroborated this experimentally. These results are interesting in the light of a study by Philippart et al. (2007) on the impact of nutrient reductions in the western Dutch Wadden Sea on phytoplankton, macrozoobenthos and estuarine birds. In their study, they used the IJsselmeer (see **Figure 1**) nutrient loads instead of the combined Rhine Maas loads used in the present study. They observed a weak correlation of P loads with the phytoplankton biomass but stronger correlations with macrozoobenthos and estuarine birds. Kerimoglu et al. (2018) studied the potential changes of phytoplankton in the North Sea under pre-industrial nutrient input scenario (about 30% of present levels) and suggested that the nutrient limitation led to disproportionately lower zooplankton biomass relative to the phytoplankton biomass (expressed as C content). This implied reduced grazing pressure, contributing in turn to a limited sensitivity of phytoplankton, and in certain regions, even to a slight phytoplankton increase under those lower nutrient input rates. Although we did not find support whether N or P was the ultimate limiting element for phytoplankton biomass formation, the impact of increasing N/P ratios on food quality definitely needs further attention.

Both proxies – summer phytoplankton and autumn NH<sup>4</sup> + NO<sup>2</sup> – are correlated, but with a large spread (r <sup>2</sup> = 0.26). This correlation does not necessarily reflect a causal relation between

the two proxies, but rather is the result of a common underlying driver: nutrient availability. Both proxies are impacted by different factors: Phytoplankton is positively influenced by light and nutrient availability but negatively by grazing. Autumn NH<sup>4</sup> + NO<sup>2</sup> is positively influenced by organic matter accumulation, but negatively by phytoplankton/microphytobenthos nutrient uptake and nitrification rates and high exchange with the North Sea during autumn. Only enhanced organic matter/nutrient availability is a common factor increasing both proxies. Other factors increasing the phytoplankton growth potential like better light conditions or decreasing grazing rates would lead to higher summer chlorophyll a levels but lower autumn levels of NH<sup>4</sup> + NO<sup>2</sup> due to increased phytoplankton nutrient uptake during autumn. Therefore, similar long-term trends of both proxies but no high correlation coefficient between the proxies is expected.

Based on the above discussion we suggest, that due to decreased riverine nutrient loads both summer chlorophyll a – as an indicator of the phytoplankton biomass and phytoplankton growth potential – and the autumn concentrations of NH<sup>4</sup> + NO<sup>2</sup> as an indicator of organic matter turnover intensity have declined over the entire Wadden Sea. Both proxies independently identify the southern Wadden Sea as having a higher eutrophication level than the northern Wadden Sea. Model data (**Figure 11**) support that import of organic matter from the North Sea to the Wadden Sea is larger along the southern Wadden Sea than along the northern Wadden Sea. An analysis of satellite data and in situ suspended matter data (Schartau et al., 2019) supports the model results by showing sharper gradients along the southern Wadden Sea than along the northern Wadden Sea. However, further model analyses including a quantification of transport rates based on multidecadal runs performed on higher resolution setups (see e.g., Gräwe et al., 2016; Androsov et al., 2019; Stanev et al., 2019) are needed to better understand the relative contribution of estuarine circulation and direct river inputs on the maintenance of horizontal gradients within the Wadden Sea.

## Regional Differences in Eutrophication Indicators

Both model data and the eutrophication proxies identify the southern Wadden Sea as having a higher eutrophication level than the northern Wadden Sea. We will discuss, whether these patterns are also observed in other ecosystem components that respond to nutrient-enrichment like green macroalgae, seagrass, or benthic filter feeders.

Green macroalgae blooms can be due to coastal eutrophication with consequences for local economies, tourism, and the ecological state of coastal ecosystems (Fletcher, 1996; Smetacek and Zingone, 2013). In the Wadden Sea, green macroalgae blooms increased from the 1970s onward (Reise and Siebert, 1994; Kolbe et al., 1995; Schories et al., 1997), reaching a maximum during the 1990s (Kolbe et al., 1995; van Beusekom et al., 2017). Since the 1990s, a general decline to very low levels is observed for the northern Wadden Sea, but the decline in the German part of the southern Wadden Sea is less clear and a high growth potential is still present (van Beusekom et al., 2017). The decline in the northern Wadden Sea and remaining blooms in the southern Wadden Sea are in line with a higher eutrophication level in the southern Wadden Sea.

Seagrass is an important component of coastal ecosystems (e.g., Costanza et al., 1997; Orth et al., 2006), but globally declining at an increased rate (Waycott et al., 2009). Many factors have contributed to the seagrass decline in the entire Wadden Sea during the 20th century, including Wasting Disease, habitat destruction through embankment, increased turbidity, increased storminess, and eutrophication (e.g., Den Hartog, 1987; de Jonge and de Jong, 1992; Kastler and Michaelis, 1997; Dolch et al., 2013). Since the late 1990s, seagrass is clearly recovering in the northern Wadden Sea (Dolch et al., 2013) but not in the southern Wadden Sea (Dolch et al., 2017), where present seagrass distribution is less than expected on the basis of habitat preferences (Folmer et al., 2016). Although multiple stressors are involved, eutrophication probably has played a major role in seagrass decline (e.g., Bittick et al., 2018). Indeed, the recovery of Chesapeake Bay submerged vegetation is directly related to reduction in nutrient loads (Lefcheck et al., 2018). We hypothesize, that if riverine nutrient loads further decrease, seagrass occurrence in the Southern Wadden Sea will increase.

Recent distribution patterns of mussel beds also show regional differences, with the southern Wadden Sea having highest shares of tidal basin surface occupied by mussel beds (Folmer et al., 2014). The authors stress that many factors are involved, but ruled out temperature and extreme wind effects. Food availability is one of the factors involved and would be in line with the regional difference in organic matter and phytoplankton availability shown in this paper. In this context, it is important to note that macrozoobenthos biomass off the German southern Wadden Sea showed a clear decline, that was linked to a decrease in eutrophication (Meyer et al., 2018).

To summarize, more opportunistic macroalgae, a higher share of mussel beds and less seagrass in the southern Wadden Sea as compared to the northern Wadden Sea are all in agreement with a higher nutrient and organic matter availability in the southern Wadden Sea.

Whereas the overall regional distribution patterns of the eutrophication proxies are in agreement with other ecosystem components that respond to nutrient-enrichment, it should be borne in mind that part of the regional differences can be due to timing of sampling with regard to the tidal phase. Grunwald et al. (2010) reported higher nutrient concentrations at low tide than at high tide. Dutch monitoring stations are sampled around high tide, whereas the other stations [Grådyb (DK) and List tidal basin(D)] are sampled independently of the tidal phase. Since 2005, the Norderney (D) time series were sampled both during low and during high tide. Whereas differences for summer chlorophyll a were small (∼6%), autumn NH<sup>4</sup> + NO<sup>2</sup> concentrations at high tide were about 22% (2–3 µmol/l) lower than at low tide. The overall effect would be that the Dutch eutrophication levels are slightly underestimated as compared to the other stations. The Norderney temporal trend is probably slightly underestimated due to slightly higher autumn values during recent years by including low tide values since 2005.

### CONCLUSION

Long time series are pivotal for understanding how ecosystems respond to both man-made and natural changes. The long history of Wadden Sea eutrophication is described best by the work done at the Royal Netherlands Institute for Sea Research (NIOZ) with a focus on the western Dutch Wadden Sea. Available data suggest that eutrophication already started from the 1950s onward (e.g., de Jonge, 1990), culminating in the 1980s (e.g., Cadée and Hegeman, 2002). Our study extends the description of Wadden Sea eutrophication to the entire Wadden Sea. The longest time series on phytoplankton/chlorophyll a are available for the Dutch Wadden Sea. They document an increase in phytoplankton from the mid 1970s to the mid 1980s and a decrease since. The Norderney and Sylt time series starting in the 1980s document the decreasing trends, whereas the Danish time series starting at the end of the 1980s only weakly reflect the decreasing eutrophication levels. Riverine nutrient loads are important drivers of phytoplankton dynamics and correlated significantly with summer phytoplankton biomass (as chlorophyll a), but only explained about 30% of interannual variability, suggesting the importance of other factors. These may include both biological factors like the balance between grazing and growth, but also physical factors like the import of organic matter (as a source of nutrients) from the North Sea.

Given the high variability of phytoplankton in response to riverine nutrient loads, it is important to have additional proxies that indicate the intensity of organic matter turnover. We show that the mean autumn concentration of the remineralization products NH<sup>4</sup> + NO<sup>2</sup> in autumn (when decreasing light levels limit phytoplankton productivity) is a useful proxy. Especially in the southern Wadden Sea, this proxy is significantly linked to riverine nitrogen loads suggesting its driving role in Wadden Sea organic matter turnover. Both proxies – summer chlorophyll a and autumn NH<sup>4</sup> + NO2<sup>−</sup> are significantly correlated. The best correlations between the eutrophication proxies and riverine nutrient loads are found for stations in tidal inlets, probably because of less stronger interactions with the sediment than at the shallower, more coastward stations.

Regional differences are apparent, with long-term summer chlorophyll levels and autumn NH<sup>4</sup> + NO2<sup>−</sup> levels clearly higher in the southern Wadden Sea than in the northern Wadden Sea. The model by Kerimoglu et al. (2017) supports, that these differences are driven by a higher organic matter import from the North Sea into the Wadden Sea. These regional differences

suggest a lower eutrophication potential in the northern Wadden Sea as evidenced by lower phytoplankton biomass, lower organic matter turnover (autumn release of NH<sup>4</sup> and NO2), seagrass recovery and decreasing green algae blooms.

The present data show, that measures taken to reduce riverine nutrient loads have led to a decrease in phytoplankton biomass in the Wadden Sea and adjacent coastal zone, and contributed to an improved ecological status. However, long, well designed long time-series with a high temporal resolution are needed to derive statistically significant correlations due to complex physical and ecological interactions, resulting in a strong interannual variability. Given the world-wide increase in the use of fertilizers (Battye et al., 2017), it can be expected that eutrophication problems will increase world-wide. The present example of the Wadden Sea underlines that measures to reduce nutrient enrichment do lead to an improvement of the coastal ecological status.

### REFERENCES


## AUTHOR CONTRIBUTIONS

This article is an extended version of a contribution to the Wadden Sea Quality Status Report on Eutrophication lead by JvB with contributions from JC, TD, AG, HL, KK, JP, HR, and JR. JvB conceived the study, analyzed the data, and wrote the first draft of the manuscript. JC, AG, LR, HR, and JR provided the chlorophyll and nutrient data. TD and KK specifically contributed to the discussion of seagrass and macroalgae. HL and JP provided the riverine input data. RH and OK contributed with the analysis of the model data. All authors contributed to the discussion of the manuscript.

## ACKNOWLEDGMENTS

We thank the three reviewers for their constructive comments that significantly improved the manuscript.




suspended diatoms in the Dutch Wadden Sea. Neth. J. Sea Res. 8, 174–207. doi: 10.1016/0077-7579(74)90016-7


**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2019 van Beusekom, Carstensen, Dolch, Grage, Hofmeister, Lenhart, Kerimoglu, Kolbe, Pätsch, Rick, Rönn and Ruiter. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

# Chesapeake Bay Dissolved Oxygen Criterion Attainment Deficit: Three Decades of Temporal and Spatial Patterns

Qian Zhang<sup>1</sup> \*, Peter J. Tango<sup>2</sup> , Rebecca R. Murphy<sup>1</sup> , Melinda K. Forsyth<sup>3</sup> , Richard Tian<sup>1</sup> , Jennifer Keisman<sup>4</sup> and Emily M. Trentacoste<sup>5</sup>

<sup>1</sup> Chesapeake Bay Program Office, University of Maryland Center for Environmental Science, Annapolis, MD, United States, <sup>2</sup> Chesapeake Bay Program Office, U.S. Geological Survey, Annapolis, MD, United States, <sup>3</sup> Chesapeake Biological

Laboratory, University of Maryland Center for Environmental Science, Solomons, MD, United States,

<sup>4</sup> Maryland-Delaware-District of Columbia Water Science Center, U.S. Geological Survey, Catonsville, MD, United States,

<sup>5</sup> Chesapeake Bay Program Office, U.S. Environmental Protection Agency, Annapolis, MD, United States

#### Edited by:

Jacob Carstensen, Aarhus University, Denmark

#### Reviewed by:

Jens Würgler Hansen, Aarhus University, Denmark Akkur Vasudevan Raman, Andhra University, India

\*Correspondence: Qian Zhang qzhang@chesapeakebay.net

#### Specialty section:

This article was submitted to Marine Ecosystem Ecology, a section of the journal Frontiers in Marine Science

Received: 30 July 2018 Accepted: 23 October 2018 Published: 21 November 2018

#### Citation:

Zhang Q, Tango PJ, Murphy RR, Forsyth MK, Tian R, Keisman J and Trentacoste EM (2018) Chesapeake Bay Dissolved Oxygen Criterion Attainment Deficit: Three Decades of Temporal and Spatial Patterns. Front. Mar. Sci. 5:422. doi: 10.3389/fmars.2018.00422 Low dissolved oxygen (DO) conditions are a recurring issue in waters of Chesapeake Bay, with detrimental effects on aquatic living resources. The Chesapeake Bay Program partnership has developed criteria guidance supporting the definition of state water quality standards and associated assessment procedures for DO and other parameters, which provides a binary classification of attainment or impairment. Evaluating time series of these two outcomes alone, however, provides limited information on water quality change over time or space. Here we introduce an extension of the existing Chesapeake Bay water quality criterion assessment framework to quantify the amount of impairment shown by space-time exceedance of DO criterion ("attainment deficit") for a specific tidal management unit (i.e., segment). We demonstrate the usefulness of this extended framework by applying it to Bay segments for each 3-year assessment period between 1985 and 2016. In general, the attainment deficit for the most recent period assessed (i.e., 2014–2016) is considerably worse for deep channel (DC; n = 10) segments than open water (OW; n = 92) and deep water (DW; n = 18) segments. Most subgroups – classified by designated uses, salinity zones, or tidal systems – show better (or similar) attainment status in 2014–2016 than their initial status (1985–1987). Some significant temporal trends (p < 0.1) were detected, presenting evidence on the recovery for portions of Chesapeake Bay with respect to DO criterion attainment. Significant, improving trends were observed in seven OW segments, four DW segments, and one DC segment over the 30 3-year assessment periods (1985–2016). Likewise, significant, improving trends were observed in 15 OW, five DW, and four DC segments over the recent 15 assessment periods (2000–2016). Subgroups showed mixed trends, with the Patuxent, Nanticoke, and Choptank Rivers experiencing significant, improving short-term (2000–2016) trends while Elizabeth experiencing a significant, degrading short-term trend. The general lack of significantly improving trends across the Bay

**342**

suggests that further actions will be necessary to achieve full attainment of DO criterion. Insights revealed in this work are critical for understanding the dynamics of the Bay ecosystem and for further assessing the effectiveness of management initiatives aimed toward Bay restoration.

Keywords: water quality standards, dissolved oxygen, criteria attainment, monitoring and assessment, Chesapeake Bay, Mann-Kendall test, ecosystem management, spatial aggregation

### INTRODUCTION

fmars-05-00422 November 19, 2018 Time: 14:41 # 2

Chesapeake Bay, the largest estuary in the United States, is an incredibly complex and productive ecosystem that provides habitats, food, and protection for thousands of species of animals and plants (**Figure 1**). This national treasure, however, has suffered cultural eutrophication for decades, largely due to anthropogenic inputs of nutrient and sediment from its multi-jurisdiction watershed (Boynton et al., 1995; Kemp et al., 2005; Hirsch et al., 2010; Zhang et al., 2013, 2015; Zhang and Blomquist, 2018). Consequently, Chesapeake Bay ("the Bay") has shown ecological degradation with symptoms such as excessive algal growth, decreased submerged aquatic vegetation acreage, reduced water clarity, and low dissolved oxygen (DO) concentrations (Hagy et al., 2004; Kemp et al., 2005; Murphy et al., 2011; Testa et al., 2017, 2018; Lefcheck et al., 2018).

To support healthy and sustainable living resources in the Bay, the Chesapeake Bay Program (CBP) partnership – which consists of the U.S. Environmental Protection Agency (USEPA), other federal agencies, local and state jurisdictions, and academic and non-governmental organizations – has been committed to the protection of water quality and habitat conditions in the Bay and its tidal tributaries. In 2003, the CBP partnership put forth a guidance framework to establish water quality criteria for DO, water clarity, and chlorophyll-a for the Bay (U.S. Environmental Protection Agency [USEPA], 2003a), which were subsequently adopted into the tidal states' water quality standards to define which waters are impaired under the Clean Water Act (**Supplementary Table S1** in **Appendix A**). In addition, this guidance framework (U.S. Environmental Protection Agency [USEPA], 2003a) has set the foundation for criteria attainment assessment procedures, which have been periodically refined as new knowledge has become available (U.S. Environmental Protection Agency [USEPA], 2003a, 2004a, 2007a,b, 2008, 2010a, 2017).

Water quality criteria are applied for five different designated uses (DUs) of aquatic habitats, namely, open water (OW), deep water (DW), deep channel (DC), migratory spawning and nursery (MSN), and shallow water (SW). These DUs reflect the nature of water column structure and the life history needs of living resources, which vary seasonally (**Figure 2** and **Supplementary Table S1**; U.S. Environmental Protection Agency [USEPA], 2003b, 2004b). In particular, the OW criterion protects diverse populations of sport fish, including striped bass, bluefish, mackerel and sea trout, as well as important bait fish such as menhaden and silversides. The DW criterion protects animals inhabiting the deeper transitional water-column and bottom habitats between the well-mixed surface waters and the deep channels, including many bottom-feeding fish, crabs, and oysters. The DC criterion protects bottom sediment dwelling worms and small clams that bottom-feeding fish and crabs consume. The MSN criterion protects migratory and resident tidal freshwater fish during the spawning and nursery season in low-salinity habitats. Lastly, the SW protects the many species that depend on vegetated shallow-water habitats. SW is part of the OW and uses the same DO criterion as the OW, although it has separate criteria on submerged aquatic vegetation/water clarity.

We recently published results for the Chesapeake Bay water quality standards attainment indicator (Zhang et al., 2018), which aggregates the estimated condition of all 92 Chesapeake Bay management segments (**Figure 1**) for DO, submerged aquatic vegetation/water clarity, and chlorophyll-a criteria that are evaluated for addressing the goal of meeting the requirements of the Chesapeake Bay Total Maximum Daily Load (U.S. Environmental Protection Agency [USEPA], 2010b). Our current work expands upon that effort by delving into the attainment results of each individual segment. Specifically, we extend the utility of the existing assessment framework beyond the binary pass/fail classification to quantify the actual amount of space-time criterion exceedance, which we call "attainment deficit." This was motivated by our observations that segments may have drastically different status and trends in the extent of their impairment while showing no state change with respect to attainment status. Thus, tracking spatial and temporal patterns of the attainment deficit has the potential to reveal further information on water quality dynamics, as compared with our prior effort employing a binary pass/fail classification.

In this work, we demonstrate the usefulness of this extended framework by applying it to all applicable Bay segments for each of the 30 3-year periods between 1985 and 2016 (i.e., 1985–1987, 1986–1988 . . . 2014–2016). This comprehensive assessment of DO criterion attainment include (1) a synthesis of DO criterion attainment deficit for three DUs (i.e., OW, DW, and DC) for the 92 segments listed in the Chesapeake Bay Total Maximum Daily Load document (the MSN DU is excluded due to data insufficiency; the SW DU is also excluded because it is part of the OW DU with respect to DO); and (2) a synthesis of DO criterion attainment deficit for aggregated subgroups in the Chesapeake Bay ecosystem. Subgroups are defined as the aggregation of all segments that belong to a specific DU (n = 3), salinity zone (n = 4), or tidal system (n = 13). These results provide essential information

[USEPA], 2004a,b).

to the Bay management and research community for (1) understanding the conditions and dynamics of the Chesapeake Bay ecosystem and (2) further assessing the effectiveness of management initiatives aimed toward Bay restoration under the influences of climatic and hydrological variability. This work also features Chesapeake Bay as a prime example where long-term monitoring network and science-based criterion assessment methods can be combined to evaluate the status and trends of complex ecosystems, which might be relevant to other coastal and inland ecosystems that are facing ecological degradation (Borja et al., 2008; Bricker et al., 2008; Patrício et al., 2016; Schiff et al., 2016; Sherwood et al., 2016; Trowbridge et al., 2016).

### THE CRITERION ASSESSMENT FRAMEWORK

#### The Existing Framework

The existing Chesapeake Bay water quality criteria attainment assessment framework is centered on the development of cumulative frequency distribution (CFD) curves that allow

for the evaluation of criteria exceedance (U.S. Environmental Protection Agency [USEPA], 2003a; Batiuk et al., 2009; Tango and Batiuk, 2013). As illustrated in **Figure 3A**, this assessment framework involves two key components, namely, "Assessment Analysis of Monitoring Data" and "Compliance Decision Framework."

For the "Assessment Analysis of Monitoring Data," the framework requires the collection of tidal monitoring data, including DO concentrations, water temperature, and salinity. These data are interpolated using the CBP's spatial-interpolation software (or "CBP interpolator") for each spatial unit (U.S. Environmental Protection Agency [USEPA], 2003a). The spatial units are defined by the intersection of Bay segments (**Figure 1**) and tidal-water DUs (**Figure 2**). In this regard, water temperature and salinity observations are used to compute the vertical density structure of the water column and delineate boundaries between the OW, DW, and DC layers, which can vary temporally due to freshwater inputs, tides, and other physical conditions. For each spatial unit, DO concentration data are horizontally and vertically interpolated and then compared with appropriate season-specific criterion values (**Supplementary Table S1**) to quantify the spatial extent of criteria exceedance for each sampling event. For each spatial unit, the estimated spatial exceedance for each sampling event is ranked from the lowest to the highest to construct a CFD curve (also called "attainment curve"), the area below which represents the

cumulative amount of space and time in which the criterion value is exceeded.

For the "Compliance Decision Framework," reference curves have been developed by the CBP Partnership to provide a scientifically based, direct measure of the allowable criteria exceedance, i.e., the amount of criteria exceedance that can occur without causing significant ecological degradation. Readers are referred to U.S. Environmental Protection Agency [USEPA] (2003a) for more details. In the non-compliance space-time assessment space, the reference curve defines the boundary of compliance and impairment. Specifically, the area below the reference curve represents the allowable criteria exceedance (U.S. Environmental Protection Agency [USEPA], 2003a; Batiuk et al., 2009).

The CFD (attainment) curve is compared with the reference curve to determine the status of the spatial unit with respect to criterion attainment. If the CFD curve is not entirely below the reference curve, then the spatial unit is considered "not attaining" the DO criterion.

#### The Analytical Extension

Here we introduce an analytical extension to the existing assessment framework (**Figure 3B**). This extension allows for further exploration of the CFD curve to quantify attainment deficit in a spatial unit, i.e., the intersection area between the attainment curve and the reference curve. This intersection area, also termed the "non-allowable criteria exceedance," is scaled by the total area of the assessment space to convert to a value in the range of 0 and 100%, which is then converted to attainment deficit by adding a minus sign. In other words, a criteria exceedance of 0% corresponds to an attainment deficit of 0%, whereas a criteria exceedance of 100% corresponds to an attainment deficit of −100%.

Attainment deficit is always in the range of 0 and −100%; see three representative examples in **Figure 3B**. An attainment deficit of 0%, which is the best possible condition, implies that the minimum water quality requirements are met for providing protection to aquatic life in the defined zones. An attainment deficit of −100%, which is the worst possible condition, implies complete non-compliance. Any other values also indicate non-compliance, with values closer to −100% implying more severe conditions that have substantial negative effects on living resources' survival, growth, and reproduction.

One major benefit of quantifying attainment deficit is to enhance our analytical capability to detect temporal changes. Many segment-DUs may not have experienced a state change using the binary pass/fail attainment classification, but they may have experienced drastically different trends in the extent of their non-compliance (or attainment deficit). This is illustrated in **Figure 4** with three simplified trajectories, which show an improving condition (i.e., declining attainment deficit), a stable condition (i.e., no significant change in attainment deficit),

and a degrading condition (i.e., increasing attainment deficit), respectively. This evolution of the extent of attainment deficit is further illustrated by the intersection area between the attainment curve and the reference curve for three timesteps. These examples clearly demonstrate the utility of attainment deficit derived from the extended assessment framework (**Figure 3B**). By contrast, under the binary pass/fail approach, these three cases would be considered equal in terms of status and trends. In other words, they are always out of attainment and they all have a zero trend over time.

## APPLICATION OF THE EXTENDED FRAMEWORK

#### Monitoring Data

Tidal monitoring data of DO, salinity, and temperature were obtained from the CBP Water Quality Database for the period between 1985 and 2016 (Chesapeake Bay Program, 2017). These data were collected by the Maryland (MD) Department of Natural Resources, the Virginia (VA) Department of Environmental Quality, and partners at more than 140 stations distributed across the Bay's middle channel, tidal tributaries, and embayments. Most of these stations have been sampled consistently since 1985, at a frequency of 12–20 times per year with limited additional synoptic sampling (U.S. Environmental Protection Agency [USEPA], 2010b; Tango and Batiuk, 2013). The sampling was done using consistent sampling and analysis protocols and complemented by a rigorous quality assurance program (U.S. Environmental Protection Agency [USEPA], 2010b; Tango and Batiuk, 2013). Most of the 92 segments contain 1–3 long-term monitoring stations and some segments contain additional stations from supplemental monitoring programs such as shallow water monitoring and citizen volunteer monitoring.

#### Attainment Deficit

The extended assessment framework was applied to the Chesapeake Bay segments (**Figure 1**) for three DO-related DUs, i.e., OW, DW, and DC (**Figure 2**), which resulted in estimates of attainment deficit for each applicable segment and DU for each running 3-year assessment period from 1985–1987 to 2014– 2016. For this work, we focused on summer results (June– September). As previously described, estimated attainment deficit falls between 0% (i.e., all space and time are in attainment for the assessment period) and −100% (i.e., all space and time are out of attainment for the assessment period).

The segment-level estimates of attainment deficit were further aggregated for each 3-year period to investigate the status and trends with different types of subgrouping. These subgroups include three different DUs (i.e., OW, DW, and DC), four salinity zones [i.e., tidal fresh (TF), oligohaline (OH), mesohaline (MH), and polyhaline (PH)], and thirteen tidal systems. For each subgroup, all applicable segments were selected and their attainment deficit values in each assessment period were averaged through surface-area weighting:

$$AD\_{subgroup\ J} = \frac{\sum\_{j}^{all\ segments} \in \,^{J}AD\_{j}^{\*}A\_{j}}{\sum\_{j}^{all\ segments} \in \,^{J}A\_{j}} \tag{1}$$

where AD<sup>j</sup> is the estimated attainment deficit value and A<sup>j</sup> is segment surface area for segment j within subgroup J. This weighting scheme was adopted for two reasons: (a) segments vary in size over four orders of magnitude (0.13–1,521 km<sup>2</sup> ; sum = 11,600 km<sup>2</sup> ) – see **Figure 1**, and (b) surface area

of each segment does not change with time or DU, unlike seasonally variable bottom water area or water volume. For certain segments in a 3-year period, monitoring data might not be available to produce attainment deficit values; those segments were excluded from the summation operations in Equation (1) to minimize bias in the aggregated result of AD for that period and correspondingly, estimated trends in AD.

## Trend Analysis

Trend analysis was conducted on the estimated attainment deficit values to determine whether DO conditions have improved over time. To do this, we used a modified version of the Mann-Kendall (MK) test that can account for autocorrelation in the time series (Hamed and Rao, 1998). This non-parametric test was chosen because the attainment deficit time series is not expected to follow any specific distribution and the values are bounded between −100 and 0%. An autocorrelation correction was needed because the assessment was conducted on monitoring data in running 3-year periods. The Sen slope was computed as well to generate an estimate of change over time (Sen, 1968). The modified Mann-Kendall and the Sen slope tests were implemented through the "mkTrend" function in the R-package "fume" (Santander Meteorology Group, 2012) to calculate the significance and slope for both a long-term trend (1985–2016) and a short-term trend (2000–2016). Following Hirsch et al. (2015), the significance level of a MK trend was not restricted to 0.05 to enhance the chance of detecting appreciable changes that are worthy of management considerations. Multiple alpha levels were considered, i.e., 0.05, 0.1, and 0.25. In addition, change-point analysis was conducted to test for a shift in the central tendency of the attainment deficit time series. The non-parametric Pettitt test was adopted (Pettitt, 1979), which was implemented using the "pettitt.test" function in the R-package "trend" (Pohlert, 2018).

## Data Availability

For the convenience of readers and end users, our results of attainment deficit are provided in the online **Supplementary Material**, including:


## RESULTS AND DISCUSSION

#### Current Status (2014–2016) of Chesapeake Bay DO Attainment Deficit Segment Patterns

The most recent (i.e., the 2014–2016 assessment period, hereafter referred to as "current") status of attainment deficit for each Chesapeake Bay segment is presented in **Figure 5**. This result and elaborations below highlight the usefulness of the attainment

deficit quantification for identifying places where patterns are different and where further evaluations are needed. Overall, there is a clear progression among the three DUs – i.e., in general, attainment status gets worse with depth as the DU goes from OW and DW to DC, which is consistent with the expectation that bottom water habitats of the tidal waters are not as healthy as surface areas in terms of DO conditions.

For OW (**Figure 5A**), 89 of the 92 applicable segments had data in the 2014–2016 assessment period. More than half of these segments (n = 48) were in full attainment in this period, including segments in the mainstem Bay and many tributaries. The status of attainment deficit was better than −4.8% for 75% of the applicable OW segments and better than −20.8% for 90% of the applicable OW segments. Overall, OW segments were dominated by zero or relatively small attainment deficit values in 2014–2016.

For DW (**Figure 5B**), all the 18 applicable segments had data in the 2014–2016 period. One third of these segments (n = 6) were in full attainment in this period, including segments in the polyhaline region of the Bay's mainstem. The status of attainment deficit was better than −3.6% for 75% of the applicable DW segments and better than −9.7% for 90% of the applicable DW segments. The largest deficit (−28.4%) was observed within segment MAGMH (Magothy River), an upper western shore tributary in MD. Notably, the mainstem segment CB4MH (Middle Central Bay) had the second largest attainment deficit (−14.6%) among all the DW segments in this assessment period. Like OW segments, DW segments were dominated by zero or minimal attainment deficit values in 2014–2016.

For DC (**Figure 5C**), 9 of the 10 applicable segments had data in the 2014–2016 period. Only one segment was in full attainment in this period – i.e., CB5MH\_VA (Lower Central Bay, VA). The status of attainment deficit was better than −15.4% for 75% of the applicable DC segments and better than −22.8% for 90% of the applicable DC segments. The largest deficit was observed with segment CB4MH, which was −40.5%. This is not surprising, since CB4MH is the region of the Bay where annual summer hypoxia develops first and lasts the longest (Testa and Kemp, 2014; Testa et al., 2018). Overall, DC segments had more occurrences of moderate or large attainment deficit in 2014–2016, as compared with OW and DW.

#### Subgroup Patterns

The segment-level attainment deficit was aggregated into different subgroups using Equation 1 based on the segments' DU, salinity zone, or tidal systems. The initial and most recent attainment deficit values calculated for each of these subgroups are provided in **Table 1** (For the complete time series, see

TABLE 1 | Estimated attainment deficit (initial and current status) and associated statistical results<sup>a</sup> for Chesapeake Bay dissolved oxygen criterion for the three designated uses, four salinity zones, and thirteen tidal systems.


<sup>a</sup>Significance levels are provided next to each estimate: p < 0.05 (∗∗∗), 0.05 < p < 0.1 (<sup>∗</sup> ), and p > 0.1 (−).

<sup>b</sup>Elizabeth does not have data in 1985–1987 or 1986–1988, so the earliest period with data (i.e., 1987–1989) was used to represent its initial deficit status.

**Supplementary Table S2**). For the three DUs, OW, DW, and DC had aggregated attainment deficit values of −0.8%, −3.2%, and −15.2%, respectively, in the 2014–2016 assessment period. This is consistent with the expectation that DC segments had generally poorer conditions than OW and DW segments.

For the four salinity zones, the 2014–2016 attainment deficit results exhibited the following ranking: PH (−0.2%) > TF (−1.4%) > OH (−1.6%) > MH (−6.0%). MH segments are generally subject to strong interactions between landward and seaward flows, which result in strong summer stratification that can prevent replenishment of oxygen from the water surface, exacerbating eutrophication effects. By contrast, TF and OH segments are generally more dominated by freshwater flow and hence less susceptible to stratification and more frequently replenished with DO-rich fresh waters. PH segments are closer to relatively DO-rich oceanic waters and tend to mix vertically in the late summer earlier than MH segments, resulting in their near-attainment status.

For the thirteen tidal systems, near-attainment status was achieved by Nanticoke (−0.3%), James (−0.3%), Choptank (−0.5%), and Tangier (−0.7%). Attainment deficit was better than −5% in Chester, Patuxent, Potomac, Rappahannock, Pocomoke Rivers, and the mainstem Bay. Attainment deficit was between −6 and −10% in upper mainstem Bay tributaries and York River. Elizabeth River is the only tidal system with a deficit worse than −10% in 2014–2016 (−22.6%).

How has the status of Chesapeake Bay's DO criterion attainment changed over time? For brevity, this question was addressed by aggregating individual segments into groups by designated use, by salinity zone, and by tidal system. Each subgroup's current aggregated attainment deficit (2014–2016) was then plotted against its initial attainment deficit (1985–1987) (**Figure 6**). The current status of each DU (**Figure 6A**) is better than its initial condition, with moderate improvements ranging between 0.8 and 2.8%. Similarly, the 2014–2016 condition of each salinity zone (**Figure 6B**) is better than its initial status, with moderate improvements ranging between 0.6 and 1.8%. The majority of tidal systems (**Figure 6C**) have better or similar current status compared to initial status. Notably, Patuxent River showed a substantial improvement of 8.4%. The Rappahannock, upper mainstem Bay tributaries, Potomac, Chester, and Elizabeth Rivers had moderate improvements in the range of 1.5∼3.6%. The mainstem Bay, Nanticoke, James, and Choptank rivers showed improvements of <1%. The York, Tangier, and Pocomoke systems were the only subgroups that showed degradation in DO attainment from 1985–1987 to 2014– 2016, but these differences were almost negligible (within <1%).

## Decadal Trends in Chesapeake Bay DO Attainment Deficit

#### Segment Patterns

The long-term (1985–2016) and short-term (2000–2016) trends in attainment deficit for Chesapeake Bay segments show strong spatial variations (**Figure 7**). The number of segments with improving and degrading trends are summarized in **Table 2**. Below, we elaborate on these trends and for brevity we focus on

trends with p < 0.1. These results highlight the effectiveness of using attainment deficit for identifying places that are associated with improving (or degrading) trends. Such information can help

guide targeting of management strategies and research to explain trend trajectories.

Among OW segments, seven had improving long-term trends and 15 segments showed improving short-term trends. However, only three segments showed consistently improving trajectories for both long-term and short-term trends, which are PAXMH (Lower Patuxent River), POCOH\_VA (Middle Pocomoke River, VA), and POTTF\_DC (Upper Potomac River, DC). The remaining four segments with long-term improving trends – CB6PH (Western Lower Bay), CB7PH (Eastern Lower Bay), SASOH (Sassafras River), and YRKPH (Lower York River) – showed no significant short-term trend. Of the remaining 12 segments with improving short-term OW trends, 10 showed no significant long-term OW trend. These included the lower portion of the


TABLE 2 | Summary of segments with improving and degrading trends in estimated attainment deficit for Chesapeake Bay dissolved oxygen criterion in the long-term period (1985–2016) and short-term period (2000–2016).

Choptank river (CHOMH1 and CHOMH2), the Corrotoman River (CRRMH), one tidal-fresh segment of the James River (JMSTF1), one oligohaline segment of the Potomac River (POTOH1\_MD), all but the lowest portion of the Nanticoke River (NANOH, NANTF\_DE, NANTF\_MD), as well as the mesohaline portions of the Rhode and West Rivers (RHDMH and WSTMH). Two segments (CHOOH and CHOTF; both in the Choptank river) with recent improving OW trends still had degrading long-term OW trend, indicating that in spite of recent improvements, conditions are still more degraded than they were in the mid-1980s. Eight of the remaining segments showed long-term degrading OW trends. Five of these segments, namely, CHSTF (Upper Chester River), PAXTF (Upper Patuxent River), POTMH\_VA (Lower Potomac River, VA), POTTF\_VA (Upper Potomac River, VA), and WICMH (Wicomico River), showed also degrading trends in the shortterm period. Moreover, additional six segments with no significant long-term OW trend showed recent degradation (ANATF\_DC, ANATF\_MD, BSHOH, EBEMH, PATMH, and WBRTF).

Among DW segments, four had improving long-term trends, namely, CB5MH\_MD (Lower Central Bay, MD), MAGMH (Magothy River), RPPMH (Lower Rappahannock River), and SOUMH (South River). For the short-term trend, improving conditions were also observed in these four segments, in addition to PAXMH (Lower Patuxent River). By contrast, degrading trends were associated with two segments (CB3MH, Upper Central Bay; CB5MH\_VA, Lower Central Bay, VA) for the long-term period and one segment (CB3MH) for the short-term period, both of which are located in the mainstem of the Bay.

Among DC segments, only one had improving long-term trend, i.e., CHSMH (Lower Chester River). For the short-term trend, three mainstem segments in addition to CHSMH showed improving conditions, namely, CB4MH (Middle Central Bay), CB5MH\_MD (Lower Central Bay, MD), and CB5MH\_VA (Lower Central Bay, VA). By contrast, degrading trends were associated with two segments (CB3MH; EASMH, Eastern Bay) for the longterm period and one segment (RPPMH; Lower Rappahannock River) for the short-term period.

Overall, the results show that many segment-DU pairs did not have significant, improving trends, suggesting that continued implementation of pollution management practices will be necessary to attain DO criterion. Further evaluation of the DO observations outside of the attainment assessment framework could very likely uncover additional trends, especially if the space-time exceedance of the DO criterion has changed in such a way that the overall attainment deficit has not changed (e.g., improvements in one part of the summer and not another). In addition, greater data resolution in space and time may provide more robust details of spatial conditions that can reduce uncertainty in assigning status and enhance the power to detect trends through time. Nonetheless, some significant trends were detected based on the metric of attainment deficit. Particularly included are some mainstem DC segments for the short-term period – i.e., CB4MH, CB5MH\_MD, and CB5MH\_VA, which are in the region of historically low summer DO (Hagy et al., 2004) and present promising evidence on the ecosystem recovery for portions of Chesapeake Bay with respect to DO criterion attainment.

#### Subgroup Patterns

fmars-05-00422 November 19, 2018 Time: 14:41 # 12

Time series of estimated attainment deficit for the subgroups are plotted in **Figure 8**. Trend results are summarized in **Table 1**. Among the three DUs (**Figures 8b–d**), only OW showed a statistically significant long-term trend with a slope of 0.04 percent/year. It was detected to have a change point at the 3-year period of 1994–1996, which is consistent with the previously identified shift in Chesapeake Bay water quality attainment indicator (Zhang et al., 2018). For DW and DC, neither the long-term nor short-term trend was statistically significant. However, their short-term trends were notable in magnitude – i.e., 0.13 and 0.24 percent/year, respectively. Also notable is the consistent and steady improvements in conditions since around 2009–2011 in OW and especially DW and DC.

Among the four salinity zones (**Figures 8e–h**), the TF zone had a negligible long-term trend and a positive short-term trend, but neither was statistically significant. MH trends behaved similarly to those in the TF zone in terms of slope and significance. By contrast, the OH zone had positive and statistically significant trends for both the long-term and short-term periods. More research is needed to test whether these improvements might be related to reductions of nutrient loads from tributaries or related to more short-term variations in hydrology. OH segments with improving trends are in the Choptank, Nanticoke, Pocomoke, Potomac, and Sassafras rivers (**Figure 1**). Finally, the PH zone had a positive and statistically significant long-term trend but a negligible short-term trend.

Among the 13 tidal systems (**Figures 8i–u**), only York had a statistically significant long-term trend, i.e., 0.15 percent/year, although it is one of the systems that showed degradation when just the 1985–1987 period was compared to the most recent period (**Figure 6C**). This disconnect appears to be due to a dip in the attainment deficit value in the last period (**Figure 8u**). An examination of the segment-level trends for York revealed that this long-term overall improvement was driven by an improvement in the OW attainment condition of the YRKPH segment (Lower York River), which has a long-term trend of 0.26 percent/yr (p < 0.1) (see **Appendix C**). More subgroups showed statistically significant short-term trends, including Choptank, James, Nanticoke, Patuxent, and Pocomoke (positive trends) and Elizabeth (negative trend), which can be attributed to specific segment-DU combinations shown in **Appendix C**. Of these tidal systems, the Patuxent had the largest short-term improvement – its aggregated attainment condition has improved with a slope of 0.98 percent/year over the short-term period. This pattern was driven by improvements in the PAXMH (mesohaline) and PAXOH (oligohaline) segments, although attainment conditions of the two tidal fresh segments (PAXTF and WBRTF) actually degraded. Another interesting case is the Nanticoke, where the short-term improvement was driven by rapidly improving conditions (p < 0.05) in the three OH and TF segments (i.e., NANOH, NANTF\_DE, and NANTF\_MD) but with no trends in the MH segment (NANMH). While attainment trends were different among salinity zones for the two systems above, the Choptank presents an example where the aggregated attainment condition represented improvements in segments distributed across all salinity zones, including CHOMH1, CHOMH2, CHOOH, and CHOTF. The Elizabeth presents a sharp contrast to the above tidal systems; attainment has degraded here in the last short-term period with a slope of −0.53 percent/year. This pattern was driven by downward trends in the OW attainment condition of EBEMH (Eastern Branch Elizabeth River), ELIPH (Mouth to mid-Elizabeth River), and WBEMH (Western Branch Elizabeth River), although only the EBEMH trend was statistically significant.

Overall, these subgroup trend results corroborate the segment trend results discussed above. Several significant, improving trends present promising evidence on the recovery for portions of Chesapeake Bay with respect to DO criterion attainment. However, these improvements are generally limited in magnitude (see **Table 1**). Overall, the general lack of significantly improving trends across the Bay over the long-term (1985–2016) and short-term (2000–2016) periods suggests that further actions will be necessary to achieve full attainment of DO criterion.

## CONCLUSION

We have introduced an analytical extension of the Chesapeake Bay water quality criterion assessment framework for quantifying the amount of space-time exceedance of DO criterion for a specific segment ("attainment deficit") and have demonstrated the usefulness of this framework by applying it to evaluate water-quality changes in the Chesapeake Bay ecosystem. With this approach, a comprehensive assessment of DO criterion attainment was conducted for Bay segments for each running 3-year period in 1985–2016. In general, the current status of attainment deficit (i.e., 2014–2016) is considerably worse for DC segments than OW and DW segments. Most subgroups show better (or similar) attainment status in 2014–2016 than their initial status (1985–1987). In terms of decadal trends, some significant trends (p < 0.1) were detected, presenting evidence on the recovery for portions of Chesapeake Bay with respect to DO criterion attainment. Over the 30 3-year periods in 1985–2016, significant, improving trends were observed in seven OW segments, four DW segments, and one DC segment. Over the recent 15 3-year periods (2000–2016), significant, improving trends were observed in 15 OW segments, five DW segments, and four DC segments. Subgroups showed mixed trends, with Patuxent, Nanticoke, and Choptank Rivers experiencing significant, improving short-term trends while Elizabeth experiencing a significant, degrading short-term trend. The general lack of significantly improving trends across the Bay suggests that further actions will be necessary to achieve full attainment of DO criterion. Overall, these attainment deficit results provided detailed information regarding the status and

trends of DO criterion attainment in Chesapeake Bay that can help target areas for further evaluation or refined management plans. Enhanced details for changes in habitat conditions are critical to the management and research community for understanding the conditions and dynamics of the Bay ecosystem and for further assessing the effectiveness of management initiatives aimed toward Bay restoration. More broadly, this work features Chesapeake Bay as an example where long-term monitoring data and science-based criterion assessment methods can be combined to evaluate complex ecosystems.

There are several directions for future research. First, the assessment can benefit from continued water quality monitoring as well as the promotion of new monitoring initiatives, such as volunteer monitoring and non-traditional partner contributions to increase station data densities as well as in situ, high resolution DO measurements. Second, the assessment approach is subject to limitations of data availability and key assumptions made to accommodate those limitations. Future work should incorporate new methods and further validate such types of assumptions to better understand short-term variability and evaluate the sensitivity of the results (particularly decadal trends) to such limitations. Third, new research should be done to tease apart the space and time aspects of the attainment deficit, so that improving or degrading trends can be more properly understood and communicated. Fourth, the segment-based attainment assessment results can be compared

with station-level DO trends. Researchers in the CBP partnership have implemented a generalized additive model (GAM) statistical approach to assess tidal-station trends. Comparison between the attainment deficit results and GAM trends to look for similarity (or dissimilarity) may provide new insights into the attainment deficit patterns as well as a deeper understanding of how and when water-quality improvements result in criteria attainment. Last but not least, clear, significant linkages between attainment status in the various segments and drivers, such as management actions (e.g., reduction of nutrient loads), internal hydrodynamic characteristics, trophic interactions, and climatic and hydrological variability, remain elusive. The relation of temporal and spatial patterns of these drivers (among others) to DO criteria attainment warrants further investigation.

## DATA AVAILABILITY

Water quality data used in this research are available through the Chesapeake Bay Program Water Quality Database – Chesapeake Information Management System Data Hub. All data supporting the conclusions of this manuscript will be made available by the authors, without undue reservation, to any qualified researcher. Most of these data are included in the manuscript and the **Supplementary Material**.

## AUTHOR CONTRIBUTIONS

QZ led the writing of the manuscript, conducted the statistical analyses, and produced the tables and figures in the manuscript. PT conceived the concept of attainment deficit and advised in the research design. RM contributed to the interpretation of statistical analyses. MF contributed to the early stage of the research. RT conducted the CFD analysis. JK advised in the

### REFERENCES


research design. ET produced the maps in the manuscript. All authors contributed to the interpretation of results as well as the writing and editing of the manuscript.

## FUNDING

This work was supported by the U.S. Environmental Protection Agency under grant "EPA/CBP Technical Support 2017" (No. 07-5-230480).

## ACKNOWLEDGMENTS

We would like to acknowledge Gary Shenk (USGS) and Elgin Perry (consultant) for helpful conversations in developing attainment deficit and guidance in selecting the trend methods applied here, respectively. We would also like to thank Scott Phillips (USGS), Tish Robertson (Virginia DEQ), and two journal reviewers for their comments on an early version of this manuscript. This work would not have been possible without the cumulative efforts that many individuals in the Chesapeake Bay Program partnership have contributed to this topic over the last two decades. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government. This is contribution no. 5526 of the University of Maryland Center for Environmental Science.

### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2018.00422/full#supplementary-material



**Conflict of Interest Statement:** The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2018 Zhang, Tango, Murphy, Forsyth, Tian, Keisman and Trentacoste. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.