# FUTURE OCEANS UNDER MULTIPLE STRESSORS: FROM GLOBAL CHANGE TO ANTHROPOGENIC IMPACT

EDITED BY : Erik Olsen, Isaac C. Kaplan, Cecilie Hansen Eide, Elizabeth A. Fulton, Michael J. Fogarty, Jamie C. Tam and Saskia Anna Otto 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 978-2-88966-308-8 DOI 10.3389/978-2-88966-308-8

#### 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

# FUTURE OCEANS UNDER MULTIPLE STRESSORS: FROM GLOBAL CHANGE TO ANTHROPOGENIC IMPACT

Topic Editors:

Erik Olsen, Norwegian Institute of Marine Research (IMR), Norway Isaac C. Kaplan, National Oceanic and Atmospheric Administration (NOAA), United States Cecilie Hansen Eide, Norwegian Institute of Marine Research (IMR), Norway Elizabeth A. Fulton, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia Michael J. Fogarty, National Marine Fisheries Service (NOAA), United States

Jamie C. Tam, Bedford Institute of Oceanography (BIO), Canada Saskia Anna Otto, University of Hamburg, Germany

Citation: Olsen, E., Kaplan, I. C., Eide, C. H., Fulton, E. A., Fogarty, M. J., Tam, J. C., Otto, S. A., eds. (2021). Future Oceans Under Multiple Stressors: From Global Change to Anthropogenic Impact. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88966-308-8

# Table of Contents

*05 Editorial: Future Oceans Under Multiple Stressors: From Global Change to Anthropogenic Impact*

Erik Olsen, Isaac C. Kaplan, Cecilie Hansen, Elizabeth Fulton, Michael J. Fogarty, Jamie C. Tam and Saskia A. Otto


Kunshan Gao, John Beardall, Donat-P. Häder, Jason M. Hall-Spencer, Guang Gao and David A. Hutchins

*70 An End-to-End Model Reveals Losers and Winners in a Warming Mediterranean Sea*

Fabien Moullec, Nicolas Barrier, Sabrine Drira, François Guilhaumon, Patrick Marsaleix, Samuel Somot, Caroline Ulses, Laure Velez and Yunne-Jai Shin

*89 Relative Impacts of Simultaneous Stressors on a Pelagic Marine Ecosystem*

Phoebe A. Woodworth-Jefcoats, Julia L. Blanchard and Jeffrey C. Drazen

*105 Arctic Sensitivity? Suitable Habitat for Benthic Taxa is Surprisingly Robust to Climate Change*

Paul E. Renaud, Phil Wallhead, Jonne Kotta, Maria Włodarska-Kowalczuk, Richard G. J. Bellerby, Merli Rätsep, Dag Slagstad and Piotr Kukliński

*119 Climate Change Vulnerability of American Lobster Fishing Communities in Atlantic Canada*

Blair J. W. Greenan, Nancy L. Shackell, Kiyomi Ferguson, Philip Greyson, Andrew Cogswell, David Brickman, Zeliang Wang, Adam Cook, Catherine E. Brennan and Vincent S. Saba


Sezgin Tunca, Martin Lindegren, Lars Ravn-Jonsen and Marko Lindroos

*155 Ecological Effects and Ecosystem Shifts Caused by Mass Mortality Events on Early Life Stages of Fish*

Erik Olsen, Cecilie Hansen, Ina Nilsen, Holly Perryman and Frode Vikebø

*168 Management Scenarios Under Climate Change – A Study of the Nordic and Barents Seas*

Cecilie Hansen, Richard D. M. Nash, Kenneth F. Drinkwater and Solfrid Sætre Hjøllo


Anne Babcock Hollowed, Kirstin Kari Holsman, Alan C. Haynie, Albert J. Hermann, Andre E. Punt, Kerim Aydin, James N. Ianelli, Stephen Kasperski, Wei Cheng, Amanda Faig, Kelly A. Kearney, Jonathan C. P. Reum, Paul Spencer, Ingrid Spies, William Stockhausen, Cody S. Szuwalski, George A. Whitehouse and Thomas K. Wilderbuer

*211 Climate Change and New Potential Spawning Sites for Northeast Arctic cod*

Anne Britt Sandø, Geir Odd Johansen, Asgeir Aglen, Jan Erik Stiansen and Angelika H. H. Renner


Ina Nilsen, Jeppe Kolding, Cecilie Hansen and Daniel Howell


# Editorial: Future Oceans Under Multiple Stressors: From Global Change to Anthropogenic Impact

Erik Olsen<sup>1</sup> \*, Isaac C. Kaplan<sup>2</sup> , Cecilie Hansen<sup>1</sup> , Elizabeth Fulton3,4, Michael J. Fogarty <sup>5</sup> , Jamie C. Tam<sup>6</sup> and Saskia A. Otto<sup>7</sup>

*<sup>1</sup> Norwegian Institute of Marine Research (IMR), Bergen, Norway, <sup>2</sup> Northwest Fisheries Science Center (NOAA), Seattle, WA, United States, <sup>3</sup> Center for Marine Socioecology, University of Tasmania, Hobart, TAS, Australia, <sup>4</sup> CSIRO Oceans & Atmosphere, Hobart, TAS, Australia, <sup>5</sup> Northeast Fisheries Science Center (NOAA), Woods Hole, MA, United States, <sup>6</sup> Bedford Institute of Oceanography, Dartmouth, NS, Canada, <sup>7</sup> Center for Earth System Research and Sustainability, Institute of Marine Ecosystem and Fishery Science, University of Hamburg, Hamburg, Germany*

Keywords: models and modeling, indicators, cumulative impacts, climate change, fisheries

**Editorial on the Research Topic**

#### **Future Oceans Under Multiple Stressors: From Global Change to Anthropogenic Impact**

## INTRODUCTION

If humanity is to achieve the ambitious targets of the UN Sustainable Development Goals (UN, 2015), we need assessments of future scenarios that evaluate combinations of natural and anthropogenic drivers that exert stress to the system as well as management actions. The current Research Topic explores futures for our oceans and coastal areas with a strong focus on effects of climate change, but also covering fishing, mass mortality events, and cumulative impacts from multiple stressors and human activities. It provides future visions for different timescales and regions, and what can be done to ameliorate negative impacts or outcomes.

The geographic scope covers regions from the coast and enclosed seas to the open oceans, and from the Arctic to Southern Ocean. Together these 20 articles paint a stark picture of the changes expected in our oceans but also present assessment methods, management paths, policies, and actions necessary to alleviate and deal with future problems.

## CLIMATE CHANGE AND FISHERIES

This Research Topic demonstrates that fisheries and climate still dominate thinking around ocean stressors. The relative impact of fishing vs. climate, and the extent to which improved fisheries management and conscious decisions about technological progress can ameliorate climate impacts (Galbraith et al., 2017) is likely to be of increasing importance in the future. For the pelagic Central North Pacific Ocean, Woodworth-Jefcoats et al. apply multi-species size-spectrum models to understand the impacts of climate scenarios on the ecosystem and fisheries. Here, climate change led to reductions in forage species and long-term declines in fisheries, though limits to fishing effort could partially offset these outcomes. Reum et al. apply a similar size-spectrum model to the Eastern Bering Sea (EBS). They noted long-term declines in ecosystem and fishery metrics. However, analysis of varying fishing rates suggested less scope for compensation or amelioration by fisheries management. Notably, the fisheries management scenarios for the EBS involved relatively small (and perhaps politically realistic) adjustments to status quo, while the central North Pacific fishery scenarios included large (50 and 80%) reductions in fishing effort. These modeling efforts are part of the Alaska Climate Integrated Modeling (ACLIM) project, which aims to analyze and

#### Edited and reviewed by:

*Susana Agusti, King Abdullah University of Science and Technology, Saudi Arabia*

> \*Correspondence: *Erik Olsen eriko@hi.no*

#### Specialty section:

*This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science*

> Received: *15 September 2020* Accepted: *29 September 2020* Published: *04 November 2020*

#### Citation:

*Olsen E, Kaplan IC, Hansen C, Fulton E, Fogarty MJ, Tam JC and Otto SA (2020) Editorial: Future Oceans Under Multiple Stressors: From Global Change to Anthropogenic Impact. Front. Mar. Sci. 7:606538. doi: 10.3389/fmars.2020.606538*

**5**

model the current and future climate driven changes to the EBS socio-economic system (Hollowed et al.).

Hansen et al. and Tunca et al. examine similar scenarios of varying fishing (pressure) under climate change for the Nordic and Baltic seas, respectively, using an end-to-end ecosystem model of the Nordic seas and a coupled bio-economic model for the Baltic. For both regions, ecosystem vulnerability to climate-change increased with increasing fishing, especially when expanding the fisheries to lower trophic levels (Hansen et al.), or for non-cooperative fisheries scenarios (Tunca et al.). In the Mediterranean, Moullec et al. use coupled modeling to project climate change scenarios on fish stocks and fisheries, showing an overall increase in fish biomass and catches but with large regional differences.

Downscaling of climate change effects has been shown to be important in evaluating species-specific effects. Sandø et al. projected a northward shift in spawning sites for Northeast Arctic cod, and Greenan et al. found overall positive effects on lobster habitat in the Gulf of Maine with increasing temperature, although the population is expected to shift toward the northeast, with associated socioeconomic impacts. In the Arctic, downscaled climate effects on benthic habitats were shown to be limited when evaluated for all taxonomic groups, but suitable habitat for 18% of the taxa studies were projected to change by more than 20%, suggesting serious ecological impact (Renaud et al.).

## CUMULATIVE IMPACTS AND MULTIPLE STRESSORS

Climate change effects need to be evaluated in conjunction with other pressures and drivers which exert stress and impair the functioning of the ecosystem (as discussed in Davies et al.). Assessing such multiple stressors and cumulative impacts is key to understanding our future oceans. Weijerman et al., illustrate one example in a model of the Mau Nui region of Hawaii. There, coral reef systems were influenced not only by fishing and climate (bleaching) but also by nutrification, sediment, and local wave action, and local management had a strong influence on performance in terms of ecosystem services. Using a statistical modeling framework, Otto et al. explore the long-term effects of cumulative impacts on a key zooplankton species in the Baltic, showing that multiple pressures were mostly additive, but that the population effects were dampened through density dependence. Consideration of cumulative effects suggests that while many multiple stressor interactions may be additive (Crain et al., 2008; Brown et al., 2013; Otto et al.), negative synergistic and positive dampening effects can lead to lower than expected outcomes (if we were assuming additive interactions). Beauchesne et al. investigate additive cumulative impacts from multiple stressors at an ecosystem level for the St. Lawrence system. They classified the cumulative impacts into six clusters depending on their relative levels within the climate, coastal, fisheries, and marine traffic drivers, with the highest exposure hotspots identified at the head of the Laurentian channel.

## TOPICAL EFFECTS OF CLIMATE CHANGE AND OTHER PRESSURES

Not all responses to climate change or other pressures and drivers are necessarily negative or linear, and topical studies exploring the effects on key trophic groups or key processes are in high demand. Andrew et al. provide vital insights into how iron availability is key for Southern Ocean phytoplankton to tolerate higher temperatures, while Gao et al. review the effects of ocean acidification (OA) on algae under multiple stressors. Responses can be highly variable, but for most calcifying algae, the combined effects of OA, UV, and increased temperatures reduce their growth. There is also need for topical knowledge of biological stressors, such as mass mortality events (MMEs) caused by pollution, natural disasters, or diseases. Olsen et al. look into the effect of MMEs on the Nordic seas using an end-toend ecosystem model, showing immediate and long-term direct and indirect effects, including potential hysteresis (Sguotti et al., 2019) that should be taken as caution when managing activities that can potentially cause MMEs. Fisheries also have nuanced effects on marine ecosystems, and detailed consideration of various management options is necessary to devise sustainable management strategies for the future. Balanced harvesting (BH) is one such alternative fisheries strategy. Nilsen et al. explore the effects of BH on the Nordic seas using an end-to-end ecosystem model, showing that for well-managed stocks, the effects were marginal, while for lower trophic level species, a BH strategy would have broadened the mix of species exploited and produce higher yields.

## INDICATORS AND ECOSYSTEM MODELS

The complexity of the socioecological marine systems and the cumulative impacts on these systems challenges us to synthesize our analytical results. Indicators of ecosystem state and the system's socio-economy (e.g., fisheries catches and revenues) have become standard in the EBM toolbox. Still, indicators need to capture trophic interactions, detect changes, and be consistent, transparent, comparable, and understandable. Designing indicators that meet these goals is difficult and should incorporate rigorous statistical testing comparable to what Kadin et al. carried out for trophic indicators of the Baltic ecosystem, where both thresholds and non-linear effects on indicators were evaluated. Measuring either the ecosystem or socioeconomic responses is not enough, however, as Fay et al. showed how ecosystem and fisheries indicators have different responses to the same fisheries management scenarios in the Northeast US. Due to the complexity and also utility of end-to-end models in EBM Tam et al. argue that indicators should be integrated into the modeling process, not estimated post-hoc as has been the typical approach.

## MANAGEMENT AND STAKEHOLDERS

Effective governance and management are necessities to deal with the challenges of the future oceans, but it has proven difficult to move from a fragmented sectoral system to an integrated management approach that addresses the complexities of cumulative impacts and multiple sectors in an adequate manner (Davies et al.). Creating a unified vision for comanagement and inclusion of indigenous rights and place-based understanding of the ecosystem knowledge are key factors to achieve holistic and effective cumulative-effects management of the Aotearoa area (New Zealand) and Great Barrier Reef Marine Park (Davies et al.). The importance of localized management is further supported by the spatial modeling of local management scenarios for the Maui Nui area in Hawai'i, where local sediment control was critical in slowing coral reef decline under climate change (Weijerman et al.). Due to existing management policies, socio-economic and political status the path to achieving an integrated management approach will look different for each country. However, the elements of robust science, public, and political support are essential elements.

## DISCUSSION

In the future, substantial (often negative) changes to our marine ecosystems are expected. However, the story is far from simple. Changes will not necessarily be negative for all species or habitats. Moreover, many of the combined effects are either dampened or synergistic, and may in some cases (Olsen et al.; Kadin et al.) show signs of hysteresis (Sguotti et al., 2019). Regionalized management options, tailor-made to the particular socioeconomic system, are overall to be recommended—be it in the design of indicators, downscaling of models or development of management and policy including local stakeholders and indigenous groups. In particular, a number of the papers in this Research Topic stress the importance of downscaling global model results (e.g., climate models) to address regional and local issues, ecosystem components, and drivers (Reum et al.; Hollowed et al.; Moullec et al.; Sandø et al.), as well as the diversity of responses (Andrew et al.). This means that while we may be forced to examine aggregated outcomes due to scarcity of data, where we can gain species-specific knowledge of system responses it will help identify the true differential outcomes across species (Renaud et al.).

The articles in this Research Topic advance and apply scenarios in various ways, ranging from relatively simple

## REFERENCES


manipulations of fishing rates (Hansen et al.; Fay et al.; Kadin et al.; Woodworth-Jefcoats et al.; and others in this issue), to development of complete policy-relevant scenarios for the Eastern Bering Sea and Hawaii (Hollowed et al.; Reum et al.; Weijerman et al.). Applying game theory models to the Baltic Sea multi-species fishing fleets, Tunca et al. illustrate the high degree to which performance of future management scenarios hinges on the level of cooperation among fishing nations. However, for the most part these consider only local or regional socio-economic and fishery drivers. Moving forward, there is a need to develop and customize the global or "broad-brush" scenarios such as the Shared Socioeconomic Pathways (SSPs) (Maury et al., 2017; O'Neill et al., 2017; Riahi et al., 2017) at a local or regional level. Application of a common set of global scenarios will also aid comparisons across ecosystems (Olsen et al., 2018; Lotze et al., 2019).

This Research Topic illustrates two transdisciplinary connections that must be made to better understand and model future ocean ecosystems. First, modeling of marine living resources should embrace cutting edge understanding regarding biogeochemical modeling and global drivers of primary production. Articles in this issue by Gao et al. and Andrew et al. illustrate the important roles of UV, iron, trace metals, stress or interactions, and responses that vary across species and broader taxa of primary producers. These responses underpin the assumptions and predictions necessary to forecast provisioning of marine ecosystem services for humanity. Second, envisioning our future oceans requires transdisciplinarity (Yates et al., 2015) including improved socio-ecological models and integrated ecosystem assessments (Holsman et al., 2017), together with better integration of stakeholders and indigenous groups (Davies et al.). This will be facilitated by transparent and easy access to data and information, for which integrated information sharing systems like eDrivers (Beauchesne et al.) are recommended.

## AUTHOR CONTRIBUTIONS

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


pathways describing world futures in the 21st century. Global Environ. Change 42, 169–180. doi: 10.1016/j.gloenvcha.2015.01.004


**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 © 2020 Olsen, Kaplan, Hansen, Fulton, Fogarty, Tam and Otto. 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 Local Stressors for Coral Reef Condition and Ecosystem Services Delivery Under Climate Scenarios

Mariska Weijerman1,2 \*, Lindsay Veazey <sup>3</sup> , Susan Yee<sup>4</sup> , Kellie Vaché<sup>5</sup> , Jade M. S. Delevaux <sup>3</sup> , Mary K. Donovan<sup>6</sup> , Kim Falinski <sup>3</sup> , Joey Lecky 1,3 and Kirsten L. L. Oleson1,3

<sup>1</sup> Joint Institute of Marine and Atmospheric Research, University of Hawai'i at Manoa, Honolulu, HI, United States, ¯ <sup>2</sup> Pacific Islands Fisheries Science Center, National Oceanic and Atmospheric Administration, Honolulu, HI, United States, <sup>3</sup> Department of Natural Resources and Environmental Management, University of Hawai'i at Manoa, Honolulu, HI, ¯ United States, <sup>4</sup> Gulf Ecology Division, U.S. Environmental Protection Agency, Gulf Breeze, FL, United States, <sup>5</sup> Biological and Ecological Engineering, Oregon State University, Corvallis, OR, United States, <sup>6</sup> Hawai'i Institute of Marine Biology, University of Hawai'i at Manoa, K ¯ anéohe, HI, United States ¯

#### Edited by:

Jamie C. Tam, Bedford Institute of Oceanography, Canada

#### Reviewed by:

Jeremy Baron Pittman, University of Waterloo, Canada Blair Greenan, Department of Fisheries and Oceans (Canada), Canada

> \*Correspondence: Mariska Weijerman mariska.weijerman@noaa.gov

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 07 August 2018 Accepted: 23 October 2018 Published: 09 November 2018

#### Citation:

Weijerman M, Veazey L, Yee S, Vaché K, Delevaux JMS, Donovan MK, Falinski K, Lecky J and Oleson KLL (2018) Managing Local Stressors for Coral Reef Condition and Ecosystem Services Delivery Under Climate Scenarios. Front. Mar. Sci. 5:425. doi: 10.3389/fmars.2018.00425 Coral reefs provide numerous ecosystem goods and services, but are threatened by multiple environmental and anthropogenic stressors. To identify management scenarios that will reverse or mitigate ecosystem degradation, managers can benefit from tools that can quantify projected changes in ecosystem services due to alternative management options. We used a spatially-explicit biophysical ecosystem model to evaluate socio-ecological trade-offs of land-based vs. marine-based management scenarios, and local-scale vs. global-scale stressors and their cumulative impacts. To increase the relevance of understanding ecological change for the public and decision-makers, we used four ecological production functions to translate the model outputs into the ecosystem services: "State of the Reef," "Trophic Integrity," "Fisheries Production," and "Fisheries Landings." For a case study of Maui Nui, Hawai'i, land-based management attenuated coral cover decline whereas fisheries management promoted higher total fish biomass. Placement of no-take marine protected areas (MPAs) across 30% of coral reef areas led to a reversal of the historical decline in predatory fish biomass, although this outcome depended on the spatial arrangement of MPAs. Coral cover declined less severely under strict sediment mitigation scenarios. However, the benefits of these local management scenarios were largely lost when accounting for climate-related impacts. Climate-related stressors indirectly increased herbivore biomass due to the shift from corals to algae and, hence, greater food availability. The two ecosystem services related to fish biomass increased under climate-related stressors but "Trophic Integrity" of the reef declined, indicating a less resilient reef. "State of the Reef" improved most and "Trophic Integrity" declined least under an optimistic global warming scenario and strict local management. This work provides insight into the relative influence of land-based vs. marine-based management and local vs. global stressors as drivers of changes in ecosystem dynamics while quantifying the tradeoffs between conservation- and extraction-oriented ecosystem services.

Keywords: trade-off, ecosystem-based management, multiple stressors, future scenarios, coral reefs, biophysical model, Hawai'i

## INTRODUCTION

Coral reef ecosystems provide valuable resources. They buffer coastal erosion, provide a cornucopia of food resources, attract tourism dollars, supply construction and pharmaceutical materials, and provide recreational opportunities for humans and essential habitat for threatened and endemic organisms (Hoegh-Guldberg, 1999; Moberg and Folke, 1999; Spalding et al., 2017). Furthermore, nature-based solutions, such as using living reefs as natural barriers for storm protection, are more cost-effective than manufactured infrastructure (Daily and Matson, 2008).

Despite the importance of reef ecosystems, they are under threat on a local scale from coastal development, overfishing, invasive species, and pollution, and on a global scale from ocean acidification, warming, and hypoxia (Carlton and Scanlon, 1985; Jokiel and Coles, 1990; Pörtner et al., 2005; Hoegh-Guldberg et al., 2011; Prouty et al., 2014). Two extensive reviews on threats to coral reefs identified ocean warming and ocean acidification as prominent threats (Burke et al., 2011; Brainard et al., 2013). Increasing carbon dioxide (CO2) emissions are slowly causing the world's oceans to become warmer and more acidic. Ocean acidification reduces calcification rates of all calcifying organisms including corals. Intense or prolonged ocean warming can result in the expulsion of the symbiotic algae that live in the coral tissue leaving them looking "bleached" and is hence called coral bleaching. Bleached corals have a higher change of mortality as they become more susceptible to pathogens (Maynard et al., 2015b). These threats are projected to intensify in coming decades (Van Hooidonk et al., 2013; Maynard et al., 2015b). Chronic stressors can lead to a more degraded reef system that has tipped to an algal dominated benthos (Bellwood et al., 2004; Hughes et al., 2010), and a replacement of top predatory fishes (large slow growing fishes) with species with a high turnover (Heithaus et al., 2008; Ruttenberg et al., 2011; Maynard et al., 2015a). These shifts are a concern because ecological functions and economic values diminish on such reef systems. Effective, long-term conservation of coral reefs and the goods and services they provide requires addressing the most critical threats.

The development of policies to address threats and promote ecosystem services is dependent on an understanding of ecosystem dynamics and responses to major stressors. Ecosystem models can synthesize the present-day condition and project changes of a system as a result of management regulations, climate conditions, or human use. Spatially explicit ecosystem models can also quantify the relative impacts of land-based vs. marine-based threats (e.g., land-based pollutants vs. fishing; (Álvarez-Romero et al., 2011; Barbier et al., 2011) and local vs. global stressors (Gurney et al., 2013; Weijerman et al., 2015). These types of models can evaluate tradeoffs of alternative courses of action to mitigate threats (Hulme, 2005; Fung, 2009; Weijerman et al., 2016). More recently, ecosystem models have been coupled with economic concepts to translate ecological outcomes in terms of human wellbeing, such as ecosystem services (Orlando and Yee, 2017).

One such spatially-explicit, biophysical ecosystem model is the COral Reef Scenario Evaluation Tool, CORSET (Fung, 2009; Melbourne-Thomas et al., 2011a; Principe et al., 2012). It includes hydrodynamics (which defines the connectivity), ecological dynamics, and land-based (nutrient and sediment pollution) and marine-based (fishing) stressors as well as global climate-related stressors (hurricanes and ocean warming). Its main use is to evaluate tradeoffs of alternative management or climate-related scenarios (Melbourne-Thomas et al., 2011b). Building on the extensive work to estimate nutrient and sediment loads and fish extraction on a 500 × 500 m scale around the main Hawaiian Islands (Wedding et al., 2017), we were able to incorporate these local stressors into the adapted CORSET model, the Hawai'i Reef dynamics Simulator or HIReefSim. Additionally, annual bleaching events were projected to start between 2035 and 2045 for the main Hawaiian Islands (Van Hooidonk et al., 2016). These projections were based on the results of an ensemble model of Intergovernmental Panel on Climate Change, Coupled Model Intercomparison Project Phase 5 (CMIP5), and as such, incorporated spatial variability in the effects of ocean warming on coral reefs. We used the projection of the Representative Concentration Pathway (RCP) 8.5 which estimates that by 2040 CO<sup>2</sup> emissions have reached 480 ppm and the onset of annual bleaching has begun (Van Hooidonk et al., 2016).

While ecological indicators are being used explicitly in management and policy (Arkema et al., 2006; Levin et al., 2013), decision-makers and the public often relate more to direct experiences, such as fishing, recreation, or coastal protection (Yee et al., 2014). Goods and services provided by coral reef ecosystems have long been acknowledged (e.g., Moberg and Folke, 1999), however, the relatively recent field of ecosystem service modeling quantifies these direct benefits to humans from functioning ecosystems (Bagstad et al., 2013). One approach uses "ecological production functions" (EPFs) to translate environmental shifts into economic implications in a way that is meaningful to decision-makers and resource managers (Nelson et al., 2009; Orlando and Yee, 2017). EPFs calculate the provision of goods and services as a function of specific ecological attributes (de Groot et al., 2002). Defining an EPF relies on an ecological understanding of which attributes are important to ecological function, as well as an economic understanding of what functions are valuable to humans. While an EPF quantifies the potential supply of ecosystem goods and services based on ecosystem condition, the realized value will depend on human demand and access (Wainger and Boyd, 2009). Economic valuation requires another relationship, ecosystem service valuation functions, to derive the value society gets from direct (e.g., food and recreation) and indirect (e.g., shoreline protection) use and nonuse (e.g., existence) of these goods and services (Compton et al., 2011; Yee et al., 2014). In this way, EPFs can be used to evaluate changes in potential provision of goods and services due to management, climate, and human use that affect the ecosystem, while valuation functions can calculate the cost/benefit of those changes.

To evaluate how different local management approaches [sediment mitigation and marine protected areas (MPA) establishment] could improve the provision of coral reef ecosystem goods and services, an ecosystem model that simulates impacts of both land- and marine-based management was parameterized for Maui Nui, Hawai'i, i.e., the islands of Maui, Lana' ¯ i, Moloka'i, and Kaho'olawe. These management approaches were also combined with two future severities of climate-related stressors. Reefs of Maui Nui served as a case study, but this tool can be used in other areas with similar local and global threats (Melbourne-Thomas et al., 2011a; Kapur and Franklin, 2017). Although several studies have shown the mitigating effects of local management on coral degradation in the face of climate change (Hughes et al., 2007; Kennedy et al., 2013; McClanahan et al., 2014), other studies have shown that under a "business as usual" greenhouse gas emissions future (IPCC RCP8.5 trajectory), local management may be unable to prevent further degradation of coral reef ecosystems (Thompson and Dolman, 2010; Selig et al., 2012; Weijerman et al., 2015; Hughes et al., 2017).

Here, we ask two questions: (1) What is the relative importance of land- and marine-based management action? and (2) Can local management mitigate the effects of climaterelated stressors? We expect that a combination of proactive local actions (sediment mitigation and fisheries controls) will attenuate declines in coral reef ecosystem goods and services delivery, but without local management, reefs will continue to decline, a trend exacerbated with more extreme future climate conditions.

## METHODS

## Study Region

Our study area encompasses ∼325 km<sup>2</sup> of shallow coral reef habitat across the Hawaiian Islands of Maui, Molokai and Lana'i, ¯ i.e., Maui Nui (**Figure 1**; the island of Kaho'olawe is excluded from this analysis due to lack of data). Of this area, 12 km<sup>2</sup> (3.6% of total reef area) are classified as MPAs, with just over 9 km<sup>2</sup> of the protected areas being designated as "no-take" area. The model domain consists of the shallow (0–30 m) reef zone around Maui Nui and is spatially represented by a 500 × 500 m grid cell network.

## HIReefSim

HIReefSim (Hawai'i Reef dynamics Simulator) is based on the framework of the Coral Reef Scenario Evaluation Tool (CORSET) developed by Fung (Fung, 2009) and adapted by Melbourne-Thomas et al. (2011a) and Principe et al. (2012). Model components include (1) 500 × 500 m gridded basemaps of the study region (see details below); (2) model dynamics (see details below); (3) larval connectivity zone delineations, which detail transition probabilities between larval sources and sinks; and (4) hurricane zone delineations, which were designed to represent grouped swaths of coastline that are similarly affected by storm events. Modeled stressors include land-based sediment and nutrient input, fishing, hurricane damage to corals and macroalgae, and climate-related coral mortalities due to coral bleaching.

## Basemaps

The HIReefSim model defines two consumer functional species groups, herbivorous (algal grazers) and piscivorous (predatory) fishes, and five benthic functional groups: macroalgae, turf algae, crustose-coralline algae (CCA), and spawner and brooder corals. Boosted regression trees generated spatial predictive maps of these ecological variables based on observations from a compilation of underwater surveys and an extensive gridded predictor dataset (**Table S1**) (Stamoulis et al., 2016; Delevaux, 2017). Unlike in the instantiation of CORSET, urchins and large (>60 cm) piscivores were not included due to very low abundance of large piscivores and a lack of urchin data preventing the creation of predictive maps.

## Dynamics

Coral reef ecosystems are extremely complex systems and influenced by a myriad of variables. HIReefSim only includes key ecological dynamics by using differential equations to estimate the interactions among the functional groups and their response to stressors in each grid cell (**Supplementary Text S2**). For example, ocean warming has led to global degradation of coral reefs with a consequent loss of structure followed by a decrease in fish biomass (Alvarez-Filip et al., 2009; Graham and Nash, 2013). Additionally, on a local scale, an increase in nutrients leads to an increase in the faster-growing macroalgae which in turn can reduce the growth of corals and impede coral recruitment. These are the key dynamics that are incorporated in the model (**Figure 2**), other stressors to coral reef ecosystems (e.g., ocean acidification, hypoxia, invasive species) are not included. Fung (2009) and Melbourne-Thomas et al. (2011a) give detailed descriptions of the model development, general model behavior, and sensitivity analyses. Kapur and Franklin (2017) describe the applicability of the CORSET model for Hawaiian reef systems. Here, only the main components of CORSET that form the basis for HIReefSim input are described (**Supplementary Text S2** has details of model equations and parameter estimates). Estimates of ecological variables represent the current (∼2004–2012) reef condition. The model has some stochasticity as it randomly selects the intensity, frequency, and region of impact to capture year-to-year variation in storms. The larval connectivity matrices were based on a pelagic duration of 45 days, calculated at the 50 m depth layer, for both corals and fishes (Wren and Kobayashi, 2016).

## Modeled Stressors

Spatially explicit stressor functions included nutrification and sedimentation (changes modeled via parameter scaling), fishing (modeled as a spatial explicit reduction in fish biomass), and coral and macroalgal mortality resulting from wave action (severity dependent on hurricane zone; **Figure 2**; **Table 1**). The projected increase in bleaching-related coral mortalities was a non-spatial stressor, affecting all corals equally.

### Model Adaptations

A regional study using CORSET (Kapur and Franklin, 2017), concluded a seemingly sustainable herbivore fishery would be possible despite the projected decline in coral cover. However, a limitation of this simplified model is its broad groupings of fishes in just herbivores and piscivores and the lack of twoway dynamics in fish size and fishing effort. For example, model results show that herbivores increase but the composition

FIGURE 1 | Modeled area of the Maui Nui complex consisting of the Hawaiian Islands: Maui, Moloka'i, and Lana'i, with the inset figure showing the location of Maui ¯ Nui in the Hawaiian Archipelago. The pink "reef" area is the 0–30 m depth range included in the model. The outer edge of open water area (blue) is defined by the 200 m depth contour. White areas interior of this indicate gaps in bathymetry data.

of this group is likely dominated by large-bodied herbivores, such as the larger surgeonfishes (e.g., Acanthurus dussiemeri, A. xanthopterus), that escape piscivore predation because of their size. However, if fishing is not restricted, these largerbodied fishes are key targets for spearfishers and an increase in spearfishing is not accounted for in the model. Additionally, reef structure is likely to erode due to coral cover decline, preventing fish recruits and juveniles from hiding (Alvarez-Filip et al., 2009; DeMartini et al., 2010; Graham and Nash, 2013) and increasing their accessibility to their predatory fishes (Rogers et al., 2014), ultimately leading to a decline in reef fish productivity (Gratwicke et al., 2005). Therefore, we included two scalars related to the structural complexity a coral reef provides for: (1) the survival of both herbivorous and piscivorous fish recruits given by the relationship survival (Frec) = aC / [1 + (a/b) <sup>∗</sup> C d ], where C is coral cover, and a, b and d are fitted parameters (**Table S2** in **Supplementary Text S2**; Gurney et al., 2013) and (2) the susceptibility of small and juvenile herbivorous fishes to predation with high coral cover leading to more hiding spaces and hence lower susceptibility to predation with the relationship refuge= min(H,Fpred <sup>∗</sup>C)], where C is coral cover and H herbivore biomass and Fpred a fitted parameter (**Table S2** in **Supplementary Text S2**) (Liu and Xing, 2012; Rogers et al., 2014).

## Model Calibration and Validation

The model was validated using a two-fold approach (Melbourne-Thomas et al., 2011a,b):


Results of model validation are presented in **Supplementary Text S3**. During calibration, it became apparent


TABLE 1 | Spatially-explicit stressors used to force scenario simulations in HIReefSim.

that the model was very sensitive to the parameters related to fish growth. For herbivores this was grazing pressure (gt, gm) and the biomass accumulation from grazing (mm, mt, me), and for piscivores this was prey availability (iph), predation pressure (gp) and biomass accumulation from predation (rp). We therefore randomly selected 50 values between the estimated ranges (**Table S2**) and assumed a normal distribution and ran each scenario 50 times to obtain uncertainty estimates related to these parameters.

## Scenario Simulations

Fifteen scenarios of separate and coupled effects of climaterelated stressors and management actions were simulated (**Table 2**). To define values for future baseline stressors (sediment and nutrient influx, herbivore and piscivore catches), annual projected population growth of 0.8% was used (DBEDT, 2016). Fishing pressure, along with sediment and nutrient runoff, was assumed to increase proportionally with the projected population growth. These stressors were further adjusted as specified within the scenario (**Table 2**).

To simulate climate change, we focused on hurricanes and bleaching-related coral mortality events. Both the frequency and intensity of cyclones in the North Pacific have increased, and sea surface temperatures (SST), which are directly correlated with storm intensity, are increasing as well (Emanuel, 2005). Murakami et al. (2013) project an average 267% increase in the number of cyclones that will reach the main Hawaiian Islands between 2075 and 2099 (0.75 annually increasing to 2 annually). To reflect these projections, we included hurricane events which directly impact Maui Nui an average of every 10 years for the running period of the climate change scenarios (**Table 2**). We specified that each hurricane event would reduce coral and macroalgae cover by 49% based on the average observed reduction in living bottom cover across the west coast of Hawai'i island (Dollar and Tribble, 1993). With the projected increase in sea surface temperatures, bleaching events will likely become annual, seasonal occurrences in the next 15–25 years (Van Hooidonk et al., 2016). Taking into consideration the variability in these estimates (Peters et al., 2017), we modeled a "severe" climate-related stressor scenario where we assumed that annual bleaching is every other year and a "less severe" climate-related stressor scenario where we assumed two annual bleaching events per decade.

## Ecological Production Functions (EPFs)

We applied four Ecological Production Functions (EPFs): "State of the Reef " (unitless), "Trophic Integrity of the Reef " (unitless), "Fisheries Production" (i.e., resource fish biomass in kg/km<sup>2</sup> ), and "Fisheries Landings" (i.e., annual fish catch in kg/km<sup>2</sup> ) to translate HIReefSim model output into values important for management applications (Principe et al., 2012; Yee et al., 2014). These EPFs represent a supporting service (first two EPFs), a potential provisioning service, and an actual provisioning service (Millennium Ecosystem Assessment, 2005), respectively, and roughly relate to biodiversity, ecosystem structure and function, conservation, and food yield outcomes. For each of the EPFs, the relative change from end to start of the simulation period (40 years) under the two climate change scenarios was calculated. To assess effectiveness, the relative change between the alternative management scenarios and current management was calculated. Based on model validation where the mean value of the 50 simulations described historical coral cover trajectory well (**Supplementary Text S3**), we used the scenario means for each parameter that described the EPF (see below) for each reef cell. We then present the resulting EPF values as a mean (and standard error) for all reef cells. We also show the spatial variation of each EPF under the different scenarios visually in maps.

#### State of the Reef

The ecological status of Maui Nui reefs was represented by the "State of the Reef," a supporting ecosystem service defined as:

$$\sum\_{i=1}^{5} w\_i \times R\_i \tag{1}$$

where w<sup>i</sup> is the weighting factor of each R<sup>i</sup> , with R<sup>i</sup> , representing the standardized value of five key indicators of reef structure: coral cover, macroalgal cover, total fish biomass, fish richness, and coral richness (**Supplementary Text S4**). An expert survey defined weighting factors as: (i) coral cover 30%; (ii) coral richness 20%, (iii) fish biomass 20%; (iv) fish richness 15% and (v) macroalgal cover 15% (Van Beukering and Cesar, 2004). Coral richness and fish richness values were based on Orlando and Yee (2017). The ecological indicator scores were scaled to the maximum value of each indicator.

#### TABLE 2 | Descriptions of 16 forecast simulations from 2010 to 2050.


Parameters are specified in Supplementary Text S2.

## Trophic Integrity

The trophic integrity was estimated by the ratio of calcifiers [corals (C) and CCA] and fleshy algae [turf (T) and macroalgae (MA)] and the trophic level of the fish community with herbivores (H) having a trophic level of 2 and piscivores (P) a trophic level of 4:

$$0.5\*\left(\frac{C+CCA}{T+MA}\right) + 0.5\*(2\*\frac{H}{H+P} + 4\*\frac{P}{H+P})\tag{2}$$

Ecological indicator scores were scaled to the maximum value of each indicator.

## Fisheries Production

Present-day predicted biomass of resource fish species (defined as species that had ≥ 450 kg of average annual harvest from 2000 to 2010 in the state of Hawai'i) was calculated as a ratio of total fish biomass (McCoy et al., 2018). This ratio was assumed to remain constant in the course of fluctuations in total fish biomass. Fisheries Production, therefore, represents the biomass of resource fish, a potential ecosystem service.

## Fisheries Landings

The model directly calculates fish catch as a function of available fish biomass (basemap input layer) maximum fishing effort, and accessibility (last two dynamics estimated for Maui Nui, **Supplementary Text S3**). Fisheries Landings is a provisional ecosystem service.

## RESULTS

## Local Management Strategies Can Attenuate Declines in Ecological Outcomes

Under the Current Management scenario, coral cover declined and was replaced by macroalgal cover (**Figure 3A**). Piscivore biomass also declined, to less than half the initial biomass but could recover in the last decade (**Figure 3B**), likely due to a shift in catch composition to predominantly herbivorous fishes (**Figures 3C,D**).

A continuation of current management attenuated the downward trend in coral cover. Comparing land-based management scenarios (high and low sediment mitigation, A scenarios) and marine-based management scenarios (additional no-take MPAs, B scenarios), sediment mitigation strongly impacted benthic composition, whereas fisheries management impacted fish biomass (**Figure 4**). For example, reduction of sediment input slowed the decline in coral cover compared to Current Management (**Supplementary Figures S5.1, 5.2**). Sediment mitigation exhibited a mixed effect on algal cover: it limited the space occupied by turf algae and CCA, but it had negligible effect on macroalgal cover (**Figures S5.3–5.5**). Fish biomass also responded to a change in benthic composition. In the high mitigation sediment reduction scenario (A1), herbivorous fish biomass declined slightly compared to Current Management (**Figure S5.6**) whereas piscivore biomass declined slightly under the low mitigation scenario (A2).

Marine-based management strategies had minimal impact on the trajectories of coral cover compared to the Current Management scenario (**Figure 4**, **Figures S5.1, 5.2**). CCA, however, did respond to MPA designations with decreases under one of the 30% MPA scenario (B1) and 20% MPA scenario (B4) but increased slightly under the other 30 and 10% MPA scenario (B2, B3 resp.; **Figure 4**, **Figure S5.5**). Interestingly, fish biomass hardly changed under MPA scenarios (**Figure 4**, **Figure S5.6**). The large error bar on the piscivore biomass also reflects the sensitivity of the model to the parameterization of fish-related variables.

The expansion of current MPA boundaries to include areas encompassing the top 10% (B3) and 20% (B4) of fish biomass and coral cover or the randomly placed 30% MPAs (B1 and B2) showed no clear pattern in the results. Striking is that the placement of the 30% no-take MPAs had opposite outcomes for fish catch which declined by up to 20% under B2 and increased to 32% under B1 (**Figure 4**) for herbivore and piscivore catches, respectively.

## Local Management Strategies Have Mixed Results Under Climate-Related Stressor Scenarios

Model projections resulted in steep declines in coral cover from current levels, especially under the more severe climaterelated stressors scenarios where hurricanes and thermal stress led to coral mortality (**Figures S5.1, 5.2**). Local management did somewhat mitigate the effects of a changing climate, with slightly more benefit (i.e., lower net loss) under the more severe scenario (**Figure 5**). Climate-related stressors were projected to have a positive effect on herbivorous fish biomass (**Figure S5.6** "base"), while local management buffered losses in piscivorous fish biomass, which then resulted in decreased herbivore biomass especially under the stricter management scenarios (E and F; **Figure S5.6** "scenario"). However, due to the large fluctuations in fish biomass, these differences in herbivore biomass were not statistically significant (**Figure 5**). Under all climate-related stressors and management scenarios (D-F), herbivores increased with 5–11 t/km<sup>2</sup> across Maui Nui from 2015 to 2050 (**Figure S5.6**). These increases represent an ecologically valid outcome, given the overall increases in turf and macroalgal cover, which are food sources for herbivorous fishes (**Figure 2**). Hurricanes did result in temporary "dips" in the overall increase in macroalgal cover under both climaterelated stressors scenarios but due to their relatively high growth rate, macroalgae recovered within a year. In general, under the climate scenarios, local management benefited corals and piscivores with a tradeoff in fisheries catches. These results were more pronounced under the stricter management scenario E and F, and under the higher climate stress scenario (**Figure 5**).

## Implications for Ecosystem Goods and Services

Under a future scenario of severe climate stress and current management (C1, **Figure 6**), Trophic Integrity declined by 15% and Fisheries Landings by 6%. However, in the entire Maui Nui area, State of the Reef and Fisheries Production increased by up to 41%, but with large spatial variation. Across all scenarios,

local management improved the Trophic Integrity of the reef (or dampened the decline) and the State of the Reef under the less severe climate change scenarios but had slightly negative effect on State of the Reef under severe climate change as well as on Fisheries Production and Fisheries Landings (**Figure 6**).

Local management could not prevent a decline in the Trophic Integrity; however, it did decrease the trajectory. Comparing Current Management (C1) to increasingly stringent local management of sediment and fishing stressors under severe climate-related stressors (D1, E1, and F1; darker colors in **Figure 6**), shows a trend of diminished decline in Trophic Integrity. Even larger improvements were evident under less severe climate-related stressors.

State of the Reef and Fisheries Production displayed improved results under all scenarios due to the increase in herbivore biomass that are components of these EPF. Model results showed a counterintuitive trend with management as both EPFs had lower values under severe climate scenarios compared to current management and Fisheries Production also decreased under less severe climate change. These lower values can be attributed to the smaller decreases in piscivore biomass due to management, leading to more predation pressure on herbivorous fish (**Supplementary Figures S5.6, S5.7**). On the other hand, State of the Reef clearly improved with local management under the less severe climate change scenarios (**Figure 6**).

Fisheries Landings were projected to decrease more compared to current management (C), especially under the scenarios including 30% MPAs (E, F). Due to the low piscivore biomass, an additional 20% reduced piscivore pressure (F) had almost no additional effect in landings (**Figure 6**).

Looking at the effectiveness of local management scenarios by comparing the end states of the EPFs of each scenario relative to current management, both the Trophic Integrity of the reef as well as the State of the Reef can benefit from additional management especially under less severe climate change (**Figure 7**). Fisheries Production fared less well, likely because of the increase in piscivore biomass (resulting in less herbivores) and the largest tradeoff was in Fisheries Landings.

Spatial variation in the results of management can provide insights into where to target local management. For example, the southern coastlines of Moloka'i and Maui showed declines for the State of the Reef and strict local management had little effect (**Figure 8**). By contrast, under less severe climate change, local management improved State of the Reef along the same coastlines (**Figure 9**). However, there were also some places (designated in blue) where management exacerbated declines. Generally, management had a positive impact along most of the coastlines (red areas in **Figures 8**, **9**).

Spatial patterns in Fisheries Production showed improvement under all scenarios and along most coastlines but less so

along the northern coastline of Moloka'i (we note that this area had relatively high biomass at the initialization of the scenario runs; **Figure 10**). Although in general many areas showed improvement, all coastlines also experienced losses; up to 100% in some places (**Supplementary Figures S6**). Management that included a 20% reduction in piscivore fish catches (Scenario F) tended to decrease overall Fisheries Production due to an increase in piscivores that preyed on herbivores (**Supplementary Figures S6**). At the same time, certain areas showed improvements with management partially due to spatial arrangement of the MPAs.

Trophic Integrity improved most along the southern coastline of Moloka'i and all around Maui (**Supplementary Figures S6**). As with the State of the Reef, the severity of climate change greatly influenced the results with much higher improvements under the less severe climate change scenarios.

## DISCUSSION

Marine resource managers are challenged with accounting for the cumulative effects of local and global stressors on coral reef ecosystems and the valuable services reefs provide to society. However, environmental conditions and human use can result in considerable spatial variability of both reef fish biomass as well as benthic community (Williams et al., 2015; Cinner et al., 2016; Gorospe et al., 2018). Managers can use spatially-explicit decision-support tools to prioritize areas of high ecosystem service value that could benefit from action. Scenarios of future conditions can guide decision-makers as to which of these actions are robust to projected climate impacts.

We evaluated the impacts of land-based vs. marine-based management and local vs. global stressors by assessing the relationships between different functional groups over time and under various management strategies and severities of climate-related stressors. As a metric for the effectiveness of these different strategies, four EPFs were used to model changes in ecosystem services; these ecosystem services best represented the economic interests of the State of Hawai'i pertaining to reef structure and resilience and reef-derived benefits.

Management of Hawai'i's valuable nearshore areas in the face of local and global stressors will need to be adaptive to changing conditions. As more people choose to live in or visit Hawai'i, local stressors will continue to mount, sometimes in unexpected places or ways. Impacts from global climate change offer another dimension of surprise. HIReefSim is an important tool to support adaptive local marine resource decision making. Our results suggest that policies are needed

to mitigate local threats, and that these policies should consider future risks of impacts from climate-related stressors, such as changing hurricane patterns and coral bleaching, and likely also sea level rise. In 2016, the governor of Hawai'i pledged to "effectively manage" 30% of the marine areas along the coastline by 2030, launching a multi-year marine spatial planning process. Key results from our analysis offer a number of suggestions for adapting near-term management to long-term conditions. First of all, strict management of all local pressures (i.e., land-based and fisheries) is needed everywhere–not just across 30% of the nearshore area–to obtain the best results for reef state and trophic integrity under all climate change scenarios. Secondly, place matters. We found that spatial variation in the effects of local management was particularly high for reef state and fisheries production, and moderate for trophic integrity and fisheries landings, suggesting that fine-tuning place-based management could improve outcomes.

## Does Local Management Matter in the Face of Climate-Related Stressors?

Yes, despite a decline in Trophic Integrity and State of the Reef EPFs, both these EPFs displayed a buffering effect of strict management on the degradation of coral reefs, mostly under the less severe climate change scenarios. However, as a tradeoff, Fisheries Landings decreased overall compared to today's levels but the maximum decrease was just under 6% (**Figure 6**). Average Fisheries Production increased between 42% (scenario C2) and 38% (scenario D2) mostly due to the increases in herbivore biomass correlated with increases in turf algae. In absolute terms, local sediment control was particularly critical in slowing the decline of reefs under severe climate change. In these scenarios, coral cover was more prone to decline rapidly, which underscores both the importance of reducing greenhouse gases and implementing proactive, stringent management policies to mitigate declines in coral cover (Ortiz et al., 2014; Hughes et al., 2017). These results correspond to other modeling

FIGURE 6 | Relative change (%) from 2010 to 2050 of each scenario including severe (darker bars) and less severe (lighter bars) climate change impacts. The blue bars (C1, C2) represent a future projection of Current Management practices, population growth, and severe (bleaching-related coral mortalities every other year; C1) and less severe (mortalities every 5 years; C2) climate-related stressors, respectively. The effects of local management are also shown under two climate scenarios: severe climate-related stressors in darker colors (denoted by a 1) and less severe stressors in lighter colors (denoted by a 2). The local management scenarios include: (D–gray bars) 10% no-take MPAs and low sediment mitigation, (E– green bars) 30% MPAs and strict sediment mitigation, and (F–red bars) 30% MPAs with an additional 20% reduced piscivore fishing effort and strict sediment mitigation.

studies. Weijerman et al. (2015) showed that reefs in Guam would experience devastating declines under projected annual bleaching events despite local management directed at reduced land-based pollution and fishing. The same result was found for a Caribbean reef where, under the projected greenhouse gas emissions of a business-as-usual scenario, corals failed to recover from frequent bleaching events (Ortiz et al., 2014). Reducing sediments is not a complete solution given that nutrients and other contaminants of concern are still present and may have a very different spatial signature. We focused on sediment impacts for this project based on the completeness and trustworthiness of available data layers, but future efforts may consider the impacts of other sources of land-based pollution.

Herbivores appeared to fare well under climate-related stressors as coral cover declined and algal cover increased (**Figures 6**, **7**). The strong correlation between algae and

herbivores could be explained by the control of bottom-up, not top-down, mechanisms. A comparison of piscivore biomass in the populated Main Hawaiian Islands (MHI) with the non-fished Northwestern Hawaiian Islands (NWHI) shows that piscivore biomass was about 10–44 times higher in the NWHI (Williams et al., 2011). Even within the MHI, relatively remote and inaccessible locations had about five times the biomass of apex predators compared to more open areas (Williams et al., 2008). Thus, the low biomass of piscivores observed in the MHI could explain why top-down effects may not be strong enough to explain some of the results, and suggests that strict fishery management is required to improve predation as an ecosystem function (**Figure 5**, **Figure S5.7**).

Alternatively, the increase in piscivores could be viewed as an undesired outcome since more piscivores led to elevated predation pressure on herbivores, reducing their biomass and

the overall Fisheries Production (**Figures 5**–**7**). Ecologically, a higher trophic level (piscivores have a trophic level of 3.5–5 compared to a level of 2 for herbivores) of the fish community represents an ecosystem that is energetically more optimal and mature (Graham et al., 2017). Hence, for ecological reasons, one might want to strive to obtain a fish community with a high composition of piscivores. Economically, this also makes sense as in general piscivores sell for a higher price (\$3.5–\$5 per pound) compared to herbivores (\$2–\$3.5 per pound). Thus, although the results indicate a lower Fisheries Production under local management scenarios, this is not necessarily a bad thing; the Trophic Integrity of the Reef improved with more stringent management (E, F; **Figure 7**), and the reduced catch may not translate into large economic losses.

Wide-scale spatial variability across Maui Nui was projected for the EPFs, with declines of up to 100% across the northern shores of Maui and Moloka'i for Fisheries Production, and large declines in State of the Reef along southern and eastern shores (where the impacts of hurricanes were projected to be highest), highlighting the importance of identifying areas where management will likely be most successful.

## How Important Is It to Consider Both Land- and Marine-Based Threats in Local Management Action?

Effectively managing the entire coral reef ecosystem requires the combination of both land-based and marine-based approaches. Land-based management was most beneficial for the benthic community, especially coral cover, whereas marine-based management not only decreased macroalgal cover, opening up substrate for reef calcifying CCA, but most notably decreased the downward trend in piscivore biomass (**Figure 3**). Of equal importance in evaluating the effectiveness of marine-based management (MPAs) are the often-conflicting extraction and conservation objectives (Dichmont et al., 2013). If the objective of MPAs is to improve coral cover, MPA establishment proved less effective than reducing land-based pollution (**Figure 3**). MPAs can enhance the resilience of coral reef ecosystems through trophic interactions (Mellin et al., 2016) or lead to small increases in coral cover (Graham et al., 2011), but reducing the main local stressors that drive coral decline directly (such as sediment inputs) is more effective (ISRS, 2004) as the model results also showed. For maintaining or increasing key ecosystem functions under the projected impacts of climate-related stressors, no single management tool was effective, but a combination of both land-based and marine-based management was needed as well as a reduction in greenhouse gas emissions (**Figures 5**– **7**). These modeling results are corroborated by other studies (Dichmont et al., 2013; Weijerman et al., 2015; Arias-Gonzalez et al., 2017).

Random placement of no-take MPAs may not benefit the overall benthic community (**Figure 3**). For example, the 30% MPA designation in scenario B1 mitigated loss of fish biomass less than B2 for fish biomass, but improved the fate of corals (**Figure 3**). Placing MPAs in areas with currently the highest 10% of fish biomass and coral cover (B3) also increased the piscivore biomass but when (almost) doubling this area by making the current MPAs also no-take MPAs (B4), piscivore biomass did not double and catches stayed very similar (**Figure 4**). Likely, the areas which currently have the highest fish biomass are already somewhat exempt from high fishing pressure, either because they are difficult to access or because only few people live close by, underscoring the importance of MPA placements. A separate issue is whether MPAs can achieve desired outcomes, as that also depends largely on strong governance (Cinner et al., 2016) and compliance (Gill et al., 2017).

## CONCLUSIONS

Overall, in the face of global threats, local management had mixed results for the ecosystem goods and services the reef provides. It abated the decline in Trophic Integrity of the reef and improved State of the Reef especially when both land-based and marine based approaches were combined but reducing greenhouse gas emissions is essential to avoid catastrophic loss of ecosystem services. Management that was more stringent required trading off Fisheries Production and Landings. By including extraction and reef resilience objectives in ecosystem goods and services, we provide a generalizable tool to clearly evaluate the tradeoff of these conflicting goals under various management and climate-related stressors. Based on the limitations of the model structure, a cautionary interpretation of the model's predicted long-term trajectory should be taken. Future work may want to improve the resource fish ratio, which we assumed to stay constant over the entire period. This assumption is unrealistic because fishers are likely to target specific species (e.g., jacks or other piscivores), causing community structure to shift. Furthermore, the results from the small number of MPA scenarios call for more thorough analysis, particularly with respect to their interaction with land-based pollution control. The model could also be used for evaluating the relative effectiveness of different fisheries management strategies, e.g., targeting different trophic groups. A similar endeavor could focus on compliance within those MPAs to test the effectiveness of restriction levels. Results of these improvements could greatly benefit the ongoing discussions of how to manage 30% of the coastline effectively.

## DATA AVAILABILITY STATEMENT

The raw data supporting the conclusions of this manuscript are available at http://olesonlab.org/data/.

## REFERENCES


## AUTHOR CONTRIBUTIONS

MW, LV, KO, JD, and SY: study design. KF, JL, and MD: data sources. MW, LV, KV, and SY: Coding. MW, LV, and KO: analysis/Interpretation. MW wrote the first draft of the manuscript. All authors wrote sections of the manuscript, contributed to manuscript revision, read and approved the submitted version.

## FUNDING

Funding for this research was provided by PICSC G13AC00361, NOAA CRCP NA13NOS4820020, and NOAA CRCP NA17NOS4820076.

## ACKNOWLEDGMENTS

The authors thank the providers of the input data used to build our models, in particular Johanna Wren for the connectivity matrix, the Hawai'i Monitoring and Research Collaborative for the underwater visual survey data, Bill Ward for regional hurricane impacts, as well as, contributions to the predictive model development by Kosta Stamoulis, Matt Poti, Bryan Costa, and Matt Kendall. The authors also thank Tayler Massey and Megan Barnes for their assistance and/or advice in the completion of this work and Brett Taylor and two reviewers for providing comments on earlier drafts. The views expressed in this article are those of the authors and do not necessarily reflect the views or policies of their respective agencies.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2018.00425/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 Weijerman, Veazey, Yee, Vaché, Delevaux, Donovan, Falinski, Lecky and Oleson. 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.

# Economic and Ecosystem Effects of Fishing on the Northeast US Shelf

Modeling tools that can demonstrate possible consequences of strategies designed

Gavin Fay† , Geret DePiper, Scott Steinback, Robert J. Gamble and Jason S. Link\*

National Marine Fisheries Service, Woods Hole, MA, United States

#### Edited by:

Cecilie Hansen Eide, Norwegian Institute of Marine Research (IMR), Norway

#### Reviewed by:

Donald F. Boesch, University of Maryland Center for Environmental Science (UMCES), United States Xiutang Yuan, National Marine Environmental Monitoring Center, China

> \*Correspondence: Jason S. Link Jason.Link@noaa.gov

#### †Present address:

Gavin Fay, Department of Fisheries Oceanography, School for Marine Science and Technology, University of Massachusetts Dartmouth, New Bedford, MA, United States

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 11 October 2018 Accepted: 04 March 2019 Published: 22 March 2019

#### Citation:

Fay G, DePiper G, Steinback S, Gamble RJ and Link JS (2019) Economic and Ecosystem Effects of Fishing on the Northeast US Shelf. Front. Mar. Sci. 6:133. doi: 10.3389/fmars.2019.00133 to operationalize ecosystem-based fisheries management (EBFM) should be able to address tradeoffs over a wide suite of considerations representing the scope of marine management objectives. Coupled ecological-economic modeling, where models for ecological and economic subsystems are linked through their inputs and outputs, allows for quantification of such tradeoffs. Here, we link the harvest output from fishery management scenarios implemented in an end-to-end ecosystem model (Atlantis) to an input–output regional economic model for the Northeast United States to calculate changes in socio-economic indicators, including the consequences of management action for regional sales, wages, and employment. We implement three simple scenarios (maintain, decrease, or increase current fishing effort), and compare model-projected values for systematic and sector-specific indicators. Systematic indicators revealed different ecological and economic outcomes, with large ecological responses and clear tradeoffs among the catch and biomass of species groups. Economic indicators for the region responded similarly to fishery yield; however, changes in total sales did not match those in landed catch. Under increased fishing effort, a lower proportional increase in sales relative to total landed catch arose due to increased yield from lower value species groups. Average fisheries income changed little among scenarios, but was highest when effort was maintained at current levels, likely a reflection of fleet and catch stability. Our results serve to demonstrate that consequences of management may be felt disproportionately among species through the region and across different fisheries sectors. With our coupled modeling approach of passing Atlantis ecosystem model outputs to an input–output economic model, we were able to assess effects of fisheries management across a broader suite of indicators that have relevance for policymakers across multiple objectives.

Keywords: Atlantis, ecosystem modeling, bioeconomic modeling, tradeoff analysis, input–output models, ecosystem-based management

## INTRODUCTION

The need for ecosystem-based fisheries management (EBFM) is well established, with a focus on managing the indirect effects of fishing across a broad set of ecological and societal factors under both tactical and strategic decision-making. While much progress has been made toward implementing EBFM, much work remains (Pitcher et al., 2009; Hilborn, 2011; Marshall et al., 2018). Evaluating options for implementing EBFM requires a better understanding of the links between marine ecosystems, the goods and services humans derive from them, and the effects of both environmental and

human pressures on these ecosystems and services (e.g., Marasco et al., 2007; Link, 2010; Kruse et al., 2012). Further, the performance of management options must be tested with respect to operational objectives that encompass both ecological and socioeconomic goals accounting for these links and pressures. It is necessary therefore to explore a range of outputs across many management scenarios when assessing indicators of management performance.

A range of ecosystem models have been developed to address the needs of EBFM (e.g., Plagányi, 2007). Many of these models have had an ecological focus, or have only evaluated economic effects for single industry sectors (commonly a single commercial fishery). Integrated economic-ecological frameworks (e.g., Arrow et al., 1995) that extend the bioeconomic approach and include models for both human economies and ecosystem dynamics offer the potential to provide critically required decision support when assessing the value of marine ecosystems (Jin et al., 2012). Some of these modeling tools have begun to be used in a management strategy evaluation (MSE; Bunnefeld et al., 2011) framework to address tradeoffs among management objectives in an ecosystem context (e.g., McDonald et al., 2008; Plagányi et al., 2013; Fulton et al., 2014). MSE has also been applied to quantify the economic risk of alternative fisheries management strategies (e.g., Little et al., 2013). Developing methods to quantify the effect of fisheries management strategies on a suite of ecosystem services is a recognized component of integrated ecosystem assessments (IEAs; Levin et al., 2009). IEAs are a recognized means for integrating and using information to implement EBFM, and modeling is a key part of them.

Atlantis (Fulton et al., 2011) is an end-to-end ecosystem model that was designed to quantify tradeoffs between economic, ecological, and societal management goals. Atlantis is well suited to evaluate ecosystem-based management strategies because it couples biophysical models of the ecological system to models for human activities (such as fishing) and incorporates models for the steps, procedures, and tools of the management decision process. Fulton et al. (2014) used Atlantis to compare the performance of fisheries management strategies against a broad range of societal indicators for a multispecies fishery in Southeast Australia. Kaplan and Leonard (2012) coupled an Atlantis model for the California Current ecosystem to a regional economic model for the United States west coast to illustrate the direct and indirect effects of alternative groundfish management strategies. This analysis extended many typical fisheries bioeconomic modeling approaches by considering industry sectors that support or are influenced by changes in fishery production, such as industry suppliers, employment, or even household spending. While an Atlantis model for the Northeast United States exists (Link et al., 2010) and has been used to assess ecological responses to management strategies (Fay et al., 2017; Olsen et al., 2018), this model has not yet been used to assess economic indicators for the region.

The Northeast United States large marine ecosystem (LME) has supported economically important fisheries for hundreds of years (Link et al., 2011a). For example, 2012 gross nominal revenue in the Northeast United States Multispecies Groundfish Fishery was \$305.5 million (Murphy et al., 2014). Bioeconomic analyses for the region have rarely focused on system-level objectives. Most models estimating the economic effects of fisheries management strategies in the Northeast United States have focused on a particular fishery and the direct impact of policy on fishermen. Examples include models for scallops (Harksever et al., 2000; Valderrama and Anderson, 2007; Hart, 2009), lobsters (Acheson and Reidman, 1982; Holland, 2011), and silver hake (Thunberg et al., 1998). Although some papers look at multiple species simultaneously, they tend to consider a subset of species of commercial and conservation interest and most have not taken an overall perspective of the effects of changes in fishery production on the larger economy in the Northeast United States (e.g., Kirkley et al., 2011; Lehuta et al., 2014; Scheld and Anderson, 2014). Hoagland et al. (2005) (see also Steinback and Thunberg, 2006) constructed a model for the coastal economy of the Northeast United States and estimated that the activity of United States marine sectors in the Northeast Shelf LME accounted for 10% of the total gross state product for the region. However, the contribution of fisheries to these grosses was low (2%). Consideration of marine sectors as a portfolio of economic activities, as well as risk related to variance of expected returns from a set of individual fish stocks via portfolio analysis, also offers opportunity for integrating economic considerations into marine management and evaluation of risk (e.g., Edwards et al., 2004; Jin et al., 2016; Link, 2018).

Dynamically interacting models of economic and ecological processes might best account for feedbacks and interactions between changes in fishery production, ecosystem state, and economic variables. Constructing such models is, however, time and data-intensive, requiring parameterization of behavioral models that include relationships between economic variables and human decision processes and necessitate a substantially reduced number of economic sectors for modeling purposes (e.g., van Putten et al., 2012). In a simpler approach, input– output models allow for coupling of ecological and economic models by quantifying both the direct and indirect economic impacts of changes in harvest rates derived from the ecological model. Input–output models for fisheries have been used at the single species (e.g., Northeast United States Atlantic herring; Kirkley et al., 2011), species groups (United States West Coast groundfish; PFMC, 2015), and ecosystem levels (United States West Coast; Kaplan and Leonard, 2012; PFMC, 2015). When applied at the ecosystem level, this approach can be used to evaluate system-wide tradeoffs across ecological, economic, and social management objectives.

Here we link the harvest from a marine ecosystem model for the Northeast United States continental shelf to an input– output regional economic model for the Northeast United States. We calculate changes in socio-economic indicators (such as jobs and earnings) and compare these changes to values of ecological indicators from the ecosystem model. We use the coupled models to explore the ecological and economic consequences of three simple fishing effort scenarios initialized to the historical range of these data with variable fishing scenarios projected over a 10 year period. A baseline scenario reflects historical conditions where fishing effort during this period was substantially lower than in previos years. This is compared with two alternative scenarios where a change in effort was implemented. In particular, we wanted to quantify the effects of changes in fishing fleet sector landings associated with these effort changes on the regional economy.

## MATERIALS AND METHODS

fmars-06-00133 March 22, 2019 Time: 16:52 # 3

This section briefly summarizes both the ecosystem and economic models, describes model coupling, and outlines the fishing scenarios tested. Rather than provide full details of model descriptions (which are referenced elsewhere), we focus on relevant details for the model coupling.

## Marine Ecosystem Model: Atlantis-NEUS

Atlantis is an end-to-end marine ecosystem model that has been applied to multiple marine systems globally (Fulton et al., 2011; Weijerman et al., 2016). Atlantis consists of biophysical, fishing dynamics, management, and assessment sub-models, and is intended to be a strategic tool for comparing the performance of management strategies under alternative scenarios (i.e., MSE; Bunnefeld et al., 2011; Fulton et al., 2014). Atlantis-NEUS, the application of Atlantis to the Northeast United States marine ecosystem (Link et al., 2010, 2011b, covers the continental shelf from the Gulf of Maine to Cape Hatteras (**Figure 1**), and is resolved into 22 spatial regions, each of which is further resolved by depth. Physical parameters and flows in the system are modeled in Atlantis-NEUS using output from a regional ocean model. The biogeochemical-based ecological model of Atlantis-NEUS consists of 45 functional groups, 24 of which are vertebrates. The exploitation sub-model of Atlantis-NEUS consists of 18 fishing fleets that are combinations of fishing gears and target species. The model was tuned to data from the Northeast United

FIGURE 1 | Map showing spatial structure of Atlantis-NEUS (white polygons) and the NERIOM models (coastal counties shaded in dark gray, coastal states light gray).

States from 1963–2004, primarily using information from the biannual Northeast Fisheries Science Center (NEFSC) bottom trawl survey (Azarovitz, 1981; NEFC, 1998), and the NEFSC commercial fisheries database (NEFSC unpublished data). Full technical details of the Atlantis-NEUS model can be found in Link et al. (2011b), and a more comprehensive summary of model details, calibration procedure, and key scenarios can be found in Link et al. (2010). Model runs to 2014 using the predictive scenario capability of Atlantis have been compared to data from 2005–2014 documenting model skill for those species groups that formed the focus of model calibration (Olsen et al., 2016).

## Economic Model: Northeast Region Input–Output Model (NERIOM)

The input–output economic model Northeast Region input– output model (NERIOM; Steinback and Thunberg, 2006) was used to quantify the regional economic effects of changes in commercial fishing landings. The NERIOM model was developed from the IMPLAN Pro system (IMPLAN Group LLC), which is based on the general Leontief approach to input/output modeling (Leontief, 1951). The NERIOM model translates seafood sector revenue to supporting industries' sales, income, and employment. NERIOM can assess the impacts of management alternatives on the entire Northeast Region's economy and on the economies of 24 specific sub-regions (**Figure 1**) that represent semi self-sufficient fishing areas with similar economic networks and attributes.

Commercial fishing activities are grouped into 18 distinct gear sectors (**Table 1**). Changes in output (e.g., sales) for the fisheries harvesting sectors associated with the fisheries management scenarios are obtained from changes in landings for each sector from the Atlantis model, using a landings-weighted average price per species group. The estimated direct changes in gross revenues for harvesters are then tracked backward to bait and ice suppliers, gear and vessel repair shops, gas stations, and the host of other service and goods providers servicing fishermen through the NERIOM multipliers. Additionally, forward-linked effects on fish exchanges/auctions, wholesale seafood dealers, and seafood processors are estimated, including the multiplier effects of their suppliers. We acknowledge the assumptions of this approach regarding fixed inputs. However, recent years of data used to inform the parameterization of NERIOM (matching the scenario period), and the scenarios we examine are from the same time period, rather than some long-term future projection during which assumptions about prices and inputs would be more tenuous, may make this less of a concern.

## Coupling Atlantis-NEUS Outputs to NERIOM

Linking the ecosystem and economic models required mapping fisheries landings by Atlantis fleets and spatial regions to NERIOM fleets and regions. Bottom trawl and scallop dredge fleets in NERIOM are defined by vessel size, with the small boat fleet encapsulating vessels < 50 ft, the medium boat fleet falling

TABLE 1 | Proportion of total landings (across all species groups) from Atlantis fleets allocated to NERIOM fleets for the baseline scenario.


Individual mappings for landings of each species group were similarly executed, but are not shown here.

between 50 and 70 ft, and large boat fleet ≥ 70 ft. Further, the lobster pot fleet in NERIOM is delineated by inshore and offshore components. These delineations are made to reflect the substantially different economic production functions associated with each type of vessel. In Atlantis, fleets are based on gear type and target species groups. We mapped landings between Atlantis and NERIOM at the species level by calculating an average proportion of catch for each species in the NEFSC commercial fisheries databases that was taken by each NERIOM vessel/gear category during the years 2007–2011. **Table 1** summarizes the proportional amount of landed catch for each Atlantis fleet that was transferred to each of the NERIOM fleets using this approach. The mapping differs substantially depending on the species and gear being considered. For example, total haddock landings by bottom trawl are historically distributed such that 85, 13, and 2% are associated with the large, medium, and small bottom trawl vessels, respectively, whereas the distribution of Atlantic cod landings by bottom trawl are 48, 32, and 20%, respectively, for the large, medium, and small vessel segments of the fleet. Our mapping explicitly accounts for such differences.

Fishing effort within Atlantis-NEUS is not directly associated with ports because a distance-to-port-based fleet dynamics model is not implemented in the effort scenarios used. The distribution of fishing effort for each fleet is allocated spatially in the Atlantis-NEUS model according to prescribed distributions (that can change over time) to be characteristic of the historical data. NERIOM requires input by state (and specific ports), which can be calculated by allocating proportions of the landings to the primary ports designated within Atlantis-NEUS. We allocated landings to ports within Atlantis by assuming that the landings of each fleet by spatial box could be assigned to ports based on the distance of the centroid of the box to the ports. The proportion pij of an Atlantis box's landings assigned to a particular port

was then:

$$p\_{\mathbf{j},\mathbf{i}} = \frac{1}{D\_{\mathbf{i},\mathbf{j}}^2} \Big/ \sum\_{\mathbf{i}=1}^{N} \frac{1}{D\_{\mathbf{i},\mathbf{j}}^2}$$

where Di,<sup>j</sup> is distance (from the centroid) of box j to port i and N is the number of ports active for each fleet in the Atlantis model. We tested the sensitivity of the assumption for this relationship by also calculating landings by port assuming inverse distance (rather than inverse squared distance). While the values for the landings by port changed slightly, these did not impact results qualitatively.

The ports defined within Atlantis and the regions modeled in NERIOM differ, meaning that landings were again mapped between the two models. **Supplementary Table S1** presents the mapping of Atlantis ports and NERIOM fleets to NERIOM regions. Port to region mapping was conducted through a hierarchical assignment algorithm. The first step assigned ports to the region of the Northeast coast in which they fell based off the original county definitions used to classify regions in the NERIOM model. For example, the Atlantis port of Chatham, MA, United States, naturally maps to the Cape and Islands region of NERIOM, while Gloucester, MA, United States, maps directly to the Gloucester, North Shore region. The second step in the algorithm then attributed landings to regions within the NERIOM model that had no corresponding port of landing in Atlantis. For example, landings to the port of Atlantic City were allocated to New York. Although this allocation may in some instances be questionable, the inverse distance squared function that allocated landings from Atlantis boxes to ports is also an approximation. At the Northeast Region level, subregional differences between observed and modeled landings are not large enough to have measurable effect on the NERIOM estimates of economic impacts. To better reflect recent patterns of landings, the final step of the port to region mapping re-allocated landings from regions with no recent history of specific fleet activity to nearby regions that have had landings from that fleet.

## Scenarios and Evaluation

fmars-06-00133 March 22, 2019 Time: 16:52 # 5

We use the coupled models to explore the ecological and economic consequences of three simple fishing effort scenarios initialized to cover the historical range of these data (1964–2004). We then started variable fishing scenarios projected for the period 1995–2014, and focused our reporting on results for the final 5 years of that time period.

Three fisheries management scenarios were considered in Atlantis to evaluate the effects of changes in landings on the regional economy: (1) a base scenario of fishing effort for 1995–2014 fixed at levels consistent with observed data for the Northeast United States from 1995–2004, (2) a reduced effort scenario where the fishing effort for 1995–2014 was halved for all Atlantis fleets compared to the base scenario values, and (3) an increased effort scenario where fishing effort for 1995– 2014 was twice that in the base scenario (again, for all fleets). We selected the fixed effort scenario (described in Link et al., 2011b) as the base scenario since this more closely represents the observed dynamics of some major invertebrate fisheries that are economically important in the region than alternatives available for the Atlantis-NEUS model. We chose the multiplicative, crossfleet effort scenarios as alternatives to the base to quantify economic effects that bracket common and reasonable largescale changes in fisheries operations which have been observed. While more complicated fisheries management scenarios could be envisaged, these simple scenarios provide an easy way to demonstrate economic impacts at the regional level. Scenarios where the magnitude of the effect size on fishing effort was even greater (e.g., fishing effort five or one-fifth times that of the base scenario) were run during exploratory analyses but are not reported here for ease of presentation.

The landings for the final 5 years of the Atlantis simulations were averaged and used as inputs to NERIOM. As NERIOM is a static model, this provided one way to moderate some of the inter-annual variability in landings within the analysis. For each of the three Atlantis scenarios, biomass and landings by species group were recorded, in addition to a set of ecological indicators that capture fundamental features of marine ecosystems related to fishery exploitation (e.g., Shin et al., 2010). Output from NERIOM is summarized in terms of effects on sales, income, and employment, both at the regional level and by individual sector. In our analyses, we focus on the changes in quantities of interest under the reduced and increased fishing effort scenarios relative to those obtained from the base scenario rather than the absolute values for metrics.

## RESULTS

## Scenario Results: Ecological Indicators

A large biological response was seen under the reduced effort scenario, with increases in biomass for many species groups associated with up to 50% reductions in the catch of many fish and invertebrate groups (**Figure 2**). These responses were variable, with large increases in biomass (>50% over base) for scallops, white hake, bluefish, benthopelagics, monkfish, cod, and silver hake, and modest increases (<20%) for many other targeted fish. Decreases in catch under the reduced effort scenario were not necessarily associated with increases in biomass, with very small changes in biomass (in some cases decreases) for lower trophic level groups, mainly as a result of increased predation pressure from the increased biomass of other piscivores (**Figure 2**). The increased fishing effort scenario resulted in increases in the catch of many species groups (**Figure 2**), with >100% increases in the catch of herring, mesopelagics, anadromous small pelagics (e.g., alewives and shad), and cephalopods. Catch declined under this scenario for a few species (notably cod and silver hake). These tradeoffs among species' yield resulted in an increase in total catch from the system over the baseline, but only of 47% (i.e., doubling effort did not double overall yield, **Figure 3**). In general, species groups that showed large increases in biomass under the reduced effort scenario compared to the base showed large decreases in biomass under the increased effort scenario (**Figure 2A**; e.g., Atlantic cod, silver hake, scallops, bluefish, small pelagics).

The effects of the changed effort scenarios are also seen at the system level. Total catch reduced to 64% of that in the baseline scenario under the reduced effort scenario (**Figure 3**). While the effect on total ecosystem biomass was much smaller than this, the fish community was impacted with a decrease in the ratio of demersal to pelagic fish under the doubled effort scenario and a concurrent increase in this indicator for the reduced effort scenario (**Figure 3**). Threatened and protected species were affected by the changes to fishing, with the biomass of seals and birds being reduced in the increased fishing effort scenario (**Figure 3**). A larger number of species groups were observed to fall below commonly used management reference points in the increased fishing effort scenario. The proportion of species groups deemed to be overfished (i.e., biomass was less than half the estimated BMSY) increased by a factor of three under the doubled effort scenario compared to the baseline, with 40% of species groups considered overfished in the increased fishing effort scenario (**Figure 3**, "PropOF"). The system-wide exploitation rate (total catch/total biomass) increased from 6 to 10% under the increased effort scenario, with the yield from the system exceeding 16% of total primary production (**Figure 3**). As described at the species level, changes in catch were more prominent for pelagic groups than demersals, with total catch from pelagics having a higher magnitude of change than that of demersals under both increased and decreased effort scenarios.

## Scenario Results: Economic Indicators

Forcing the changes in landings from the Atlantis model to the NERIOM model had large and variable effects on sales for the fishing sectors (**Figure 4**). Under the reduced effort scenario, sales for many sectors decreased up to 50% (**Figure 4A**), though the scallop dredge and bottom trawl sectors had decreases smaller than this. In contrast, large increases (e.g., >50%) in sales for only some sectors (lobster traps, small dredge, and surf

clam/ocean quahog dredge) were observed under the increased effort scenario over baseline. This translated to a disproportionate effect on the sales from seafood processing, seafood dealers, and fish exchanges/auctions, with 30–50% reductions in value under the reduced effort scenario but less than 20% increases under the increased effort scenario. Consequently, overall economic indicators for the region responded similarly, with total sales being reduced by 26% under the reduced effort scenario and increasing 19% over base under the increased effort scenario. Similar effects were seen with respect to total income and total employment for the entire Northeast Region, resulting in very small changes to overall average fisheries income (total income/total employment, **Figure 3**). Although the magnitude of the differences was small, average incomes under both the increased and decreased effort scenarios were lower than that in the baseline. The average income for the fishing sectors was 60% that of all sectors included in the analysis. Average incomes for the fishing sectors were also less than the baseline in the changed effort scenarios (\$491 less per year than baseline for the increased effort, and \$833 less per year than baseline for the decreased effort scenario).

The number of jobs for some of the fishing sectors was more sensitive to the increased effort scenario than total sales (e.g., hand/mobile gear, demersal longline, midwater trawls, **Figure 4**). There were distinct regional differences in the magnitudes of effects for the changed effort scenarios, reflecting the differences in species and fleets associated with the various ports. Most notably, the changes in sales and employment for the scallop dredge sectors that occurred during the changed effort scenarios were completely a result of changes to the New England economy, with very small changes to these sectors in the Mid-Atlantic (**Figures 5**, **6**). In contrast, decreases in total sales from the

FIGURE 3 | Levels of response for ecological and economic indicators to the three fishing effort scenarios. Ecological indicators are the average value from the terminal 5 years of the run, the period used to calculate the economic indicators. TotBio = total biomass, Prop OF = proportion overfished, DemBio/PP = demersal fish biomass relative to primary production, Bio/PP = total biomass relative to primary production, MTLCat = mean trophic level of catch, MTLBio = mean trophic level of biomass, PelCat = pelagic catch, DemCat = demersal catch, DemPelFish = demersal to pelagic fish ratio, CatBio = catch to biomass ratio, and TotCat = total catch.

midwater trawl sector under the decreased effort scenario were driven by changes in New England, but increases in sales from this sector under the increased effort scenario were due to increases in the Mid-Atlantic (**Figures 5**, **6**). Because some of the nuances of these changes are associated with the assumptions made when mapping fleets to ports, we do not overly highlight these and instead focus on system-wide indicators that are more robust. However, these results serve to demonstrate that the consequences of management scenarios for individual sectors may be felt disproportionately through the region in addition to across sectors, an issue of importance to managers.

The changes to economic indicators from NERIOM are consistent with the changes in landed catch from the Atlantis model. Under the decreased effort scenario, the reduction in total catch from the baseline scenario means there is less demand from seafood processors and traders for goods and services required to handle the catch. Similarly, the increase in total catch from the increased effort scenario provides more business for seafood processors and subsequently more demand for industries supplying these sectors. However, the proportional increase in sales under the increased fishing effort scenario (+19%) does not match the increase in total landed catch (+46%), as increases in yield under this scenario are generally for lower value species groups. Taking the fishing fleet sectors alone, the economic system consequences mirror those of the regional indicators shown in **Figure 3**, even though individual sectors showed more varied responses to the scenarios and by region (**Figure 4**).

## DISCUSSION

We linked the output of a marine ecosystem model to a regional economic model for the Northeast United States, and estimated the impacts of simple management strategies on both ecological and economic indicators. The value of using a coupled modeling approach to quantify these effects is that it is possible (a) to make use of extant tools facilitating relatively rapid analysis and (b) to retain the detail associated with both the ecological and economic systems. Such detail is often lost when using a single model approach that bridges across disciplines and spatial and temporal scales (e.g., Fulton, 2010). Coupling existing models, even in a one-way fashion as we did here, greatly facilitates the simultaneous consideration of multiple management objectives.

A key element of ecosystem-based management of marine resources is the development of analytical tools for quantifying tradeoffs associated with human activities (Leslie and McLeod, 2007; Link, 2010). We quantified tradeoffs among ecological groups associated with alternative fishing scenarios, with shifts in ecosystem composition and resulting changes to both magnitude and composition of landed catch. Under our increased effort scenario, the total amount of fisheries landings increased (but not linearly with effort), leading to a higher proportion of species groups overfished compared to the baseline and reduced effort scenarios. At the system level, our increased fishing effort scenario increased sales, income, and employment, yet there was very little change to the average income. This implies that the dynamics of the entire ecological and economic system may have some inherent stability despite individual taxa or fleet dynamics (Link, 2018). Our analyses suggest that the economic impacts of fishing scenarios on individual industry sectors, particularly harvesting sectors, can be large and variable even though system level properties were predictable and robust. These large, systemic effects were observed even during the relatively short time period for our model projections; consequences would potentially be amplified if viewing these scenarios over the long term. This highlights the need to consider relative resilience of individual system components in addition to systematic indicators when evaluating management strategy performance.

There was a disproportionate effect of the fishing scenarios on the ecological versus economic components of the modeling framework. While the economic indicators tracked in the direction expected (increased landed catch led to more dollars and jobs, decreased landed catch resulted in less value and fewer jobs), the magnitude of the change at the system level was not the same as for the ecological system. For the increased fishing effort scenario, economic gains were smaller than the ecological losses in terms of proportionality, with this scenario appearing to have greater magnitude of effects on the biological system. In this scenario, values for ecological indicators approached threshold values known to be associated with perturbed systems (e.g., Shin et al., 2010; Large et al., 2013; Pranovi et al., 2014). There are also undoubtedly threshold values in economic indicators that would define departure from safe operating space (e.g., levels of revenue from a sector that would force it to go out of business), which would also constrain the feasibility of management options. However, these thresholds have not yet been fully developed.

Although welfare analysis would be necessary to understand optimal tradeoffs, this analysis suggests some potential for gains from management regimes aimed at system stability (Link, 2018).

These disproportionate impacts pose questions as to what policy objectives to prioritize. The primary economic impacts associated with the fishing effort scenarios were on the fishery sectors, with smaller impacts on jobs and earnings at the scale of the Northeast United States, consistent with the results of Kaplan and Leonard (2012). Our coupled model is a tool to at least address quantitatively what the changes associated with alternative actions are for different sectors, which seems preferable to ignoring such questions and tradeoffs even though they exist. Undoubtedly, these questions and tradeoffs are being made, even if implicitly (e.g., Stephenson et al., 2017). Consequently, this tool can help elicit policy priorities and the viability (or not) of actions given constraints of satisfying management objectives, which would include societal goals such maintaining employment in individual business sectors. Clearly quantifying the tradeoffs among a range of objectives (by using MSE) will be critical for advancing EBFM implementation.

Input–output models can be useful to elucidate a broad suite of system dynamics and impacts, with considerable detail on linkages among industries. The spatial resolution of the NERIOM model provides information at scales relevant to fisheries management decision-making in the Northeast United States, although we mainly focused on the larger regional scale here. Because the input–output analysis is static, there is no feedback mechanism from the economic back to the ecological sub-model. For example, the model does not include market corrections such as price changes or behavioral responses to changes in supply, even though both economic theory and empirical analysis (e.g., Lee and Thunberg, 2013) suggest these might occur. Effects are best interpreted as the immediate/short-term impacts of the change being analyzed. Dynamic economic models, such as computable general equilibrium models (CGE; e.g., Jin et al., 2012) can model behavioral responses to changes in the economy. A dynamic economic model feeding back into the ecosystem model might also enable investigation of non-linear effects such as regional sectors going out of business, or redistribution of fishing effort and associated employment and investment as a result of changes in seafood supply, be it due to fluctuations in biomass or changes in spatial availability of various fished species. Other, related methods to estimate the value of ecosystem goods and services with respect to ecosystem status in response

to human use in a particular sector (e.g., Costanza et al., 1997) have their place. Other ecosystem goods and services could be derived from the modeling approach highlighted here (e.g., recreational opportunities and ecotourism). And it would certainly be plausible to pair the tradeoff analyses presented here with those obtained from other modeling approaches that might better represent other ecosystem goals (e.g., non-market value of taxa such as charismatic megafauna, biodiversity, etc.). We thus reiterate the important role of multidisciplinary datasets, multimodel inference, and multiple modeling objectives in tradeoff analysis that inform management. We also acknowledge that other, prior studies focused only on one species or one fleet and/or sub-region have indeed shown the benefits of using the input–output approach (e.g., Briggs et al., 1982; Steinback, 1999, 2004; Kirkley et al., 2011) that was then expanded to consider other factors. Of particular emphasis here is that, although it could always be expanded, we contrasted the responses of an entire system simultaneously, ensuring consistency in the treatment of ecosystem dynamics across a range of ecosystem goods and services.

We applied blanket multipliers to fishing effort across all fishing fleets in our scenarios. By presenting these results, it is not our intent to suggest that one should double fishing effort to increase catch (indeed the ecological costs associated with this are at odds with current fisheries policy objectives). Nor do we suggest that halving effort would not affect the economy (although our results suggest the cumulative impacts of doing so are disproportionately lower). Tradeoffs between conservation objectives and economic impacts were apparent, but these were non-linear. Rather, our scenarios provide some contrast for identifying non-linearities and second order, trickle-through effects that would have not otherwise been identified, highlighting why this type of coupled full system modeling needs to be done. Kaplan and Leonard (2012) considered a set of more plausible management options associated with some specific objectives, and Fulton et al. (2014) specifically included a stakeholder scoping process when developing objectives and management scenarios to evaluate. Alternative management scenarios that varied the level of change to mortality or fishing effort across fishing fleets and/or species groups could be applied to see what the tradeoffs of these management actions are. While additional detail may be required to focus on some combinations of management action and indicators of performance (for ecological, economic, and societal objectives), our scenarios show

that impacts do transfer and that these are not always linear or straightforward. Further, the tools to conduct such detailed analysis are extant for this region, and are growing around the world, highlighting the development of such MSE tools to explore a range of tradeoffs among various management objectives.

An advantage of our approach is that it is able to incorporate complexities of both ecological and economic system components and generally affords the ability to test management options (e.g., via MSE; Punt et al., 2016), particularly noting a range of responses across a suite of performance measures. A large number of performance measures have been suggested and used for MSE (Punt, 2017). Ideally these ought to reflect the full set of management objectives against which performance of options needs to be compared, which includes societal objectives. Thus, the combination of ecological and economic indicators is sorely needed. Summarizing results using integrative, systematic metrics rather than analyzing at the individual species or fishing fleet level makes it easier to visualize quantities and tradeoffs that appropriately reflect larger scale, strategic goals for management (e.g., Shin et al., 2010; Coll et al., 2016). Indicators reflecting performance with respect to societal objectives of fisheries are also increasingly available and calculated at a range of cultural and governance scales (e.g., Melnychuk et al., 2012; Colburn et al., 2016; Costello et al., 2016). Coupled modeling approaches provide a formal means of calculating values for economic indicators of changes associated with human and environmental pressures within frameworks typically used to derive indicators quantifying biological management performance. Our approach thus offers opportunity to extend the range of performance measures considered when evaluating the effects of management strategies that extend beyond fishing (though that was the focus of our analyses). Including the indicators derived from this coupled approach into the performance measures considered in a more formal MSE that includes feedbacks from human system of activities and management decisions on socio-ecological system dynamics is a natural extension of this work.

Coupled ecological-economic models can help to identify system-level responses to management alternatives in a manner otherwise impractical. The coupled approach presented here has the detail necessary to identify which fleets and communities warrant additional investigation through more refined modeling to more rigorously assess changes in welfare and benefits. Our results show how information on both the economic and ecological consequences of alternative management actions can more clearly illustrate benefits and pitfalls of alternative management options. As we continue to implement EBFM, it is the judicious use of extant tools as noted herein that will escalate broader, systematic management and serve to better identify the management choices needed.

## DATA AVAILABILITY

fmars-06-00133 March 22, 2019 Time: 16:52 # 11

The datasets generated for this study are available on request to the corresponding author.

## AUTHOR CONTRIBUTIONS

GF, RG, and JL ran the ecological model. GD and SS ran the economic model. GF and GD coupled model outputs. JL and GF conceptualized the work. GF, JL, and GD wrote the initial draft. All authors contributed to subsequent drafts. JL procured funding.

## REFERENCES


## FUNDING

JL obtained a Fisheries and the Environment project to support GF for part of this work.

## ACKNOWLEDGMENTS

We thank Sarah Gaichas and Scott Large for their good advice in discussions leading up to this manuscript. The manuscript was improved following the helpful comments of Cam Ainsworth, Isaac Kaplan, and the reviewers.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00133/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 Fay, DePiper, Steinback, Gamble and Link. 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.

fmars-06-00133 March 22, 2019 Time: 16:52 # 12

Edited by:

Dorte Krause-Jensen, Aarhus University, Denmark

#### Reviewed by:

Henrique Cabral, Irstea Centre de Bordeaux, France Kevin David Friedland, National Oceanic and Atmospheric Administration (NOAA), United States

#### \*Correspondence:

Martina Kadin martina.kadin@nrm.se; mkadin@uw.edu orcid.org/0000-0002-4706-7267

#### †Present address:

Martina Kadin, Swedish Museum of Natural History, Stockholm, Sweden Maria Angeles Torres, Instituto Español de Oceanografía, Centro Oceanográfico de Cádiz, Cádiz, Spain ‡orcid.org/0000-0002-6991-7680 §orcid.org/0000-0003-4910-5236 ||orcid.org/0000-0003-1803-0622 ¶orcid.org/0000-0003-4922-1894 #orcid.org/0000-0001-7780-1322

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 31 January 2019 Accepted: 24 April 2019 Published: 21 May 2019

#### Citation:

Kadin M, Blenckner T, Casini M, Gårdmark A, Torres MA and Otto SA (2019) Trophic Interactions, Management Trade-Offs and Climate Change: The Need for Adaptive Thresholds to Operationalize Ecosystem Indicators. Front. Mar. Sci. 6:249. doi: 10.3389/fmars.2019.00249

# Trophic Interactions, Management Trade-Offs and Climate Change: The Need for Adaptive Thresholds to Operationalize Ecosystem Indicators

#### Martina Kadin<sup>1</sup> \* † , Thorsten Blenckner<sup>1</sup>‡ , Michele Casini2§, Anna Gårdmark3|| , Maria Angeles Torres<sup>3</sup>†¶ and Saskia A. Otto1,4#

<sup>1</sup> Stockholm Resilience Centre, Stockholm University, Stockholm, Sweden, <sup>2</sup> Institute of Marine Research, Department of Aquatic Resources, Swedish University of Agricultural Sciences, Lysekil, Sweden, <sup>3</sup> Department of Aquatic Resources, Swedish University of Agricultural Sciences, Öregrund, Sweden, <sup>4</sup> Institute of Marine Ecosystem and Fishery Science, Center for Earth System Research and Sustainability, University of Hamburg, Hamburg, Germany

Ecosystem-based management (EBM) is commonly applied to achieve sustainable use of marine resources. For EBM, regular ecosystem-wide assessments of changes in environmental or ecological status are essential components, as well as assessments of the effects of management measures. Assessments are typically carried out using indicators. A major challenge for the usage of indicators in EBM is trophic interactions as these may influence indicator responses. Trophic interactions can also shape tradeoffs between management targets, because they modify and mediate the effects of pressures on ecosystems. Characterization of such interactions is in turn a challenge when testing the usability of indicators. Climate variability and climate change may also impact indicators directly, as well as indirectly through trophic interactions. Together, these effects may alter interpretation of indicators in assessments and evaluation of management measures. We developed indicator networks – statistical models of coupled indicators – to identify links representing trophic interactions between proposed food-web indicators, under multiple anthropogenic pressures and climate variables, using two basins in the Baltic Sea as a case study. We used the networks to simulate future indicator responses under different fishing, eutrophication and climate change scenarios. Responsiveness to fishing and eutrophication differed between indicators and across basins. Almost all indicators were highly dependent on climatic conditions, and differences in indicator trajectories >10% were found only in comparisons of future climates. In some cases, effects of nutrient load and climate scenarios counteracted each other, altering how management measures manifested in the indicators. Incorporating climate change, or other regionally non-manageable drivers, is thus necessary for an accurate interpretation of indicators and thereby of EBM measure effects. Quantification of linkages between indicators across trophic levels is similarly a prerequisite for tracking effects propagating through the food web, and, consequently, for indicator interpretation. Developing meaningful indicators under climate change calls for iterative indicator validations, accounting for natural processes such as trophic interactions and for trade-offs between management objectives, to enable learning as well as setting target levels or thresholds triggering actions in an adaptive manner. Such flexible strategies make a set of indicators operational over the long-term and facilitate success of EBM.

Keywords: zooplankton, forage fish, networks, coupled Generalized Additive Models, Baltic Sea, Marine Strategy Framework Directive

## INTRODUCTION

fmars-06-00249 May 18, 2019 Time: 16:3 # 2

Reduced impacts of human activities and sustainable use of marine natural resources is an urgent calling when other severe pressures on coastal and ocean systems, such as climate change, can only be curbed on long time-scales (Dayton et al., 1995; Worm et al., 2006; Field et al., 2014; Cloern et al., 2016). Ecosystem-based management (EBM) makes sustainable use achievable by taking an integrated perspective on multiple uses and different components of ecosystems (Rosenberg and McLeod, 2005; Leslie and McLeod, 2007). To ensure success, EBM needs to include initial and regular update assessments of the status of the ecosystem as well as evaluate the response to management measures and their efficiencies. Integrated ecosystem assessments (IEAs) provide a scientific basis for decision-making within EBM (Levin et al., 2009) where carefully selected indicators constitute the basis for status assessments and management strategy evaluations.

Typically, IEAs make use of several indicators to assess the state of the ecosystem, each representing some component or aspect of the ecosystem (e.g., Ottersen et al., 2011). The strong ecological linkages present in many ecosystems will often make indicators interlinked, e.g., in a food web due to species interactions (Håkanson and Blenckner, 2008). Therefore, management measures or pressures do not only act on one monitored indicator directly, but also indirectly on others (Torres et al., 2017). Consequently, when developing indicators and IEA frameworks, it is rarely sufficient to understand relationships between single pressures and single indicators, but joint analyses of multiple indicators are needed. This is true particularly for food-web indicators that represent different trophic guilds, which may integrate direct as well as indirect effects of pressures propagating through the food web. Overfishing of predatory fish for example, often results in marked increases of pelagic forage fish and benthic macroinvertebrates, in turn reducing biomass of zooplankton and subsequently consumption of phytoplankton (Frank et al., 2005; Casini et al., 2008). Other trophic cascades induced by fisheries have led to loss of, e.g., kelp forests as well as sea grass meadows and make ecosystems sensitive to disturbance (Jackson et al., 2001). Such mechanisms may amplify or exaggerate effects of eutrophication, thereby interfering with efforts to reduce nutrient loads, as well as signals of their success, typically tracked by indicators. Conversely, bottomup dynamics can result in positive impacts of eutrophication on species benefitting from higher ecosystem productivity (Laursen and Moller, 2014). Trophic interactions may thus create trade-offs between management objectives, constrain achievable target levels for indicators or affect evaluations of management strategies as well as of specific measures to move the ecosystem toward a healthy state (Shelton et al., 2014; Punt et al., 2016).

Climate conditions are key pressures acting on ecosystems, with potential to influence management pathways or the effort required to improve their status (Niiranen et al., 2013). Quantitative evaluations of the interplay between climate and management measures on indicators are yet sparse, despite international and national legislations requiring the implementation of EBM and the large number of frameworks to develop operational indicators. Projected climate change may further amplify or dampen effects, and is thus necessary to account for when assessing the benefits of management measures (Lynam and Mackinson, 2015). Existing studies of indicators that account for climate change typically focus on effects of single drivers (Gårdmark et al., 2013), but EBM strives to balance multiple objectives, and management measures motivated by different objectives will thus be implemented simultaneously rather than in isolation. How do indicators respond to management alternatives when measures affecting top-down and bottom-up processes are implemented simultaneously under a changing climate? Indirect effects mediated by food-web interactions may be particularly difficult to foresee and could have a substantial impact on the interpretation of indicators.

In this study, we examine how indicators track effects of management strategies targeting different pressures on marine food webs under climate change, accounting for the interactions among species that constitute them. Using food-web indicators proposed for the Baltic Sea under the European Marine Strategy Framework Directive (MSFD), we apply advanced statistical modeling tools to identify indicator networks (Llope et al., 2011; Blenckner et al., 2015; Lynam et al., 2017, **Figure 1**). The indicator networks allow us to evaluate links between indicators, suggesting potential for cascading effects of management measures, and if these links magnify or counterbalance indicator responses to (single and multiple) pressures. Our approach illustrates how effects of climate change may interfere with effects of potential measures as manifested in the indicators. Such interactions change the interpretation of indicator responses in relation to reference points. By including indicators representing different trophic guilds and simultaneously modeling management measures targeting different pressures, our indicator networks aid in identification and quantification of potential trade-offs in EBM. We discuss strategies in EBM to detect and handle conflicts between objectives that arise due to trophic interactions, as well as modulating effects of climate change. Adaptive targets and thresholds are proposed as one approach to these challenges.

## MATERIALS AND METHODS

## Study System and Selected Zooplankton and Fish Indicators

The study focused on food-web indicators of trophic functions in the pelagic food web of the Central Baltic Sea, a relatively simple brackish-water ecosystem, where recent impacts of fisheries and eutrophication in combination with climate factors have resulted in substantial ecological changes (Österblom et al., 2007). A regime shift occurred in the Central Baltic Sea in the early 1990s with effects cascading through the food web (Casini et al., 2008; Möllmann et al., 2009). Environmental gradients are prominent in the Baltic Sea and the recent changes have had, quantitatively and qualitatively, different impacts in the basins (Casini et al., 2011). We therefore used indicators of food-web status developed separately for the Bornholm and Gotland Basins (corresponding to ICES subdivisions 25 and 28, respectively, **Supplementary Figure S1**). At the trophic levels of zooplanktivorous and piscivorous fish, indicators of trophic functions correspond to single or only a few species in the speciespoor system of the Baltic Sea (ICES, 2015c,d; Torres et al., 2017). The zooplankton community includes a substantially larger number of species and we derived indicators of the community representing aspects of quantity and quality of food for upper trophic levels. The indicators were considered for assessment under the 'Baltic Sea Action Plan' and descriptor 4 – food webs – of good environmental status in the MSFD and showed a good performance when studied individually (HELCOM, 2013a,b; Otto et al., 2018). Assessments under descriptor 4 are based on trophic guilds and our indicators covered three guilds: apex fish predators, planktivorous fish, and secondary producers; and three of four assessment criteria (EU Decision 2017/848; ICES, 2015b).

## Historical Data and Indicator Time Series Construction

Data on piscivorous and planktivorous fish, representing the apex predator guild and the planktivore trophic guild, respectively, were obtained from the autumn Baltic International Acoustic Survey (BIAS, ICES, 2015a) and historical acoustic surveys by the Swedish University of Agricultural Sciences and the former Swedish Board of Fisheries. We calculated fish indicators based either on actual abundance (for sprat Sprattus sprattus and herring Clupea harengus) or modeled Catch Per Unit of Effort (CPUE) (for cod Gadus morhua; Casini et al., 2019) in the surveys (Cod, Sprat, Herring – collectively referred to as abundance-FI) and two indicators based on body size (Small Prey Fish, SPF, forage fish <10 cm; and Large Predatory Fish, LPF; piscivores > 38 cm – as size-based FI), according to the approach described in Torres et al. (2017). One initially considered indicator, CPUE of three-spined stickleback Gasterosteus aculeatus (Stickleback), was tested as a pressure variable, representing additional competition or predation depending on trophic level. Zooplankton-based indicators, corresponding to the secondary producer trophic guild, are collectively referred to as ZPI. This set of ZPI included total zooplankton abundance (TZA), zooplankton mean size (MS), and abundance ratio of cladocerans to copepods (excluding nauplii, RCC). Bornholm Basin ZPI were based on summer samples (average of July–August) taken at the BY5 station (55.25◦N, 15.98◦E) collected by the Leibniz Institute for Baltic Sea Research Warnemünde, Germany. Gotland Basin ZPI were calculated from summer samples (average of July–August) taken at multiple stations in seven ICES rectangles (42G8, 42G9, 42H0, 43G9, 43H0, 44G9, 44HO) close to station BY15 (57.32◦N, 20.05◦E) by the Institute of Food Safety, Animal Health and Environment (BIOR), Latvia.

Missing values in the ZPI, Sprat and Herring time series were replaced by interpolation – the average of the 2 years before and the 2 years after the year without sampling replaced the missing value. Data were missing in 1 year in the Bornholm Basin SPF time series and in 4 years in the Gotland Basin SPF series. Both time series had high variability and the numbers of years with missing data were relatively many in relation to the length of the time series. These factors increase the risk of introducing bias when replacing missing values by interpolation. We therefore

opted for not replacing the missing values and instead removed these years from the analysis. The abundance and ratio ZPI as well as all FI, except SPF, were ln-transformed prior to analysis. The SPF indicator was transformed using the formula ln(SPF + 1) as 1 year in the original data had a value of 0.

After combining the time series (see **Supplementary Information: Appendix I**), we had datasets covering 1979–2008 in the Bornholm Basin, and 1979–2011 in the Gotland Basin, for developing indicator networks based on ZPI and abundance-FI. For networks including size-based FI the datasets covered the period 1984–1992, 1994–2008 in the Bornholm Basin, and 1984–1990, 1992, 1994, 1996, 1998–2011 in the Gotland Basin.

## Pressure Data

We evaluated climate (summer sea surface temperature (TempSummer), winter deep-water salinity (SalinWinter) and, for cod, oxygen concentration (O2Cod), each with speciesspecific time-lags based on prior knowledge), fishing mortality (FCod, FSpr, FHer), and chlorophyll a (ChlSummer, a proxy for primary production and here, thereby for eutrophication and nutrient load) as pressures potentially affecting indicators (**Figure 2**). Pressure data were obtained from the Baltic Environment Database at the Baltic Nest Institute, Sweden; IFM Geomar (Helmholtz-Zentrum für Ozeanforschung Kiel), Germany and ICES (ICES, 2015a), see further **Supplementary Information: Appendix I**.

Climate-related pressures (TempSummer, SalinWinter, O2Cod) were viewed as non-manageable pressures at the regional level, as future climate trajectories will largely depend on decisions made at other scales. Fisheries and eutrophication were considered manageable, as management decisions are made within regional governance structures or by national bodies incorporating regional agreements or plans (e.g., HELCOM, 2013a). O2Cod is an effect of both large-scale climate variations influencing water inflows from the North Sea to the Baltic Sea and regional eutrophication, and as such is partly related to ChlSummer and was not included as pressure in the future predictions.

## Statistical Modeling

Our statistical modeling approach included three steps (**Figure 1**): (1) Fitting statistical models for each indicator and basin, including potential pressures and links to other indicators identified a priori based on existing knowledge and plausible relationships (**Figure 2**). (2) Building indicator networks by combining the relationships identified in 1. (3) Simulating the effects of different management scenarios under climate change on the indicators using the statistical indicator networks. Statistical modeling was carried out in R 3.0.2 and 3.2.4 (R Core Team, 2016).

### Development and Selection of Indicator Models

The statistical models for individual indicators were Generalized Additive Models (GAMs) (Wood, 2006) or their threshold formulation (tGAMs) (Ciannelli et al., 2004), developed using the mgcv library (Wood, 2006).

A variance inflation factor (VIF) test was performed on each set of potential covariates (pressures and links to other indicators) to detect and avoid issues of multicollinearity (Zuur et al., 2010). We excluded one variable at a time (the one with the highest VIF) when any VIF > 3 until we had reduced sets of covariates with all VIF ≤ 3. All a priori identified covariates were thus tested for effects in at least one set.

We did not model autoregressive effects but included only external pressures as explanatory variables, as our focus

was not on autoregressive processes and several indicators were aggregated across species or groups. No interactions between explanatory variables were evaluated due to the size of the datasets and tGAMs representing a form of interaction between covariates.

### **Selecting GAMs**

All possible additive combinations of covariates were compared for each individual indicator in the GAM analyses. GAM comparisons were made based on Generalized Validation Criterion (GCV; Wood, 2006), where a lower GCV means a more parsimonious model. We checked model diagnostics of candidate models (see **Supplementary Information: Appendix I**), and ultimately selected the statistically best model that had sensible ecological effects (i.e., not contradicting existing ecological knowledge, for example a positive direct effect of higher fishing mortality or of a competitor).

## **Selecting tGAMs**

For tGAMs, we constructed starting models, using as large as possible sets of covariates, taking into account VIF results and ensuring that the model degrees of freedom did not exceed the length of the time series. Models including size-based FI had simpler structure (**Figure 2**), making it feasible to try all potential pressures and use threshold variables defined a priori. Models including abundance-based fish indicators had a higher number of potential trophic links and threshold variables, so we used the results of the GAM analysis and results from single-pressure analyses carried out by Otto et al. (2018) to inform choice of covariates and threshold variables (see details in **Supplementary Information: Appendix I**).

Each starting tGAM was reduced in a step-wise manner, by excluding the explanatory variable with highest p-value until all explanatory variables had p < 0.05, after which model GCV was minimized to identify the most parsimonious model. We examined the effects of explanatory variables above and below the threshold, confirming that these indeed were qualitatively different. If not, the model was simplified by removing the threshold effect on that explanatory variable until variables with thresholds had qualitatively different dynamics. After examination of model diagnostics, we calculated the genuine cross validation score (genuine CV; Ciannelli et al., 2004) of the tGAM, which equals the average squared leave-one-out prediction errors and accounts for the grid search needed to find the value of the threshold. This was compared to the genuine CV of the corresponding GAM, i.e., with the same model structure except for the threshold. If the tGAM had a lower genuine CV, it was added to the list of candidate tGAMs. After completing this process with all starting models, we compared the candidate tGAMs using GCV and examining the ecological relationships. As for GAMs, we selected the statistically best tGAM that had sensible ecological effects.

### **Selecting final model**

If the selected tGAM had a higher genuine CV than its corresponding GAM, i.e., being a less suitable model, a GAM would be the best model for the indicator and we picked the selected GAM as the final model. If the selected tGAM had a lower genuine CV than its corresponding GAM, i.e., being a more suitable model, the selected tGAM was picked when the covariates in the models were the same as the models' genuine CV are comparable in this situation. If the selected GAM and tGAM differed in their covariates we could not use the genuine CV to choose the final model. Instead, we picked the model with the most simple structure (a GAM in all cases).

We did not find statistically sound models with reasonable ecological effects among models for the Cod and LPF indicators in the Gotland Basin, or for Stickleback in the Bornholm Basin. In these cases, the observed indicator time series was only used as a covariate when relevant for the indicator networks.

### Construction of Indicator Networks

The selected indicator models were coupled into an indicator network, where the dynamics were driven by the external covariates (environmental and climate variables, fishing) and trophic interactions as identified by the individual indicator models. We used the predicted value of an indicator to feed into any other model component where it had an effect, until all indicators in the network had been predicted. Noise, in the form of resampled residuals from the individual indicator models, was added to test the robustness of predictions and generate confidence intervals.

Indicators at two trophic levels simultaneously affected each other in some of the networks (**Figures 3**, **4**). In this case, we started the coupling by adding the observed value of one indicator into the model predicting the other, then using the modeled value to predict the first indicator and repeating until convergence was reached.

Lastly, we validated the indicator networks by predicting the last 3 years of the time series – that were not used for fitting individual indicator models – and compared predicted versus observed values. Poor performance during the validation period did, however, not disqualify networks from simulations of future scenarios, as we were interested in seeing if robustness of relationships affected conclusions.

### Simulations of Management Alternatives and Climate Change

Future scenarios for regionally manageable pressures covered years 2012–2040 and included high and low levels of fisheries exploitation for cod and clupeids and three levels of nutrient loads: reductions following the Baltic Sea Action Plan (HELCOM, 2013a), reference levels (PLC 5.5, HELCOM, 2015) and increase due to intensified agriculture in the catchment area. A realistic climate change scenario, corresponding to SRES emission scenario A1B, was simulated by two global models, HadCM3 and ECHAM5, to illustrate uncertainty (Gordon et al., 2000; Roeckner et al., 2006). Regionally downscaled climate variables from the two climate projections and nutrient load scenarios were modeled by the coupled physical-biogeochemical model BALTSEM (Gustafsson, 2003; Gustafsson et al., 2012) to simulate these future pressures at the basin scale. BALTSEM runs thereby generated simulated time series of sea surface temperature, deep-water salinity and of chlorophyll α subsequently used in


TABLE 1


Summary of model selection and coupling of individual indicator model components

 into indicator networks for the Bornholm (BB) and Gotland Basins (GB) in the Baltic Sea.

mortality

(species-specific);

 eutroph.,

eutrophication.

∗Indicator not modeled, only linked as driver for another indicator in this network. Gray, no link established

 to any other indicator in this set of food-web indicators.

our scenario simulations (**Supplementary Figure S2**). Sensible models for piscivorous FI in the Gotland Basin were not identified, and we instead constructed time series to investigate impacts of a range of future Cod levels on other indicators. Details about the climate projections, scenarios for nutrient load and fisheries exploitation as well as the BALTSEM model and Cod future time series are found in **Supplementary Information: Appendix I**.

The quantitative relationships of the networks and the scenario data were used to project the indicators by running a thousand Monte Carlo simulations for each indicator and year. Noise was added in each simulation by sampling from the residuals (from model component runs on observed data). We calculated a mean and 95% confidence intervals based on bootstrapping of estimated values for each indicator and year.

Effect sizes and interactions between pressures under the scenarios were evaluated by running multi-factorial ANOVA or Generalized Least Squares (GLS) models. Since simulated time series tend to show less stochasticity and higher autocorrelation GLS were applied when temporal autocorrelation was detected, using auto-regressive error structures of order 1 or 2, depending on the detected autocorrelation (Pinheiro and Bates, 2000). We started with a full model, including higher-order interaction terms, and applied a backward-selection based on Akaike's Information Criterion (AIC).

## RESULTS

## Individual Indicator Models

In the first step (**Figure 1**), sensible models were found for 19 of 22 food-web indicators (**Supplementary Table S1** and **Supplementary Figure S3**). GAMs were often sufficient to capture observed variation in the indicators, where nonlinear relationships were present in about half of all responses (**Supplementary Table S1**). Threshold formulations (tGAMs) performed better in a few cases, mostly related to Sprat dynamics (**Supplementary Table S1**).

## Indicator Networks

The majority of the individual indicator models suggested links between indicators (**Supplementary Table S1**), making it possible to couple individual indicator models to each other into eight indicator networks (**Table 1** and **Figures 3**, **4**). All networks reproduced the overall pattern of observed indicator time series used for fitting, but did not always capture temporal variation

(**Figures 3**, **4** and **Supplementary Figure S4**). Observations during the time periods used for validation were replicated well by some networks (e.g., **Figure 3B**), but other networks had worse performance when predicting data points not previously considered in the individual models (e.g., **Figure 3C**).

The complexity of indicator networks varied. For example, the network of size-based ZPI and FI in the Gotland Basin, had low complexity; with unidirectional trophic control and summer sea surface temperature being the only external pressure (**Supplementary Figure S5B**). Other models included multiple pressures, threshold effects and mixed trophic control (e.g., abundance ZPI and FI models in the Bornholm Basin, **Figure 3A**).

Links between indicators representing different trophic levels were key explanatory factors. On the other hand, links between indicators at the same tropic level – corresponding to competition, which we investigated for planktivore-based FI – were not detected (**Supplementary Table S1**). Sprat and the size-based SPF were often linked to ZPI as well as piscivorous FI, making tri-trophic indicator networks the most common configuration (**Figures 3**, **4**). There was only one case of a link between Herring and another indicator. The piscivorous FI Cod and LPF always exerted a top-down control on planktivore-based FI, except in one interaction network (**Figure 3B**). Direction of coupling between ZPI and planktivore-based FI differed between networks, and bidirectional linkages were found as well (**Figures 3, 4** and **Supplementary Figure S5**). Five of the indicator networks included a direct effect of a manageable pressure on one indicator; which in turn had a relationship with another indicator (**Figures 3, 4B** and **Supplementary Figure S5A**), suggesting that trophic interactions could introduce indirect links between pressures and indicators. The same type of pattern involving climate variables existed in five networks.

## Impacts of Regionally Manageable Pressures

A few key pressure variables had similar effects across the different indicator networks. Chlorophyll a – as a proxy for eutrophication – emerged as a central pressure variable in the Bornholm Basin where we found significant relationships with all three types of ZPI (**Supplementary Table S1**). In the Gotland Basin, only TZA responded to this pressure (**Supplementary Table S1**). Indicators responding to fishing pressure variables were relatively fewer. In the Bornholm Basin significant effects of FCod on Cod and LPF (direct effect, **Supplementary Table S1**) were detected, which indicator network structure suggested would cascade onto Sprat (as an indirect effect, **Figures 3A,B**). FHer affected Herring in the Gotland Basin (**Figure 4B**). Ecologically meaningful (i.e., negative) effects of FSpr on sprat were not detectable.

The scenario simulations highlighted indirect and cascading effects of management measures suggested by the structures of the indicator networks (**Tables 1**, **2** and **Figures 5**, **6**). While the network structures suggested that management measures may have indirect effects on quite many indicators, the simulations revealed that detectable indirect effects involved fewer indicators.


additive effects of climate. The line and radar charts show significant differences between two or more scenarios, detected by GLS analysis, in the Bornholm Basin, where letters (A–C) correspond to the models shown in Figure 3. Blue represents climate projections from the HadCM3 model and green represents projections from the ECHAM5 model. Black indicates no effect of climate on the indicator. Each axis in the radar charts represents one fishery, cod stock or eutrophication scenario. The outer ring in radar charts correspond to the maximum effect size, the inner ring corresponds to minimum effect size and the center 90% of the minimum effect size. The differences between scenarios are thus small, when the two rings are close to each other. Interactions between climate and manageable pressures were found for the lower trophic levels, but not the piscivorous level where effects were additive. Future climate conditions had consistent effects across all indicators in the Bornholm Basin. For clear illustration of the effects of climate, we only show the interaction between climate and one pressure for Sprat in (A), but effects of a second pressure, not interacting with climate, are not shown. Supplementary Tables S1, S2 and Supplementary Figures S5–S11 present complete results of the analysis. TZA, total zooplankton abundance; MS, zooplankton mean size; RCC, abundance ratio of cladocerans to copepods (excluding nauplii); SPF, small prey fish; LPF, large predatory fish; F, fishing mortality (species-specific, FClup refers to the clupeids sprat and herring); MSY, maximum sustainable yield; ChlSummer, chlorophyll a during summer; TempSummer, summer sea surface temperature; SalinWinter, winter deep-water salinity (the latter two with species-specific time-lags).

In the Bornholm Basin, we found significant indirect effects of nutrient input on Sprat and SPF values, two planktivore FI (see **Supplementary Table S2**). Both indicators showed lower values under decreased nutrient input (**Figure 5B** and **Supplementary Figures S6, S7**), suggesting a strong bottom-up control of these interlinked indicators. However, when accounting for the coupling of Sprat to the ZPI RCC, the nutrient effect was reversed (**Figure 5C** and **Supplementary Figure S8**). In the Gotland Basin, the size-based ZPI MS was affected by a top-down effect from clupeid fisheries, acting via Herring (**Figures 4B**, **6B**). While this effect was significant, it had a weak response (**Table 2**).

No ecologically meaningful model was identified for Cod in the Gotland Basin, but the indicator networks showed a topdown effect of Cod on Sprat, in turn affecting the ZPI TZA as well as RCC, in their respective networks (**Figures 4A,C**). Any pressure acting on Cod could thereby have cascading effects on lower trophic level indicators in the Gotland Basin.

Interactions between regionally manageable pressures, i.e., an indicator responding to two or more pressures, were rather sparse: these were only found for Sprat in the coupled model of abundance ZPI and FI in the Bornholm Basin. Here, we found a cascading effect of cod fisheries, which was modulated by the threshold effect of clupeid fishing mortality: The positive

effect of higher cod fishing pressure, leading to lower cod stock size and hence predation pressure for sprat, occurred only under reduced clupeid fisheries (i.e., at 0.5∗FMSY) and was even reversed when increasing clupeid fishing pressure to FMSY (**Figure 5A** and **Supplementary Figure S6**). Other than this interaction, the simulations suggested that the structural links between indicators across trophic levels did not result in detectable trade-offs between management objectives, for the pressures and indicators we studied.

## Role of Climate and Interaction With Management Measures

Future climate was projected to have marked impacts on the performance of indicators and indicator relationships to pressures (**Figures 5**, **6**). In the Bornholm Basin indicator values were overall higher under the HadCM3 model (projections resulted in overall higher temperatures, higher salinity and lower Chl a than in the projections from ECHAM5, see **Supplementary Figure S2**). The effect was small on TZA and LPF, but there were strong effects on all other indicators (**Supplementary Table S2** and **Figure 5**). The two climate projections had less pronounced impacts on the results for the Gotland Basin (**Supplementary Table S3**). Effects were found on all three ZPI and on Herring. In this basin, there was no consistent difference between climate models with respect to indicator values (**Figure 6**).

Importantly, we found a striking pattern in terms of how climate modified responses to other pressures (**Table 2** and **Figures 5**, **6**): the climate variables interacted almost exclusively with the nutrient scenarios. Climate variables either modulated the magnitude of the indicators' response to nutrient load reductions, as found for Sprat (**Figure 5C** and **Supplementary Figure S8**) and the ratio ZPI RCC (**Figure 5C** and **Supplementary Figure S8**), or counteracted the nutrient load effect on the abundance-based ZPI TZA (**Figure 5A** and **Supplementary Figure S6**; with a very small effect: **Figure 4A** and **Supplementary Figure S9**).

## DISCUSSION

## Understanding Indicator Responses Under Cumulative Pressure Regimes

A key element needed for traditional management to evolve into agile EBM is an holistic approach that recognizes the full array of interactions between single species and ecosystem components (Slocombe, 1993). Yet, when it comes to developing indicators as tools for status assessment of marine food webs and management strategy evaluations, such an integrated approach is often ignored for wider ranges of ecosystem- and resource-use objectives than fisheries (Sainsbury et al., 2000). Independent of whether indicators are based on single species, species groups

or aggregated metric such as mean community size (Teixeira et al., 2014), they will integrate system-specific interactions and environmental effects (Torres et al., 2017). To ensure accurate interpretation of indicator responses to manageable as well as unmanageable pressures, and to account for cascading effects of management measures, potentially resulting in conflicting management objectives, it is essential to disentangle and quantify such interactions. An approach that integrates several pressures and multiple indicators representing different trophic guilds is thus needed when developing food-web indicators. In our case study, most long-term observed trends of indicators were explained by the combined effect of system-internal variables, i.e., other food-web indicators, and external variables relating to fishing pressure, eutrophication or climate. These results were also independent of the type of indicator (single species-based or aggregated) or trophic level. We did, however, find that indicator responses differed between the Baltic Sea basins.

## Appropriate Spatial Scales for Indicators and Evaluations

There was substantial variation in indicator responses to pressures, but in some cases clear patterns emerged. In one basin (Bornholm), responses to climate were consistent across all indicators despite substantial differences in the performance of indicator networks. In the other basin (Gotland), responses to climate were not consistent while differences in performance between indicator networks were smaller. These types of results confirm the need to carry out system-specific or spatially explicit performance evaluations (Shin et al., 2018).

However, in the Gotland basin, we struggled to identify ecologically sensible models for the two indicators of the apex predator guild, Cod and LPF. This could be due to missing an important covariate or a mismatch between data (indicator or covariate) and spatio-temporal dynamics (Bartolino et al., 2017). Fish-based indicators (for the species studied herein) could potentially be more meaningful if they instead are estimated at the broader central Baltic Sea level. The most abundant fish species in the Baltic Sea – which the indicators are based upon – move seasonally and have occupied different ranges over time, mainly regulated by density-dependence (Casini et al., 2011; Bartolino et al., 2017). Our results did neither reveal interactions between climate and fishing pressure as found in previous studies (Gårdmark et al., 2013). This may be an artifact of the different spatial scales (fishing mortality estimates were only available for the entire Baltic Sea and thus did not constitute basinspecific covariates).

These results suggest that alternating between region-wide and basin-specific application of indicators may therefore be required for comprehensive sets of indicators under cumulative pressure regimes.

### Advantages and Challenges With the Indicator Networks

Our indicator networks were based on statistical models (GAMs and tGAMs), enabling us to test for non-linear effects and even threshold effects of pressures, i.e., if ecosystem dynamics above and below the threshold value are qualitatively different. Threshold dynamics in marine ecosystems have been observed in many different systems, for example coral reefs (Knowlton, 1992) and pelagic food webs (Conversi et al., 2015). However, they are rarely accounted for in EBM indicator evaluations, perhaps due to quantitative validation schemes only being recently developed (Queirós et al., 2016; Otto et al., 2018). Non-linear effects were relatively common in our indicator networks, while threshold effects were more rare. Threshold dynamics and different ecosystem configurations have been identified in the Baltic Sea (Casini et al., 2009), so it was not surprising that a few networks functioned similarly. As a consequence of threshold dynamics some indicator responses may be challenging to connect to pressures, if not accounted for. For example, it was not possible to adequately model the Ratio of cladocerans to copepods indicator in one basin using GAMs only, but when incorporating threshold dynamics a wellperforming model was found. However, sometimes responses are difficult to interpret even after threshold dynamics have been identified, as exemplified by the intricate effects of cod and clupeid fisheries on the Sprat indicator in one of our indicator networks (see **Figure 5A** and **Supplementary Figure S6**). Such results may appear discouraging and could rise from spurious relationships between the particular monitoring time series used, but without this type of quantitative approach, anticipation of these relationships appears close to impossible.

A few of the indicator networks had worse performance during the validation period than the period used for fitting individual indicator models, suggesting that these relationships were not very robust (i.e., non-stationary). The insights from our study – regarding substantial impacts of climate on the indicators as well as abundant links between indicators representing different trophic levels – rest, however, on the results from several indicator networks.

The indicator network approach was well suited to identify broad-scale factors affecting interpretation of indicator values, such as climate, and the presence of trophic interactions modifying indicator interpretation as well as feasibility of achieving multiple objectives. As any quantitative approach it is sensitive to availability and quality of data. The lack of long-term datasets on key ecosystem components may impede application in data-poor systems. The Baltic Sea is relatively data-rich and we were able to include threshold dynamics, but not other types of interactions between pressures, as that would have led to overfitting of models. However, as a minimum, statistical modeling and indicator networks provide information that data availability or quality may be insufficient. For identifying operational foodweb indicators, the ability to examine relationships and responses at the scale of ecosystem assessments, or finer, constitute a major improvement over expert opinion and panaceas for all marine ecosystems.

The challenges highlighted by relatively data-hungry and potentially complex underlying models emphasize at the same time the main advantage of the indicator networks: the ability to test for multiple pressure effects and indicator linkages simultaneously, which may be essential as some relationships may only be revealed by modeling all time series jointly (Torres et al., 2017, this study: **Figure 3** and

**Supplementary Information: Appendix II**), and without an a priori specification of the shape of relationships. When relationships between indicators representing different trophic levels or guilds have been identified, the indicator networks enable further investigation of how effects of single or multiple management measures may propagate through the food web.

Another potential constraint of our presented approach can be time-consuming analyses related to the complexity of the food web and the number of indicators to test. The GAM/tGAM coupling approach to build the network relies on individual models for each indicator. With increasing food web complexity, there are more models to fit and couple, the more timeconsuming this approach becomes.

## Trophic Interactions and Achievable Management Targets

Relationships across trophic levels or between guilds may have strong implications for management measures, ranging from effectiveness of single measures to human-induced trophic cascades and conflicts between management objectives (Estes et al., 2011; Reilly et al., 2013).

Our example suggested bottom-up control of planktivorous FI in one basin and top-down control on most indicators in the other basin. This pattern has strong implications for management strategies. Interpreted together with the scenario simulations, the results point to that nutrient load reductions (as planned in the region, HELCOM, 2013a) may have limited effectiveness in one area (Gotland Basin, where top-down control was more prevalent) while these measures are likely to impact even the state of the indicators at intermediate trophic levels in the other area (Bornholm Basin, with bottom-up control). The net effect of management measures is difficult to predict with simple modeling approaches, as overall effects of multiple stressors may be additive, synergistic or antagonistic depending on the response level (population or community), trophic level as well as the stressors involved (Crain et al., 2008). This applies especially when different pressures are targeted simultaneously, as in our scenarios and many real cases. The net effect may, however, be estimated through simulations, after indicator relationships, across trophic levels or between individual species, have been quantified. Such quantification of propagating effects is essential for determining if trade-offs between management objectives are likely to occur.

### Potential Trade-Offs Between Management Objectives

A profound challenge in management governed by policies with multiple objectives is to account for trade-offs that can exist between individual management objectives (McClanahan et al., 2011). Ambitions to eliminate effects of eutrophication may have negative effects on biodiversity, for example: the abundance of benthic-feeding ducks declines as fewer, smaller and less-nutritious mussels are available to feed upon, concurrent with lowered nutrient levels in the ecosystem (Laursen and Moller, 2014). Trade-off directions in multi-objective EBM are tightly linked to the direction of trophic control in the ecosystem, which is not necessarily static or unidirectional (Lynam et al., 2017). During top-down forcing, management measures targeting indicators representing higher trophic levels, e.g., reductions in fishing pressure on piscivorous fish, are also likely to influence indicators representing a lower level of the food web, such as forage fish. Such effects on fish-based indicators were also detected in our case study (**Supplementary Tables S2, S3** and **Figure 5**). Trade-offs between objectives and difficulties defining targets and thresholds may emerge already when two trophic levels are involved. This includes cases when there should be no adverse effects on balances between trophic guilds or on population characteristics of fished stocks (see e.g., criteria for good environmental status in EU decision 2017/848).

Mixed trophic control may lead to substantial conflicts between objectives. If there are other forage fish predators, in such an ecosystem as above, that instead are food-limited, indicators of their status will illustrate adverse effects of the implemented measures for piscivorous fish. Reilly et al. (2013) describe how reduced fishing pressures on North Sea haddock and whiting are likely to result in increased abundance of these species, followed by higher predation pressure on sandeels reducing their abundance, with potentially detrimental impacts on kittiwake populations. This situation represents incompatible sets of objectives as the indicators and their target levels currently are defined (Reilly et al., 2013). We did not find indications of marked trade-offs between objectives in our case study, but our models included relatively few species. They did for example not include other forage fish predators, e.g., auks or seals, which potentially could give different results. When there is a risk of conflicts between objectives, multiple covariate models that couple several indicators, such as our indicator networks, have the advantage to allow for multidirectional trophic control and quantification of links between indicators representing different trophic levels, or guilds. Such quantification enables management bodies to anticipate trade-offs as well as to adjust targets and action thresholds to accommodate them, once priorities between the conflicting objectives have been decided politically (Martin et al., 2009).

## Management Under Modulating Effects of Climate Change

The influence of current and future climate is essential to consider when evaluating management options and setting thresholds, as large-scale environmental pressures may provide critical context for decisions in EBM, despite not being directly controllable in the short term (Samhouri et al., 2017). Our simulations depict this situation and illustrate strong modulating effects of climate change on the food-web indicators, and hence, on effects of other management measures. Both temperature and salinity were linked to the observed long-term development of most indicators we tested. Most importantly, future climate magnified and in other cases interfered with effects of management measures in our simulations. This is expected given the links between indicators representing trophic levels in the indicator networks. In our example, interactions between climate and nutrient

load reductions were suggested, affecting the development of zooplankton-based indicators across regions. For the aggregated indicator TZA, the results suggest that this interaction has even the potential to counteract effects of eutrophication mitigation measures. This implies that the meaning of the indicator, for evaluation of management options, evolves with a changing climate.

Adaptive approaches appear central to handle such effects of climate change, as explicit recognition of the uncertainty and unpredictability is needed, along with a structured process for management response. A key feature of adaptive management is the ability to respond to environmental feedback through monitoring, (re)-assessments and new management options (Allen et al., 2011; Williams, 2011). Quantification of links between multiple pressures and indicators across interacting trophic levels, followed by simulations including climate change, is central for management strategy evaluations and management design. This enables identifying potential tradeoffs and dependencies between manageable and unmanageable pressures. To learn about current meaning of indicators, evaluation and validation of indicators need to be done in an iterative manner. More frequent evaluations, using approaches like our indicator networks that also allow for threshold-shaped relationships to pressures and each other, of the present and projected near-term climate development, could lead to regular re-adjustments of indicator target and threshold values. Such processes could be one way to apprehend the uncertainty related to potentially modulating effects of climate, or other large-scale environmental pressures (Samhouri et al., 2017). We foresee that this kind of flexible strategies make a set of indicators operational long-term and provide a route-map to navigate impacts of climate change. Incorporation of these approaches in EBM is essential if we want to ensure human wellbeing while preserving our ocean ecosystems in an increasingly uncertain future.

## AUTHOR CONTRIBUTIONS

AG and SO conceived the original idea. MK, TB, and SO designed the research. MC and SO prepared the data. MK and SO analyzed

## REFERENCES


the data. MK led the writing with contributions from TB, MC, MT, AG and SO.

## FUNDING

This study is a contribution to the project 'Ecosystembased approach for developing and testing pelagic food-web indicators' financially supported by the Swedish Environmental Protection Agency (Grant No. 20704801). The scenario work was additionally supported by the Swedish Agency for Marine and Water Management (Grant No. 1862-16). MC and TB were also partially financed by the BONUS INSPIRE project supported by BONUS (Art 185), funded jointly by EU and the Swedish Research Council FORMAS, AG was partially financed by a Swedish Research Council FORMAS (Grant No. 217-2013-1315). TB and SO were also partially financed by the BONUS BLUEWEBS project supported by BONUS (Art 185).

## ACKNOWLEDGMENTS

We would like to thank colleagues from the Department of Fish Resources Research, Institute of Food Safety, Animal Health and Environment (BIOR) in Latvia, the Finnish Environment Institute (SYKE), the Leibniz Institute for Baltic Sea Research Warnemünde (IOW) in Germany, and the Swedish University of Agricultural Sciences (SLU) for maintaining these extensive long-term monitoring programs. Threshold functions for tGAMs were supplied by Lorenzo Ciannelli and Kung-Sik Chan. BALTSEM runs were generated as part of the BONUS ECOSUPPORT project and provided by the Baltic Nest Institute, Stockholm University.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00249/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 Kadin, Blenckner, Casini, Gårdmark, Torres and Otto. 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.

# Effects of Ocean Acidification on Marine Photosynthetic Organisms Under the Concurrent Influences of Warming, UV Radiation, and Deoxygenation

Kunshan Gao<sup>1</sup> \*, John Beardall1,2, Donat-P. Häder<sup>3</sup> , Jason M. Hall-Spencer4,5 , Guang Gao<sup>1</sup> and David A. Hutchins<sup>6</sup>

#### Edited by:

Elizabeth Fulton, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia

#### Reviewed by:

Laurie Carol Hofmann, Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research (AWI), Germany Shiguo Li, Research Center for Eco-Environmental Sciences (CAS), China

> \*Correspondence: Kunshan Gao ksgao@xmu.edu.cn

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 01 April 2019 Accepted: 28 May 2019 Published: 18 June 2019

#### Citation:

Gao K, Beardall J, Häder D-P, Hall-Spencer JM, Gao G and Hutchins DA (2019) Effects of Ocean Acidification on Marine Photosynthetic Organisms Under the Concurrent Influences of Warming, UV Radiation, and Deoxygenation. Front. Mar. Sci. 6:322. doi: 10.3389/fmars.2019.00322 <sup>1</sup> State Key Laboratory of Marine Environmental Science and College of Ocean and Earth Sciences, Xiamen University, Xiamen, China, <sup>2</sup> School of Biological Sciences, Monash University, Clayton, VIC, Australia, <sup>3</sup> Department of Biology, University of Erlangen-Nuremberg, Erlangen, Germany, <sup>4</sup> School of Biological and Marine Sciences, University of Plymouth, Plymouth, United Kingdom, <sup>5</sup> Shimoda Marine Research Center, University of Tsukuba, Shimoda, Japan, <sup>6</sup> Marine Environmental Biology, Department of Biological Sciences, University of Southern California, Los Angeles, CA, United States

The oceans take up over 1 million tons of anthropogenic CO<sup>2</sup> per hour, increasing dissolved pCO<sup>2</sup> and decreasing seawater pH in a process called ocean acidification (OA). At the same time greenhouse warming of the surface ocean results in enhanced stratification and shoaling of upper mixed layers, exposing photosynthetic organisms dwelling there to increased visible and UV radiation as well as to a decreased nutrient supply. In addition, ocean warming and anthropogenic eutrophication reduce the concentration of dissolved O<sup>2</sup> in seawater, contributing to the spread of hypoxic zones. All of these global changes interact to affect marine primary producers. Such interactions have been documented, but to a much smaller extent compared to the responses to each single driver. The combined effects could be synergistic, neutral, or antagonistic depending on species or the physiological processes involved as well as experimental setups. For most calcifying algae, the combined impacts of acidification, solar UV, and/or elevated temperature clearly reduce their calcification; for diatoms, elevated CO<sup>2</sup> and light levels interact to enhance their growth at low levels of sunlight but inhibit it at high levels. For most photosynthetic nitrogen fixers (diazotrophs), acidification associated with elevated CO<sup>2</sup> may enhance their N<sup>2</sup> fixation activity, but interactions with other environmental variables such as trace metal availability may neutralize or even reverse these effects. Macroalgae, on the other hand, either as juveniles or adults, appear to benefit from elevated CO<sup>2</sup> with enhanced growth rates and tolerance to lowered pH. There has been little documentation of deoxygenation effects on primary producers, although theoretically elevated CO<sup>2</sup> and decreased O<sup>2</sup> concentrations could selectively enhance carboxylation over oxygenation catalyzed by ribulose-1,5-bisphosphate carboxylase/oxygenase and thereby benefit autotrophs.

**52**

Overall, most ocean-based global change biology studies have used single and/or double stressors in laboratory tests. This overview examines the combined effects of OA with other features such as warming, solar UV radiation, and deoxygenation, focusing on primary producers.

Keywords: algae, global warming, light, multiple stressors, nutrients, hypoxia, phytoplankton, primary productivity

## INTRODUCTION

fmars-06-00322 June 17, 2019 Time: 15:18 # 2

Fossil fuel burning, tropical deforestation, and altered land use have caused the atmospheric CO<sup>2</sup> concentration to rise from 280 ppmv before the Industrial Revolution to 409 ppmv in 2018. These increases in atmospheric CO<sup>2</sup> of about 0.5% per year<sup>1</sup> result in global warming and, after dissolving in surface seawater, also cause ocean acidification (OA). Ocean warming induces stratification and shoaling of the upper mixed layer (UML). This hinders the injection of nutrients from deeper layers and also causes the organisms dwelling in the UML to be exposed to increased daily exposures of PAR and UV radiation (Gao et al., 2012a; Hutchins and Fu, 2017; **Figure 1**). O<sup>2</sup> solubility in seawater decreases due to increasing temperatures, leading to global ocean deoxygenation and contributing to the expansion of anoxic zones. Intensified stratification also results in deoxygenation of deeper water due to lack of ventilation, and human nutrient over-enrichment can deplete dissolved oxygen in coastal regions by promoting excessive levels of primary production and consequent elevated bacterial respiration (Schmidtko et al., 2017; Breitburg et al., 2018).

Most investigations into the biological and ecological effects of OA have been based on experiments using single species

<sup>1</sup>https://www.esrl.noaa.gov/gmd/ccgg/trends/gl\_trend.html

radiation [redrawn based on Gao et al. (2012a) and Hutchins and Fu (2017)].

under controlled conditions, and data on the effects of OA on complex communities in natural environments are relatively limited (Riebesell and Gattuso, 2015; Boyd et al., 2018), although experiments at volcanic seeps and using mesocosms have shown that increased pCO<sup>2</sup> benefits some algae more than others and that lower carbonate saturation states harm some algae more than others. So while most benthic algae are tolerant of OA conditions, changes in carbonate chemistry cause large changes in algal community composition and ecosystem function (Hofmann et al., 2012; Hall-Spencer and Harvey, 2019). The combined effects of OA with other drivers, such as warming, increased UV exposure or deoxygenation, are expected to alter marine communities with effects that differ regionally. Further research is needed to plan for the risks associated with the ways in which OA is interacting with warming, UV radiation, and deoxygenation (Boyd et al., 2018).

Marine photosynthetic organisms account for about half of global photosynthetic carbon fixation (Falkowski and Raven, 2013). Across most of the oceans, the dominant photoautotrophs are mainly microalgae and cyanobacteria, with macroalgae and seagrasses making a higher proportional contribution in coastal environments. These organisms are influenced by both increasing pCO<sup>2</sup> and declining pH with ongoing OA. While increased CO<sup>2</sup> availability can potentially stimulate photosynthesis, stress due to the associated pH drop may be deleterious. This is especially a problem during night in the absence of photosynthetic CO<sup>2</sup> removal and with respiratory CO<sup>2</sup> release, which may lead to more pronounced impacts of acidification during the night period.

While there are many reports in the literature on the responses of primary producers to OA [see recent reviews by Hutchins and Fu (2017) and Boyd et al. (2018) and references therein], contradictory results have often been documented in relation to different phytoplankton species or communities (Gao and Campbell, 2014; Hong et al., 2017). In macroalgae, most studies show algal tolerance of diel pH fluctuations and benefits from rising pCO<sup>2</sup> (Gao et al., 1991, 1993; Cornwall et al., 2012, 2015); in calcifying micro- and macro-algae, it is widely accepted that OA reduces their calcification (Gao et al., 1993; Riebesell et al., 2000; Feng et al., 2008; Gao et al., 2009; Gao and Zheng, 2010; Sinutok et al., 2011). Nevertheless, changes in temperature, light, UV, and nutrient availability are known to modulate the responses of photosynthetic organisms to OA. In this review, we first briefly consider the individual effects of major components of global climate changes (OA, warming, etc.) and then go on to focus on recent advances in our understanding of the effects of OA when combined with other factors, including results from different experimental protocols or algae from different biogeographic regions, and analyze underlining processes and/or mechanisms with special reference to interactive effects of OA with warming, UV, nutrient availability, and dissolved O<sup>2</sup> levels.

## Ocean Acidification

fmars-06-00322 June 17, 2019 Time: 15:18 # 3

Since the Industrial Revolution the ocean has been absorbing anthropogenic CO<sup>2</sup> emissions, and this absorption is currently running at an average rate of over 1 million tons per hour (Sabine et al., 2004), a process that leads to OA. This alters marine carbonate chemistry, changing the concentration ratios of different inorganic carbon species (CO2, HCO<sup>−</sup> 3 , CO2<sup>−</sup> 3 ) and reducing the saturation state of CaCO<sup>3</sup> ().

$$
\Omega = \langle [Ca^{2+}] \times \ [CO\_3^{2-}] \rangle / K\_c,
$$

where K<sup>c</sup> is the product of [Ca2+] <sup>×</sup> [CO2<sup>−</sup> 3 ] when the CaCO<sup>3</sup> solution is saturated.

Since the Ca2<sup>+</sup> concentration of seawater is relatively stable (about 10 mM), mainly depends on the concentration of CO2<sup>−</sup> 3 , as shown in the above equation. Increased dissolution of CO<sup>2</sup> into seawater results in increased concentrations of dissolved CO2, HCO<sup>−</sup> 3 , and H<sup>+</sup> and decreased concentration of CO2<sup>−</sup> 3 , leading to a decrease in the saturation state of CaCO3. The concentration of CO2<sup>−</sup> 3 varies in different regions, depending mainly on temperature and salinity; for instance, the concentration of CO2<sup>−</sup> 3 in polar waters is only about 41% of that in tropical waters, with the former decreasing faster than the latter under the influence of OA (Orr et al., 2005). In the geological past, mass extinctions of marine life have been associated with OA events, and the contemporary seawater pH and carbonate saturation state are declining faster than it has been in about the last 300 million years (Hönisch et al., 2012; Garbelli et al., 2017). Therefore, OA has been ranked as a major research priority in the NOAA Ocean Exploration 2020 report<sup>2</sup> .

Ocean acidification is known to reduce calcification of many calcifying organisms. In coralline algae, elevated pCO<sup>2</sup> reduces calcification of Corallina pilulifera due to lowered pH and decreased concentration of carbonate (Gao et al., 1993). This species exhibits higher calcification rates during daytime than during the night period and increases its calcification with increased dissolved inorganic carbon concentration when pH is maintained constant (Gao et al., 1993). In contrast, another coralline alga, Jana rubens, when transplanted to lower pH site near a CO<sup>2</sup> seep, showed much reduced photochemical efficiency, implying its calcification might be reduced due to reduced photosynthetic energy supply, though the short-term exposure did not bring about significant change in calcification (Porzio et al., 2018). It is clear that OA causes a decrease in competitiveness and fitness for coralline algae compared to brown and turf algae.

Ocean acidification can reduce the content of biogenic silica in diatoms (Milligan et al., 2004; Tatters et al., 2012; Xu et al., 2014), although this response to high CO<sup>2</sup> does not appear to be shared by all diatoms (Qu et al., 2018; Li et al., 2019). Cellular silicification is known to influence predation rates by zooplankton as well as the number of fecal pellets (Liu et al., 2016). Acidification thus might reduce the settling of particulate organic carbon (POC) by lowering diatom silicification and density. With increasing pCO<sup>2</sup> and acidification, surface diatom abundance in the oligotrophic South China Sea decreases, resulting in decreased primary productivity (Gao et al., 2012b). On the other hand, under nutrient replete conditions, growth of diatoms tends to be enhanced (Wu et al., 2014). In coastal eutrophic water, increased partial pressure of CO<sup>2</sup> increased primary production of phytoplankton assemblages dominated by diatoms in a mesocosm study (Huang et al., 2018).

Different diatom species may have entirely different responses to OA, mostly due to differences in species or phenotypes. Under OA conditions, the coastal diatom Thalassiosira weissflogii shows a faster particulate carbon production rate, whereas a pelagic species, Thalassiosira oceanica, grows slower, implying that under diel pH fluctuations the large coastal diatom appears to be more tolerant of the pH decline (Li et al., 2016). While there are species-specific responses to OA, diatoms with lower CCM efficiencies showed more pronounced responses to OA in terms of growth rate (Shi et al., 2019).

Natural gradients of CO<sup>2</sup> at volcanic seeps show that a reduction in mean seawater pH from 8.1 to 7.8 results in around a 30% reduction in animal biodiversity compared with only a 5% reduction in algal diversity, although the algal communities are completely changed due to dissolution of calcified algae and shellfish (Hall-Spencer et al., 2008; Fabricius et al., 2011; Agostini et al., 2018). Ecological impacts seen at CO<sup>2</sup> seeps contrast to those caused by OA events in the geological past when the species and communities of marine organisms were unlike those that we see today (Hönisch et al., 2012).

Seawater uptake of CO<sup>2</sup> at volcanic seeps changes seawater chemistry in a process that is a natural analog for anthropogenic OA (González-Delgado and Hernández, 2018). The seeps create steep gradients in seawater carbonate chemistry with falling pH, falling carbonate, and increasing bicarbonate levels with increasing proximity to the seabed release of CO<sup>2</sup> (Linares et al., 2015). The biological effects of this acidification are found at seep systems worldwide, allowing investigations into the underlying mechanisms that drive community level responses (Cornwall et al., 2017; Connell et al., 2018). As phytoplankton drift into these seep systems calcareous coccolithophores are damaged and dissolve, but this is a shock response to rapid changes in seawater carbonate chemistry (Ziveri et al., 2014). The microphytobenthos at shallow-water seeps, on the other hand, is exposed to long-term acidification over multiple generations. Diatom- and cyanobacteria-dominated biofilms proliferate at shallow-water marine CO<sup>2</sup> seeps. Broad scale analysis of diversity has shown that microphytobenthic communities in high CO<sup>2</sup> water are substantially modified compared with ambient conditions, with a stimulation of diatom growth on sand, mud, bedrock, glass, and plastic substrata (Johnson et al., 2015).

Marine volcanic seeps reveal ecological responses to moderate increases in pCO<sup>2</sup> that retain natural pH variability (Foo et al., 2018). They show the consequences of long-term exposure to acidified waters as well as the implications of more frequent low pH excursions for a wide range of organisms

<sup>2</sup>https://oceanexplorer.noaa.gov/oceanexploration2020

(Agostini et al., 2018). At Mediterranean CO<sup>2</sup> seeps, invasive macroalgal taxa such as Sargassum, Caulerpa, and Asparagopsis species thrive, whereas calcareous macroalgae are sensitive to acidification (**Figure 2**). This causes algal community shifts, fleshy algal dominance, and a loss of coastal biodiversity (Hall-Spencer et al., 2008). Although most macroalgae are resilient to the effects of OA, with only around a 5% loss in species diversity at levels projected under IPCC RCP 8.5, shifts in algal community composition greatly alter habitats (Porzio et al., 2011; Enochs et al., 2015). Increased availability of bicarbonate and pCO<sup>2</sup> stimulates primary production in seagrasses and in some algae but not others, depending on their physiology (Cornwall et al., 2017). In areas sheltered from wave action, this increases carbon fixation and the standing stock of seagrasses and large phaeophytes (Linares et al., 2015; Sunday et al., 2017). However, acidification lowers ecosystem resilience in more exposed conditions, such that only microalgal biofilms and weedy small turf algae persist at high CO<sup>2</sup> levels after storms (Hall-Spencer and Harvey, 2019).

Studies in areas with naturally high levels of CO<sup>2</sup> show that coastal ecosystems are altered especially if elevated DIC is combined with high levels of nutrients (Celis-Plá et al., 2017; Rastrick et al., 2018). In upwelling areas such as Namibia and Peru, these acidified waters have productive food webs that support major fisheries. However, if the acidification is strong enough to cause periods of carbonate under-saturation this is detrimental to shellfish fisheries and calcareous biogenic reefs. Very similar patterns are seen in tropical, sub-tropical, and temperate coastal systems, with stimulated diatom growth and opportunistic algal dominance, calcareous biogenic habitat degradation, and loss of biodiversity. This potentially lowers the resilience of coastal habitats to a cluster of other drivers associated with climate change (global warming, sea level rise, increased storminess), increasing the risk of marine regime shifts and the loss of critical ecosystem functions and services, though it should be noted that in some areas, such as the coastal waters of

FIGURE 2 | Bubbles of CO<sup>2</sup> rising from the seafloor at a depth of 1 m at a volcanic seep off Ischia Island, Italy (40◦43.84<sup>0</sup> N, 13◦57.08<sup>0</sup> E). At a mean pH of 7.8 turf algae (T), seagrass Posidonia oceanica (P) and green seaweed Caulerpa prolifera (C) cover the seabed. Epiphytic and epilithic coralline algae are scarce, yet they are abundant 300 m away from the CO<sup>2</sup> seep where the seawater returns to the normal mean of pH 8.1. (Photo taken by J. Hall-Spencer).

California, upwellings bring elevated levels of nutrients as well as CO<sup>2</sup> into surface waters, leading to high levels of primary productivity (Ryther, 1969; Pauly and Christensen, 1995). Modeling suggests enhancement of some coastal upwellings in the future, and also indicates that the ability to predict flow on effects to ecosystems is complicated by the range of contributing factors and scales involved (Xiu et al., 2018).

There are obvious diel and seasonal variations in the chemistry of coastal seawaters (Dai et al., 2009; Wang et al., 2014), and coastal seawater acidification often coincides with hypoxia, which is frequently enhanced after algal blooms (Zhai et al., 2012). For example, in the Bohai Sea off northern China, pH drops by 0.29 units (H<sup>+</sup> concentration doubles) during the summer period (Zhai et al., 2012). Nevertheless, it is predicted that by the end of the 21st century OA will cause the average pH of the oceanic surface waters to decrease from the preindustrial level of 8.2 down to 7.8 (H<sup>+</sup> concentration rises by 150%) (Gattuso et al., 2015; Guy et al., 2015), endangering marine organisms and ecosystem services (Broadgate et al., 2013).

## Ocean Warming

Since the Industrial Revolution, increasing global average temperature has been linearly correlated with atmospheric CO<sup>2</sup> concentrations (IPCC, 2013a). Since the 1970s, the oceans have absorbed over 90% of Earth's heat gain, leading to ocean warming (Gattuso et al., 2015). Ocean warming can be detected to a depth of 1,000 m (Levitus et al., 2000), and during the past century the global ocean surface temperature has increased by about 1◦C (Fischetti, 2013). By the end of the 21st century, the average Earth surface temperature is predicted to rise by 2–4 ◦C (Gattuso et al., 2015). Compared with many other regions in the world, the marginal waters in the East China Sea and the South China Sea have shown a faster rate of temperature increase over the past 50 years (Bao and Ren, 2014; Williams et al., 2016; Cai R. et al., 2017), though polar regions have warmed fastest (Wassmann, 2015). Global warming is predicted to have other consequences such as increasing storm activity that can influence marine biota. For instance, a rise in temperature of 1◦C has been shown to increase the number of typhoon events by up to 25% (Bigg and Hanna, 2016). At the same time, the frequency of marine heat waves, prolonged discrete anomalously warm water events (Hobday et al., 2016), has been predicted to increase tens of fold with a 3.5◦C rise in sea surface temperature (Frölicher et al., 2018). Marine heat waves are postulated as drivers of ecological disasters such as the triggering of toxic algal blooms along the North American west coast (McCabe et al., 2016).

Temperature is a key factor affecting enzyme activity and metabolism. Metabolic rates usually increase with temperature up to a maximum value and thereafter rapidly decline, exhibiting an energy activation to deactivation transition. Shallow water marine organisms can be exposed to diurnal, seasonal, and current-driven sudden changes in temperature. These changes involve the thermocline, tides, typhoons, and cloud cover as well as the long-term changes due to natural climate cycles and human activities (IPCC, 2013b). Ocean warming affects organism physiology (Pörtner, 2008; Sinclair et al., 2016) and changes biogeographic boundaries,

community composition, and phenology (Hutchins and Fu, 2017). In laboratory experiments, increasing temperature below optimal values can increase phytoplankton growth (Fu et al., 2014; Summers et al., 2016; O'Donnell et al., 2017; Jiang et al., 2018). However, on a global scale, warming may decrease primary productivity due to nutrient limitation driven by enhanced stratification of the upper mixing layer (UML); it may also enhance fixed carbon remineralization by heterotrophic microorganisms and so reduce the strength of the ocean carbon sink (Danovaro et al., 2016; Cavan et al., 2019).

## Deoxygenation

One additional effect of ocean warming is that it decreases the solubility of oxygen (and gases generally) in seawater. In oceanic and coastal waters the dissolved oxygen content has declined over the past 50 years (Schmidtko et al., 2017; Breitburg et al., 2018). In addition, shoaling of the UML or enhanced ocean stratification reduces ventilation from the surface to deeper layers, further exacerbating the deoxygenation phenomenon. As POC sinks, marine bacteria feed on it, consuming O<sup>2</sup> and releasing CO2. This results in an oxygen minimum zone at depths of 500–600 m in the eastern tropical Pacific (Brewer and Peltzer, 2009). Over the past 50 years the depth of a hypoxic layer (<2 mg O<sup>2</sup> L −1 ) in the Pacific Ocean has shoaled from 400 to 300 m, and the dissolved oxygen content has decreased significantly (Whitney et al., 2007) due to sea surface warming (Keeling et al., 2010); and ocean deoxygenation has been recently shown to cause deterioration in a number of processes including biogeochemistry and ecosystem function (Breitburg et al., 2018). In coastal regions deoxygenation is driven mainly by eutrophication, leading to excess oxygen consumption and the development of hypoxia and "dead zones." However, decreasing solubility of O<sup>2</sup> through warming can exacerbate this process.

Most marine organisms need O<sup>2</sup> in their metabolic processes. When the dissolved O<sup>2</sup> concentration is below a certain value, these organisms suffer from stressed respiration, and hypoxic events may lead to the death of many organisms. For different organisms, the half lethal concentration (LC50) of dissolved O<sup>2</sup> is different, and the critical value for one organism might be several times that for others. In typical anoxic zones, the oxygen content is below 2 mg L−<sup>1</sup> and the hypoxia is usually associated with high pCO<sup>2</sup> and low pH. In coastal ecosystems, the frequency of hypoxia is increasing at a rate of about 5.5% per year due to the interaction between deoxygenation and eutrophication (Vaquer-Sunyer and Duarte, 2008), which is responsible for a faster rate of OA in coastal, compared to oceanic waters (Cai et al., 2011; Breitburg et al., 2018). Eutrophication frequently leads to algal blooms. When these blooms collapse, the algae die, leading to enhanced organic carbon loads, the breakdown of which, through microbial action, contributes to anoxia and in turn renders spawning grounds for fish such as herring untenable. In terms of biogeochemical impacts, deoxygenation promotes denitrification, reducing the concentration of nitrate and affecting the ocean N cycle, primary productivity, and the efficiency of the marine biological carbon pump (Hutchins and Fu, 2017).

Lower oxygen levels would be expected to have the most effect in photoautotrophic species on the ratio of oxygenase to carboxylase activity of the CO2-fixing enzyme ribulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco). Diminished oxygen concentrations favor the carboxylase activity of Rubisco and diminish photorespiratory activity. However, in practice low oxygen concentrations, i.e., below air-equilibrium levels, appear to have little effect on physiology and growth of algae, though levels above air-equilibrium can be damaging (Black et al., 1976; Drechsler and Beer, 1991; Beer et al., 2000; Kitaya et al., 2005, 2008; Kliphuis et al., 2011; Raso et al., 2012; Haas et al., 2014). Kim et al. (2018) showed a stimulation of photosynthetic rates of the seagrass Zostera marina at sub-saturating, but not saturating, light intensities when oxygen concentrations dropped from 231 to 8 µmol O<sup>2</sup> L −1 . At the same time respiration rates dropped fivefold at the lower oxygen level. In part, the lack of effect of low oxygen is likely to be due to the presence in most algae of active CO<sup>2</sup> concentrating mechanisms (CCMs), which maintain a sufficiently high CO2:O<sup>2</sup> ratio at the active site of Rubisco to minimize oxygenase activity and photorespiration (Giordano et al., 2005). Of the other oxygen consuming reactions found in algae and cyanobacteria, only the Mehler reaction and glycolate oxidase (in those organisms that possess that enzyme rather than glycolate dehydrogenase) have a K0.<sup>5</sup> O<sup>2</sup> high enough to be affected by oxygen concentrations around air equilibrium (Beardall et al., 2003; Raven and Beardall, 2005). Cytochrome oxidase for example has a K0.<sup>5</sup> O<sup>2</sup> of ≤2 µmol O<sup>2</sup> L −1 (see Table 3.2 of Raven and Beardall, 2005).

However, to date there have been no studies on the interaction between low oxygen concentration and other stressors associated with ocean global changes. In the study of Kim et al. (2018) low oxygen resulted in a 2.8-fold downregulation of γ-carbonic anhydrase (γ-CA) genes, which they suggest may play a role in bicarbonate transport rather than as a standard CA. It might be expected that elevated CO<sup>2</sup> (OA) and diminished O<sup>2</sup> concentrations lead to further downregulation, though this has yet to be tested. Anoxia also stimulates nitrogen fixation in the marine cyanobacterium Trichodesmium (Staal et al., 2007), with aerobic conditions leading to N<sup>2</sup> fixation being limited by the availability of reducing equivalents, a result interpreted by Staal et al. (2007) as suggesting a competition for electrons between nitrogenase and respiration when oxygen is present.

## Stratification and Increased Exposure to UV

It is commonly known that the UML is much shallower during summer than in winter. For example, in pelagic areas of the South China Sea, the UML can be over 30 m shallower in summer than in winter (Liu et al., 2007). In addition to less wind and storm mixing than in winter, the higher temperatures in summer are mainly responsible for the enhanced stratification.

On a global scale, ocean warming enhances stratification in the upper oceans due to reduced advective mixing, leading to shoaling of the UML. Such stratification hinders the upward transport of nutrients from deeper layers. Consequently, phytoplankton cells are exposed to a reduced availability of

nutrients, provided that the contribution from other sources such as nitrogen fixation or deposition into the oceans from run-off and atmospheric sources is unaltered by global climate changes. Phytoplankton cells dwelling in a shallower UML are exposed to higher doses of UV radiation since they cannot actively move downward to deeper waters (Jin et al., 2013). Since the vertical migration path for cells within a shallower UML is shorter, they experience much higher daytime average or integrated levels of solar radiation [see reviews by Riebesell and Tortell (2011) and by Gao et al. (2012b)]. Since UV irradiances can penetrate as deep as 80 m in the pelagic ocean, down to the middle or lower layers of the euphotic zone, phytoplankton cells are inevitably exposed to increased levels of UV radiation due to enhanced stratification [see the review by Gao et al. (2012b) and literature cited therein].

UV radiation is divided into UV-A (315–400 nm), UV-B (280–315 nm), and UV-C (<280 nm). The stratospheric ozone layer completely absorbs UV-C (the most biologically harmful band). The major portion of the UV-B is also absorbed by ozone, mainly in the stratosphere. Only a small fraction of UV-B reaches the Earth's surface through the stratosphere and the upper troposphere. On the other hand, most of the UV-A and visible light reaches the surface of the Earth. In the subtropics, UV-B radiation reaching the sea surface usually accounts for <1% of the total sunlight energy; UV-A and visible light (PAR: 400–700 nm) account for 7–8 and 40–50%, respectively, with infrared radiation making up the balance. At noon, the UV-B:UV-A:PAR ratio is about 0.5:16:100. The UV-B:PAR ratio is highest at noon, while the UV-A:PAR ratio remains unchanged during the day (Gao, 2018). Solar radiation is reduced by moisture in the atmosphere, sea surface reflectivity, and inorganic and organic substances in the water, and has obvious latitudinal, diel, and seasonal dependencies. UV-B radiation increases by 2% for each 1% decrease in stratospheric ozone concentration (Kerr and McElroy, 1993). Ozone destruction has been curbed, with stratospheric ozone expected to return to pre-1980 levels by the mid-21st century (Plummer et al., 2010). However, global warming, the presence of trace gases (Ossó et al., 2011) and changes in atmospheric circulation might increase UV-B level at low latitudes by 2–3% (Williamson et al., 2014).

Many oceanographic biological observations of primary productivity have neglected the influence of UV radiation because of traditional recommendations for use of transparent vessels, which, unless made of quartz and UV-transparent materials, are mainly glass or polycarbonate bottles that are opaque to UV wavelengths. With progressive ocean climate changes, more attention should be given to the interactive impacts of UV with other global environmental change drivers (Häder and Gao, 2015).

Even though the irradiance of UV-B is only a few percent (in tropical and subtropical areas, <1%) of the total solar UV reaching the Earth's surface, it is the most deleterious wavelength band encountered by photosynthetic organisms dwelling in the photic zone. UV-B can damage proteins, lipids, and other bioactive components of the cells (Ganapathy et al., 2017). One of the main targets is photosynthetic electron transport, since UV-B damages the D1 protein in photosystem II, reducing the photosynthetic capacity (García-Gómez et al., 2016). This damage is repaired by removing the faulty protein and replacing it with a newly synthesized molecule. This enzymatic process is favored by increased temperature (Gao et al., 2008). When moving actively or passively in the UML, cells are affected most near the surface and use the time in deeper water closer to the thermocline for repair (Helbling et al., 2003). Another important target of solar UV-B is DNA (Gao et al., 2008; Rajneesh et al., 2018). UV-B radiation mainly induces the formation of cyclobutane pyrimidine dimers, leading to mutations and cell death if these are not removed by a photolyase which uses the energy of UV-A and blue light to split the dimers (García-Gómez et al., 2014).

Many prokaryotic and eukaryotic phytoplankton as well as macroalgae synthesize UV-absorbing substances, such as scytonemin (only cyanobacteria) and mycosporine-like amino acids (MAAs), which convert the energy of short-wavelength UV radiation into heat before it can hit sensitive targets (Singh et al., 2008). Another mechanism of damage by UV radiation is the production of reactive oxygen species (ROS) such as peroxides, superoxide, hydroxyl radicals, and singlet oxygen which can oxidatively damage proteins, lipids, and other cell structures (Williamson et al., 2019). ROS can be detoxified either enzymatically (e.g., by superoxide dismutase, catalase, peroxidase) or non-enzymatically by antioxidants (e.g., thiols, ascorbic acid, glutathione, carotenoids) (Martínez, 2007; Goiris et al., 2015).

In addition to the negative effects of UV radiation, it is known that moderate levels of UVA radiation enhance primary production by enhancing photosynthetic light use efficiency, as shown in the photosynthetic rate vs. PAR curves determined in the presence and absence of UV radiation (Gao et al., 2007a). When PAR was filtered out, UVA stimulates carbon fixation of phytoplankton assemblages in coastal waters of the South China Sea, and addition of UV-B reduces the carbon fixation (Gao et al., 2007a). UV-induced inhibition of photosynthetic carbon fixation in surface seawater increases with distance from the coast toward the open ocean, probably due to differences in attenuation coefficient, phytoplankton community structure, and nutrient availability as well as mixing rate (Tedetti and Sempéré, 2006; Li et al., 2011). Faster mixing reduces UV-induced photosynthetic inhibition (Helbling et al., 2003; Jin et al., 2013).

## INTERACTIONS BETWEEN FACTORS ASSOCIATED WITH CLIMATE CHANGE

## OA and Solar Radiation (PAR and UVR)

Changing PAR levels modulate diatom responses to OA. For instance, when three diatom species were grown under different levels of sunlight, lowered pH with elevated CO<sup>2</sup> stimulated growth under low to moderate levels of light, but inhibited it under high levels (Gao et al., 2012a; **Figure 3**). However, such a reversed response to OA with increasing light levels was not found in coccolithophores, which showed enhanced growth rates (with reduced calcification) under the elevated pCO<sup>2</sup> under all tested PAR levels (Jin et al., 2017;

**Figure 3**). Likewise, Feng et al. (2008) found that the interactive effects of high light and elevated CO<sup>2</sup> on a subtropical coccolithophore increased growth rates but strongly decreased calcification.

sunlight levels, though its calcification decreases. [Derived from Gao et al.

(2012b) for panel "A" and from Jin et al. (2017) for panel "B"].

Solar UV radiation may interact with OA to affect primary producers. To date, most OA effects have been observed under UV-free light conditions, either under indoor artificial light or under solar radiation with vessels that are opaque to UV irradiances. Therefore, relatively little information is available on this aspect. OA exacerbates the impact of UV on calcification in the coccolithophore Emiliania huxleyi (Gao et al., 2009) and on calcification of a coralline alga (Gao and Zheng, 2010). Decreased thickness of the calcified layer increased cellular exposure to UV radiation and consequently led to enhanced photoinhibition of photosynthesis by UV (Gao et al., 2009; Gao and Zheng, 2010). Elevated CO<sup>2</sup> increased the sensitivity of freshwater lake phytoplankton populations to UV-B (Sobrino et al., 2009) and also exacerbated the harmful effect of UVR on PSII function in the marine diatom T. weissflogii through reducing the PsbD removal rate and the ratio of Rubisco to PsbA during UVR exposure (Gao et al., 2018c). In contrast, the marine microalga Nannochloris atomus increased its growth in response to elevated CO<sup>2</sup> with an insignificant photosynthetic response to UVR (Sobrino et al., 2005). Furthermore, the rETRmax of Corallina officinalis was stimulated by elevated CO2, and exposure to UVR led to further stimulation (Yildiz et al., 2013). The divergence between these findings may be due to the differences in species or light intensity.

## Interactions Between OA Effects and Nitrogen Fixation and With the Availability of Other Nutrients

Marine diazotrophs, N2-fixing cyanobacteria such as Trichodesmium spp., play an important role in remediating N-limitation in oligotrophic waters (Capone and Carpenter, 1982; Sohm et al., 2011; Hutchins et al., 2015). Trichodesmium has a high abundance in the oligotrophic China Sea, frequently forming blooms in the East and South China Seas (Chen et al., 2014). Because of the N<sup>2</sup> fixation capacity of Trichodesmium and subsequent microbial cycles, biologically available nutrients can be replenished, which promotes growth of phytoplankton, thus enhancing primary and secondary productivities, playing an important role in driving the marine biological CO<sup>2</sup> pump (Hutchins and Fu, 2017).

Numerous studies have demonstrated that N<sup>2</sup> fixation increases under elevated CO<sup>2</sup> levels in nutrient-replete cultures of both the filamentous species Trichodesmium (Barcelos e Ramos et al., 2007; Hutchins et al., 2007, 2013, 2015; Kranz et al., 2009, 2010; Levitan et al., 2010a,b; Eichner et al., 2014; Walworth et al., 2016a,b) and the unicellular Crocosphaera (Fu et al., 2008; Garcia et al., 2013a,b). However, there are a few studies showing that OA decreased N<sup>2</sup> fixation of the diazotroph (Hong et al., 2017; Luo et al., 2019). While there are disputes over the positive and negative effects of elevated CO<sup>2</sup> concentrations on N<sup>2</sup> fixation in Trichodesmium (Hutchins et al., 2017), it is generally agreed that severe iron (Fe) limitation cancels out (or even reverses) the positive effects of CO<sup>2</sup> on cyanobacterial diazotrophs (Fu et al., 2008; Shi et al., 2012; Walworth et al., 2016a; Hong et al., 2017). In contrast, growth limitation by phosphorus (P) appears to operate independently, regardless of any effects of changing pCO<sup>2</sup> (Hutchins et al., 2007; Garcia et al., 2013b; Walworth et al., 2016a). A nearly decade-long experimental evolution study with Trichodesmium selected under two CO<sup>2</sup> levels yielded an unexpected result: the increases in growth and N<sup>2</sup> fixation commonly observed under exposure to short-term high CO<sup>2</sup> conditions became irreversible after selection under OA conditions for several years. Thus, high CO2-adapted cultures retained their increased N<sup>2</sup> fixation and growth rates indefinitely, even when moved back to lower pCO<sup>2</sup> levels (Hutchins et al., 2015; Walworth et al., 2016a). This constitutive upregulation occurs through the process of "genetic assimilation," whereby following extended natural selection, traits such as elevated nitrogen fixation rates transition from being physiologically plastic responses to being genetically fixed ones (Walworth et al., 2016b).

Nevertheless, OA can sometimes show different effects on other nitrogen fixing taxa. For heterocystous cyanobacteria, studies showed that OA can either promote nitrogen fixation (Wannicke et al., 2012) or inhibit it (Czerny et al., 2009). Ship-board studies during research cruises have shown that OA treatment can either stimulate N<sup>2</sup> fixation (Lomas et al., 2012), did not bring any significant effects, or negatively affected

cyanobacterial N<sup>2</sup> fixation (Böttjer et al., 2014; Gradoville et al., 2014). These field experiments are difficult to interpret though, as N2-fixing cyanobacteria such as Trichodesmium are famous for surviving poorly in shipboard incubations, and Gradoville et al. (2014) were unable to demonstrate consistent limitation of N<sup>2</sup> fixation not only by CO<sup>2</sup> but also by any other potentially limiting resources, including iron and P. There are inherent differences in CO<sup>2</sup> affinities and growth responses among strains and species of N2-fixing cyanobacteria (Hutchins et al., 2013), and different culturing or light conditions can lead to discrepancies in experimental results. Nevertheless, it is clear that species-specific differences in responses of N2-fixing cyanobacteria to OA may lead to changes in community structure and diversity of marine diazotrophs in future oceans (Hutchins et al., 2013; Gradoville et al., 2014).

UV radiation has been shown to adversely affect cyanobacterial N<sup>2</sup> fixation, including inhibiting development of heterocysts and reducing chlorophyll content in the fresh water N<sup>2</sup> fixer Anabaena sp. (Gao et al., 2007b). In the marine diazotroph Trichodesmium, UV radiation reduces rates of photosynthesis, N<sup>2</sup> fixation, and growth (Cai X. et al., 2017), though moderate or low levels of UV-A may be stimulating. Another recent study showed that growth, N<sup>2</sup> fixation, and CO<sup>2</sup> fixation rates of the Trichodesmium strain IMS 101 and Crocosphaera WH0005 were both negatively affected by UV-B exposure (Zhu et al., 2019). This inhibition was much greater for Trichodesmium IMS 101, which fixes N<sup>2</sup> during the day, than for Crocosphaera, which fixes N<sup>2</sup> only during the night; however, another Trichodesmium isolate (GBR) was much more resistant to UV-B inhibition (Zhu et al., 2019). To the best of our knowledge, however, nothing has been documented on the interactive effects of OA and UV radiation on marine diazotrophs.

## OA and Warming

Laboratory studies have shown that combined OA and warming have different ecological and physiological effects on different species. For example, in the picoplanktonic cyanobacterium Synechococcus, warming and OA together promote its growth, but they have no effects on Prochlorococcus (Fu et al., 2007). Similarly, simultaneous OA and warming by 4◦C promoted the growth of the diatom Skeletonema (Kremp et al., 2012), but had no obvious effects on Thalassiosira and Chaetoceros (Hyun et al., 2014). Under acidification and warming conditions, the growth rate and calcification of the coccolithophore Coccolithus decreased (Schlüter et al., 2014), and OA and warming together lowered the optimum growth temperature and maximum growth rate of E. huxleyi (Listmann et al., 2016). However, other work shows that warming combined with partial pressures of 20–6,000 µatm CO<sup>2</sup> increases the production rate of particulate inorganic carbon and POC of E. huxleyi and Gephyrocapsa oceanica (Sett et al., 2014). In a natural North Atlantic Bloom community, simultaneous OA and warming stimulated coccolithophore growth while at the same time significantly reducing their calcification (Feng et al., 2009).

In general, acidification and warming have usually been found to promote the growth and nitrogen fixation rates of open ocean diazotrophic cyanobacteria such as Trichodesmium and Crocosphaera (Hutchins et al., 2007, 2019; Fu et al., 2014; Hutchins and Fu, 2017). Warming alone also favors the growth of globally important open ocean nitrogen-fixing cyanobacteria taxa, as long as temperatures do not exceed ∼32–33◦C. The stimulatory effects of higher temperatures and pCO<sup>2</sup> together have been suggested to be roughly additive (Hutchins et al., 2007; Levitan et al., 2010a). Temperatures in the low-latitude tropical oceans are predicted to eventually exceed optimum growth levels in many regions, leading to potential future thermal exclusion of these N<sup>2</sup> fixers from core parts of their current biogeographic range (Breitbarth et al., 2007; Thomas et al., 2012; Fu et al., 2014; Jiang et al., 2018).

Ocean acidification and warming are known interactively to affect key physiological performances of macroalgae. For instance, the respiratory coefficient (the rate of change of the respiratory rate as the temperature increases) in Sargassum fusiforme increases under OA conditions, indicating that acidification and warming additively increase the respiratory rate (Zou et al., 2011). Furthermore, warming and OA synergistically enhance reproduction events and shorten the generation span in the "green tide" alga Ulva rigida (Gao et al., 2017, 2018b), suggesting more severe green tides may occur in future oceans.

Research on reef-building coralline algae shows that acidification causes a reduction in their calcification, and warming and acidification together further lower the calcium quality (Martin and Hall-Spencer, 2017). Acidification and warming together have been shown to improve the nitrogen fixation of nitrogen-fixing cyanobacteria (Hutchins et al., 2007). In the Bay of Kiel, warming is believed to drive changes in the phytoplankton community structure and an increase in the zooplankton biomass; under in situ water temperature conditions, the predation rate of zooplankton decreases due to the negative effects of acidification (Paul et al., 2016). For some marine algae, the combined effects of acidification and warming are manifested as acidification changing their capacity to cope with temperature changes and alterations in the range of their optimum survival temperature. For example, the calcification rate of the coral Acropora pulchra is controlled by temperature and acidification (Comeau et al., 2016). For the warming and acidification effects, there may be regional differences due to different biotic populations and physical and chemical environments. This phenomenon needs to be further verified and its mechanisms need to be discussed.

Temperature also influences the responses of algae to UV radiation. We might expect a priori that temperature would exert an effect on net UV-B impacts by affecting enzyme-driven repair mechanisms (which are strongly temperature dependent) more than photochemical damage (since photochemistry is temperature independent). In the model diatom Phaeodactylum tricornutum, elevated temperature increased the repair rate of PSII either under ambient or elevated CO<sup>2</sup> levels in the presence of UV-A or UV-B (Li and Gao, 2012). On the other hand, elevated temperature increased inhibition of photochemical quantum yield by UVR in freshwater phytoplankton assemblages from Patagonia in Argentina (Villafane et al., 2013); however, increased

temperature in this region helped to counter the magnitude of daytime yield decrease during the onset of a phytoplankton bloom (Helbling et al., 2013).

Recently, it has been suggested that ocean warming would affect the marine biological carbon pump (MBP) and microbial carbon pump (MCP) as well as their interactions (Jiao and Zheng, 2011). However, it has been shown that acidification has no effect on the composition of dissolved organic carbon (DOC) in a plankton ecosystem (Zark et al., 2015), while other findings show that acidification promotes the production of POC (Czerny et al., 2013). Whether acidification and warming will produce combined effects and further affect the composition and lability of DOC and the production of POC in different waters is not yet known. In the oceans, most DOC will be converted to CO<sup>2</sup> in the short term through bacterial action; however, some DOC resists bacterial breakdown and survives for a long time (for hundreds or even thousands of years), playing a steady role in the carbon sink (Jiao et al., 2010). Therefore, it is particularly important to understand the carbon sink/source processes in marine organisms and the response of assimilation and dissimilation to acidification and warming. Nevertheless, most of the present research findings on the combined effects of the warming and acidification have been obtained under constant laboratory conditions, which may not reflect natural conditions where factors such as light exposure, temperature, and concentrations of dissolved CO<sup>2</sup> can vary even on short time scales. They are still quite controversial, and the combined effects and their mechanisms are still unclear.

## Nutrient Availability and Interactions With Other Components of Global Change

With progressive ocean warming, intensified stratification of the UML is expected, and consequently, upward transport of nutrients across the thermocline will be reduced. Therefore, phytoplankton cells dwelling in this layer are expected to be exposed to decreasing availability of nutrients. Levels of nutrients and/or ratios of different nutrients are known to affect algal physiology. Interactions among nutrient availability, warming, and OA are thus to be expected. For instance, in the diatom Thalassiosira pseudonana, nitrate limitation interacts with OA and warming, leading to increased respiration and decreased photosynthetic rates and growth (Li et al., 2018). Likewise, in the large centric diatom Coscinodiscus, OA and warming together reduced growth affinity for nitrate, and so increased nitrate concentrations required for growth (Qu et al., 2018).

In the diatom Chaetoceros didymus, OA had no effect on growth rate under N-replete conditions, but stimulated growth under N-limitation. N-limitation and OA also interacted to increase sinking rates, which were less affected under normal CO<sup>2</sup> conditions when N supply was high (Mannfolk, 2016). Toxic microalgae are influenced by interactive effects of elevated CO<sup>2</sup> and nutrient supply. Fu et al. (2010) showed that under P-limited conditions, toxin production by Karlodinium veneficum was greatly stimulated by elevated CO2. Tatters et al. (2013) found saxitoxin production by the dinoflagellate Alexandrium was greatly increased by the combination of OA and P limitation, but warming to some extent counteracted these effects. Tatters et al. (2012) reported that elevated CO<sup>2</sup> enhanced domoic acid production by the harmful bloom diatom Pseudo-nitzschia fraudulenta under nutrient-limited conditions. In general, it seems that the combination of OA with nutrient limitation could lead to much more toxic and damaging blooms of harmful algae in the future ocean.

In addition to microalgae, coastal eutrophication can also lead to macroalgal blooms, including green tides and golden tides. Ulva-dominated green tides and Sargassum-dominated golden tides have increased worldwide in recent years (Smetacek and Zingone, 2013). Kang and Chung (2017) reported that ammonium enrichment stimulated growth of Ulva pertusa, with further stimulation when combined with OA. Gao et al. (2017, 2018b) showing that the combination of eutrophication, OA, and warming enhanced the settlement, germination, and growth of U. rigida. The combination of phosphate enrichment and OA also enhanced growth rate in the golden tide alga Sargassum muticum (Xu et al., 2017). These findings suggest that the rising trend of macroalgal blooms that occurred in Chinese coastal waters (Liu et al., 2013) may correlate with ocean climate changes. Eutrophication is also likely to interact with other stressors, negatively impacting coral reef systems (Bell et al., 2014).

Nutrient limitation favors smaller-celled organisms such as the picophytoplankton Micromonas and Ostreococcus and nanophytoplankton such as the coccolithophorid E. huxleyi by virtue of their higher surface area to volume ratio. Nutrient supply and concentrations in the UMLs of the oceans are often reported as being drivers of variation in the size structure and taxonomic composition of phytoplankton communities (Eppley and Peterson, 1979; Chisholm, 1992; Coale et al., 1996; Finkel et al., 2010). With warming-enhanced stratification, such cells will be subjected to higher average exposures to solar radiation, including periodic acute exposure to high fluxes of UV-B close to the surface, which is known to do more harm to smaller cells (Wu et al., 2015).

As stated earlier, it is generally agreed that severe iron (Fe) limitation cancels out (or even reverses) the positive effects of CO<sup>2</sup> on cyanobacterial diazotrophs (Fu et al., 2008; Shi et al., 2012; Walworth et al., 2016a; Hong et al., 2017). In contrast, growth limitation by P appears to operate independently, regardless of any effects of changing pCO<sup>2</sup> (Hutchins et al., 2007; Garcia et al., 2013b; Walworth et al., 2016a). Trichodesmium cultures simultaneously co-limited by both Fe and P grow considerably faster than either Fe-limited or P-limited cells (Garcia et al., 2015; Walworth et al., 2016a). However, their phenotype relative to acidification is more similar to Fe-limited than P-limited cultures, in that high CO<sup>2</sup> does not increase their physiological rates. These high CO2, Fe/P co-limited cells express a unique complement of proteins that is quite distinct from those seen in low CO<sup>2</sup> or single nutrient-limited cultures, representing perhaps the manifestation of a genetically determined compensatory mechanism (Walworth et al., 2016a).

Limitation by the micronutrient iron (Fe) is a key control of marine nitrogen fixation and, along with other important variables, Fe supplies are also changing with a shifting climate

(Hutchins and Boyd, 2016). Unlike CO<sup>2</sup> and P, though, interactions between Fe availability and temperature in N<sup>2</sup> fixers do not appear to be linear and additive (Jiang et al., 2018). Warming from 27 to 32◦C decreases growth and N<sup>2</sup> fixation rates in Fe-replete Trichodesmium, but 32◦C is the temperature at which maximum rates are observed in Fe-limited cultures. Consequently, warming across this temperature range nearly erases the negative effects of Fe limitation by greatly increasing cellular iron use efficiencies (moles N fixed per mol cellular Fe per hour) by up to 470%. The net effect of this non-linear interaction between Fe and warming in a global biogeochemical model is a predicted ∼22% increase in total global marine N<sup>2</sup> fixation over the next 90 years. This potential alleviation of Fe limitation by warming throughout large parts of the ocean (particularly the oligotrophic Pacific) could also enable currently Fe-limited diazotroph populations to respond positively to concurrently increasing atmospheric pCO<sup>2</sup> (Jiang et al., 2018).

Nutrient limitation appears to increase the sensitivity of algae to UV-B radiation. N-limitation for instance can impair the capacity of cells to synthesize UV-B-screening compounds such as MAAs or influence the ability of cells to carry out repair [particularly of the D1 protein in PSII (Shelly et al., 2002)]. P-limitation also appears to decrease the capacity for repair processes under UV-B (Heraud et al., 2005). Similar effects of nutrient availability on UV sensitivity have also been reported in macroalgae. Zheng and Gao (2009) reported higher MAA concentrations in Gracilaria lemaneiformis under nitrate enrichment, which led to significantly decreased UVR-induced inhibition of growth and photosynthesis. In the same organism UV-B significantly reduced the net photosynthetic rate, but this was alleviated by enrichment with ammonia (which also stimulated the accumulation of UV-absorbing compounds) (Xu and Gao, 2012). Similar data have been reported for Porphyra (Korbee et al., 2005) and Gracilaria tenuistipitata (Barufi et al., 2011). In the red alga G. lemaneiformis, UV-induced inhibition of photosynthesis and growth was exacerbated under P-limited conditions (Xu and Gao, 2009). These data and other work cited by Figueroa and Korbee (2010) imply that in macroalgae as well as microalgae, low N-availability enhances UVR sensitivity of photosynthesis and growth, perhaps by decreasing the capacity to minimize damage through MAA synthesis. There are many reports in the literature on this topic, as reviewed by Beardall et al. (2014). However, how such nutrient-related responses to UVR are modulated under the additional stress of OA is not yet known.

## OA Effects Under Multiple Stressors

Ocean acidification is a global phenomenon but is usually overlaid by pronounced regional variability modulated by local physics, chemistry, and biology (Boyd et al., 2018; Hurd et al., 2018). Biotic responses to multiple environmental drivers depend on the response to the single dominant driver, and the chance of a driver of larger effect being present increases with the number of drivers (Brennan and Collins, 2015). In addition to the effect of OA together with one additional stressor, the combined effects of OA with two drivers have also been investigated, although relatively fewer papers addressing this can be found. As mentioned above, enhanced stratification due to warming leads to decreased upward transport of nutrients from deeper layers and this in turn can lead to nutrient limitation. Therefore, the effects of OA and warming under changing nutrient levels are of general significance. Li et al. (2018) showed that OA or warming did not affect the specific growth rate of T. pseudonana under nitrate-replete conditions, but both conditions reduced its growth rate under nitrate-limited conditions, suggesting that a decreased upward transport of nutrients due to enhanced stratification could act synergistically with OA and warming to reduce its growth. Under influences of natural environmental conditions, a mesocosm test showed that the toxic microalga Vicicitus globosus has a selective advantage under OA, increasing its abundance in natural plankton communities in oligotrophic subtropical waters, which has had a dramatic impact on the plankton community, disrupting trophic transfer of primary produced organic matter (Riebesell et al., 2018).

In regard to macroalgae, both OA and warming increased germination and juvenile growth in a green tide alga U. rigida regardless of nutrient availability, while the stimulatory effect of OA and warming was reduced by nitrate limitation (Gao et al., 2017). In addition to physiological performance, OA, warming, and nitrate abundance synergistically increased fatty acid content in U. rigida (Gao et al., 2018a), affecting the food quality of this green macroalga. Some macroalgae growing near CO<sup>2</sup> seeps appear to use more CO<sup>2</sup> rather than bicarbonate except coralline algae, whose abundance declined (Cornwall et al., 2017). In coastal waters, where nutrient availability is high and diel pH fluctuation is typical, most fleshy macroalgae appear to benefit from increasing CO<sup>2</sup> concentration, while calcifying algae are harmed by OA (Gao et al., 1993; Gao and Zheng, 2010; Sinutok et al., 2011). It seems that algae respond differentially to OA under influences of multiple stressors. Obviously, more studies are needed in order to visualize comprehensive pictures of how marine photosynthetic organisms in different ecosystems respond to the concurrent ocean climate changes.

## FUTURE PERSPECTIVES

Microalgae are known to exhibit evolutionary responses to elevated CO<sup>2</sup> over hundreds of generations, including downregulated CCMs in a green alga (Collins et al., 2006), irreversible capacity of reduced calcification in a coccolithophorid (Tong et al., 2018), smaller cell size and decreased respiration and photosynthesis in a model diatom (Li et al., 2017), and increased cellular N/C ratio in E. huxleyi (Jin et al., 2013). The evolutionary response of a diazotroph to OA treatment led to irreversible enhancement of growth and N<sup>2</sup> fixation over hundreds of generations, driven by "genetic assimilation" of plastic traits into adaptive ones (Hutchins et al., 2015; Walworth et al., 2016a). Adaptation to high CO<sup>2</sup> is also associated with the imprint of cytosine methylation of the Trichodesmium genome (Walworth et al., 2016b).

Phytoplankton also shows obvious evolutionary responses to warming. Cultured diatoms from tropical waters increased their optimal growth temperature, maximum growth rate, or maximum critical thermal limit after 200–600 generations of

acclimation to elevated temperature, suggesting a fast adaptation to ocean warming (Jin and Agustí, 2018). A decade-long experiment in outdoor mesocosms showed warm-adapted green alga, Chlamydomonas reinhardtii, had higher optimal growth temperatures, higher competitive fitness, and increasing rates of net photosynthesis compared to its control (Schaum et al., 2017).

Ocean acidification and warming can interactively disturb the advection of nutrients and trace elements in the ocean systems, thereby affecting biogeochemical processes and ecosystem stability. Coastal ecosystems and their services are often degraded due to the effects of human activities and global ocean changes. However, these possible macro-changes need micro-mechanism-based explanations to enhance forecasting reliability. Organisms in the upper pelagic zone of the oceans face multiple environmental stresses, such as OA, warming, lack of nutrients (except in coastal waters near human influences), and increased exposures to UVR (**Figures 1**, **4**). OA and warming together could additively increase respiratory rate, since respiration increases with temperature, and acclimation to OA-induced acid–base disturbance requires extra energy to allow cells to cope with. As a result, OA and warming may cumulatively affect the marine biological pump (MBP) and the MCP, which is suggested to be sensitive to warming (Jiao and Zheng, 2011). Nevertheless, the net effects of OA and warming on MBP/MCP are still quite uncertain.

Deoxygenated waters typically coincide with low pH (Cai et al., 2011); therefore, the combined effects of OA and deoxygenation are of ecological relevance. Along with progressive OA and warming, a consequent drop in the pO2/pCO<sup>2</sup> of seawater is unavoidable (**Figures 1**, **4**). This ratio has been defined as the respiratory index for animals and may be defined as the photorespiratory index for photosynthetic organisms. Although lower values of pO2/pCO<sup>2</sup> are harmful or fatal to many aerobic

Primary producers provide basic energy flows for ecosystems, and consequently, they are closely related to activities of other organisms. The habitable waters for marine organisms are degrading under progressive OA and other ocean climate changes, such as ocean warming, stratification, deoxygenation, and enhanced exposure to UV radiation (**Figure 4**; Gao et al., 2012a; Hutchins and Fu, 2017). Based on documented data and theoretical reasoning, a habitable niche degradation hypothesis can be established as follows (**Figure 4**): phytoplankton food quality is supposed to decline in response to OA (Riebesell et al., 2007; Jin et al., 2015), leading to reduced energy supply to grazers; at the same time, motile organisms dwelling within the UML, such as zooplankton or fish, tend to move to deeper layers to avoid exposure to harmful solar UV radiation (Rhode et al., 2001) and high levels of PAR that traps more heat; however, stressful suboxic or hypoxic and acidified seawater below the thermocline may act as a barrier preventing some species from moving deeper. As a result, the habitable niches for such organisms will shrink, and their habitat will degrade. This might exacerbate the effects of other anthropogenic stressors such as over-fishing, and so contribute to declining fish catches in previously productive waters. As a result, more algal blooms will be expected in eutrophicated waters with increasing levels of

FIGURE 4 | Ocean climate changes and the habitat degradation hypothesis. Ocean warming, acidification, and deoxygenation associated with increasing atmospheric CO<sup>2</sup> rise. A shoaled UML due to warming exposes organisms dwelling there to higher levels of solar radiation. The habitable niche degradation hypothesis: phytoplankton abundance and community structure can be altered within the UML under multiple stressors associated with ocean climate changes; and motile organisms dwelling within the UML are stressed due to increased exposure to solar UV radiation and high levels of PAR, which traps more heat; however, the low O<sup>2</sup> and pH waters below the UML hamper downward migration. (Re-drawn based on Gao, 2018).

FIGURE 5 | Illustration of the hypothesis that OA and UV synergistically induce carbon loss in surface primary producers, based on the observational data on UV impacts (Li et al., 2011) and results that smaller diatoms decrease their growth rate under OA and nutrient-limitation conditions (Li et al., 2018). Diel pH changes in highly productive coastal waters are shown with a sun and a moon symbol to indicate pH rise with increasing photosynthetic C removal during daytime and pH decline with respiratory CO<sup>2</sup> release during night. Note that benthic macroalgae contribute greatly to the diel pH fluctuations, and that their growth and photosynthesis are usually stimulated by OA. Furthermore, they are more tolerant of solar UV radiation [see the review by Gao et al. (2012) and literature therein]. The symbols "+" and "–" indicate more positive effects due to OA and UV in coastal non-nutrient-limited waters and negative effects due to OA and UV in oligotrophic offshore waters, where shoaling of UMLs owing to warming reduces upward transport of nutrients from deeper layers.

ocean climate changes and decreasing top-down pressure. This hypothesis could be tested by laboratory microcosm or mesocosm experiments in combination with in situ investigations.

Effects of OA can be positive, neutral, and negative under influences of other drivers or in different ecosystems where chemical and physical environments are contrastingly different. As discussed above, in coastal waters, fleshy macroalgae such as Porphyra sp. (Gao et al., 1991), Gracilaria sp. (Gao et al., 1993), Ulva (Gao et al., 2016), and Sargassum sp. (Xu et al., 2017), usually increase their growth in response to elevated pCO<sup>2</sup> levels up to 1,000–2,000 µatm, showing tolerance to acidification. Larger diatom species are found to show stimulated growth (Wu et al., 2014) or enhanced carbon assimilation rate (Li et al., 2016), while smaller diatoms tend to show reduced photosynthetic and growth rates under influence of OA (Wu et al., 2014), nitrate limitation, and temperature rise (Li et al., 2018). On the other hand, smaller phytoplankton cells in open ocean showed higher photosynthetic inhibition caused by UV radiation (Li et al., 2011). It is most likely that the combined effects of OA and solar UVR induce more carbon loss from the cost to open ocean in phytoplankton assemblages (**Figure 5**). Such a hypothesis needs observational and experimental data to be tested.

In summary, OA effects under multiple stressors have been documented but only in association with a very limited number of drivers. Future work toward understanding the ecological impacts of ocean climate changes should include both scenario-oriented and mechanism-directed studies. Considering

## REFERENCES


the impractical nature of the number of treatments and replicates needed to accommodate all combinations of possible drivers in experimental designs, it is highly recommended to refer to the recently published guide for multiple drivers marine research (Boyd et al., 2018, 2019). With increasing understanding of multiple stressors effects using controlled experiments in combination with field observational findings, OA effects in different ecosystems under multiple stressors can thus be understood both mechanistically and predictively.

## AUTHOR CONTRIBUTIONS

KG contributed to the theoretical designs, data analysis, and writing of the manuscript. JB, D-PH, JH-S, GG, and DH contributed to the analysis of the data and writing of the manuscript.

## FUNDING

This study was supported by the National Key R&D Program (2016YFA0601400), National Natural Science Foundation (41430967, 41720104005, and 41721005), Joint Project of National Natural Science Foundation of China and Shandong Province (No. U1606404), and the U.S. National Science Foundation (OCE 1538525, 1638804, 1657757, and 1851222).


wind field and ocean forcing on interdecadal time scales. J. Clim. 30, 8987–9004. doi: 10.1175/JCLI-D-16-0016.1



metabolism during a bloom: a coastal mesocosm study. Sci. Total Environ. 633, 618–629. doi: 10.1016/j.scitotenv.2018.03.222




coccolith morphogenesis along a natural CO<sup>2</sup> gradient. Biol. Bull. 226, 282–290. doi: 10.1086/bblv226n3p282

Zou, D., Gao, K., and Luo, H. (2011). Short- and long-term effects of elevated CO<sup>2</sup> on photosynthesis and respiration in the marine macroalga Hizikia fusiformis (Sargassaceae, Phaeophyta) grown at low and high N supplies. J. Phycol. 47, 87–97. doi: 10.1111/j.1529-8817.2010. 00929.x

**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 Gao, Beardall, Häder, Hall-Spencer, Gao and Hutchins. 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.

# An End-to-End Model Reveals Losers and Winners in a Warming Mediterranean Sea

Fabien Moullec<sup>1</sup> \*, Nicolas Barrier<sup>2</sup> , Sabrine Drira<sup>1</sup> , François Guilhaumon1,3 , Patrick Marsaleix<sup>4</sup> , Samuel Somot<sup>5</sup> , Caroline Ulses<sup>4</sup> , Laure Velez<sup>1</sup> and Yunne-Jai Shin1,6

<sup>1</sup> MARBEC, CNRS, Ifremer, IRD, Université de Montpellier, Montpellier, France, <sup>2</sup> MARBEC, CNRS, Ifremer, IRD, Université de Montpellier, Sète, France, <sup>3</sup> Laboratoire d'Excellence CORAIL, ENTROPIE, IRD, CNRS, Université de La Réunion, Saint-Denis, France, <sup>4</sup> Laboratoire d'Aérologie, CNRS, UPS, Université de Toulouse, Toulouse, France, <sup>5</sup> CNRM, CNRS, Météo-France, Université de Toulouse, Toulouse, France, <sup>6</sup> Department of Biological Sciences, Marine Research Institute, University of Cape Town, Rondebosch, South Africa

#### Edited by:

Elizabeth Fulton, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia

#### Reviewed by:

Cameron Ainsworth, University of South Florida, United States Nova Mieszkowska, University of Liverpool, United Kingdom

> \*Correspondence: Fabien Moullec fabien.moullec@ird.fr

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 31 March 2019 Accepted: 05 June 2019 Published: 25 June 2019

#### Citation:

Moullec F, Barrier N, Drira S, Guilhaumon F, Marsaleix P, Somot S, Ulses C, Velez L and Shin Y-J (2019) An End-to-End Model Reveals Losers and Winners in a Warming Mediterranean Sea. Front. Mar. Sci. 6:345. doi: 10.3389/fmars.2019.00345 The Mediterranean Sea is now recognized as a hotspot of global change, ranking among the fastest warming ocean regions. In order to project future plausible scenarios of marine biodiversity at the scale of the whole Mediterranean basin, the current challenge is to develop an explicit representation of the multispecies spatial dynamics under the combined influence of fishing pressure and climate change. Notwithstanding the advanced state-of-the-art modeling of food webs in the region, no previous studies have projected the consequences of climate change on marine ecosystems in an integrated way, considering changes in ocean dynamics, in phyto- and zoo-plankton productions, shifts in Mediterranean species distributions and their trophic interactions at the whole basin scale. We used an integrated modeling chain including a high-resolution regional climate model, a regional biogeochemistry model and a food web model OSMOSE to project the potential effects of climate change on biomass and catches for a wide array of species in the Mediterranean Sea. We showed that projected climate change would have large consequences for marine biodiversity by the end of the 21st century under a business-as-usual scenario (RCP8.5 with current fishing mortality). The total biomass of high trophic level species (fish and macroinvertebrates) is projected to increase by 5 and 22% while total catch is projected to increase by 0.3 and 7% by 2021–2050 and 2071–2100, respectively. However, these global increases masked strong spatial and inter-species contrasts. The bulk of increase in catch and biomass would be located in the southeastern part of the basin while total catch could decrease by up to 23% in the western part. Winner species would mainly belong to the pelagic group, are thermophilic and/or exotic, of smaller size and of low trophic level while loser species are generally large-sized, some of them of great commercial interest, and could suffer from a spatial mismatch with potential prey subsequent to a contraction or shift of their geographic range. Given the already poor conditions of exploited resources, our results suggest the need for fisheries management to adapt to future changes and to incorporate climate change impacts in future management strategy evaluation.

Keywords: biodiversity scenario, climate change, ecosystem model, end-to-end model, OSMOSE, fishing, Mediterranean Sea

## INTRODUCTION

fmars-06-00345 June 21, 2019 Time: 16:38 # 2

Climate change and ocean acidification are altering oceans at rates that have been unprecedented over the last millennia (IPCC, 2014; Howes et al., 2015; Weatherdon et al., 2016). Such changes in ocean conditions have numerous impacts scaling from individuals up to ecosystems, jeopardizing ecosystem goods and services as well as human societies (Brown et al., 2010; Ainsworth et al., 2011; Cheung et al., 2013; Pecl et al., 2017). Following environmental changes, the physiology of marine organisms, population dynamics, ecological interactions, and entire marine food webs are or will be, directly or indirectly, impacted (Parmesan and Yohe, 2003; Cheung et al., 2013; Albouy et al., 2014; Poloczanska et al., 2016; Henson et al., 2017; Miller et al., 2018; Selden et al., 2018). Climate change will affect all ocean organisms and primary productivity, change the composition of marine communities, and alter ecosystem functions such as the production of marine living resources (Brown et al., 2010; Hoegh-Guldberg and Bruno, 2010; Blanchard et al., 2012; Holt et al., 2016).

With growing human populations, rising incomes, and changing dietary preferences, the global demand for fish is expected to increase in the future while climate-induced changes are expected to change future fisheries production patterns dramatically, either by shifting spatial patterns of production as species tend to track their suitable environmental niche or as a result of changes in net primary production (Perry et al., 2005; Brander, 2007; Cheung et al., 2010; Merino et al., 2012; Barange et al., 2014; IPCC, 2014; Cheung, 2018; FAO, 2018). For example, the spatial distribution of fish has been shown to shift toward higher latitude regions or into deeper waters, with rates of range shift of ca. 30–130 km decade−<sup>1</sup> toward the poles and 3.5 m decade−<sup>1</sup> to deeper waters (Cheung et al., 2010; Cheung, 2018). Regarding global primary production, 10 earth system models projected a mean global decrease of 8.6% (±7.9%) under the highest emission scenario RCP8.5 (Representative Concentration Pathway) and a decrease of 2% (±4.1%) under the high mitigation scenario RCP2.6 by 2090, with large regional differences (Bopp et al., 2013). These changes are likely to trigger a global redistribution of the maximum catch potential (MCP) of fishing areas, with MCP and global revenue projected to decrease by 7.7 and 10.4%, respectively, by 2050 relative to 2000 when considering RCP8.5 (Lam et al., 2016). Using a dynamic size-based food web model forced by a physical-biogeochemical model, Blanchard et al. (2012) predicted a decline of 30–60% in potential fish production in some tropical and upwelling areas and an increase in the production of pelagic predators by 28–89% in some high latitude shelf seas by 2050 under the SRES A1B scenario (Special Report on Emissions Scenarios). According to Carozza et al. (2018), climate change could decrease the global fish biomass by as much as 30% by 2100 (RCP8.5), because of changes in primary production and a temperature-driven increase of natural mortality. From an ensemble of ecosystem models included in the Fisheries and Marine Ecosystem Model Inter-comparison Project (Fish-MIP), a 15–30% decline of the total marine animal biomass in the North and South Atlantic, Pacific, and Indian Ocean is projected by 2100, whereas polar ocean basins would experience a 20–80% increase under a high emission scenario (RCP8.5) (Bryndum-Buchholz et al., 2019). It turns out that climate change can significantly alter the availability and composition of commercial fisheries catches, thereby having socioeconomic implications for fisheries, markets, and consumers worldwide (Weatherdon et al., 2016).

The Mediterranean Sea, bordered by Africa, Europe, and Asia (**Figure 1**), is one of the most responsive regions to climate change (Giorgi, 2006; Marbà et al., 2015), with various sources of disturbance interacting synergistically (Coll et al., 2012; Micheli et al., 2013; Ramírez et al., 2018). Several studies conducted in the region have already explored the impacts of climate change on marine populations, species assemblages and ecosystem structures (Ben Rais Lasram and Mouillot, 2008; Ben Rais Lasram et al., 2010; Lejeusne et al., 2010; Albouy et al., 2012, 2013, 2014; Coll et al., 2012; Tsikliras and Stergiou, 2014; Halpern et al., 2015; Marbà et al., 2015; Hattab et al., 2016). Under a high emission scenario (SRES A2 scenario), Albouy et al. (2013) showed that by the end of the century, 54 out of 288 coastal fish species are expected to lose their climatically suitable habitat, species richness would decrease across 70.4% of the continental shelf area and mean fish body size would increase over 74.8% of the continental shelf area. Under the same climate change scenario, Ben Rais Lasram et al. (2010) suggested that the coldest areas of the Mediterranean Sea (i.e., the Adriatic Sea and the Gulf of Lions) would first act as refuges for cold-water species then would become a "cul-de-sac," driving those species toward extinction by the end of the century.

Most of the future projections conducted so far at the Mediterranean basin scale have been based on climate niche models, and none have projected future changes in trophic and ecosystem functioning as well as in biomass evolution or fisheries catch at the whole basin scale under climate change. Several local scale scenarios of climate change impacts involved the Ecopath with Ecosim modeling approach (Coll and Libralato, 2012; Libralato et al., 2015; Corrales et al., 2018), focused on trophic fluxes within food webs, with most of the dynamics being non-spatially explicit (Brander, 2010; Perry et al., 2010; Urban et al., 2016). A few global scale models provided some quantification of the climate-induced changes to be expected for the Mediterranean Sea, but these were typically developed using physical and biogeochemical models at a spatial resolution probably too low to properly reflect the very complex Mediterranean dynamics (e.g., Cheung et al., 2010, 2016, 2018). Yet, not all these studies have studied the consequences of climate change on Mediterranean marine ecosystems in an integrated way: considering changes in ocean dynamics, in plankton production, shifts in species distributions, their life cycles and their trophodynamic interactions. Consideration of these processes is critical to fully address the future of marine biodiversity, and to explore robust mitigation and adaptation strategies in response to global changes. Representing the strength of food web connections and developing holistic approaches are fundamental to project the response of ecosystems under bottom-up and top-down forcing, such as climate-driven changes or over-exploitation of living marine resources (Perry et al., 2010; Grimm et al., 2017;

Seidl, 2017; Cheung, 2018; Nicholson et al., 2018; Peck et al., 2018; Selden et al., 2018). Such scientific progress is needed to support an ecosystem-based approach to marine resources management (EAM) (Garcia et al., 2003; Pikitch et al., 2004; Coll and Libralato, 2012; Coll et al., 2013) and to advance the sustainable use and conservation of the oceans (UN Sustainable Development goal 14; Pecl et al., 2017).

In this context, we used an integrated modeling chain including a high-resolution regional climate model, a regional biogeochemistry model and a food web model (Moullec et al., 2019) to project the potential effects of climate change on biomass and catches for a wide array of species in the Mediterranean Sea, by the middle and end of the 21st century under the high-emission RCP8.5 socio-economic scenario and "business as usual" fisheries management (i.e., current fishing mortality). With this modeling chain, primary and secondary production changes and spatial distribution shift of species induced by climate change were considered. We aimed to explore how climate-induced changes could affect the Mediterranean marine biodiversity, as well as the ecosystem structure and functioning, by using a set of ecological indicators relevant at different scales, from individuals to communities and from ecoregions to the whole Mediterranean Sea.

## MATERIALS AND METHODS

## General Structure of the End-to-End Modeling Chain

In this study, we used a consistent end-to-end modeling chain from global climate to regional marine ecosystem under the RCP8.5 scenario. The RCP8.5 scenario from the IPCC AR5 is characterized by increasing greenhouse gas emissions (GHG) over time, leading to high GHG concentration levels in 2100 (Riahi et al., 2011). It assumes a high population growth rate and relatively slow income growth with modest rates of technological change and energy intensity improvements (Riahi et al., 2011). In terms of expected global temperature increase by the end of the century, the RCP8.5 scenario can be considered close to the IPCC SRES A1F1 and A2 scenario (Rogelj et al., 2012). It was chosen here as the range of projected temperatures were the most frequently explored in the region (Albouy et al., 2014; Hattab et al., 2016; Benedetti et al., 2018; Corrales et al., 2018), hence facilitating comparisons with previous findings.

Our modeling chain includes:


In this chain, CNRM-RCSM4 is driven one-way by atmosphere and ocean lateral boundary conditions extracted from CNRM-CM5 (see **Supplementary Material S1** for details). Eco3M-S is itself driven one-way by the atmosphere and ocean outputs of CNRM-RCSM4. Finally, OSMOSE is driven one-way by the biogeochemistry outputs of Eco3M-S (i.e., by phyto- and zoo-plankton biomass). This end-to-end

<sup>1</sup>http://www.osmose-model.org

model is fully described in Moullec et al. (2019), only a brief presentation of the structure and parameterization is given in the present study. Details on the OSMOSE model can be found at https://documentation.osmose-model.org/index.html.

Despite the complexity of the modeling chain, we consider that it represents the best solution to date to combine in a consistent way and at high-resolution all the drivers required to assess the future evolution of the Mediterranean upper trophic species. We acknowledge, however, that we are exploring here only one possible modeling chain among a large ensemble of possibility and in particular, we do not explore the uncertainty related to the choice of each modeling block.

## The Regional Biogeochemistry Model Eco3M-S

Eco3M-S is a biogeochemical model which simulates the lower trophic part of the food web. It represents several elements' cycles such as carbon, nitrogen, phosphorus and silica in order to reproduce the different limitations and co-limitations observed in the Mediterranean Sea and the dynamics of different plankton groups (Auger et al., 2011). Seven planktonic functional types (PFTs), characterized by a specific size range and representing the main PFTs of the Mediterranean Sea were modeled: Pico- (0.7–2 µm, mainly Synechococcus spp.), nano- (2–20 µm, mainly dinoflagellates), and micro-phytoplankton (20–200 µm, mainly diatoms); nano- (5–20 µm, mainly bacterivorous flagellates and small ciliates), micro- (20–200 µm, mainly ciliates and large flagellates), and meso- zooplankton (>200 µm, mainly copepods and amphipods) and heterotrophic bacteria (not considered in the present study). All features, formulations, and parameterization of biogeochemical processes integrated in the mechanistic Eco3M-S model were described in details by Auger et al. (2011) and Ulses et al. (2016).

The coupling between Eco3M-S and OSMOSE was realized through (i) the spatial distribution of high trophic level species (HTL) in OSMOSE and (ii) the predation process with the planktonic organisms from Eco3M-S serving as potential prey fields (in the form of biomass) for the HTL species in OSMOSE. As within OSMOSE, predation upon planktonic groups is an opportunistic size-based process (Travers-Trolet et al., 2014a) controlled by a minimum and a maximum predation size ratio parameter.

## The High Trophic Level Model OSMOSE

OSMOSE is a multispecies and individual-based model, spatially explicit and representing the whole life cycle of several interacting marine species from eggs to adult stages. Major processes of the life cycle, i.e., growth, predation, reproduction, natural, and starvation mortalities as well as fishing mortality, are modeled with a time step of 2-weeks in this study. Species interact through predation in a spatial and dynamic way (Shin and Cury, 2001, 2004). The predation process occurs when there are both spatio-temporal co-occurrence and size compatibility between a predator and its prey. The model is forced by species-specific spatial distribution maps (one unique map per species in this study) (see section "Current and Future Species Geographic Distributions"). A maximum and a minimum predator/prey size ratio are defined to constrain predator prey interactions. The food web structure thus emerges from these local individual interactions (Travers et al., 2009; Travers-Trolet et al., 2014b).

OSMOSE covers the whole Mediterranean basin with a regular grid of 20 km × 20 km counting 6229 cells. It represents the Mediterranean food web from plankton production to main apex predators on the 2006–2013 period (Moullec et al., 2019). Ninetyseven high trophic level species (82 fish species, 5 cephalopod species, and 10 crustacean species, mainly shrimps) were modeled, accounting for around 95% of total declared catches in the region during the 2006–2013 period. Modeled species were selected according to their ecological and economic importance and Data Availability (Moullec et al., 2019). For this study, three amphihaline fish species (i.e., Alosa alosa, Alosa fallax, and Anguilla anguilla) were removed from the previous version of the model because their complex life cycle characterized by movements between fresh-water and salt-water has not been modeled as being influenced by climate change. A benthos compartment was added and modeled with just a few parameters (i.e., size range, trophic level, and biomass) to take into account the diet specificity of some HTL species that partly feed on benthic invertebrates (e.g., crustaceans, polychaetes) (Moullec et al., 2019). The biological parameters linked to somatic growth (Von Bertalanffy parameters, length-weight relationship parameters), mortalities (longevity, additional natural mortality that is not explicitly represented in OSMOSE, age/size at fisheries recruitment), reproduction (size at maturity, relative fecundity) and predation (minimum and maximum predation size ratios, maximum ingestion rate), along with their sources, are detailed in the **Supplementary Tables S1, S2**.

## Implementation of the Future Scenario

Current and Future Species Geographic Distributions A niche modeling approach based on environmental data was used to generate species presence/absence maps in the Mediterranean Sea and drive species spatial distributions in OSMOSE (Moullec et al., 2019). Environmental predictor variables, i.e., temperature, salinity, were extracted from the World Ocean Atlas 2013 version 2<sup>2</sup> which provides observed climate data over the 1975–2012 period. To take into account the vertical distribution of species in the water column, six environmental metrics were derived from monthly temperature and salinity climatologies: mean sea surface temperature and salinity (0–50 m depth), mean vertical temperature and salinity (0–200 m depth), and mean sea bottom temperature and salinity (50 m – maximum bathymetry depth).

Current geographic distributions were modeled using an ensemble forecasting approach involving eight climate suitability models embedded in the freeware BIOMOD2 R package (Thuiller et al., 2009; R Core team, 2015) (see Moullec et al., 2019 for details on models parameterization and assumptions). The niche models developed and calibrated under present conditions were then used to project the environmental niche of species to the 2021–2050 and 2071–2100 periods using future environmental

<sup>2</sup>https://www.nodc.noaa.gov/OC5/woa13/woa13data.html

predictors. A threshold approach maximizing the fit with current species distribution was used to predict the geographical range of the species under both current and future environmental conditions. In this study, only the current and future geographic distributions of species already present in the Mediterranean Sea before the year 2013 and for which the biological knowledge necessary for the parameterization of the OSMOSE model is available were modeled. We included potential invasive species, which distribution centroid and main abundance are located outside the Mediterranean Sea, but which have been observed in Mediterranean waters, even in small numbers. No new introduction of non-indigenous marine species has been considered.

Sea temperature and salinity values at different depth strata were obtained for the historical period (1970–2005), the middle (2021–2050), and the end of the 21st century (2071–2100) from CNRM-RCSM4. For projecting the future species geographical distribution, a deltas method was used: anomalies between the historical simulated period (1970–2005) and the future projected periods were calculated and applied to current climate temperature and salinity climatologies to create future environmental conditions.

#### Future Plankton Productions

The same deltas approach was followed for the biogeochemistry forcing extracted from Eco3M-S: anomalies between the historical and future time periods (2021–2050 and 2071–2100) were calculated and applied to current plankton biomasses.

Transient biogeochemical simulations were performed over historical (1950–2005) and future (2006–2100, RCP8.5 scenario) periods using the Eco3M-S model, forced by the physical model CNRM-RCSM4. The historical simulation was initialized using the MEDAR-MEDAtlas database (Manca et al., 2004) as in Kessouri (2015). The final state of this simulation was then used to initialize the scenario simulation. Terrestrial, atmospheric inputs and nutrient concentrations in the Atlantic have been kept constant from 1950 to 2100. An average of nutrient loads over the period 1960–2000, based on regional estimates by Ludwig et al. (2010), was imposed at the mouths of the 173 rivers considered. The atmospheric deposition of dissolved inorganic nitrogen has been determined on the basis of studies by Ribera d'Alcalà (2003), Powley et al. (2017), and Richon et al. (2018); and the phosphate deposition has been derived from a climatology of Saharan dust deposits simulated by the regional model ALADIN-Climat (Nabat et al., 2015; Richon et al., 2018). Nutrient profiles applied in the Atlantic buffer zone were prescribed using monthly profiles from the World Ocean Atlas 2009 climatology (Garcia et al., 2006).

### Assessing Climate Change Effects on Mediterranean Marine Biodiversity With OSMOSE

We used OSMOSE to project potential changes in biomass and catch of high trophic level species (fish, cephalopods, and crustaceans) by the middle (2021–2050) and end of the 21st century (2071–2100) under the high emission RCP8.5 scenario and current fisheries exploitation level (fishing mortality and size of recruitment were held constant). For each future time period (each spanning 30 years), climate and biogeochemical forcing variables were used as climatologies. All the parameters relating to growth, reproduction, predation or mortality of the modeled species were kept similar between scenarios (except predation mortality which varies dynamically and is an outcome of the model). Given the inherent stochasticity of OSMOSE (the main source of stochasticity lies in the species movement within their habitat and the order at which schools interact (through predation) (Moullec et al., 2019), ten replicated simulations by time period were run and averaged. For each of the three time slices (current, 2021–2050 and 2071–2100), simulations were run for 110 years to ensure sufficient spin-up time and only the last 10 years were averaged to analyze the outputs.

To assess climate change impacts on Mediterranean marine biodiversity, a range of output indicators, including total biomass and catch, were analyzed and compared between the current (2006–2013) and future time periods, 2021–2050 and 2071–2100. Trophic indicators were used to assess potential changes in food web structure and functioning: the Mean Trophic Level of the community (MTLc) (Pauly et al., 1998), the High Trophic Indicator (HTI), which represents the proportion of biomass of predators with a trophic level higher or equal to 4 (Bourdaud et al., 2016). The percentage of biomass within different body size-classes (<10 cm; 10–20 cm; 20–30 cm; 30–40 cm, and >40 cm) was also used to assess climate change impacts on ecosystem structure. Note that the proportion of total biomass that exceeds a threshold length of 40 cm is equivalent to the Large Fish Indicator (LFI) that is a key indicator monitored in European waters to assess ecosystem impacts of fishing (Modica et al., 2014). All analyses were performed using R version 3.5.1 (R Core Team, 2018).

## RESULTS

## Current and Future Environmental Conditions

During the historical period (1970–2005), CNRM-RCSM4 estimated that the annual mean Sea Surface Temperature (SST; 0–50 m depth) and the mean Sea Surface Salinity (SSS; 0–50 m depth) of the Mediterranean Sea were 17.6◦C (±1.3◦C; standard deviation) and 37.9 practical salinity unit (±0.7 PSU; standard deviation), respectively (see **Supplementary Figures S1, S2**). The Gulf of Lions and the Northern Adriatic Sea were identified as the coolest areas (with a mean SST of 15.3 and 15.6◦C, respectively) while the Levantine Sea and the Gulf of Gabes were identified as the warmest areas (with mean SST of 19.4 and 18.9◦C, respectively). Under the RCP8.5 emission scenario, CNRMS-RCSM4 projected a spatially homogeneous warming and a more regionally contrasted salinification of the Mediterranean Sea by the end of the century (**Supplementary Figures S1, S2**). The Mediterranean Sea was projected to warm by 0.9◦C (±0.05◦C) globally for 2021–2050 and by 2.51◦C (±0.16◦C) for 2071–2100 with respect to 1970–2005. By the end of the century, the projected increase in mean SST was highest in the Levantine Sea and the Western Ionian Sea (+2.7◦C). In parallel, the SSS is expected to increase, with marked regional differences, by 0.13

PSU (±0.13 PSU) for the 2021–2050 period and to come back to its current global climate value (±0.01 PSU) for the 2071–2100 period (**Supplementary Figures S1, S2**). By 2021–2050, the SSS of the Adriatic Sea was projected to increase by 0.35 PSU while that of the Alboran Sea (next to the Strait of Gibraltar) was projected to decrease by 0.1 PSU. By the end of the century (2071–2100), due to the evaporation increase, the precipitation decrease and the strong decrease in the Po freshwater input, SSS was found to increase by 0.55 PSU in the Adriatic Sea, while due to changes in Atlantic waters inflow characteristics, SSS may decrease by 0.65 PSU in the Alboran Sea.

## Current and Future Plankton Productivity

Under RCP8.5, projections of Eco3M-S showed a relative stability of the overall biomass of phytoplankton by mid-century (2021–2050) compared to the current period (**Supplementary Figure S3**). This global stability conceals a biomass increase of the smallest groups of phytoplankton such as the picoand nano-phytoplankton (by 10 and 4%, respectively) and a decrease of ca. 6% in the biomass of the largest size group (i.e., microphytoplankton). The biomass of zooplankton followed similar trends with a slight increase of the smallest size groups (3 and 4% increase for nano- and micro-zooplankton, respectively) and a very low increase of biomass of 1% for the mesozooplankton group (**Supplementary Figure S4**). Significant changes of primary and secondary productions appeared toward the end of the century. Projections showed an overall increase of phytoplankton biomass at the whole Mediterranean scale, due to a large gain of biomass for the smaller sized organisms (pico- and nano-phytoplankton biomass were projected to increase by 28 and 13%, respectively), but a decrease by 15% for microphytoplankton biomass was expected (**Supplementary Figure S5**). Likewise, climate changes are projected to favor the most opportunistic zooplankton class in the model with an increase of biomass of 8, 19, and 7% for nano-, micro-, and meso-zooplankton groups, respectively (**Supplementary Figure S6**). For both time periods, changes of plankton productivity were spatially heterogeneous and a more pronounced increase of plankton productivity was projected in the eastern basin, compared to the western part.

## Current and Future Species Geographic Distribution

By 2021–2050, under the RCP8.5 scenario, the geographic range of 12 species (12.4%) was projected to shrink whereas 16 species (16.5%) were projected to increase their geographic range (**Supplementary Table S3**). By the end of the century (2071–2100), while the number of species gaining in geographic range remained relatively stable (14 species), the proportion of species projected to lose suitable habitat increased by fifty percent (24 species) to reach almost a quarter of the Mediterranean modeled fauna. Among the 24 "losers," Micromesistius poutassou was expected to contract its geographic range by 95%, with a distribution becoming extremely fragmented. The projections reported high variations in the size of species distribution areas through time. By 2021–2050, the average loss and gain in species distributional range were 22 and 32%, respectively, whereas by 2071–2100 they were 26 and 174%, respectively. Gains in range size were mainly due to some thermophilic alien species (i.e., Etrumeus teres, Caranx crysos, Sphyraena viridensis, Stephanolepis diaspros, and Upeneus moluccensis) originally restricted to small areas of the Mediterranean Sea and which found, with changes in environmental conditions, new suitable habitats across the basin. Some species of high commercial interest exhibited contrasted evolution of their range between the two time periods. For instance, European hake (Merluccius merluccius) was expected to gain up to 9% of potential suitable climatic habitat at first (2021–2050), but then to experience a range reduction of 15% toward the end of the century (2071–2100).

## Projected Changes in Biomass of the High Trophic Level Species

At the Mediterranean scale, considering projected changes of plankton productivity and species geographic distribution under the high emission scenario RCP8.5, climate change is projected to increase the total biomass of all high trophic levels species by 5 and 22% by 2021–2050 and 2071–2100, respectively (**Figures 2A,B**). Changes in biomass globally reflected the changes in primary and secondary productions. For both future periods, the gain in biomass was more important in the eastern basin and especially in the Levantine Sea (**Figure 2A**). In this area, some thermophilic exotic species, here qualified as "winner" species, such as E. teres, Saurida undosquamis, S. diaspros, and U. moluccensis, benefited from an increase in their geographic range, as well as an increase in plankton productivity, especially for the planktivorous fish species such as E. teres. Biomass of this latter species has been found to potentially boom 70-fold by the end of the 21st century while the biomass of C. crysos or S. undosquamis could be multiplied by 2 and 48, respectively.

Future changes in biomass are expected to slightly differ depending on the vertical distribution of species in the water column (**Figure 2B**). By the middle of the century, the biomass of demersal species could increase by ca. 3% whereas benthic biomass could decrease by 2%. Pelagic species, with an increase in biomass of 7%, could benefit the most from the increase in plankton productivity (**Figure 2B**). Nevertheless, the global gain of biomass by 2021–2050 masked some loser species. For instance, biomass could be reduced by 6% for Dicentrarchus labrax, by 4% for M. merluccius, by 20% for Spondyliosoma cantharus, by 6% for Octopus vulgaris, by 7% for Scomber scombrus, and by 5% for Diplodus vulgaris. In addition, three species, for which a reduction in range had not been predicted by niche models, were projected to be on the verge of collapse, with a decrease of more than 50% in their biomass, and four species were projected to become quasi extinct with a 90% decrease in their biomass by 2021–2050.

By the end of the 21st century, with a projected increase of ca. 3%, the biomass of demersal species could remain

stable compared to 2021–2050 (**Figure 2B**). However, the biomass of pelagic species was projected to increase by more than 25% and that of benthic species by 32% compared to the baseline period. Despite the global increase, the biomass of some species of high commercial interest are expected to decline, for instance, M. merluccius and Scomber scombrus biomass could decrease by 26 and 15%, respectively. Among the losers, ten species were projected to suffer from a drastic reduction exceeding 50% of their current biomass, and among these species, five were projected to become extinct, following a reduction exceeding 90% of their current biomass. On the other hand, the biomass of other species of commercial interest, mainly pelagic species such as Engraulis encrasicolus, Coryphaena hippurus, Thunnus thynnus, or Sardina pilchardus, are expected to increase by 35, 34, 9, and 6%, respectively.

Regional contrasts in biomass changes can be observed in the projections (**Figure 3**). By the middle of the century, along the longitudinal gradient, from 0◦ to 32◦E, total biomass is expected to increase very moderately by ca. 2.5%. Changes could likely be more pronounced in the most western part, between 0◦ and 6◦E, and in the most eastern part of the Mediterranean Sea with a total biomass gain of up to 27 and 90%, respectively. Overall, by the end of the century, projected changes in biomass showed similar spatial patterns, but with higher magnitudes of changes (**Figure 3**). With the continued northward and westward expansion of the ranges of some exotic species and higher planktonic productivity toward the end of the 21st century, biomass gain could continue and reach up to + 50% between 15◦E and 21◦E (South Ionian Sea) and up to + 61% in the Levantine Sea (26◦E). Analysis of biomass changes along the latitudinal gradient revealed an increasing trend of biomass from north to south (**Figure 3**). As with longitudinal changes, projected changes of biomass by 2021–2050 and 2071– 2100 showed a similar pattern but of different magnitude. Between 30◦N and 35◦N, the increase in biomass is projected to reach up to 25% by 2021–2050 and up to 66% by 2071–2100.

By mid-century, changes in biomass were rather homogeneous over the continental shelf and the offshore area and along longitudinal and latitudinal gradients except for some local zones. For instance, between 5◦E and 7◦E, a decrease in biomass of up to 39% was projected on the continental shelf (mainly Balearic island and Algerian coastal zone) while a relative stability in biomass was projected in the offshore area (**Figure 4**). By the end of the century, the continental shelf and the offshore area exhibited more pronounced differences along longitudinal and latitudinal gradients, especially in the easternmost regions where biomass increases were greater in offshore areas than on the continental shelf. Along the latitudinal gradient, between 36◦N and 45◦N, biomass increases were found to be generally higher on the continental shelf than in the offshore area. This trend is reversed at latitudes below 36◦N where the increase in biomass was much higher offshore.

## Projected Changes in Size Structure

The analysis of the proportion of biomass within different size-classes showed no substantial change by 2021–2050 but a very slight increase (+3%) of medium-sized individuals (20–30 cm) and a slight decrease (−6%) of very large-sized individuals (>40 cm) (**Figure 5**). By the end of the 21st century, the proportions of biomass in the two smallest sizeclasses (<10 cm; 10–20 cm) were projected to increase by 3 and 7%, respectively, while the proportions of biomass of medium-sized individuals, large-sized individuals and very largesized individuals were projected to decrease by 8, 15, and 15%, respectively (**Figure 5**).

## Projected Changes in Trophic Indicators

The two trophic indicators, namely the High Trophic Indicator (HTI) and the Mean Trophic Level of the community (MTLc) showed the same downward trend for the two future periods (**Figure 6**). The HTI is projected to decrease by 5 and 15% by 2021–2050 and 2071–2100, respectively. Logically linked to the increase in the biomass of pelagic species (mainly planktivorous fish species), the MTLc is predicted to decrease by 0.4 and 2% by the middle and end of the 21st century, respectively.

## Projected Changes of Catch

Annual fisheries catches simulated by OSMOSE amounted to 788 043 t for the current period. By 2021–2050, under RCP8.5 and "business as usual" fisheries management, the total projected catches in the Mediterranean Sea are expected to remain stable (**Figures 7A,B**). By the end of the century, the total catches could rise by ca. 7% to reach 840 008 t. However, this projected increase

FIGURE 4 | Projected relative changes in total biomass (all high trophic levels species confounded) between the current (2006–2013) and future periods (2021–2050, top; 2071–2100, bottom) in continental shelf (depth ≤200 m) and offshore (>200 m) under emission scenario RCP8.5. The dotted line indicates no change in total biomass.

hides a substantial heterogeneity between species and between management units (i.e., Geographical Sub-Areas; GSA).

By the middle of the century, simulated catches showed either a downward trend in most GSAs (up to −22% in South Tyrrhenian Sea (GSA 10), −9% in Balearic Island and in Southern Adriatic Sea (GSA 5 and 18, respectively) or −7% in South of Sicily (GSA 16), or a relative stability (i.e., increase of less than 2%) (**Figure 7A**). As with the projections of total biomass, it was in the Levantine Sea (GSA 27) that catches are expected to increase the most (up to +42%), mainly due to the biomass explosion of two exotic species (E. teres and S. undosquamis). In the Alboran Sea (GSA 1 and 3), Northern Spain (GSA 6), Gulf of

Lions (GSA 7), and Aegean Sea (GSA 22), catches were projected to increase by between 7 and 9% mainly due to an increase in the catch of small pelagic species such as E. encrasicolus (+6%).

The spatial patterns of catch are projected to change radically by the end of the century (**Figure 7A**). Three regions could be distinguished: the western Mediterranean, the eastern Mediterranean and the Adriatic Sea. By 2071–2100, in all the western Mediterranean Sea, catches are expected to decrease by between 2 and 22% (−22% in Balearic Island, −19% in the southern Tyrrhenian Sea, −14% in Northern Spain and −13% in Algerian and Tunisian waters, for instance). In the Adriatic Sea, catches were projected to remain stable with an increase of ca. 2% in the northern part (GSA 17) and a decrease of ca. 3% in the southern part (GSA 18). By contrast, due to a large increase of catches of some exotic species in the eastern Mediterranean Sea, all the GSAs of this part of the basin were projected to experience an increase in catch by between 8% (eastern Ionian Sea) and 47% (Cyprus Island).

Depending on the vertical distribution of species, differential responses to future climate change could be observed (**Figure 7B**). Projections suggested a moderate to low increase in the catches of demersal and pelagic species, of 2 and 0.6%, respectively, and a decrease in the catches of benthic species of ca. 10% by 2021–2050 (**Figure 7B**). Among demersal catches, those of M. merluccius, one of the main exploited species, are expected to decrease by 4% while Boops boops catches are expected to increase slightly by 2%. Among pelagic species, E. encrasicolus catches were projected to increase by 6% while Sarda sarda catches were projected to decrease by 7%. Finally, among benthic species, Mullus barbatus catches are expected to increase by 3% while Mullus surmuletus catches were projected to decrease by up to 2%. By 2071–2100, some trends are expected to be reversed or amplified with a reduction in demersal catches of about 2%, an increase in pelagic catches of 9% and a substantial increase in the catches of benthic organisms by nearly 16% (**Figure 7B**). Among the main exploited species, M. merluccius catches are expected to fall by 26% compared to current catches, while E. encrasicolus catches could increase by nearly 35% and Mullus barbatus catches are expected to decrease by just over 3%. The increase in catches of thermophilic and/or exotic species is the main cause of the overall increase in projected catches by the end of the century. According to the business-as-usual fishing mortality scenario considered in this study, the catches of exotic species modeled in OSMOSE are expected to increase by an average factor of 40.

## DISCUSSION

## Advances, Limits, and Perspectives

Under climate change, the Mediterranean climate is getting warmer and drier, causing large-scale changes in the Mediterranean Sea and associated marine biodiversity with significant implications for marine ecosystems and the livelihoods that they support (Somot et al., 2006; Coll et al., 2010; Macias et al., 2014, 2015; Adloff et al., 2015; Marbà et al., 2015; Ramírez et al., 2018). Many studies have already shown, assessed or modeled potential impacts of climate change on Mediterranean marine ecosystems (Galil, 2000; Giorgi and Lionello, 2008; Lejeusne et al., 2010; Albouy et al., 2013; Cramer et al., 2018). Most of them focused

on a specific compartment, whether biotic (e.g., Ben Rais Lasram et al., 2010; Benedetti et al., 2018) or abiotic (e.g., Richon et al., 2019). Most of them were conducted at local scales, at the scale of the continental shelf (e.g., Albouy et al., 2014; Hattab et al., 2014) or for specific ecosystems (e.g., Libralato et al., 2015; Corrales et al., 2018). To our knowledge, the

present study is the first attempt to project the effects of climate change at the whole Mediterranean scale in an integrated way, considering explicit and consistent changes in regional climate, ocean dynamics, nutrient cycle, plankton production, shifts in species distributions, their life cycles and their trophodynamic interactions. Nearly one hundred high trophic level species were explicitly modeled in the modeling chain set for this study. Despite the significant progress that our end-to-end modeling chain represents to project the potential effects of climate change on populations, communities, and ecosystems structure, some limits still remain in the model projections as the results presented here are subject to several sources of uncertainty.

The first uncertainty lies in the choice of specific physical (RCSM4) and biogeochemistry (Eco3M-S) models to project the future evolution of the regional climate, the Mediterranean Sea physics and the plankton productivity that were used to force the high trophic level model OSMOSE. These choices were constrained by the existence of a very limited number of consistent hydrodynamic-biogeochemical projections developed at the Mediterranean scale, at high resolution, and for which the most up-to-date greenhouse gas emission scenarios (i.e., IPCC RCPs) were implemented. The low trophic model Eco3M-S simulated a significant increase in phytoplankton (1 and 11% in the western and eastern sub-basins, respectively) and zooplankton (5 and 15% in the western and eastern sub-basins, respectively) biomass, with an increasing contribution of small phytoplankton by the end of the 21st century. The simulated evolution of phytoplankton community structure in response to the extension of the stratified period is consistent with previous observational and modeling studies (Karl et al., 2001; Bopp et al., 2005; Morán et al., 2010; Herrmann et al., 2014). The increase of total plankton biomass obtained can be attributed to an increase in metabolic processes due to surface water warming, as well as by an increasing water inflow and associated nutrient supply at the Gibraltar Strait, which accelerated after the 1950s. Primary production, grazing and recycling processes are temperature sensitive in Eco3M-S model (Auger et al., 2011). Their rates are influenced directly by temperature through an Eppley-type

formula (Eppley, 1972) of the form of Q (T−T1 ) T2 <sup>10</sup> (where Q<sup>10</sup> and T were empirical constants, T<sup>1</sup> = 14 and T<sup>2</sup> = 10). Eco3M-S results are consistent with previous studies in which integrated primary production increased with the direct effect of temperature and an increasing stratification (Sarmiento et al., 1998; Karl et al., 2001; Taucher and Oschlies, 2011; Herrmann et al., 2014). In particular, Taucher and Oschlies (2011) showed that the response of primary production to climate change strongly varies according to the temperature sensitivity in model equations of primary production and recycling processes, with a change of direction in primary production evolution if temperature influence is directly taken into account or not. However, other studies obtained a decline in primary production due to reduced vertical nutrient supply into the photic layer with the weakening of vertical mixing (Steinacher et al., 2010; Bopp et al., 2013). Here, the direct effect of temperature prevailed over the decline of vertical nutrient supply. This is consistent with the study of Herrmann et al. (2014) who obtained no significant change in phytoplankton biomass, but significant increase in zooplankton biomass and primary production in the north-western Mediterranean Sea where a weakening of deep convection was projected under the SRES A2 scenario. This evolution is, however, in contrast to that simulated by Richon et al. (2019) with a decline in zooplankton biomass for the 21st century in the whole Mediterranean basin, under the SRES A2 scenario. The discrepancies of Eco3M-S results with the latter study may be partly explained by differences in nutrient supply at the Gibraltar Strait. In our study, the annual input of nutrients at the Gibraltar Strait was increasing over the whole future period. Thus, the impacts of climate change on the Mediterranean Sea could be modulated by the choice of the near-Atlantic surface water evolution, an uncertain element in General Circulation Models (Adloff et al., 2015). Furthermore, in this study variations over the last decades and future changes in nutrient river loads were not taken into account in the Eco3M-S simulation as no consistent projections until the end of the 21st century exist, partly due to the difficulties of predicting socio-economic decisions (Ludwig et al., 2010). However, the study of Lazzari et al. (2014) showed that an increase in nutrient terrestrial inputs could lead to an increasing primary production close to river mouths. More complex scenarios will be assessed with the Eco3M-s model in future works. Outputs trends of Eco3M-S, related to the structure and parameterization characteristics of the model, influence the overall trends of our results. One approach to overcome individual model uncertainties and limitations would be to force OSMOSE with an ensemble of several hydrodynamicbiogeochemical coupled models when they are available for the Mediterranean sea to estimate mean future trends and associated inter-model spread (Lotze et al., 2018).

Despite the many ecological processes integrated explicitly in OSMOSE, a number of simplifications were mandatory to render the parameterization, the calibration of the model and the simulations tractable. For example, the effects of changes in temperature, oxygen content or pH, on the ecophysiology as well as the feeding and intrinsic mortality rates and behavioral capabilities of marine organisms were not considered in our projections (Pauly, 2010; Cheung et al., 2011, 2013). Yet, such ecophysiological changes could affect life history traits, life cycles and key ecological processes such as predator-prey interactions (Cheung et al., 2013; Mazumder et al., 2015; Allan et al., 2017) and thus could dampen or exacerbate the projected effects of climate change on ecosystem structure and functioning (e.g., Beaugrand and Kirby, 2018). Likewise, OSMOSE does not consider the adaptive potential, whether phenotypic or evolutionary, of marine organisms to climate change stressors. When the magnitude and velocity of changes are moderate, adaptation can buffer substantially the effects of climate change on marine organisms and ecosystems (Crozier and Hutchings, 2014; Boyd et al., 2016; Beaugrand and Kirby, 2018).

In our study, a source of uncertainty also lies in the choice, for methodological reasons, to not model and consider adaptive behavior of fishermen to potential changes in species abundance and distribution. We caution that a constant fishing mortality scenario, implying no changes in fishing effort, technology, management and conservation, is simplistic and could influence

projected biomass and catch trends (Lotze et al., 2018) but this type of scenario allows to focus and isolate climate change effects on marine animal biomass (Bryndum-Buchholz et al., 2019). As mentioned by Cheung et al. (2010), expliciting changes in fishing dynamics is yet important in evaluating climate change impacts and needs to be incorporated in future analyses. Our results are most likely conservative with regard to the projections of biomass and catches toward the end of the century. Climate change is only one component of global change. In the Mediterranean Sea, perhaps more than elsewhere, climate change is likely to act in synergy with other increasing anthropogenic disturbances such as pollution, eutrophication, overexploitation of resources and habitat modification and destruction, all of which playing a major role in altering the structure and functioning of ecosystems (Crain et al., 2008; Ben Rais Lasram et al., 2010; Pörtner, 2010; Pörtner and Peck, 2010). Our projections did not consider the effects of climate change on key fish habitats such as seagrass beds which act as nurseries for several species of high commercial interest and are already threatened by the rapid warming of the Mediterranean Sea (Hoegh-Guldberg and Bruno, 2010; Marbà and Duarte, 2010; Jordà et al., 2012). Changes in the biomass and geographical distribution of benthic invertebrates were also overlooked in the present study, yet subject to climate change effect and playing a major role in marine biogeochemistry and as food source for many high trophic level species (Hiddink et al., 2015). In addition, a recent study suggests that species distribution models such as those used here for forcing OSMOSE may underestimate the potential spread of invasive species (i.e., Lessepsian species) in the Mediterranean Sea thus leading to an underestimation of the subsequent changes on marine biodiversity (Parravicini et al., 2015). Finally, our projections did not consider potential ingression of Atlantic thermophilic species through the Gibraltar Strait or future settlement of new Lessepsian species through the Suez Canal. With the expected changing environmental conditions by the end of the century, it is most likely that the number of invasive species would increase and may have significant environmental, socio-economic and human health impacts (Ben Rais Lasram and Mouillot, 2008; Mannino et al., 2017).

## Structure and Functioning of the Mediterranean Sea Ecosystem Under Climate Change

Our results show that the high greenhouse gas emission scenario RCP8.5 could lead to a warmer Mediterranean Sea with large variations of salinity conditions toward the end of the 21st century relative to the current period. Such physical changes are expected to change the biogeography of marine organisms with many species expanding or shifting their distribution areas northward and westward. These results are in line with previous studies projecting future spatial distributions of fish species on the Mediterranean continental shelf based on global warming scenarios (e.g., Ben Rais Lasram et al., 2010; Albouy et al., 2012, 2013).

The rise of plankton productivity which is projected by Eco3M-S, mainly in the Alboran Sea and in the southeastern of the Mediterranean Sea, associated with species' range shift, could lead to an increase in biomass and total catches at the Mediterranean scale. Two processes were at the origin of these changes: in the most western part, a higher planktonic productivity allowed, by bottom-up effect, an increase in the biomass of high trophic levels species while in the eastern part, the increase of biomass resulted from a higher planktonic productivity combined to the extension of the distribution areas of thermophilic and/or exotic species. Several studies have already shown the importance of bottom-up control of the Mediterranean ecosystem, generally considered as an oligotrophic system in which productivity of higher trophic levels is under the control of primary productivity (Macias et al., 2014; Lynam et al., 2017). Macias et al. (2014) have for instance demonstrated that during the last 50 years the control of marine productivity in the Mediterranean Sea, from plankton to fish, was principally mediated through bottom-up processes.

The general projected increase in total biomass and catch is principally due to the increase in biomass of small pelagic species and thermophilic exotic species such as the lizardfish S. undosquamis and the red-eye round herring E. teres, indicating that climate change may produce "winners" and "losers" among Mediterranean species. Winners are clearly thermophilic planktivorous species that are projected to benefit both from an increase of their spatial range and an increase of available food within their range. This favorable association of thermal and trophic niches partly explains the projected evolution of biomass and catch in the Levantine Sea. With another trophic model, Corrales et al. (2018) have also shown that according to climate change scenarios, primary producers and alien fish species were expected to increase the total biomass on the Israeli continental shelf, masking the reductions of the biomass of native species. Based on our integrated modeling, two major processes of change emerged, i.e., the meridionalization and the tropicalization of the Mediterranean Sea during the 21st century, in line with previous findings (e.g., Boero et al., 2008; Azzurro et al., 2011).

According to our results, pelagic species, mainly the small ones, would be the main winners of climate-induced changes. This finding is in accordance with the study of Hattab et al. (2016) who projected that the future food webs of the Gulf of Gabes would be composed of smaller-sized species under a high emission scenario. Smaller sized species, with higher biomass turn-over rate, tend to show larger changes in biomass in response to environmental modifications than larger species with slower biomass turn-over (Brown et al., 2010). It has also been shown that short life span species have benefited from the increase in water temperature in the basin over recent decades (Tzanatos et al., 2014). The increase in the prevalence of low trophic levels and small sized species in the ecosystem by the end of the century may have consequences for both ecosystem functioning and fishing sustainability. Indeed, planktivorous fish species play a central role in food webs and have the potential to initiate complex cascading effects across and between trophic levels thus modifying the trophic functioning of ecosystems. Small pelagic species are more sensitive to climate variability and are subject to more pronounced variability in recruitment under environmental fluctuations (Hsieh et al., 2006; Ottersen et al.,

2006; Perry et al., 2010). With climate change, the mean turnover rate of marine communities is expected to increase due to the relative increase in the proportion of smaller individuals with higher metabolic rates. Thus, by favoring the dominance of shortlived prey populations and strengthening the already important bottom-up control in the basin, climate change might increase the vulnerability of the Mediterranean Sea in synergy with other drivers of change, in a context where fishing pressure has already led to an alteration of the life history traits and demographic structure of exploited populations in the Mediterranean Sea (Colloca et al., 2013, 2017).

There will be winner but also loser species under climate change. In our study, the variation in biomass of loser species can be explained by a shift or contraction of their geographic range leading to spatial mismatch between previously interacting predators and prey. Indeed, climate-induced changes have a strong potential to alter interspecific trophic interactions by modifying the degree to which predators and prey overlap in space and by creating or eliminating prey spatial refugia (Schweiger et al., 2008; Chevillot et al., 2017; Selden et al., 2018). This suggests the importance of considering trophic interactions for improving predictions of biodiversity under climate change (Urban et al., 2016; Selden et al., 2018). As an example, according to our niche models, the geographic range of the European hake (M. merluccius), one of the main commercial species in the basin, could be reduced by 15% by the end of the century, but when considering trophic interactions, it is a reduction of almost 26% in its biomass and catches that is projected over this period. Our model results thus suggest that trophic interactions can amplify the direct effects of climate on species as already shown locally by Libralato et al. (2015). In addition, even if species distribution models have the potential to predict the westward and northward expansion of thermophilic species, the increase in biomass in the southeastern Mediterranean Sea could not be anticipated without taking into account trophic interactions in the projections.

Under the high emission scenario RCP8.5, with changes in biogeography and productivity of modeled marine organisms, the species composition of communities and the functioning and structure of Mediterranean marine ecosystems are expected to change significantly. There will likely be a reorganization of species assemblages and associated food webs by the end of the century, both on the continental shelf and in offshore area. Other projections focusing on the continental shelf of the Mediterranean sea showed the same patterns but it is the first time that projections are performed on the offshore area of the basin (Ben Rais Lasram et al., 2010; Albouy et al., 2012, 2013, 2014; Hattab et al., 2014, 2016). Our results suggest an increase in the biomass of low trophic levels species, a higher proportion of small-sized individuals, a decrease in toppredators' biomass as evidenced by the decrease in the HTI indicator and associated decline of the mean trophic level of the community. Several studies have already shown such trends in the Mediterranean Sea (e.g., Ben Rais Lasram et al., 2010; Albouy et al., 2012, 2014; Hattab et al., 2014, 2016; Libralato et al., 2015; Corrales et al., 2018) and at global scale (Cheung et al., 2010, 2011; Blanchard et al., 2012; Carozza et al., 2018; Lotze et al., 2018; Bryndum-Buchholz et al., 2019) but not reporting at the same level of resolution in both species responses and spatial scales. In addition, the way species interactions are handled in OSMOSE, i.e., opportunistic and mechanistically formulated (vs. correlative or fixed trophic interactions), makes it appropriate to explore the impacts of future environmental changes, and allows to explore shifts in trophic structure.

The response of the Mediterranean Sea to climate change could have significant consequences for ecosystem productivity and biodiversity and hence for the overall goods and ecosystem services they provide, especially the production of living marine resources. Although our business-as-usual fishing scenario is simplistic (management and conservation plans will most likely be applied before the end of the century and fishing strategies will change), our catch projections showed contrasted patterns during the 21st century. By the middle of the century, most Geographical Sub-Areas (GSAs) exhibited a slight decline in catches as the loss of catch of native species was not compensated by gains in catch of thermophilic and/or exotic species. By the end of the century, the western and eastern part of the Mediterranean showed opposite trends with an increase of catch in all the eastern basin due to an increase in catch of thermophilic/exotic species and a decrease in catch in all the western basin due to the decrease of the biomass of several main exploited native species and the non-replacement by warm-water species. Our results suggest a tropicalization of catch composition in eastern GSAs of the Mediterranean Sea as already shown by Tsikliras and Stergiou (2014). In a context where one-third of the Mediterranean human population is concentrated along the coasts and is projected to grow, the question of the availability of food resources is crucial, especially in the southern countries where food demand is projected to increase most. With the proliferation of nonindigenous invasive species there is a need to explore market options for non-target species currently of low or no economic value (Weatherdon et al., 2016). Moreover, as shown by Lam et al. (2016), due to the increasing dominance of low value marine resources in the total world catches, an increase in catch does not necessarily translate into increases in revenues for fishing communities. The economic consequences of climate change on fisheries might manifest through changes in the price and value of catches (Sumaila et al., 2011). However, climate-induced changes may also offer new opportunities to some Mediterranean fisheries, with increased landings of warmwater species, some of which of high commercial interest (e.g., C. hippurus).

The projected increase in plankton production could provide opportunities to rebuild some overfished stocks, but climate change questions the relevance of current stock assessment models and management strategies to reach sustainable exploitation of all living marine resources. Several studies have indeed shown the potential synergistic effects of climate change and fishing on exploited populations and ecosystem functioning in the Mediterranean Sea and other regions of the world (e.g., Scheffer et al., 2001; Hsieh et al., 2006; Ottersen et al., 2006;

Perry et al., 2010; Hidalgo et al., 2011; Quetglas et al., 2013; Tu et al., 2018). For instance, Hidalgo et al. (2011) showed that the erosion of the age structure of harvested hake populations in the Mediterranean Sea may drastically alter their capacity to dampen environmental fluctuations. Ignoring the effects of climate change in stock assessment could compromise the validity of stock forecasts and affect the robustness of several biological reference points such as the Maximum Sustainable Yield (MSY) (Brander, 2010; Grafton, 2010; Link et al., 2011; Galbraith et al., 2017; Serpetti et al., 2017). However, improved fisheries and ecosystems management in a highly overexploited Mediterranean Sea could have the potential to offset many negative effects of climate change (Roberts et al., 2017; Gaines et al., 2018).

## CONCLUSION

This study projects climate change impacts on the biomass and fisheries catch at the whole Mediterranean scale under the high emission scenario RCP8.5. It is the first attempt to project future marine biodiversity over the whole Mediterranean Sea at fine resolution, and by explicitly considering climate-induced changes in plankton production, shifts in species distributions and their trophic interactions. Despite various uncertainties associated with projections, our results suggest that the high emission scenario RCP8.5 could result in an increase in total fish and macroinvertebrate biomass by 5 and 22%, and in fisheries catch by 0.3 and 7% by 2021–2050 and 2071–2100, respectively, overall mirroring changes in primary and secondary production in the Mediterranean Sea. These global increases masked several "losers" among modeled species while "winners" were mainly small pelagic species, thermophilic and/or exotic species, of smaller size and of low trophic levels. Projected increase in biomass and catch were expected in the southeastern part of the basin whereas significant decreases are most likely in the western Mediterranean Sea. We also showed that changes in the biogeography of species, associated with changes in productivity, could result in changes of Mediterranean ecosystem structure and trophic functioning by the end of the century. Combined with fishing pressure, climate change has the potential to render marine ecosystems more vulnerable to invasions by non-indigenous species. Finally, our results emphasized the importance of considering trophic interactions to improve predictions of biodiversity changes. The strong spatial contrasts in the projections also call for improved spatial management of marine resources across GSAs and collaboratively among states at the whole Mediterranean scale in order to mitigate global change effects in the region and create new opportunities for fisheries.

## REFERENCES

Adloff, F., Somot, S., Sevault, F., Jordà, G., Aznar, R., Déqué, M., et al. (2015). Mediterranean Sea response to climate change in an ensemble of twenty first century scenarios. Clim. Dyn. 45, 2775–2802. doi: 10.1007/s00382-015-2507- 2503

## DATA AVAILABILITY

The raw data supporting the conclusions of this manuscript will be made available by the authors, without undue reservation, to any qualified researcher.

## AUTHOR CONTRIBUTIONS

FM developed the model, acquired the data, and analyzed and interpreted the data. LV, NB, and Y-JS contributed to the data analysis and interpretation. FG and SD helped in the development of species distribution models. CU, PM, and SS provided the data on primary and secondary productions (from the biogeochemical model) and on climate environment (from the CNRM-RCSM4), respectively. NB helped with the programming code of OSMOSE and use of the HPC cluster DATARMOR. FM led the drafting of the manuscript with the contributions and revisions from all the authors.

## FUNDING

FM was funded by a Ph.D. grant from the French Ministry of Higher Education, Research and Innovation. This work was partially funded by the USBIO project of the LabEx CeMEB, an ANR "Investissements d'avenir" program (ANR-10-LABX-04-01) and the SOMBEE project of the joint BiodivERsA and Belmont Forum call "BiodivScen 2018" (ANR-18-EBI4-0003-01).

## ACKNOWLEDGMENTS

The authors acknowledge the Pôle de Calcul et de Données Marines (PCDM) for providing DATARMOR computational resources (http://www.ifremer.fr/pcdm) and the CALMP computation center (Grant P1331) for the HPC resources, and the support of the SIROCCO team (http://sirocco.obs-mip. fr/). The authors also acknowledge the CNRM for making the physical data available. The climatic simulations used in this work were downloaded from the Med-CORDEX database (www.medcordex.eu).

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00345/full#supplementary-material


climate change impacts on coastal Mediterranean fish assemblages. Glob. Change Biol. 18, 2995–3003. doi: 10.1111/j.1365-2486.2012.02772.x


food webs: implications for fisheries and conservation. Glob. Change Biol. 16, 1194–1212. doi: 10.1111/j.1365-2486.2009.02046.x


in the Mediterranean. Nat. Clim. Change 8, 972–980. doi: 10.1038/s41558-018- 0299-292


and biological processes in the oceans. Front. Mar. Sci. 2:36. doi: 10.3389/fmars. 2015.00036



and management responses. J. Mar. Syst. 79, 427–435. doi: 10.1016/j.jmarsys. 2008.12.017



upwelling ecosystem: an end-to-end modelling approach reveals dampened effects. PLoS One 9:e94286. doi: 10.1371/journal.pone.0094286


**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 Moullec, Barrier, Drira, Guilhaumon, Marsaleix, Somot, Ulses, Velez and Shin. 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.

# Relative Impacts of Simultaneous Stressors on a Pelagic Marine Ecosystem

#### Phoebe A. Woodworth-Jefcoats1,2 \*, Julia L. Blanchard<sup>3</sup> and Jeffrey C. Drazen<sup>2</sup>

<sup>1</sup> Pacific Islands Fisheries Science Center, NOAA Fisheries, Honolulu, HI, United States, <sup>2</sup> Department of Oceanography, University of Hawai'i at Manoa, Honolulu, HI, United States, ¯ 3 Institute for Marine and Antarctic Studies, University of Tasmania, Hobart, TAS, Australia

Climate change and fishing are two of the greatest anthropogenic stressors on marine ecosystems. We investigate the effects of these stressors on Hawaii's deep-set longline fishery for bigeye tuna (Thunnus obesus) and the ecosystem which supports it using a size-based food web model that incorporates individual species and captures the metabolic effects of rising ocean temperatures. We find that when fishing and climate change are examined individually, fishing is the greater stressor. This suggests that proactive fisheries management could be a particularly effective tool for mitigating anthropogenic stressors either by balancing or outweighing climate effects. However, modeling these stressors jointly shows that even large management changes cannot completely offset climate effects. Our results suggest that a decline in Hawaii's longline fishery yield may be inevitable. The effect of climate change on the ecosystem depends primarily upon the intensity of fishing mortality. Management measures which take this into account can both minimize fishery decline and support at least some level of ecosystem resilience.

Keywords: climate change, fishing, pelagic, bigeye tuna, size-based model, food web model

## INTRODUCTION

Climate change and fishing are two of the greatest anthropogenic stressors on marine ecosystems and commercial fisheries. Additionally, these stressors are impacting marine systems simultaneously, potentially exacerbating one another. Given that current carbon emissions are outpacing the most emission-heavy scenario being used in climate models (RCP8.5; Riahi et al., 2011) and that a growing human population derives nearly one-sixth of its animal protein from the sea (Pentz et al., 2018), it is imperative that we understand the effects of these joint stressors now and in the future (Perry et al., 2010). Furthermore, we need to do so in an ecosystem context in order to understand the full ramifications of these stressors' effects (e.g., Pikitch et al., 2004; Brander, 2007). In this study, we examine the effects of climate change and fishing on Hawaii's longline fishery for bigeye tuna (Thunnus obesus) and its supporting ecosystem. This fishery operates largely outside the United States EEZ, extending from equatorial waters to the northern limits of the North Pacific subtropical gyre (35–40◦N) and from the dateline to the outer limits of the California Current region (roughly 125◦W), excluding the eastern tropical Pacific's oxygen minimum zone (**Figure 1**). Yet, a sizeable portion of the fishery operates in waters with little to no international competition (Woodworth-Jefcoats et al., 2018). This means that local

#### Edited by:

Isaac C. Kaplan, Northwest Fisheries Science Center (NOAA), United States

#### Reviewed by:

Guillem Chust, Centro Tecnológico Experto en Innovación Marina y Alimentaria (AZTI), Spain Autumn Oczkowski, United States Environmental Protection Agency, United States

#### \*Correspondence:

Phoebe A. Woodworth-Jefcoats phoebe.woodworth-jefcoats@ noaa.gov

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 28 March 2019 Accepted: 19 June 2019 Published: 04 July 2019

#### Citation:

Woodworth-Jefcoats PA, Blanchard JL and Drazen JC (2019) Relative Impacts of Simultaneous Stressors on a Pelagic Marine Ecosystem. Front. Mar. Sci. 6:383. doi: 10.3389/fmars.2019.00383

management measures have the potential to effect broad ecosystem change. Additionally, Honolulu ranks 6th among United States commercial fishing ports in terms of the value of fish landed (\$106 million; NOAA Fisheries, 2017) and over half the nation's tuna landings are from this fishery (NOAA Fisheries, 2018). These factors create a strong incentive to ensure the fishery's future ecological and economic viability.

Commercial fishing has reduced the abundance of large high-trophic level predators in this ecosystem by over 20% (Ward and Myers, 2005) and at the same time has led to increasing catch rates of smaller mesopredator species (Polovina et al., 2009). Modeling studies have replicated these historical observations using both species-based (Cox et al., 2002; Kitchell et al., 2002) and size-based (Polovina and Woodworth-Jefcoats, 2013) models. Similar modeling approaches have projected future effects of fishing and/or climate change over the 21st century. These approaches range from highly specific single species models (Lehodey et al., 2010, 2013; Del Raye and Weng, 2015) to multi-species ecosystem (Howell et al., 2013; Woodworth-Jefcoats et al., 2015) and dynamic bioclimate envelope (Cheung et al., 2010) models to size-based approaches without specieslevel resolution (Woodworth-Jefcoats et al., 2013, 2015; Lefort et al., 2015). Collectively, they suggest climate-driven declines in food availability may reduce fish body size (Lefort et al., 2015; Woodworth-Jefcoats et al., 2015) and biomass (Howell et al., 2013; Dueri et al., 2014; Lefort et al., 2015; Woodworth-Jefcoats et al., 2015), as well as future fishery yields (Howell et al., 2013; Woodworth-Jefcoats et al., 2015). The location of spawning and fishing grounds may also change with climate change (Cheung et al., 2010; Lehodey et al., 2010, 2013; Dueri et al., 2014; Erauskin-Extramiana et al., 2019). A number of these studies included the effects of increasing temperatures. Multi-species or species-blind approaches relied on statistical relationships (Cheung et al., 2010; Erauskin-Extramiana et al., 2019) or monotonically increasing costs of metabolism (Woodworth-Jefcoats et al., 2013), while species-specific models were able to incorporate more complex temperature effects. These include linking spawning to ocean temperature (Lehodey et al., 2010, 2013) and incorporating temperature into physiological rates (Dueri et al., 2014; Lefort et al., 2015).

Despite the array of approaches discussed above, there has not been, to our knowledge, a multi-species approach that includes both size and species resolution as well as the physiological effects of rising ocean temperatures. In this study, we use a food web model that integrates both size and species. This approach allows us to examine species-specific change in terms of biomass, abundance, and size structure. The model also incorporates temperature's effects on metabolism as well as aerobic scope, providing more realistic future projections. Aerobic performance is closely linked to temperature (e.g., Pörtner and Peck, 2010; Pörtner, 2012) and affects fishes' ability to forage. Our simulations include climate change's effects on two variables which most directly affect fishes' fitness: food supply, via changes to the plankton community, and temperature. We also examine a range of future fishing scenarios. Our results offer insight into the simultaneous effects of these stressors, and the modeling framework we developed offers a new tool for supporting strategic management decision-making in this and other regions.

## MATERIALS AND METHODS

## Model

We developed the size-based food web model therMizer, which is a modification of mizer, a well-documented multi-species size spectrum model (Blanchard et al., 2014; Scott et al., 2014). Such models describe predation, mortality, reproduction, and physiological processes at the individual level and scale them up to population and community levels (Blanchard et al., 2017). They track the flow of biomass through fully resolved body size classes (size measured in mass) via growth and size-based predation (Blanchard et al., 2017). In mizer, the smallest fish size classes feed upon a background resource size spectrum that exhibits semi-chemostat growth dynamics (Blanchard et al., 2014; Scott et al., 2014). Our model therMizer contains two key modifications from mizer. The primary modification was incorporating the effect of ocean temperature on metabolic scope. Temperature dependencies are absent in mizer. We also replaced mizer's semi-chemostat background resource with a resource that is input at each time step.

The effect of temperature on metabolic scope was determined by including temperature's effect on both metabolic rate and prey encounter rate. This was incorporated into the model by scaling both rates as described below and illustrated in **Figure 2**. In all cases, temperature was averaged over each species' depth range.

As temperature increases, metabolic rate increases. To capture this relationship, we modeled temperature's effect on metabolic rate, TEM, following Eq. (1):

$$T \text{EM} = e^{25.22 - \frac{\mu}{M}} \tag{1}$$

where T is vertically averaged temperature in Kelvin, k is Boltzmann's constant (8.62 × 10−<sup>5</sup> eV K−<sup>1</sup> ), and E is activation energy (0.63 eV; Brown et al., 2004; Jennings et al., 2008). TEM was then scaled to TEM<sup>0</sup> , a value ranging from 0 to 1, following Eq. (2):

$$TEM' = (TEM - Min\_{\rm sp}) / R\_{\rm sp} \tag{2}$$

where Minsp and Rsp are the minimum value and range, respectively, of TEM for each species (**Figure 2**). TEM<sup>0</sup> was then used as a multiplier for standard metabolic rate. This has the effect of standard metabolic rate being at its minimum at the lower limit of a species' thermal range and at its maximum at the upper limit of a species' thermal range.

In addition to influencing metabolic rate, temperature also influences aerobic scope and fishes' ability to successfully forage. To capture this relationship, we incorporated temperature into prey encounter rate. While species-specific thermal performance parameters are largely lacking in the literature, the relationship between temperature and aerobic scope is broadly understood (Pörtner and Peck, 2010). Therefore, we modeled the effect of temperature on encounter rate, TER, using a generic polynomial rate equation (van der Heide et al., 2006):

$$TER = T\left(T - T\_{\min}\right)\left(T\_{\max} - T\right) \tag{3}$$

where T is vertically averaged temperature, Tmin is a species' minimum thermal tolerance, and Tmax is a species' maximum thermal tolerance (**Figure 2**). All temperatures in Eq. (3) are in ◦C. TER was then scaled to TER<sup>0</sup> , a value ranging from 0 to 1, by dividing by Maxsp, the maximum value of TER for each species (**Figure 2**). TER<sup>0</sup> was then used as a multiplier for encounter rate. This has the effect of species being able to realize peak aerobic performance and encounter the maximum amount of prey possible when they are at their optimal temperature. Foraging success declines to either side of this temperature.

The joint effects of temperature on metabolic rate and prey encounter rate (TEM<sup>0</sup> and TER<sup>0</sup> , respectively) are shown in **Figure 3**. At temperatures outside species' thermal range, both TEM<sup>0</sup> and TER<sup>0</sup> were set to 0 representing local extinction. Species' thermal and vertical ranges are listed in **Table 1**.

## Model Parameters and Input

We attempted to include as many species as possible of the top 20 species caught by the Hawaii deep-set longline fishery, regardless of species' commercial value. The 12 species listed in **Table 1** are those for which there was sufficient life history and thermal tolerance information available to parameterize the model. Together, these species account for 76% of the fishery's observed catch.

## Parameters and Calibration

Global model parameters were left unchanged from the default mizer settings (Blanchard et al., 2014; Scott et al., 2014), with the exception of kappa (κ) which we set at 1012. This variable

is used in determining species' initial size spectra (Blanchard et al., 2014). Also as in Blanchard et al. (2014), all teleosts enter the model as larvae weighing 1 mg. Blue sharks enter at 354 g, an average of mean male and female birth weights (344 and 362 g, respectively; Shark Working Group Report, 2017). The additional species-specific parameters are listed in **Table 1**. Values in **Table 1** were taken from the literature as noted, with the exception of the Brody growth coefficient, kvb, for lancetfish. Estimates of this parameter for lancetfish were not available in the literature. Based on available values for similar species (Morales-Nin and Sena-Carvalho, 1996; Lorenzo and Pajuelo, 1999; Harada and Ozawa, 2003; Figueiredo et al., 2015; Froese and Pauly, 2017), we used the median value of the lower quartile of teleost kvb values.

Predation in therMizer is both species- and size- specific. All fish have a log-normal prey size preference that is dependent upon predator body size, species' predator-prey mass ratio (100 for teleosts, Blanchard et al., 2014; 400 for blue sharks, Barnes et al., 2008), and the width of the prey selection window (1 for all species, Blanchard et al., 2014). Prey selection is further informed by the interaction matrix (**Supplementary Table 1**). Interaction, θij, between species i and j ranges from

#### TABLE 1 | Species-specific model parameters.


Weights (w) are in grams. Unless otherwise indicated, weight-at-maturity (wmat), and maximum weight (wmax) are calculated using the length-weight conversions detailed in Supplementary Table 3. kvb is the Brody growth coefficient. Maximum recruitment (Rmax) is scaled from maximum size as 10<sup>11</sup> × wmax <sup>−</sup>1.5 following Blanchard et al. (2014). wF0 and wF1 are the sizes at which species are initially and fully susceptible to fishing mortality. Species are listed in rank order of their numeric abundance in catch of Hawaii's deep-set longline fishery for bigeye tuna (1995–2016, pooled). <sup>∗</sup>Average of male and female size, calculated using the values found in Supplementary Table 3. ∗∗Average of male and female values. <sup>a</sup>Estimated from Portner et al., 2017. <sup>b</sup>Uchiyama and Kazama, 2003. <sup>c</sup>Nicol et al., 2011. <sup>d</sup>Boyce et al., 2008. <sup>e</sup>Boettiger et al., 2012 and Froese and Pauly, 2017. <sup>f</sup>Howell et al., 2010. <sup>g</sup>Uchiyama and Boggs, 2004. <sup>h</sup>Uchiyama et al., 1986. <sup>i</sup>Shark Working Group Report, 2017. <sup>j</sup>Stevens et al., 2010. <sup>k</sup>Maunder, 2001. <sup>l</sup>Bayliff, 1988. <sup>m</sup>Schaefer and Fuller, 2007. <sup>n</sup>Wild, 1986. <sup>o</sup>Billfish Working Group Report, 2014a. <sup>p</sup>Hawn and Collette, 2012. <sup>q</sup>Francis et al., 2004. <sup>r</sup>Polovina et al., 2008. <sup>s</sup>Zischke et al., 2013. <sup>t</sup>Sepulveda et al., 2011. <sup>u</sup>Billfish Working Group Report, 2015. <sup>v</sup>DeMartini et al., 2007. <sup>w</sup>Abecassis et al., 2012. <sup>x</sup>Shimose et al., 2015.

0 to 1. Previous size spectrum models have determined the interaction matrix values based on horizontal overlap as inferred from bottom trawl surveys (Blanchard et al., 2014; Reum et al., 2019). Here, we determined interaction based on species' vertical overlap following Eqs (4) and (5) and illustrated in **Figure 4**:

$$
\Theta\_{i\bar{j}} = D\_{i\bar{j}} ^2 / D\_i D\_{\bar{j}} \tag{4}
$$

$$D\_{ij} = a - (a - b) - c \tag{5}$$

where D<sup>i</sup> and D<sup>j</sup> are the depth ranges of species i and j, respectively; Dij is the range of overlapping depths for species i and j; and a is the greater maximum depth, b is the lesser maximum depth, and c is the greater minimum depth for the pair of species i and j. All species have a minimum depth of 0 m, with the exception of opah which has a minimum depth of 50 m (Polovina et al., 2008). For all species pairs, the interaction matrix determines the proportion of total prey biomass of the appropriate size that is available to the predator.

Fishing mortality increases linearly from 0 to F over a size range unique to each species. Fishing mortality is phased in over a range of body sizes to account for longline gear's inefficiency in catching smaller body sizes (Polovina and Woodworth-Jefcoats, 2013). To establish these sizes, we used time-averaged (2006– 2016, pooled) catch records from the Pacific Islands Region Observer Program, which since 2006 has recorded the size of every third fish caught by Hawaii's longline bigeye tuna fleet. Roughly 20% of this fishery's effort is observed, and observer records have been found to correlate well with vessel logbooks (Woodworth-Jefcoats et al., 2018). We binned observed sizes of fish caught into equally spaced logarithmic size classes as in therMizer (Scott et al., 2014; Edwards et al., 2017). Each species is initially susceptible to fishing mortality at the size which contributes at least 1% toward that species' total observed catch. Fish are fully susceptible to fishing mortality at the size which contributed the most to that species' total catch. The sizes at which each species is first and then fully susceptible to fishing mortality are listed in **Table 1**.

## Climate Forcing Variables

We used output from a suite of earth system models included in the 5th phase of the Coupled Model Intercomparison Project (CMIP5; Taylor et al., 2012; **Supplementary Table 2**). CMIP is a coordinated international climate and earth system modeling approach that centers around common model forcings and output variables (Taylor et al., 2012). Phyto- and zooplankton densities (**Figure 5**) were vertically integrated over the upper 200 m of the water column. Numerical abundance within each size class was determined by dividing density by mean

are calculated. θij, interaction between species i and j. D<sup>i</sup> , D<sup>j</sup> , depth ranges of species i and j, as determined from species' minimum and maximum depths (e.g., imin and imax). Dij, range of overlapping depths for species i and j.

size. Plankton size spectra were created as linear fits to logtransformed abundances and sizes. Model-specific plankton size classes are listed in **Supplementary Table 2**.

As with the original mizer model, some calibration of the background resource was required (Blanchard et al., 2014). To this end, we compared the above described plankton spectra with the background spectrum generated by the semi-chemostat resource model to determine appropriate scaling for the slope (×1.2) and intercept (×0.8) of the CMIP5-generated plankton spectra. These scaled spectra were extended to therMizer's full size range to determine the background resource at each time step. Initial spectra for individual fish species were determined as in the original mizer model (Scott et al., 2014).

Ocean temperature for each species was determined by averaging across each species' depth range. Initial temperatures are from World Ocean Atlas 2013 v2 data (Locarnini et al., 2013). Temperature changes from the CMIP5 models were then applied to these initial temperatures. This approach accounts for potential bias in the CMIP5 models.

Model input plankton densities are summed and temperatures averaged over the footprint of the Hawaii-based deep-set longline fishery targeting bigeye tuna: 0◦–40◦N from 180◦–150◦W and 15◦–36◦N from 150◦–125◦W (**Figure 1**; Woodworth-Jefcoats et al., 2018). Across the CMIP5 models used, phytoplankton densities declined by an average of 6% by mid-century and 12% by 2100. Declines in zooplankton density were twice that of phytoplankton, i.e., 12% by mid-century and 24% by 2100 (**Figure 5**). All species, regardless of vertical range, were projected to encounter rising ocean temperatures (**Figure 3**). For the deepest-living species modeled which have a maximum depth of over 1000 m, temperatures increased by about 0.5◦C by midcentury and 1◦C by 2100. For the shallowest-living species which live within 100 m of the ocean's surface, temperatures increased by nearly three-times this amount or roughly 1.5◦C by 2050 and 3 ◦C by 2100.

## Model Verification

Model output from a run forced with a static climate (1986– 2005 mean) and constant fishing mortality (F = 0.2) was compared to time-averaged records of observed catch (see description of the observer data above). Observed sizes were binned as in therMizer to create size spectra of catch. Modeled and observed catch size spectra were well correlated, with Pearson's correlation coefficient, r, ranging from 0.39 to 0.85 (**Supplementary Figure 1**). We used a value of F = 0.2 for model verification because, for the species for which there are stock assessments, most of these assessments estimate fishing mortality to be close to this value (e.g., Billfish Working Group Report, 2014a,b, 2016; McKenchie et al., 2017; Shark Working Group Report, 2017; Xu et al., 2018).

## Scenarios Modeled

We evaluated the individual and joint effects of climate change and fishing on the ecosystem and on fishery catch. In all scenarios, the model was run for 600 years with a static climate (1986– 2005 mean) and constant fishing mortality (F = 0.2) to account for spin-up effects and allow the model to reach equilibrium. Projections run from 2006 through 2100. To assess the impact of climate change alone, we held fishing mortality constant at F = 0.2. To assess the impact of fishing alone, we used a static climate scenario. In all cases where a variable was held static, we held the spin-up value constant over the 21st century.

We examined four scenarios in which fishing mortality changed linearly over the projection period (2006–2100): doubling from F = 0.2 to 0.4, increasing five-fold to 1, halving to 0.1, and declining to one fifth or 0.04 (hereafter referred to as 2F, 5F, 0.5F, and 0.2F, respectively). These scenarios were chosen based in part on trends in effort of Hawaii's deep-set longline fishery. Over the logbook record, effort has risen more than five-fold from 8.4 million hooks set in 1995 to over 47 million hooks set in 2015 (Woodworth-Jefcoats et al., 2018). Fishing effort does not translate equally to fishing mortality,

and therefore we consider 5F to be a fairly aggressive future fishing scenario. We also considered the effect of fishing mortality doubling (2F) as a more moderate scenario. We simply used the reciprocals of the fishing increase scenarios to model a decline in fishing mortality. This facilitated scenario comparison. To further facilitate scenario comparison, we used the same value of F for all species. This approach eliminated potential confounding influences of fishing different species at different levels of intensity and replicated observed catch reasonably well (see section "Model Verification" above). However, we note that therMizer is capable of incorporating species-specific F values (Scott et al., 2014).

We evaluated several measures of ecosystem structure and fishery performance. Total biomass and abundance provide species-specific measures of the fishery's catch and its relation to the ecosystem. We refer to ecosystem biomass as "biomass" and catch in weight as "yield." The large fish indicator (LFI; Blanchard et al., 2014) is a broad measure of the numerical proportion of fish ≥15 kg (Polovina and Woodworth-Jefcoats, 2013; Woodworth-Jefcoats et al., 2015). The LFI provides insight into both the size structure of the ecosystem as well as the potential value of fish catch, as larger fish are generally more valuable. As a complementary measure to the LFI, we also examined the change in species' mean size.

We assessed these measures both through time series over the projection period as well as with 20-year averages in an effort to minimize the confounding influence of interannual variability. We averaged results over three 20-year time periods to capture the beginning, middle, and end of the 21st century: 1986–2005, 2041–2060, and 2081–2100 (hereafter referred to as 2000, 2050, and 2100, respectively). The 1986–2005 average corresponds to the equilibrium value at the start of the therMizer projections.

## RESULTS

We find that, taken as individual stressors, climate change and increasing fishing mortality act to reduce fish biomass and size across all species. The effects of reduced fishing mortality are generally of the opposite sign. However, when modeled jointly, there were no scenarios in which yield increased. Results for the ecosystem supporting the fishery are slightly more optimistic, with reduced fishing mortality somewhat offsetting the negative effects of climate change.

## Total Biomass and Yield

Climate change, with constant F, acts to reduce bigeye biomass by 7% by 2050 and by 20% by 2100. Across all species modeled,

these declines range from 3% (skipjack) to 14% (blue shark) by 2050 and from 7% (skipjack) to 37% (wahoo) percent by 2100. Declines in yield reflect declines in ecosystem biomass (**Figure 6**).

For all species, in the absence of climate change, decreasing F leads to increasing biomass, and vice versa. This is because lower levels of F result in less biomass being removed as yield. Both scenarios with increasing F lead to declining yield for all species, due to declining biomass. Likewise, the 0.2F scenario also leads to declining yield, due to less fishing effort, for all but the largest species (swordfish, blue shark, and blue marlin). The yield of these three largest species increases an average of 7% by 2050 and 8% by 2100 (**Figure 6**). The 0.5F scenario leads to similarly little change in yield by 2050 (<10% change). By 2100, roughly half the species modeled see an increase in yield of 25% or less, while two see no change, and three see small (<10%) declines (**Figure 6**).

We find that when changes in F are paired with climate change, reducing F can compensate somewhat the climate-driven biomass declines for all species. Bigeye biomass increases to within 10–12% of what it would be in the absence of climate change by 2050 under the 0.5F + climate change and 0.2F + climate change scenarios. Across all species, this value ranges from 4 to 23% (**Figure 6**). By 2100, biomass of all species except wahoo more than doubles (bigeye biomass increases 136%) when climate change is incorporated into the 0.2F scenario. When climate change is included in the 5F scenario, yield increases over the initial ∼15 years and then declines. Other than this short-term increase, there is a decline in yield for all species under all fishing scenarios; none of the modeled fishing scenarios were able to compensate for the climate-driven declines in yield. Furthermore, climate change amplified the biomass declines seen under scenarios with increasing fishing mortality.

## Total Abundance

Climate change, in the absence of changing F, increases the abundance of a number of species (**Figure 7**). By 2050, all species except blue shark experience an increase in abundance of 1–9%. By 2100, all species except blue shark, yellowfin, wahoo, striped marlin, and swordfish experience increases in abundance of 1–17%. Blue shark abundance declines by 10 and 21% across these time points. Yellowfin, wahoo, and striped marlin abundance decline by 9, 9, and 10%, respectively. Swordfish abundance is unchanged by 2100, despite increasing earlier in the century (**Figure 7**).

The effects of changing F on abundance are essentially the same as those on biomass: declining fishing mortality leads to increased fish abundance and vice versa. The effects on the number of fish caught, however, are different than those of biomass (i.e., decreasing fishing mortality leads to a decline in the number of fish caught, **Figure 7**).

The effects on abundance of pairing climate change and changes in F varied by species. For species that saw abundance increase under climate change, the climate effect somewhat dampened the abundance declines resulting from increasing F and amplified increases in abundance under decreasing F. For species that saw abundance decline under climate change, these declines were exacerbated by increasing F. When F was reduced, climate change dampened the expected increases in abundance (**Figure 7**).

## Large Fish Indicator

The effect of climate change on the large fish indicator (LFI) was small in the absence of changing F. LFI declines from 0.129 to 0.119 by 2050 and to 0.105 by 2100. Catch LFI declines as well, falling from 0.218 to 0.201 by 2050 and to 0.173 by 2100 (**Figure 8**).

The effects of changing F on LFI were greater than those from climate change. Reducing F led to LFI increasing from 0.129 in 2000 to 0.143–0.162 by 2050 and to 0.153–0.191 by 2100, across both the 0.5F and 0.2F scenarios. Increasing F had a greater effect on LFI, reducing it to 0.069–0.107 by 2050 and to 0.046–0.091 by 2100, across both the 2F and 5F scenarios. The effects on catch LFI were similar (**Figure 8**).

We found that when paired with climate change, halving F almost equally offset the decreased LFI caused by climate change alone (**Figure 8**). Climate change acted to undermine the increase in LFI caused by decreasing F to one fifth the initial value. Climate change also exacerbated the decline in LFI caused by increasing F. When looking at modeled catch, we found that neither modeled decrease in F was able to offset the decline in LFI after 2050. By 2100, catch LFI declined to 0.208 under the 0.2F + climate change scenario and to 0.109 under the 5F + climate change scenario.

## Mean Size

Mean size declined for all species under climate change alone. By 2050, declines in mean size range from 4% (blue shark) to 13% (yellowfin, wahoo, striped marlin, and swordfish) across species, with bigeye mean size declining by 11%. By 2100, mean size declines by 8–38% across species, with blue shark experiencing the least and wahoo experiencing the greatest decline in mean size (bigeye declines by 23%). Declines in the mean size of fish caught are slightly smaller (**Figure 9**).

Because fishing targets a species' largest body sizes, the effects on mean size of changing F are fairly straightforward: In the absence of climate change increasing F leads to mean body sizes decreasing by 11–62% by 2050 across both the 2F and 5F scenarios, with bigeye size decreasing by 19–48% across these scenarios. By 2100, increasing F leads to mean body size decreasing by 19–77% (bigeye by 32–64%). Decreasing F has the opposite effect on mean size. By 2050, the increase is somewhat less than opposite that of the reciprocal fishing scenario. However, by 2100, reciprocal fishing scenarios result in nearly opposite effects on mean size. As with other indicators, these effects are somewhat dampened in the catch relative to the ecosystem due to the size-selective nature of fishing (**Figure 9**).

The joint effect of fishing and climate change on species' mean size varied by species. Reduced F was able to offset the climateinduced decline in mean size, to some degree, for all species. By 2100, the 0.2F + climate change scenario led to increases in mean size for all species except wahoo. The 0.5F + climate change scenario allowed mean size to increase for roughly half the species modeled. These results were dampened in the modeled catch. By 2050, the mean size of fish caught changed by −8–+11% across species under the 0.5F + climate change and 0.2F + climate change scenarios. The change in size of bigeye caught in 2050 ranged from −2–+2% across these two scenarios. By 2100, the 0.5F + climate change scenario allowed mean size of fish caught to increase in four species (lancetfish, blue shark, swordfish, and blue marlin). The 0.2F + climate change scenario allowed mean size caught to increase in all but four species (mahi, yellowfin, wahoo, and striped marlin; **Figure 9**).

## DISCUSSION

We used therMizer, a size-structured food web model with individual species that is capable of capturing the metabolic effects of rising ocean temperatures, to assess the effects of climate

change and fishing on Hawaii's deep-set longline fishery and its supporting ecosystem. Our results show that while a decline in this fishery's yield seems likely, this may mask resilience in the ecosystem supporting the fishery. The contrast between changes in catch and changes to the ecosystem is particularly noteworthy as it highlights the limited ability of some fishery dependent data to fully capture ecosystem trends.

## Outlook for Future Yield and Ecosystem

Our results show that as the climate continues to change, a decline in the yield of Hawaii's bigeye tuna fishery seems inevitable. None of the changes in fishing mortality that we modeled, whether increasing or decreasing, allowed yield to increase after more than about 15 years. These results reinforce those of Howell et al. (2013), who found that climate change is projected to reduce the Hawaii longline fishery's target yield even when fishing mortality is halved. Their study used an Ecopath with Ecosim model to simulate food web and fishery response to climate change. That two dissimilar modeling methods produce similar projections for declining yield should be noted by regional fishery managers. Additional modeling (e.g., Cheung et al., 2016; Fu et al., 2018; Queirós et al., 2018) and empirical (Watson et al., 2012) studies of other ecosystems have led to similar projections.

In addition to total yield declining, we also find that the proportion of large fish in the catch declines in all scenarios after 2050. This suggests that not only will yield be reduced, but all else being equal, the fish caught may be less valuable because there will be fewer large fish. That said, increasing fishing mortality does lead to increased numbers of fish caught for some species (**Figure 7**). This is likely because therMizer models fishing

mortality as a removal of a numeric percentage rather than a biomass percentage (i.e., as F increases, a greater number of individuals is removed, though yield may still decline if those individuals' mean size declines).

Despite the poor outlook for fishery yield, we find that the ecosystem may be more resilient under specific future scenarios. Biomass of all species increases when climate change is modeled jointly with a reduction in fishing effort (**Figure 6**). This result reinforces calls from previous authors that reduced fishing can help reduce the effects of climate change (e.g., Brander, 2013). We also find that halving fishing mortality allows the LFI to remain essentially unchanged over the 21st century, and that reducing fishing mortality to one fifth initial values allows the LFI to increase. Ultimately, the decision of whether to lower fishing mortality in favor of ecosystem resilience comes down to societal values. Models such as therMizer can help fishery managers and other stakeholders understand a broad range of fishery management consequences (Blanchard et al., 2014; Cheung et al., 2016).

## Mechanisms Driving Change

One value in modeling studies is that they allow for investigation of the mechanisms driving change. This is particularly valuable when different stressors have the same effect; without being able to examine the underlying mechanisms it can be easy to assume that they are the same. We find that both climate change and increasing fishing mortality have similar effects on the central North Pacific's ecosystem and fishery yield: reduced biomass and a decline in mean body size. However, the mechanisms driving this response are different. The declining plankton biomass projected as a result of climate change reduces the amount of energy (food) available to all predators. This leads to reduced growth and, in turn, lower biomass. The shift in the plankton community's size structure also propagates through the food web, with proportionally less food available to larger body sizes, further reducing growth at larger body sizes. This disproportionate allocation of limited resources shifts the size structure toward smaller body sizes, resulting in a decline in mean body size across species (see also the discussion of speciesspecific effects below). Further, the disproportionate allocation of resources favoring smaller body sizes, paired with the inverse relationship between abundance and body size, explains why climate change leads to increased numerical abundance for some species.

Fishing, on the other hand, selectively removes the largest individuals from the population. Because a single large individual can be orders of magnitude larger than smaller individuals, removal of numerous large fish reduces both total biomass and mean size. Conversely, allowing more large individuals to remain in the ecosystem by reducing fishing effort more than counteracts the effect of removing them (**Figures 6**, **9**).

Modeling climate change and fishing jointly highlights the different mechanism at work to drive ecosystem change. Regardless of how fishing mortality changes, climate change acts to lower the system's carrying capacity, thereby reducing potential biomass, abundance, and yield. This interaction of stressors is only apparent when they're modeled together. Such interaction may explain the diminished impact of climate change as fishing mortality increases. As fishing increases, its effects may overshadow the lower carrying capacity resulting from climate change (Blanchard et al., 2012). This result is somewhat surprising given that a number of studies have found that the effects of climate change are stronger on more heavily fished systems (e.g., Blanchard et al., 2012; Brander, 2013). One possible explanation for this disparity may be tied to model structure (Woodworth-Jefcoats et al., 2015). Application of mizer to another ecosystem produced results similar to ours. Fu et al. (2018) found that higher trophic level fish were more likely than those at lower trophic levels to see dampened effects when fishing and climate change were combined. The species considered in our study are nearly all high trophic level species. We encourage further ecosystem modeling comparisons across modeling frameworks and ecosystems to help separate model structure from ecosystem structure (e.g., Tittensor et al., 2018). We also encourage further studies to consider the joint effects of stressors, especially in the open ocean beyond the limits of EEZs and LMEs given the relative paucity of studies doing so (Ortuño Crespo and Dunn, 2017).

Another mechanism that we investigate in this study is the role that temperature plays in driving species' response to climate change. We find that shallower-living species, most notably wahoo, see the greatest effect from climate change. On the other hand, species projected to see the least warming (e.g., lancetfish, swordfish, and blue shark) experience an increase in mean body size under both scenarios where decreasing fishing mortality is paired with climate change. Rising temperatures exacerbate the effect of reduced food availability by both increasing metabolic demand and reducing aerobic scope. This means that as climate change progresses fish will need more food despite there being less available, and that they'll be less able to successfully forage for this food. The large effect that rising temperature has on wahoo and, to some degree, on mahi mahi, suggests that shallowerliving species may be bellwethers of larger ecosystem changes. It also creates the potential for a shifting species composition of both the ecosystem and catch as species are differentially affected by rising ocean temperatures. Conducting additional therMizer simulations with more spatially discrete temperature projections, both vertically and horizontally, or with temperature exposure varying across life stages could provide further insight into how species may be affected by the ocean's warming.

Our method for incorporating temperature's effect on metabolic demand and aerobic scope requires only minimal parameterization (universal constants and species' thermal tolerance limits). This potentially increases the utility of the approach across other modeling frameworks. Similarly, it could provide an independent first approximation of how individual marine species may be affected by climate change. Others have highlighted the need to better incorporate aerobic scope into projections of climate effects (e.g., Pörtner, 2012). If a similarly simple approach could be applied to the relationship between oxygen or carbon dioxide and aerobic scope, this would significantly enhance our abilities to meet this challenge.

Food web interactions are also an integral mechanism in the response to fishing and climate change. We find that the impact

of warming is somewhat offset by the effect of body size on predation. For example, blue marlin, the largest species in our simulations, experiences an increase in mean size when climate change is paired with decreasing fishing mortality, despite being a fairly shallow-living species (**Table 1**). This is likely due to the lack of competition between blue marlin and other species for prey, as its maximum body size exceeds those of other species (Kitchell et al., 2002). Conversely, yellowfin tuna, which has a maximum body size nearly one-fifth that of blue marlin sees its mean size decrease or remain constant under these scenarios despite having a deeper vertical range. This might be a result of yellowfin tuna being both predator and prey simultaneously (Cox et al., 2002; Kitchell et al., 2006). We note also that food web interactions would perhaps be more important in scenarios where different species are subject to different levels of fishing mortality, as they are in real systems. In this case, food web interactions could act to amplify or dampen fishing effects or the effects of climate change.

## Sources of Uncertainty

Three primary sources of uncertainty emerged in this study. The first is linked to the range of the CMIP5 models' plankton densities. While there is broad agreement across CMIP5 models regarding change in temperature, these models vary substantially in their values for plankton densities (**Figure 5**; Woodworth-Jefcoats et al., 2017). We've presented the multi-model mean across CMIP5 models in this study for clarity. However, the range of plankton values and change in plankton values leads to quite a wide range in therMizer output forced by different CMIP5 models. To some degree, this is expected as CMIP5 was the first CMIP to include zooplankton among the output variables. Skill will likely improve in future generations of earth system models and CMIP6 has intercomparisons planned toward this goal (Eyring et al., 2016). We note, though, that a reliable baseline to which modeled changes could be applied (which is how temperature is treated in this study) would be valuable to future earth system and ecosystem modeling efforts. It could also help reconcile differences in the magnitude of observed and modeled size spectra (**Supplementary Figure 1**). Such an empirical baseline exists for physical oceanographic variables in the World Ocean Atlas. While there are global plankton databases (e.g., COPEPOD; O'Brien, 2010), their coverage is fairly limited.

The second major source of uncertainty is the speciesspecific model parameters. For example, the effect of rising temperature depends in part on where thermal habitat places species' metabolic scope (Rountrey et al., 2014). For species with narrow thermal ranges (e.g., wahoo), a small change in temperature can have a large impact on metabolic scope. We note that our modeled metabolic scope is dependent on the accuracy of species' thermal tolerance limits. For well-studied fish such as tuna, these tolerance limits are likely accurate. However, for other species, particularly those of no commercial value, these tolerance limits are inferred from data such as diet or vertical range. Better understanding of how species use their full three-dimensional habitat would reduce model uncertainty.

Uncertainty around other species-specific parameters such as maximum recruitment, growth rate, and size-at-maturity also likely influences the model's results. A mizer sensitivity analysis found uncertainty around life history parameters to be the second greatest source of model uncertainty (Zhang et al., 2015). Furthermore, we know very little about how these parameters may change as climate changes. Improved understanding of species' life history and its relationship with environmental influences would not only reduce model uncertainty, but also improve fisheries management more broadly by enabling it to incorporate the effects of climate change (Brander, 2007; Koenigstein et al., 2016; Pentz et al., 2018). Such information would also better inform ecosystem-based approaches to fisheries management by allowing for more accurate parameterization, especially for non-target and bycatch species.

The third source of uncertainty is that linked to fundamental assumptions about the nature of the central North Pacific's pelagic ecosystem. The most critical assumption is that this is a food-limited system. If this weren't the case, then declines in biomass at the base of the food web wouldn't necessarily result in reduced biomass across the food web. A number of factors may be contributing to this assumed food limitation. Competition and prey switching can result in bottom-up forcing and aren't well captured in therMizer. It's also possible that there's a benefit to be had for fish being less than fully satiated. Perhaps they're better able to evade predators (MacLeod et al., 2007). Or perhaps feeding to a level below that of satiation optimizes the risks and benefits of foraging (Heithaus et al., 2008) or the balance of energy gained from food ingested with that needed to forage further (Enberg et al., 2012). While delving further into this question is beyond the scope of the present study, it is important to highlight that this assumption underpins this and likely many other projections about the ecosystem impacts of climate change. Additionally, uncertainty around feeding levels was found to be the greatest source of uncertainty in a set of mizer simulations (Zhang et al., 2015). Ecosystem models such as mizer and therMizer are one tool that can be used to evaluate the validity of this assumption and others. Future work on this topic is encouraged.

## Model Limitations and Future Directions

Our results raise several interesting questions that therMizer's limitations make challenging to address in this study. For example, food supply (via plankton) and temperature are only two variables shaping pelagic habitat. Oxygen concentration is important and isn't included in this variation of mizer. Beyond shaping pelagic habitat, oxygen concentration also influences aerobic scope, as do both carbon dioxide concentration and pH (Pörtner, 2012). Including any of these variables may provide a clearer picture of how different species will respond to climate change. Additionally, marine species can move in response to environmental change (Pinsky et al., 2013; Montero-Serra et al., 2015), and climate change has the potential to redistribute marine species (Cheung et al., 2010; Lehodey et al., 2010, 2013; Jones and Cheung, 2014; Woodworth-Jefcoats et al., 2017; Erauskin-Extramiana et al., 2019). Incorporating two or three spatial dimensions into therMizer would allow us to address questions related to fish movement. For example, can fish simply exploit deeper depths to escape rising temperatures, or will decreasing

light levels at depth diminish their foraging success? How might spatial changes in species' pelagic habitat affect their catchability? Finally, our representation of the fishery is quite simplistic in that it does not include fisher behavior. We recognize that this is a critical aspect of modeling fishery response to climate change (Haynie and Pfeiffer, 2012), and look forward to exploring this dimension in future work.

This study models the effects of declining food availability and rising ocean temperatures on species caught by Hawaii's deepset longline fishery for bigeye tuna. We show how these climate effects interact with a range of changes in fishing mortality. While increasing the yield of Hawaii's longline fishery may not be possible, projections for potential ecosystem resilience are encouraging.

## DATA AVAILABILITY

The data generated and analyzed for this study, as well as the therMizer model code, are available on GitHub at: https://github. com/pwoodworth-jefcoats/Size-Based-Modeling.

## AUTHOR CONTRIBUTIONS

JB and PW-J created the therMizer variation of the mizer model. PW-J conceived of and designed the study, performed data analysis, and wrote the first draft of the manuscript. All authors contributed to the interpretation of results and manuscript revision, and read and approved the submitted manuscript.

## REFERENCES


## FUNDING

PW-J was funded in part by the Fernando Gabriel Leonida Memorial Scholarship and the Denise B. Evans Fellowship for Oceanographic Research at the University of Hawaii.

## ACKNOWLEDGMENTS

We thank J. Raynor for asking the question that shaped this manuscript, J. Polovina for suggesting fruitful avenues for data analysis, and A. Andrews for guidance in compiling species parameters. J. O'Malley helped explain how uncertainty around broad ecological questions affects the estimation of life history parameters. M. Peck provided valuable insight into the role temperature plays in metabolic scope. This manuscript benefitted from thoughtful reviews by J. Reum, J. Polovina, J. O'Malley, and M. Donahue. We acknowledge the World Climate Research Programme's Working Group on Coupled Modelling, which is responsible for the CMIP. We also thank the climate modeling groups listed in **Supplementary Table 2** for producing and making available their model output. This is SOEST contribution 10727.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00383/full#supplementary-material


and effects on tuna dynamics. Can. J. Fish. Aquat. Sci. 59, 1736–1747. doi: 10.1139/f02-138



of spawning and possible migration in the central North Pacific. Mar. Fish. Rev. 68, 19–29.


**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 Woodworth-Jefcoats, Blanchard and Drazen. 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.

# Arctic Sensitivity? Suitable Habitat for Benthic Taxa Is Surprisingly Robust to Climate Change

Paul E. Renaud1,2 \*, Phil Wallhead<sup>3</sup> , Jonne Kotta<sup>4</sup> , Maria Włodarska-Kowalczuk<sup>5</sup> , Richard G. J. Bellerby3,6, Merli Rätsep<sup>4</sup> , Dag Slagstad<sup>7</sup> and Piotr Kuklinski ´ 5

<sup>1</sup> Akvaplan-niva, Tromsø, Norway, <sup>2</sup> Department of Arctic Biology, The University Centre in Svalbard, Longyearbyen, Norway, <sup>3</sup> Norwegian Institute for Water Research, Bergen, Norway, <sup>4</sup> Estonian Marine Institute, University of Tartu, Tartu, Estonia, 5 Institute of Oceanology, Polish Academy of Sciences, Sopot, Poland, <sup>6</sup> SKLEC-NIVA Centre for Marine and Coastal Research, State Key Laboratory of Estuarine and Coastal Research, East China Normal University, Shanghai, China, <sup>7</sup> SINTEF, Trondheim, Norway

#### Edited by:

Elizabeth Fulton, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia

#### Reviewed by:

Yong Jiang, Ocean University of China, China Kristina Øie Kvile, University of Oslo, Norway

> \*Correspondence: Paul E. Renaud pr@akvaplan.niva.no

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 30 March 2019 Accepted: 15 August 2019 Published: 03 September 2019

#### Citation:

Renaud PE, Wallhead P, Kotta J, Włodarska-Kowalczuk M, Bellerby RGJ, Rätsep M, Slagstad D and Kuklinski P (2019) Arctic ´ Sensitivity? Suitable Habitat for Benthic Taxa Is Surprisingly Robust to Climate Change. Front. Mar. Sci. 6:538. doi: 10.3389/fmars.2019.00538 Arctic marine ecosystems are often assumed to be highly vulnerable to ongoing climate change, and are expected to undergo significant shifts in structure and function. Community shifts in benthic fauna are likely to result from changes in key physicochemical drivers, such as ocean warming, but there is little ecological data on most Arctic species to support any specific predictions as to how vulnerable they are, or how future communities may be structured. We used a species distribution modeling approach (MaxEnt) to project changes over the 21st century in suitable habitat area for different species of benthic fauna by combining presence observations from the OBIS database with environmental data from a coupled climate-ocean model (SINMOD). Projected mean % habitat losses over taxonomic groups were small (0–11%), and no significant differences were found between Arctic, boreal, or Arcto-boreal groups, or between calcifying and non-calcifying groups. However, suitable habitat areas for 14 of 78 taxa were projected a change by over 20%, and several of these taxa are characteristic and/or habitat-forming fauna on some Arctic shelves, suggesting a potential for significant ecosystem impacts. These results highlight the weakness of general statements regarding vulnerability of taxa on biogeographic or presumed physiological grounds, and suggest that more basic biological data on Arctic taxa are needed for improved projections of ecosystem responses to climate change.

Keywords: benthic invertebrates, climate warming, multiple stressors, ocean acidification, species-distribution modeling

## INTRODUCTION

Ongoing climate change affects multiple environmental drivers, with direct and indirect implications for marine ecosystems. For example, reductions in sea-ice volume and extent, increasing air and sea-surface temperatures, and freshwater and biogeochemical discharge in the Arctic have already been noted to be occurring at rates higher than the global average (IPCC, 2014; AMAP, 2017) with consequences for regional ocean acidification rates (Bates et al., 2012; Bellerby, 2017). Anthropogenic impacts from fossil fuel CO<sup>2</sup> emissions are linked with changes in these drivers and have led to changes in ocean chemistry. These impacts include reductions in both

ocean pH and carbonate mineral saturation states, which have been implicated in decreased performance of many marine taxa (Pörtner, 2008; Kroeker et al., 2011). Models project that ongoing ocean acidification (OA) will continue over the coming decades, and the degree of OA will be dependent mainly on social and policy decisions (e.g., Bellerby et al., 2014, 2018). The amplification of both warming and acidification observed in the Arctic, in combination with other expected physical and biological impacts, has led to the common viewpoint that Arctic ecosystems are particularly vulnerable to on-going climate changes (Smetacek and Nicol, 2005; Renaud et al., 2007; Brierley and Kingsford, 2009; Doney et al., 2012; Hoppe et al., 2018). Changes in community composition that may result are not merely of academic interest: they may reflect underlying functional responses with broad-reaching implications for ecosystem processes and services.

Laboratory experiments and field sampling have generated numerous but often contradictory predictions of how changes in climate-change-related drivers will affect marine life, from individuals to ecosystem processes. Physiology, phenology and life-history, community structure, and system productivity (Pörtner et al., 2005; Ardyna et al., 2014; Renaud et al., 2015, 2018) are just several examples of impacts that may be felt: from the very basic processes of life to the provisioning of ecosystem services. Until now, however, most studies have focused on effects of single factors, and usually in laboratory or isolated field settings (Pedersen et al., 2016; AMAP, 2018). Recent syntheses have called for investigations of broader thematic and spatial scope, including studying multiple factors in combination for arriving at more realistic predictions (e.g., AMAP, 2018), and adopting a pan-Arctic perspective when identifying trends in ecosystem response (Wassmann, 2015). These perspectives are needed for developing mitigation, adaptation, and management strategies for marine ecosystems across the Arctic (Halpern et al., 2012). Whereas some Arctic taxa are clearly threatened by changing climatic conditions (Kovacs et al., 2011; Doney et al., 2012; Grebmeier, 2012; Beaugrand and Kirby, 2018), a multifactor investigation of vulnerability of Arctic taxa in general has not been performed.

One of the most intuitive changes that could be expected due to climate warming is the poleward expansion of boreal organisms. This has already been observed for both benthic and pelagic organisms, and is suggested to be largely related to thermal tolerance (e.g., Wethey and Woodin, 2008; Sunday et al., 2012; Neukermans et al., 2018). Indeed, dominance of boreal and Arctic taxa in seafloor (benthic) communities of the Barents Sea have been shown to fluctuate with temperature in the region over the past 100+ years (Blacker, 1965; Matishov et al., 2012). Niche theory is useful for making predictions as to historical or future distributions in the context of climate change. Single-variable models (thermal-envelope models) have been used to predict, for example, changes in fish distribution with climatic warming (Cheung et al., 2009). In Arctic systems, this technique is problematic due to high uncertainty in either fundamental or realized niches of nearly all taxa in the ecosystem. On-line databases can provide one way around this problem by identifying the characteristics of an organism's realized niche. This technique has been used to identify benthic taxa that may expand or contract under warming scenarios (Renaud et al., 2015). Although biased by the geographical range of data included in databases, and non-random sampling in general, the method can also provide information for multi-dimensional realized niches. Once current distributions are mapped, a suite of species distribution models may be applied to project future distributions in response to multiple environmental drivers. Such modeling activities have been identified as critical next steps in understanding climate change effects on benthic community structure in Arctic environments (Renaud et al., 2015).

A key requirement for implementing these models is a well resolved and validated hydrographic/ocean chemistry model with realistic climate forcings. This allows for accurate physical characteristics to be assigned to each sampling location in both hindcasting and forecasting modes. One such model has recently been expanded to include the needed spatial and temporal perspectives for the carbonate system, and has been used to upscale results of a mesocosm experiment to other Arctic areas (Bellerby et al., 2012).

This study investigates the vulnerability of benthic fauna found in the Arctic and its marginal seas to end-of-century climate change. We use outputs from a state-of-the-art ecosystem model to drive species-distribution modeling for a broad range of characteristic macro- and mega- benthic taxa characteristic of the Arctic's shelf seas. We analyse results to determine if Arctic taxa are more susceptible to changes in a combination of environmental drivers than boreal or more widespread taxa. We also investigate whether calcification status or higher taxonomic classification (phylum, class) could indicate vulnerability. Implications of the presence or absence of such patterns are discussed.

## MATERIALS AND METHODS

## Environmental Projections

We ran the ocean biogeochemical model SINMOD (Slagstad et al., 2011, 2015), which includes a CO<sup>2</sup> system module (Bellerby et al., 2012), to estimate bottom water temperature and aragonite saturation state ((ar)). The SINMOD domain is pan-Arctic (see **Figure 1**) with 20 km horizontal grid resolution and 25 fixed depth levels, and calculations were based on biweekly saved output. Saturation state describes the seawater concentrations of carbonate and calcium ions relative to the equilibrium concentrations for that mineral; conditions near or below saturation ( = 1) are generally considered unfavorable for production or maintenance of normal calcification. SINMOD was first run to produce a hindcast simulation (1979–2008), and then to produce a projection (2001–2099) under the SRES (Special Report on Emissions Scenarios) scenario A1B (see Slagstad et al., 2015; Wallhead et al., 2017, for more details). SRES A1B is a mid-range business-as-usual scenario assuming a "balanced" use of fossil vs. non-fossil energy sources (Nakicenovic et al., 2000) that results in end-of-century global responses that are comparable to the more recent RCP6.0 scenario and

A1B scenario. Left and middle figures show averages over the hindcast period (1978–2008) and the projection period (2090–2099), respectively; figures on the far

substantially more conservative than the SRES A2 or RCP8.5 scenarios (Collins et al., 2013). SINMOD bottom-water output was corrected using bias estimates, calculated as a function of spatial position (temperature) or of model bathymetry ((ar)) using a compilation of in situ observations and matched model output (see Wallhead et al., 2017). Water depth over the SINMOD grid was estimated by interpolating high-resolution bathymetry products (the 500 m resolution IBCAOv3, Jakobsson et al., 2012, where available, otherwise the 2 min ETOPOv2, National Geophysical Data Center, 2006).

right show the differences. Black lines show 50, 200, and 500 m depth contours.

## Niche Description and Climatic Change

Occurrence records for 95 benthic taxa characteristic of Arctic shelf infaunal and epifaunal communities were extracted from the Ocean Biogeographic Information System database (OBIS<sup>1</sup> ; extracted February 14, 2017; **Supplementary Table S1**). We chose taxa that were common, ecologically relevant (e.g., ecosystem engineers) and/or characteristic taxa of Arctic shelves, and where biogeographic affinities are well-known (see below). In addition, taxa were chosen to cover the range of factors studied here (biogeographic affiliation, calcification status, and higher taxonomic level). Further, we required that each taxon have at least 30 records within the geographic domain of SINMOD in order to maintain MaxEnt model performance and confidence level. This cutoff appeared to be sufficient since the model performed quite well for taxa with close to 30 observations (e.g., Parastichopus tremulus and Serripes groenlandicus, see **Supplementary Tables S2, S3**). Bottom water temperature and aragonite saturation state from the bias-corrected SINMOD hindcast output were matched to the georeferenced benthic fauna distributional records by interpolating from the model grid to the faunal data positions and averaging over the 5 years preceding each faunal observation date. Water depth at the sampling points was mostly (61%) from measurements when plausible values were provided; the remaining 39% were estimated using the high-resolution bathymetry products. Each species was assigned status as calcifying or not, biogeographic affinity (Matishov et al., 2012; World Register of Marine Species)<sup>2</sup> , and higher taxonomic grouping (mollusk, echinoderm, polychaete, crustacean, bryozoan).

In locations where species data have been collected systematically, for example through biological monitoring, both presence and absence of species at each site have been recorded. However, most observations of the OBIS database have been collected non-systematically and are available as presence-only records, and the different gear types deployed and study foci make it difficult to infer absence data from these records. We therefore employed a species distribution modeling approach that requires only presence data, in order to maximize the utility of the database.

From the water depth, temperature, and aragonite saturation state associated with each species record, we estimated three-dimensional realized niches and related them to the projected changes in these parameters through the end of the century. The contribution of each environmental variable to species occurrence probabilities in the Arctic was calculated using the MaxEnt method. MaxEnt is a machine-learning algorithm for modeling species distributions from presence-only

<sup>1</sup>www.iobis.com

<sup>2</sup>www.marinespecies.org

species records. In brief, MaxEnt identifies the environmental characteristics of locations with species occurrences that differ from the environmental setting across the whole geographical region of interest. Based on the observed mismatch, a species' suitable habitat is defined. More specifically, the MaxEnt model minimizes the relative entropy between two probability densities (one estimated from the presence data, and one from the landscape) defined in covariate space compared to a uniform distribution null model (see Merow et al., 2013 for details). MaxEnt's predictive performance is consistently competitive with the highest performing methods. Since becoming available in 2004, it has been utilized extensively for finding correlates of species occurrences, mapping current distributions, and predicting to new times and places across many ecological, evolutionary, conservation and biosecurity applications (Elith et al., 2006).

## Sampling Biases

By default, MaxEnt models assume that: (1) all locations on the landscape (or "background locations") were equally likely to be sampled, and (2) any focal species would have been recorded at the sampled landscape points if it had actually occurred there. Assumption (1) is likely violated in our occurrence datasets, because of logistical and practical constraints on sampling in the Arctic (e.g., due to ice cover). Given such spatiotemporal biases, one cannot differentiate whether species are observed in a particular environmental niche because those locations are preferable or because they receive the largest search effort. Therefore, in order to relax assumption (1), landscape points were only drawn from the "target group sampling" (TGS) locations, which we defined as the total set of sampled locations for all species. However, this does not relax assumption (2), which may be violated where sampling gear and study focus have excluded certain species (Phillips et al., 2009). In theory the gear/study focus biases could be accounted for by partitioning the TGS, but this was not considered feasible for our particular dataset due to insufficient samples.

## Model Fitting and Validation

Multicollinearity can be an issue with MaxEnt when answering if and when environmental variables are of ecological interest. Thus, prior to modeling, a correlation analysis was conducted for environmental variables and the final MaxEnt models included variables that were not significantly correlated with each other at p < 0.05 (aragonite saturation state, temperature, water depth). Calcite saturation state was excluded as it correlates highly with aragonite saturation state.

In this study MaxEnt models were fitted as combinations of basic functions and features. MaxEnt had six feature classes: linear, product, quadratic, hinge, threshold, and categorical. Products were all possible pairwise combinations of covariates, allowing simple interactions to be fitted. Threshold features allowed a "step" in the fitted function, hinge features were similar except they allowed a change in gradient of the response. Many threshold or hinge features were fitted for one covariate, giving a potentially complex function. Nevertheless, the MaxEnt program was allowed to simplify the associations between species and the environment and in many models only one or a few features were used, e.g., hinge, linear, and quadratic.

Segment-based (non-gridded) data were modeled using SWD (samples-with-data) format in MaxEnt for both presence and background sites. A separate model was run for each species. A 10-fold cross-validation was used to obtain out-of-sample estimates of predictive performance and estimates of uncertainty around fitted functions. In order to reduce model overfitting, a balance between accurate prediction (model fit) and generality (model complexity) was sought by maximizing the penalized maximum likelihood function, i.e., the gain function. When doing so, regularization or the LASSO penalty was applied by exploring a range of regularization parameter values and choosing a value that maximizes measures of fit on a crossvalidation data set. The LASSO penalty is based on the rationale that features with larger variance should incur a larger penalty and, thus be less likely to be included in the model (Hastie et al., 2009). For model validation a random selection of 25% of the overall localities of species occurrences were used. The percent contributions of individual variables to the final model were identified with jackknife tests. The jackknife test evaluates how each variable contributes to the "gain" of the MaxEnt's model (i.e., improvement in penalized average log likelihood compared to null model) (Elith et al., 2011). We used the raw output (default) as this output is the closest estimate of the probability that the species is present (Elith et al., 2011).

## Model Performance and Predictions

We used the area under the receiver-operating-characteristic curve (AUC) as a single measure of overall model accuracy that is not dependent upon a particular threshold. The value of the AUC is between 0.5 and 1.0 with AUC = 1.0 indicating that the model has a perfect match and AUC = 0.5 indicating that model is no better than random (Fielding and Bell, 1997).

The MaxEnt model was used to predict suitable habitat for each species under current and future environmental conditions, and differences between these predictions were used to assess the response to climate change of potential habitat area for each species. We are aware that projecting future species distributions usually involves extrapolating models to novel combinations of environmental variables, and such projections should be treated with extreme caution. However, a comparison of current and future environmental niche space indicated that only 20% of the whole study area would enter a completely novel environment, as defined by moving beyond the 3D convex hull of the present-day niche space distribution. We therefore consider any artifacts of extrapolation will have only a minor effect on our results.

## Statistical Analyses of Model Results

Change in suitable habitat area from the present (1978–2008) to the future (2090–2099) was calculated from the MaxEnt model for each taxon. One-way analyses of variance (ANOVA) were performed for each of the three factors [calcification (as a t-test), biogeography, higher taxonomic level] on the change in habitat area (both percentage and absolute) after testing for homogeneity of variances and normality among factor levels (in each case these results were non-significant). Since single-factor analyses

may obscure interaction effects or even effects of one factor when effects of a second factor dominate the variability, we also ran a 3 factor ANOVA on the data. Since we lacked degrees of freedom to run a fully factorial test, we used a main-effects model. Analyses were performed in Statistica v. 13.

## RESULTS

## Environmental Settings Projections

Projections for bottom water temperature (T) and aragonite saturation state ((ar)) indicate substantial changes over the 21st century (**Figure 1**). The warmer Atlantic Water-influenced zone expands northward and eastward from its current core in the southern and southwestern Barents Sea (**Figures 1A,B**), and benthic habitat in the Barents Sea warms by up to 6◦C between 1979–2008 and 2090–2099 (**Figure 1C**). Over the same period, bottom (ar) on the shelves (<500 m depth) decreases, mostly by 0.6–1.1 units from around 1–2 (weakly saturated) to 0–1 (undersaturated) (**Figures 1D–F**). Exceptions include the shallow Russian shelf regions (<50 m depth), which are already strongly undersaturated ((ar) < 0.4; **Figure 1D**) and cannot become much more undersaturated in the future (**Figures 1E,F**). The largest decreases in (ar) (∼1 unit) are in the North, Bering, and Chukchi Seas, and on the East Siberian Shelf at ∼50–500 m (**Figure 1F**). Arctic water-dominated areas of the northern and eastern Barents/Kara Seas remain relatively stable in temperature (**Figure 1C**), but experience (ar) reductions as large as in the central Barents Sea (**Figure 1F**). Deeper areas (>500 m) in the Arctic and North Atlantic basins, and in Baffin Bay, see relatively small changes in benthic environment (**Figures 1C,F**).

## Depicting Realized and Predicted Niches of Benthic Taxa

Of the 78 taxa with sufficient (>30) records for evaluation, the mean change in suitable habitat was small: approximately a 5% loss in suitable habitat based on MaxEnt. The range, however, was high, from a 42% increase in habitat (the amphipod Byblis gaimardii) to a 53% decrease (the sea cucumber P. tremulus) according to the 3-variable model (**Table 1**, **Supplementary Table S2**, and **Figures 2**, **3**). MaxEnt performed very well for most (∼70%) taxa, with AUC > 0.9 (**Supplementary Table S3**). Temperature contributed most to defining the niche space for 72% of the taxa, while depth and aragonite saturation state were the largest contributors for 19 and 8% of the taxa, respectively (**Supplementary Table S3**).

Surprisingly, both calcifiers (in general and by phylum) and Arctic (compared to Arcto-boreal or Boreal) taxa were found to be resilient to the projected environmental changes in terms of group mean response. Calcifiers were slightly more vulnerable on average than non-calcifiers (6.1 vs. 2.8% mean habitat loss) but the difference was not significant due to high intra-group variability (t-test, P = 0.18, **Figure 4**). Neither biogeographic affinity nor higher taxonomic level had a significant influence on the group-mean change in habitat (one-way ANOVA tests, P = 0.39 and 0.80, **Figure 4**). When Arcto-Boreal and Arctic taxa were combined, their mean habitat


TABLE 1 | Winners (>20% increase in suitable habitat, upper panel) and losers


The number of records extracted from OBIS for each taxon (N), higher taxon, calcification status, and biogeographical zones are indicated for each taxon. See Figures 2, 3 for maps of present (1978–2008) and future (2090–2099) distributions for the two biggest winners and losers.

loss was less than for Boreal taxa (2.9 vs. 10.9% habitat loss, P = 0.043). Typically, heavily calcified taxa from the Mollusca and Echinodermata phyla have been suggested to suffer greatest from acidification (Hale et al., 2011). Our analysis suggested that only 24 and 25% of the taxa in these phyla, respectively, would experience range reductions, compared to 45–100% for other phyla (**Supplementary Table S2**). Similarly, only 12% of Arctic taxa lost habitat, compared with nearly 50% of Arcto-boreal and Boreal taxa (**Supplementary Table S2**). Results were the same for all analyses whether the absolute or percent change in suitable habitat was analyzed. The 3-factor main effects ANOVA also indicated no significant factor effects (biogeography: F = 2.1, df = 2, P = 0.14; calcification: F = 1.9, df = 1, P = 0.17; higher taxonomic level: F = 0.70, df = 4, P = 0.59).

The data we extracted from the OBIS database were at both the species and genus levels. In most cases, genus-level taxa were assigned to broader biogeographic zones (e.g., ArctoBoreal) because they include several species. We tested for this potential bias by analyzing only the data for taxa identified to the specieslevel or where it was reasonably certain that the taxa found could be ascribed to either Boreal or Arctic zones. Results of these analyses with 23 fewer taxa were nearly identical to those for the entire data set (3-factor main-effects ANOVA: biogeography: F = 2.4, df = 2, P = 0.10; calcification: F = 0.89, df = 1, P = 0.35; higher taxonomic level: F = 0.23, df = 4, P = 0.92).

(+40%): (C,D)] in the study area based on the results from the MaxEnt model (see Table 1). Maps show present (A,C) and predicted future (B,D) distributions for the two species. All models predict the expected probability of occurrence from 0 (0% probability; blue) to 1 (100% probability; red).

## DISCUSSION

We found little evidence to support the belief that Arctic taxa are more sensitive than taxa with more southerly or cosmopolitan distributions, nor did we find that calcifiers as a group are more sensitive to projected climate-related changes than non-calcifiers. Models did indicate, however, that individual taxa can experience considerable changes in suitable habitat under future scenarios, with potentially significant ecological consequences.

## Roles of Niche-Defining Parameters

When fitting MaxEnt models, temperature was the most important factor determining current suitable habitat for more than two-thirds of the taxa studied. This should come as little surprise as temperature is one of the main factors used to delineate biogeographical provinces (Longhurst, 1998), so studies over a large spatial scale with strong temperature gradients such as this are likely to show an important temperature effect. Temperature was also identified (by MaxEnt modeling) as the most important factor for mollusk and crustacean species distributions at local scales in an Arctic fjord (Drewnik et al., 2017). Temperature can act in many ways, and may be viewed as a "master parameter" that works directly through thermal tolerance for survival and/or reproduction (e.g., Wethey and Woodin, 2008), or indirectly through its impacts on, e.g., primary production, ice cover, and zooplankton and microbial grazing pressure, all of which could affect benthic food supply (Maar and Hansen, 2011). Shifts in community structure of Barents Sea benthos and fish have been noted, with the southern limit of Arctic communities moving northward during warming and retreating again southward during cooling periods (Blacker, 1965; Matishov et al., 2012; Fossheim et al., 2015). These changes were attributed to temperature, but it is unclear whether they were direct tolerance effects or something more indirect. It is important to note, however, that some high-latitude taxa have very narrow temperature ranges over which they occur, making them particularly vulnerable to climate-driven changes (Pörtner et al., 2014; Morley et al., 2019).

We used aragonite saturation state to represent the degree of potential species response to ocean acidification, but we acknowledge that this may not be the only relevant carbonate system driver. Whereas most of the calcifying taxa we studied produce aragonitic skeletons (Mollusca: Kuklinski, unpublished ´ data), there is also a large number of organisms included in our study which use calcite with varying degrees of Mg incorporation (Echinodermata, Bryozoa, Cirripedia: Kuklinski and Taylor, 2009; Iglikowska et al., 2017, 2018). In addition, some species may be more sensitive to pH or pCO<sup>2</sup> or combinations of OA stress.

However, we suspect that using only one parameter (aragonite saturation state) provides a reasonable first approximation, since most changes are expected to be quite strongly correlated between different carbonate system parameters (e.g., Figure 9 in Wallhead et al., 2017).

We included depth in our distribution models, even though it will not change significantly in the near future. Depth can often reflect a number of linked parameters important for an organism's habitat state, such as sediment grain size, bottom currents, and vertical flux of particulate organic carbon (i.e., a proxy for food for many benthic organisms). Maintaining homeostasis under conditions of ocean acidification, for example, is thought to require more energy, and, as such, may only be possible when sufficient resources are available for affected organisms (Pansch et al., 2014). Increased metabolic rates due to higher temperatures may enhance this challenge (Pörtner, 2008). Primary production in the Barents Sea is predicted to change, both quantitatively and spatially, due to climate change (Slagstad et al., 2015). Clearly, inclusion of depth in the model will not account for all of this, but given the strong relationship between depth and vertical flux, we expect use of depth as one of the niche parameters to some extent accounted for spatial differences in food supply, in addition to partially controlling for other habitat parameters.

## Accuracy of Environmental Data

The pan-Arctic SINMOD model, used here to provide bottom water temperature and aragonite saturation state, has been extensively tested and bias-corrected using field measurements (Slagstad et al., 2015; Wallhead et al., 2017). We therefore consider that the hindcast data used to define faunal niches are unlikely to have been a major source of bias. Projected changes for the 2090s are, however, subject to numerous model uncertainties and an ensemble approach would likely be required to quantify these. Nevertheless, we know that SINMOD gives bottom water projections that are comparable with global and other downscaling models, and it is sufficient for this study that the model gives plausible projections for the chosen climate change scenario (SRES A1B). For end-of-century simulations, the dominant source of uncertainty is likely to be the climate change scenario itself (Hawkins and Sutton, 2009). Water depth estimates for the pan-Arctic region are also not without uncertainty (e.g., the absolute differences between plausible measured values and the interpolated bathymetry products had median = 6 m, 95th percentile = 103 m) but it seems

unlikely that such errors could significantly bias our estimates of mean habitat loss.

## Surprising Robustness of Habitat Suitability

Given the reported higher vulnerability of polar organisms to ocean warming (IPCC medium confidence, Pörtner et al., 2014) and of highly calcified organisms to ocean acidification (IPCC high confidence, Ibid), the fact that our projections for benthic taxa at high northern latitudes show no clear increase in habitat loss for Arctic or calcifying species may seem surprising. It leads us to question whether something could have been lost in the complexities of the MaxEnt analysis, and if the results might be understood or corroborated by simpler analyses based on the statistics of the environment where the different species were found.

A first reason to expect higher Arctic vulnerability is that environmental changes for Arctic species may be larger under climate change, due to rapid loss of ice cover/thickness and stronger warming in polar regions (Smetacek and Nicol, 2005; Brierley and Kingsford, 2009; Doney et al., 2012). However, if we consider the median projected bottom water temperature change at the sampled presence points for each species (see **Figure 5A**), the species-mean warming is slightly higher for non-Arctic species (2.3 vs. 1.8◦C, P = 0.04, t-test, unequal variances). **Figure 5A** also shows that the warming is much more consistent for Boreal species than for non-Boreal species (standard deviation = 0.15 vs. 0.87◦C; P < 10−<sup>11</sup> , F-test). Regarding the median decreases in aragonite saturate state at sampled points (**Figure 5B**), the mean decreases are again slightly larger for non-Arctic species (0.87 vs. 0.63, P < 10−<sup>3</sup> , t-test). So at least for bottom water temperature and saturation state, it seems the environmental changes are not larger for Arctic species, as

(blue), and for both calcifiers (triangles) and non-calcifiers (circles). X-axes show the median projected changes in temperature (A) and aragonite saturation state (B), the 95% ranges in hindcast temperature (C) and aragonite saturation state (D), the 95th percentiles of hindcast temperature (E), and the 5th percentiles of hindcast aragonite saturation state (F). Projected changes are from the SINMOD projections (2090s vs. 1978–2008 under SRES A1B) interpolated to the presence locations; hindcast variables are from the bias-corrected SINMOD hindcast, interpolated to the presence locations and averaged over the 5 years preceding each sample date.

one might have anticipated from **Figure 1**. Furthermore, there are no significant correlations between the median environmental changes and the % habitat loss (**Figures 5A,B**).

A second basis for high Arctic sensitivity is the idea that polar organisms have narrow environmental tolerance ranges, e.g., a "stenothermal" characteristic due to low energy-demand lifestyles and specific adaptations to cold water (Pörtner et al., 2014; Morley et al., 2019). We can test this by considering the 2.5−97.5% ranges in bottom water temperature and aragonite saturation state at the sampled presence points as tolerance metrics for each species (**Figures 5C,D**). Here we find that Arctic species do have somewhat narrower ranges that non-Arctic species on average (6.0 vs. 9.6◦C, 1.2 vs. 1.8; P < 10−<sup>3</sup> , t-tests). However, variability in tolerance ranges is large within each group (**Figures 5C,D**), and there are no significant correlations between % habitat loss and tolerance ranges for temperature or saturation state (or indeed water depth, not shown). The rather moderate differences in mean tolerance range may in part reflect a higher level of historical variability in Arctic vs. Antarctic systems, at least for temperature (Pörtner et al., 2014). Arctic shelf systems have experienced, even in the past 150 years, several cycles of warming and cooling within the range we are observing today, and Arctic benthic fauna have expanded or retreated on decadal scales accordingly (e.g., Blacker, 1965).

A third basis for high Arctic sensitivity is the idea that polar organisms will have their habitat "squeezed out" under global change, either because they cannot move to higher latitudes in response to warming (Pörtner et al., 2014; Renaud et al., 2015), or because critical thresholds of saturation state will be approached first in the Arctic under ocean acidification (Feely et al., 2009; Steinacher et al., 2009; Ciais et al., 2013). If we consider the 95th percentile of bottom water temperature at the presence locations as a simple metric of upper thermal limit (**Figure 5E**), we find that Arctic species do indeed have a significantly lower upper thermal limit on average (3.3 vs. 9.2◦C, P < 10−12, two-sample t-test). However, there is only a marginally significant correlation between % habitat loss and upper thermal limit (Pearson's r = 0.23, P = 0.04). Considering the 5th percentile of aragonite saturation state at presence locations as a lower tolerance limit and potential predictor of vulnerability to acidification, we find no significant difference between Arctic vs. non-Arctic species, and no significant overall correlation with % habitat loss (**Figure 5F**).

With regard to the paradigm of higher sensitivity of calcifying species, we find no significant difference in species-mean 5th percentiles of aragonite saturation state at presence locations for calcifiers vs. non-calcifiers (0.71 vs. 0.67, P = 0.76, t-test), and both categories contain species with 5th percentiles much less than 1, and in several cases < 0.2 (**Figure 5F**). Such tolerance of corrosive water may be partly explained by physiological mechanisms associated with the calcification process, particularly in these often heavily calcified taxa, which are already adapted for internal control of carbonate saturation state at the site of precipitation (Hendriks et al., 2015). Considerable variability in response to ocean acidification has been shown in experimental manipulations (Kroeker et al., 2011; Vihtakari et al., 2016), and

naturally high environmental variability in some habitats (shallow waters, vegetated substrate) may also help to promote tolerance via phenotypic plasticity and local adaptations (Hofmann et al., 2010). Such mechanisms may make the species less susceptible to modest decreases in saturation state, even at under-saturation levels, although food supply may mediate the ability of an organism to regulate conditions for calcification (Ramajo et al., 2016).

## Analysis of Habitat Losses in Niche Space

It is worth noting that neither large environmental changes nor narrow tolerance ranges will necessarily result in high species habitat loss; if climate change produces enough new habitat, the net loss of habitat may be zero (or indeed negative) no matter how large the changes or narrow the tolerances. Rather, large changes and narrow tolerances will likely result in large (negative or positive) changes in suitable habitat area. To better understand the projections, it can be helpful to consider the calculation of % habitat loss in niche space – here a 3D space of (temperature, saturation state, water depth). The niche space can be divided into a grid of subvolumes, where the ith subvolume "contains" a physical area of habitat A (1) i in the present and A (2) i in the future. The MaxEnt algorithm essentially determines an occupancy probability Pji for the jth species and ith nichespace subvolume by fitting a complex preference function to the present-day niche occupancy data (Merow et al., 2013). Assuming no change over time in this niche-space preference function (ruling out e.g., genetic adaptation), the net change in habitat area Hj for species j is given by:

$$
\Delta H\_{\vec{j}} = \sum\_{i} P\_{\vec{j}i} \left( A\_i^{(2)} - A\_i^{(1)} \right),
$$

It follows that the projected habitat losses can in large part be explained by the changes in niche subvolume area content 1A<sup>i</sup> = A (2) <sup>i</sup> − A (1) i . These latter are entirely independent of species and the MaxEnt procedure, and depend only on the present and future environment (here based on the SINMOD biogeochemical model and the high-resolution bathymetry products).

The habitat area changes over niche space can be visualized on a 2D grid over temperature (T) and aragonite saturation state ((ar)) by integrating over the water depth dimension (**Figure 6A**). The largest losses are in the red cells around (- 2 ≤ T < 1 ◦C, 1 ≤ (ar) < 1.5), while the largest gains are in the blue cells around (−2 ≤ T < 1 ◦C, 0 ≤ (ar) < 0.2) and (0 ≤ T < 5 ◦C, 0.4 ≤ (ar) < 1). In order to be a "winner" under the projected environmental change, a species needs to have an occupancy probability or niche preference function Pji that favors the niche subvolumes that gain habitat area. This requires that its present-day sampling distribution over niche space gives evidence that the species can tolerate the winning blue cells. For example, the biggest winner, the amphipod B. gaimardii (+42% habitat area), has a widespread niche space distribution (magenta crosses). Comparing this to the sampling distribution over all species (the TGS, black dots), the MaxEnt algorithm infers a strong tolerance of low (ar) and a thermal window that comfortably contains the winning range (0 ≤ T < 5 ◦C). This allows B. gaimardii to make large area gains in the future Barents Sea, Bering Strait, Chukchi Sea, and on the East Siberian shelf between 50 and 500 m (**Figures 1B,E**, **2A,B**). By contrast, the biggest loser, the sea cucumber P. tremulus (-53% habitat area), has a much more confined sampling distribution in niche space (magenta circles), suggesting limited tolerance of undersaturated water ((ar) < 1) and a thermal window of roughly 4–10◦C. Consequently, P. tremulus loses habitat in the North Atlantic and becomes increasingly confined to the Norwegian shelf shallower than 500 m (**Figures 3A,B**).

**Figure 6A** also sounds a note of caution regarding our projections: the sampling distribution of presences (black dots) is highly non-uniform over niche space, and in particular the winning niche cells (blue colors) are strongly undersampled. For several winning cells, the present-day niche area is zero (indicated by a gray "x"); these are new niches that do not exist yet in the pan-Arctic region. Although the MaxEnt procedure is rarely strictly extrapolating with respect to the 3D convex hull (see section "Materials and Methods") it is often being used to make bold interpolations based on limited data. This issue, plus other potential sources of bias (e.g., gear biases), reduce our confidence in the projections for individual species habitat loss, although we suspect a lesser impact of sampling bias on the significance of differences between group means. Our first attempt at running the analysis used the default MaxEnt assumption of uniform sampling effort over the pan-Arctic grid (a clearly untenable assumption, see **Figure 6B**); this led to much larger group-mean habitat losses (59–82%), but mean losses of Arctic species (75%) were still not significantly different to those of boreal (59%) or Arcto-boreal species (82%). The results presented in this paper account for this first assumption.

Further insights into vulnerability and sampling needs are provided by interpolating the 3D niche space area changes onto the present-day (1978–2008 average) values over the SINMOD grid (**Figure 6B**). Here, any species whose presentday geographical distribution is confined to areas of niche area loss (yellow/red colors) is set to be a loser under the projected changes. Hence, we can project, independently of any MaxEnt analysis, that species with environmental tolerances that confine them to e.g., the northern Barents Sea, or the East Greenland Shelf, are likely to lose habitat under climate change. Conversely, any species that tolerates present-day conditions in the Kara Sea near the Gulf of Ob, or the shallow areas (<50 m) of the eastern Laptev Sea and East Siberian Sea, is set to win. It is thus unfortunate that relatively few species presence records from these latter areas have made it into the OBIS database, and future attempts to project future species distributions would likely be well served by more sampling effort in and/or data recovery from these particular regions.

## Ecological Consequences of Distributional Shifts

Approximately half the taxa were projected to experience substantial changes in suitable habitat area (>10% increase or decrease), and some of these species are characteristic

or dominant on Arctic continental shelves. The polychetes Chaetozone setosa and Spiochaetopterus typicus, the bivalve Medicula ferruginosa, the barnacle Balanus balanus, and the decapods Pagurus pubescens, Hyas araneus, and Pandalus borealis are common and characteristic taxa in the open Barents Sea and adjacent fjords. S. typicus (12% increase in habitat) and B. balanus (26% loss) are key ecosystem engineers in that the physical structure they provide is important for ecosystem function. Pandalus is an important commercial shrimp in Arctic and boreal waters, and a 17% increase in habitat was predicted. If these taxa are as strongly affected as our analyses suggest, the impact on ecosystem structure and services in the Barents region could be significant over the 21st century.

Surprisingly, one of the biggest "winners" in our analyses, the Arctic bivalve Yoldiella frigida (40% habitat increase), was described as one likely to experience substantial range retraction in an analysis based solely on temperature ranges (Renaud et al., 2015). However, the Renaud et al. (2015) projection considered only the Barents Sea. The MaxEnt algorithm predicts a much more modest increase for the Barents Sea (9%), with most of the pan-Arctic habitat increase driven by gains on the Chukchi/Russian shelves and coastal waters off Canada, Greenland, and Iceland (**Figure 2D**). Also, Renaud et al. (2015) assumed a hard upper temperature limit of 4◦C; the MaxEnt algorithm infers some tolerance of warmer temperatures (thus allowing for the slight increase in the future Barents Sea) based on the limited presence records (41) within the hindcast timeframe (1978–2008). In fact the histogram shown in Figure 5 of Renaud et al. (2015) also suggests some tolerance to warmer waters, although this latter is based on Augustonly SINMOD output for 2000–2009 without bias correction or matching to sample date.

One caveat of our study is the exclusive focus on bottom water environment, while many benthic organisms reproduce via planktonic larvae that live for hours to months in the pelagic zone. A recent study of an Antarctic barnacle indicated the need to consider both stages of multi-phasic species since conditions are likely to change differently in the two ocean realms (Gallego et al., 2017). Registered presence in the OBIS database, as used in our study, integrates environmental influence on multiple life-history stages/processes from fertilization, to development and dispersal, to recruitment and juvenile stages, any of which can be acted upon by ecological (including acidification) interactions in both the water column and sea floor. Another possible caveat is our use of only two physical/chemical parameters. Many other such parameters (e.g., carbon flux, bottom-water oxygen content) may define where an organism can exist, and ecological interactions cannot be expected to be unchanged as species ranges change. Kroeker et al. (2012) found ocean acidification to alter competitive hierarchies of macroalgae, and this could have consequences for associated fauna. In contrast to what has been observed for Barents Sea fish communities (Fossheim et al., 2015), benthic community change is not likely to progress via wholesale replacement of Arctic fauna with boreal fauna. Highly mobile fish are expected to move northward and southward in response to changing extent of Atlantic Waters, whereas benthic taxa have limited mobility and range changes are more likely to vary by taxa as a consequence of life-span, dispersal ability, etc. New species combinations will likely lead to novel and unpredictable species interactions, further affecting species distributions (Pecl et al., 2017). Finally, our models are based on fixed niches so that populations follow the environmental change by shifting their geographical distributions. However, populations may adapt to

tolerate novel conditions through selection by evolving traits that provide tolerance in the new environment. Although it is tempting to project future distributions that take into account evolutionary responses, scientists seldom know if, or how quickly, the climate-sensitive traits of populations can evolve (Merilä and Hendry, 2014; Jezkova and Wiens, 2016).

## CONCLUSION

Despite acknowledged issues and uncertainties with using species distribution models, especially in poorly sampled areas like the Arctic, such models are accepted as useful in providing a first approximation of how climate change will affect biodiversity (Pearson and Dawson, 2003; Wiens et al., 2009). This study is not meant to emphasize fine-scale species-specific changes, but rather to delineate expected shifts in broad patterns. The findings reported here are surprising in the context of current paradigms regarding susceptibility to climate change. In particular, we found that for benthic species under end-of-century ocean warming and acidification: (i) mean habitat losses over taxonomic groups were generally small (0–11%), (ii) Arctic benthic species were not significantly more vulnerable than boreal or Arcto-boreal species, (iii) calcifying species were not significantly more vulnerable than non-calcifiers. The lack of sensitivity of our results to such taxonomic groupings was consistent with simpler analyses based on magnitude of environmental change and observed tolerance windows, although such criteria were individually poor predictors of the habitat loss results from MaxEnt. Sampling bias in the presence data was a major issue, and in general we expect projections from benthic species distribution models for high northern latitudes to be strongly sensitive to how this is dealt with. Analysis of niche space area changes, dependent only on the environmental data, suggested that largest future habitat losses will be for taxa now inhabiting cold/oversaturated niches (−2 < T < 3 ◦C, 1 < (ar) < 1.5, e.g., the present-day northern Barents Sea and East Greenland Shelf) while the largest gains will be for those currently found in cold/undersaturated niches (−2 < T < 5 ◦C, (ar) < 1, e.g., the present-day eastern Laptev and East Siberian Seas). The latter "winning" niches were undersampled in our dataset, and we expect that further field sampling and data recovery for these niches will be especially helpful for projecting future species distributions.

There are few similar studies from high latitudes (but see Morley et al., 2019 for Antarctic waters), mainly due to limited information about species ranges and physiological tolerances, and regional models powerful enough to project environmental parameters into the future. As more data become available and models are developed, the tools we used here can be improved, and predictions refined. As in other predictive modeling approaches, MaxEnt is a process that uses data mining and probability to forecast outcomes. When doing so, the established relationships are only based on statistical dependence between environmental and distributional data, and therefore may fail to account for physiological limits and biological interactions. Development of hybrid modeling platforms that integrate species distribution models and experimental results can facilitate higher resolution distribution models for individual taxa, which can be particularly useful for economically or ecologically influential species (Kotta et al., 2019). This work is important as changes in ranges of some of these species may have substantial impacts on ecosystem services provided by benthic communities (Pecl et al., 2017).

## DATA AVAILABILITY

The datasets generated for this study are available on request to the corresponding author. Model output data are available on Zenodo (https://zenodo.org/record/3245446).

## AUTHOR CONTRIBUTIONS

PR, MW-K, and PK developed the study idea. JK and MR performed MaxEnt modeling. RB and DS developed the SINMOD simulations. PW analyzed and bias-corrected the SINMOD output. PK was the project leader. PR led work on the manuscript. All authors contributed to the interpretation of results and editing of the manuscript.

## FUNDING

This work was supported by the Polish-Norwegian Research Programme operated by the National Centre for Research and Development under the Norwegian Financial Mechanism 2009–2014 in the frame of Project Contract No. Pol-Nor/196260/81/2013. Additional funding was provided by the Norwegian Foreign Ministry and its Arctic 2030 program (CoArc project), the Norwegian Research Council ("Adapting Coastal Zone Management to Ocean Acidification (ACIDCOAST)"; Grant No. 255748), the FRAM – High North Research Centre for Climate and the Environment under the Ocean Acidification Flagship, and the BONUS MARES project funded by the European Union's Seventh Framework Programme for research, technological development, and demonstration through BONUS, the Joint Baltic Sea Research and Development Programme (Art 185), and Akvaplan-niva.

## ACKNOWLEDGMENTS

The authors are grateful for technical assistance provided by Hector Andrade. Species occurrence and bias-corrected model output data used in this manuscript are openly accessible at Zenodo (https://zenodo.org/record/3245446). Further SINMOD output from the runs used in this manuscript can be found at https://doi.org/10.5281/zenodo.886077.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars.2019. 00538/full#supplementary-material

## REFERENCES

fmars-06-00538 August 30, 2019 Time: 17:21 # 13


Proc. Natl. Acad. Sci. U.S.A. 108, 14515–14520. doi: 10.1073/pnas.110778 9108


**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 Renaud, Wallhead, Kotta, Włodarska-Kowalczuk, Bellerby, Rätsep, Slagstad and Kuklinski. 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.

# Climate Change Vulnerability of American Lobster Fishing Communities in Atlantic Canada

Blair J. W. Greenan<sup>1</sup> \*, Nancy L. Shackell <sup>1</sup> , Kiyomi Ferguson<sup>1</sup> , Philip Greyson<sup>1</sup> , Andrew Cogswell <sup>1</sup> , David Brickman<sup>1</sup> , Zeliang Wang<sup>1</sup> , Adam Cook <sup>1</sup> , Catherine E. Brennan<sup>1</sup> and Vincent S. Saba<sup>2</sup>

*<sup>1</sup> Fisheries and Oceans Canada, Bedford Institute of Oceanography, Dartmouth, NS, Canada, <sup>2</sup> NOAA National Marine Fisheries Service, Geophysical Fluid Dynamics Laboratory, Northeast Fisheries Science Center, Princeton University Forrestal, Princeton, NJ, United States*

#### Edited by:

*Elizabeth A. Fulton, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia*

#### Reviewed by:

*Kate Barclay, University of Technology Sydney, Australia Christopher James Brown, Griffith University, Australia*

\*Correspondence:

*Blair J. W. Greenan blair.greenan@dfo-mpo.gc.ca*

#### Specialty section:

*This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science*

> Received: *30 April 2019* Accepted: *30 August 2019* Published: *13 September 2019*

#### Citation:

*Greenan BJW, Shackell NL, Ferguson K, Greyson P, Cogswell A, Brickman D, Wang Z, Cook A, Brennan CE and Saba VS (2019) Climate Change Vulnerability of American Lobster Fishing Communities in Atlantic Canada. Front. Mar. Sci. 6:579. doi: 10.3389/fmars.2019.00579* Climate change impacts on fisheries will undoubtedly have socio-economic impacts on coastal communities and the seafood market. However, it is a challenge to integrate climate change information in a form that can be used efficiently by adaptation planners, policy makers, and fishery managers. In this study, we frame a climate change impact assessment using a geographical perspective based on the management units of the dominant fishery, in this case, American lobster in Nova Scotia, Canada. The information considered here includes economic dependence on the fishery, population size, diversity of the fishery revenue, status of harbor infrastructure, total replacement cost of each harbor, increased relative sea level and flooding, and the vulnerability of offshore lobster to ocean warming and changes in zooplankton composition and anticipatory changes in fishery productivity across management borders. Using two ocean models to provide multi-decadal scale projections of bottom temperature, changes in offshore lobster distribution are projected to have a neutral, or positive impact on the region as a whole. However, when lobster vulnerability is combined with climate change related vulnerabilities of coastal fishing communities, it is evident that adaptation planning is needed for long-term sustainability. This impact assessment provides both a framework and information for further in-depth analyses by climate change adaptation planners and fishery managers.

Keywords: climate change, lobster, thermal habitat, coastal vulnerability, harbor infrastructure, climate projections, fishery management

## INTRODUCTION

Globally, coastal regions and communities have been identified as particularly vulnerable to climate change. This is recognized in Canada in the Pan-Canadian Framework on Clean Growth and Climate Change (https://www.canada.ca/en/services/environment/weather/climatechange/ pan-canadian-framework.html). Hence, there is a need to adapt and build resilience in these communities so that they are adequately prepared for climate risks like coastal flooding, extreme weather events, and shifting fish populations. The Government of Canada has recently appointed a Minister of Rural Economic Development with the primary goal of creating a Canadian Rural Economic Development Strategy. At the provincial level in Nova Scotia, a goal was set to double the value of exports (relative to a 2012 baseline) from the fisheries (including aquaculture) and the agricultural sectors on a sustainable basis by 2024 (https://onens.ca/goals/goal-15 fisheries-and-agriculture-exports/). The determination of this sustainability will be contingent upon accounting for the impacts of climate change on the fisheries and fishing communities.

Ocean temperatures on the continental shelves off the Northeast USA and southern Atlantic Canada have increased over the past half century (Jewett and Romanou, 2017; Greenan et al., 2018b), consistent with the global trend of increasing ocean heat content resulting from climate change (Cheng et al., 2017). The resulting biological impacts vary both regionally and by species (Fogarty et al., 2007; Wernberg et al., 2011; Pinsky and Fogarty, 2012; Shackell et al., 2014; Stortini et al., 2015, 2017; Kleisner et al., 2016, 2017). As warming continues, conditions may become uninhabitable for some species while others may flourish (Sorte et al., 2010). These long-term changes can have a major impact on commercial fisheries if waters become unsuitable for species of economic importance and range shifts lead to a decrease in local abundance (Fogarty et al., 2007). Alternatively, an area may experience a change in abundance of certain species if warming improves habitat suitability (Fogarty et al., 2007; Sorte et al., 2010; Shackell et al., 2014; Stortini et al., 2015).

American lobster (Homarus americanus) is Canada's most valuable fishery (\$1.3 B in 2016), and contributed 44% of the total commercial value of all fisheries in Atlantic Canada in 2016 (DFO, 2018). Lobster landings have been trending upward in recent decades among the 45 directed fisheries in the Atlantic provinces and Quebec (Bernier et al., 2018). Many small rural communities in Atlantic Canada rely heavily on lobster for their economic well-being, although snow crab (Chionoecetes opilio) is important in the colder northeast region (**Figure 1**). Ocean temperatures above an optimal thermal range can reduce lobster survival, growth, and reproduction as a result of stress, decreased recruitment, and increased disease (Aiken and Waddy, 1986; ASMFC, 2015). The scale and characteristics of lobster response to warming varies across its range (Boudreau et al., 2015). Lobster abundance in the Gulf of Maine was at record high in 2015 (ASMFC, 2015) where the lobster industry has initially benefitted from a loss of lobsters' predators, warming bottom temperature (Boudreau et al., 2015) as well as strong conservation measures (Le Bris et al., 2018). In contrast, lobsters are declining in the warmer southern New England where conservation measures are fewer (Le Bris et al., 2018).

The observed boom in the Gulf of Maine may not be permanent given that warming-induced changes in molting and timing of migration have extended the fishing season while increasing the number of individuals eligible to the fishery and may lead to over-fishing (Mills et al., 2013) alongside a declining trend in young of year in the Gulf of Maine and Georges Bank since 2012 (Carloni et al., 2018). While ocean temperature change and fishing pressure are important factors impacting lobster distributions, the processes, and interactions that occur at early life cycle stages are also highly relevant. Recently hatched, planktonic lobster spend the first 6–10 weeks of life in nearsurface waters, during which planktonic organisms make up the bulk of lobster diets (Lawton and Lavalli, 1995; DFO, 2009). In the Gulf of Maine, a correlation has been identified between the abundances of post-larval and young-of-year lobster, and the copepod species Calanus finmarchicus, suggesting a link between the declining trend in lobster recruitment and deviations in zooplankton assemblages in this region (Carloni et al., 2018). As ocean temperatures warm and habitat suitability decreases in some areas, ensuring that management and fishing practices regionally are tailored to support the future of the stock can help moderate the effects of ocean warming (Le Bris et al., 2018).

In this paper, we present an analysis of coastal vulnerabilities to climate change (physical environment, socio-economic, and infrastructure), alongside potential responses of adjacent lobster populations given increased ocean temperatures. This is performed at the management unit scale as a means to identifying variation between regions, and where/how preparation for investment and/or adaptation strategies may be beneficial in boosting local resiliency to future changes (Colburn et al., 2016). Our analysis of the projections for the lobster fishery is a first step in considering this information in local fishery management decisions and longer term economic development strategies.

## DATA AND METHODS

Providing climate change adaptation tools that integrate changes in the physical environment with fisheries response and the potential socio-economic impacts presents a significant challenge. However, such tools are needed by adaptation planners, policy makers, and fishery managers. In the case of fishery managers, they generally make decisions at the scale of the stock management unit as opposed to the entire range of the species. The approach adopted in this paper is to generate two climate change vulnerability indices (one for coastal communities and one for lobster) and then aggregate this information at the scale of fishery management units (**Figure 2**).

In the first case, the Coastal Infrastructure Vulnerability Index (CIVI) will provide a measure of the relative vulnerability of DFO Small Craft Harbor (SCH) locations to climate change (Cogswell et al., 2018; Greenan et al., 2018a). The lobster fishery in Nova Scotia is supported by an extensive network of coastal infrastructure, which is a nationwide SCH program responsible for the maintenance of more than 1,000 harbors with infrastructure (e.g., wharves, breakwaters, buildings) valued at approximately \$5.6 billion. While the harbors are locallyoperated and managed by not-for-profit Harbor Authorities, the SCH program provides the property, infrastructure, liability insurance, and budget for major and minor repairs. These harbors are critical to the fishing industry and the economy of rural coastal communities in Canada.

**Abbreviations:** BNAM, Bedford Institute of Oceanography (BIO) North Atlantic Model; CIVI, Coastal Infrastructure Vulnerability Index; CM2. 6, NOAA Geophysical Fluid Dynamics Laboratory's Climate Model 2.6; DFO, Department of Fisheries and Oceans; ESI, Exposure Sub-Index; ISI, Infrastructure Sub-Index; LFA, Lobster Fishing Area; LVI, Lobster Vulnerability Index; NOAA, National Oceanographic and Atmospheric Administration; SCH, Small Craft Habors; SESI, Socio-Economic Sub-Index; SVD, Species Value Diversity.

FIGURE 1 | Value of fisheries landings associated with each Lobster Fishing Area (LFA) in the Maritimes Region of Fisheries and Oceans Canada (DFO). Landings are reported at the community level, and aggregated at the level of LFA. The size of pie charts in each LFA is scaled by the total value of landings and the proportions presented for lobster, snow crab and all other species. DFO Small Craft Harbor locations are represented by the black dots.

While CIVI is comprised of exposure (natural forces), infrastructure, and socio-economic indicators, it does not incorporate biological impacts of climate change. In the second part of the analysis, a species vulnerability index will be calculated as a function of exposure (gain/loss in suitable habitat) and sensitivity (measures of abundance, potential, and food availability) (Stortini et al., 2015, 2017), based on projections of change in suitable habitat (Shackell et al., 2014). The objective of this paper is to integrate relevant information at the spatial scale of a stock. The result is a Climate Change Impact Assessment by Fishery Management Units (**Figure 2**), and this highlights spatially distinctive characteristics in vulnerability and identifies needs for customized adaptation planning (Colburn et al., 2016).

## Coastal Infrastructure Vulnerability Index

The Coastal Infrastructure Vulnerability Index (CIVI) was developed as a national-scale adaptation tool for SCH to provide a numerical indicator of relative vulnerability that incorporated the effects of climate change (Greenan et al., 2018a). This vulnerability index was designed with three component subindices: Exposure (natural forces), Infrastructure, and Socioeconomics. Each of the sub-indices incorporates three to five component variables which were scored on a 1–5 scale (least vulnerable to most vulnerable) depending on the harbor's vulnerability to that particular variable. The scoring is a relative measure for the variable over the geographical area of this study. Most of the variables are scored objectively using the methodology described in this section, however, some variables required expert judgment and, in those cases, we have provided information on who undertook this assessment. A detailed description of the criteria for scoring each of the variables in CIVI is provided in Greenan et al. (2018a). The individual sub-index scores were calculated as the geometric mean of the constituent variables (Cogswell et al., 2018) and the final vulnerability index is the geometric mean of the three sub-indices for each harbor.

## Exposure Sub-Index (ESI)

The Exposure Sub-Index includes five component variables: relative sea level change, maximum wind speed, mean significant wave height, coastal materials (shoreline type/susceptibility to erosion), and change in sea ice duration.

## **Sea level change**

For the Atlantic Canada region considered in this study, relative sea level is rising (and is projected to continue to rise) faster

than the global rate, in part as a result of land subsidence due to glacial isostatic adjustment (Greenan et al., 2018b). The relative sea level change data were derived from the Intergovernmental Panel on Climate Change Fifth Assessment Report Coupled Model Intercomparison Project, Phase 5 (IPCC, 2014). Using the IPCC RCP8.5 scenario, the change in relative sea level at each SCH location was computed as the difference between the projected mean water level in the year 2100 relative to 2010. The RCP8.5 scenario was used because it represents the high emission scenario (i.e., business as usual, no mitigation of greenhouse gases) and so for planning purposes is a more conservative option. This scenario is also consistent with the ocean climate model projections that were used for the lobster vulnerability analysis. The scoring of this variable is 1 (5) in the locations of the study with the smallest (largest) projected relative sea level rise.

## **Wind and wave climate**

A wind and wave climate for the Canadian coastline was generated for the years 1990–2012 using wave model data generated from the French Research Institute for Exploitation of the Sea (IFREMER) wave hindcasts using the WAVEWATCH III model (Rascle et al., 2008; Rascle and Ardhuin, 2013) and wind data from the National Centers for Environmental Prediction (NCEP) Climate Forecast System Reanalysis (CFSR) (Saha et al., 2010). Two high resolution (10 min) grids of Atlantic and Pacific maximum modeled wind speeds were used for southern Canadian coastal areas while a coarser (30 min) worldwide grid was used for the Arctic areas. From these datasets the mean annual maximum wind speed and the mean maximum significant wave height were calculated. Future projections of changes in the wind and wave climate along Canada's coastline have low confidence at this point (Greenan et al., 2018b), so the decision was made to use the present wind and wave climate. The scoring of these variables is 1 (5) in the areas with the areas of the study with the lowest (highest) wind speed and smallest (largest) wave heights.

## **Sea ice**

Sea ice data were acquired from the Canadian Ice Service providing percent ice coverage for each week over four decades (1970s through 2000s). For each decade, a single dataset was calculated as the sum of all weeks with ice coverage in excess of 50% along the coastline, with a maximum possible score of 52 weeks for each decade. As a measure of change in ice duration (number of weeks), the data from the 2000s was subtracted from the 1970s. A positive number represented a reduction in weeks of ice coverage, a negative number an increase in ice coverage. This variable is scored one (5) for the locations in the study area that have experienced the smallest (largest) absolute change in the number of weeks of ice coverage.

## **Coastal materials**

The base layers from which the coastal materials layer was derived were the Fulton surficial geology (Fulton, 1995) and the Wheeler bedrock geology (Wheeler et al., 1996), both at scales of 1:25 million. At locations where the surficial geology was greater in thickness than veneer, a score of 3–5 was assigned, with 5 being most erodible (muds, marine clay, materials that will flow) and three being less erodible (sands, gravels). Where surficial materials had the same thickness as veneer or less, the bedrock geology was used as the basis for the score. Scores based on bedrock geology were assigned 2 if the geology was sedimentary, and 1 if igneous or metamorphic (G. Manson, personal communication, 2015, Natural Resources Canada (NRCAN), Bedford Institute of Oceanography, Dartmouth, NS, B2Y 4A2).

#### Infrastructure Sub-Index (ISI)

A harbor's vulnerability to the potential impacts of climate change depends in large part on their ability to physically withstand the forces associated with these impacts. Three variables were selected for the Infrastructure Sub-Index:

#### **Harbor condition**

This variable is based on the SCH engineering evaluation of harbor condition, which incorporates each separate harbor facility at the component level (e.g., breakwater, wharf, building) and assigns a numerical score between 1 and 5 for vulnerability. This index is a weighted average of the score for the individual components. For each harbor, the average of all the individual facility conditions are weighted against the harbor's replacement cost. The scoring of this variable was based on the expert judgment of SCH Regional Engineers responsible for maintenance of these facilities.

#### **Degree of facility protection**

The degree of facility protection is an assessment of the degree to which a harbor is naturally protected or has manufactured protection from storm surge, wind, and other natural forces. The elements included in this assessment were: the basin, wharves, floats, shore protection, slipways, and breakwaters, but not buildings, roads, or parking lots. The variable is a function of the presence or absence of protective assets (such as breakwaters or natural topographical features) and their orientation (i.e., positioned such to withstand primary wave direction). This value was assigned by SCH Regional Engineers with a low score (1) for a fully enclosed harbor and a high score (5) for a completely exposed harbor.

#### **Total replacement cost**

The value of infrastructure at a harbor can itself be an indication of a harbor's vulnerability to the impacts of climate change. The larger the asset holdings at a harbor, the greater the opportunity for financial losses in the event of a major weather event associated with the exposure-related impacts of climate change. While the degree of facility protection provides an indication of the harbor's ability to protect users and infrastructure from the impacts of climate change, harbor facility replacement cost provides an indication of the potential liabilities related to major financial losses. This variable is scored 1 (5) for the harbor infrastructure with the lowest (highest) replacement cost in the region of this study.

#### Socio-Economic Sub-Index (SESI)

A Socio-Economic Sub-Index was developed to assess the harbors' economic vulnerability, as well as the harbors' role within the local economy. The Socio-Economic Sub-Index is comprised of the following four variables:

#### **Quantity landed per vessel**

The quantity landed per vessel is derived from landings (kg) divided by the number of vessel per harbor. The landings data were retrieved from DFOs Integrated Catch and Effort System. The landings reported here are the aggregated record of landings of all species returned to a particular Small Craft Harbor over the period of 2009–2013.

The number of vessels served by each harbor as port of landing is an estimate of vessel activity by harbor for the purposes of landing harvest from the fisheries. This variable is different from the number of vessels using a specific harbor as home port. Vessel activity of a Small Craft Harbor as the port of landing was used in lieu of harbor as home port as it more reliably captures the presence of economic activity at each harbor location. The number of active vessels that landed harvest in each SCH location was obtained from DFOs Integrated Catch and Effort System. We assume that landings at a particular harbor are associated with the LFA closest to that harbor. This is not necessarily the case, but for this analysis it is a reasonable assumption. This variable is scored 1 (5) for the SCH locations in the study area they have the highest (lowest) quantity landed per vessel.

#### **Percent income from fishing**

Fishing income is aggregated at the level of the census subdivision by the Canada Revenue Agency from the reported income of the following four fishing-related sectors: self-employed fish harvesters, wage earning-fish harvesters, fish processing employees, and aquaculture employees. Only individuals who reported a positive amount of income in any of these fishing sectors were included in the analysis. All other employment income is considered non-fishing income. The average fishing income by the census subdivision of each Small Craft Harbor (2009–2013) was calculated as a percentage of total average employment income. This socio-economic variable provides an indicator of the weight of fishing-related incomes in the census subdivision associated with the SCH location. This variable is scored 1 (5) for the community surrounding each SCH facility with the lowest (highest) percent income from fishing in the study area.

### **Population**

The spatial area of Statistics Canada (StatsCan) census subdivisions are too large to properly represent populations associated with many SCH sites. Hence, the population linked to SCH sites was assigned using the smaller StatsCan geographical units of Population Centers, Dissemination Areas, and Dissemination Blocks (largest to smallest). The process used for assigning population to SCHs is as follows:

1. If a harbor falls in a Population Center (or within a distance of 10 km), the population for that Population Center will be assigned to the harbor (Population Center is a StatsCan delineation used for municipal areas).


This variable is scored 1 (5) for the communities surrounding the SCH facilities with the smallest (largest) populations in the study area.

### **Species value diversity**

We refer to the diversity of fishing revenue as an Species Value Diversity (SVD). We recognize that fishing revenue is only one aspect of economic diversity in coastal communities, but it does provide a measure of whether communities depend on the revenue from a few or several species. For this analysis, Pielou's evenness index (Pielou, 1996) was used to compute the SVD where values range from 0 to 1, with 1 representing a community with similar sized proportions of landed value (\$CAD) of each species. While measures of diversity typically use species counts for their calculation, we opted to use landed value for each species at each harbor as a proxy for species counts. These values were then reclassified into five equal intervals and redistributed to a range of 1–5 as has been done with other CIVI variables. The outcome of this calculation is that the SVD directly represents the capacity of fishing communities to adapt to climate change impacts should they result in the failure of a commercial fishery.

## Lobster Vulnerability Index Input: Habitat and Zooplankton

The majority of lobster is fished in inshore areas, where inshore is defined as waters up to 12 nautical miles from shore. The only available data for modeling lobster habitat in the region of this study is derived from regular research vessel (RV) surveys, which do not sample inshore areas (see delineation on **Figure 1**). Without inshore data, it is not possible directly predict the inshore habitat suitability for lobster on the Scotian Shelf, but it is important to note that the RV offshore survey often serves as a proxy for lobster regional population dynamics and habitat preferences (Cook et al., 2017). In LFAs 27, 31A, 31B, and 32 the lobster fisheries are almost completely inshore, and there are too many uncertainties and insufficient information in trawl data to make reliable predictions in the offshore in these LFAs. In this study, a decision was made to limit the analysis to projections in offshore lobster habitat only. It will not be possible to draw conclusions about the inshore areas on the Scotian Shelf until geo-referenced lobster abundance data become available.

## Research Vessel Survey Data

Since the early 1960s and 1970s, respectively, the U.S. National Oceanographic and Atmospheric Administration (NOAA) National Marine Fisheries Service and Fisheries and Oceans Canada (DFO) have conducted seasonal and/or annual research vessel (RV) bottom- trawl surveys to monitor the distribution and abundance of groundfish in Northwest Atlantic waters. These surveys range from Newfoundland and Labrador to the Gulf of Maine, and include the Gulf of St. Lawrence and the Bay of Fundy. For each trawl set, species presence, abundance, and biomass are recorded along with several ocean variables (temperature, salinity, depth, date, geographic coordinates). To inform our species distribution model on American lobster habitat preferences, we used presence, depth, bottom temperature, season, and location data from a subset of the RV survey data that includes data from: 1990 to 2016, depths shallower than 450 m, temperatures below 19.5◦C, and winter and summer months (January, February, March, July, August, September).

## Ocean Model Projections

In this study, we model offshore habitat suitability and measure exposure to climate change as the percent change in suitable habitat availability given a projected increase in ocean bottom temperature from computer models for the months with RV survey data. The two ocean models used in this study included a regional ocean model that has high resolution in the region of the Scotian Shelf and Gulf of Maine [BIO North Atlantic Model (BNAM); (Brickman et al., 2016)], and a global climate model [NOAA Geophysical Fluid Dynamics Laboratory's Climate Model 2.6 (CM2.6); (Saba et al., 2016)].

The BNAM model is run in two basic modes: a hindcast simulation from 1990–present, and future climate projections. Output from the former has been applied to a number of ecosystem related and ocean variability studies in Atlantic Canadian waters (Brickman et al., 2015, 2018; Wang et al., 2016; Shackell et al., 2019). The future climate predictions have been used in studies of impacts on the marine ecosystem (Lowen and DiBacco, 2017; Stanley et al., 2018).

The BNAM provides simulated projections of ocean variables for two future climate bi-decadal periods (2055: 2046–2065, 2075: 2066–2085) using two IPCC AR5 scenarios, RCP4.5 and RCP8.5. In this study, we use the bottom temperature output from the 2055 RCP8.5 model run as it provides the scenario most similar to the CM2.6 doubled CO<sup>2</sup> experiment (Meinshausen et al., 2011). The BNAM model simulated the present day climate using the Coordinated Ocean-ice Reference Experiments (CORE) atmospheric forcing dataset (Griffies et al., 2009). Forcing for the future climate simulations was created by adding anomalies derived from six Coupled Model Intercomparison Project Phase 5 (CMIP5) Earth System Models to the present climate CORE forcing for RCP8.5. A representation of the projected Greenland glacier ice melt was also included in the simulations (Brickman et al., 2016). From these simulations, spatial fields representing predictions of mean monthly differences in temperature between the 2055 period and present climate were derived.

The CM2.6 model projected ocean temperature change in response to a 1% per year increase in atmospheric carbon dioxide (CO2) concentrations (Saba et al., 2016). This model was initialized using present day ocean conditions, a 100 year spin up under pre-industrial (1860) atmospheric CO<sup>2</sup> concentration levels (Saba et al., 2016). After the initialization period, atmospheric CO<sup>2</sup> concentrations were increased by 1% each year for 70 years (with CO<sup>2</sup> doubling at 70 years and then remains fixed for 10 more years). A parallel control simulation was run for which after the initialization, CO<sup>2</sup> concentrations were fixed to pre-industrial (1860) levels for 80 years. To calculate the change in temperature, control simulation years were subtracted from the corresponding projection months years, producing 80 years of monthly temperature change projections. In our analysis, we used the average of years 61–80 as our temperature change values.

### Plankton Monitoring Data

The American Lobster has a complex lifecycle throughout which their diet and habitat needs change. During their planktonic stage, they inhabit near-surface waters where they complete three molts by feeding mostly on other planktonic organisms (DFO, 2009). C. finmarchicus is biomass dominant in the zooplankton assemblages in this region of the northwest Atlantic (Reed et al., 2018). Since phytoplankton are the primary source of nutrients for C. finmarchicus, their lifecycle is highly dependent on the timing and magnitude of seasonal blooms. Therefore, shifts in the spatial and temporal components of the bloom are highly relevant to C. finmarchicus distributions and abundances (Record et al., 2019a,b; Staudinger et al., 2019). Shifts in C. finmarchicus abundance can also lead to timing mismatches for migratory species and reduced availability of anticipated food (Record et al., 2019b; Staudinger et al., 2019). In recent years, this has been observed among the North Atlantic right whale (Eubalaena glacialis) whose diet consists primarily of copepods (Brennan et al., 2019) and is correlated with fish and lobster recruitment (Perretti et al., 2017; Carloni et al., 2018).

To incorporate biological interactions and prey availability into the assessment of stock status, C. finmarchicus abundance data were compiled from the DFO Atlantic Zone Monitoring Program (AZMP), and the spatial abundance was simulated by a coupled bio-physical model (Brennan et al., 2019). The AZMP characterizes oceanographic variability through measurements of temperature, salinity, nutrients, chlorophyll, and zooplankton. Seasonal opportunistic sampling occurs along set sections of the Scotian Shelf, but bi-weekly sampling also occurs at easily accessible fixed sites, and also incorporates ocean data collected by other monitoring programs such as fisheries surveys. Zooplankton samples are collected through ring net vertical tows from near-bottom to surface waters. These samples are split in half, one to determine wet–dry weight, the other is subsampled to identify and count taxa. To represent early life stage food availability in this analysis, we use AZMP C. finmarchicus Scotian Shelf data (individuals per m2) collected from consistently sampled months (April, September, and October) between the years 1999 and 2018.

## Lobster Vulnerability Index Formulation

In the context of climate change, the vulnerability of a species balances on the extent of the exposure, the population's sensitivity to this exposure, and its ability to adapt to the associated change (IPCC, 2014). In vulnerability assessments, exposure, sensitivity, and adaptive capacity are quantified using combinations of multiple variables, with specific application to human systems (IPCC, 2014). In the analysis for this study, sensitivity and adaptive capacity have been grouped into a single category "stock status," which is typical for vulnerability assessments for natural systems (Stortini et al., 2015; Hare et al., 2016).

The Lobster Vulnerability Index formula consisted of two sub-indices: exposure and stock status. Exposure consisted of the percent change in suitable habitat for American lobster in response to projected changes in bottom temperature from the two ocean climate model projections, and stock status was comprised of four component variables: potential suitable habitat, occupancy, abundance status, and early life stage food availability. The gain/loss of suitable habitat per lobster fishing area (LFA) is computed and then combined with stock status within each LFA to arrive at a vulnerability index that represents regional vulnerability to climate-related changes in American lobster habitat. The Lobster Vulnerability Index per LFA will be analyzed in combination with Small Craft Harbor vulnerability indices to characterize an LFA's overall vulnerability to climate change.

#### Habitat Exposure

For this study, exposure is defined as the percent change in suitable habitat for American lobster between current and projected temperature scenarios. Note that this is different from the Exposure sub index used in CIVI that represents natural forces that can act on infrastructure. Using the mgcv package (Wood, 2004, 2017) available for R statistical programming, a generalized additive model (GAM) was used to measure habitat suitability across our study region. On a scale from 0 to 1, the habitat suitability model predicts the likelihood of a species being present at a location, based on the distributions of known presence and absence locations across the range of environmental conditions. The GAM was selected using a forward stepwise procedure, and comparing the Akaike information criterion (AIC) and Analysis of Variance (ANOVA) diagnostics (R Core Team., 2018). The model with the lowest AIC, lowest residual deviance and highest deviance explained, served as the null model before adding a new variable. The final model incorporated bottom temperature and depth as thin plate splines, longitude and latitude as a tensor product with an exponential spatial correlation, and season and year as factors.

To quantify exposure per LFA, three versions of the same model were run, one for each temperature scenario (current: associated with the RV survey data, future: BNAM, and CM2.6) while holding all other variables constant. For each temperature scenario, the model was run 100 times using a random selection of 85% of the data, habitat suitability was assigned back to each survey point, and to visually represent the results a habitat suitability map was interpolated using inverse distance weighting. To calculate exposure, for each temperature scenario, survey points were grouped by LFA and then the percentage of the data that received a habitat suitability score >0.3 was calculated (Cook et al., 2017). The percent gain/loss was computed by differencing the current estimate with each of the projected temperature change scenarios.

#### Stock Status

Stock status was calculated as a function of potential suitable habitat, occupancy, abundance status, and early life stage food availability. We measured potential suitable habitat as the percent of the total surveyed LFA area that is suitable (area >0.3 suitability/total surveyed LFA area) when habitat suitability is modeled excluding location. For occupancy, the realized suitable habitat was calculated the same way, but included longitude and latitude in the model, and then divided realized suitable habitat by potential suitable habitat to arrive at an occupancy ratio. Low numbers suggest that there is a lot of unused suitable habitat while high numbers suggest that high occupancy in suitable areas is leading to population growth in less suitable areas. Finally, for abundance status we computed the 5 year (2013–2016, inclusive) mean weight of landings (a proxy abundance index) and divided this by the maximum of the time series. The resulting value represented recent landings as a percent of the highest in the time series, if landings are currently peaking, this value would be 100. There was an exception for LFAs 37, 40, and 41 where we used the RV survey values for landings estimates because LFA 37 is a shared area between LFA 36 and 38. LFA 40 is closed to lobster fisheries, and LFA 41 is a total allowable catch controlled lobster area.

Finally, for early life stage food availability, we combined two values: mean C. finmarchicus abundance per LFA, and the trend in abundance. The trend was defined as the percent change in mean abundance (individuals per m<sup>2</sup> ) in AZMP surveys between the periods of 1999–2009 and 2010–2018 using only data from consistently sampled months (April, September, and October). LFAs that did not contain survey stations were assigned values from the nearest neighboring LFA. Data were divided into two time periods corresponding to the shift circa 2010 in right whale migration patterns (Brennan et al., 2019). As Calanus spp. are the main prey of right whales, this likely represents an upper trophic level response to the documented changes circa 2010 in the zooplankton community in the northwest Atlantic shelf (Devine et al., 2017; Johnson et al., 2018; Meyer-Gutbrod et al., 2018). Mean abundance per LFA was calculated from a coupled bio-physical model prediction of the spatial distribution of average May-November 2008 C. finmarchicus late stage abundance (copepodite stages CV and CVI) on the eastern Canadian shelf (Brennan et al., 2019).

#### Scoring Matrix

We factored the sub-indices and component variables into 1–5 scores using intervals that were based on the potential range for each variable (**Table 1**). The final stock status score was calculated using the geometric mean of the four component variables. The vulnerability assessment used a 5 × 5 matrix to describe the relation between exposure and stock status, and assigned vulnerability scores to American lobster per LFA (**Table 2**). In the TABLE 1 | Definitions of the bins used to factor indicators into scores that range from 1 to 5 (SH, suitable habitat).


TABLE 2 | Vulnerability assessment scoring matrix.


*Rows represent exposure scores (gain/loss in suitable habitat availability), and columns represent stock status scores. White represents neutral vulnerability, darkening shades of red represent increasing vulnerability, and darkening shades of blue represent decreasing vulnerability.*

matrix, exposure ranged from maximum gain (1) to maximum loss (5) in suitable habitat, and stock status ranged from strong (1) to weak (5).

If exposure is low and stock status is strong, vulnerability will be low, and if exposure is high and stock status is weak, vulnerability will be high. However, some combinations of exposure and stock status can be assigned similar vulnerability scores. An area with high exposure and weak stock status will be scored similarly to a place with low exposure and strong stock status. Because of this, mid-ranged vulnerability scores will require deeper analysis should a more detailed understanding be required. Additionally, when suitable habitat is gained in areas where it is currently relatively scarce, a small absolute gain can lead to a large percent gain, while large spatial gains in areas with high existing availability of suitable habitat will have more muted percent gains. Similarly, when starting points for stock status indicators are low, this can be interpreted as a population under stress, or the absence of an established population (which we parallel with uncertainty). We assessed that in both of these scenarios, weak stock status scores are justified, and for interpretation, the nature of the stock status can be deduced from the known footprint of the fishery and population.

## RESULTS

## Future Climate Simulations From Two Ocean Models

The future climate simulations from the two ocean models have been used to produce projections of lobster vulnerability in Atlantic Canada. These species distribution models projected similar spatial patterns of habitat suitability for American lobster. This is encouraging given the different methods the two models use to simulate future climates, and supports the use of CM2.6—a global climate model–in similar studies across pan-Canadian waters and elsewhere. The similarity in spatial pattern can be attributed to the high resolution of the models—something that is required to correctly simulate ocean variability and circulation in this region (Brickman et al., 2018). The difference in intensity of the species distribution model responses can be directly related to larger bottom temperature changes predicted by the CM2.6 model (see **Figure S1**).

## Lobster Response to Projected Climate Change

Current habitat suitability is higher and more widespread in the western waters of Nova Scotia, including the Bay of Fundy and Browns Bank, and is moderate throughout the Gulf of Maine and on the outer part of the Scotian Shelf from ∼62 to ∼65◦ W (**Figure 3**). By mid-century, both future climate model scenarios project an increase and expansion of habitat suitability (**Figure 3**). While the spatial pattern of change in suitability is similar using both ocean climate models, the suitability trend was more pronounced under the CM2.6 temperature change projections.

The absolute and percent change in projected habitat suitability between the current conditions and the two future scenarios demonstrates similar increases in most LFAs (**Figure 4**). For both scenarios, despite suitability remaining high, there is a decrease in suitability in LFA 35 (Bay of Fundy), suggesting that these waters will begin to warm beyond optimal temperatures by mid-century (**Figure 4**). For both scenarios, habitat suitability in LFAs 34, 38, and 40 increases considerably. In general, the projected change is larger for CM2.6 than BNAM, with a larger range of both increase and decrease in habitat suitability in some areas including LFAs 33 and 41 (**Figure 4**). However, for LFAs 41 and 33, the gains outweigh the loss, which leads to a net gain in habitat suitability.

The percent increase in suitable habitat per LFA (exposure) for the two future scenarios yielded similar patterns, with CM2.6 projecting consistently higher gains (**Figure 5**). LFA 33 is projected to experience a large percent increase, more than doubling the habitat suitability in the CM2.6 scenario and almost doubling in the BNAM scenario. This is because the current habitat suitability in this LFA is quite low, and the increase in habitat suitability, although relatively small compared to other

FIGURE 3 | Current and projected habitat suitability maps for American lobster *(Homarus americanus*), based on measured and projected ocean temperatures. Habitat suitability is measures on a 0–1 scale, where 0 = low likelihood of the species being present and 1 = a high likelihood of species presence. (A) Reference map with Lobster Fishing Areas (LFAs) labeled. (B) Current prediction for habitat suitability distribution modeled using temperature, depth, season, year, and location from research vessel (RV) survey data. (C) Habitat suitability modeled using the same formula as the current prediction, but projected temperatures from the BNAM ocean forecast model instead of measured temperatures. (D) Habitat suitability modeled using the same formula as the current prediction, but projected temperatures from the CM2.6 ocean forecast model instead of measured temperatures. The black dashed line along the coast of Nova Scotia delineates the extent of the inshore fishery where no data are available.

LFAs (34, 38, and 40), is large relative to its starting point (**Figure 5**). Percent gains in LFAs 34, 38, 40, and 41 were considerable, and negligible in LFAs 35, 36, and 37. LFA 33 is located between LFAs 34 (with considerable gain) and 32 (presently with no offshore lobster fishery). The substantial gain in LFA 33, may represent the edge of a range shift, where habitat suitability is increasing northeast as temperatures warm.

Using the scoring matrix described in Section Scoring matrix, we assigned vulnerability scores to each LFA. For both projected temperature change scenarios, the vulnerability scores were

by 3/2 times, whiskers (vertical lines) represent the maximum and minimum scores (excluding outliers). Boxplots represent distributions of Exposure Sub-Index (ESI), Infrastructure Sub-Index (ISI), and Socio-Economic Sub-Index (SESI) scores among Small Craft Harbors within each LFA. Background shading represents the Lobster Vulnerability Index (LVI) per LFA. There is no LVI for LFAs 27, 31A, 31B, and 32 because there are no offshore lobster fisheries in these regions.

similar in each LFA, except in the case of LFA 41. Values ranged from 2 to 2.5, with LFAs 33, 34, and 38 scoring the lowest (2) and LFAs 35 and 36 the highest (2.5) (**Figure 6**). LFA 41 scored 2.5 under the BNAM scenario and 2 under the CM2.6 scenario. None of the LFAs were predicted to experience a net loss of

suitable habitat, although in LFA 35, the gain only marginally outweighs the loss. LFAs with higher scores, were typically unlikely to see any significant gain in habitat suitability, and their stock status was scored lower due to component variables (specifically occupancy and potential). LFAs that scored 2, did so due to a combination of reasons, some were predicted to experience large gains so had very low exposure scores, while others had very strong stock status scores, with high potential, occupancy, abundance, and/or food availability.

## Integrated Climate Change Impact Assessment

The primary objective of this paper is to attempt to integrate future projections of lobster vulnerability by LFA with assessments of the climate change vulnerabilities for the respective fishing communities (**Figure 2**). This could provide a tool for fisheries resource managers to inform decisions to incorporate climate change considerations and adaptation plans into management decisions. The results of this assessment are presented in **Figure 6**.

The present-day economic dependency on fisheries, as indexed by the SESI, varies across the region with LFA 32 being the least dependent on fisheries and 31B, 34, and 38 being the most dependent (**Figure 1**). The reason that the LFA 32 SESI is low is that this area is very close to Halifax (the largest city in Nova Scotia) and, therefore, the percentage of income derived from fishing is relatively low. The proximity to Halifax may mask the importance of the lobster industry in the small coastal communities in this LFA. LFAs 31B, 34, and 38, are comprised of many communities with small populations which are highly dependent economically on the lobster fishery (**Figure 1**, **Figure S2**).

Two areas with relatively high total value from landings are LFA 34 and LFA 27 (**Figure 1**). Although lobster vulnerability is not indexed for LFA 27, this region relies heavily on fisheries with both inshore lobster and snow crab contributing heavily (**Figure 1**). LFA 34 has a very low LVI with lobster accounting for more than three quarters of the landed value from all fisheries (**Figures 1**, **6**). To illustrate the contrast in more detail, the frequency distributions of CIVI sub-index scores in LFAs 27 and 34 are compared with the overall distribution for the region (**Figure 7**).

The environmental forcing resulting from climate change (represented by the ESI) indicates that the distribution of vulnerability scores across the harbors in LFA 34 is primarily in the moderate range with no site assessed above 3. In LFA 27, the range of the ESI is much broader and skewed toward the high end of the distribution when compared to both LFA 34 and the overall distribution. The distribution of the ISI is similar for LFA 27 and 34, and consistent with the overall distribution for the region. There is a strong contrast in the SESI scores for LFA 27 and 34 indicating that these regions have much different economic dependencies on the fishing industry. For LFA 34, the SESI is highly right skewed indicating that a large percentage of the communities in this area are predominantly dependent on fishing income and likely have low adaptive capacity. For LFA 27, the SESI distribution is relatively flat with equivalent numbers of low, medium, and high vulnerability. The contrast in SESI scores LFA 27 and 34 is likely explained by the economic diversity in LFA 27, with the SCH sites in the northern part of the area being remote rural communities with small populations and low

business diversity while the southern part of LFA 27 encompasses the Cape Breton Regional Municipality with a large population base and diversified commercial sector. LFA 27 also has a more diverse fishery with income from the snow crab fishery being almost equivalent to that of the lobster fishery. Nonetheless, the Scotian Shelf represents the southern range of the snow crab fishery; therefore, future warming of the ocean in LFA 27 is likely to change the balance between the snow crab and lobster fisheries

Socio-Economic Sub-Index (SESI).

(Stortini et al., 2015; Zisserson and Cook, 2017). The SESI scores for LFAs 27 and 34 are further compared in **Figure 8** by displaying the four indicator variables that

comprise the index. It is evident from the percent income derived from fishing that LFA 34 is highly dependent on this industry at almost all SCH locations. In LFA 27, the percent income from fishing is split between bins 1 and 4, demonstrating that this LFA has SCH sites that are either highly dependent on fishing, but an equivalent number that are not very dependent. The population indicator is right skewed for LFAs 27 and 34, this matches the overall distribution which reflects the fact that most SCH sites are in small rural communities and this does not vary much across the region. The quantity landed per vessel at SCH locations is slightly left skewed distribution and would seem to imply that this indicator is not the dominant factor in the SESI. Relative to the overall distribution, LFA 34 is slightly right skewed indicating that there is a relatively high amount of economic activity occurring in this region and the opposite is observed for LFA 27. Finally, the SVD for both LFAs 27 and 34 show high dependence on a few species, which is consistent with the overall distribution.

Although LFAs 27, 31A, 31B, and 32 were not assigned LVI scores, these regions maintain active and profitable inshore lobster fisheries. The considerations of their Exposure, Infrastructure, and Socio-Economic Sub-Index scores can help describe the capacity of these regions to adapt should lobster productivity change (**Figure 6**). The ESI and ISI scores are moderate in all of these LFAs, indicating that they are likely to experience some degree of change in environmental conditions, and infrastructure damage/costs. There is a wider range in SESI scores: low in LFA 32, moderate in LFA 27, and on the higher end in LFAs 31B and 31A. The SESI describes the vulnerability of the community and their dependence on the fisheries so a high SESI score combined with a decrease in lobster abundance, would provide reason to look into other confounding factors, and the potential need for adaptation strategies. It is important to note that this model cannot resolve the inshore lobster fishery (which is highly active in these LFAs), and it is possible that the inclusion of the inshore fishing data could provide useful additional information for future development of Lobster Vulnerability Indices in these regions.

## DISCUSSION

A climate change impact assessment by fishery management unit has been presented by integrating coastal and economic vulnerability with projected changes in offshore lobster populations. This assessment projects an overall increase in suitable habitat across the shelf, most notably in the southwest with expansion to the northeast (including LFAs 33, 34, 38, 40, and 41). However, habitat suitability is predicted to decline in some parts of the Bay of Fundy where the ambient bottom temperature is already relatively warmer than elsewhere. Rheuban et al. (2017) performed a similar analysis on the southern New England to the Gulf of Maine range of this stock (directly south of our study region) and projected habitat expansion northward and offshore with a loss of habitat in southern inshore areas where temperatures will begin to exceed the thermal range (New England). Although our study was unable to assess inshore areas, we assume that inshore habitat is not vulnerable to warming in the near-term because this has only been projected for the southern-most part of the range. Using survey data collected in both inshore and offshore in the Gulf of Maine, lobster habitat has been shown to be increasing in both areas (Tanaka et al., 2019a), and is projected to continue to increase under the RCP8.5 scenario (Tanaka et al., 2019b). These results, with patterns similar to our own, highlight how climate change-informed habitat suitability projections can help prepare communities to adapt to potential changes in their fisheries (Rheuban et al., 2017; Tanaka et al., 2019a).

Higher vulnerability in areas with increasing suitable habitat is largely a reflection of either poor coastal infrastructure or socioeconomic factors. This reflects the economic risks associated with a community dependence on revenue from a single fishery or the fishery as a whole. While the fishing industry in DFO Maritimes Region is supported by a diverse range of species, when it comes to value, it is dominated by a few key species (**Figure 1**). Coupled with high socio-economic dependence and a moderate state of infrastructure, the region would be at great risk if it were dependent on a cold-water species. Importantly, the entire province is heavily reliant on lobster, and adaptation planners will have to address this dependency on one species. Here, we've used simple economic indicators but a more in-depth analysis would be informative.

Overall, the ISI scores are similar across LFAs and of moderate vulnerability (**Figure S3**). This suggests that at the scale of the LFAs, Small Craft Harbors are in similar (moderate) condition across the region. The range of harbor conditions is highest in LFA 34, and partially reflects a higher number of SCH sites in that LFA. The assessment of the vulnerability of the coastal infrastructure can be enhanced by considering the ESI, which represents the expected forcing of the physical environment resulting from climate change. The general trend in the ESI is that it increases from west to east. One of the primary drivers of this trend is the projection of a reduced number of weeks of winter sea ice under a warmer climate. In the present climate, landfast ice in harbors provides some protection for infrastructure from large waves in the winter. Sea ice does not form in the harbors in the western part of Nova Scotia and, therefore, no change is expected in the future climate scenarios. While the ESI is lowest in the LFAs in the western part of Nova Scotia, this should not be interpreted as the environmental forcing not being an issue for climate change vulnerability. Indeed, all of this region is expected to experience relative sea level rise at a rate faster than the global average, in part due to land subsidence in southern Atlantic Canada (Greenan et al., 2018b).

In general, our study suggests that offshore lobster is not imminently vulnerable to the projected warming of bottom waters. However, lobster do spend the early part of their life cycle (when mortality rates are highest) in the surface waters (DFO, 2009). Carloni et al. (2018) suggested that changes in zooplankton assemblages due to ocean warming may be transferring up the food web as they found post-larval abundance and lobster recruitment to be correlated with C. finmarchicus abundance and not temperature. To account for bottom up foodweb variability, we included current trends an abundance of C. finmarchicus as a measure of early life cycle food availability in the lobster stock status sub-index. We did not include future projections for C. finmarchicus availability, or indices for other potential prey species including at other lifecycle stages. This could be a beneficial next step should appropriate data become available. There is also some indication that Atlantic Canada is experiencing more frequent and intense heat waves (Oliver et al., 2018). With climate change, these extremes are expected to become more common. This presents a challenge to management because while animals may be able to adapt over the longer term, extremes can present short term disruptions when marine heat waves exceed the animal's thermal tolerance (Mills et al., 2013). For example, in 2011/2012, the commercial snow crab fishery on the Scotian Shelf suffered as a result of a positive temperature event that negatively impacted the snow crab juvenile stages (predominantly), resulting in a temporary decrease in abundance (Zisserson and Cook, 2017).

Rising temperatures and heat waves have been linked to an increase in susceptibility and prevalence of epizootic shell disease that became noticeable among US lobster populations since the mid 1990's (Castro et al., 2006; Glenn and Pugh, 2006). The relation with temperature is complicated, but in general, rising temperatures can reach levels that increase physiological stress and enable bacterial to grow (Smolowitz et al., 2005; Glenn and Pugh, 2006; Tlusty and Metzler, 2012). Epizootic shell disease is most damaging at ∼15◦C (Tlusty and Metzler, 2012), and lobsters begin to physiologically stress at temperatures above 20◦C (Glenn and Pugh, 2006; Fogarty et al., 2007). Spatially, shell disease is most pervasive US waters, increasing in prevalence toward southern extent of the range (Smolowitz et al., 2005; Glenn and Pugh, 2006), where higher temperatures are more consistently sustained. As temperatures continue to rise along the Northeast US and southern Atlantic Canada, we can expect the occurrence of shell disease to continue to spread north. While it is current practice to include this consideration in New England natural mortality rate assessments (Correia et al., 2006), the model projections for changes in ocean temperature on the Scotian Shelf and Bay of Fundy do not suggest that this will be an important issue by mid-century if temperature is a primary driver.

Marine heat waves can also promote shifts in distribution and disrupt the timing of seasonal events, leading to a mismatch between the seasonal/spatial aspects of the targeting fishery (Mills et al., 2013). This was observed among lobster populations in the Gulf of Maine in 2012 when record high temperatures spurred an increase in molting/growth rates alongside an early migration into the inshore fisheries (Mills et al., 2013). This created a longer fishing season with proportionally high landings of new recruits. The combination of these factors makes the population vulnerable to overfishing (Mills et al., 2013). The potential effects of marine heat waves are of specific interest on the Scotian Shelf, because this region has a relatively high proportion of species at the edge of their thermal range and is, therefore, more susceptible to marine heat waves (Smale et al., 2019). Developing an understanding of the potential timing, severity, and effects of marine heat waves on populations of economic importance, is an additional element that should be considered in adaptation planning in fisheries management.

It is important to recognize that temperature change is not the only factor that will impact the spatial development of offshore lobster populations. Other factors include predator and prey species, acidification, environmental degradation, invasive species, and fishing pressure. The abundance of both predator and prey species as well as fishing pressure will factor into lobster condition, mortality, and stock status (Shackell et al., 2014; Le Bris et al., 2018). A population that is shifting due to warming temperatures can be further constrained if key lower trophic level prey species are not available within the new range. This can occur as these short-lived species have a stronger capacity to adapt to change, and thus may not require range shifts to cope with warming temperatures (Friedland et al., 2018). Le Bris et al. (2018) observed that in southwest New England (the southern limit of the range), shell disease has made a significant contribution the collapse of the stock; however, alongside rising temperatures in the Gulf of Maine, fishing practices that protect the more fecund females (and are not practiced in southwest New England), combined with the removal of key predators through other fisheries led to a boom in population. Overall, effective fishery management will require dynamic and regionally tailored planning with constant consideration of border issues and interactions between compounding factors, this can be supported with ongoing monitoring and collaboration with science.

The overall projected changes in offshore lobster habitat for the region are positive. Nonetheless, to promote the prolonged sustainability of this stock these changes should still be considered in resource management. As the population shifts toward the northeast, it will become less abundant around its southwest limit. Through the identification of areas where changes in lobster populations are projected to occur alongside regional vulnerabilities to climate change, the tool presented here can help inform decisions on locations where adaptation planning strategies may need to be developed and or implemented. This could involve preparing a region for changes in potential catch, through adjustments in licensing and quotas, preparing to adapt to a decrease in productivity by encouraging/assisting fishers to diversify (i.e., where they fish, targeted species, to non-fisheries-related income), or supporting projected increases in catch through the investment in upgrades and upkeep of coastal infrastructure where needed Finally, it is important that management anticipate and prepare to adapt to changes as a lag in updates to regulatory constraints (such as licensing and quotas) can lead to an unrepresentative fishing footprint which can lead to overfishing or economic struggles for license holders (Pinsky and Fogarty, 2012; Mills et al., 2013). The LFA borders within the DFO Maritimes Region impose constraints on the fishing industry and present challenges as a warming ocean increases productivity in areas such as LFA 41, which currently has a single license holder.

Climate-induced ocean warming is leading to an accelerated redistribution of marine species catch potential (Cheung et al., 2010; Pinsky and Fogarty, 2012; Poloczanska et al., 2013). Regulations and resources will limit the fishing industry's ability to adapt to changes in stock availability. If fishers are accustomed to long distance fishing trips, they can reroute to follow the stock. Nonetheless, depending on the extent of the range shift, borders may become a limiting factor, and if they do not have the resources to follow range shifts, or if there is a drop in abundance, they will be forced to adapt their practice and target new species which will be accompanied by more overhead costs (Pinsky and Fogarty, 2012). Knowing how animals will shift distribution, and what to do about shifts across management borders, both regional, and international will be critical to their plans (Link et al., 2011). Considerations of climate change are not common in stock assessment models. Tanaka et al. (2019a) combined American lobster recruitment dynamics with spatio-temporal variability in habitat suitability, and demonstrated how the inclusion of dynamic environmental variables can improve the performance of a stock assessment model. Here, we present a tool for anticipating change, that could potentially be incorporated into stock assessment models and help fishers and resources managers with long term planning.

## CONCLUSIONS

The results of this study highlight the importance of coastal adaptation planning, and flexible fisheries management that is capable of making adjustments in a dynamic environment impacted by climate change. This assessment which integrates information on coastal community and lobster vulnerability provides both a framework and information for further indepth analyses by climate change adaptation planners and fishery managers. In the USA, the NOAA Climate Change Science Strategy is being used to guide development of regional action plans. The fact that our study focused on one commercial species in one DFO administrative region points to the need for additional research in this area, as has been highlighted in the adaptation and resilience pillar of the Pan-Canadian Framework on Clear Growth and Climate Change. The integrative approach presented in this paper can be adapted for other species (commercial, depleted, etc.) to help support management and planning decisions.

## DATA AVAILABILITY

The raw data supporting the conclusions of this manuscript will be made available by the authors, without undue reservation, to any qualified researcher.

## AUTHOR CONTRIBUTIONS

BG, NS, KF, PG, and ACog contributed to the development of the concept, analysis of the data, and writing of this paper. DB, ZW, and VS provided the future ocean climate model projections and contributed to writing those sections of the paper related to this. ACoo contributed to the interpretation of lobster thermal habitat model and provided general background on the lobster biology and fishery. CB provided zooplankton data and model analysis and contributed to writing related sections of the paper.

## FUNDING

This research was funded by Fisheries and Oceans Canada through the Aquatic Climate Change Adaptation Services Program (ACCASP).

## ACKNOWLEDGMENTS

The authors would like to thank Dr. Jamie Tam and two reviewers for their generous and useful suggestions on an earlier text that improved the paper.

## REFERENCES


## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00579/full#supplementary-material

Figure S1 | Comparison of BNAM and CM2.6: seasonal changes in bottom temperature from two ocean models. The magnitude of temperature change varies between models, so they have been plotted on different color scales, however, overall patterns are similar. (A) Projected change in summer (July, August, and September) bottom temperature (BNAM). (B) Projected change in summer bottom temperature (CM2.6). (C) Projected change in winter (January, February, and March) bottom temperature (BNAM). (D) Projected change in summer bottom temperature (CM2.6).

Figure S2 | The Coastal Infrastructure Vulnerability Index (CIVI) for Small Craft Harbor locations. The scoring of 1(low vulnerability) to 5 (high vulnerability) has been grouped into three categories and then presented in the pie charts for each Lobster Fishing Area (LFA).

Figure S3 | Sub-indices of the Coastal Infrastructure Vulnerability Index (CIVI) for Small Craft Harbor locations are as follows: (1) Environmental Sub-Index (ESI), (2) Infrastructure Sub-Index (ISI), and (3) Socio-Economic Sub-Index (SESI). The scoring of 1 (low vulnerability) to 5 (high vulnerability) has been grouped into three categories and then presented in the pie charts for each Lobster Fishing Area (LFA).

potential in the global ocean under climate change. Glob. Chang. Biol. 16, 24–35. doi: 10.1111/j.1365-2486.2009.01995.x


americanus, H. Milne Edwards 1837. J. Shellfish Res. 24, 749–756. doi: 10.2983/ 0730-8000(2005)24[749:ADOTPO]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 © 2019 Greenan, Shackell, Ferguson, Greyson, Cogswell, Brickman, Wang, Cook, Brennan and Saba. 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.

# Better Together: The Uses of Ecological and Socio-Economic Indicators With End-to-End Models in Marine Ecosystem Based Management

#### Jamie C. Tam<sup>1</sup> \* † , Gavin Fay<sup>2</sup> and Jason S. Link<sup>1</sup>

<sup>1</sup> NOAA-Fisheries, Woods Hole, MA, United States, <sup>2</sup> Department of Fisheries Oceanography, School for Marine Science and Technology, University of Massachusetts Dartmouth, New Bedford, MA, United States

#### Edited by:

Dorte Krause-Jensen, Aarhus University, Denmark

#### Reviewed by:

Donald F. Boesch, University of Maryland Center for Environmental Science (UMCES), United States Christopher James Brown, Griffith University, Australia

> \*Correspondence: Jamie C. Tam jamiectam.phd@gmail.com

†Present address: Jamie C. Tam, Bedford Institute of Oceanography, Fisheries and Oceans Canada, Dartmouth, NS, Canada

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 28 February 2019 Accepted: 26 August 2019 Published: 18 September 2019

#### Citation:

Tam JC, Fay G and Link JS (2019) Better Together: The Uses of Ecological and Socio-Economic Indicators With End-to-End Models in Marine Ecosystem Based Management. Front. Mar. Sci. 6:560. doi: 10.3389/fmars.2019.00560 Ecological and socio-economic indicators are used as proxies for attributes of ecosystems and human communities, respectively. End-to-end models are used to predict how ecosystems will respond to alternative management actions and changing environmental conditions. Despite the importance of these two tools for Ecosystem-Based Management (EBM), there has been limited integration of ecological indicators directly into end-to-end models; the former are typically calculated post hoc with output from the latter. Here we explore how ecological indicators can be better incorporated into end-to-end models and examine the importance of this union with regards to cumulative impacts and indirect effects, setting management objectives, practical indicator selection, and applications to management. We conclude that the inclusion of ecological indicators in end-to-end models is not only feasible, but provides needed guidance on describing ecosystem status relative to strategic as well as tactical ecosystem-level management goals, and will escalate the implementation of EBM.

Keywords: Atlantis, Ecopath with Ecosim, indicators, emergent properties, ecosystem-based management, cumulative impacts, ecosystem-level reference points

## INTRODUCTION

Marine resource management at the ecosystem level is becoming a useful approach to complement single-species, single-sector, and single-impact assessments (Link et al., 2012; Gaichas et al., 2017; Link and Browman, 2017). Given the theoretical advancements in understanding whole ecosystems, there is a growing global imperative to implement ecosystem-based management (EBM) – a management approach that recognizes the full array of interactions within an ecosystem and accounts for multiple human uses, complexity, cumulative impacts, indirect effects, emergent properties, and tipping points (Christensen et al., 1996; McLeod et al., 2005; Link et al., 2015). The rationale and descriptions thereof are replete (e.g., Leslie and McLeod, 2007; Link and Browman, 2014) and not repeated here; the salient point is how to operationally implement EBM. Current strategic policies are actively aiming to use ecosystem-level science to guide decisions in marine management: e.g., the European Union Marine Strategy Framework Directive (MSFD:

Palialexis et al., 2014), Commission for the Conservation of Antarctic Marine Living Resources (Constable et al., 2000; Constable, 2011), Australia's Ocean Policy (Smith et al., 2007), and National Oceanographic and Atmospheric Administration's Ecosystem-Based Fisheries Management Policy (Noaa-Fisheries, 2016).

The development and uses of both ecological and socioeconomic indicators in current ecosystem-level ocean management have greatly increased over the last decade. The importance of ecological indicators to marine ecosystem management has been made evident in global projects such as IndiSeas (Shin and Shannon, 2010), International Council for the Exploration of the Sea working groups (ICES: Tam et al., 2017b) and the Food and Agriculture Organization of the United Nation's (FAO) approach to sustainable fisheries management (Garcia et al., 2000; Punt et al., 2001). Operationally, ecological indicators have been selected for use directly in marine policy by a number of countries to determine the state of ecosystems (Fulton et al., 2005; Thrush et al., 2011; Levin et al., 2014; Shephard et al., 2015; Jackson et al., 2016). Socio-economic indicators have been used to identify communities vulnerable to fishing collapses and climate change (Pollnac et al., 2015; Colburn et al., 2016). Indicators act as proxies to simplify complicated trends in multiple biological, environmental or anthropogenic variables and are immensely useful to conservation and resource management (Methratta and Link, 2006; Link et al., 2009; Blanchard et al., 2010; Shin and Shannon, 2010; Coll and Lotze, 2016). Ecological indicators can help to reveal overarching patterns in ecosystems, while socio-economic indicators can help to quantitatively define management objectives and determine the achievement of those objectives. By capturing the emergent properties, cumulative impacts, and indirect effects of ecosystems and human communities through indicators (Link et al., 2015; Pollnac et al., 2015) and developing ecosystem-level reference points (ELRPs; Large et al., 2015b; Samhouri et al., 2017a; Tam et al., 2017a), it is possible to avert negative scenarios such as the loss of jobs, overfishing, hypoxia, or stock collapse (Rabalais et al., 2002; Fay et al., 2015).

End-to-end ecosystem models are important tools to collate, understand and predict key features of marine ecosystems (Travers et al., 2007; Fulton, 2010; Rose et al., 2010; Collie et al., 2014; Tittensor et al., 2017; Lotze et al., 2019). The most commonly used marine end-to-end model is Ecopath with Ecosim and Ecospace (Polovina, 1984; Walters et al., 1999; Christensen and Walters, 2004). Ecopath is a mass-balance model of energy flows in an ecosystem, while Ecosim produces time dynamic simulations of the initial Ecopath model, and has been used primarily for fisheries policy exploration. Ecospace allows for the consideration of spatial management by including habitat dependency, migration, and fisheries distributions among other spatially explicit parameters. Large biogeochemical-based models such as Atlantis (Fulton et al., 2011) incorporate human dynamics which have model applications to concentrate on questions spanning all parts of the adaptive management cycle (Jones, 2009). Atlantis connects the biophysical system, human users (primarily industry), monitoring, assessment, management decision processes, and socio-economic drivers of human use and behavior (Fulton et al., 2011). The Atlantis model is capable of addressing policy needs, balancing socio-ecological objectives of human activities and clearly presenting potential trade-offs (Weijerman et al., 2016). Most importantly, Atlantis can test the feasibility of management strategies before they are implemented in reality (Fulton et al., 2011), which is an important step to exploring fisheries, coastal zone, and related ocean-use management actions up to and including full EBM scenarios (e.g., Fulton et al., 2014). Agent-based models such as OSMOSE (Shin and Cury, 2001) and in vitro (McDonald et al., 2006) include individual-based, age-structured fish or predator population and trophic interaction models, biogeochemical plankton production models, hydrodynamic and environmental models, habitat models and representations of human activities. These models use decision algorithms that allow for fluid representation of processes like movement, growth, phenotypic expression, and evolution, making them a useful tool for examining fine scale interactions and responses to the impacts of large scale drivers (Fulton, 2010; Rose et al., 2010).

Many end-to-end models are designed to perform ecosystemlevel scenario analysis and Management Strategy Evaluations (MSEs) where ecosystem dynamics can be explored under a variety of plausible management, climate, oceanographic and human use conditions (Fulton et al., 2014; Masi et al., 2018). The MSE process ultimately aims to explore the results from a set of management strategies to compare how well they meet specified management objectives. Other end-to-end models also exist, but the increasing global use of end-to-end models as a common tool is the direct result of public, scientific, and management concerns and interests in examining system-level or indirect effects in multiple use scenarios. Furthermore, the fast pace with which ecosystem level end-to-end models are evolving is strong indication of their importance in furthering EBM.

The importance of both indicators and end-to-end modeling to the future of EBM is clear, yet there has been limited integration of indicators directly into end-to-end models. Currently, ecological indicators are typically calculated as post hoc analyses from end-to-end model simulations (Coll and Steenbeek, 2017; Masi et al., 2017). In doing so, the ability to use information from indicators on cumulative impacts and indirect effects to adjust management actions is lost and not captured within model simulations. Here we aim to describe how indicators and end-to-end models are mutually beneficial to each other, how they can be better integrated to improve the understanding of ecosystem dynamics and, in turn, facilitate more successful management actions. Ultimately, these two tools together will increase the accuracy of end-to-end model predictions to provide operational management advice at the ecosystem-level.

## VALUE OF INDICATORS FOR END-TO-END MODELS

To execute EBM, ecosystem modeling tools are advisable to collate, synthesize and predict ecosystem dynamics related to cumulative impacts (Korpinen and Andersen, 2016), indirect

effects (Crain et al., 2008), emergent properties (Link et al., 2015) and ELRPs (Tam et al., 2017a). There is widespread empirical support that ecosystem-level emergent properties and reference points can be calculated for multiple marine ecosystems using thresholds from empirically derived ecological indicators (Link et al., 2012, 2015; Large et al., 2015a,b; Samhouri et al., 2017a; Tam et al., 2017a). These key fundamental features have been revealed through ecological indicators, adding to strong theoretical support of global ecosystem patterns. Link et al. (2015) found common sigmoidal cumulative biomass-trophic level curves and "hockey stick" cumulative production-biomass curves across 120 marine ecosystems that can help delineate when marine ecosystems are perturbed or recovered. Large et al. (2015a), Samhouri et al. (2017a) and Tam et al. (2017a) have found common multivariate ELRPs for ecological indicators along both anthropogenic and environmental pressure gradients. This comparative work of ecological indicator ELRPs showed that, generally, total landings above ∼2–4 t km−<sup>2</sup> yielded significant changes in ecosystem state which was consistent with surplus production models for multiple marine ecosystems (Bundy et al., 2012; Link et al., 2012; Tam et al., 2017a). Friedland et al. (2012) and Tam et al. (2017a) found that there was a notable increase in monthly fishery yield and ecosystem shift, respectively, when primary production was above ∼0.7 mg m−<sup>3</sup> .

Many of these ecosystem dynamics can be captured with end-to-end models, and thus can be used to help evaluate the consistency and skill of ecosystem model structures (e.g., do indicators from end-to-end models respond to ecosystem drivers similarly as observed). This often requires post hoc processing. For example, the ECOIND plug-in for Ecopath with Ecosim (Coll and Steenbeek, 2017) can calculate a number of indicators (biomass, catch, trophic, size, and species based), but does so after Ecopath, Ecosim or Ecospace has been run. Masi et al. (2017) used the Gulf of Mexico Atlantis model to define indicators that are sensitive to changes in fishing mortality through post hoc calculations of indicators from model outputs under differing fishing scenarios. Olsen et al. (2018) calculated a suite of indicators from the output of a set of Atlantis models and used them to compare ecosystem responses to fishing, spatial management, and ocean acidification both within and between marine ecosystems. In all these examples, the end-to-end model outputs were handled post hoc, not as an integrated part of the modeling and analytical efforts.

While these applications are useful for isolating indicators that are sensitive to particular pressures (e.g., fishing mortality, pollution, etc.), incorporating calculation of these indicators directly into end-to-end models will better reveal overall ecosystem dynamics. Doing so will better facilitate examination of cumulative effects and elucidate indirect impacts, by capturing unintended consequences, exploring synergistic and antagonistic dynamics, and integrating scales and multiple biological features in an end-to-end model setting. This is because calculating indicators post hoc from model output only allows analysts to view indicator snapshots that might not fully detail the way in which values for indicators change over time and space within the model domain. This is not a problem solely with indicators. Essentially having these indicators embedded as part of endto-end models will enable model users to better account for the second order, non-linear, and indirect effects common in ecosystem models (Kaplan et al., 2010, 2013; Fay et al., 2017). Fay et al. (2017) examined the impacts of ocean acidification on the Atlantis-Northeast US ecosystem model. They determined that impacts to the Northeast US food web extended beyond groups that were thought to be most vulnerable, however, the precise nature of these post hoc analyses were difficult to interpret.

Calculating indicators outside of model simulations captures some of the ecosystem dynamics, but removes the possibility of including feedback loops from indicator values to system dynamics within the model runs, say as the result of management action. Post hoc calculation also negates the ability to track behavioral responses of human activities within the model domain by using ELRPs as part of the model dynamics. Incorporating indicators directly within models will better capture the nuances of these dynamics and allow for the exploration of synergistic management action which has been shown to be more efficient and effective at restoring depleted populations (Crain et al., 2008; Smith et al., 2015; Samhouri et al., 2017b). Embedding indicators as part of end-to-end models allows for "real-time" testing and use of ELRPs for management action. Including these indicators with ELRPs facilitates the ability to test and track their performance, thus advancing the recommended levels and use of ELRPs (Samhouri et al., 2010; Large et al., 2015a,b; Link, 2017; Tam et al., 2017a). This in turn will facilitate the validity and uptake of ecosystem model output by using these standard decision criteria that have been tested and validated.

Indicators also play an important role in evaluating endto-end model skill (i.e., calibration, validation, and how much confidence to have in the model) and performance. There are numerous methods that can be used to assess end-to-end model skill, but most commonly, model parameters are adjusted to plausible levels (changes made within confidence limits of observed monitoring or assessment data) or are matched to estimates from time series data. This is frequently an iterative process that is unique to each model type, but general guidelines and best practices are documented in the literature (Shin and Cury, 2001; Link et al., 2011; Heymans et al., 2016; Steenbeek et al., 2016). Indicators act as a pathway to assess model skill by additionally including emergent properties that reflect the interactions between model components. Incorporating indicators in the initial development of an end-to-end model would ensure that observed emergent properties and dynamics of the ecosystem will be captured. Olsen et al. (2016) found that indicators were an important consideration when assessing model skill because they could examine emergent properties of ecosystems across a range of spatial levels and metrics (i.e., an indicator of broad system properties is total biomass, an indicator of narrower system properties is charismatic megafauna biomass). By using indicators alongside other data sources (i.e., single species and human use metrics) end-to-end models can be evaluated on their utility for making predictions (hindcast or forecast) for whole socio-ecological systems. Having indicators directly in the models further facilitates this skill evaluation.

A consistent challenge in EBM and MSEs is the task of setting and defining management objectives. While ELRPs for ecological indicators quantitatively define tipping points in ecosystems that translate to avoidance points for managers, setting management objectives for human uses of marine ecosystems beyond a fisheries lens can be difficult. International biodiversity targets such as Aichi or the trade restrictions by the Convention on International Trade in Endangered Species of Wild Fauna and Flora can be used to define ocean policy (CITES, 1973; CBD, 2011; Juffe-Bignoli et al., 2016), but considerations for human coastal community health and well-being are seldom considered in management scenarios. This is not surprising, since measures of well-being have been difficult to quantify (e.g., cultural attachment, job satisfaction, health, and safety) even despite the establishment of specific limits on pollutant concentrations in coastal communities. Recently, however, there has been development of socio-economic indicators that track patterns of community vulnerability and well-being, further elucidating some of the complexity of the human dimension of EBM (Bowen and Riley, 2003; Pollnac et al., 2015; Colburn et al., 2016; Auad et al., 2018). These indicators embedded in end-to-end models will add nuance to the more conventional management considerations (e.g., total allowable catches to commercial fisheries, recreational fishing opportunities) by incorporating patterns of human behavior and overall community health. Furthermore, developing ELRPs for socio-economic indicators will help to quantitatively set management objectives. For example, system-level optimal yields can be calculated as ELRPs for socio-economic indicators that define the amount (or ranges) of resource extraction (for fisheries) needed to maintain community health (avoiding, for example, long-term poverty). With simultaneous explorations of both ecological and socio-economic management objectives within end-toend models we can begin to quantitatively assess tradeoff spaces (Rockstrom et al., 2009; Link, 2010; Dearing et al., 2014) that avoid ecological regime shifts (e.g., stock collapses) and undesirable shifts to human coastal communities (e.g., increased outmigration).

## VALUE OF USING END-TO-END MODELS FOR INDICATORS

Ecosystem-Based Management is reliant on the use of ecological indicators to assess ecosystem status. This is evident from the inclusion of ecological indicators in frameworks for a number of EBM programs including the Integrated Ecosystem Assessments (Levin et al., 2009, 2014; Walther and Mollmann, 2014) and the MSFD (Palialexis et al., 2014). Several international efforts have been made to determine a pragmatic set of ecological indicators to assess marine ecosystem status (Fulton et al., 2005; Shin et al., 2010; Tam et al., 2017b; Fu et al., 2019). While much of the development and selection of indicators has been done through time-series and pressure-response relationships with human or environmental pressures (Methratta and Link, 2006; Large et al., 2013), there is much to be gained from using end-to-end models to advance the uses of indicators for ecosystem management.

In many cases, scientists and managers have to work within specified budgets with which to develop research and monitoring to meet specific objectives in EBM (Fulton et al., 2005; Niemeijer and de Groot, 2008). The number of ecological indicators found in the literature to evaluate marine ecosystems can be overwhelming, and determining the "Goldilocks" number of indicators (i.e., not too few, not too many, but just right) can be a challenge. There are numerous methods to reduce the number of indicators needed for practical use in EBM and to limit bias in the representation of an ecosystem attribute (Link et al., 2002; Rice and Rochet, 2005). Tam et al. (2017b) used expert opinion to develop selection criteria to determine a standard set of five food-web indicators from 60 potential indicators. Bundy et al. (2017, 2019) reduced 358 possible indicators used to represent the Scotian Shelf region to a set of 30 indicators through a series of qualitative (selection criteria) and quantitative (redundancy analysis) screening. They also examined these indicators across multiple spatial levels (the strata, regional, and ecosystem level) and determined that different sets of indicators were most effective at detecting changes at each spatial level. To build upon this work, it would be beneficial to adjust endto-end model parameters to identify which suites of indicators best represent impacts to major concerns or priorities (i.e., maintaining fisheries yields, maximizing biodiversity). Model based approaches to selecting indicators and assessing them against known pressures (human activities, climate, etc.) have a substantial advantage over other methods (e.g., expert opinion, time series trends, multivariate dimension reduction, etc.) as they are not as heavily reliant on up-to-date field data and can be more cost effective.

While there is a wide breadth of indicators that have been vetted for current use in EBM, there is a continuous stream of indicators being conceived and developed. Tam et al. (2017b) identified a number of proposed ecosystemlevel food-web indicators that were underdeveloped or lacked necessary data to be considered operational. They suggested that these indicators be re-evaluated for operational use in light of new information. As such, this iterative process to develop indicators is a key step in many EBM frameworks to better understand ecosystem dynamics and ensure that management objectives are met (Levin et al., 2014; Walther and Mollmann, 2014; Queirós et al., 2016). End-to-end models are the perfect platform to test the validity (Which ecosystem attribute is this indicator a proxy for?), sensitivity (What is the capacity of this indicator to detect change in the ecosystem attribute?), and specificity (What is the level of confidence with which the variation of an indicator can be attributed to a particular pressure?) of these un-vetted indicators (Houle et al., 2012; Ortega-Cisneros et al., 2018; Shin et al., 2018). Fulton et al. (2005), Samhouri et al. (2009), Kaplan et al. (2013), Olsen et al. (2018) and Ortega-Cisneros et al. (2018) used endto-end model simulations to examine the impacts of fishing and climate on indicators. In these studies model simulations were projected at different levels of fishing or environmental variability and indicators were calculated from these outputs

and examined against ecosystem attributes (to examine validity and sensitivity) or pressures (specificity). These post hoc analyses of indicators are useful and informative, however, including some of these indicators into existing end-to-end models directly could identify the utility of underdeveloped indicators or to screen for less useful indicators by examining any changes to model performance when adding or subtracting new indicators. This would remove the need for potentially subjective expert opinion from indicator selection processes and allow for more objective, quantitatively based evaluation of indicator performance and selection.

## SUMMARY AND CONCLUSION

Successful EBM requires the ability to account for cumulative effects and indirect impacts of human and environmental pressures at the ecosystem level while also accounting for single sector assessments such as fishing mortality from stock assessments or risk analysis for energy exploration (McLeod et al., 2005; Link, 2010; Stelzenmüller et al., 2018). Indicators and end-to-end models are both ecosystem-level management tools already in use that can account for the complexity of interactions within ecosystems alongside single sector assessments that need to be operationalized for EBM (Fulton et al., 2011; Patrick and Link, 2015; Weijerman et al., 2015; Link and Browman, 2017). Evidence from multispecies models using control rules derived from indicator thresholds suggests these models were able to perform better against catch and biodiversity objectives than when harvests were based solely on single-species advice (Fay et al., 2015; Fulton et al., 2019). Kaplan et al. (2013) found that there was a mix of additive and non-additive impacts to fish in model simulations when using indicators as performance measures. Some fleets had a direct impact on target and Bycatch species without extending to other parts of the food web, while other fleets showed unintended impacts on groups beyond the targeted species. These examples of indicators used with end-to-end models reveal that ecosystemlevel examinations of pressures on systems necessitates dynamic and mixed approaches that cannot be achieved through single sector management alone.

By integrating indicators into end-to-end models, the patterns, properties and impacts on indicators within simulations can be tracked and also used to help make "virtual" management decisions. Fundamental changes to the model structure and behavior would likely occur through the incorporation of indicators to end-to-end models compared to current versions that operate without. For example, including existing indicator time series in the model fitting process (e.g., Scott et al., 2015) or Monte Carlo routine (e.g., Steenbeek et al., 2018) would constrain model outputs, change simulation results, and potentially improve model uncertainty. Much like the integration of harvest control rules for fisheries into end-to-end model MSE simulations, management actions based on systemic properties through indicators can be used in the simulation process to track management actions based on ELRPs (**Figure 1**). Fay et al. (2015) describes how the incorporation of ELRPs can improve model management performance using a multispecies model. Incorporating indicators into more complex, end-toend models would increase the ability to leverage a fuller suite of indicators that span a broader range of objective types. Indicator-based management decisions can be made in "virtual" real-time, thereby better tracking the feedback that such decisions will make on the ecosystem features being monitored and managed. This would give managers and policy makers a tool that incorporates cumulative impacts and indirect effects to fully explore the tradeoffs required to balance the needs of both people and ecosystems, with a better sense of the "non-delayed" (i.e., "real-time") ramifications of such decisions.

There have been numerous advancements in indicator development and end-to-end modeling over the last decade, with an increasing interest from policy makers and stakeholders to move toward EBM (Patrick and Link, 2015). While there has been increasing joint use of indicators with end-to-end models, there has yet to be a true merger of these two tools. We recommend (1) that direct integration of indicators into end-to-end models should be used to improve model

skill and performance, (2) that end-to-end models be used to test un-vetted indicators and to develop suitable indicator suites that effectively represent both ecosystem state and community well-being and (3) that these two tools be used together to develop both strategic and tactical management advice using ELRPs of ecological and socio-economic indicators in addition to testing feasible management strategies. We assert that the benefits of integrating these tools will be greater than the sum of its parts and will further the ability of scientists and managers to implement EBM.

## AUTHOR CONTRIBUTIONS

fmars-06-00560 September 13, 2019 Time: 16:56 # 6

JT, GF, and JL contributed to the ideas in this manuscript. JT drafted the text. GF and JL contributed considerable edits and input during the writing process.

## REFERENCES


## FUNDING

JT was supported by a NOAA Postdoctoral Research Fellowship. GF was supported (in part) by the National Oceanic and Atmospheric Administration (NOAA) through the Cooperative Institute for the North Atlantic Region (CINAR) under Cooperative Agreement NA14OAR4320158.

## ACKNOWLEDGMENTS

The ideas in this manuscript grew out of break-out group discussions at the 1st International Summit on the Atlantis end-to-end ecosystem model. The authors thank the attendees of the Atlantis Summit for providing the genesis of this work.



**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 Tam, Fay and Link. 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.

fmars-06-00560 September 13, 2019 Time: 16:56 # 8

# Cooperative Fisheries Outperform Non-cooperative Ones in the Baltic Sea Under Different Climate Scenarios

Sezgin Tunca<sup>1</sup> \*, Martin Lindegren<sup>2</sup> , Lars Ravn-Jonsen<sup>3</sup> and Marko Lindroos<sup>1</sup>

<sup>1</sup> Department of Economics and Management, University of Helsinki, Helsinki, Finland, <sup>2</sup> Centre for Ocean Life, National Institute of Aquatic Resources, Technical University of Denmark, Lyngby, Denmark, <sup>3</sup> Department of Sociology, Environmental and Business Economics, University of Southern Denmark, Esbjerg, Denmark

#### Edited by:

Isaac C. Kaplan, Northwest Fisheries Science Center (NOAA), United States

## Reviewed by:

Xiutang Yuan, National Marine Environmental Monitoring Center, China Philipp Neubauer, Independent Researcher, Wellington, New Zealand

> \*Correspondence: Sezgin Tunca sezgin.tunca@gmail.com; sezgin.tunca@hotmail.com

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 03 December 2018 Accepted: 20 September 2019 Published: 11 October 2019

#### Citation:

Tunca S, Lindegren M, Ravn-Jonsen L and Lindroos M (2019) Cooperative Fisheries Outperform Non-cooperative Ones in the Baltic Sea Under Different Climate Scenarios. Front. Mar. Sci. 6:622. doi: 10.3389/fmars.2019.00622 Game theory has been an effective tool to generate solutions for decision making in fisheries involving multiple countries and fleets. Here, we use a coupled bio-economic model based on a Baltic Sea dynamic multispecies food web model called BALMAR and, we compare non-cooperative (NC) and cooperative game (grand coalition: GC) solutions. Applications of game theory based on a food web model under climate change have not been studied before and the present study aims to fill this gap in the literature. The study focuses on the effects of climate variability on the biological, harvest and economic output of the game models by examining two different climate scenarios, a first scenario characterized by low temperature and high salinity and a second scenario by high temperature and low salinity. Our results showed that in the first scenario sprat spawning stock biomass (SSB) and harvest dropped dramatically both in the NC and the GC cases whereas, herring and cod SSBs and harvests were higher compared to a base scenario (BS) keeping temperature and salinity at mean historical levels. In the second scenario, the sprat SSB and the harvest was higher for both GC and NC cases while the cod and the herring SSBs and harvests were lower. The total GC payoffs clearly outperformed the NC payoffs across all scenarios. Likewise, the first and second scenario GC payoffs for countries were higher except for Poland. The findings suggested the climate vulnerability of Baltic Sea multi-species fisheries and these results would support future decision-making processes of Baltic Sea fisheries.

Keywords: Baltic Sea, fisheries, game theory, climate change, food web model

## INTRODUCTION

Game theory has been an effective tool to generate solutions for decision making in many fields (e.g., policy making, military methodologies, environmental and natural resource economics and management) (Eatwell et al., 1989). In general, the nature of game theory is highly suited for management problems in fisheries, as the fishers want to increase their economic profits from their activity that generates positive and negative externalities for the resource users and nonusers (Bailey et al., 2010). One of the important management problems around the world is open access use of the fisheries resources without cooperative agreement. Non-cooperation is quite

common among fishing states. In particular, many conflicts arise concerning fishing rights on highly migratory and range shifting species (Pinsky et al., 2018), often as a response to climate change (Perry et al., 2005; Pinsky et al., 2013). Cooperative agreements provide resilience through time so that the states can react flexibly to the impact of unexpected shifts in, e.g., biology, climate, and economy (FAO, 2016). Therefore, there is a clear need to understand and predict the impacts of climate change. Otherwise, disputes over cooperative agreements can be inevitable (Miller et al., 2001; Sissener and Bjørndal, 2005). One instrument to provide flexibility for such conflicts is side payments that prevent the losses generated by the inequality raised by these shifts (induced mainly by climate change) (Miller and Munro, 2004). Previously, single-species game theoretic studies (e.g., Diekert et al., 2010; Bjørndal and Lindroos, 2012; Kulmala et al., 2013) and multispecies game theoretic models have been utilized for various fisheries management issues concerning climatic and other environmental variations in the literature (Nieminen et al., 2012, 2016).

Baltic Sea fisheries constitute prime examples of common pool fisheries managed by European Union Common Fisheries Policy (EU-CFP). The EU fishing nations in the Baltic Sea jointly agree on an annual total allowable catch (TAC) for each commercially important stock. The TACs are shared among participating nations considering the relative stability principle that determines harvest quotas based on historical catch records of the EU member states whereas, the nationwide TAC is shared among fishermen according to country specific rules (Nieminen et al., 2016). Denmark, Poland, and Sweden have been the dominant cod (Gadus morhua callarias) fishing nations in the Baltic Sea for the last two decades. These countries are also actively involved in fishing sprat (Sprattus sprattus) and herring (Clupea harengus membras) that are key prey for the Eastern Baltic cod. During the past decades the Baltic Sea has experienced pronounced changes in hydrographic conditions, notably a marked long-term increase in temperature, decrease in salinity and deep-water oxygen concentrations (Meier, 2006; Neumann, 2010), as well increased eutrophication causing widespread algae blooms (Mackenzie et al., 2007; Markus Meier et al., 2011; Neumann et al., 2012). These abiotic changes have led to large-scale ecosystem changes, i.e., regime shifts, occurring in the late 1980s (Möllmann et al., 2009; Casini et al., 2012; Blenckner et al., 2015) that particularly affected the recruitment of the commercially important species, cod, sprat, and herring (Cardinale et al., 2009; Margonski et al., 2010; Thøgersen et al., 2015).

In the literature there are few applications of game theory in different environmental variation problems. For example, management implications of sprat, herring, and cod in the Eastern Baltic Sea under changing climate scenarios were studied by Thøgersen et al. (2015). The authors represented the bioeconomic output of three management scenarios based on a multi-species multi-fleet bioeconomic model. They concluded that the management plan in practice for cod have negative impact on the cod abundance and on the economic gains of fishermen and, this negativity can be eradicated by a reduction in fishing mortality. In another study, Wang and Ewald (2010) highlighted the positive output for competing species survival in a prey-predator system under cooperative management with climate variation whereas, non-cooperation resulted in stock collapse. Brandt and Kronbak (2010) represented the changes in stability of fishery agreements under different scenarios. The authors investigated the stability of fishery agreements under climate uncertainty based on an age-structured bioeconomic model and concluded that climate change has negative impact on the payoffs by decreasing the likelihood of establishment of stable cooperative agreements. Nieminen et al. (2012) evaluated Baltic sprat, herring, and cod fisheries for changing salinity scenarios including a species interaction function into a bioeconomic model. They found that lower fishing mortality would result in higher economic input whereas, under a high salinity scenario, cod stock achieved better levels of recruitment. Nieminen et al. (2016) investigated a multispecies partition function game among three asymmetric nations bordering the Baltic Sea. They showed that the full cooperation among three nations can be stabilized if the dominant nation compensates the other nations. They also presented higher revenue under cooperation if the cod biomass declined under climate change.

In this study, we applied a novel multi-species game theory approach for Baltic Sea fisheries based on a food web model to investigate and compare non-cooperative (NC) and cooperative game (grand coalition: GC) solutions under different climate scenarios. Our study focus on the effects of climate variability on the biological, harvest, and economic output of the NC and GC approaches by examining two different climate scenarios representing favorable and unfavorable temperature and salinity conditions for the stock status of cod, sprat and herring compared to a base scenario (BS) with climate conditions maintained unchanged at their mean historical levels. To assess the sensitivity of the model we used varying economic parameters including cost, price and discount rate.

## MATERIALS AND METHODS

## Biological Model: Setup and Validation

We established the bioeconomic models based on the previously developed Baltic Sea food-web model called BALMAR (Lindegren et al., 2009; **Supplementary Figure S1**). BALMAR represents the food web dynamics of sprat, herring and cod (the three ecologically and commercially most important species in the Central Baltic Sea) and accounts for their pair-wise species interactions, as well as climate and fishing impacts on their recruitment and survival. This statistical model is based on a theoretical approach for modeling long-term population dynamics (Ives et al., 2003) and is given by:

$$X(t) = BX(t-1) + CU(t-\chi) + E(t) \tag{1}$$

$$Y(t) = ZX(t) + V(t) \tag{2}$$

where X are spawning stock biomass (SSB) values of cod, sprat and herring derived from a multi-species fish stock assessment model (ICES, 1996) at time t and t − 1, respectively, and B is a 3 × 3 matrix of pair-wise species interaction parameters. The covariate vector U contains time-series of estimated mean annual

fishing mortalities (F) and a number of selected environmental variables known to affect recruitment of cod, sprat, and herring, respectively (Köster et al., 2003, 2005; Nissling, 2004; Möllmann et al., 2005; Dickmann et al., 2007). These include time series of summer bottom (80–100 m) salinity, spring surface (0–10 m) temperature and the log(abundance) of the key zooplankton prey (Pseudocalanus acuspes) for herring (Möllmann et al., 2005). The data was provided by the ICES/HELCOM Working Group on Integrated Assessments of the Baltic Sea (ICES, 2008). The effect of the fishing and environmental variables on each species are represented by the diagonal parameters in the matrix (C). Regression parameters were found by maximum likelihood estimation using a Kalman filter (Harvey, 1989). E is the process error, V is the observation error of the covariance matrix of the normal random variable. Y is the true observed state. The fitted model parameters captured accurately the known mechanisms of species interactions (**Supplementary Figure S1**), including density-dependence, competition between sprat and herring and cod predation on both sprat and herring (Köster and Möllmann, 2000; Neuenfeldt and Köster, 2000; Möllmann et al., 2005). Furthermore, the model parameters also illustrate the negative effect of fishing and the positive effect of the environmental variables including temperature, salinity, and zooplankton on sprat, cod, and herring, respectively (MacKenzie and Köster, 2004; Köster et al., 2005; Möllmann et al., 2005).

A number of diagnostics were applied to assess whether the final food-web model (**Supplementary Figure S1**) gave a reasonable description of the food-web dynamics (**Supplementary Figure S2**). The assumption of normality of the error terms is supported by an analysis of the residuals (**Supplementary Figure S3**). A partial autocorrelation analysis of the residuals further indicates that the model errors were independent for all species and lags. Finally, a stability analysis of the final parameters of the community matrix, B reveal a dominant eigenvalue below one (λ1 = 0.93), indicating a stable food-web model dynamic. The predictive capabilities of the food-web model was validated by a sequential refitting procedure where the model was initially fitted to only the first 10 years of the data set and then refitted on a yearly basis, producing a prediction for each consecutive year. The predicted values and associated 95% prediction intervals were compared with the observed values to assess the predictive accuracy of the model. Additionally, the food-web dynamics was simulated using only the first-year values as initial conditions. This procedure is fundamentally different from a simple fit to the data, as the observed values from the second year onward are not used in forward predictions. Simulations were run 1,000 times with random process noise added at each time step. Mean values and a 95% confidence interval of the hindcast predictions were computed. To assess the relative contribution of environmental and species interactions in affecting the food-web dynamics, an additional hindcast simulation was performed using a simpler single-species model fitted only to fishing mortalities and biomasses of each individual species separately. Both the sequential refitting and the simulated dynamics demonstrated a distinct ability to "recreate the past" dynamics of cod, herring and sprat (**Supplementary Figure S4**). The hindcast simulations without accounting for environmental forcing and species interactions, however, did not at all explain nor recreate the past dynamics of the three species, especially in the case of cod and sprat (**Supplementary Figure S1**). Consequently, the food-web model including both species interactions and climate effects was used in the original publication by Lindegren et al. (2009) and in our bio-economic simulations.

## Economic Model

To explore game theoretic scenarios, we investigated two strategic interactions between players, here represented by different fishing fleets (rather than individual vessels) as agents. The first is NC interactions where each fleet take its fishing decision by itself and the second is a fully cooperative (grand coalition: GC) interaction where all fleets cooperate by a binding agreement. Three fishing states, Denmark, Poland, and Sweden were considered where each state has its own pelagic fleet for sprat and herring, as well as a demersal fleet for cod. Hence, a total of six fishing fleets were considered for the models. In the NC games, the six fleets act independently and exploit the sprat, herring, and cod stocks whereas, in the GC game, the fleets act depending on a binding agreement. Additionally, the catch of these three dominant fishing nations amount to 70% of the total catch. Consequently, we focused on these three states as they also historically exploit the resource dominantly.

The economic parameters of the model were obtained from the literature. Following Nieminen et al. (2016), the species' prices, pi,<sup>j</sup> , are constant over time and asymmetrical for the countries. Here, i is country and j is the species, discount rates for each country is constant over time, r<sup>i</sup> , were applied from Nieminen et al. (2016) and c<sup>i</sup> is the cost parameters (constant over time) for each species. In our model, use of dynamic prices and costs would be useful to evaluate our case study closer to the real-life case; however, such dynamic cost and price taking into account stock size are to our knowledge missing for all nations and species except for the Danish cod fishery (Röckmann et al., 2008). So that, we utilized the constant cost and price parameters over time. In our model, the costs are only depended on fishing mortalities (**Table 1**). All models were simulated in R Program (R Core Team, 2019).

Harvest costs were calculated depending on the following equational relationships. Ei,<sup>j</sup> is the effort in number of fishing days, fi,<sup>j</sup> is the fishing mortality per fleet per species and q<sup>j</sup> is the catchability parameter of the species.

$$E\_{i,j(t)} = \frac{f\_{i,j}(t)}{q\_j} \tag{3}$$

Harvest per species and per fleet can be derived by

$$h\_{i,j(t)} = q\_j E\_{i,j(t)} X\_{j(t)} \tag{4}$$

The cost function can be rewritten as

$$C\_{i,j}\left(t\right) = c\_i \times f\_{i,j}\left(t\right) \tag{5}$$

where c<sup>i</sup> is the cost parameter for species. Here, the costs are depended on fishing mortalities and cost parameters as well as effort.


TABLE 1 | The economic parameters used in the coupled bio-economic model in terms of market prices, fishing costs, and catchability coefficients for each country and target species.

<sup>1</sup>Cost Parameters for human consumption for sprat and herring; <sup>2</sup>Catchability coefficient for human consumption herring; <sup>3</sup>Catchability coefficient for fodder herring and sprat.

In the NC case, each country maximizes its long-term profits independently. Term πi,<sup>j</sup> denotes the sum of discounted profits of each country i, from each species j across the years t. The countries maximize their economic benefits according to the formula below and, the profit maximization formula was subjected to the population dynamics explained in the biological part of the model above. We used a closed-loop Nash equilibrium where each player can observe the play of the others in the game.

When i denotes the fleetsi = 1,3 and j denotes for speciesj = 1,3 and the objective function of the NC game for each country, i, is

$$\pi\_{\text{NC}\left(\left\|\right\|\right)} = \max\_{f\_{i,j}} \sum\_{t=1}^{\text{so}} \frac{p\_{i,j}h\_{i,j}\left(t\right) - c\_{i,j}E\_{i,j}(t)}{\left(1+r\right)^{t-1}} \tag{6}$$

The objective function of the GC is maximizing the joint discounted profit across countries and species as follows:

$$\pi\_{GC} = \max\_{f\_{i,j}} \sum\_{t=1}^{80} \sum\_{t=1}^{3} \frac{p\_{i,j} h\_{i,j}(t) - c\_{i,j} E\_{i,j}(t)}{(1+r)^{t-1}} \tag{7}$$

## Climate Scenarios

For both the NC and GC cases, we forced the BALMAR food-web model with three climate scenarios reflecting time periods with different hydrographic conditions highly favorable or unfavorable for recruitment of cod, herring and sprat (**Supplementary Figure S2**), namely: (i) a BS keeping temperature and salinity at mean historical levels (1975–2010), (ii) a first scenario (S1) keeping temperature and salinity at the mean levels observed prior to the regime shift in the 1980s (1975–1980), a period with low temperatures and high salinities favorable for cod and herring, (iii) a second scenario (S2) keeping temperature and salinity at the mean levels observed after the regime shift (1990–1995), a period with high temperatures and low salinities favoring sprat recruitment. Here, we mainly aimed to see how NC or GC behavior was impacted by the changes in climate variables, temperature and salinity by comparing the preand post-regime shift scenario relative to the BS in terms of SSB, yield and revenue. In order to account for ecological uncertainty (arising from the food web model), we performed multiple (N = 100) stochastic simulations for each scenario by introducing multivariate random errors into the food web model for each realization (see Eq. 1). Furthermore, we performed a sensitivity analysis on the economic parameters by varying the discount rates, cost and price for all cases.We decided to leave out some of the results regarding sensitivity tests for the deterministic and stochastic simulations to reduce the length and complexity of the paper. However, we introduced **Supplementary Text** and figures that show these results (**Supplementary Figures S6–S24**).

## RESULTS

## Biological and Harvest Outputs

In the first scenario, sprat SSB dropped dramatically both in the NC and the GC cases whereas, herring and cod stocks were higher compared to the BS (**Figure 1**). Furthermore, the herring SSB is the only one that was greater in the GC compared to the NC. In the second scenario, sprat SSB was higher for both GC and NC cases while cod and the herring were lower compared to the BS (**Figure 1**).

In the first scenario, the NC and the GC harvest changes were highest for Poland, while the smallest changes in the NC and the GC were observed for Sweden. In the first scenario, the GC harvests of Denmark and Sweden were higher than their NC harvests, in contrast to Poland that got significantly higher harvest in the GC compared to the NC. In the second scenario, the GC harvest for Poland solely surpassed the NC harvest. Moreover, second scenario GC harvest changes of Denmark and Sweden stayed below the NC harvest (**Figure 2**).

In the first scenario, sprat harvests declined under the NC and the GC whereas herring and cod harvests increased. The NC sprat and herring harvest changes were less than their GC harvest changes while the GC cod harvest change was considerably higher than the NC cod harvest change. In the second scenario, in general, there were positive change in sprat harvest whereas, herring and cod harvests decreased markedly. Also, as in the first scenario, NC sprat and herring harvest changes were smaller compared to their GC harvest changes (**Figure 3** and **Supplementary Table S2**).

## Economic Output

The total payoffs of the GC clearly outperformed the NC case across all scenarios (**Table 2**). Likewise, the payoffs for each country separately were generally higher under GC, except for Poland that showed a slightly higher net present value under NC compare to GC for scenario 1 and 2 (**Supplementary Table S1**).

If we compare payoffs between climate scenarios, scenario 1 yielded considerably higher total revenues compared to the BS for both NC and GC, while scenario 2 demonstrated considerably lower payoffs (**Figure 4**). For scenario 1, the gains were equally distributed between countries, while for scenario 2 Denmark show considerably lower payoffs compared to Poland and Sweden. In the first scenario, the GC solutions outperformed the NC case only for Denmark, whereas Poland and Sweden did not get higher economic benefits by joining the GC. In the second scenario, GC outperformed NC only for Poland while Denmark and Sweden demonstrated lower payoffs under GC compared to the base scenario (**Supplementary Figure S5**).

TABLE 2 | Country level aggregated net present values (millions €) for the non-cooperative (NC) and the grand coalition (GC) games under the three climate scenarios considered.


Finally, we also assessed the economic performance of NC and GC by sprat-herring and cod fleets separately. For the spratherring fleets, the Danish and the Swedish sprat-herring fleets were negatively impacted under the first scenario for both GC and NC whereas the Polish sprat-herring fleet showed almost equal payoffs compared to the BS under GC and NC (**Figure 5**). In the second scenario, all sprat-herring fleets show higher payoffs compared to the BS. For the Danish and the Swedish spratherring fleet NC payoffs were found higher than the GC payoffs in contrast to the Polish sprat-herring fleet. In general, the total payoffs of the GC exceed the total payoffs of the NC in the first scenario whereas, in the second scenario, the total NC payoff was higher than the GC. For the cod fleets, the NC and GC payoffs of the first scenario were higher compared to the BS and the second scenario. In the first scenario, the GC payoff of the Danish cod fleet surpassed the NC payoff whereas the NC payoffs of the Polish and the Swedish cod fleets were higher than the GC payoffs. Lastly, the first scenario total GC payoff was greater than the total NC payoff. In the second scenario, the GC payoffs were greater than the NC payoffs for all countries and the payoffs were significantly lower than the BS.

## DISCUSSION

In this study, we assessed the NC and GC payoffs for three asymmetric countries that optimize their rents from Baltic Sea cod, sprat and herring fisheries. In general, the GC payoffs were found to be much higher than the NC payoffs. Both in the GC and the NC games, Denmark is apparently the most profitable country, especially regarding the cod fishery. The GC benefits of Denmark is much higher than the benefits of Poland and Sweden and, so that, to provide equal share of the excess benefits in the GC, Denmark would pay compensation to Sweden and Poland as it has the highest profit among countries (Nieminen et al., 2016). Having said that, this would still be an issue of debate as the profits of Sweden is as similar as the profits of Denmark.

The payoffs of all the sprat-herring fleets were highest in the second scenario that is considered as favorable for sprat recruitment due to high temperatures but unfavorable for cod due to the low salinity levels (Lindegren et al., 2009). Hence, this scenario resulted in higher recruitment and survival of the sprat and herring stocks compared to cod and increased economic returns of the sprat-herring fleets. On the contrary, when the temperature is low and the salinity high, as in the first scenario, cod is benefited and the conditions where the GC payoffs are higher than the NC payoffs are better for Denmark and Sweden because of mainly, Danish and Swedish fishery is economically dependent on the cod fishery. These results contrast with previous findings (except for Poland) that showed weakened GC payoffs (Nieminen et al., 2016). Our results are in line with previous finding by Brandt and Kronbak (2010) that suggested lower cooperation or cooperative agreements with the

negative impact on the resource rent. As mentioned in Nieminen et al. (2016), the impact of fluctuated cod, herring or sprat recruitment on fishery agreements would be better understood if other species, as well as environmental conditions are considered within the models.

[scenario 1 (S1): low temperature-high salinity; scenario 2 (S2): high temperature-low salinity].

Discount rate sensitivity of the NC model were found to be substantial. Especially, the economic returns of the cod fleets sharply increased with decreasing discount rates. In addition to weighting future payoffs higher, the lower discount rates also favor a more precautionary exploitation level allowing the stock to rebuild to a higher level. This in turn significantly increase long-term yields and reduce operating costs, as the same yield can be achieved with lower effort (Döring and Egelkraut, 2008; Lindegren et al., 2009). In contrast, a 100% increase in the discount rate yielded much lower cod and herring SSBs resulting in a considerably lower total net present value. The GC cases also followed the similar trend in discount rate sensitivity with relatively higher payoffs compared to the NC case (**Supplementary Table S3**). The sensitivity of the fish prices via 20% decrease or increase in prices resulted in very different economic output, especially for the NC case. Such volatility in payoffs due to changes in fish prices would not be wanted by the fishermen or industry. However, the GC results for the same price sensitivity intervals yielded much more positive output compared to the NC results. So that, the GC case reduced the price volatility compared to the NC case. Likewise, a 20%

decrease or increase in costs also disproportionately decreased or increased the final economic output (**Supplementary Table S4**). But, the changes in the cost parameters did not affect the economic output as much as the prices. Interestingly, in the GC case, the countries, Denmark and Poland, did not get much benefit from the decrease in cost parameters as these countries in the GC were substantially lower compared to the NC case. Having said that, increase in the cost parameters resulted in relatively higher payoffs in the GC compared to the NC (**Supplementary Table S5**). To summarize, the discount rate, price and cost parameters' sensitivity did have substantial impact in the economic shares of the NC and GC games. So that, this variation would likely to have additional increase or decreases in compensation amounts that the dominant fishing nation would likely to pay. So that, increase benefits with low discount or cost and high price would not only be good for the dominant fishing nation but also, good for the other fishing nations.

## Policy Considerations

The recent reform of the EU-CFP states that "the CFP shall ensure that fishing and aquaculture activities contribute to long-term environmental, economic and social sustainability" (European Council [EC], 2013). Quota allocation schemes, such as TAC are commonly used around the world, including the Baltic Sea. The precautionary principle, that emphasizes the management of fish stocks within safe biological limits, has been the basis of TACs allocations. However, EU has previously been unsuccessful in meeting the precautionary approach, leading to overexploitation of fish stocks, partly driven by overcapacity and poor profitability of the fishing fleets. Consequently, costs and benefits of the fleets should be considered when determining the TAC. TAC allocation would be considered according to the relative stability principle to be accepted by the all member states. A sharing rule would be solution in sharing the resource benefits. In the current study, the GC did not result in positive returns for all the fishing states. For example, Poland (in all scenarios) and Denmark (in scenario 2) received no, little or negative economic outputs from the GC compared to the NC. In this case, the countries that receive positive economic return from the GC would compensate the countries that cannot yield positive economic return. As a solution, Nash Bargaining equal sharing rule could be useful to allocate the payoff increases (Nash, 1953; see e.g., Kaitala and Lindroos, 1998; Li, 1998). The allocation should be based on compensation schemes created collaboratively in which the dominant fishing state would be transferring surplus benefits to states with negative economic returns. Further cooperation among the Baltic Sea states would be inevitable, especially given the forecasted changes in fish

## REFERENCES


stocks expected under climate change (Lindegren et al., 2010; MacKenzie et al., 2012; Bartolino et al., 2014; Blenckner et al., 2015). Costs of measures on the mitigation of the climate change would be provided by the surpass benefits. The area closure can also be substantial tool for the management of the stocks.

## CONCLUSION

The cooperative management once again demonstrated to be fundamental in defining economically optimal use strategies for shared fish resources. In our case, considering the multi-species and multi-fleet nature of the fisheries, the effectiveness of the cooperative approach would be essential in the decision-making process. Furthermore, this effectiveness of the cooperation was not only limited with the existing climate conditions but also under changing climatic conditions that would be mitigated with the cooperative agreements. Finally, it is essential to increase the number of game theoretical studies focusing on the biological and economic externalities under changing environmental conditions.

## AUTHOR CONTRIBUTIONS

ST was the leading author with substantial support from principal investigators. MtL, LR-J, and MkL involved in software coding, editing, and revising the manuscript.

## FUNDING

The MARMAED project has received funding from the European Union's Horizon 2020 Research and Innovation Program under the Marie Skłodowska-Curie grant agreement no. 675997.

## ACKNOWLEDGMENTS

The authors would like to thank Dr. Emmi Nieminen who provided fundamental insights in programing of the model.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00622/full#supplementary-material




**Disclaimer**: The results of this publication "Game Theory Applications to the Baltic Sea Fisheries under Climate Change" reflects only the authors' view and the Commission is not responsible for any use that may be made of the information it contains.

**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 Tunca, Lindegren, Ravn-Jonsen and Lindroos. 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.

# Ecological Effects and Ecosystem Shifts Caused by Mass Mortality Events on Early Life Stages of Fish

Erik Olsen<sup>1</sup> \*, Cecilie Hansen<sup>2</sup> , Ina Nilsen<sup>2</sup> , Holly Perryman<sup>2</sup> and Frode Vikebø<sup>3</sup>

<sup>1</sup> Demersal Fish Research Group, Institute of Marine Research, Bergen, Norway, <sup>2</sup> Ecosystem Processes Research Group, Institute of Marine Research, Bergen, Norway, <sup>3</sup> Marine Processes and Human Impacts Research Program, Institute of Marine Research, Bergen, Norway

#### Edited by:

Jeffrey William Krause, Dauphin Island Sea Lab, United States

#### Reviewed by:

Donald F. Boesch, University of Maryland Center for Environmental Science (UMCES), United States Moritz Mathis, Helmholtz-Zentrum Geesthacht, Germany

> \*Correspondence: Erik Olsen eriko@hi.no

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 07 June 2019 Accepted: 14 October 2019 Published: 29 October 2019

#### Citation:

Olsen E, Hansen C, Nilsen I, Perryman H and Vikebø F (2019) Ecological Effects and Ecosystem Shifts Caused by Mass Mortality Events on Early Life Stages of Fish. Front. Mar. Sci. 6:669. doi: 10.3389/fmars.2019.00669 Mass mortality events (MMEs) are a key concern for the management of marine ecosystems. Specific stages and species are at risk and the causes may be single or cumulative pressure from a range of sources including pollutants, anthropogenic climate change or natural variability. Identifying risk and quantifying effects of plausible scenarios including MMEs are key to stakeholders and a quest for scientists. MMEs affect the whole ecosystem, but are traditionally only studied in relation to specific species, disregarding ecological feedbacks. Here we use an end-to-end ecosystem model adapted to the Nordic and Barents seas to evaluate the species-specific and ecological impacts for 50 years following an MME. MMEs were modeled as 10, 50, or 90% reduced recruitment for cod, herring and haddock, individually or in combination. The MME scenarios were compared to a base case model run that includes the current fishing mortality. All species showed declines in population biomass following an MME, increasing in duration and severity with increasing mortality. Cod biomass rebounded to the base case level within 3–13 years post the MME independent of scenario, while neither haddock nor herring fully rebounded to base case levels within the considered time horizon. Haddock responded much more variably to the mortality scenarios than cod or herring, with some scenarios yielding much higher levels of biomass than the base case. Herring responded negatively to all scenarios, leading to lower herring biomass and a steeper decline of the species than seen in the base case due to persistent harvest pressure. Corresponding responses showed that the demersal guild biomass increased substantially, while the pelagic guild biomass declined. Few effects were observed on the other guilds, including the top predators. Ecosystem effects as measured by ecological indicators were greatest after 5 years, but persisted through the entire model run. Fishery indicators showed the same features, but the responses were stronger than for the ecosystem indicators. Taken together this indicates longterm, ecological response to MMEs that can be described as regime shifts, highlighting the importance of using ecosystem models when evaluating effects of MMEs.

Keywords: mass mortality events, ecosystem model, Barents sea, Norwegian sea, oil spill, cod, haddock, herring

## INTRODUCTION

fmars-06-00669 October 25, 2019 Time: 17:16 # 2

Rare events with large negative consequences aka. 'Black Swans' (Taleb, 2011) occur in natural ecosystems, human societies and economies (Sornette, 2002). Black Swans are caused by events like earthquakes, storms, oil spills, or disease outbreaks and their consequences on ecosystems can include severe die-offs or population crashes (Anderson et al., 2017), called mass mortality events (MMEs). In fact, catastrophic MMEs affecting entire populations have been increasing in frequency during the last decades (Fey et al., 2015). Events that severely affect only parts of the life-history. are also considered as Black Swans and may cause MMEs with severe short- or long-term effects on the populations and ecosystems (species examples in Lawrence, 1996; López et al., 2008), similar to the MMEs (Langangen et al., 2017). In the face of climate change and increases in other human-induced external pressures the effects of such events may be exacerbated through cumulative impact (Fey et al., 2015).

The acute effects of 'Black Swan' events, are often dramatic and extensive but the long-term effects also give great cause for concern as exemplified by the 1989 Exxon Valdez oil spill in Alaska (Peterson, 2001). There is also concern that they may lead to regime shifts (Mollmann et al., 2014), altering the structure of the ecosystem on a long-term or even permanent basis.

Several marine fish populations show sporadic recruitment patterns with many years passing between successful year-classes contributing substantially to the long-term reproductive part of the population. Therefore, there is concern that mortality events in the early life stages may have severe and long-lasting effects on the population (Ohlberger and Langangen, 2015; Langangen et al., 2017). Most long-term effect studies have focused on single species, whereas the wider effects on the entire marine ecosystem have been neglected, mainly because of a lack of adequate tools for such complex integrated analyses.

Offshore petroleum extraction is the largest export industry in Norway, having boosted the country's economy since 1969 when the first oil deposit was discovered in the North Sea. The environmental consequences, especially of acute spills caused by accidents, are at the heart of the ongoing debate over whether regions should be opened to oil and gas extraction (Olsen et al., 2016). Environmental risk analysis has been used to evaluate the effects of potential oil spill and large research projects have focused on understanding and quantifying these (Carroll et al., 2018), although with a focus on a few key species and their interactions. Quantifying the full ecological effects of such accidental spills has so far not been done for any of the Norwegian marine ecosystems.

Complex models combining the ecology and human pressures of marine ecosystems are now being developed for many major sea and ocean areas, allowing researchers to explore the ecological effects as well as the species-specific responses of perturbations and changing pressures on marine ecosystems. Such models have been used e.g., to specifically explore the ecosystem-based fisheries management strategies for fisheries in Australia (Fulton et al., 2014) and the United States West coast (Kaplan et al., 2012; Kaplan and Marshall, 2016), but also the integrated effects of a range of human activities on the Great Australian Bight in general (Fulton et al., 2018), and future fisheries, protection and ocean acidification scenarios across a range of global ecosystems (Olsen et al., 2018). Ecosystem models have been used to evaluate the ecological effects of the Deepwater Horizon oil spill in the Gulf of Mexico (Ainsworth et al., 2018). Ainsworth et al. (2018) used the end-to-end ecosystem model Atlantis to simulate impacts from the Deepwater Horizon oil spill on fish guilds within the Gulf of Mexico. By using Atlantis, Ainsworth et al. (2018) were able to discern that the biomass of fish within areas most affected by the oil spill decreased between 25 and 70%, loss of prey caused starvation amongst predators, and recovery of stocks took 10–30 + years.

In the present, study we utilize an Atlantis end-to-end ecosystem model (Fulton et al., 2011) for the Norwegian – Barents sea ecosystem (NoBa, **Figure 1**; Hansen et al., 2019a) to explore the species specific as well as the wider ecosystem effects of MMEs leading to reduced recruitment of key commercial fish species through increased mortality of fish eggs and larvae.

## MATERIALS AND METHODS

In the present study, we analyzed the ecosystem effects of MMEs on key fish species in the Norwegian and Barents sea: Northeast Arctic (NEA) Cod (Gadus morhua), Northeast Arctic haddock (Melanogrammus aeglefinus) and Norwegian Spring Spawning (NSS) Herring (Clupea harengus) all of which have major overlapping spawning grounds in the Lofoten – Vesterålen – Senja area or early life stage (ELS) drifting by Olsen et al. (2010) and Misund and Olsen (2013). The MMEs could be caused by a number of factors. However, we assumed that the pressure only occurred in a single year, which is plausible e.g., for the acute phase of an oil spill, but not for ocean acidification. That is, oil may be enclosed in ecosystem compartments such as sediments and result in releases also in subsequent years (Peterson, 2001), but at lower levels than those that are needed

FIGURE 1 | Bathymetric map of the Northeast Atlantic showing the NOBA Atlantis model polygons.

to cause a MME. Hence, we did not consider pressures lasting multiple years, but only direct mortality effects occurring in a single year in the model. While the study was motivated by risk analysis considering oil spill impacts the results are valid independent of the cause of MMEs in fish ELS. This approach allowed us to evaluate the long-term effects over a 50-year post-MME modeling period, both on the species directly affected by added mortality, but also the indirect ecological effects through trophic interactions.

The Nordic and Barents Seas Atlantis model (NoBa) is a deterministic end-to-end ecosystem model parameterized for the Nordic and Barents seas (**Figure 1**). The Atlantis framework (Fulton et al., 2005, 2011) includes a variety of modules ranging from sunshine to fisheries, socio-economy and possibilities for management strategy evaluations. NoBa includes 53 species and functional groups (hereafter components), representing the ecosystems of the areas. The components are connected through a diet matrix, where a prey availability is defined. The prey availability represents the overlap between the prey and its predator, but as Atlantis is spatially resolved, they do need to be overlapping in time and space to have any interactions. The size of prey compared to the size of the predator is also taken into consideration. Important life history processes such as growth rates, consumption rates, recruitment and mortality rates were all defined based on published literature and verified to reproduce observations of biomass and population structures (Hansen et al., 2016). Atlantis also requires daily input of sea water temperature, salinity and currents. NoBa was forced by a Regional Ocean Modeling System (Shchepetkin and McWilliams, 2005) covering the Nordic and Barents seas. The forcing was received from three different set-ups of this system, covering three different time slots; 1981–2001, 2002–2005, and 2006–2068. By using three different setups for forcing, we introduced uncertainties tied to the differences in these. However, in this study, we only analyze output from the model driven by the last forcing. This specific setup is based on emission scenario RCP 4.5, which has been shown to only include a weak climate signal during this period (Skogen et al., 2018).

The simulated biomass and catches of pelagic (including Norwegian Spring Spawning herring, mackerel and blue whiting) and demersal guilds (including Greenland halibut, Northeast Arctic cod, haddock, beaked redfish and golden redfish) were comparable to observations (Hansen et al., 2019b). The biomass levels were in good agreement, the timing of peaks and troughs though fitted better for the demersal guild than for the pelagic guild. Moreover, the model was able to reproduce the ratio found by Hansen et al. (2019a) that the individual weights in each age class in over 70% of the components (including the three key species in this study) was within 50% change of the initial values, a measure often applied for tuning of Atlantis models.

Enhanced mortality on the eggs and larvae of ecologically important fish species in the Lofoten – Barents Sea ecosystem due to oil spill has been assessed over the past decade through numerous risk assessments and fate-modeling synthesized in Carroll et al. (2018). Although Atlantis is capable of explicitly simulating an oil spill (Ainsworth et al., 2018), the coarse spatial structure of the NoBa Atlantis model (**Figure 1**) is not conducive for realistic modeling of oil spill drift. Instead, we refer to published values of significant enhanced mortality on ELSs following large oil spills and initiated scenarios where recruitment to year 1 of NEA Cod, NEA Haddock and NSS Herring span out a range of single-year recruitment failures. Recruitment to age 1 occurs in all spatial boxes where a species occur, representing the distribution of ELS. Hence an MME event occurring in limited spatial area was assumed to affect the entire population. The 43% ELS reduction reported in Carroll et al. (2018) corresponding to 12% reduction in recruitment to the adult stock was the most severe outcome of scenarios investigated in that study. However, as stated above there is significant uncertainty when projecting additional ELS mortality onto changes in recruitment. Hence, as discussed by Langangen et al. (2017), depending on how the precautionary principle is implemented the 43% loss of ELS might also be interpreted as a larger loss of recruits. We therefore investigated losses of 10, 50, and 90% of recruits, both individually and cumulatively. The 90% scenario was chosen to include a worst case MME higher than the max levels established in the Norwegian planning context.

The three mortality levels were imposed on each of the species in 12 different model runs, either by applying the added mortality to one, or to all three species simultaneously (**Table 1**). We also ran a base-case simulation without adding any mortality to serve as comparison for the MME scenarios. The mortality was applied in the model by first running a base case with no MME. Thereafter, the mortality to the first cohort of the species was applied by multiplying the number of that specific cohort by a scalar (0.9, 0.5, 0.1), e.g., by reducing the numbers of the specific cohort by 10, 50, or 90%. This was only the case for the single year where we assumed that the MME would happen. The simulated years after the MME was performed without any forcing. All runs included active fisheries, following historical fishing pressures until 2017, thereafter applying Fmsy levels (fishing pressures achieving maximum sustainable yield) for all commercial stocks (Hansen et al., 2019b).

Each model run was 110 years with 24 years for spin-up followed by 36 year reflecting the 1982–2017 historical period before any perturbation was introduced. The spin-up period was forced by looping the physical forcing from ROMS representing 1981 24 times. The mortality was introduced in one time-step in January of year 62 (2018). The exact timing of the MME during a year would have little effect on model outputs due to the MME being modeled as a reduction of the incoming cohort,

TABLE 1 | Scenarios of reduced survival at the egg and larval stages of Northeast Arctic cod (Gadus morhua), Northeast Arctic haddock (Melanogrammus aeglefinus), Norwegian spring spawning herring (Clupea harengus), evaluated using the Norwegian – Barents Sea Atlantis model.


not a disruption of a spawning event. In the years after the event mortality was not adjusted as we assume the acute effects of the MME to be transient and out of the system well before the next annual recruitment event. This might not be true if there are additional environmental changes at spawning grounds that may affect future spawning. However, we currently do not have information indicating that this may be the case.

The effects of the 12 scenarios (**Table 1**) were evaluated at three levels: (1) the direct responses on an index of biomass (dividing all biomass values by 250 000) and age structure of the species affected in each scenario and the indirect effect on the other two of the three focus species, (2) the indirect effects on other ecosystem components as evaluated by functional guilds (**Supplementary Table S1**), and (3) the effects on ecosystem structure and emergent properties as measured through previously vetted indicators of ecosystem and fisheries status (**Supplementary Table S2**; Fulton et al., 2005; Olsen et al., 2018). The guild responses and ecosystem indicators were evaluated as averages of the first 5, 10, and 20 years post the MME.

## Model and Code for Analysis

The scenarios were run with Atlantis version 6205. Model outputs were analyzed in R studio version 3.5.2. Scripts used to generate plots and the raw model output data are available upon request.

## RESULTS

## Total Biomass

All scenarios for reduced recruitment (reduced survival of fish eggs and larvae) had clear effects on the total stock biomass of the species directly affected (**Figure 2**), as well as on all species included in the NoBa model (**Supplementary Figure S1**). Although the focal species of the present analysis has been cod, haddock and herring, the other species included in the NoBa Atlantis model also reacted to the MMEs indirectly through food-web interactions. In general, the magnitudes of their reactions were lower than for the study species (**Supplementary Figure S1**), with some notable exceptions like capelin, beaked redfish, killer whale, squid, large and medium zooplankton (Calanus finmarchicus) that all showed responses exceeding 10% for some of the scenarios.

There was an immediate drop in biomass for all three focus fish species, largest for cod and smaller for the haddock and herring. Cod rebounded fairly quickly from the mortality event, in 3 years for the CO10 scenario, 8 years for CO50 and 13 years for CO90, with cod total biomass stabilizing at base case levels (**Figure 2A**). None of the scenarios for haddock or herring had any discernible effect on the total biomass of cod (**Figure 2A**) as all lines overlapped completely with the base case.

The loss of juvenile fish due to a MME could affect dietary responses across the ecosystem, such as shifting predation pressures due to the loss of important sources of food or promoting the growth of other juvenile groups due to the reduction in competition for resources. The NOBA Atlantis model uses nitrogen to track the flow of nutrients through the marine ecosystem, so the condition of age-structured functional groups can be tracked through the energy allocation to structural (e.g., hard tissue; bone) and reserve (e.g., soft tissue; fat) nitrogen. To explore changes in organismal condition following a MME, we considered changes in residual nitrogen relative to the residual nitrogen just before a MME (RNt/RN0; where t is the annual time step) and structural nitrogen relative to the structural nitrogen just before a MME (SNt/SN0). Nitrogen ratios of cod were about (∼1.6) for all scenarios, both for residual and structural Nt/N<sup>0</sup> (**Supplementary Figures S2, S3**), indicating increased individual growth and fewer individuals toward the end of the model run compared to the time of the MME.

The impacts of reduced recruitment to haddock had very different effects on the total biomass than for cod (**Figure 2B**). The HA90 and HA50 scenarios both led to immediate and discernable drops in biomass, although the relative changes from the base case and between scenarios were much lower than for cod. Interestingly, the HA10 scenario showed no immediate drop in biomass as compared to the base case, only starting to diverge from the base case 18 years after the MME, and staying below the base case until 33 post the MME, after which it fluctuated above and below the base case. Eight years past the MME both the HA90 and HA50 scenarios had rebounded to the base case level, and thereafter followed the base case trajectory for 10–18 years, when biomass became higher than the base case for the next 10–15 years. 10–28 years past the MME the HA50 and HA90 scenarios showed variable performance compared to the base case, but at the end of the model run both HA50 and HA90 showed biomass levels above the base case. Thus, the haddock scenarios were much more chaotic than the cod scenarios, possibly reflecting the naturally more stochastic recruitment events and stock fluctuations in this stock compared to the cod (Olsen et al., 2010). The haddock also showed clear responses to reduced recruitment in the other two species (**Figure 2B**), with all herring and cod scenarios leading to an increased haddock stock above the baseline for at least 20 years past the mortality event. Haddock nitrogen ratios at the end of the model run were lower than one for both residual and structural nitrogen for all scenarios except HA90 and MX90, indicating a reduced individual growth toward the end of the modeling period compared to the time of the MME (**Supplementary Figures S2, S3**).

The modeled herring population steadily declined under the present level of recruitment and harvest control rules (HCR), with or without a mortality event. Added mortality exacerbated this decline. There were however slightly higher levels of herring for some of the cod scenarios compared to the base case due to reduced predation by cod (see **Figure 2C**). The HE10 scenario showed no immediate drop in total biomass level and this scenario followed the baseline only with minor variability through the whole model run. In contrast, both the HE50 and HE90 showed an immediate drop in total biomass index relative to the baseline. Both the HE50 and HE90 scenarios had lower total herring biomass for the entire run, compared to the base case, but none had completely merged with the base case. Herring responded only slightly to reduced recruitment for the other two species, and only from 9 years after the MME. The exception was the CO50 scenario

that led to lower total biomass of herring from nine years past the MME and onward. From 22 years after the MME the CO50 scenario showed the same levels of total biomass as the HE50 and HE90 scenarios. After 30 years the other non-herring scenarios diverged from the base case. The HA10 and HA90 scenarios led to slightly higher herring biomass while the CO10 and CO90 cod and HA50 haddock led to reduced total biomass compared to the base case for the rest of the run. Herring also had nitrogen ratios below 1 for both structural and residual nitrogen for all scenarios except HE90 and MX90, indicating a reduced individual growth toward the end of the modeling period compared to the time of the MME (**Supplementary Figures S2, S3**).

For cod the combined scenarios followed the same trajectory as the species-specific scenarios for the entire model run (**Figure 2A**). For herring most scenarios lead to generally lower biomasses depending on the imposed mortality, with a similar impact of both species-specific or combined MME scenarios. For haddock the response to the combined scenarios was more variable, although at the start all three showed declines comparable to the haddock species-specific scenarios (**Figure 2B**). Most haddock scenarios lead to distinctly higher biomasses, without any clear trend toward rebound to the base case.

## Changes in Age Structure

Since all vertebrate components in the Atlantis models are represented in an age-structured way we could evaluate the changes to each specific age group under each scenario. In **Figure 3**, the age structure is shown for all combined scenarios: MX10, MX50, MX90. From **Figure 3**, a clear rippleeffect showing how an initial decline in recruits due to the mortality event could be followed as a decline in one-year-olds after one year, two-year-olds after two years and so on (see **Figures 3A,B,H,I** for the clearest examples).

The increasingly negative effect when reducing recruitment was clearly seen for all three species with the cod rebounding to levels close to the base case 3–13 years post the mortality event (**Figures 3A–C**). For haddock the initial decline was much smaller, and there was no rebound – rather, the stock levels increase, although with large variability especially for the MX10 scenario (**Figures 3D–F**). For herring there was more variability around the base case, especially for the MX10 scenario that showed increasing fluctuations toward the end of the run (**Figure 3G**), while the MX50 and MX90 herring scenarios showed a steady decline compared to the base case (**Figures 3H,I**).

## Ecological Responses

### Trophic Guild Responses

We observed clear effects of the MME on the demersal fish and pelagic fish guilds, but varying with scenario and time after the MME. Only the effects of the most severe scenarios (CO90, HA90, HE90, MX10, MX50 and MX90) are shown in **Figures 4A–C** below, the trophic guild responses of the remaining six scenarios are shown in **Supplementary Figure S4**. Average guild effects were small (<5%) for the first 5 year period past the MME event (**Figure 4** and **Supplementary Figure S5**). They increased over time for the demersal fish guild, while the average responses of the pelagic fish guild remained comparable for 5, 10, and 20 years past the MME. The average guild responses for the demersal and pelagic guilds were driven by increased responses in one species, while the other guild species remained at the base case, or in some cases even showed decreased responses. Responses of individual species increased for all guilds over time, and after 10 years one could start noticing effects on other guilds than those directly affected by the MME.

Some very minor effects could be seen on the epibenthos (food for demersal fish), and zooplankton (food for pelagic fish) for the 5-year period past the MME event, but these effects increased by 10 and 20 years. By year 20 there were some positive effects on squid for the CO90 and MX50 scenarios, possibly linked to the large increase in biomass of prey from the demersal guild.

0 is the year of the mortality event.

The average effects on the demersal and pelagic guilds were slightly negative for all the combined scenarios (**Figure 4A**) in the first 5 years past the mortality event, except for the demersal fish guild which responded slightly positively to the HE90 scenario and the pelagic fish guild which responded positively to the CO90 scenario. For each scenario there was a spread in responses

(triangles within the colored bars of **Figure 4A**) with only a single species showing a marked response (in either direction) while the rest remained at or close to the base case.

After 10 years (**Figure 4B**), the demersal fish guild still showed a positive response under the HE90 scenario, but now there was a slight positive response in the MX10 scenario. The average response for the CO90, MX50 and MX90 scenarios were the same as the base case, but for these three scenarios some species responded positively while others responded negatively.

In the pelagic guild the average guild and individual species responses were similar to the situation 5 years earlier. Epibenthos responded slightly negatively to all scenarios, while the primary production responded slightly positively to all scenarios.

Over the 20-year period past the MME event (**Figure 4C**), the demersal fish guild had responded positively under all scenarios, driven by a very strong positive response in haddock. In the demersal guild there were some slight negative responses of single species for the CO90, MX50 and MX90 scenarios although the average response was positive. The pelagic guild again showed a similar response as 5 and 10 years past the MME, although the extent of single species responses had been slightly dampened. Epibenthos showed less responses than after 10 years, while primary production was slightly higher for all scenarios. Zooplankton responses were also stronger than after 10 years, we saw increased biomass of squid for the CO90 and MX50 scenarios. It is also interesting to note the small average increase (<5%) in the primary producer guild.

## Ecosystem and Fishery Indicators

There were stronger responses in the fishery indicators than the ecological indicators to the MME event at all three-periods (5, 10, and 20 years) past the event (**Figures 5A–C** and **Supplementary Figure S6**). Response levels for the ecological indicators were minor (<5%), while the fishery indicators showed some stronger responses (up to 25%, for the 5 year period past the MME). The response of the ecological indicators changed over time, with the strongest responses appearing in the 20 year period past the MME event (**Figures 5A–C**). However, yearly responses varied much more (**Supplementary Figure S6**), with the largest yearly deviations from the base case occurring more than 20 years after the MME. Contrary, the effects on the fishery indicators were dampened after the MME and lowest over the 20 year period post the MME. Still, even 20 years past the MME event there were up to 6% negative effects on some of the fisheries indicators Some of the fisheries indicators (Total Catch, Value of catch, Demersal catch) showed a similar pattern to the ecological indicators with increased deviation from the base case (with large variability between the scenarios) during the period from year 20 to the end of the model run (**Supplementary Figure S6**). These changes in both ecological and fisheries indicators can mainly be linked to the large changes in the haddock population during this period.

The ecological indicator response was very similar for the 5 and 10 year periods past the MME (**Figures 5A,B**), with only the demersal – pelagic fish ratio (Dem.pel.fish) decreasing more than 2% compared to the base case in the CO90, HE50, HE90 and MX 90 scenarios. Five years past the MME the proportion of predatory fish increased above the base case for the CO90, MX90 and MX50 scenarios, while the demersal fish ratio decreased below the base case for the MX90, CO90, MX50, HE90, HE50 scenarios. This pattern remained 10 years past the MME, although less intense responses were observed for the CO90 and MX90 scenarios. Then, over the 20 year period past the MME the pattern changed dramatically, with all ecological indicators, except the proportion of predatory fish, showing divergences >1% for one or more of the scenarios (**Figure 5C**). Now the pelagic biomass – primary production ratio (Pel.bio.pp) showed the greatest variability around the base case with the CO90 scenario exceeding the base case by 4.3% while the MX90 was 2.1% below the base case. Similar diverging patterns of responses were seen for both the Bio.pp, Dem.pel and Dem.pel.fish indicators. It was also interesting to note that the various scenarios had different indicator responses. Whereas the CO90 scenario had the lowest response for the Dem.pel indicator, 3% below the base case, as noted above it exceeded the base for the Pel.bio.pp indicator.

All fisheries indicators, except the mean trophic level of the catch (MTL.C, **Figures 5A–C**) showed responses diverging >1% for one or more of the scenarios. Although the strongest responses were seen over the first 5 years past the MME, and then decreasing 10 and 20 years past the MME, the fisheries indicators never rebounded completely to the base case level for any scenario. Even 20 years after the MME the fisheries indicators of the CO90 and MX 90 scenarios were below the base case level (except for MTL.C).

## DISCUSSION

By using a vetted Atlantis end-to-end ecosystem-model of the Norwegian and Barents Sea (Hansen et al., 2019a) we were able to simulate the ecosystem effects of increased mortalities in fish egg and larvae of cod, haddock and herring. All species show immediate declines in biomass following the mortality event, from low, but measurable effects at 10% added mortality, increasing with the severity of the event, and highest for the 90% mortality scenarios affecting cod (CO90), or all three species (MX90). This immediate post-event decline was expected, both based on our understanding of species life cycles as implemented in the NoBA model, as well as on previous modeling studies (Ohlberger and Langangen, 2015; Langangen et al., 2017; Carroll et al., 2018). However, we see that the effects on cod persisted for longer than the previous studies have indicated, caused widely fluctuating biomass of haddock, and accelerated the population decline of herring (**Figure 2**). In the NoBa model, cod had a very limited response to increased mortality for the haddock or herring, while both herring and haddock (to the greatest degree) showed strong responses to changes in the mortality of the other two species. This clearly shows the importance of including species interactions when studying the effect of perturbations to ecologically important species. It also indicates differences in resilience between the three species to MMEs, with the cod population being resilient, returning to biomass levels similar to the base case 13 years after the mortality event. Herring

FIGURE 5 | Averages of the ecological (left) and fisheries (right) indicators within the years following a mass mortality event. (A) 5 years post the mortality event, (B) 10 years post the mortality event, (C) 20 years post the mortality event. Displayed values are relative to the baseline run and are based on interannual snapshots of the indicator values. Pel.bio.PP: Pelagic biomass as ratio of primary production, Dem.bio.PP: Demersal biomass as ratio of primary production, Dem.Pel: Demersal – Pelagic biomass ratio, Dem.pel.fish: Demersal – Pelagic fish biomass ratio, Predfish.Prop: Proportion of predatory fish, MTL.bio: mean trophic level of biomass, Bio.pp: total biomass as ratio of primary production, Pel.C: catch of pelagics, Dem.C: catch of demersals, Fish.C: catch of fish, Val: value of catch, Exp.rt: exploitation rate of all, Fish.exp.rt: exploitation rate of fish, MTL.C: mean trophic level of catch, Tot.C: total catch (see Supplementary Table S2 for details).

and haddock showed no similar tendency to resilience, rather the haddock population showed a very fluctuating population biomass for all mortality scenarios, while the herring biomass under the mortality scenarios were consistently under the base case biomass for the entire model run. Our results correspond to the results of the sensitivity analysis of a different NoBa Atlantis

model implementation (Hansen et al., 2019a), which showed that top predators such as whales and seabirds are less impacted by changes in other species, as they are generalists and therefore are able to switch prey if a prey group decreases while species that are at the center of the trophic web, in particular herring, are much more vulnerable to changes at both their own level, but also at other components.

Cod experienced an immediate decline in overall biomass, and we observed a clear ripple-effect of reduced survival of the affected year-classes (**Figure 3**) that varied with the intensity of the mortality. Although the immediate drop in biomass was compensated and the stock biomass returned to base case levels in 10–15 years, showing stability and indicating low internal variability of the cod stock. This was in stark contrast to the haddock which exhibited signs of large internal variability (i.e., sensitivity to initial conditions) making it a potential significant factor in the simulated biomass trajectory. The immediate decline compared to the base case ranged from ∼5% for the MX10 (and CO10) scenario to ∼55% for the MX90 scenario. The effect on year-classes subsided with time, but at age 3, when cod mature and start entering the fishery (Durant et al., 2008) they were ∼5% below the base case for the MX50 scenario and ∼12% below in the MX90 scenario. A decline of 5–12% of year-class 3 of cod translates to 9 570–22 969 tonnes, based on the 2018 NEA cod stock assessment (ICES, 2018). For the fishery the effects of a MME first become noticeable after the year class enters the fishery. Three years past the MME, the effect on cod would be marginal as few of the 3-year olds are caught in the fishery. However, the effect will increase as the fish grow and become more available to the fishery, and the affected year class becomes a more substantial part of the catch. We see clear effects on the demersal fishery (which is dominated by cod) of the scenarios including 50% or 90% mortality of cod (CO50, CO90, MX50 and MX90) when evaluating the fishery indicators (**Figure 5**) throughout the model run, although at the end (20 years past the event) the effects are less than 5% compared to the base case. Immediately after the event (5 years) we observed a reduction in demersal catch of up to 25% leading to a 10–20% reduction in the value of the catch for the 50 and 90% mortality scenarios. Such effects are substantial and dramatic for the fishing industry in Norway and Russia that dominate the demersal fisheries in the Barents Sea.

Since cod, haddock and herring are key species in the Norwegian Sea and Barents Sea ecosystems, it is no surprise that changes in the biomass of these species have strong effects on the total biomass of the demersal and pelagic fish guilds that they are part of **Figure 4**. In particular the increases in the demersal fish guild after 10 and 20 years was driven by the large increases in the haddock population caused by very good recruitment events at three to four periods after the MME (depending on the scenarios – see **Figure 3B**). Such sporadically good recruitment events are a characteristic of the Northeast Arctic haddock stock (Fogarty et al., 2001; Olsen et al., 2010), indicating that the NoBa model gives a realistic representation of the haddock recruitment potential. A sporadic recruitment potential also implies that the species reproduction is more opportunistic than species with a more stable recruitment, being able to take advantage optimal physical conditions or opportunities with more prey that may arise from perturbations to the system (even pulse perturbations as in the present analysis), potentially caused by a MME. It may be hypothesized that this is a more general characteristic of sporadic recruiting species – that they are better able to take advantage of perturbations than less opportunistic species. However, at the guild level the effects of the MMEs were most discernable at the 50 and 90% mortality levels, indicating a dampening effect of perturbations stemming from varying effect of different guild members, and increased resilience of ecological guilds comparable to what has previously been observed for similar ecosystem modeling studies (Olsen et al., 2018). Even so, the ecological effects that were observed continued for the entire model run, and for several of the guilds increased 10–20 years after the mortality event. The changes in guild biomass indicate changes in ecosystem structure, which was clearly seen in the ecosystem and fishery indicators (**Figure 5**).

Taken together the effects on species biomass, ecological guilds, ecosystem and fishery indicators point to complex ecological effects of MMEs, effects that may be temporary blips for some species, but that may lead to long-lasting or even permanent changes in the absolute population size and trophic role. The ecosystem indicators also point to longterm secondary ecological effects impacting guilds not directly affected by the event, but that are prey or predators of the species affected. Indicators of ecosystem state and fisheries outputs thus give insight into how interactions in the ecosystem respond to perturbations to single or combined components of the system. Here they indicate long-term and measurable changes in the dynamics of the ecosystem. The ecosystem becomes more dominated by the demersal fish guild under all mortality scenarios compared to the unperturbed base case run. Such changes are persistent alterations of the structure of the ecosystem, also apparent in the species trajectories of all species in the NoBa model (**Supplementary Figure S1**), and thus can be classified as regime shifts according to Mollmann et al. (2014).

Neither the marine mammal, seabird nor shark guilds showed any response to any of the mass mortality scenarios at any time step (**Figure 4**). These three guilds are the top-predators of the system and find much of their food in the pelagic guild, and to a less extent the demersal fish guild. It is therefore a bit surprising that there is no feedback effect on these guilds from changes in the biomass of their prey. Because the changes in the pelagic and demersal guilds mainly affected cod, haddock and herring, while the top predators remained at base case level the most probable explanation is that the top predators in the model changed their feeding on other members of their prey guilds when the biomass of their preferred prey changed. When the recruitment of NE Arctic cod experienced an MME, the predation upon cod reduced amongst all of the primary predators (**Supplementary Figure S7A**). However, there were not any other changes in predation from these predators (**Supplementary Figure S7B**). This was most likely because NE Arctic cod was not a key prey item for any of the indicated predators. A similar situation could be observed for NE Arctic haddock

(**Supplementary Figures S8A,B**). When the recruitment of herring experienced an MME, decreases in predation on herring was observed immediately for boreal seabirds and after 25 years for killer whales (**Supplementary Figure S9A**). However, the direct impact from the MME appears to have little influence on the predation of these key predators on other prey groups (**Supplementary Figure S9B**).

For the historical period, NoBa showed reasonable biomasses and development for both the pelagic and demersal guilds, comparable to Hansen et al. (2019b) runs (**Supplementary Figure S10**). For the demersal guild, the correlation between observed and modeled biomass was particularly high (r = 0.89), whereas the correlation between observed and modeled pelagic biomass was not as high (r = 0.47). However, it needs to be kept in mind that these comparisons are not completely independent, as the harvest pressure for the historical period unquestionably had an impact on the development of the stocks/guilds. Although the model proved to be within reasonable bounds for these guilds, there are still large intrinsic uncertainties in complex end-to-end models (Lehuta et al., 2016; Hansen et al., 2019a).

## CONCLUSION

MMEs on early life stages of fish have strong negative responses in the short term. In our modeled system some species rebounded, and some even to higher levels than before. Systemic effects could be traced in the ecosystem for the entire model run which lasted 50 years post the event, indicating that such events perturb the system into a new state. In the present scenarios this state is one where demersal fish species dominate more than in the base case without the mortality event. The main driver for the demersal dominance were the several successful recruitment events of haddock leading to large increases in this population, indicating that the sporadically recruiting haddock was better able to take benefit from the perturbations to the ecosystem caused by the MME. It may be hypothesized that the ability to take advantage of ecosystem perturbations is a characteristic shared by many marine teleost species with sporadic recruitment pattern.

Our model analysis supports the view that cod as a species is resilient to MMEs, but found a longer impact duration than that of Durant et al. (2008) and Ohlberger and Langangen (2015), and more similar to those found by Ainsworth et al. (2018) in the Gulf of Mexico. Our modeled impacts are stronger than those modeled for oil spills in the Great Australian Bight (Fulton et al., 2018), but then again, our scenarios of 50% and 90% reduction in recruitment were higher than what was predicted in the Australian oil spill scenario. Also, we found significant impact on the fishery following a 50% recruitment reduction, which is a lower level than that explored by Ohlberger and Langangen (2015). Our scenario of 50% reduced recruitment was chosen based on the worst case scenarios of 43% simulated by Carroll et al. (2018), and similar levels modeled by Ohlberger and Langangen (2015), that in turn lead to <3–12% decreases in adult cod (spawning stock) biomass (**Figures 2**, **3**).

Our motivation for initiating this study was the ongoing discussion of the ecological effects of major oil spills on marine ecosystems, one potential cause of a MME. Oil spill environmental impact assessments are an integral part of the Norwegian Integrated management plan for the Lofoten – Barents sea areas (BSMP, Olsen et al., 2007), but the magnitudes of the spills assessed are debated. In the 2011 revision of the BSMP a maximum spill rate of 4500 m<sup>3</sup> /day was used (Hauge et al., 2014), which is low compared to oil spills like the Deepwater Horizon with a spill rate up to 8332 m<sup>3</sup> /day (McNutt et al., 2012). Limiting the analysis to 4500 m<sup>3</sup> /day would potentially underestimate the worst-case oil spill for the region. Concerned by the critique of the maximum spill size (Hauge et al., 2014), we therefore chose to include higher mortality levels (90%) than those used in the risk analysis of the management plan, and by Carroll et al. (2018) in their integrated modeling study.

Previous studies have limited their scope to a single species, i.e., cod, which seems to be more resilient to oil than its sibling species haddock (Sørensen et al., 2017). Adding to this the different life histories and recruitment patterns (Olsen et al., 2010) for cod, herring and haddock leads to overlooking speciesspecific and ecological responses that become apparent when investigating the effects of MMEs using end-to-end ecosystem models. One would have expected all species to respond like cod, to have an immediate effect of the MME, but then return to the base case level in a few years, and remain stable for the rest of the model run. Our model system did not behave like this neither for haddock, nor herring, and this is a very important result to be aware of when evaluating the consequences of human activities that can lead to MMEs, e.g., petroleum activities. Still, our results only show potential effects of an MME event given the limitations of the structure and parameterization of our Atlantis model. Thus, our results need to be checked with other models, as well as empirical and analytical studies, to verify the potential effects we have uncovered. Even so, under the precautionary principle, potential long-term ecosystem effects such as these we have discovered should be included in the considerations taken by managers of the Norwegian and Barents Sea ecosystems.

Understanding the effects of MMEs, be it on entire populations or the recruitment to the population, is vitally important in order to predict the possible changes to species and ecosystems from external perturbations ranging from pollution events such as off-shore oil spills, climate change leading to increased temperature and increased ocean acidification, or epidemics. In the Norwegian – Barents sea ecosystem the potential effects of oil spills are at present the most important to management as it is at the heart of the ongoing political discussions of what areas should be opened to the petroleum industry (Misund and Olsen, 2013). However, in the longer term (50–100 years) climate change impacts such as ocean acidification (Olsen et al., 2018) and temperature increases may overshadow the effects of direct anthropogenic activities. This makes our present ecological modeling approach of direct relevance to the ongoing and future management of the Norwegian and Barents Sea ecosystem.

## DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request to the corresponding author.

## AUTHOR CONTRIBUTIONS

fmars-06-00669 October 25, 2019 Time: 17:16 # 12

EO, CH, HP, and FV conceived the idea for this study and defined the scenarios. CH, IN, and HP carried out the modeling of the scenarios and produced the results. All authors contributed to the analysis and writing of the manuscript.

## REFERENCES


## FUNDING

This work was carried out as part and funded by the Institute of Marine Research Strategic Project 'Reduced Uncertainty in Stock Assessment' (2016–2020), project number 3680\_14809.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00669/full#supplementary-material


between a global and a regional model system. ICES J. Mar. Sci. 75, 2355–2369. doi: 10.1093/icesjms/fsy088


**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 Olsen, Hansen, Nilsen, Perryman and Vikebø. 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.

# Management Scenarios Under Climate Change – A Study of the Nordic and Barents Seas

Cecilie Hansen\*, Richard D. M. Nash, Kenneth F. Drinkwater and Solfrid Sætre Hjøllo

Institute of Marine Research (IMR), Bergen, Norway

#### Edited by:

Paul E. Renaud, Akvaplan-niva, Norway

#### Reviewed by:

Nis Sand Jacobsen, National Marine Fisheries Service (NOAA), United States Marta Coll, Superior Council of Scientific Investigations, Spain Vidette McGregor, National Institute of Water and Atmospheric Research (NIWA), New Zealand

#### \*Correspondence:

Cecilie Hansen cecilie.hansen@hi.no

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 02 June 2019 Accepted: 14 October 2019 Published: 30 October 2019

#### Citation:

Hansen C, Nash RDM, Drinkwater KF and Hjøllo SS (2019) Management Scenarios Under Climate Change – A Study of the Nordic and Barents Seas. Front. Mar. Sci. 6:668. doi: 10.3389/fmars.2019.00668 The effects of increasing fishing pressure in combination with temperature increases in the Nordic and Barents Seas have been evaluated using an end-to-end model for the area forced by a downscaled RCP 4.5 climate scenario. The scenarios that have been applied have used four different fractions of fisheries mortality at maximum sustainable yield (Fmsy); 0.6, 0.8, 1.0 and 1.1 × Fmsy. As it is highly likely that more ecosystem components will be harvested in the future, the four scenarios have been repeated with fishing on a larger number of ecosystem components, including harvesting of lower trophic levels (mesozooplankton and mesopelagic fish). The zooplankton biomass had an increasing trend, regardless of the increase in fishing pressure on their predators. However, when introducing harvest on the lower trophic levels, this increase was no longer evident. When harvesting more components, the negative response in biomass of pelagic and demersal fish to increasing harvest became more prominent, indicating an increasing vulnerability in the ecosystem structure to stressors. Although harvest on lower trophic level led to an immense increase in the total catch, it also resulted in a decrease in the total catches of pelagic and demersal fish, despite more species being harvested in these guilds.

Keywords: ecosystem based management, Atlantis end-to-end ecosystem model, climate change, fisheries management, socio-political pathways

## INTRODUCTION

Facing climate change and a growing human population, the world's ocean resources will be put under even higher pressures in the future than they already are. The increased demand for food from the oceans for human consumption has already led to declining stocks in several marine ecosystems (e.g., FAO, 2018a), and cumulative impacts of climate and increased fishing pressure may cause further declines (Halpern et al., 2015; Armstrong et al., 2019). There are however, strong spatial differences in how climate change affects ecosystems. Cheung et al. (2010) showed that highlatitude areas such as the Norwegian and Barents Seas are likely to experience an increase in total catch potential in the future, based mostly on calculations of future primary production, trophic level of the species and its geographic range. This is supported by observations from the Barents Sea over the last decades, where increasing temperatures have been beneficial for e.g., Northeast Arctic cod Gadus morhua (Kjesbu et al., 2014). However, changes in fisheries management strategies can be as or more important to the ecosystem as climate change (Groeneveld et al., 2018).

To meet the United Nations sustainability development goal (SDG) 2 (zero hunger), catches from marine ecosystems should likely increase. This has to occur along with SDG 14 (life below water), securing sustainable harvests of all marine resources. As seen in FAO (2018a), the commercial stocks being harvested above sustainable levels are currently at 33%, whereas only 7% are underfished. In the Norwegian and Barents Seas, most stocks are being harvested at or close to maximum sustainable yield (ICES, 2017, 2018a). Hence, a further increase of the fishing mortality would likely lead to declining stocks and decreasing catches.

Another possibility is to harvest more ecosystem components, including lower trophic levels such as zooplankton and mesopelagic fish. The production and biomass at this ecosystem level are considerably higher compared to pelagic fish and toppredators. However, by being the most important food source for juvenile and pelagic fish, they form the basis for nearly all life in the ocean. Therefore, harvest at this trophic level needs to be managed in a way that does not disrupt the balance in the system. In the Norwegian Sea, a trial fishery for the copepod Calanus finmarchicus has been active for the last decade or so (Grimaldo and Gjøsund, 2012). Recently, a commercial quota was set for this species but with area-based restrictions. However, the quota is low (165 000 tonnes) compared to the standing stock biomass (∼31 million tonnes), following a strict precautionary approach to the fishery (Broms et al., 2016).

Exploring the effect of changes in management strategies, including the number of ecosystem components being harvested, in combination with climate changes can be undertaken by applying end-to-end ecosystem models. End-to-end models include 'everything' from sunshine to fishing vessels and harvest control rules, and are built for studying ecosystem effects resulting from almost any kind of perturbation or disturbance (Plagányi and Food and Agriculture Organization of the United Nations [FAO], 2007). They also provide the only way of testing indirect and direct impact of increased harvest on multiple components in an ecosystem, without applying the change on the real system. Complex end-to-end models such as Atlantis (Fulton et al., 2011) give an overview of tradeoffs that need to be considered, especially when introducing harvest at lower trophic levels. However, their complexity and level of uncertainty makes them inappropriate for setting quotas (FAO, 2008; Link et al., 2010).

Future management regimes are very uncertain, and changes in harvest pressures or strategies can outweigh climate changes (Frank et al., 2016). The EU Climate change and European aquatic RESources (CERES) project approached the complex problem of investigating climate change effects using a scenario approach (CERES, 2016), and utilized the community-derived Shared Socio-economic Pathways (SSPs) from the Intergovernmental Panel on Climate Change (Riahi et al., 2017) which inspired a simplified version of four sociopolitical scenarios, including changing fishing pressure only. To examine the potential consequences of these simplified scenarios on the fisheries of the Barents and Norwegian Sea in a future climate, we utilized a complex end-to-end ecosystem model developed specifically for the Nordic and Barents Seas (Hansen et al., 2016). For the environment, we applied the results from a regional downscaled IPCC RCP4.5 scenario (Sandø et al., 2014; Skogen et al., 2018). Combined changes in climate, fishing pressure and the number of harvested components in the system are explored, and the uncertainties in this set of scenarios are discussed.

## METHODS AND MODELS

Atlantis (Fulton et al., 2011) is a deterministic end-to-end ecosystem model consisting of multiple modules, including ecology, physics and fisheries. The Nordic and Barents seas Atlantis model represents the ecosystems of the respective areas based upon 53 species and functional groups (Hansen et al., 2019). These species and functional groups are biomass dominant, key species (e.g., predator–prey relationships), vulnerable and/or commercially important (Hansen et al., 2016). The model domain consists of 60 polygons that cover the Nordic and Barents seas, an area of close to 4 million km<sup>2</sup> (**Figure 1**). Vertically, the model is relatively coarse, resolving the water column with seven depth levels and one sediment layer.

The 53 species and functional groups (hereafter components) are connected through a flexible diet matrix, where the predator–prey interactions are defined as fractions of prey available for the predator. However, if the prey is an inadequate size for the predator, or not overlapping in time and space, the predator will switch to another prey. Likewise, the biomass of the prey will have a large impact on the diet of the predator. In Hansen et al. (2019), it was found that of the five most important key life history parameters (Pantus, 2006), the growth rates of particularly the lower trophic levels are important for the behavior and responses in the model.

As physical forcing, the Nordic and Barents Seas Atlantis model (hereafter NoBa) applied information on temperature, salinity and currents from three different set-ups of the Regional Ocean Model System model (Shchepetkin and McWilliams, 2005), covering the periods 1981–2000 (reference period), 2001– 2005 (for comparison of observed and modeled present day biomass) and 2006–2068 (future climate; regional downscaled IPCC RCP4.5 scenario (Sandø et al., 2014; Skogen et al., 2018). The differences between the three configurations of the ROMS model are likely to introduce changes in the physical forcing, and we have therefore chosen to mainly focus on results from the latter part of the simulation (2006–2068). There are, however, comparisons between historical and model biomasses for the period 2001–2015. The temperature trend in the period 2006– 2068 is relatively weak, 1SST ∼0,02◦ year−<sup>1</sup> , although with regional differences. No trend in primary production is seen, but the timing of spring bloom is 1 month earlier at the end of the period (Skogen et al., 2018). Despite the limitations in using one single possible future climate evolution as forecasts of future conditions, we see this as a novel approach to enhance our understanding of marine species' and systems responses to multiple climate drivers and pressures, which further can be fed into the development of climate adaptation strategies and actions.

The NoBa model has a spin-up time of 24 years, for which the model only applies the physical forcing from 1981. Thereafter, the simulation ran for 87 years, including a 36 year period following historical fishing levels for the main commercially important stocks (**Table 1**), from 1981 to 2016. The historical fishing levels were calculated using assessment catches and total stock biomass data for the larger commercial



For the currently harvested species, the median fishing mortality is given together with the first and third quartile. These are calculated for the entire time series of applied fisheries in NoBa, all in all, 36 values. <sup>∗</sup>The level of fishing mortality for capelin Mallotus villosus is not at the Fmsy level, but the applied fishing mortality based on a representative average over the last decade.

stocks (ICES 2017, 2018b). As opposed to using the total catches from the assessments directly in the model, we chose to apply time series of fishing mortality (yr−<sup>1</sup> ). These were calculated by dividing total catch (t yr−<sup>1</sup> ) by total stock biomass (t). The last part of the model simulations (2016–2068) applied the fishing mortality representing maximum sustainable yield (**Table 1**). The fishing mortality changes only once per year and transfers from the previous year's mortality to current mortality over the course of 1 day. The fishing mortality was applied on the whole stock, evenly across the model domain. We only briefly present a modelobservation comparison, as a full skill assessment of the model would be a paper in itself.

To identify additional new components for potential fishing, we used fisheries statistics from the period 1980–2010. These data were obtained from the Norwegian Directorate of Fisheries, and have a higher spatial resolution compared to ICES official areas (Directorate of Fisheries, 2019). Only areas overlapping with the model domain were included, all data collected in areas outside of the model domain were excluded. The data were resolved at species-level, thus the first step was sorting the species into their respective components (either species-specific or functional groups – **Supplementary Table S1**). The species where harvest was already implemented in the model was removed from the list, as the aim was to identify new groups for harvesting. The resulting list represented commercial species of less economic importance not currently being harvested in the model. Based on this, we ended up with 103 components/species not currently being harvested in the model. Based on their average catches, the functional groups, small pelagic fish, large demersal and other demersals, made up 67% of the total catch of the top six groups in terms of average catches over the 30-year period represented in the data. As these three functional groups were already present in the model, we decided to use these as three of our additional groups.

In addition to these three, mesopelagic fish and mesozooplankton were added to the list of additional components that should be harvested. The reason for the last two is the increased interest in these resources, and that they most likely will be fished at a larger scale in a not too distant future (e.g., Hidalgo and Browman, 2019 and references therein).

The fishing mortality at maximum sustainable yield (Fmsy; **Table 1**) was calculated for each of the harvested components by performing multiple simulations. In these simulations, the catches of other species were held constant at current levels, whereas the catch level of the species of interest was increased until the stock collapsed. Each simulation was run for a period of 55 years, and biomass and catches were averaged over the last 10 years of each simulation. Fmsy was then the fishing mortality that gave the highest average catches at the end of the simulation while at the same time avoiding biomass collapse. The high Fmsy for mesozooplankton (4.5; **Table 1**) emerges from the high zooplankton production and corresponds to productivity levels found in e.g., the Norwecom.e2e model (Morten D. Skogen, personal communication).

The scenarios were then classified as one of four scenarios (Global sustainability, World markets, Local stewardship, and National enterprise; Groeneveld et al., 2018), each involving a

different fishing strategy. In the case of 'Global sustainability' the scenario includes lower fish and meat consumption, reductions in fishing areas and the introduction of lower impact fishing gears. The result is a reduced fishing pressure which is interpreted here as 0.6 × Fmsy. In the 'World markets' scenario, fish are obtained from the cheapest source, decommissioning reduced, few legal and technical restrictions on fishing with a greater competition for resources. This is manifested as a lower fishing pressure (0.8 × Fmsy) but not to the same extent as the previous scenario. The 'Local stewardship' scenario is considered, amongst other things, under local or regional governance of fish resources and a mosaic of different management measures. In this scenario, the result is that F occurs at MSY (1.0 × Fmsy). The last scenario (National enterprise) entails relying on national supplies with decreased imports thus a greater pressure on local stocks. This results in an increase in the fishing pressure above MSY (1.1 × Fmsy). These applications were discussed within the ICES community before being used (Katell Hamon, personal communication).

A future with both changed environmental forcing and harvesting strategies were simulated by using the downscaled RCP4.5 forcing and the four fishing pressure scenarios (0.6, 0.8, 1.0, 1.1 × Fmsy). Hence, the model was run with historical fishing pressure until 2017, from then on the Fmsy value and its multiplier was applied. The simulations were divided into two batches, those including Fmsy for commercial species only, and those including the additional five species mentioned above. For the five additional species, the fishing pressure changed from 0 to Fmsy in 2017. Capelin is a special case. This is a short-lived fish, fished almost exclusively on the spawning fraction of the stock and with total spawning mortality. The fishable biomass of capelin varies by an order of magnitude, and the concept of a "constant Fmsy" fits poorly to this stock. In high capelin years, the stock can sustain a high fishing mortality and still achieve full reproductive success, while in low capelin years any fishing level could impair future recruitment. The stock is managed with an escapement strategy for precisely this reason. Therefore, an average catch rate calculated over the last decade was applied for capelin. This fishing mortality level was multiplied with the fractions defined above. However, due to a recent collapse in the population, in addition to a large predator cod population, the fishing pressure applied to the capelin population was very low (**Table 1**).

The mesozooplankton does not have this total spawning mortality, and a fishery would likely harvest a greater variety of life stages. Furthermore, multiple species were combined into a single "mesozooplankton" stock within the model, which would make the level of specificity in the capelin case difficult to achieve. We therefore considered a simplified "constant Fmsy" appropriate for the mesozooplankton.

For each set of socio-political simulations (global sustainability, local stewardship, national enterprise, world market; Groeneveld et al., 2018), there are 28 different runs. 14 of these include 'all' species, 14 include only those currently harvested in the model system (**Table 1**). Each set of 14 simulations followed an individual pattern of mesozooplankton growth. The variability was based on a time series of mesozooplankton biomass in the Norwegian sea for the period 1995–2017 (Broms et al., 2016). The mesozooplankton growth was calculated by the mesozooplankton biomass pattern from this time series, twelve of these started in a different year (see example in **Supplementary Figure S1**), one had random variation based on the time series and the last replicate did not apply any zooplankton growth forcing at all. The reasoning for applying 13 different time series for mesozooplankton growth in the model was that complex end-to-end ecosystem models incorporate a large degree of uncertainty. Performing the simulations this way meant that an envelope of possible results was created for a larger part of the components, within which ranges we were confident. All in all, this gives us 112 simulations. We use the scenario with 1 × Fmsy as a base case run, all other runs will be compared to this one.

All results are presented using ecosystem guilds, where we have chosen to use pelagic fish, demersal fish, non-harvested lower trophic levels and harvested lower trophic levels. We also chose to split between harvested and non-harvested lower trophic levels. The reason for doing so was that these components have a tendency to increase when others at the same level decrease, hence disguising any changes.

Simple ecological and fisheries indicators were calculated for comparing the differences between the scenarios (**Table 2**). Pelagic and demersal catches, and the relationship between

TABLE 2 | Ecological indicators, explanation and abbreviations used in figures and text.


pelagic and demersal biomass are all indicators that also were used in the study by Olsen et al. (2018).

this change in harvest pressure will in itself lead to a decrease in biomass when compared to historical levels of the stocks.

## RESULTS

The fourteen different time series of zooplankton applied for each set of management scenarios created an envelope of solutions for the components in the model. This was the case for both the pelagic and the demersal fish, which ended up by being almost as vulnerable to the bottom-up effects of mesozooplankton growth. Some top predators such as cod, who have a larger spectrum of prey, were not as dependent on one single prey source. The variability at lower trophic levels following the changes in mesozooplankton time series were larger than the effects of any of the changes in the harvest regimes (**Figure 2**).

Overall, the impact of changing the harvest in the model from historical levels to the fractions of Fmsy applied in the different scenarios, was more evident in the 'all in' scenarios compared to the 'commercial' scenarios (**Figures 3**, **4**). This was the case both for the demersal and pelagic guilds. The time series of the guild shows no significant deviations between the eight management scenarios in time (**Figures 3**, **4**).

The reader should be aware that we are only applying the fishing pressure described in the socio-political scenarios, and have for simplification chosen not to follow the other details given in Groeneveld et al. (2018). Also, as several of the stocks in Norwegian waters have been fished below Fmsy in recent decades,

## Evaluation of Historical Biomass Levels

NoBa Atlantis was built using input information on biomasses, weights, abundances, distributions and other life history parameters from assessment reports, literature and gray literature (Hansen et al., 2016). When first initializing the model based on all available information, it was run toward equilibrium repeating a 1-year cycle (daily) of physical forcing (Hansen et al., 2016). After the spin-up of 24 years, the majority of the components were between 0.5 and 1.5 of the initial values (Hansen et al., 2019). Here, the set-up is changed to run the model with continuous daily physical forcing. In all simulations, we applied the historical harvest pressure for all the commercial stocks. The pelagic and demersal guilds were compared to observations from ICES assessments; WGWIDE and AFWG, respectively (**Figures 3**, **4**). The observations from 1981–2000 was used for tuning, therefore only observations from 2001 to 2015 were used for comparison. For this period, the pelagic guild biomasses fit well (**Figure 3**), but not so well in terms of timing of events, resulting in a correlation of 0.49 (p = 0.07). The demersal biomass fit well and had a much better match in terms of timing (**Figure 4**, resulting in a correlation of 0.87 (p = 0). It has to be mentioned that the observations that were used for comparison were not entirely independent, as they were a result of the harvesting regime which was also applied in the model.

## Guilds in a Future Climate and With Different Harvesting Regimes Global Sustainability (0.6 × Fmsy)

For all guilds except the pelagic and the mammals, the reduced fishing pressure (0.6 × Fmsy) that was applied in all global sustainability simulations with harvesting of commercial species led to a positive response, compared to historical levels (**Figures 2**, **5**). Introducing more components to the harvest had, however, a negative effect on the already established pelagic and demersal guilds, due to direct and indirect predator–prey effects. The decrease shown in the pelagic guild was both caused by a decline in Norwegian spring spawning herring, and an increased predation pressure from top predators. The biological indicators (**Table 2**) showed the same picture, with the biomasses of the global sustainability scenarios being highest compared to the other scenarios for the demersal, pelagic and lower trophic level harvested biomass (**Figure 5**; DemB, PelB, LTLhB, respectively). The fraction of pelagic to demersal biomass (**Figure 5**; PelDem parameter) was highest for both the 'commercial' simulations and the 'all in' simulations, compared to the three other scenarios. However, the historical period had a higher fraction than all the other projections. In the 'commercial' simulations, an increase in both the catches and biomass of the demersal guild (**Figure 5**; DemB and DemC, respectively) was seen, in contrast to a lower demersal biomass when more ecosystem components were harvested. Considering the lower trophic level biomass (**Figure 5**; LTLnhB and LTLhB parameter), this was at its lowest value for the commercial scenario, completely opposite of the 'all in' scenarios. This was a direct result of the removal of LTL, leaving a larger share of the phytoplankton for the non-harvested LTL components. This was seen in all scenarios. Compared to the historical levels, both the non-harvested and the harvested LTLs experienced an increase in the biomass in the 'commercial' scenario. The same was valid for catches of all guilds, including the lower trophic levels (**Figure 5**). Catches of demersal species were higher in the 'all in' simulations compared to the 'commercial' simulations, whereas the catches of pelagic fish decreased by 9% when more ecosystem components were harvested, despite the higher number of harvested groups included in the pelagic guild. The total catch biomass was 37 times higher in the 'all in' simulations compared to the commercial simulations (**Figure 6**).

## World Markets 0.8 × Fmsy

In a future climate, both demersal and pelagic guilds showed a decrease compared to historical levels (**Figure 2**), due to an increase in the fishing pressure compared to the global sustainability scenario (0.6 × Fmsy). The decrease in the demersal guild was around 10%, while the pelagic guild experienced a difference between 17 and 20% compared to historical levels. As in most of the other scenarios, the pelagic and demersal guild experienced a significant reduction in their biomass levels when more components were harvested

(**Figure 2**). The lower trophic levels not harvested experienced their largest biomass for this scenario, and the variability was higher for the simulations that only included the currently harvested components (**Figure 2**). The variation in the lower trophic levels canceled out any significant differences between the two sets of scenarios, indicating that the forced variation in mesozooplankton growth was as important as changes at the harvest level. The marine mammals (**Figure 5**; MamB parameter) had their highest biomass in the 'all in' simulations, although the differences between the scenarios in total marine mammal biomasses were small. The same shift in the demersal catches compared to the pelagic catches as in the global sustainability scenario was seen here, with the 'all in' simulations providing a higher demersal catch compared to the 'commercial' simulations (**Figure 5**; DemC and PelC, respectively). Likewise, the demersal and pelagic biomass (**Figure 5**; DemB and PelB, respectively) was higher in the 'commercial' simulations compared to the 'all in' simulations. In total, the pelagic catches decreased by 10% when additional species were being harvested, despite the pelagic guild now including more species. When comparing the total average catches, there is an increase of close to 49 times between the 'commercial' simulations and the 'all in' simulations (**Figure 6**).

### Local Stewardship 1.0 × Fmsy

The local stewardship (1.0 × Fmsy) scenario can also be interpreted as representing the effect of environmental changes only. Here, for the future climate, the demersal and pelagic guilds experienced a decline compared to the two previous scenarios and compared to historical levels (**Figure 2**). The difference between the simulations only including currently harvested species and those including additional species increased for the demersal guild. There was no such significant difference between species/guilds for the lower trophic levels for the two sets of simulations. However, while the scenarios that include harvest on the additional components experienced a future decline in the non-harvested lower trophic levels, compared to the global sustainability scenario, the scenarios excluding the additional components showed a small increase. Local stewardship showed a surprisingly high biomass of lower trophic levels for the 'all in' scenario (**Figure 5**; LTLhB parameter). Apart from that, it scored low on the total biomass, and high on the total catches of multiple guilds, not unexpected (**Figure 5**). At the Fmsy level, the same pattern in the catches as in the scenarios applying lower harvest mortalities was seen, with a higher demersal catch in the 'all in' simulations

FIGURE 5 | Ecological indicators for the four scenarios investigated. (A) Results from the scenarios when only commercial species are harvested. (B) Results from the scenarios when all components (commercial + five additional components) are being harvested. For the four scenarios, gs is the global sustainability (gray), ls is local stewardship (blue), ne is national enterprise (orange) and wm is world market (green). In addition, the indicators from the historical period of the simulations are added to the figures, for comparison between future projections and historical results. Biomass indicators are grouped on the left and catch indicators on the right side of the radar charts. All values increase from the center of the radar plot, to the edge.

simulated average total catches per year for the period 2005–2015 (Hist) and observed (Obs.). (B) Simulated average catches for the period 2055–2065 for the four scenarios including the additional components. Both figures show catches for the same guilds; pelagic (Pel - gray), demersal (Dem - green) and harvested lower trophic levels (LTLh - blue). Notice the difference in catch levels between panels (A) and (B), and that lower trophic level is missing from the observed bar.

compared to the 'commercial' (**Figure 5**; DemC). The pelagic catches, on the other hand, decreased by 6% when more species were harvested, despite slightly more pelagic components being harvested. The total catches in the 'all in' scenario was roughly 49 times higher than the total catches in the commercial scenario (**Figure 6**).

## National Enterprise 1.1 × Fmsy

The difference between the 'commercial' simulations and the 'all in' simulations was at its highest in the demersal guild for this scenario. The demersal guild experienced a decrease of between 15 and 17% compared to the historical levels, much lower than the 20–30% decrease seen in the pelagic guild. Both lower trophic

levels were at the lowest biomasses for this scenario, including the components not being harvested (**Figure 2**). The 'commercial' simulations showed a slight positive development in the lower trophic levels, whereas adding harvest at this magnitude on the lower trophic levels decreased their biomasses by close to 20% compared to historical levels (**Figure 2**). National enterprise showed by far the largest catches of all the harvested components of the ecosystem (**Figure 5**; DemC, PelC, LTLhC, respectively, **Figure 6**). It had a higher biomass at the lower trophic level (non-harvested) compared to both global sustainability and the local stewardship. This can be explained by the effect of both indirect and direct predator–prey interactions caused by the removal of top and mid-level predators, such as demersal and pelagic fish. There was a smaller difference between the local stewardship scenario and the national enterprise for the pelagic catches compared to the demersal catches when only the commercial species were harvested. This was not as clear in the 'all in' simulations. However, the pelagic catches decreased by 5% when additional species were harvested, even though the number of harvested pelagic species increased slightly. The average catches in the commercial simulations were more than 50 times lower than in the 'all in' simulations for this particular scenario (**Figure 6**).

## DISCUSSION

In this study, we explored the ecosystem effects of changes in management regimes for eight different scenarios in two environmental settings (historical and future climate). In each of the scenarios, multiple simulations were performed to introduce bottom-up variability by forcing the growth rate of mesozooplankton. All the scenarios applied different fractions of Fmsy mortality for the major commercial fisheries in the Norwegian and Barents seas. In four of the eight scenarios, additional components were harvested. The additional species were chosen based on catch statistics (small pelagics, large demersals and other demersals) and increased commercial interest in the component (mesopelagic fish and mesozooplankton).

## Model Uncertainty

Structural uncertainties in large end-to-end ecosystem models can be quantified in different ways, for instance by running ensembles of models, or by testing multiple model set-ups within the same model framework (Lehuta et al., 2016). There are other end to end models representing the Norwegian and Barents seas, for example the NORWegian ECOlogical Model system End-To-End (NORWECOM.E2E), a coupled physical, chemical, biological model system (Skogen et al., 2007), covering primary and secondary producers as well as key fish species. The model has been validated by comparison with field data in the Nordic and Barents seas (Skogen et al., 2007; Hjøllo et al., 2012; Utne et al., 2012). The NORWECOM.E2E model has been run using the same physical forcing as in this study, and showed a slight future decline in some of the pelagic fish components (herring), using the present level of the fishery (S. S. Hjøllo, personal communication), and a slight increase in the zooplankton component. Having the same trend in two so different ecosystem model supports the findings in this study. Uncertainties within the model system originates from multiple sources, e.g., parameter settings, initial conditions, simplifications of ecological processes and the application of fisheries. Given the complexity and computational costs of running NoBa, we eventually ended up with the 14 simulations for each scenario. As we already were aware of the models sensitivity to mesozooplankton (Hansen et al., 2019), we were confident that adding this variability creates reasonable envelopes around the components in the model, something also shown in **Figures 3**, **4**. For further studies, uncertainties connected to recruitment and/or biomass levels of the three potentially most important pelagic species in the systems (polar cod, capelin and Norwegian spring spawning herring) could be explored.

## Ecosystem Effects of Changes in Management and Climate

The cumulative effect of fisheries and climate has, for a majority of studies, showed declines in catches and productivity in commercially important stocks (see e.g., Cheung et al., 2016; FAO, 2018b), although this also depends on the physical projection applied. While climate change might decrease the production in some regions of the oceans, the world is also facing a growing requirement for food (United Nations, Department of Economic and Social Affairs, Population Division, 2019). This is a direct result of a growing human population and the world health goal of zero hunger, indicating that there will be continued demand for increasing catches from the marine fisheries as the world population continues to grow. Enhancing the number of components and also spreading the harvest across multiple trophic levels has been suggested as one solution (Sethi et al., 2010; Garcia et al., 2012; Kolding et al., 2016). However, going to a fully balanced harvest might not be the preferred solution (Howell et al., 2016; Nilsen, 2018).

The Barents and Norwegian Seas ecosystems are among the more balanced harvested ecosystems of the world, harvesting on most trophic levels from copepods (commercial quota given from 2019) to marine mammals (Howell et al., 2016). Increasing the harvest on lower trophic levels can introduce unforeseen predator–prey effects, lowering the catches and/or economic yield at higher trophic levels (Smith et al., 2011). Despite not being a surprise based on previous sensitivity studies of the model system (Pantus, 2006; Hansen et al., 2019), it was interesting to note the strong effect of implementing an Fmsy fishery for the mesozooplankton and mesopelagic fish. Not only was the effect evident in the increase in total catches, but also in the decrease in biomass of other components in the system due to direct and indirect predator– prey interactions. Even though the sensitivity of the model to the zooplankton components might be a structural effect of the Atlantis framework (Hansen et al., 2019), the importance of mesozooplankton (in particular Calanus finmarchicus) as a food source for pelagic fish in the Norwegian Sea is unquestionable (Bachiller et al., 2016).

Contrary to (Smith et al., 2011), where harvesting on forage fish had a strong impact on parts of the ecosystem, the Norwegian and Barents seas ecosystems seem to handle fisheries targeting pelagic fish at Fmsy well. However, there is an important piece that is missing in Fmsy scenarios applied here, namely the capelin. For all eight scenarios, this component is fished rather lightly. Capelin, being short-lived and semelparous, does not translate into an Fmsy easily, and it can be debated whether or not it is possible, or at all correct, to calculate an Fmsy for this component. Based on this we chose to leave the catches at a representative average for the last decade. Due to the total catch allowances over recent years being low, corresponding to the high biomass of Northeast Arctic cod (Gjøsaeter et al., 2015), this resulted in a low harvest rate of the capelin, leaving it rather unresponsive to changes in the fishing pressure. Despite this missing piece, there is an increasingly negative response in catches of the demersal guild with increasing fishing pressure (from 0.6 to 1.1 × Fmsy) in the 'all in' simulations compared to the commercial simulations. This indicates a growing vulnerability in the whole system when a larger number of components are being harvested. It is possible that this increasing vulnerability would be even stronger if the capelin had been harvested at a higher rate. This is based on empirical observations showing severe ecosystem effects caused by simultaneously low biomasses of pelagic fish and zooplankton in the Barents Sea (Gjøsaeter et al., 2009). For future studies, we suggest that the harvest pressure on capelin should be implemented using a version of the escapement rule that leaves at least 200,000 t of capelin to spawn (Gjøsaeter et al., 2015).

With the large changes in the management strategies for the historical and future projections, it was particularly difficult to track down changes related to increasing temperatures alone. However, **Figure 5** showed an increase in biomasses at the lower trophic levels for 2055–2065 compared to the historical time slot, regardless of fishing pressure at the higher trophic levels. This is in agreement with findings in other papers exploring climate change in the high latitudes (e.g., Steinacher et al., 2010), although model systems applying different physical forcing have reported declines in primary production (Slagstad et al., 2015). Both these studies explain their differences based on changes in the environment, such as reduced light and temperature limitation and advection. Unfortunately, RCP 4.5 was the only down-scaled scenario available for the Nordic and Barents Seas. Skogen et al. (2018) showed that downscaling of low-resolution global models to regional models is important for this area, which is the reason why we chose not to use a global model representing any of the other RCPs. For future work, we strongly recommend that multiple physical scenarios should be available, as this will provide the opportunity to differentiate between the effect of climate change alone compared to the impact of changes in management.

## Impact on Catches Following Changes in Management and Climate

Calculating the Fmsy levels for the additional components resulted in the same value for several of the non-commercial stocks (**Table 2**). This might emerge from these components not being as thoroughly evaluated and parameterized as the large, commercially important stocks. For the commercially important stocks there are much more data available. These conclusions are also supported by Nilsen et al. (this issue), who found that the harvest levels of these components were generally lower than for the commercially important stocks, although the production levels one would expect from these components should not be very different. From **Table 1**, the Fmsy level for saithe, beaked redfish and Greenland halibut diverges from the historical levels of fishing mortalities.

Applying an Fmsy on the mesopelagic fish and mesozooplankton is relatively unrealistic, due to both the technical aspects with these fisheries, and the role of these species as key prey in the ecosystem. Based on the Commision for the Conservation of Antarctic Marine Living Resources work, a precautionary approach is applied to set the quota for the mesozooplankton in the Norwegian Sea. This gives a total allowable catch of 165 000 tonnes from a total biomass estimated to be above 30 million tonnes, and it also includes area restrictions on where the harvest can take place. The area restrictions are meant to prevent issues with bycatch of eggs and larvae of other species. The aspect of bycatch is a concern with both mesopelagic fish and copepods, and was thoroughly discussed when the management plan for Calanus finmarchicus in the Norwegian Sea was developed (Broms et al., 2016). The main concern in this sense is the bycatch due to mesh size and target areas. We consider the technical details on these fisheries outside the scope of our study and have therefore only considered the ecosystem effects of fishing mortality as such.

The Fmsy multipliers used in this study were taken from the predefined CERES scenarios (Groeneveld et al., 2018). Singlespecies assessments are, to a large degree, managed with the aim of applying a Fmsy (ICES, 2018a). However, there have been studies showing that applying multiple single-species Fmsys simultaneously is not ecosystem-friendly (e.g., Gaichas et al., 2012), as the cumulative pressure of each of these might result in a total harvest above the threshold of the system. This is explained by predator–prey interactions between the species, which are not taken into account in assessments for the majority of the stocks worldwide (Walters et al., 2005;. Skern-Mauritzen et al., 2016). Although we do not experience a collapse in the system when applying multiple Fmsy simultaneously, we do notice a decrease in both the pelagic and demersal guilds when the Fmsy's are introduced (**Figures 3**, **4**). We explain this by the combined effect of introducing a higher F for some of the species, in combination with the decrease in biomass at lower trophic levels for half of the scenarios. In addition, there is a significant difference between applying a flat Fmsy on a stock, compared to the harvest control rules (HCR) that presently are a part of a majority of the commercial fisheries in the Norwegian and Barents seas. While an HCR possibly could prevent the collapse of the herring stock (not shown) that was experienced here, this was not the case with the invariant Fmsy.

Extending the number of harvested components, increased the total catches significantly due to the biomass at the lower trophic levels. This resulted in decreasing catches of

the traditionally targeted pelagic and demersal guilds. The decrease was not dramatic, 6–10%, but it must be kept in mind that for these scenarios, the number of harvested species also increased for these guilds. Isolated, we would therefore expect a total increase, not a decrease. However, the catches of mesozooplankton increase with several magnitudes, introducing strong bottom-up effects. For the demersal guild, there was a clear connection between increased harvest pressure, increased number of components being harvested, and a decline of demersal catches. Comparing the historical (2005–2015) and the future (2055–2065) biomass, we found a clear decline in the future biomass of pelagic, demersal and LTL-harvested ('allin' scenarios) components, caused mainly by the change in harvest regimes.

## CONCLUSION

There are multiple important tradeoffs being identified within and between the management scenarios described here. The most emergent ones are related to the fishing of mesozooplankton and mesopelagic fish. Here, the catches of pelagic and demersal guilds decrease whereas the new fisheries at lower trophic levels can increase a lot. It also has to be kept in mind that the majority of the catches from the pelagic and demersal guilds (in particular the demersal) are being used directly for human consumption, whereas the newer lower trophic level catches with present day consumer behavior cannot fill that role. However, their use for food in aquaculture might be important in terms of food security.

Large end-to-end models such as NoBa should never be used for tactical management (Link et al., 2010; Fulton et al., 2014). However, they can be useful and informative for exploring ecosystem responses to cumulative changes as, for example, in management regimes and fisheries and environmental conditions. As the oceans are facing increasing pressure from global warming and food production industry, we need to explore the possible effects of changes in our traditional fisheries management. Here, using one realization (NorESM RCP4.5 scenario downscaled to regional resolution with ROMS) out of a range of possible future climate projections, we conclude that increasing harvest on lower trophic levels could be beneficial in terms of total catches, but multiple trade-offs need to be considered and discussed across the different sectors involved in both the traditional fisheries and potential new fisheries, before such management changes are implemented.

## REFERENCES


## DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request to the corresponding author.

## AUTHOR CONTRIBUTIONS

CH performed the simulation, created the figures, and is main author of manuscript. SH provided invaluable discussions on results and significant contributions to the manuscript and figures. RN and KD provided invaluable discussions on results and significant contributions to the manuscript.

## FUNDING

SH acknowledges support from the European Union's Horizon 2020 Research and Innovation Programme under grant agreement no. 677038 (ClimeFish). CH, RN, and KD acknowledge support from the European Union's Horizon 2020 Research and Innovation Programme under grant agreement no. 678193 (CERES).

## ACKNOWLEDGMENTS

We thank Dr. Daniel Howell (Institute of Marine Research, Bergen, Norway) for comments on a draft of this manuscript. We would also like to thank the three reviewers and the editor for all their effort and their valuable critique of the manuscript.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00668/full#supplementary-material

FIGURE S1 | Normalized zooplankton biomass time series from (in black) the unforced replicate, and (in orange, whole line) one of the 13 replicates that is forced with the time series of zooplankton growth. The dotted orange line is the fraction that eventually was multiplied by zooplankton growth and applied in the model system to create variability for each scenario.

TABLE S1 | Overview over functional groups and species included in NoBa Atlantis.


catch potential. Glob. Chang. Biol. 16, 24–35. doi: 10.1111/j.1365-2486.2009. 01995.x



**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.

The reviewer NJ declared a past co-authorship with one of the authors CH to the handling Editor.

Copyright © 2019 Hansen, Nash, Drinkwater and Hjøllo. 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.

# Iron Availability Influences the Tolerance of Southern Ocean Phytoplankton to Warming and Elevated Irradiance

Sarah M. Andrew<sup>1</sup> \*, Hugh T. Morell<sup>1</sup> , Robert F. Strzepek<sup>2</sup> , Philip W. Boyd2,3 and Michael J. Ellwood<sup>1</sup>

<sup>1</sup> Research School of Earth Sciences, Australian National University, Canberra, ACT, Australia, <sup>2</sup> Antarctic Climate and Ecosystems Cooperative Research Centre, University of Tasmania, Hobart, TAS, Australia, <sup>3</sup> Institute for Marine and Antarctic Studies, University of Tasmania, Hobart, TAS, Australia

#### Edited by:

Elizabeth A. Fulton, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia

#### Reviewed by:

Matthew McGinness Mills, Stanford University, United States Katherina Petrou, University of Technology Sydney, Australia

> \*Correspondence: Sarah M. Andrew sarah.andrew@unc.edu

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 04 April 2019 Accepted: 21 October 2019 Published: 01 November 2019

#### Citation:

Andrew SM, Morell HT, Strzepek RF, Boyd PW and Ellwood MJ (2019) Iron Availability Influences the Tolerance of Southern Ocean Phytoplankton to Warming and Elevated Irradiance. Front. Mar. Sci. 6:681. doi: 10.3389/fmars.2019.00681 The Southern Ocean is responsible for approximately 40% of oceanic carbon uptake through biological and physical processes. In the Southern Ocean, phytoplankton growth is limited by low iron (Fe) and light supply. Climate model projections for the Southern Ocean indicate that temperature, underwater irradiance and Fe supply are likely to change simultaneously in the future due to increasing anthropogenic carbon dioxide emissions. The individual effects of these environmental properties on phytoplankton physiology have been extensively researched, and culturing studies using Southern Ocean phytoplankton have shown that temperature and Fe will play a key role on setting growth under future conditions. To explore the potential responses of Southern Ocean phytoplankton to these environmental changes, we cultured the haptophyte Phaeocystis antarctica and the diatoms Chaetoceros flexuosus, Proboscia inermis, and Thalassiosira antarctica under two light and iron combinations and over a range of temperatures. Our study revealed that the thermal response curves of key Southern Ocean phytoplankton are diverse, with the highest growth rates measured at 5 ◦C (the annual temperature range at the isolation sites is currently 1–4◦C). Warming had species-specific effects on the photochemical efficiency of photosystem II (PSII; Fv/Fm), the functional absorption cross-section of PSII (σPSII), carbon:nitrogen ratio and cellular Chlorophyll a concentrations. Iron availability increased species' ability to tolerate warmer conditions by increasing the upper limit for growth and subsequently increasing the thermal niche that each species inhabit.

Keywords: temperature, climate change, photosynthesis, evolution, multiple stressors, carbon

## INTRODUCTION

Phytoplankton productivity in the Southern Ocean plays an important role in the transfer of carbon from the atmosphere to the ocean's interior, in a process called the biological carbon pump. The strength of the biological carbon pump is important in regulating global climate. Southern Ocean productivity, in turn, is regulated by the availability of iron (Fe), light, and temperature, which influence the efficiency of the carbon pump (Sunda and Huntsman, 1997; Boyd et al., 2010).

The effect of temperature on model diatom species has been well documented (Sunda and Huntsman, 2011); however, the interaction between temperature, light, and Fe on phytoplankton growth rate has only recently been explored in Southern Ocean species (Zhu et al., 2017; Boyd, 2019). Studies of subantarctic phytoplankton have identified Fe and temperature as key controls on phytoplankton growth with light, macronutrients and CO<sup>2</sup> playing a lesser role (Boyd et al., 2016). Future warming of the Southern Ocean is expected to shift thermal niches poleward resulting in an associated shift in the biogeographical range of species, as they accommodate environmental changes (Thomas et al., 2012). Limited information is available on the thermal tolerance of specialized Southern Ocean phytoplankton, especially when temperature varies concurrently with other environmental variables (Boyd et al., 2013; Coello-Camba and Agustí, 2017).

Phytoplankton generally have an optimum growth temperature above the average mid-summer water temperature in which they grow, thus protecting them against short-term temperature fluctuations (Thomas et al., 2012). The growth rate of phytoplankton generally increases with temperature, until an optimum temperature is reached (Eppley, 1972). Once this temperature optimum has been exceeded, growth rate decreases and eventual mortality occurs (Kudo et al., 2000; Boyd et al., 2013; Zhu et al., 2017). Unlike tropical phytoplankton, which are already at or near at their thermal capacity, cold-adapted phytoplankton display optimum growth temperatures higher than the temperature of their current environment (Thomas et al., 2012). Thus, Southern Ocean phytoplankton may have a thermal safety net that will allow them to withstand the expected global warming associated with increasing CO<sup>2</sup> concentrations.

In addition to temperature, Fe supply plays a key role in controlling productivity in the Southern Ocean (Boyd and Law, 2001; Blain et al., 2007). The biochemical importance of Fe in photosynthesis has been demonstrated by laboratory experiments (Greene et al., 1991, 1992; Sunda and Huntsman, 1997; Strzepek et al., 2011) and extensively explored through theoretical calculations by Raven (1990). The greatest metabolic requirement for Fe in phytoplankton is photosynthetic electron transport (Strzepek and Harrison, 2004). Furthermore, Rose et al. (2009) showed that while warming increased phytoplankton productivity, there is also evidence for an increased Fe demand and an earlier onset of nutrient and Fe limitation when compared to a lower temperature control group.

Future changes in net productivity under climate change scenario RCP 8.5 were explored by Laufkötter et al. (2015), who found large discrepancies in primary productivity between simulations. Consequently, these large uncertainties suggest that a dedicated and sustained effort should be undertaken to provide greater certainty in models by providing quality datasets that underpin model development. Defining the thermal tolerances of marine biota from a range of latitudes is central to improved biogeochemical modeling as most species have important roles in the cycling of nutrients (nitrate, silicic acid, phosphate), trace element cycling (e.g., zinc, copper, and Fe), and CO<sup>2</sup> sequestration and export to the oceans' interior. Additionally, temperature-induced floristic changes will have wide-scale impacts on global food webs. For example, Southern Ocean diatoms support most krill-based food webs (Feng et al., 2010); thus a shift in the community composition away from diatoms could have negative consequences for higher trophic levels.

The goal of this work is to identify the thermal thresholds of phytoplankton isolated from the Southern Ocean and provide basic physiological data on how Southern Ocean phytoplankton may respond to predicted ocean warming. We report the physiological responses of three Southern Ocean diatoms (Chaetoceros flexuosus, Proboscia inermis, and Thalassiosira antarctica), as well as the Southern Ocean haptophyte Phaeocystis antarctica, grown in low and high Fe and light treatments, and over a range of temperatures (3–14◦C). Because the diatoms used in this study were isolated from the same locale, we hypothesize that they will share similar responses to light, Fe and temperature due to the unique photosynthetic specializations needed to grow in this area (Strzepek et al., 2019).

## MATERIALS AND METHODS

## Culturing Conditions and Temperature Manipulation

The haptophyte P. antarctica (Clone SX9) was isolated from water collected in the Australasian sector of the Southern Ocean (62◦ 08.72<sup>0</sup> S and 174◦ 08.940E) in December 2004 (Strzepek et al., 2012). The Southern Ocean diatoms grown for this research were C. flexuosus, P. inermis, and T. antarctica; all of which were isolated from seawater collected in November 2001 from 57◦ 51.10<sup>0</sup> S and 139◦ 50.700E at a temperature of 3◦C (**Figure 1**; Strzepek et al., 2011).

The artificial seawater medium Aquil (Price et al., 1989) was used to culture Southern Ocean phytoplankton. The Aquil medium was microwave-sterilized and enriched with filtersterilized (0.2 µm, Gelman Acrodisc PF) trace metals and

FIGURE 1 | Map showing isolation temperature and location of strains used in this study.

vitamins, and chelexed macronutrients (nitrate = 300 µmol L−<sup>1</sup> , silicate = 100 µmol L−<sup>1</sup> , phosphate = 10 µmol L−<sup>1</sup> ). Previously reported experiments were used to select high Fe (Fe-replete) versus low Fe (Fe-limiting) conditions (Strzepek et al., 2011) and are listed in **Supplementary Table 1**. For the high Fe treatments, the synthetic ligand Ethylenediaminetetraacetic acid (EDTA) was used to buffer Fe and the other trace metals (e.g., Zn, Cu, Ni, etc.) in solution. The final EDTA concentration within the media was 10 µmol L−<sup>1</sup> . Iron was added to the media in a 1:1 complex with EDTA to prevent Fe precipitation. The final Fe concentration of the high Fe treatments was either 58 nmol L−<sup>1</sup> or 4.4 nmol L−<sup>1</sup> . Values for [Fe<sup>0</sup> ] were calculated using Visual MINTEQ (version 3.0; default thermodynamic database) and the calculations of Sunda and Huntsman (2003) as described in Strzepek et al. (2012).

For the low Fe treatments, the siderophore desferrioxamine B mesylate (DFB) was added to culture media because of its ability to strongly complex to Fe. Low Fe media were prepared using a premixed solution containing 3.5 nmol L−<sup>1</sup> FeCl<sup>3</sup> complexed with 4, 40, or 400 nmol L−<sup>1</sup> of the siderophore DFB and added to Aquil medium containing 10 µmol L−<sup>1</sup> of EDTA to buffer the other trace metals as described in Strzepek et al. (2011). The low Fe media were designed to reduce phytoplankton growth rates by ∼50% at growth saturating irradiance; therefore, the Fe:DFB ratios (nmol L−<sup>1</sup> : nmol L−<sup>1</sup> ) of the Aquil media differed between species: P. antarctica – Fe:DFB 3.5:400; C. flexuosus and P. inermis – Fe:DFB 3.5:40; and T. antarctica – Fe:DFB 3.5:4.

## Growth Rate Measurements

Cultures were maintained in exponential growth with a 1:100 dilution by transfer into fresh media as required. Growth rate data were collected from cultures acclimated to experimental temperature, irradiance or Fe treatments for a minimum of 15 generations. Each transfer was grown for at least 10 generations. In vivo chlorophyll a fluorescence of acclimated cultures was measured using a Turner Designs model 10-AU (Brand et al., 1981) and specific growth rates (µ) were determined from leastsquares regressions of ln in vivo fluorescence versus time during the exponential phase of growth.

## Seawater Temperature Conditions

Current surface seawater temperatures for the Southern Ocean were obtained from satellite data<sup>1</sup> and match the temperature at which each species was isolated (**Figure 1**). To estimate future surface sweater temperatures in the Southern Ocean, we increased temperature by 2◦C, in line with simulated predictions by Boyd et al. (2015) and Ito et al. (2015) and predicted by the RCP8.5 climate change scenario estimated in the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC; Stocker et al., 2013).

In order to mimic the warming that polar phytoplankton might encounter over the coming century, Southern Ocean cultures were exposed to higher temperatures in a stepwise fashion over 2–4 months. After acclimation to each temperature, a subset of cultures was moved to a higher temperature, allowing at least ten generations of growth at each temperature. The cultures were subsequently acclimated to experimental conditions for a further 4 weeks before the experiment was initiated, and growth responses recorded. The higher temperature experimental cultures in this study were grown at 8, 10, 12, and 14◦C in 3 L water baths under a mean continuous irradiance of 17 ± 5 µmol photons m−<sup>2</sup> s <sup>−</sup><sup>1</sup> or 90 ± 10 µmol photons m−<sup>2</sup> s <sup>−</sup><sup>1</sup> using LED lights. The temperature of each of the water baths was regulated using a PID temperature controller with a k type thermocouple thermometer. The lower temperature experimental cultures were grown at 3 and 5◦C in an incubator under a continuous irradiance of 20 ± 5 µmol photons m−<sup>2</sup> s −1 (low light treatment), 90 ± 5 µmol photons m−<sup>2</sup> s −1 (saturating light treatment), or 200 ± 10 µmol photons m−<sup>2</sup> s −1 (high light treatment); using Philips Alto II fluorescent tubes (**Supplementary Table 1**). The spectra for both light sources were measured using an Ocean Optic USB4000 spectrometer. Both light sources span the same wavelength range with a similar intensity. Growth saturating irradiances were determined in previous studies (Strzepek et al., 2012; Boyd et al., 2013; Boyd, 2019).

## Cellular Physiology

Cell dimensions and culture density for C. flexuosus, P. inermis, and T. antarctica were determined by microscopy and cell volumes were calculated assuming a cylindrical geometry. Triplicate samples (1 mL) for cell diameter and culture density for P. antarctica were determined by Coulter Counter <sup>R</sup> (Model MS4) and cell volumes were calculated assuming a spherical geometry. This information was obtained from mid-exponential phase cultures at the same time as cells were collected for in vitro Chlorophyll a, cellular carbon and nitrogen, and Fast Repetition Rate fluorometry (FRRf) analyses.

## Particulate Organic Carbon/Particulate Organic Nitrogen

Samples for cellular carbon (POC) and nitrogen (PON) were collected by filtering 25 ml of cells on to pre-combusted 13 mm GF/F filters (Merck Millipore). All filter holders and funnels were washed with 10% HCl, rinsed with deionized (Milli-Q <sup>R</sup> ) water and then dried before use. The filters were then placed in sterile plastic wells and left to dry in an oven at 50◦C for 2 weeks. The filters were then wrapped and stored at −80◦C. A Sercon-Callisto CF-IRMS stable isotope analysis system was used for total organic carbon and organic nitrogen determination.

## Chlorophyll a Analysis

Cells were collected for Chlorophyll a (Chl a) analysis from midexponential phase of growth. Cells (25–50 mL) were collected under a low vacuum (<100 mm Hg) onto 25 mm glass fiber filters (Whatman GF/F), rinsed three times with synthetic ocean water to ensure all cells were collected onto the filter and then stored frozen at −20◦C in 15 mL centrifuge tubes. Pigments were extracted in 10 mL of 90% acetone and Chl a concentrations determined by in vitro fluorometry (Wright et al., 2005).

<sup>1</sup>https://coastwatch.pfeg.noaa.gov/

## Fast Repetition Rate Fluorometry

fmars-06-00681 October 31, 2019 Time: 17:34 # 4

Photochemical measurements of PSII were determined using a LIFT-FRR Fluorometer (Soliense, United States). After low light (∼15 µmol photons m−<sup>2</sup> s −1 ) acclimation for ∼30 min, samples were exposed to 140 flashes of 470 nm light every 2.5 µs (saturation sequence) in order to saturate photosystem II (PSII) and the first stable electron acceptor, QA. The time interval between flashes was then increased exponentially (relaxation sequence) for 90 flashes to relax PSII and determine the rate of electron transport between the first stable electron acceptor, QA, through to Photosystem I (PSI). Fv/F<sup>m</sup> and σPSII were determined from the mean of 200 iterations of the fluorescence induction and relaxation protocol.

## Statistical Analyses

Statistical analysis was completed with R version 3.4. A threeway ANOVA was used to analyze statistical differences. Factors were Fe with two levels (Fe<sup>+</sup> and Fe−), light with three levels (200 µmol photons m−<sup>2</sup> s −1 , 90 µmol photons m−<sup>2</sup> s −1 and 20 µmol photons m−<sup>2</sup> s −1 ) and temperature with six levels (3, 5, 8, 10, 12, and 14◦C; **Supplementary Tables 4**, **5**). To obtain specific information for each temperature, light and Fe condition a pairwise comparison of variables was undertaken using the Tukey post hoc test. All testing was done at the 95% confidence level.

## RESULTS

This study outlines the responses of four Southern Ocean species to multiple stressors or drivers, namely Fe, light, and temperature. Due to the complex interaction between experimental treatments, the temperature responses of each species will be separated into two sections. The first section will examine the thermal tolerance curves of the haptophyte P. antarctica, and the two diatoms C. flexuosus and T. antarctica grown under sub-saturating and saturating irradiances from 3 to 14◦C in both high Fe and low Fe media (**Supplementary Table 2**). The second section compares the responses of four species grown at present-day (isolation temperature = 3◦C) and future (projected 1.5–2◦C increase in the Southern Ocean = 5◦C) temperature conditions combined with variations in Fe and light availability. This section will also focus on the impact of temperature on growth rates and changes in cell physiology coupled with measurements of photochemical physiology (**Supplementary Table 3**).

## Temperature Response Curves Under Combined Fe and Light Limitation

Consistent with our hypothesis that all species will respond in a similar way to warming, P. antarctica, C. flexuosus, and T. antarctica all responded positively to temperature increases up to 5◦C, after which growth rates declined with subsequent temperature increases. This optimal growth temperature was observed at 5◦C under both sub-saturating and saturating light (**Figure 2**). Despite having a temperature optimum comparable to the other species, C. flexuosus exhibited a different response to low Fe treatments when temperature was increased under saturating light. While P. antarctica and T. antarctica growth ceased at temperatures above about 10◦C, and a greater tolerance to increased temperature was observed in high Fe cultures, C. flexuosus was able to withstand warming up to 12◦C regardless of Fe supply.

Phaeocystis antarctica cultures grew more rapidly under saturating irradiance under high Fe conditions, with growth approximately 2-fold higher than when grown under low light conditions (p < 0.05; **Figures 2A,D**). Under low Fe conditions the growth rate for P. antarctica declined sharply under high light conditions beyond a temperature of 5◦C. Under low light conditions at comparable iron and temperature conditions growth declined but not as dramatically, at 8◦C growth rates were ∼0.1 ± 0.01 d−<sup>1</sup> under low light and low Fe conditions, compared to the high light and low Fe treatment at 8◦C where growth rates were reduced to ∼ 0.02 ± 0.01 d−<sup>1</sup> .

Chaetoceros flexuosus was unable to grow at 14◦C when grown under high irradiance in both high Fe and low Fe conditions (**Figures 2B,E**). Under sub-saturating irradiance, C. flexuosus was only grown across a temperature range of 3–8◦C. At temperatures above 5◦C, Fe limitation was alleviated as growth rates for the low Fe and high Fe treatments converged (**Figure 2E**). C. flexuosus cultures grew more rapidly under saturating irradiance and under high Fe conditions, with growth approximately twofold higher at 5◦C (0.46 ± 0.05 d−<sup>1</sup> ) compared to the low light treatment (0.29 ± 0.03 d−<sup>1</sup> , p < 0.05; **Figures 2B,E**).

In contrast to the two other species, T. antarctica displayed an inconsistent growth response to the decrease in Fe concentration. Instead of the expected reduction of growth rate with the decrease of inorganic Fe concentration (Fe<sup>0</sup> ), T. antarctica tended to have a higher growth rate when the Fe<sup>0</sup> was reduced from 100– 500 pmol L−<sup>1</sup> (FeEDTA medium) to 3.7 pmol L−<sup>1</sup> (FeDFB medium). This positive growth response to the reduction in Fe<sup>0</sup> upon DFB addition to media in T. antarctica was observed under high and low irradiances (p < 0.05; **Figures 2C,F**). T. antarctica cultures were unable to grow at 12◦C under high light, however, mortality of low Fe (FeDFB media) T. antarctica cultures was not realized until 14◦C when grown under low light.

Contrary to our hypothesis that species will share the same response to warming, trends in the thermal performance of PSII (Fv/F<sup>m</sup> and σPSII) differed between species. The photochemical efficiency (Fv/Fm) measurements of P. antarctica revealed different trends in treatments grown under high and low light, however, these differences due to changes in light are not significant (p > 0.05; **Figure 3**). Generally, in P. antarctica, Fv/F<sup>m</sup> was unchanged between 3 and 5◦C; above this temperature, Fv/F<sup>m</sup> declined with increasing temperature (p < 0.05; **Figure 3D**). In contrast, Fv/F<sup>m</sup> values for C. flexuosus did not change with increasing temperature when grown under high light (p > 0.05). Fv/F<sup>m</sup> values for C. flexuosus grown under low light showed a subtle increase with increasing temperature (**Figures 3B,E**). In contrast to these two species, the Fv/F<sup>m</sup> of T. antarctica cultures was highest under low light conditions and responded significantly to the individual and combined effects of light, Fe and temperature (p < 0.05; **Figure 3F**).

The functional absorption cross-section of PSII (σPSII) also varied between species in response to temperature light and

FIGURE 2 | Thermal tolerance curves of P. antarctica (A,D); C. flexuosus (B,E); and T. antarctica (C,F); grown under low and high Fe concentrations at (A–C) growth saturating light (90 µmol photons m−<sup>2</sup> s −1 ) or (D–F) growth limiting light levels (20 µmol photons m−<sup>2</sup> s −1 ). Errors are standard deviation, n = 3.

Fe (**Figure 4**). High Fe concentrations decreased σPSII in all species except P. antarctica, as did increasing light intensity (p < 0.05). For cultures of P. antarctica, σPSII increased with increasing temperature (p < 0.05; **Figures 4A,D**). In contrast, σPSII for C. flexuosus remained constant under high light when temperature increased in both low and high Fe cultures (p < 0.05), but increased ∼2-fold in low light, low iron cultures in response to temperature (**Figure 4E**). In all treatments, σPSII for

FIGURE 4 | Functional absorption cross-section of PSII (σPSII) of P. antarctica (A,D); C. flexuosus (B,E); and T. antarctica (C,F); grown under Fe limited and Fe replete conditions at (A–C) growth saturating light (90 µmol photons m−<sup>2</sup> s −1 ) or (D–F) growth limiting light levels (20 µmol photons m−<sup>2</sup> s −1 ). Errors are standard deviation, n = 3.

T. antarctica decreased with increasing temperature (p < 0.05; **Figures 4C,F**). Regardless of experimental conditions, σPSII was large, as previously observed for Southern Ocean species (Strzepek et al., 2012, 2019).

## Interactive Effects of Fe, Light, and Temperature on Southern Ocean Phytoplankton at 3 and 5◦C

In vivo measurements of Chl a indicated uniform growth rates for all species over the 24 months the work was undertaken. These rates are largely comparable to those previously reported for treatments with each species grown under similar conditions (Strzepek et al., 2011, 2012; Boyd et al., 2013), except for T. antarctica which unexpectedly responded by increasing their growth rate in Fe deficient media (**Figure 5**). In this section, the growth responses of P. antarctica, T. antarctica, and C. flexuosus will be directly compared to each other and to P. inermis at ecologically relevant temperatures: isolation temperature of 3◦C and a future warming scenario of 5◦C, at high (200 µmol photons m−<sup>2</sup> s −1 ) and low (20 µmol photons m−<sup>2</sup> s −1 ) light.

The highest growth rates for P. antarctica, P. inermis, and C. flexuosus were observed in cultures grown at 5◦C under high light and Fe replete conditions (0.67 ± 0.05, 0.71 ± 0.02, and 0.66 ± 0.03 d−<sup>1</sup> , respectively; **Figure 5**). In comparison, at 3◦C growth rates for these species were significantly lower (p < 0.05) with values: 0.43 ± 0.02, 0.39 ± 0.02, and 0.38 ± 0.05 d−<sup>1</sup> for the same Fe and light conditions. P. inermis was the only species observed to have growth negatively affected by increased

phytoplankton species under high (5◦C) and low (3◦C) high temperature, high (200 µmol photons m−<sup>2</sup> s −1 ) and low (20 µmol photons m−<sup>2</sup> s −1 ) light, grown in high and low Fe media. Errors are standard deviation, n = 5.

temperature in some treatments (**Figure 5**). Under low light conditions at 5◦C, the growth rate for P. inermis decreased in both the high Fe and low Fe treatments in comparison to the low temperature high Fe and low Fe treatments. The optimum

growth rates for T. antarctica were observed in cultures grown in low Fe media (p < 0.05). This unexpected growth rate under low Fe was tested across multiple batches of FeEDTA (high Fe) media and FeDFB (low Fe) media at 3◦C over a 24-month period with consistent results across experiments.

Increasing temperature from 3 to 5◦C generally increased C:N ratio across all treatments in P. antarctica, P. inermis, and C. flexuosus but not T. antarctica (p < 0.05; **Figure 6A**). In contrast, C:N mostly decreased or remained the same when temperature increased in T. antarctica. Warming elevated C:N in T. antarctica when grown under low light and high Fe treatment, however, this relationship was not significant (p > 0.05). C:N ratio was also found to be significantly correlated with Fv/F<sup>m</sup> and Chl a in cultures of C. flexuosus (p < 0.05).

The three fastest growing species in this study, P. antarctica, P. inermis, and C. flexuosus displayed similar relationship in their Chl a concentration per cell volume. The highest Chl a concentrations were observed for the 5◦C low light high Fe treatments (p < 0.05; **Figure 6B**). The lowest Chl a concentrations were generally observed for the high light and low Fe treatments. Increasing temperature generally increased Chl a concentrations regardless of light or Fe supply, for P. antarctica, P. inermis, C. flexuosus, and T. antarctica with a couple of exceptions (**Figure 6B**). Chl a concentrations of T. antarctica followed a similar pattern to the other study species despite the high growth rate observed in low Fe medium. Increasing temperature under low light and low Fe resulted in a decrease in Chl a concentration for P. antarctica and P. inermis (p < 0.05).

Fv/F<sup>m</sup> values for P. antarctica, P. inermis, and C. flexuosus were generally elevated in high light and high Fe treatments (at 3◦C, 0.57 ± 0.01, 0.44 ± 0.01, and 0.31 ± 0.01 d−<sup>1</sup> ). In contrast, the highest Fv/F<sup>m</sup> values were measured for T. antarctica in the low Fe treatments (at 3◦C: 0.30 ± 0.02, high light; 0.28 ± 0.03, low light). Light had a significant effect on Fv/F<sup>m</sup> in all species except P. antarctica (p > 0.05; **Figure 6C**). Warming also had a significant effect on Fv/F<sup>m</sup> in all species except T. antarctica (p < 0.05). A significant positive relationship was observed between Chl a and Fv/F<sup>m</sup> but only in C. flexuosus and P. inermis (p < 0.05).

σPSII values for P. antarctica, P. inermis and C. flexuosus generally increased with higher temperature with some important exceptions (p < 0.05; **Figure 6D**). At high light and high Fe σPSII decreased due to temperature in cultures of P. inermis, C. flexuosus, and T. antarctica. σPSII was consistently larger in P. inermis than other study species when experimental treatments were compared, with the largest σPSII measured in cultures of all species grown at 5◦C under low light and low Fe conditions. A significant negative relationship was observed between growth and σPSII but only in P. inermis (p < 0.05).

## DISCUSSION

The Southern Ocean possesses distinct flora due to its unique environmental conditions and geographical isolation. The principal goal of this research is to predict how such isolated populations will fare in a rapidly changing ocean by first understanding how these primary producers respond to the three major environmental factors that limit primary productivity in the Southern Ocean – low iron supply, low temperatures, and low underwater light levels. Previous research has revealed systematic differences occur between "model" diatom species (typically coastal, temperate isolates) and Southern Ocean diatoms, pointing to adaptations in Southern Ocean species that allow them to overcome the increase in cellular Fe requirements that generally occurs with decreasing light (Strzepek et al., 2012, 2019). Similar to Boyd (2019), we found that low Fe conditions resulted in mortality at a lower temperature than high Fe cultures while the optimal growth temperature in our study agrees with the calculated growth temperature optimum of 5.2◦C for Southern Ocean phytoplankton (Coello-Camba and Agustí, 2017).

The temporal succession of Southern Ocean phytoplankton is not fully resolved, with multiple suggestions as to what triggers the change from one dominant group over another (Petrou et al., 2016). It is suggested that over winter, the deep mixed layer regenerates the surface nutrients and Fe supply, while in spring as the sea surface temperatures increases and winds decrease, the mixed layer shallows – therefore increasing the total photosynthetic active radiation (PAR) allowing early spring blooming cells to bloom (Boyd et al., 2010). Following this, and as Fe supply declines due to Fe uptake by larger cells, smaller cells with lower Fe requirements would tend to dominate (Tagliabue et al., 2014). While temperature is projected to increase in the future, irradiance is also predicted to increase due to a decrease in mixed layer depth (Boyd et al., 2015). However, there are still uncertainties surrounding future Fe supply, which hinders our understanding of primary productivity in the Southern Ocean (Laufkötter and Gruber, 2018).

## The Peculiarities of T. antarctica With Respect to Fe Supply

The optimal growth rate for T. antarctica is less clear due to its difficulty to be cultured at elevated Fe concentrations, i.e., Fe:EDTA media (**Figure 5**). The growth rates of T. antarctica grown in high Fe media and 90 µmol photons m−<sup>2</sup> s −1 at 3◦C in our study are similar to those reported by Strzepek et al. (2011). However, previously reported growth rates of T. antarctica grown in low Fe media are significantly lower than our results even though the experimental conditions in our study are comparable to those used in past studies (Strzepek et al., 2011).

Surprisingly, Fv/F<sup>m</sup> values for T. antarctica did not change appreciably in cultures grown under high light conditions and at 10◦C in high Fe medium, or at 12◦C in low Fe medium, even though growth rates were severely reduced in these treatments (p < 0.05; **Figures 3C,F**). This finding suggests that at these higher temperatures, T. antarctica cells appear photosynthetically competent but for some unknown reason are unable to divide. We cannot exclude the possibility that cells were undergoing sexual reproduction and hence stopped dividing asexually.

It has been shown that T. antarctica lack the gene ferritin (FTN) (Moreno et al., 2018). This ability to store Fe via FTN may mean that Fe may be "toxic" at higher concentrations.

The higher growth rates observed for T. antarctica under Fe deficient conditions may also be due to the presence of DFB in the media. We propose that this key difference in chelator, i.e., EDTA vs. DFB, may cause Fe to be held in solution in a more desirable form for uptake by the cell. We suggest that growth experiments should be considered to understand the response of T. antarctica cultures grown in different media containing an assortment of ligands to test our hypothesis regarding preference for Fe uptake.

## Thermal Response Curves of Southern Ocean Phytoplankton

To understand changes in the ability to sequester and export carbon due to climate-induced alterations in net primary productivity it is crucial to understand the underlying mechanisms that phytoplankton use to cope with environmental change (Padfield et al., 2016). Photophysiological measurements are used routinely to quantify the quantum yield of photochemistry in PSII (Fv/Fm) and are derived from the three possible paths for solar energy absorbed by photosynthesizing organisms (Falkowski and Raven, 2007). Absorbed photons can be (a) used for charge separation events leading to carbon (organic) synthesis, (b) be dissipated as heat, or (c) emitted back to the environment as fluorescence (Butler and Strasser, 1977). Such measurements have been used to suggest that compared to temperate coastal diatoms, Southern Ocean phytoplankton compensate for low Fe conditions by modifying their photosynthetic machinery to capture light when irradiance is limiting by modifying the size rather than the number of photosynthetic units (Strzepek et al., 2019).

Typically, σPSII (PSII absorption cross-section area) increases under low Fe conditions, either due to an increase in the number of PSII antennae complexes relative to reaction center complexes or an apparent increase in absorption cross-section area due to the decoupling of light harvesting antennae from PSII reaction center complexes (Greene et al., 1991; Behrenfeld et al., 2008; Strzepek et al., 2012). While σPSII was observed to increase for C. flexuosus and T. antarctica under low Fe conditions, σPSII

increased under high conditions for P. antarctica. The coupled increase in photochemical efficiency (Fv/Fm) and antennae size (σPSII) for P. antarctica (**Figures 3**, **4**) cultured under high Fe conditions at all temperatures are inconsistent with published literature and unexpected due to the decreased probability of excitation energy transfer to the reaction centers associated with larger antennae size (Kolber et al., 1994; Strzepek et al., 2012). The combination of increased Chl a and σPSII size may provide insight into these conflicting results, as the increase of photosynthetic pigments energetically coupled to PSII reaction centers in high Fe cells may exceed the degree of decoupling of the PSII reaction center from light harvesting antennae (reflected as an increase in σPSII) typically observed under Fe limitation. By reducing the amount of photons that are absorbed by the PSII reaction center, this unexpected increase in σPSII and Fv/F<sup>m</sup> may suggest that P. antarctica can withstand higher light intensities without compromising the photosynthetic efficiency of PSII.

While our study shows that biological activity increases with temperature up to an optimal temperature of 5◦C, most photochemical reactions are thought to be temperature independent (Somero and Hochachka, 1976). That said, thermal stress has been shown to damage the photosynthetic apparatus (Schreiber and Berry, 1977) and cause the PSII reaction centers and light-harvesting complex of PSII to functionally and physically dissociate (Armond et al., 1980). Thermal damage to PSII was most obviously observed in P. antarctica, shown as a decrease in Fv/F<sup>m</sup> and an increase in σPSII with warming (**Figures 3**, **4**). This was observed under both experimental irradiances. The photochemical physiology of T. antarctica was less affected by temperature, with a slight negative relationship between Fv/F<sup>m</sup> and σPSII with warming. In contrast, warming had little effect on Fv/F<sup>m</sup> and σPSII of C. flexuosus, expressed as a flat response curve, suggesting the photosynthetic apparatus of C. flexuosus and T. antarctica are thermally stable (Baker et al., 2016).

## Southern Ocean Phytoplankton Responses to Climate Change

In addition to predicted changes in light and Fe availability in the future (Boyd et al., 2015), the results from this study suggest that the 2◦C temperature increase estimated for the Southern Ocean at the end of this century (Ito et al., 2015), will result in significant ecological and biochemical changes to the Southern Ocean. P. antarctica plays a major role in the export of carbon in the Southern Ocean (Arrigo et al., 1999), while P. inermis has been shown to dominate austral spring/summer blooms, often outlasting other species into late summer (Annett et al., 2010; Lin et al., 2017). As the ocean warms, this will stimulate growth under high Fe conditions and high light (**Figure 6**). However, each species responds differently to other combinations of temperature, Fe and light supply – suggesting extensions or decreases in temporal bloom formation (**Figures 1**, **6**). It is also apparent that populations of P. inermis would diminish as they do not grow as well at higher temperatures (**Figure 7**).

The Southern Ocean has a distinct diatom community due to ocean circulation barriers, e.g., the Antarctic Circumpolar

Current (ACC) (Malviya et al., 2016). But recently it has been shown that large temperate macroalgae (southern bull kelp: Durvillaea antarctica) can frequently disperse across the ACC and it is the extreme conditions of the Southern Ocean that prevent the establishment of temperate-adapted taxa (Fraser et al., 2018). Thus as the Southern Ocean warms, it may become more habitable for species with a higher thermal tolerance while becoming less hospitable for species with a lower thermal tolerance (Griffiths et al., 2017). If the Ross Sea warms above 8 ◦C, the distribution and abundance of phytoplankton would change to favor species that retain the ability to photosynthesize optimally at elevated temperatures. The generalist Chaetoceros genus is one of the most abundant globally (Suto, 2006), thus it is likely that this genus can inhabit empty niches much faster than specialist species (Sriswasdi et al., 2017). This is due to the high speciation rates measured in generalist species and their ability to tolerate environmental change (Birand et al., 2011; Vamosi et al., 2014). It is likely that our study species C. flexuosus recently evolved from a temperate ancestor and has adapted to tolerate the low light and Fe environment of the Southern Ocean. Thus, it is likely that its heat resistant cellular functions will give it a competitive advantage over polar diatoms as anthropogenic CO<sup>2</sup> emissions and sea surface temperatures increase.

Like temperature, the light climate that phytoplankton will be exposed to in the future will vary. As the oceans warm and stratify, phytoplankton will be exposed to increased irradiances and damaging radiation (Doney, 2006). This could reduce rates of photosynthesis, growth and survival in phytoplankton species that are unable to successfully tune their photosynthetic repair and protection strategies to regulate excitation pressure. In turn, high temperatures can increase enzymatic turnover rates (therefore productivity) through increasing activity of thermally sensitive enzymes or accelerating protein synthesis (Boyd et al., 2016; Wagner et al., 2016). Indigenous Southern

Ocean phytoplankton have specific molecular and physiological adaptations to this low Fe and low light polar region (Strzepek et al., 2019), however, their narrow temperature tolerances (e.g., 0–8◦C) will limit their ability to adapt to increasing sea surface temperatures (Libralato et al., 2015).

Studies on the intraspecies variation in resident Antarctic phytoplankton species show there is some phenotypic plasticity to temperature (Reusch and Boyd, 2013) and light and Fe availability (Luxem et al., 2017), thus phytoplankton may not require physiological or molecular adaptations to photosynthesize optimally in the cold, low light and low Fe environment (Boyd, 2019). We propose that invading species with generalist lineages and high temperature tolerances may dominate blooms under future Southern Ocean conditions instead of species that have more specialized adaptations. However, due to the lack of evolutionary studies on Antarctic phytoplankton, the implications of physiological plasticity of polar species to environmental conditions are limited, therefore, it is difficult to extend predictions to the ecosystem level.

## CONCLUSION

In this study, we demonstrate the importance of investigating the response of non-model phytoplankton species isolated from the same region to assess their thermal optima. Phytoplankton species were cultured using a matrix of environmental conditions to determine any synergistic or antagonistic effects of light, temperature and iron limitation on phytoplankton physiology. Our results clearly show that Fe and light availability affect not only growth rate, but also the temperature range over which growth can occur. Iron stressed cells generally reached mortality at lower temperatures than high Fe cells.

The unique temperature reaction norms observed in this study show that polar phytoplankton utilize different specializations that allow them to coexist in the Southern Ocean, which is probably indicative of their different evolutionary history. Furthermore, Fe deficiency decreases the upper thermal limit for growth, which agrees with recent studies on temperate and polar

## REFERENCES


phytoplankton (Thomas et al., 2017; Boyd, 2019). The high light and low Fe conditions expected in the austral spring/summer conditions in 2100 will limit the productivity of some species while enhancing the growth of others.

The question about the variation in thermal tolerance raised in this study suggests that understanding a species evolutionary history will play a key role in predicting the success of future Antarctica phytoplankton populations. More information on genetic differences between phytoplankton displaying stark differences in their thermal tolerances would help to establish more accuracy in understanding how species will respond to future climate changes.

## DATA AVAILABILITY STATEMENT

All datasets generated for this study are included in the article/**Supplementary Material**.

## AUTHOR CONTRIBUTIONS

SA, RS, HM, and ME designed the research. SA and HM performed the research. SA wrote the manuscript with contributions from ME, RS, and PB.

## FUNDING

This work was supported by an Australian Government Research Training Program Stipend Scholarship (SA) and Australian Research Council's Discovery Program (DP170102108 and DP130100679).

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2019.00681/full#supplementary-material




Zhu, Z., Qu, P., Gale, J., Fu, F., and Hutchins, D. A. (2017). Individual and interactive effects of warming and CO2 on Pseudo-nitzschia subcurvata and Phaeocystis antarctica, two dominant phytoplankton from the ross sea, Antarctica. Biogeosciences 14, 5281–5295. doi: 10.5194/bg-14-5281-2017

**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 Andrew, Morell, Strzepek, Boyd and Ellwood. 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.

# Integrated Modeling to Evaluate Climate Change Impacts on Coupled Social-Ecological Systems in Alaska

Anne Babcock Hollowed<sup>1</sup> \*, Kirstin Kari Holsman<sup>1</sup> , Alan C. Haynie<sup>1</sup> , Albert J. Hermann2,3 , Andre E. Punt<sup>4</sup> , Kerim Aydin<sup>1</sup> , James N. Ianelli<sup>1</sup> , Stephen Kasperski<sup>1</sup> , Wei Cheng2,3 , Amanda Faig2,4, Kelly A. Kearney1,2, Jonathan C. P. Reum1,5, Paul Spencer<sup>1</sup> , Ingrid Spies<sup>1</sup> , William Stockhausen<sup>1</sup> , Cody S. Szuwalski<sup>1</sup> , George A. Whitehouse2,4 and Thomas K. Wilderbuer<sup>1</sup>

### Edited by:

Jamie C. Tam, Bedford Institute of Oceanography (BIO), Canada

#### Reviewed by:

Nancy Shackell, Bedford Institute of Oceanography (BIO), Canada Daniel Howell, Norwegian Institute of Marine Research (IMR), Norway

\*Correspondence:

Anne Babcock Hollowed Anne.Hollowed@noaa.gov

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 20 August 2019 Accepted: 02 December 2019 Published: 14 January 2020

#### Citation:

Hollowed AB, Holsman KK, Haynie AC, Hermann AJ, Punt AE, Aydin K, Ianelli JN, Kasperski S, Cheng W, Faig A, Kearney KA, Reum JCP, Spencer P, Spies I, Stockhausen W, Szuwalski CS, Whitehouse GA and Wilderbuer TK (2020) Integrated Modeling to Evaluate Climate Change Impacts on Coupled Social-Ecological Systems in Alaska. Front. Mar. Sci. 6:775. doi: 10.3389/fmars.2019.00775 <sup>1</sup> Alaska Fisheries Science Center, National Marine Fisheries Service, National Oceanic and Atmospheric Administration, Seattle, WA, United States, <sup>2</sup> Joint Institute for the Study of the Atmosphere and Ocean, University of Washington, Seattle, WA, United States, <sup>3</sup> Pacific Marine Environmental Laboratory, Oceans and Atmospheric Research, National Oceanic and Atmospheric Administration, Seattle, WA, United States, <sup>4</sup> School of Aquatic and Fishery Science, College of the Environment, University of Washington, Seattle, WA, United States, <sup>5</sup> Centre for Marine Socioecology, Institute for Marine and Antarctic Studies, College of Sciences and Engineering, University of Tasmania, Hobart, TAS, Australia

The Alaska Climate Integrated Modeling (ACLIM) project represents a comprehensive, multi-year, interdisciplinary effort to characterize and project climate-driven changes to the eastern Bering Sea (EBS) ecosystem, from physics to fishing communities. Results from the ACLIM project are being used to understand how different regional fisheries management approaches can help promote adaptation to climate-driven changes to sustain fish and shellfish populations and to inform managers and fishery dependent communities of the risks associated with different future climate scenarios. The project relies on iterative communications and outreaches with managers and fishery-dependent communities that have informed the selection of fishing scenarios. This iterative approach ensures that the research team focuses on policy relevant scenarios that explore realistic adaptation options for managers and communities. Within each iterative cycle, the interdisciplinary research team continues to improve: methods for downscaling climate models, climate-enhanced biological models, socio-economic modeling, and management strategy evaluation (MSE) within a common analytical framework. The evolving nature of the ACLIM framework ensures improved understanding of system responses and feedbacks are considered within the projections and that the fishing scenarios continue to reflect the management objectives of the regional fisheries management bodies. The multi-model approach used for projection of biological responses, facilitates the quantification of the relative contributions of climate forcing scenario, fishing scenario, parameter, and structural uncertainty with and between models. Ensemble means and variance within and between models inform risk assessments under different future scenarios. The first phase of projections of climate conditions to the end of the 21st century is complete,

including projections of catch for core species under baseline (status quo) fishing conditions and two alternative fishing scenarios are discussed. The ACLIM modeling framework serves as a guide for multidisciplinary integrated climate impact and adaptation decision making in other large marine ecosystems.

Keywords: climate change, fishery management strategy, Bering Sea, walleye pollock, Pacific cod, climate projections

## INTRODUCTION

Significant increases in sea surface temperature (SST) over the next century are projected for most ocean systems (IPCC, 2014, 2018). Global warming is expected to have strong impacts on ocean temperature and ocean productivity in high latitude systems (Arrigo and Van Dijken, 2015; Smith et al., 2019). However, the effect of warming climate conditions on marine ecosystems and species are expected to be system- and speciesdependent, and the footprint of environmental change may exhibit considerable variation across space and time (Poloczanska et al., 2013; Cheung et al., 2016; Spencer et al., 2019). High latitude marine ecosystems such as the Bering Sea are expected to experience large deviations from historical ocean conditions (Hermann et al., 2019; Spencer et al., 2019). Indeed, increased ocean temperature has already impacted the Bering Sea marine ecosystem through shifts in trophic demand and overwinter survival (Heintz et al., 2013), species interactions (Spencer et al., 2016), shifting spatial distributions (Barbeaux and Hollowed, 2018; Stevenson and Lauth, 2019; Thorson, 2019), and overall system productivity (IPCC, 2014; Hermann et al., 2019). Bering Sea ecosystems are also threatened by the effects of ocean acidification on valuable crab stocks and important pelagic prey species (Comeau et al., 2010; Long et al., in press).

In anticipation of these changes, the US National Marine Fisheries Service (NMFS) established a Climate Science Strategy (NCSS) that called for scientists from each management region to conduct research to understand, prepare for, and respond to, climate impacts on the distribution and abundance of managed species and the ecosystems in which they reside (Buser et al., 2016). In response to this national call to action, an interdisciplinary team of researchers was formed in 2015 to develop the Alaska Climate Integrated Modeling (ACLIM) projec<sup>1</sup> . The goals of ACLIM were to:


We addressed these goals by: (1) applying a multi-model approach (sensu Kaplan et al., 2019; Lotze et al., 2019) to project the distribution and abundance of commercially important fish and fisheries in the EBS under various climate change and fishing scenarios (**Figure 1**); (2) evaluating the economic and biological performance of the fishing scenarios for consideration by the North Pacific Fishery Management Council (NPFMC), the federal fisheries management body for the region (sensu Holsman et al., 2019); and (3) decomposing uncertainty in future climate projections according to structural, scenario, and parameter uncertainty sources (Cheung et al., 2016; Payne et al., 2016; Reum et al., in press).

This paper describes the ACLIM research framework, its approach to quantifying uncertainty and multi-model inference, and the program's approach to interfacing science with regional fishery management councils. This paper is designed to provide an overview of the research framework. For in-depth details of each modeling approach, the reader is directed to relevant publications.

## HISTORY

The ACLIM Team selected the EBS as a case study for the development and implementation of a regional climate impact, assessment, and management planning enterprise. The EBS supports abundant fish and shellfish resources that are of considerable economic and social value to the region, the United States, and the world. For example, the estimated 2017 first-wholesale value for commercial harvest of all species (crab, groundfish, clams, scallops, salmon, halibut) in the United States shelf and slope regions of the EBS was \$2.68 billion (Fissel et al., 2019). In addition, the major physical and biogeochemical processes governing ecosystem production have been studied for at least 40 years, providing opportunities for formulation and parameterization of the responses of marine species to changes in atmospheric, oceanographic, and biogeochemical drivers (Sigler et al., 2016b; Stabeno et al., 2016). An Ecosystem Approach to Fisheries Management (EAFM, Dolan et al., 2015) is used in the region (Stram and Evans, 2009) and managers and stakeholders are actively seeking improved climate- and ecosystem-based information for decision making.

Key features of the EBS include: seasonal ice cover, distinct biophysical domains driven by surface forcing and tidal mixing, ice associated algal and phytoplankton blooms, and fall phytoplankton blooms (Hunt et al., 2011; Stabeno et al., 2012, 2017; Wang et al., 2012; Baker and Hollowed, 2014; Cheng et al., 2015; Hermann et al., 2019). The role of temperature and sea ice on the seasonal availability of high energy content planktonic prey (large zooplankton) has been shown

<sup>1</sup>https://www.fisheries.noaa.gov/alaska/ecosystems/alaska-climate-integratedmodeling-project

to be associated with overwintering survival of walleye pollock (Gadus chalcogrammus) and Pacific cod (Gadus macrocephalus), two abundant and economically important species (Heintz et al., 2013; Duffy-Anderson et al., 2017). Laboratory studies have quantified key bioenergetic responses for commercially important groundfish and the impacts of ocean acidification on key developmental processes of commercially important crab stocks (Long et al., 2013, 2019).

The origins of the ACLIM project can be traced back to a long legacy of interdisciplinary research programs. Early versions of the current high spatial and temporal resolution oceanographic model [a Regional Ocean Modeling System (ROMS) model with 10 km horizontal resolution for the Bering Sea, Bering10K] were developed as part of the US GLOBEC Northeast Pacific program, a partnership between the National Science Foundation (NSF) and the National Oceanic and Atmospheric Administration (NOAA) (Curchitser et al., 2005). Early food web models were developed as part of the Southeast Bering Sea Carrying Capacity research program, a partnership between the Coastal Ocean Program, the Pacific Marine Environmental Laboratory, and the NMFS (Aydin et al., 2007). Development of climate-enhanced single-species projection models (CE-SSM, Ianelli et al., 2016) and fully coupled end-to-end ecosystem models (Hermann et al., 2013, 2016; Ortiz et al., 2016), and fisher's choice models (Haynie and Pfeiffer, 2013) were all developed as part of the Bering Sea Project (BSP); a partnership between the NSF, the North Pacific Research Board (NPRB), and the NOAA (Wiese et al., 2012). The Climate-Enhanced Age-based model with Temperature-specific Trophic Linkages and Energetics (CEATTLE, Holsman et al., 2016) was funded directly by NMFS research initiatives focused on the development of integrated ecosystem assessments and stock assessment improvement. This legacy of collaborative research led to a mechanistic understanding of key biophysical linkages governing fish production (Sigler et al., 2016a) and completed model performance verification studies that served as the foundation for the ACLIM project. Briefly, the ACLIM framework generates dynamically coupled, regionally downscaled projections of the oceanography and biogeochemistry of the EBS ecosystem derived from earth system models (ESMs) driven under contrasting future

emission scenarios [representative concentration pathways (RCPs)] (**Table 1**). Recognizing that there are strengths and weaknesses to every modeling approach (e.g., Hollowed et al., 2013), investigators proposed a multi-model inter-comparison approach (Stock et al., 2011; Stouffer et al., 2017). Projected ocean and biogeochemical conditions are directly or indirectly utilized to project the future of marine species and fisheries in the region (**Figure 1** and **Table 1**) using a suite of population dynamics models with various levels of complexity (**Figure 2** and **Table 2**).

Using the multi-model approach, projections of fish and shellfish distribution and abundance are assessed for the current (2006–2020), mid-century (2030–2050), and end-ofcentury (2080–2100) time periods under suites of potentially viable fishing scenarios and management strategies that are vetted through the NPFMC. Projected stock conditions (e.g., size-at-age, abundance, reproductive potential, reproductive success, and distribution) for six core species (walleye pollock; Pacific cod; yellowfin sole, Limanda aspera; northern rock sole, Lepidopsetta polyxystra; arrowtooth flounder, Atheresthes stomias; and snow crab, Chionoecetes opilio) are compared to assess the performance of current and alternative fishing scenarios with respect to the ecosystem, social and economic goals of the NPFMC. Projections of key ecosystem status indicators (e.g., species diversity, mean trophic level of the catch) are derived from ecosystem models and are evaluated under current and alternative fishing scenarios. In this context, fishing scenarios include both the suite of constraints imposed by a given fisheries management strategy and external processes influencing fishing behavior [e.g., allocation of total allowable catch (TAC) between fishing sectors, fuel costs, world markets] (Groeneveld et al., 2018; Fulton et al., 2019). The implications of shifting spatial distributions of commercial species on the coupled socialecological system are assessed using spatially explicit models (**Table 2**). Collectively, these projections provide the scientific information needed to identify thresholds for management action and viable adaptation strategies. For example, many regions monitor proxies for reproductive potential (i.e., spawning stock biomass) and establish biological reference points for reductions in fishing mortality or the development of rebuilding plans. Examination of the performance of current and alternative fishing scenarios and management strategies helps to identify climate-ready harvest control rules that are robust to changing climate, and to inform the public and management of trade-offs associated with different options for the management of marine resources under a changing climate.

The origins of the EAFM approach used in the region and the keen interest of managers and stakeholders in improved climate- and ecosystem-based information for decision making can be traced to the iterative communication between scientists, managers, and stakeholders. The NPFMC was one of the first Councils in the United States to adopt an ecosystem considerations report (Livingston et al., 2001). This report has evolved over the years and is now considered an integral part of the NPFMC's annual assessment reviews (Zador et al., 2017) and in 2018, the NPFMC adopted a Fishery Ecosystem Plan for the Bering Sea<sup>2</sup> . The shared recognition of the scientific and management community of the potential risks of changing climate conditions on sustainable fishery management in the region underscored the need for a climate module within the FEP which would serve as a strategic planning tool for the

<sup>2</sup>https://www.npfmc.org/bsfep/


TABLE 1 | Summary of global or earth system models and scenarios used in ACLIM and Bering Sea Project (BSP) experiments.

<sup>∗</sup>An example of one of the six global models that will be used in ACLIM phase 2 included to illustrate the progression to finer spatial resolution.


NPFMC. The ACLIM modeling framework is designed to fill this strategic need.

## CLIMATE MODEL STRUCTURAL UNCERTAINTY AND EMISSION SCENARIOS

Projections are completed in phases that are tied to the availability of updated ESM projections and funding. Phase 1 is nearing completion. Phase 1 utilized output from six ESMs developed for the IPCC fourth or fifth Assessment Reports (based on output from the third or fifth Coupled Model Intercomparison Project, "CMIP3" or "CMIP5") (**Table 1**). In this phase, the ACLIM framework was developed and tested, providing multimodel projections of the impacts on fish, invertebrates, and fisheries under status quo and two fishing scenarios. In phase 2, the entire suite of biological models (**Table 2**) will be updated with ESM output developed for the IPCC Sixth Assessment Report (i.e., CMIP6) and evaluated under an expanded suite of fishing scenarios that will include alternative management strategies. Comparison of results from phase 1 and 2 will allow analysts to explore how improvements in ESMs affect projected impacts on fish and fisheries. This paper focuses on phase 1 of the project.

Structural differences among global climate models and uncertainty regarding emission scenarios were addressed in phase 1 of the ACLIM project by comparing outcomes based on multiple climate models under several emission scenarios (Van Vuuren et al., 2011; Van Vuuren and Carter, 2014; **Table 1**). In this phase, output from six global climate models (three from CMIP3 and three from CMIP5) were selected from the full suite of global climate models considered by IPCC (**Table 1**). The CMIP3 suite was selected during the BSP to span a broad range of potential sea ice dynamics (Hermann et al., 2016). Selected models from the CMIP5 suite included: the Geophysical Fluid Dynamics Laboratory (GFDL) – ESM 2M (ESM2M) (Dunne et al., 2012); the National Center for Atmospheric Research (NCAR) Community Earth Systems Model (CESM) (Kay et al., 2015); and the MIROC ESM (Watanabe et al., 2011) (**Table 1**). These three models were selected because they projected a broad range of global patterns for precipitation and SST, and provided contrasting views of future ocean conditions in the EBS. Output from these models under two RCPs (4.5 and 8.5; Van Vuuren et al., 2011; Van Vuuren and Carter, 2014) were used to drive the Bering10K regional model. RCP 8.5 and 4.5 represent a high-emission business-as-usual scenario and an intermediate scenario, respectively.

## CLIMATE PROJECTION DOWNSCALING

Previous analysis of the skill of global scale ESMs over the historical period revealed that these coarse resolution models are unable to adequately resolve the seasonal spatial patterns of sea ice extent and bottom water temperatures that are key structural features of the EBS shelf (Vancoppenolle et al., 2013). Additionally, model intercomparisons of 21 global biogeochemical models' abilities to reproduce observed net primary productivity in the Arctic Ocean revealed several limitations that varied by region (Lee et al., 2016). Many of these deficiencies related to mixed layer depth and sea ice concentration in the simulations.

To address these potential deficiencies, the ACLIM framework deploys the Bering10K ROMS model (Hermann et al., 2016, 2019) to dynamically downscale CMIP5 projection simulations for the Bering Sea region. In this framework, the ROMS ocean model is forced at the surface by heat fluxes, freshwater fluxes, and wind stress values derived from prescribed atmospheric states based on the global model projections and modeled surface temperature (SST), and at the lateral boundaries by temperature, salinity, and current speeds from the ocean component of the global model projections. In two simulations, nitrate and ammonium values from the biogeochemical component of the ocean model in the global projections were also used in the lateral boundary condition variables; in these instances, simulations were run under projected nutrient boundary conditions and alternatively with World Ocean Atlas-derived climatological nutrient boundary conditions to contrast the relative impact of temporal trends in projected nutrients. This downscaling framework allows for approximately a 100-fold increase in the number of horizontal grid points compared to that of the global models (**Figure 3**). The Bering10K model also includes its own sea ice and biogeochemical models through which the climate model forcing data influence the local dynamics. When forced in hindcast mode with surface and lateral boundary conditions from the data-assimilating Climate Forecast System Reanalysis (CFSR), the Bering10K model has demonstrated significantly improved representation of features such as advection pathways, mixed layer depth, sea ice extent, and the strength and interannual variability of the Bering Sea cold pool compared to that seen in CFSR itself (Hermann et al., 2016; Kearney et al., in press).

Since phytoplankton and zooplankton are responding to both physics and nutrients, the Bering10K ocean model is coupled to a biogeochemical model (BESTNPZ) that simulates lower trophic level dynamics for the Bering Sea (Gibson and Spitz, 2011; Kearney et al., in press). Within the ACLIM framework, the use of this single biogeochemical model to derive biological indices for all the downscaled climate simulations is the one place where neither structural nor parameter uncertainty is addressed via an ensemble approach. During development of the ACLIM modeling framework, the merits of including multi-model- or parameter-varying ensembles of biogeochemical models were discussed but ultimately not included due to the need to limit the permutations assessed in the project. Therefore, for phase 1, the only uncertainty related to lower trophic dynamics that is quantified is that relating to projected trends in nutrient boundary conditions.

Relevant biogeochemical and physical properties (**Tables 3**–**5**) were projected for the period 2006–2100 for scenarios based


TABLE 3 | Fish- and fisheries-relevant output variables from the Bering10K-BESTNPZ model<sup>∗</sup> .

The "2D/3D" column indicates whether the variables are two- or three-dimensional, and "Grid" refers to the positions of the variables on the ROMS grid: u = east/west, v = north/south, and ρ = center of grid cells. <sup>∗</sup>The mix of units for biological state variables is for output purposes only; internally, all model calculations use nitrogen as the primary currency, with a constant 106:16 C:N ratio.

on CMIP5 models. Since CMIP3 did not cover the entire time period 2006–2100, we used CMIP3 for those periods to the extent possible (2003–2040). In addition, a hindcast simulation spanning the period of 1970–2018, forced by a combination of version 2 forcing for Coordinated Ocean-Ice Reference Experiment, i.e., COREII (Large and Yeager, 2009) (1970–1994), the CFSR (Saha et al., 2010) (1995–March 2011), and the Climate Forecast System Operational Analysis, i.e., CFSv2-OA (April 2011–Sep 2018) was performed for use in calibrating the various upper trophic level models for past decades. Comparison of these hindcasts using different ESM boundary conditions revealed potential temperature biases. Several methods have been used to address systematic biases in global model temperatures relative to current observed temperatures (Piani et al., 2010). In comparative studies, bias can be accounted for by evaluating relative changes in mean state between time periods (Hermann et al., 2019). However, when animals respond to environmental thresholds, relative environmental changes may not be adequate when downscaled variables are used to drive responses of secondary producers and higher trophic levels (Small et al., 2015; see section "Bias Corrections for Biological Responses"). To address this issue, projections with and without bias corrections are compared.

## EXPLORATIONS OF THE POTENTIAL ROLE OF BIOLOGICAL COMPLEXITY

The ACLIM framework employs a multi-model approach for projection of biological responses to explore the trade-offs between computational speed and ecosystem realism inherent in the selection of higher trophic level models (Hollowed et al., 2013). When models of varying complexity are considered jointly (some with high spatial resolution and species interactions and others with well-defined distributions for key parameters), multi-model projections can provide a more complete suite of future projections for evaluating climate change impacts on ecosystems and resource-dependent human communities (Plagányi et al., 2011; Tittensor et al., 2018). Within the ACLIM framework the suite of models range from minimally realistic single-species climate-enhanced stock projection models (CE-SSM) that are capable of detailed treatment of process error and measurement error, to whole ecosystem models that track

#### TABLE 4 | Eastern Bering Shelf-derived indices.

fmars-06-00775 December 31, 2019 Time: 13:37 # 8


These index variables are extracted from the Bering10K-BESTNPZ output as yearly time series. The spatial and temporal reduction is applied in two ways: (1) survey replication (SR): variables are sampled at the same location and day-of-year as in the annual groundfish survey, then averaged across each year, and (2) seasonal (seas.): values within the survey sampling strata polygons are averaged spatially, then in time for each season (fall = Oct–Nov, spring = Apr–Jun, summer = Jul–Sep, winter = Dec–Mar) and annually.

potential structural changes within the ecosystem that may emerge from complex ecosystem interactions (**Figure 2** and **Tables 2**, **5**; Plagányi et al., 2011; Stock et al., 2011). The diverse multi-model projection suite provides a reasonable range of representative futures (with sufficient contrast in climate scenarios) that can be used to evaluate short- and long-term implications of management actions under future climate change.

The ACLIM framework leveraged eight types of stock or ecosystem projection models (**Table 2**):

• Trait-based vulnerability analyses (Hare et al., 2016; Spencer et al., 2019). The VA model utilizes expert judgment to assess sensitivity, exposure and vulnerability to climate change and does not project the specific outputs shown in **Table 5**.



The ACLIM framework enables analysts to evaluate the contributions of different sources uncertainty. The inclusion of MICE assessment (Plagányi et al., 2014) in the ACLIM framework provides opportunities to explore the contribution of process error and scenario uncertainty in single- and multi-species projections. Two MICE models in particular are included in ACLIM; the CEATTLE model (Holsman et al., 2016), and a CEversion of the spatiotemporal models of intermediate complexity for ecosystems (i.e., "MICE-in-space" model; Thorson et al., 2019). These models can be run relatively quickly, allowing sensitivity testing of the implications of uncertainty in climate linkages to: predator–prey overlap (and hence mortality rates); prey switching, prey availability, and metabolic rates (growth and maturation rates); and reproductive success (via the spawner–recruit relationships). Of these processes, the linkages between climate variability and future fish production are the most influential in terms of projected stock status and the most challenging to parameterize correctly (Szuwalski et al., 2015) because the processes governing climate impacts on fish and crabs are temporally varying and stage-specific (Bailey, 2000). The inclusion of food web, size spectrum, and end-to-end models provides an opportunity to evaluate the relative contributions of structural uncertainty, species interactions, fishing, and ecosystem changes to future states of nature.

Techniques for assessing the predictive skill of ecosystem models are emerging and they reveal a modest ability to reconstruct observed dynamics in stock status (Olsen et al., 2016). FEAST is a spatially explicit end-to-end ecosystem model that tracks core species in space and time (**Table 2**). Movements are determined from an evaluation of the relative quality of the current location with respect to foraging needs (demands on metabolic rate and prey availability) to adjacent cells within the ROMS Bering 10K grid (**Figure 3**). In longterm projections, small errors can accumulate in a spatial model of this complexity (Punt et al., 2016b). To address this issue, FEAST can be nudged by initiating the model using the projected environmental conditions at mid-century (2030–2050) and end of century (2080–2100) and seeding the starting abundance of

fmars-06-00775 December 31, 2019 Time: 13:37 # 9

TABLE 5


core species using output from simpler single- or multi-species climate enhanced models (**Table 1**).

## BIAS CORRECTIONS FOR BIOLOGICAL RESPONSES

A critical element of the ACLIM framework is the demonstration that the modeling suite used for projections is skillful in reconstructing observed population dynamics of core species and catch. To confirm this skill, it was necessary to demonstrate that when driven by hindcasts of observed oceanographic and biogeochemical conditions, the projected higher trophic level and fishing responses were consistent with observed historical interannual fluctuations. Functional forms and parameters used in the ACLIM were derived from a combination of retrospective studies external to the model (e.g., laboratory studies of metabolic rate or consumption, Holsman et al., 2016), retrospective data analysis based on observed data (e.g., climate envelope studies, Spencer et al., 2016), or retrospective analysis based on output from previous Bering10K hindcasts (e.g., spawner–recruitment relationships). In the case of CE-SSMs and the CE-MSM, environmentally linked age- or size- based statistical assessment models were used to derive functional forms and base parameters (see Holsman et al., 2016; Ianelli et al., 2016; Spencer et al., 2016 for examples). For all models, once parameterized, we drove our hindcast period (2006–2017) with reanalysis-based (e.g., Saha et al., 2010) and data-assimilating input products, which successfully tied our hindcast simulation time series to their real-world counterpart datasets. These hindcasts incorporate observed variations due to radiative forcing changes from natural and anthropogenic sources and internal natural variability.

When incorporating downscaled physical and biogeochemical indices (**Tables 3–5**) into hindcast-trained models, it was necessary to account for both systematic biases in each global model as well as mismatch at any given time due to internal variability of each model compared to the hindcast period. While we have identified methods to address the former, the latter is left unaddressed during phase 1 of the ACLIM simulations. The existing simulations include only a single decade of overlap (2006–2017) between the hindcast and projection simulations. This time period is not long enough to separate model bias from decadal variability due to internal oscillations such as ENSO or the PDO. For the phase 2 of ACLIM simulations, we will extend the downscaled climate model projections to include several decades from the historical period (i.e., 1970 present). This will allow for the diagnosis of model biases vs. the hindcast and observations, and allow for smoother forcing

of the upper trophic level models as they cross the hindcastto-projection threshold. Given the lack of a sufficient overlap period for phase 1 calculations, we assumed that the mean and variance during the reference overlap period are representative for both the hindcast and projections under present-day radiative conditions, and that conditions during the reference period are not anomalous.

Bias corrections must be considered carefully when projected environmental data are used to drive biological responses. Global climate model ensemble projections routinely apply additive bias correction (e.g., the "delta method"; Ho et al., 2012; Hawkins et al., 2013). The procedure adjusts projections based on mean differences between the hindcast and projection variable in a period of overlap. However, the procedure is not straightforward to apply to biological projections such as biomass densities that are bounded by zero because the resulting values can take negative values. Instead, a proportional correction can be applied. As with the additive correction method, the biomass density projection is re-centered based on the mean difference between the projection and hindcast overlap period, but the proportional change observed in the uncorrected projection (that is, variable level relative to the mean value from the overlap) is carried through to the recentered projection (Buser et al., 2009; Haerter et al., 2011; Reum et al., in press).

## FISHING SCENARIOS AND MANAGEMENT STRATEGIES

There are myriad pathways through which climate change can impact marine industries (Allison and Bassett, 2015). Bounding the range of possible management futures within the context of global shared socio-economic pathways is challenging and requires strong communication between management and modeling teams (O'neill et al., 2014; Groeneveld et al., 2018). In phase 1, a narrow suite of fishing scenarios was selected which represented two variations in TAC allocations across fishing sectors within the existing constraints of the NPFMC's existing EAFM management strategy. These fishing scenarios reflected two alternatives to status quo that have a significant impact on stakeholders (Ono et al., 2017).

The NPFMC's EAFM in the EBS employs a complex suite of management measures that are designed to sustain fisheries using science-based precautionary harvest control rules that are designed to sustain the reproductive potential of the stocks, preserve essential fish habitats, maintain a sustainable forage base for fish and other top trophic level consumers, and preserve ecosystem structure by limiting the overall removal of groundfish from the ecosystem (Stram and Evans, 2009; Hollowed et al., 2011). Under the US guidelines for the Magnuson Stevens Fishery Management Act, the TAC must be less than or equal to the Acceptable Biological Catch (ABC) and the combined TACs for federal groundfish fisheries in the Bering Sea Aleutian Island (BSAI) region cannot exceed the 2 million t system level cap. Groundfish fisheries are constrained by bycatch [Prohibited Species Caps (PSC)] that limit on non-groundfish species targeted by other commercial, recreational, and subsistence harvester (Pacific halibut, Hippoglossus stenolepis; Pacific herring, Clupea pallassi, salmon, and crab). The management system also includes catch share provisions and sector limitations designed to ensure that: a diverse suite of fishing sectors and communities have access; gear conflicts are avoided; and prey for protected species (such as marine mammals) is protected (Stram and Evans, 2009; Abbott and Haynie, 2012; Reimer and Haynie, 2018; Kroetz et al., 2019). In 2018, the NPFMC adopted a Fisheries Ecosystem Plan for the Bering Sea that specifically calls for the exploration of climate impacts on EBS fisheries (NPFMC, 2018). These features of the management system needed to be adequately represented in the suite of models employed by the ACLIM project.

In phase 1, all fishing scenarios employed the NPFMC's EBFM Management Strategy with respect to estimation of biological reference points, prevention of overfishing, and prohibitions on fishing forage fish. The fishing scenarios explored four alternatives of the groundfish TAC across fishing sectors under the 2 million t cap: (a) no fishing; (b) the status quo; (c) a 2 million t cap which allows for the expansion of flatfish fisheries (10% increase in the total cap allocation to different flatfish species under the overall system level cap); and (d) a shift in the groundfish TAC allocation across species such that potential pollock and/or cod TAC at high stock sizes could expand despite its impact on fishing opportunities for non-pollock and cod groundfish fishers under the cap (10% increase in the allocation of gadids under the cap). This suite of alternative management strategies allows the NPFMC to explore trade-offs between harvesting more pollock and cod or more flatfish.

Stakeholder engagement is a critical element of successful management strategy evaluations (MSEs; Colenbrander and Sowman, 2015; Jones et al., 2016; Punt et al., 2016a). The selection of the initial suite of fishery scenarios that only modified allocations of groundfish TAC across species and fishing sectors had the benefit of being easily understood by managers and stakeholders. This provided an excellent opportunity to introduce the utility of the ACLIM framework for management planning in a public forum through multiple workshops. A benefit of the workshops was the two-way communication between stakeholders, managers, and the scientific community. The current scenarios are also valuable as they provide insight to managers of the trade-offs of sustained increases in allocation to one group of species. As of this submission, the first phase of projections of climate conditions to the end of the 21st century are complete for CE-SSM, CE-MSM, EwE, and MIZER models, including projections of catch for core species under no fishing, baseline (status quo) fishing conditions, and two alternative fishing scenarios. Projections based on FEAST and IBM models are in preparation. FEAST runs will not include the fisher response capability. Incorporating fisher's responses within the spatial–temporal ecosystem model would require fleet/sector level data and predictions that was beyond what could be done in phase 1. In phase 2, scenarios generated from the multi-model framework will be expanded to explore

fishing scenarios where the performance of alternative climateinformed fishery management strategies will be tested to identify and avoid maladaptive pathways (Wise et al., 2014) and explore climate resilient options (Holsman et al., 2019; Karp et al., 2019). This is challenging given the broad suite of potential fishery management strategies that could be considered by the modeling team (Fulton et al., 2019) and possible shifts in societal choices regarding marine resources (Groeneveld et al., 2018). An integrated approach of vetting strategies with the public based on results from climate-enhanced single- and multispecies models will help to narrow the suite of candidate fishery management approaches applied to the fully integrated spatial– temporal ecosystem model (FEAST). Broader suites of fishery management strategies and parameter settings can be explored across other biological models.

## FISHERY-DEPENDENT COMMUNITY IMPACTS ASSESSMENTS

Several models have been developed to assess the economic impacts of changing groundfish distributions and abundance. Climate envelope assessments coupled with spatial impacts on vital rates have been used to assess climate impacts on regional (Le Bris et al., 2018) and global (Cheung et al., 2019) fisheries. Models have been used to assess climate change impacts on regional economies (Seung et al., 2015; Seung and Ianelli, 2016) and global supplies of fish on fishery dependent communities (Merino et al., 2012). Evaluation of community impacts and adaptation options examine multiple pathways through which changes in the quantity, location, and value of harvest translate into regional economic impacts on communities (Seung and Miller, 2018).

In phase 2, shifting species ranges can impact fisheries in multiple ways (Pinsky et al., 2018). Fisher responses can be directly included in FEAST. Alternatively, expected shifts in the spatial distribution of fish and shellfish can be predicted from spatiotemporal models informed by size-specific and nonlocal responses to climate projections (Thorson et al., 2017; Thorson, 2019) and implications for shifting distribution on core parameters in single- or multi-species models can be modeled as a function of projected environmental conditions (Spencer et al., 2016). For example, the CEATTLE multi-species model uses a climate-specific overlap index (Carroll et al., in revision) for predator and prey species to inform annually varying predation mortality. In phase 2 of ACLIM, the Spatial Economics Toolbox for Fisheries (FishSET) could be linked with the spatiotemporal model to project variables that influence fishers' choices regarding where and when to fish (Haynie and Pfeiffer, 2012).

## DATA SYNTHESIS AND INFERENCE

The outcomes of projections derived from the ACLIM framework can be synthesized using techniques commonly used in the stock assessment community. Multi-model inference can be used for several purposes, including evaluating the extent to which general conclusions regarding management actions are robust to structural assumptions (e.g., Murawski et al., 2010; Payne et al., 2016; Kaplan et al., 2019), using simple models (such as CE-SSM and CEATTLE, MICE-in-space, and perhaps even EwE) to identify the most consequential uncertainty which can then be used to prioritize sensitivity runs for more complex models such as FEAST, and improve stability in management advice (Stewart and Hicks, 2018). Structural uncertainty is a source of uncertainty that is usually hard to qualify but can be amongst the largest sources of uncertainty when providing management advice (Hill et al., 2007). Consequently, structural uncertainty has become a major focus for MSEs (Punt et al., 2016a).

Given the multiple sources of potential uncertainty in climate impact assessments, multi-model ensembles within and across models of different complexity are of interest. Ensemble modeling approaches will be applied to synthesize the information to derive overall system level trends (Ianelli et al., 2016; Rosenberg et al., 2018; Lotze et al., 2019). Selection of models used to derive ensemble estimates may be informed by an analysis of among model correlations (Stewart and Hicks, 2018). Current decisions relevant to formulating the ensemble include selection criteria for model inclusion in the ensemble (e.g., Butterworth et al., 1996) and weighting criteria for the models included in the ensemble (Anderson et al., 2017) among others. In applications for stock assessments, the results from each model could be weighted by its fit to the available data using a Bayesian approach (e.g., Butterworth et al., 1996; Hill et al., 2007) to create probability distributions for model outputs. Model selection and weighting for ecological projections is more challenging due to the lack of observations for tuning. Our approach has been to tailor ensemble syntheses within each application and question, and different approaches are illustrated in recent publications (e.g., Hermann et al., 2019; Kearney et al., in press; Reum et al., in press). Another key outcome of this analysis will be an evaluation of which parameters and processes within the linked model most determine uncertainty; such parameters and processes could be the target of future research (see Reum et al., in press); such parameters and processes could be the target of future research.

The performance of each "climate scenario/biological model/fishing scenario" combination relative to the goals of EAFM can be evaluated using indicators of social, industry, and ecosystem status (Long et al., 2015; Levin et al., 2018). Evaluation of the performance of fishing scenarios from the multi-model suite in the ACLIM project will involve two approaches. Initially, output from each "climate scenario/biological model/fishing scenario" combination will be evaluated relative to an agreed upon suite of standard indicators previously selected by the NPFMC and its advisory bodies (Zador et al., 2017; Fissel et al., 2019; **Table 6**). Output from vulnerability assessments, whole ecosystem models, and FEAST can be used to calculate ecosystem and socioeconomic indicators (**Table 6**). Subsequently, indicators derived from ensemble of "climate scenario/biological model/fishing scenario" combinations will be evaluated. This two-step process will enable analysts to contrast the synthesized projection relative to the range of possible outcomes from models of different complexity.

#### TABLE 6 | Suite of candidate performance indicators for ACLIM.

fmars-06-00775 December 31, 2019 Time: 13:37 # 13


## INITIAL RESULTS

The results of the first phase of the ACLIM suite revealed substantial differences in projected spatially averaged air temperature in the Bering Sea based on the GFDL and the MIROC ESM with projected air temperatures differing by approximately 5◦C at end of century (see Figure 2 in Hermann et al., 2019). This result illustrates the importance of considering the ensemble projections. Under RCP 8.5, Bering Sea shelf average mean bottom temperatures may warm by as much as 5◦C by 2100, with associated loss of large zooplankton (**Figure 4**), whereas, under the lower emission scenario, bottom temperatures will warm by approximately 2.5◦C (**Figure 4**).

Results from a sub-set of the full ACLIM multi-model suite illustrate the utility of applying the ACLIM framework. Comparison of projections of future status of walleye pollock and Pacific cod from three different modeling approaches under the status quo fishing scenario under RCP 8.5 using the size spectral model (Reum et al., in press), the CE-MSM projection model (CEATTLE), and the vulnerability assessment (Spencer et al., 2019) provide an interesting contrast and exemplify the importance of the multi-model approach employed by the ACLIM team. Projections from the size spectral model that incorporated bioenergetics and predator–prey interactions indicated that future status of walleye pollock will decline, while results were more modest and mixed for Pacific cod (see Figure 4 in Reum et al., in press). The CE-MSM model incorporated temperature effects on growth and recruitment of walleye pollock and Pacific cod. This model projected warm ocean conditions will negatively impact both stocks through impacts on survival to age-1. In contrast, Spencer et al. (2019) utilized a trait based approach and expert judgment of 34 experts to assess the vulnerability of walleye pollock and Pacific cod to changing climate. Their results indicate that walleye pollock and Pacific cod exhibit numerous traits that would allow these populations to adapt to a changing climate (e.g., broad spatial distribution, mixed prey, large populations, relatively high rate of production). This preliminary comparison illustrates the importance of contrasting outcomes using a multimodel approach.

## SUMMARY

Alaska Climate Integrated Modeling is a novel multidisciplinary modeling study designed to quantify the impacts of climate change on Bering Sea species and fisheries. The management strategies used to project future capture of marine species, processing, and distribution represent a realistic suite of future options. The evaluation of alternative management strategies allows analysts to assess the performance under a range of climate change scenarios. The ACLIM framework is designed to quantify the contributions of climate forcing scenario, model parameter, and management uncertainty in projected impact assessments.

The operationalized framework developed through ACLIM aligns global projections of climate impacts on the physical biogeochemical environment with assessments of the impacts of these changes on ecosystems and humans. Ideally, the ACLIM framework would be re-employed in parallel with new climate assessments to provide climate ready fisheries

management advice that enables resiliency to a rapidly changing climate. During each climate assessment cycle, the generalized ACLIM approach will involve three steps for rapidly generating updated climate change assessments for the Bering Sea following: (1) release of new IPCC emission projections of global climate models will be downloaded and interpolated to generate boundary conditions for the high resolution regional ocean model (Bering 10K or its successor); (2) identify novel management challenges that require climatespecific MSEs. In each of these cases (or both combined) a new climate assessment will be initiated; and (3) with input from stakeholders and fisheries management councils, various harvest and management strategies will be used to iteratively develop and refine MSEs. This will enhance the global assessment of climate impacts on the world's oceans as well as regional management actions to ensure climate resilience for the Bering Sea ecosystem and the fishing industry it supports.

Identifying harvest strategies that perform well under nonstationary environmental conditions is a challenging (Szuwalski and Hollowed, 2016). Recent studies indicate that ecosystem dynamics can substantially influence optimal harvest strategies in multi-species fisheries (Kasperski, 2015) and impact the cost of harvesting commercial species (Haynie and Pfeiffer, 2013), thus climate-driven changes to predation and production could alter future optimal harvest strategies. The proposed iterative ACLIM framework conducted on a ∼5 year cycle is modeled after the annual stock assessment cycle in the region; the approach should ensure that fisheries management decisions account for climate-driven changes to fish production and distribution and that climate-ready fisheries management in the region reflects the most recent global climate and carbon emission projections and best available ecosystem and socioeconomic science.

The ACLIM modeling framework is designed to inform the NPFMC of the performance of current and alternative management approaches in a changing climate. The scenarios will help to identify and test climate-resilient management options (Holsman et al., 2019). The Climate Action Module within the Bering Sea FEP provides a vehicle for communicating results to managers, stakeholders and the public. The scope of the framework serves to integrate the socio-ecological research community providing a forum for improving and adapting the framework. The near future projections (2030–2050) provide useful information regarding how changing climate will affect peoples' livelihoods, longer term projections inform the public of what is at risk in the region.

Five key messages have emerged from first phase of ACLIM. Structural uncertainty in ESMs used as boundary conditions for the ROMs model is a key consideration in the assessment of climate impacts on marine resources. Comparison of projected change based on boundary conditions from different earth systems models differs strongly with differences being comparable in scale to differences stemming from different RCPs in a single model. Structural uncertainty in ecosystem complexity should be considered in regional impact assessments. Results from a subset of models from phase 1 of ACLIM revealed alternative response pathways for walleye pollock and Pacific cod. Ensuring the conservation measures currently in place in an existing EAFM system was critical to managers and stakeholders because these measures do preserve resources into the future. Aligning fishing scenarios with the evolving EAFM approach of the NPFMC requires strong community engagement.

## DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request to the corresponding author.

## AUTHOR CONTRIBUTIONS

fmars-06-00775 December 31, 2019 Time: 13:37 # 15

ABH was the lead author and a principal investigator for the project. KH was a principal investigator for the project and was also responsible for the CE-MSMs. ACH was a principal investigator for the project and was also responsible for fishing scenarios and economic modeling. AJH was a principal investigator for the project and was also responsible for downscaling ESM models using ROMS. AP was the coinvestigator responsible for CE-SSM, and CE-MSM models and management strategy evaluations. KA was the co-investigator responsible for FEAST model. JI was the co-investigator responsible for CE-SSM models of pollock. SK was the coinvestigator responsible for community impact assessment and economic modeling. WC was the co-investigator responsible for ESM model selection and climate downscaling. AF was post-doc responsible for the development of function to estimate catch estimates under status quo and alternative fishing scenarios. KK was a research scientist responsible for nutrient phytoplankton zooplankton model. JR was postdoc responsible for size spectral model. PS was the coinvestigator responsible for spatial CE-SSM for pollock that includes predator overlap impact. IS was the co-investigator responsible for CE-SSM models for flatfish. WS was the coinvestigator responsible for IBM for snow crab. CS was the co-investigator responsible for spatial CE-SSM for snow crab. GW was a graduate student responsible for CE food-web model. TW was the co-investigator responsible for CE-SSM models for flatfish.

## REFERENCES


## FUNDING

Multiple NOAA National Marine Fisheries programs provided support for ACLIM including Fisheries and the Environment (FATE), Stock Assessment Analytical Methods (SAAM) Science and Technology North Pacific Climate Regimes and Ecosystem Productivity, the Integrated Ecosystem Assessment Program (IEA), the Economics and Human Dimensions Program, NOAA's Research Transition Acceleration Program (RTAP), the Alaska Fisheries Science Center (ASFC), the Office of Oceanic and Atmospheric Research (OAR), and the National Marine Fisheries Service (NMFS). Additionally, the International Council for the Exploration of the Sea (ICES) and the North Pacific Marine Science Organization (PICES) provided support for Strategic Initiative for the Study of Climate Impacts on Marine Ecosystems (SI-CCME) and the Strategic Initiative on the Human Dimension (SIHD) workshops, which facilitated development of the ideas presented in this manuscript. This publication is partially funded by the Joint Institute for the Study of the Atmosphere and Ocean (JISAO) under NOAA Cooperative Agreement NA15OAR4320063, Contribution No. 2019–1043. This is IEA publication number 2019\_9.

## ACKNOWLEDGMENTS

We thank James Thorson and Martin Dorn for their helpful comments and suggestions that improved the manuscript. We also thank the journal reviewers for their helpful comments and suggestions. We thank Christine Stawitz for her contributions to the snow crab IBM. The scientific views, opinions, and conclusions expressed herein are solely those of the authors and do not represent the views, opinions, or conclusions of NOAA, the Department of Commerce, ICES, or PICES.



forecasts (2010-2040). Deep Sea Res. Part II Top. Stud. Oceanogr. 94, 121–139. doi: 10.1016/j.dsr2.2013.04.007


resource for studying climate change in the presence of internal climate variability. Bull. Am. Meteorol. Soc. 96, 1333–1349. doi: 10.1175/bams-d-13- 00255.1


39-year retrospective analysis of seasonal processes on the eastern Bering Sea shelf and slope. Deep Sea Res. Part II Top. Stud. Oceanogr. 134, 390–412. doi: 10.1016/j.dsr2.2016.07.009


juvenile walleye pollock in the southeastern Bering Sea. Deep Sea Res. Part II Top. Stud. Oceanogr. 134, 223–234. doi: 10.1016/j.dsr2.2016.01.003


**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.

The handling Editor declared a past co-authorship with one of the authors KH.

Copyright © 2020 Hollowed, Holsman, Haynie, Hermann, Punt, Aydin, Ianelli, Kasperski, Cheng, Faig, Kearney, Reum, Spencer, Spies, Stockhausen, Szuwalski, Whitehouse and Wilderbuer. 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.

# Climate Change and New Potential Spawning Sites for Northeast Arctic cod

Anne Britt Sandø\*, Geir Odd Johansen, Asgeir Aglen, Jan Erik Stiansen and Angelika H. H. Renner

*Institute of Marine Research, Bergen, Norway*

In this study we investigate both historical and potential future changes in the spatial distribution of spawning habitats for Northeast Arctic cod (NEA cod) based on a literature study on spawning habitats and different physical factors from a downscaled climate model. The approach to use a high resolution regional ocean model to analyze spawning sites is new and provides more details about crucial physical factors than a global low resolution model can. The model is evaluated with respect to temperature and salinity along the Norwegian coast during the last decades and shows acceptable agreement with observations. However, the model does not take into consideration biological or evolutionary factors which also have impact on choice of spawning sites. Our results from the downscaled RCP4.5 scenario suggest that the spawning sites will be shifted further northeastwards, with new locations at the Russian coast close to Murmansk over the next 50 years, where low temperatures for many decades in the last century were a limiting factor on spawning during spring. The regional model gives future temperatures above the chosen lower critical minimum value in larger areas than today and indicates that spawning will be more extensive there. Dependent on the chosen upper temperature boundary, future temperatures may become a limiting factor for spawning habitats at traditional spawning sites south of Lofoten. Finally, the observed long-term latitudinal shifts in spawning habitats along the Norwegian coast the recent decades may be indirectly linked to temperature through the latitudinal shift of the sea ice edge and the corresponding shift in available ice-free predation habitats, which control the average migration distance to the spawning sites. We therefore acknowledge that physical limitations for defining the spawning sites might be proxies for other biophysically related factors.

Keywords: Northeast Arctic cod, spawning site, shift, climate change, downscaling

## 1. INTRODUCTION

The recent warming of the oceans (Levitus et al., 2009; IPCC, 2013) has resulted in shifts in the geographical distribution of marine fish (Perry, 2005; Dulvy et al., 2008; Fossheim et al., 2015). Several factors and mechanisms, both physical and biological, determine geographical distributions, and subsequently distribution shifts, of fish stocks (Planque et al., 2011). The Northeast Arctic (NEA) cod (Gadus morhua) is one of the most important fish stocks in the Barents Sea, both because of its role in the North Atlantic ecosystem and as a major fishing resource. The

#### Edited by:

*Michael J. Fogarty, National Marine Fisheries Service (NOAA), United States*

#### Reviewed by:

*Keith Brander, Technical University of Denmark, Denmark Joanne Morgan, Department of Fisheries and Oceans, Canada*

> \*Correspondence: *Anne Britt Sandø anne.britt.sando@hi.no*

#### Specialty section:

*This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science*

> Received: *29 March 2019* Accepted: *16 January 2020* Published: *11 February 2020*

#### Citation:

*Sandø AB, Johansen GO, Aglen A, Stiansen JE and Renner AHH (2020) Climate Change and New Potential Spawning Sites for Northeast Arctic cod. Front. Mar. Sci. 7:28. doi: 10.3389/fmars.2020.00028* spawning sites have sustained a coastal fishery for thousands of years and a potential shift in the spatial distribution of the spawning sites is likely to impact current fishing activities. NEA cod has undergone distribution shifts involving most of its life history stages, parallel to the observed ocean warming during the last 3 decades (Kjesbu et al., 2014; Ingvaldsen et al., 2015; Fall et al., 2018). Considerable expansion of its distribution limits north- and eastwards in the Barents Sea has been observed following increased inflow of warm Atlantic Water into the region (Eriksen et al., 2011; Kjesbu et al., 2014; Fossheim et al., 2015). Further distribution shifts (Stenevik and Sundby, 2007; Drinkwater, 2011) and changes in total stock biomass (Årthun et al., 2018) in the NEA cod stock as a whole have been predicted due to continued ocean warming.

Fish may exhibit a remarkable variability in geographical distribution patterns on population level, and the mechanisms behind are often interacting throughout the life cycle (Planque et al., 2011; Ciannelli et al., 2014). This complicates the identification of important mechanisms, their relative importance and interactions, and thus the projection of future distribution patterns (Loots et al., 2010, 2011; Planque et al., 2011). However, during the spawning season, fish tend to aggregate at spawning sites defined by a narrower set of physical and biological factors, compared to during other life history stages (Planque et al., 2011; Ciannelli et al., 2014). NEA cod is an example of such a species (**Figure 1**), exhibiting a limited degree of spawning plasticity on the population level (Ciannelli et al., 2014; Michalsen et al., 2014). This means that it is likely to track favorable environmental conditions on local scale for spawning and subsequent survival of the offspring (Righton et al., 2010). The variability in the physical conditions at the spawning sites of NEA cod is therefore an interesting candidate for studying reasons for shifts in these sites. In a warmer future climate this relationship could result in a northward migration of the NEA cod and potentially to immigration of other cod populations adapted to other sets of environmental conditions. The center of the geographic distribution and the outer fringes of the main spawning site of NEA cod along the coast of Norway have fluctuated throughout at least the last century, where we have reliable observations showing both interannual variability in the use of specific spawning sites, as well as multidecadal distribution shifts (Opdal et al., 2008; Sundby and Nakken, 2008; Opdal and Jørgensen, 2015; Langangen et al., 2018). The alternative to shifting spawning sites in a warmer future could be an evolutionary change of the local population (Mieszkowska et al., 2009).

The NEA cod has a substantial spawning migration from the feeding grounds in the Barents Sea to the west coast of Norway. The choice of spawning location will be a trade-off between the cost of migration and reaching favorable spawning locations. This also involves the need of the larvae to drift back to suitable nursery areas while having survivable conditions on the journey. The success criteria for this cycle involves a broad spectrum of biological and environmental factors (Ellertsen et al., 1987; Sundby, 2000). In this exercise we address some of the physical properties, while acknowledging that biological and other environmental factors both are important and even dominating in a complete picture. However, it is beyond the scope of this paper to give a full evaluation, and we focus instead solely on selected physical parameters as temperature, salinity, depth, and bathymetry. From hydrographic properties sampled annually at fixed stations along the Norwegian coast since the middle of the last century we know that the temperatures have been steadily increasing over the last 3–4 decades (Skagseth et al., 2015). Effects of climate change are particularly noticeable in the Barents Sea and in the Arctic Ocean, where surface air and sea surface temperatures have increased at twice the global rate (Hansen et al., 2006; Skagseth et al., 2015; Iz, 2018). However, air and ocean temperatures show also strong multidecadal variability on timescales of 50–80 years. Under a future ocean warming scenario it is likely that the spatial distribution of spawning sites will change and follow a northeastward displacement of the isotherms. The development of regional ocean models enables precise description of the physical environment at local scale (Melsom et al., 2009). These new modeling tools therefore enable detailed analyses of physical characteristics of potential spawning sites for NEA cod that can reveal some of the physical mechanisms behind localization of NEA cod spawning sites and possible shifts of these due to climatic variability.

The general physical characteristics such as hydrography, spawning depths and bathymetry at different spawning sites of NEA cod are presented in a number of papers (Bergstad et al., 1987; Ottersen and Sundby, 2005; Righton et al., 2010; Höffle et al., 2014), as well as the large-scale variability in spawning locations (Sundby and Nakken, 2008; Opdal, 2010; Opdal and Jørgensen, 2015; Langangen et al., 2018). Two main hypotheses for the observed large-scale variability in spawning location have been presented, suggesting multidecadal climate variability (Sundby and Nakken, 2008), or demographic processes as main drivers (Opdal and Jørgensen, 2015). Langangen et al. (2018) did only find support for the climate hypothesis by combining economic and genetic data. Observations from Data Storage Tags (DST) on ambient depth and temperature experienced by individual cod contain information about the prevailing conditions at the spawning sites (Godø and Michalsen, 2000; Michalsen et al., 2014). The information about physical conditions at the spawning sites from these sources are summarized in **Table 1**. To investigate the implications of future climate change on stock development, model results are needed. Some attempts to model the future distribution of cod have been done using global climate models coupled with physiological characteristics of cod (Dahle et al., 2018). However, global climate models do not have the horizontal resolution that is needed to properly resolve relevant circulation features, hydrographic conditions, and constraints such as bottom topography at local spawning sites in the Norwegian and Barents Seas (**Figure 1**). Vestfjorden on the inside of Lofoten will for example not be more than one grid cell in a global model with the commonly used resolution of 1 degree. Global climate models must therefore be downscaled to provide the detailed hydrographic descriptions needed to project the potential spawning habitat of NEA cod in the future. To our knowledge, the use of a high resolution ocean model with abilities to describe and project variability in spawning sites of fish has not yet been

TABLE 1 | Criteria on salinity, temperature, depth and position used for NEA cod spawning in different papers of interest and in this study.


*Note that Ottersen and Sundby (2005) and Langangen et al. (2014) refer to the water masses in the transition layer* between *the cold and fresh coastal waters and the warm and saline Atlantic Water.*

explored. Hopefully modeled physical characteristics of spawning sites from downscaled models can be used to give future projections of the spatial distribution of cod spawning sites under ocean warming.

In this study we address the main question: In the future, will there be other areas with physical properties corresponding to potential spawning habitat for NEA cod which may be chosen as new spawning sites? In particular, we will address the following subquestions: (1) How do we describe the current physical characteristics of the spawning site of NEA cod with detailed numerical model output? (2) Is it possible to reproduce variability in observed spawning sites based on a physical habitat criteria from a numerical climate model? (3) What are the future projections of the spatial distribution of NEA cod spawning sites under ocean warming? To answer these questions, we combine existing knowledge about hydrographic conditions at various spawning sites and output from a regional ocean model to see to which areas the preferred water mass conditions have been and will be shifted. The paper describes the spatial distribution of potential spawning habitats for NEA cod based solely on physical factors from a downscaled climate model. The novel approach to use a high-resolution regional model that provides more detail and spatial information of the physical factors for habitat description, gives results that lead to new hypotheses for the projection of spawning sites for NEA cod under a warming ocean climate in the Barents Sea during the next 50 years. We are mainly interested in the optimum spawning habitat criteria where NEA cod prefers to spawn according to probability density functions (Righton et al., 2010), and not necessarily in the extreme spawning habitat limits. Nevertheless, to explore the sensitivity of our choice of spawning habitat criteria, we perform a sensitivity study where we step by step test the response of using different criteria, all based on numbers from the literature as given in **Table 1**.

## 2. METHODS

In this paper we define water masses along the Norwegian coast from a downscaled ocean model corresponding to the physical conditions temperature, salinity and depth characterizing the physical conditions of the spawning habitat of NEA cod. The geographic distribution of these water masses is then compared to the observed geographic distribution of NEA cod spawning sites to evaluate their ability to predict this distribution. Then the potential future shift of spawning sites under climate change are studied using output from an ocean model projection.

## 2.1. Spawning Site Characteristics for North East Arctic cod (Gadus morhua)

Physical conditions characteristic of the spawning sites of NEA cod are described and reviewed by Bergstad et al. (1987), Ottersen and Sundby (2005), Righton et al. (2010), Höffle et al. (2014), Langangen et al. (2014), and Michalsen et al. (2014). The values from the literature study are summarized in **Table 1**. The main spawning sites of NEA cod are observed along the Norwegian coast from Møre to Finnmark, but mainly centered on and along the shelves off Nordland and Troms as can be seen in **Figure 1** and Sundby and Nakken (2008). Spawning takes place in near-shore areas in March and April. Most intensive spawning is reported to occur in the transition layer between the cold and fresh coastal waters and the warm and saline Atlantic Water masses, except when this layer is very thin (Eggvin, 1933, 1934). Temperature and spawning depth are a common constraint in all the studies listed in **Table 1**. Salinity is not, and may be listed in some of them as salinity usually correlates well with temperature and therefore appears to be a varying constraint on the spawning water mass.

Our choice of physical habitat descriptors for high intensive spawning is defined as water masses within a temperature range of 4–6◦C, a salinity range of 34.0–34.9, within a depth range of 50–150 m, and limited by a maximum sea floor depth of 180 m. Modeled temperature and salinity are averaged for March and April. The spatial distribution of spawning sites is defined by model cells satisfying the characteristics given above. Volumes of spawning water masses in each horizontal grid point are determined by vertical integration of each of these grid areas. That means that the volumes of every model cube that satisfy the above criteria are summed up vertically at every horizontal grid cell in the model. Thus, based on hydrography and depth constraints applied to the output from a high resolution regional ocean model (section 2.2.1) that is downscaled from a global climate model (section 2.2.2), we can get projections of potential spawning sites 50 years into the future.

## 2.2. Model Descriptions

### 2.2.1. Regional Model

The model used for downscaling here is the regional ocean model system, ROMS, described in Shchepetkin and McWilliams (2005). The regional model is initialized from a medium resolution version of the Norwegian Earth System Model (NorESM1-M) (Bentsen et al., 2013), and results from this model are also used at the open boundaries and as atmospheric forcing. A weak relaxation with a time scale of 360 days toward NorESM sea surface salinity is also applied. ROMS is run on a stretched orthogonal curvilinear grid with an average horizontal resolution of 10 km and covers the Arctic and the Atlantic Ocean south to about 20◦ S (see Figure 2 in Sandø et al., 2014b). There are 40 generalized sigma (terrain following) levels in the vertical, applying the scheme of Song and Haidvogel (1994), with stretching that enhances the vertical resolution toward the bottom and the surface. Lateral motions and diffusive energy losses induced by small-scale processes are related to the gradients of the mean velocities and tracers by eddy viscosity and diffusivity coefficients (Smagorinsky, 1963). For advection, we use the third-order upwind biased scheme proposed by Shchepetkin and McWilliams (2008). ROMS employs split-mode explicit time stepping, and in this study, the baroclinic mode time step is 100 s, while the barotropic mode time step is 10 s.

To evaluate ROMS results directly against observed time series, a hindcast simulation forced with atmospheric forcing from the CORE2 reanalysis (Large and Yeager, 2009) from 1958 to 2008 was performed in a parallel study parallel study (Sandø et al., in preparation). Simulated volume and heat transports in different sections were compared to observation based estimates with respect to mean values and variability. The modeled mean inflows to the Nordic Seas were shown to be close to the observed mean inflows Thereafter, the same model was used to downscale the future RCP4.5 scenario from NorESM1-M for the period 2006–2070.

### 2.2.2. Choice of Global Climate Model for Downscaling

The Coupled Model Intercomparison Project Phase 5 (CMIP5) offers many global climate models that can be used for downscaling, but it is important to be aware that every model has strengths and weaknesses. Although the latest IPCC report (AR5) (IPCC, 2013) confirms the results from the previous IPCC report (AR4) about projected strong decreases in sea ice extent in the Arctic toward the end of this century, the inter-model spread is considerable. It is therefore important to evaluate different models in the region of interest, before downscaling the model that is closest to the observed values of the most relevant variables, both with respect to mean values and variability. To get an estimate of the uncertainty in the results, it is desirable to downscale an ensemble of models, but time and computational resources often put constraints on this. For this study, where heat content and sea ice extent strongly influence the regional ecosystem, evaluation of the heat transport into the Barents Sea and Arctic Ocean is of particular importance. Sandø et al. (2014a) evaluated three coupled climate models (CNRM-CM5, MRI-CGCM3, and NorESM1-M) against multiple estimates from the literature with respect to poleward heat transport through four gateways to the Arctic, and NorESM1-M transports were found to be closest to the observed mean in both Barents Sea Opening between Svalbard and Norway and in the Fram Strait between Greenland and Svalbard. These gateways are closest to the region of interest in this study, and NorESM1-M is therefore chosen for downscaling in this analysis.

The future climate is strongly dependent on the future emissions of greenhouse gases. Four different representative concentration pathways (RCPs) are used to describe a set of greenhouse gas concentration trajectories adopted by the IPCC for its fifth Assessment Report (AR5) in 2014 (IPCC, 2013). These are RCP2.6, RCP4.5, RCP6, and RCP8.5. Of these, the RCP4.5, in which the emissions peak around 2040, decline, and stabilize at an increased radiative forcing of 4.5 W m−<sup>2</sup> relative to preindustrial time, is the one used for downscaling in this study.

## 3. RESULTS

In this section, we first evaluate the hindcast simulation and its ability to reproduce the observed southerly offset in spawning water masses (**Table 1**) in cold years and the northerly offset in warm years (Sundby and Nakken, 2008; Langangen et al., 2018). Thereafter we apply the same criteria on the results from the future projection to see how spawning sites may change in the future, given that there is a close link to the hydrographic properties.

## 3.1. Spawning Sites in the Past and Associated Shifts

The main spawning site has traditionally been in Nordland and Troms, with secondary areas at the coasts of Møre in the south and Finnmark in the north as indicated in **Figure 1**. During February and March 2004 and 2005, the fishing industry reported large numbers of mature and pre-spawning cod at the fishing grounds along the coast of East Finnmark. This came after an extended period of high temperature, starting in the early 1990s (Sundby and Nakken, 2008). These observed spawning sites are comparable to the simulated sites from the hindcast simulation shown in **Figure 2**, where the potential spawning sites for cod in the cold 1960s (1965–1970) and the warm 2000s (2003– 2008) are shown by colors that indicate the volumes of water masses that correspond to the hydrography and depth criteria listed **Table 1** (4<T<6 ◦C, 34.0<S<34.9, 50 m<depth<150 m, sea floor depth<180 m). As indicated in **Figures 1**, **2** shows how the spawning sites are shifted southwards in cold years and northwards in warm years. In the cold 1960s there are larger areas at Mørebanken and along the coast up to Lofoten islands compared to the warm 2000s, and in the warm 2000s the figure indicates more spawning in Troms and also spawning sites in a narrow belt close to the coast between 20◦E and 30◦E that are not present in the cold 1960s. These southward and northward shifts in potential spawning sites are in agreement with results in Sundby and Nakken (2008) which were based on cod roe indices. In some regions, especially in Troms, our results show larger volumes of spawning water further from the coast compared to the official spawning map (**Figure 1**). These offshore spawning sites are to some degree related to the chosen maximum sea floor depth.

Time series of temperature from the hindcast simulation at locations close to Bud at Møre and Eggum outside the Lofoten

islands are compared to observations in the depth interval from 50 to 150 m in **Figure 3**. These time series show that the simulated temperatures are close to the observed ones both at Bud and Eggum, with temperatures at Eggum lying about 1◦C below those at Bud. The simulated salinities are too high compared to observations, about 0.5 higher at Bud and about 0.75 higher at Eggum. Insufficient impact or rendering of the river runoff is one explanation, and detected inaccurate instrumental measurements is another (Carvajalino-Fernández et al., 2018). Furthermore, the observed time series of salinity in **Figure 3** suggest that these salinities were too low several years compared to our criteria, but as these time series are averages from different standard depths which contain values well within our criteria, there are values at selected depths satisfying the criteria most of the time (not shown). On the other hand, the modeled salinity profiles show little variability with depth due to the insufficient representation

of coastal water in the upper water column, and the time series is therefore representative for most layers in the chosen depth interval of 50–150 m. Modeled and observed salinities in the southern part of the Barents Sea Opening, the Fugløya-Bjørnøya section, are shown in **Figure 3**. These salinities are more consistent with each other, at least with respect to mean values and biases. The instruments used here are known to be more exact, and last but not least, the area is farther from coast and less affected by runoff and fresh coastal waters. Based on the information collected in **Table 1** and used as criteria for spawning water mass in **Figure 2**, it seems plausible that Bud was not very well suited for spawning in the early 1990s and in the 2000s, while the area outside the Lofoten islands was well suited the whole period.

The observed shifts in spawning sites, may also be linked to temperature through the shifts of sea ice edge in the Barents Sea in cold and warm periods (**Figure 4**). Such shifts in the sea ice edge change the area available for predation (Sundby and Nakken, 2008; Drinkwater and Kristiansen, 2018). To find the mean position or center of geographic distribution for the spawning sites along the coast in March and April, we use the masks satisfying the spawning mass criteria for this study given in **Table 1**. Likewise, we find the mean position for ice free waters available for predation in the Barents Sea in September (**Figure 4**) by masking out grid cells of non-zero sea ice concentration between 68–82◦ N and 20–70◦ E. The distance between the average position of ice free waters in the Barents Sea during summer in cold and warm periods is here calculated to be 319 km. A corresponding shift in the spawning site center of geographic distribution is calculated to be 278 km.

## 3.2. Potential Future Change in Climate and Spawning Sites During 2010–2070

A similar analysis for potential future cod spawning sites is done for the RCP4.5 scenario. Results from the last decade of the downscaled model run, 2060–2069, are compared to the decade representing the present climate, 2010–2019, from the same model run. This future simulation is initialized and run with an atmospheric forcing from a global climate model that has a another natural variability than the hindcast September, respectively.

simulation (1958–2008), simply because the hindcast simulation is forced with an observation based atmospheric forcing. The two simulations may therefore have different biases in temperature and salinity. The results from the hindcast simulation and future projection are therefore not directly comparable, and the projection can not be considered as a continuation of the hindcast simulation.

**Figure 5** (upper) shows the simulated spawning sites in the first decade, the 2010s. Compared to the warm 2000s in the hindcast run, there are no longer any spawning sites at Møre, and the easternmost limit at the coast of Finnmark is now even further east. Looking at the last decade, the 2060s, (**Figure 5**, lower) the spawning site around the Lofoten islands has now disappeared, but the area off Finnmark has extended eastwards to the longitude of Murmansk.

From the literature study on criteria for spawning water masses summarized in **Table 1**, we experience that not all studies consider salinity at the spawning site to be important, and also the lower and upper temperature limits vary. Hoegh-Guldberg and Bruno (2010) indicate a lower limit of 3◦C, Righton et al. (2010) find that NEA cod experiences and tolerates temperatures during spawning time up to 7◦C while Michalsen et al. (2014) observe that cod during spawning time aggregate at temperatures between 4◦C and 8◦C with an average temperature around 5.5◦C, and at depths between 30 and 200 m. To test the consequences for using such criteria, we perform different sensitivity calculations with temperature limits of 3◦C and 8◦C, without any limitations with respect to salinity, and finally with a combination of these two new criteria. The resulting extent and change of spawning water masses between the 2010s and 2060s for the reference (4<T<6 ◦C, 34.0<S<34.9) and the sensitivity cases can be seen in **Figure 6**. The figure shows how the altered temperature limits change the spawning sites at Møre and in the Russian sector (**Figure 6**, lower left), while salinity has minor implications along the Norwegian coast, except for the Lofoten area where omission of this criterion imply less reduction of spawning water masses between the two periods (**Figure 6**, upper right). A combination of a longer temperature interval and no salinity criterion gives more spawning water masses in the Russian sector and less at Møre toward the 2060s (**Figure 6**, lower right). The choice of spawning depth range used here is not found to be sensitive (not shown). Time series of salinity and temperature from the projection at Bud and Eggum are presented in **Figure 7**. These show that the temperatures off Møre are too high compared to the temperature criteria in **Table 1** during this period and that also the temperatures outside the Lofoten islands at Eggum become too high after short time. The salinities at these locations are within the range most of the time with some exceptions in the middle and at the end of the integration period, and like the time series for the hindcast study, these time series reveal large interannual to decadal variability.

## 4. DISCUSSION

Our approach in this study is to reproduce historical spawning sites and shifts in these based purely on the physical criteria from a literature study summarized in **Table 1**, and thereafter apply the same method to downscaled projections of the future climate. The methodological limitations to this kind of analysis are that it is based on an exercise where the physical climate is the only explanation for the variability in spawning at the sites in consideration. There are probably both direct and indirect

causes for temperature to be important, as well as biological mechanisms. An argument supporting that hydrography is essential for spawning is that the transition layer between the relatively fresh coastal water and the more saline Atlantic Water was used as an indicator for the typical depth where the spawning NEA cod arrived the eastern Lofoten and Vestfjorden (Eggvin, 1933; Ellertsen et al., 1981). The vertical position of this layer was therefore used by the fishermen to find the depth of spawning cod.

## 4.1. Historical and Present Spawning Sites

The simulated temperatures at the fixed stations at Bud and Eggum reproduce the observed values well, at both interannual and decadal timescales (**Figure 3**), and likewise for temperatures and salinities in the Fugløya-Bjørnøya section in the Barents Sea Opening. As described in section 3.1, **Figure 2** shows how the spawning sites are shifted southwards in cold years and northwards in warm years, in agreement with observations presented in Sundby and Nakken (2008). The same figure also shows yellow spots on the periphery or outside the observed spawning sites in **Figure 1**, where the bathymetry is relatively deep compared to the depth at the traditional spawning sites. The mismatch in these areas can be associated to the method of calculating the spawning water masses which is dependent on the topography as explained in section 2.1, and the volume will therefore increase by depth as long as the conditions with respect to hydrography and maximum bottom depth are fulfilled. Apart from that, using the criteria specified in section 2.1, we are able to reproduce the latitudinal shifts of spawning sites along the Norwegian coast with shifting climate.

## 4.2. Future Potential Areas for Spawning

Repeating the analysis with output from the future scenario, the most remarkable results here are the total disappearance of the specific spawning water masses in the Lofoten area as indicated in **Figure 5**, a region known for its cod fisheries related to spawning migration through centuries. Less surprising are the new potential spawning sites outside Murmansk, based on the current knowledge about the recent Barents Sea warming and reports about large numbers of mature and pre-spawning cod at the fishing grounds along the coast of East Finnmark in the early 2000s (Sundby and Nakken, 2008). Filina and Trostyanskii (2007) also report about spawning individuals in the coastal waters of Murmansk in 1999–2003, when the temperature in the Kola section exceeded 4◦C for the first time since the late 1930s (Tereshchenko, 1999). In other words, this can be viewed as an extension of the previous trend. Shifts in the spawning center of geographic distribution following the slow, large scale changes in temperature and position of sea ice edge were well documented by Sundby and Nakken (2008), but a termination of spawning in the southern site at Møre in warm periods was not previously found. This is also in contrast to our simulated results from the cold 1960s and the warm 2000s (**Figure 2**), where there are indications of reductions and increases of spawning water masses at the respective spawning sites with temperature and salinity anomalies, not a disappearance. **Figure 7** shows that modeled temperatures at Bud are outside our defined range throughout the entire projection, and therefore explain the vanished spawning site. From the sensitivity analysis in **Figure 6** it can also be concluded that a temperature limit of 8◦C will also counteract spawning south of 63◦ N in the future. That said, the absolute upper temperature limit for Atlantic cod to spawn in the Celtic and North Sea is 9.6◦C (Meeren and Ivannikov, 2006; Kjesbu et al., 2010), so based on this, the Atlantic cod stock may still have some extra years left to spawn at Møre before it meets this limit (**Figure 7**).

Next, what is the cause for the future depletion of spawning waters in the Lofoten area? The time series for temperature at Eggum is outside the range given in **Table 1** most of the time (**Figure 7**), while that for Skrova is within, especially at the end. It is therefore reasonable to test the salinity criteria, which are suggested by only 2 of 6 studies. Omission of the salinity criteria on the water masses gives the same result as the reference case east of North Cape (**Figure 6**), but for the region

outside Lofoten the reductions in spawning sites are less without any salinity criteria, and at Skrova there is even an increase, meaning that the future climate and higher temperatures give more favorable temperatures for spawning there. So therefore, based on the reference case and the sensitivity experiments herein, the only clear conclusion to be drawn is that increased temperatures in the southern Barents Sea will lead to more suitable spawning conditions along the coast in that region, and especially outside Murmansk in the Russian sector. Dahlke et al. (2018) assessed the embryonic ranges of thermal tolerance under different RCP scenarioes and mapped the corresponding spawning habitat suitability by using CMIP5 ensemble median of maximum potential egg survival. For the RCP4.5 scenario they found that the thermally suited spawning habitat was reduced by up to 20% along the Norwegian coast and further east to about 40◦E at the Russian coast. It should be noticed that the horizontal resolution of the CMIP5 models are only 1◦ × 1 ◦ and is therefore a limitation to reproduce realistic circulation and hydrographic features at specfic spawning sites along the coast.

## 4.3. Uncertainties

A crucial uncertainty regarding the results of this study is the use of hydrography and depths as a choice of method for describing spawning sites. In addition to the direct effect of hydrography on spawning sites (Bergstad et al., 1987; Ottersen and Sundby, 2005; Hoegh-Guldberg and Bruno, 2010; Righton et al., 2010; Langangen et al., 2014), indirect effects suggest other mechanisms for changes in spawning sites (Sundby and Nakken, 2008; Opdal and Jørgensen, 2015). As shown in section 3.1, these can be related to the migration distance from the feeding area of high food abundance at the ice edge in the Barents Sea to a suitable spawning site close to the Norwegian coast (Sundby and Nakken, 2008). The distribution of cod catches from bottom trawls shown in Kjesbu et al. (2014) indicates increased catches in the northern and eastern parts of the Barents Sea where sea ice retreats in warm years. So, in warm years when the simulated sea ice edge is further north and east (**Figure 4**), the increased migration distance may be a limiting factor of how far south cod can reach at constant speed before the spawning season peaks around April 1st. The distance between the average position of

ice free waters in the Barents Sea during summer in cold and warm periods is here calculated to be 319 km. This is comparable to the shift of 278 km in the simulated spawning center of geographic distribution, and therefore supports the idea of an indirect temperature effect in terms of sea ice extent as suggested by Sundby and Nakken (2008). Another indirect effect can be faster gonad maturation in warm years (Kjesbu et al., 2010), limiting the distance cod can migrate at constant speed before it is ready to spawn.

The observed temperature, salinity and depth intervals for the transition layer between coastal and Atlantic waters at the spawning sites of NEA cod may also be a proxy of where there is sufficient food available for survival of early life history stages. From observations (Drinkwater, 2011) and modeling studies (Slagstad and Tande, 2007) Calanus finmarchicus is known to be the dominant zooplankton species along the Norwegian Shelf in spring, where they are held by eddies and mean circulation and is important as it constitutes the prey for larval and early juvenile cod (Sundby, 2000).

Another uncertainty by using this kind of method is that we don't take into account adaptation of cod to spawn at higher temperatures than the observed limits of today. If cod or its prey adapt to climate change faster than the period of interest here, then our assumptions will break down. There is no doubt that adaptation has played an important role in developing different stocks of Atlantic cod that now lives and spawn in very different habitats, but this evolution have probably happened over much longer time scales than those considered here (Mieszkowska et al., 2009).

Common to the explanations listed here, is that they are all, directly or indirectly, dependent on temperature variability. So, taking into account that the resulting spawning sites are affected by different factors, involving hydrography, distance of migration from feeding grounds or gonad maturation, we argue that the hydrography, and in particular the temperature, can be used as an indicator for potential changes in the future. Anyway, such sources of uncertainty should be kept in mind when concluding on the effects of future warming on spawning sites.

There are also uncertainties with respect to the simulated future climate. According to Hawkins and Sutton (2009), such uncertainties strongly depend on three parts, namely model errors, internal or natural variability in the climate system, and future scenarios on emissions of greenhouse gases. On interannual to decadal time scales, the natural variability is much bigger than the effects of anthropogenic emissions of greenhouse gases on climate change, but as the contributions from anthropogenic emissions are positive every year, the effects of these emissions are substantial after some decades. Hawkins and Sutton (2009) therefore find that on regional scale, the internal variability and model errors dominate in the first period of about 20 years. After this, the uncertainties due to internal variability are strongly reduced, and toward the end of a century-long projection, uncertainties due to future emissions are totally dominating.

## 4.4. Possible Impacts of Spawning Site Shifts

A question rising from the analysis done here, is how increased temperature will impact successful spawning and further survival of 0-group cod. The recent warming in the Barents Sea has both led to a shift in spawning sites, and to a change in the spatial distribution of fish communities with a northward expansion of boreal species at a pace reflecting the local climate change (Kjesbu et al., 2014; Fossheim et al., 2015). As for spawning sites, indications about altered distributions of species at different trophic layers in the future can be found based on a combination of changes in water masses as simulated herein and already known effects of climate on ecosystem dynamics as described in e.g., Drinkwater (2011), Johannesen et al. (2012), Kjesbu et al. (2014), and Fossheim et al. (2015). That said, can such effects from present day climate be extrapolated into the future?

Prerequisites for survival of cod larvae is that they are spawned in an upstream water mass where they can drift into a suitable nursery area (Ådlandsvik, 1989), and that there is sufficient food for them as they are drifting (Ellertsen et al., 1987). Lofoten has up to now been such an appropriate place with optimal hydrographic spawning conditions, subsequent drifting by the ocean current into the Barents Sea, and plenty of food on their way in terms of Calanus finmarchicus (Ellertsen et al., 1987). Sundby et al. (2016) define the North Atlantic adjacent to the Polar Circle with its spring bloom system as a critical region due to the seasonal light cycle which sets particular demands on planktivorous species. Planktivorous species such as Calanus finmarchicus deposits lipids during the short spring bloom period and are therefore able to overwinter at great depths during winter when phytoplankton is insufficient. Therefore, if the spawning and drifting areas are invaded by more temperated species from further south that are not able to adapt to such a seasonal life cycle, it might become a problem for the drifting larvae. Furthermore, what will be the fate of the eggs that potentially will be spawn in Russian waters outside Murmansk in the future? Will they drift into an area of sufficient food abundance? **Figure 4** indicates that eggs spawn in that area will drift northeastwards west of Novaya Zemlya toward the sea ice edge and remain in the Barents Sea. In a parallel study, (Sandø et al., in preparation), results from an end-to-end ecosystem model, NORWECOM.e2e (Skogen et al., 2018), forced with the same physical output from the RCP4.5 scenario as analyzed in this study, show that areas where sea ice concentration decreases will have an increase in both primary and secondary production (Sandø et al., in preparation). The simulations show that there will be a change to more Atlantic characteristics (T>3 ◦C) in the eastern Barents Sea up to the northern tip of Novaya Zemlya. These results therefore indicate that spawning and drifting, and subsequent survival may be successful in the eastern Barents Sea, but a more comprehensive study needs to be done to conclude on this. Such a combination of physics from a downscaled climate model for a future scenario with chemical and biological model components as in NORWECOM.e2e (Skogen et al., 2018) is an example of how further knowledge about potential climate impacts on the marine ecosystem can be gained and will be one of the main perspectives in our future work.

## 5. CONCLUSIONS

A regional ocean model, ROMS, has been used to describe the physical characteristics of NEA cod spawning sites for the period 1958–2008, and similarly for projections into the future. The physical criteria are collected from a literature study and include hydrographic properties of spawning waters, spawning depth and bottom topography. Based on this method we are able to reproduce a long-term spatial shift of the mean position of the spawning sites, with a southern displacement in cold years (1965–1970) and a northern displacement in warms years (2003– 2008). Applying the same method on results from a downscaled future scenario we find that the spawning sites are shifted further northeastwards, and with new locations at the coast close to Murmansk 50 in years. Dependent on whether salinity is important for the spawning habitat or not, future freshening may lead to additional reduction of the spawning habitat in the Lofoten area.

The mechanisms for these shifts can be linked to the temperature change in two ways. Low temperatures have up to now been a limiting factor east of Finnmark in the southern Barents Sea during spawning in March and April. In the future scenario, global warming leads to increased occurrences of waters warmer than 4◦C in this region, and spawning will probably take place more often and to a greater extent than today. Dependent on the maximum temperature for spawning, temperature may be a limiting factor for spawning habitats at Møre. It has also been shown that the observed and simulated long-term shift in spawning habitats along the Norwegian coast can be linked to temperature through the latitudinal shift of the sea ice edge and the corresponding shift in predation habitats in the Barents Sea in September, which in that way control the maximum southward migration distance. Therefore, while acknowledging that the location of spawning sites can be indirectly related to biophysical processes as migration distance and appropriate larval prey, our results indicate that direct physical limitations may work as criteria in future projections of spawning sites in a moderate emission scenario and at time scales as considered here.

## DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request to the corresponding author.

## AUTHOR CONTRIBUTIONS

AS has contributed with evaluation of model and model analyses, literature studies, making figures, and essential writing of paper. GJ has contributed to the model analyses, literature studies, writing and structuring of the paper. AA has contributed with constructive discussions and feedback on the paper writing. JS has contributed to the model analyses, constructive discussions, and feedback on the paper writing. AR has contributed with constructive discussions and feedback on the paper writing.

## FUNDING

This work was supported by the Norwegian Research Council Projects ArcChange (Grant: 257630) and Stockshift (Grant: 257614), by the Trond Mohn Foundation (Project number: BFS2018TMT01), and by the UNINETT Sigma2 AS through a grant of computing time.

## ACKNOWLEDGMENTS

We would like to thank Per Arne Horneland for contributions to the map on observed spawning sites.

## REFERENCES


**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 © 2020 Sandø, Johansen, Aglen, Stiansen and Renner. 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.

# Trans-Tasman Cumulative Effects Management: A Comparative Study

Kathryn K. Davies<sup>1</sup>† , Karen T. Fisher<sup>2</sup> , Gemma Couzens<sup>3</sup> , Andrew Allison1,2,4 , Elizabeth Ingrid van Putten5,6, Jeffrey M. Dambacher<sup>5</sup> , Melissa Foley<sup>7</sup>† and Carolyn J. Lundquist1,4 \*

<sup>1</sup> National Institute of Water and Atmospheric Research, Hamilton, New Zealand, <sup>2</sup> School of Environment, The University of Auckland, Auckland, New Zealand, <sup>3</sup> Ministry for the Environment, Wellington, New Zealand, <sup>4</sup> Institute of Marine Science, The University of Auckland, Auckland, New Zealand, <sup>5</sup> Oceans and Atmosphere, Commonwealth Scientific and Industrial Research Organisation, Hobart, TAS, Australia, <sup>6</sup> Centre for Marine Socioecology, University of Tasmania, Hobart, TAS, Australia, <sup>7</sup> Research and Evaluation Unit, Auckland Council, Auckland, New Zealand

### Edited by:

Erik Olsen, Norwegian Institute of Marine Research (IMR), Norway

#### Reviewed by:

Miguel Dino Fortes, University of the Philippines Diliman, Philippines Andrew Kenny, Centre for Environment, Fisheries and Aquaculture Science (CEFAS), United Kingdom

> \*Correspondence: Carolyn J. Lundquist carolyn.lundquist@niwa.co.nz

#### †Present address:

Kathryn K. Davies, University of Utah, Salt Lake City, UT, United States Melissa Foley, San Francisco Estuary Institute, Richmond, VA, United States

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 01 May 2019 Accepted: 14 January 2020 Published: 18 February 2020

#### Citation:

Davies KK, Fisher KT, Couzens G, Allison A, van Putten EI, Dambacher JM, Foley M and Lundquist CJ (2020) Trans-Tasman Cumulative Effects Management: A Comparative Study. Front. Mar. Sci. 7:25. doi: 10.3389/fmars.2020.00025 Managing the cumulative effects (CE) that arise from human and natural stressors is one of the most urgent and complex problems facing coastal and marine decision makers today. In the absence of effective processes, models, and political will, decisionmakers struggle to implement management strategies that effectively tackle cumulative effects. Emerging efforts to address cumulative effects provide a timely opportunity to assess the efficacy of a range of management strategies operating at different scales and in different legislative and cultural contexts. Using primarily qualitative methodologies including literature reviews, focus groups, and workshops, this paper compares cumulative effects approaches within the Reef 2050 Plan for the Great Barrier Reef Marine Park (GBRMP), Australia, with those in Aotearoa New Zealand (Aotearoa NZ). Both case studies illustrate that cumulative effects management is especially complicated by: fragmented legislative regimes and institutions that cannot account for cross-scale or cross-sector interactions; chronic data scarcity and high levels of uncertainty that make system-based assessments and predictions challenging; and often conflicting societal and economic expectations, values, and rights that are poorly integrated into management decision-making. By considering how these two cases align with transformational change characteristics, we draw several conclusions and establish priority actions regarding (1) how to mobilise resources and political will to address CE, (2) how to deal with data scarcity and uncertainty, and (3) how to promote comprehensive and inclusive CE management of coastal and marine areas.

Keywords: Aotearoa, Australia, cumulative effects, cumulative impacts, ecosystem-based management, governance, Great Barrier Reef, New Zealand

## INTRODUCTION

Managing for cumulative effects (CE) in coastal and marine systems is confounded by many issues operating across a variety of spatial and temporal scales. There is a general recognition that collaboration between key institutes and stakeholders is needed to produce successful CE governance and management (Halpern and Fujita, 2013; Murray et al., 2014; Mach et al., 2015; Lundquist et al., 2016), but there is little discussion of the aspirational or negotiated elements

**224**

of the relationships between institutes and stakeholders required to develop common visions for CE management. Meanwhile ongoing degradation of marine environments resulting from the inadequate management of CE has led to the degradation or loss of resources (Foley et al., 2017), and created uncertainty for investors (Davies et al., 2018a). To adequately respond to CE, a more strategic approach is needed that considers, and where possible aligns, legislative frameworks and institutional practices, data collection and assessment, and consideration of values and rights in decision making across multiple scales and sectors (Le Heron et al., 2016).

For the purposes of our research, and in light of the myriad definitions of cumulative effects (or cumulative impacts) in international research, we define CE here as the effects of stressors that overlap in space and/or time (e.g., caused by a single repeated stressor or multiple stressors) (Davies et al., 2018a). This high-level definition, we argue, provides us enough clarity and flexibility to guide the cross-boundary discussions that are needed to address CE management in coastal and marine areas. Moreover, we distinguish CE assessment from CE management; while assessment of CE is becoming increasingly common in human-environmental systems (Halpern et al., 2015; Korpinen and Andersen, 2016; Stelzenmüller et al., 2018), CE assessment is often tenuously linked to, and does not necessarily instigate, CE management of coastal and marine systems (Stelzenmüller et al., 2018).

Reviews of CE suggest three key challenges to the implementation of CE management: (1) fragmented legislative regimes, which makes consideration of multi-scale interactions difficult (Therivel and Ross, 2007; Canter and Ross, 2010); (2) a lack of standardised, long-term ecological-scale data and modelling capability, which makes system-based assessments and predictions challenging to undertake (Sheaves et al., 2016); and (3) poor integration of socio-economic and cultural values, and Indigenous rights into management decision-making, which can lead to short-term planning horizons and high levels of conflict (Goldberg et al., 2016). One study identified a shared vision, the capacity to work across institutions, and a set of national scale guidelines as some of the perceived key transformative elements needed to effectively address CE in Aotearoa New Zealand (Aotearoa NZ) (Davies et al., 2018a). While none of these CE challenges are surprising, and they are well documented in the international literature, this paper focuses on understanding why these issues still exist, why there is not more impetus to address them, and how we might prompt more effective action on CE management in the future.

Building on (Davies et al., 2018a), we conducted a comparative study of CE policies and practices in Aotearoa NZ and in the Great Barrier Reef Marine Park (GBRMP), Australia, to elucidate how challenges to cumulative effects management might be overcome and to identify leverage points that are likely to apply across international contexts. These two studies operate on similar geographic and spatial scales and both have undergone and are undergoing processes of co-management of coastal and marine areas between government and Traditional Owners (in Australia) and Iwi (Indigenous tribes in Aotearoa NZ). However, a number of significant differences between the two cases, including the level, nature and role of government, governance practices, management approaches, public and commercial access, and resource use and allocation make for instructive comparisons. The case studies also facilitate comparison of how a comprehensive CE policy could be implemented both in the presence of a well-publicised and acknowledged environmental disaster, and in the absence of a clear driver (where CE management is potentially a low political and/or social priority).

The Aotearoa NZ case study involved a co-developed research project entitled "Navigating the implementation impasse: enabling interagency collaboration on cumulative effects," funded as part of the Sustainable Seas National Science Challenge<sup>1</sup> . In this case, research partners from Aotearoa NZ universities, research institutes, Maori consultancies and charitable trusts, government ¯ agencies, ministries, and private enterprises were mobilised to look more closely at how to undertake CE management in Aotearoa NZ. The Australian case study (GBRMP) illustrates a primarily reactive rather than proactive CE approach, where significant efforts have been made in recent years to address CE as a result of several crises that have focussed more attention on the degradation of the reef (e.g., McCook, 1999; De'ath et al., 2012; Ainsworth et al., 2016), as well as the economic and cultural effects of degradation (De Valck and Rolfe, 2018; Marshall et al., 2019).

Our comparative analysis seeks to identify some impetus for changing behaviours and management in the marine environment, particularly in cases where political will is lacking [political will is defined broadly here as "the extent of committed support among key decision makers for a particular policy solution to a particular problem" (Post et al., 2010)], or no significant environmental disaster has occurred to force the acknowledgement of ecosystem degradation. From this work, inferences can be drawn regarding (1) how to mobilise resources and political will to address CE, (2) how to deal with data scarcity and uncertainty, and (3) how to promote comprehensive and inclusive CE management of coastal and marine areas.

## BACKGROUND

Ecosystem-based management (EBM) of coastal and marine areas aims to enhance the resilience, health and productivity of an interconnected social-ecological system (SES) through integration of policy and management of multiple uses and users (McLeod et al., 2005; Cormier et al., 2017; Gelcich et al., 2018). Consideration of governance practices is a primary consideration when looking to implement EBM, allowing for separation of the "governing system" from the "system being governed" (Fanning et al., 2007). However, while EBM principles are regularly referred to in national and international policy documents (Hewitt et al., 2018), references to EBM principles do not necessarily lead to their implementation in national policies (Gelcich et al., 2018; Sander, 2018).

Within the principles of EBM (Arkema et al., 2006; Gelcich et al., 2018), this paper focusses on the governance and

<sup>1</sup>https://sustainableseaschallenge.co.nz/programmes/our-seas/navigatingimplementation-impasse

management of CE in coastal and marine areas, with a particular interest in understanding how to transform current arrangements so that they enhance the resilience, health and productivity of the SES. To transform the current system of CE management and governance, a novel suite of configurations must be introduced; this new system would necessarily consist of new components and ways of governing CE, and thus have the potential to change system state variables, scales of key cycles, and the structures and processes that provide feedback (Olsson et al., 2006). Transformational change is likely to involve changes in perceptions, meanings and configurations of networks including leadership, power relations and institutional arrangements and structures (Folke et al., 2010).

Potential and actual CE can vary both spatially and temporally. In Aotearoa NZ, CE include a wide range of impacts on the coastal and offshore marine environment (MacDiarmid et al., 2012). In both coastal and offshore marine ecosystems, commercial, recreational and customary fishing directly impact on species and food webs through resource extraction, and for some fishing methods, result in significant disturbance to biogenic habitats on the seafloor. Other resource industries (oil and gas, mineral extraction, sand mining) and aquaculture as well as non-extractive industries (e.g., tourism, transport) may also impact on marine ecosystems. Climate change (temperature, sea level rise, increasing storm and wave events, ocean acidification) also has large impacts. Coastal marine systems are also impacted by sediments, nutrients and other pollutants derived from landbased activities. In the GBRMP, CE operate on local and global scales (Ortiz et al., 2018), including nitrogen inputs (Fraser et al., 2017), crown-of-thorns starfish outbreaks (Fraser et al., 2017; Vercelloni et al., 2017; Ortiz et al., 2018), climate change and extreme weather events such as cyclones (Fuentes et al., 2011; Ortiz et al., 2018), and degraded water quality and warming leading to coral bleaching (Ortiz et al., 2018).

## Statutory Context

In Aotearoa NZ, coastal and marine management is covered by 25 statutes across 14 agencies and seven spatial jurisdictions (Bremer and Glavovic, 2013; Brake and Peart, 2015). Under the Resource Management Act 1991 (RMA) (Ministry for the Environment [MfE], 1991), responsibility for the sustainable and integrated management of marine natural resources (with the exception of fisheries) in the territorial sea (low water to 12 nautical miles) is devolved to regional and district councils (Severinsen and Peart, 2018). Sustainable management of natural resources (again with the exception of fisheries) within the Exclusive Economic Zone and on the continental shelf (from 12 to 200 nautical miles) is regulated by the Exclusive Economic Zone and Continental Shelf (Environmental Effects) Act (EEZ Act) 2012 (Ministry for the Environment [MfE], 2012). Other activities not covered by these two acts include maritime transport, submarine cables, and marine reserves; these activities are addressed under a variety of other Acts. While the need to avoid, remedy or mitigate cumulative effects is legislated within many of these Acts (e.g., the RMA, EEZ Act, and Fisheries Act), coordinated and consistent definitions and response to CE is lacking in Aotearoa NZ. Central government support for implementation, coordination, and collaboration across all levels of government is required to address CE in the marine environment (Bess, 2010), however, the fragmented approach to CE management in Aotearoa NZ makes consideration of multi-scale interactions challenging. This fragmentation is found globally; the scale at which management occurs often is not in alignment with the scale at which a problem occurs (Cumming et al., 2006; Guerrero et al., 2013). Short-sighted decision making is enhanced by the relatively short timeframe of political cycles, which hinder the formation and implementation of long-term management plans (Guerrero et al., 2013; Weeks et al., 2015).

In Australia, the GBRMP was established under the Great Barrier Reef Marine Park Act 1975; an intergovernmental agreement, the Offshore Constitutional Settlement, was put in place in 1979 to help protect the GBRMP (Hassan and Alam, 2019). The GBRMP was designated as a United Nations Educational, Scientific and Cultural Organization (UNESCO) World Heritage site in 1981. Due to continuing declines in reef health, UNESCO raised concerns over the state of the reef in 2012. This instigated the preparation of the Reef 2050 Long-Term Sustainability Plan (Reef 2050 Plan), which involved three levels of government working together to develop joint policies to control all activities that impact the marine park, including activities located outside the boundaries of the GBRMP. Strategic assessments, an integrated monitoring framework, and a review of protection mechanisms were associated with the development of the Reef 2050 Plan; these were carried out in partnership with traditional owners (TOs) and industry between 2012 and 2015 (Commonwealth of Australia, 2018). The following 2 years (2016/2017) saw devastating coral bleaching events occur within the GBRMP. In response to the bleaching events and the impact of Tropical Cyclone Debbie in 2017, an updated Reef 2050 Plan was expedited and released by the Australian and Queensland governments in July 2018 and is the overarching framework for protecting and managing the Reef until 2050 (Great Barrier Reef Marine Park Authority, and Queensland Government, 2018). The Australian Government, the Great Barrier Reef Marine Park Authority (GBRMPA) and the Queensland Government will lead implementation of the Reef 2050 Plan to protect the Outstanding Universal Value of the Reef. The Plan builds upon, and does not replace, the existing statutory and foundational management arrangements for the World Heritage site.

The Reef 2050 Plan responds to the pressures facing the Reef and aims to address cumulative impacts and increase the Reef's resilience to longer-term threats such as climate change. In recognition of the need to manage cumulative impacts (as outlined in the legally mandated Strategic Environmental Assessment undertaken jointly by GBRMPA and the Queensland Government), a review of current understanding with respect to cumulative impact management and application for management was undertaken in 2017. A suite of supporting policies and programms include the Reef 2050 Cumulative Impact Management Policy and Net Benefit Policy passed in July 2018 (Australian Government, 2018); these two documents, along with the Good Practice Management for the Great Barrier Reef document, are guidance materials to support implementation of the Reef 2050 Plan. In addition, the Reef 2050 Integrated

Monitoring and Reporting Programm is a key part of the Reef 2050 Plan and will track the progress of the Reef 2050 Plan's outcomes and targets.

## Co-governance and Co-management Arrangements

The involvement of Indigenous peoples in the management of CE is a feature of resource management in both Aotearoa NZ and GBRMP. In Aotearoa NZ, the Treaty of Waitangi signed in 1840 shapes the nature of the relationship between Iwi (Indigenous Maori tribes) and the Crown, whereby the Crown has obligations ¯ to recognise and provide for the rights of Maori under the Treaty. ¯ According to the Treaty and contemporary resource management regulations, tangata whenua (local Indigenous people) have the right to exercise kaitiakitanga (Maori stewardship according ¯ to their own aspirations and practices). Maori, therefore, have ¯ the right to be included in planning and decision making for natural resources through co-management and co-governance arrangements (Harmsworth et al., 2016; Lundquist et al., 2016; Webster and Cheyne, 2017), though statutory requirements vary substantially from consultation to co-governance across different legislative Acts and institutional practices (Joseph et al., 2018). Maori rights under the Treaty extend to the right ¯ to redress for Crown breaches of the Treaty (Harmsworth et al., 2016). Treaty claim settlements can include a formal apology, financial reparations for loss of land and resources, the right to first purchase of government infrastructure (such as airports and public land), and recognition of the groups' cultural association with specific lands and waters. Evolving recognition of Treaty rights and obligations toward stronger co-governance arrangements present opportunities for both matauranga M ¯ aori (M ¯ aori Indigenous knowledge systems) ¯ and scientific knowledge to contribute to the evolution and enhancement of sustainable management goals and practices (Jollands and Harmsworth, 2007; Henwood and Henwood, 2011). Matauranga M ¯ aori offers a holistic world view that ¯ emphasises relationality, interconnectedness, and the cultural and metaphysical dimensions of place (Harmsworth and Awatere, 2013; Clapcott et al., 2018).

Indigenous rights in Australia have a different history of implementation. Since the mid 1970's, there have been many important events that have contributed to the current state of sea-country management in and around the Great Barrier Reef (see **Figure 1** in Dale et al., 2016). The ground-breaking "Mabo" decision (1992) acknowledged the rights of Indigenous peoples as the original occupants of Australia in the court system; the Native Title Act (1993) and "The Croker Island" decision (2001) established Indigenous rights to traditionally owned seacountry (Nursey-Bray and Rist, 2009; Dale et al., 2016). In the late 1990's, when plans were proposed to stop declining dugong populations which impacted on Traditional Owners abilities to harvest dugong from their sea-country, Traditional Use Marine Resource Agreements (TUMRAs) arose as a possible resolution to the tension around dugong and sea turtle harvesting (Dale et al., 2016). The establishment of the first TUMRA with the Girringun Community in 2005 enacted a co-management regime for the first time on the Great Barrier Reef (Nursey-Bray and Rist, 2009; Dale et al., 2016).

TUMRAs and Marine Park Indigenous Land Use Agreements (ILUAs) continue to be enacted to provide space for consideration of Indigenous perspectives and practices in the management, monitoring and compliance programms. TUMRAs in the GBRMP operate for a set timeframe after agreement between traditional owner groups, the GBRMPA (the lead authority responsible for the management of GBRMP) and the Department of National Parks, Recreation, Sport and Racing. In practice, these agreements tend to be descriptive documents outlining the role of traditional owner groups in management rather than facilitating co-management of Indigenous groups' traditional land and sea country.

## MATERIALS AND METHODS

We conducted a comparative study of CE policies and practices in GBRMP and Aotearoa NZ using qualitative methods to collect and analyse data. Workshops and focus groups were employed to explore CE assessment and management frameworks, especially policies and practices, and the Trans-Tasman similarities and differences in addressing this global challenge. A review of international literature, focussed on understanding existing CE assessment and management frameworks, informed the design of workshops, focus groups and subsequent data analysis. Specifically, a range of policy documents were analysed using qualitative content analysis (e.g., Irvine et al., 2013; Takala et al., 2019; **Table 1**) to understand the governance and legislative arrangements regulating activities and the management of CE in Aotearoa NZ and the GBRMP, and to provide further context for the analysis (Charmaz, 2014).

The workshops and focus groups that formed the fieldwork portion of this research were designed to provide spaces for collaboration, co-learning and co-production of knowledge among scientists, practitioners, Maori and stakeholders with ¯ expertise and interests in CE in the marine environment (Le Heron et al., 2016). Study participants therefore comprised a purposive sample who could provide meaningful reflections on the topic of Trans-Tasman CE management. Utilising this approach meant that while meaningful and robust findings could emerge in relation to the study context, care had to be taken when making broad generalisations as a result of the research (Yin, 2011). Details of the workshop and focus group procedures are summarised in **Table 2**. All research procedures were approved by the NIWA Human Research Ethics Process prior to fieldwork and were performed in compliance with relevant human research ethics laws and institutional guidelines. All workshops and focus groups were led by experienced facilitators.

The primary Aotearoa NZ case study workshop was conducted in Wellington, New Zealand, and included 14 representatives from diverse backgrounds in local government, central government, industry, research organisations, Maori ¯ organisations, and Maori interests. This workshop focussed ¯ on understanding how CE are currently managed in light of legislative requirements and mechanisms, as well as identifying

FIGURE 1 | Reef 2050 Plan adaptive management framework (Adapted from Commonwealth of Australia, 2018: 77).

TABLE 1 | Policy documents analysed to understand the governance and legislative arrangements regulating activities and the management of cumulative effects in Aotearoa NZ and the GBRMP.




management practices that fall outside of formal requirements; for example, the practice and exercise of kaitiakitanga among Maori, and non-statutory co-governance and co-management ¯ arrangements. This workshop also sought to investigate some of the key drivers, pressures, values, and possible responses to CE that might align across scales, including identification of factors that can help or hinder resource managers in managing CE.

The Australian case study was comprised of two components; the first was a Trans-Tasman collaborative workshop held in Hobart, Australia which involved 10 representatives with different disciplinary backgrounds and expertise (seven from the Australian context and three from the Aotearoa NZ context). This workshop explored a broad CE research agenda, focussing on understanding scientific approaches to assessing CE, as well as legislative and governance arrangements to specifically manage CE in the GBRMP. One of the breakout sessions focussed on identifying and comparing helping and hindering factors in GBRMP and Aotearoa NZ to begin to elucidate commonalities and differences across the two case study locations.

The second Australian component was a focus group conducted in Brisbane, Australia, with five representatives from GBRMPA. Due to the interests and expertise of the participants, the Brisbane focus group concentrated primarily on questions regarding CE management and governance implementation. The discussion centred on the governance and management of the GBRMP and implementation of the Reef Plan, including the various actions and management plans preceding the Reef Plan. This included a discussion of the inter-governmental and regulatory arrangements (Federal Government and Queensland State Government) to manage impacts of activities in the GBR and catchment area.

Summary notes from each workshop/focus group were taken by the researchers and/or a research assistant, compiled into a single document and provided to participants for clarification and amendment. This document was then analysed using constructivist grounded theory (Charmaz, 2014) to guide the comparative analysis of CE approaches across the two cases. The data were coded using QSR International NVivo 12 software and memoing was used to record the process of emergence and relationships between themes (Strauss and Corbin, 1990). Four central themes associated with CE management policies and practices emerged from the data analysis and are considered throughout this paper (see **Tables 3**, **4** for details):


Each theme links to at least one of the three key CE management challenges described in the introduction. Our initial analysis of these themes highlighted factors that hinder and/or help progress effective CE management across diverse cycles, stages, timeframes, actors, scales, and cultures of governance and decision making (after Jann and Wegrich, 2007). We then conducted a further analysis to consider whether and how CE


TABLE 3




management and/or governance transformations have occurred and in what ways these transformations have addressed the three key CE management challenges. The phases and characteristics of transformational change that we evaluate our data against are outlined in **Table 5**.

## RESULTS

## Legislative Framings

Participants at the focus group and workshops for both case studies identified legislative and sectoral fragmentation as a key challenge for CE management, but also identified a number of other aspects of legislative framing that influenced their ability to effectively manage for CE (either positively or negatively). Participants spent time discussing the need for political will to address CE (**Tables 3**, **4**), describing it as a crucial element of CE management implementation. Although the term "political will" is a notoriously slippery concept to define, we are confident that participants were using the term in alignment with the definition provided by Post et al. (2010). In the GBRMP context, the importance of having an external driver (UNESCO) and major environmental events such as coral bleaching were repeatedly emphasised in relation to generating the political will to address CE. In the Aotearoa NZ context, the lack of political will was perceived to be a major constraint to making progress on CE management; this was linked to the absence of external drivers that might influence CE management, and the influence of short-term economic drivers. However, a growing awareness of the problems associated with CE was seen as a helping factor that could encourage the development of more political will to address CE.

Having the legal mandate to address CE was another key factor that participants raised at all of the workshops/focus groups (**Tables 3**, **4**). Although there is a legal mandate in Aotearoa NZ for CE management under the RMA, guidance on how this should be undertaken was missing, and it has therefore remained a low priority. Participants suggested this was due to the previously mentioned lack of political will. Some participants also pointed out that mismatches between having a legal imperative to progress CE management, having a moral inclination to do so, and having the support of the government to take action on CE were difficult to navigate. However, participants pointed out that the fact that some sense of obligation did exist was probably a positive sign. In the GBRMP, a legislative requirement designed to specifically address CE was not in place until the Reef 2050 Plan was passed in 2018. Again, the passing of this plan was largely attributed to the increased political will that emerged in recent years as a result of several external drivers. Participants from Aotearoa NZ and Australia rued the fragmented management and governance regimes that have made it difficult to coordinate a CE mandate that stretches across scales. They also recognised that practitioner expertise was often local/regional, while the guidance came from a higher scale (state/national). This disconnect has been difficult to overcome except through the establishment of hard-earned, long suffering collaborative efforts (as in the efforts to establish the Reef 2050 Plan).

TABLE 4


Continued

TABLE 5 | Evaluation criteria associated with transformational change in CE management and governance systems.


The role of independent voices and leadership was also discussed by many participants, especially in the Australian contexts (**Tables 3**, **4**). GBRMPA is an independent organisation, which means that while it must operate within the constraints of its mandated authority, it can raise issues of importance without facing political repercussions (e.g., release position statements on issues that are technically outside its jurisdiction such as climate change). Participants described this independence as crucial to getting the best possible CE management plan in place. From the Aotearoa NZ perspective, there has been little discussion about CE management until recently and therefore no political champion has yet emerged to drive action on CE management. Approaches to CE assessment and management have thus far largely relied on case law<sup>2</sup> related to resource decisions made through the courts (Milne, 2008) and regional efforts by councils or community groups.

## Data, Systematic Assessments, and Uncertainty

The kinds of reporting needed to develop CE management plans, guidelines, and assessments were discussed at length in all workshops and focus groups (**Tables 3**, **4**). The GBR context is much further along this path than Aotearoa NZ. The Australian Institute of Marine Science (AIMS) Long-Term Monitoring Program (LTMP) for the GBR has been going since 1993 and a long-term monitoring programm that focusses on social and economic data for the GBR was instigated in 2013 (Marshall et al., 2013). This concerted effort in data collection has enabled the GBR to make relatively speedy progress on developing assessments, risk management plans, and other related CE guidelines. The GBRMPA is required to produce a summary report that collates this monitoring data at least every 5 years.

While the New Zealand Ministry for the Environment (MfE) has started to produce national-scale marine reports every 3 years (Ministry for the Environment [MfE], 2015), there is little cohesive long-term data or other standardised monitoring data for these reports to collate, making these efforts an important but limited step for Aotearoa NZ. The 2016 MfE report, Our Marine Environment 2016, begins with a discussion of the uncertainty facing coastal and marine systems in Aotearoa NZ due to a lack of data:

"We cannot quantify the state of marine habitats at a national level, or the full ecological impacts of commercial, recreational or customary fishing on coastal and open ocean ecosystems" (Ministry for the Environment & Statistics NZ, 2016: 8).

The lack of credible metrics/indicators associated with CE has constrained efforts in both countries (**Tables 3**, **4**). The Secretariat of the Pacific Regional Environment Programme (SPREP) has instigated a large-scale project to create core indicators, which has started to address this gap in the GBR. Meanwhile, a range of projects are in development in Aotearoa NZ to investigate the suitability of different indicators for environmental health that could be used to inform CE management and assessment (e.g., Department of Conservation, 2000). However, participants from both Aotearoa NZ and GBR contexts admitted there was likely more data available than anyone realised, and an essential part of the work will be to take adequate time to collate and analyse this data.

In addition to reporting on coastal and marine status and trends, a series of strategic assessments undertaken in the GBR context have provided crucial information that has driven the CE management process forward (Anthony et al., 2013). The Cumulative Impact and Structured Decision-Making (CISDM) framework was designed to understand the cumulative impacts of multiple stressors and incorporate this knowledge into management decisions. The GBRMPA focus group pointed out that having a long term, sequential series of efforts addressing CE helped build a case for and socialise the implementation of the Reef 2050 Plan.

The cumulative impacts policy provides for a strategic, systematic and consistent approach for managing and reducing cumulative impacts on the GBRMP (Stelzenmüller et al., 2018;

<sup>2</sup>Notable cases that have influenced decision-making in relation to CE in Aotearoa NZ include: Dye v Auckland Regional Council (2002), RJ Davidson Family Trust v Marlborough District Council (2016, 2017) and Okura Holdings Limited v Auckland City Council (2018).

**Figure 1**). The policy outlines a Drivers-Pressures-State-Impact-Response (DPSIR) framework for assessing condition and trend, along with values and attributes of the GBR. It also provides guidance on how to deliver net benefit outcomes for the reef using a range of approaches including working collaboratively with stakeholders at local, regional, national and international scales.

## Values and Rights in Decision Making

Participants acknowledged that under current circumstances it is difficult to determine responsibility for CE, and that this is a significant challenge for CE management (**Tables 3**, **4**). Mechanisms to promote either an individual or collective sense of responsibility are actively being sought in both the Aotearoa NZ and GBR contexts. The GBRMPA focus group described several mechanisms (e.g., workshops, public forums) that have been utilised in an attempt to gain consensus and move management decision making forward. In the Aotearoa NZ context, participants referred to the "tragedy of the commons" (Hardin, 1968) to summarise many of the challenges faced by proposals that relied on collective management of coastal and marine areas.

All participants agreed that the need to develop and socialise a clear and unifying vision and objectives is needed to guide any cohesive CE management process. Those in the Hobart workshop agreed that the current State of the Environment reports (e.g., Ministry for the Environment & Statistics NZ, 2015) do not provide this clarity or vision. However, there are signs that CE policies are gaining interagency and multi-scalar support, although satisfactory implementation is still in progress. For example, the Reef 2050 Plan was recently passed into law, and the Plan's implementation is overseen through the Great Barrier Reef Ministerial Forum (which includes representation from both the Australian and Queensland governments) (Great Barrier Reef Marine Park Authority, and Queensland Government, 2018). At the core of the Reef 2050 Plan is an outcomes framework that may drive progress toward an overarching vision:

"To ensure the Great Barrier Reef continues to improve on its Outstanding Universal Value every decade between now and 2050 to be a natural wonder for each successive generation to come" (Commonwealth of Australia, 2018: 1).

Participants agreed that establishing cohesive CE management protocols requires a substantial investment in participatory processes. The Brisbane focus group and the Hobart workshop participants described the value in having expert and/or independent facilitators involved in running participatory processes (**Tables 3**, **4**). The need to get buy-in and agreement from both conservation and industry groups was a significant hurdle for the Reef 2050 Plan and required extensive stakeholder engagement and conflict management. In the Aotearoa NZ context, participatory processes are perhaps more standardised than in Australia due to RMA consultation requirements that have been in place since 1991, but evidence suggests that there is considerable diversity in how long-term, large-scale engagement with diverse and often conflicting interests plays out (Davies et al., 2018b). Another crucial factor related to participation is the rights of Indigenous partners in CE management decision making. While in Aotearoa NZ, Treaty agreements provide some guidelines and protections regarding how co-management and co-governance practices should unfold, participants in the GBRMPA focus group described the need to recognise TOs as partners and provide support for participation and capacity building as a major gap in terms of including values and rights in CE management and decision making, even under the recently implemented Reef 2050 Plan.

Knowing the distribution of human values and impacts on marine environments, spatially and temporally, is key to successful CE management (Jones et al., 2018). The absence of large-scale social and cultural data sets in both Aotearoa NZ and Australia complicates attempts to incorporate Indigenous values into ecosystem-based management (**Table 4**). Much of the information available is in an unquantified or unquantifiable form and seen as unsuitable for inclusion in traditional western scientific monitoring and management programms. Some traditional use and value data are deemed not sharable for cultural reasons by Indigenous communities. As with the existing ecological data sets, the social and cultural information available is not conducive to facilitating assessment or management of large scale coastal and marine systems facing issues across multiple spatial and temporal scales, but perhaps could be viable as part of a more localised or regional CE management cluster.

## Linking Across Scales – The Role of Metaphors and Models

The approach taken to CE assessment and decision making can differ depending on who is undertaking the assessment, the purpose for which it is intended and whether it is to be done at a strategic level or a project level. In Aotearoa NZ, operators in the marine environment are required to consider the CE of proposed activities. However, the method and scope of the assessment that is undertaken can change dramatically between applications, and decision makers must be able to evaluate and make a determination on these assessments. Decisions about the scope of the assessment (activity vs. receptor) and the spatial and temporal scale all influence CE assessments and their reliability (Natural England, 2014).

Both case studies emphasised emerging, future-focussed, innovative initiatives that bring some optimism to the CE management discussion (**Table 3**). In the Aotearoa NZ context, the Sustainable Seas National Science Challenge is providing funding for work on CE through to 2024. In GBR, the Reef Guardians Schools Programm (Day and Dobbs, 2013) is an action-based sustainability education programm that builds relationships and recognises opportunities. Participants and teachers are said to change the way they do things and how they think about their effect on the reef (Evans, 2011). There was some admission, however, that work in the social sciences is a large gap in many of these efforts (**Table 3**).

Both case studies have developed metaphors, models, and tools to address the many complexities associated with CE management (**Table 3**). The DPSIR framework (Commonwealth of Australia, 2018) has been used extensively to deal with CE in GBRMP. Participants touted DPSIR as a tool that compromises

some of the details preferred by scientists but makes sense to most stakeholders. The Aotearoa NZ Team is also using a modified DPSIR framework to progress discussions across groups because of its flexibility and transparency. The tool is not perfect; issues have been identified in terms of an over-emphasis on pressurestate interactions, with inadequate treatment or integration of management responses and impacts to human well-being (Lewison et al., 2016; Patrício et al., 2016), but it does help to identify where pressures and threats are in a system, and therefore possible options for intervention. In Aotearoa NZ, ki uta ki tai—from mountain to sea—is a Maori concept that emphasises ¯ the interconnectedness of ecosystems inclusive of people (Schiel and Howard-Williams, 2016; Tipa et al., 2016; Kainamu-Murchie et al., 2018). This concept aligns closely with the commitment to EBM required to effectively manage cumulative effects in complex marine systems. The ki uta ki tai strategy provides a way to conceptualise and manage linkages across scales and cultures. This is an important unifying metaphor that is somewhat missing from the GBR context.

## DISCUSSION

This research has thus far provided important insights into the barriers and enablers associated with CE management and why they persist (or not) across a range of scales, legislative framings, and cultures. The discussion that follows will use this empirical data to tease out lessons about how to prompt more effective action on CE management in the future. We do this by considering how well the Aotearoa NZ and GBRMP case studies align with qualities associated with transformative management and governance (**Table 5**). Although our findings come from an Oceania context, by viewing this data through a transformative lens we illuminate the broader implications of this research for CE management in other contexts around the world.

## Preparing for Change

Building knowledge and networking are key aspects of successful implementation of CE governance and management transformations. Not surprisingly, results from this study indicate that actors in the GBRMP case study have done more to prepare for a CE management transformation than the actors in the Aotearoa NZ case study. Study participants pointed out that the long-term monitoring programms in GBRMP have enabled relatively speedy progress on developing assessments, risk management plans, and other related CE guidelines. Mechanisms that promote participation and a sense of responsibility when it comes to CE governance and management have also been actively developed in GBRMP, with a substantial amount of work aimed at sharing information across scales and building trust among actors. There is no equivalent work in the Aotearoa NZ context beyond the co-developed project associated with this study.

The collective efforts associated with CE management in GBRMP have paid off, and the governance and management of the area has progressed through at least the early stages of a transformation (Folke et al., 2010). However, data that are available for CE assessments and management in both cases are still considered by study participants to be fragmented, not standardised and often combined with high levels of uncertainty, making system-based assessments and predictions difficult. Failure to adequately address uncertainty has directly led to many cases of failed management around the world (Ludwig et al., 1993; Ralls and Taylor, 2000).

While government agencies in Aotearoa NZ have recognised the existence of CE and the role that local coastal and marine systems play in the wider global ocean ecosystem, long-term ecological-scale data is needed in order to manage these systems effectively. There is a better record of collecting ecological data and conducting strategic analyses in GBRMP, but both case studies reveal substantial gaps in the social and cultural data needed to effectively link ecological data to behaviour changes and implementation that would be effective for CE management. Further, understanding of interactions between stressors are limited, and better understanding of whether component and system interactions are additive or synergistic (i.e., total effect is great than the sum of the parts) can assist in more efficient CE management (Burkepile and Hay, 2006; Crain et al., 2008; Harvey et al., 2013; Przeslawski et al., 2015).

Another key component of preparing for change is individual and/or organisational leadership that develops strategies for exploring new configurations for CE management (Olsson et al., 2006). GBRMPA has been touted for its willingness to reorganise internal structures and test innovative management strategies to gain more traction on GBRMP management challenges (Folke et al., 2010). In contrast, in Aotearoa NZ, there has been little cohesive leadership taken by any government authority on CE management beyond the recognition in recent State of the Environment reports that it is a substantial concern (Ministry for the Environment & Statistics NZ, 2015, 2016).

## Windows of Opportunity

Policy changes are affiliated with a "window of opportunity" by Kingdon (1995); he argues that transformations are most likely when problems, solutions, and politics all converge at a critical moment. Ecological crises and other periods of rapid change can similarly provide windows of opportunity that trigger the emergence of new networks and promote new forms of governance (Folke et al., 2005; Olsson et al., 2006). In this study, we are interested in understanding how to promote windows of opportunity even when there is no immediate crisis or other substantial external driver.

## Problem Awareness

Problem awareness can become very challenging when the problems are as complex and cross-scale as something like climate change or CE management. People often struggle to connect their mainly land-based activities to outcomes in the oceans (McKinley and Fletcher, 2012). This contributes to the inability of traditional activity-led management to account for the interconnected nature of terrestrial, atmospheric and coastal and marine systems, let alone the social, political, economic and cultural (SPEC) elements of these systems (Wu et al., 2015; Allison et al., 2018). In the GBRMP case study, the unifying

entity of the reef has provided a focal point (albeit associated with multiple diverse meanings, identities and values) to guide the development of problem awareness and subsequent work on CE governance and management transformations.

In Aotearoa NZ, the biggest challenges to transforming CE governance and management could arguably arise in relation to developing a cohesive problem awareness, and thus linking knowledge, values, and actions across scales (Davies et al., 2018a). This process requires coordination of CE management among institutions and agencies, a notoriously difficult task (Lundquist et al., 2016). The RMA should supply some level of overarching coordination to legislative frameworks and the agencies responsible for CE management in Aotearoa NZ, but many instances of fragmentation remain, not the least of which are the distinctions between the territorial sea and the EEZ, and the exclusion of fisheries management from consideration (Severinsen and Peart, 2018). However, emerging efforts to conduct mission-led inter- and transdisciplinary science (such as the Sustainable Seas National Science Challenge) indicate that there is an increasing awareness in some sectors that CE governance and management transformations are needed. Similar integrated science and society approaches have also emerged elsewhere around the world to address complex, crossscale challenges (e.g., Collins et al., 2007). It has also been suggested that taking a local/regional level approach and then linking to principles-based approaches that apply at higher scales could be a more efficient way to address CE governance and management challenges (Crease et al., 2019).

### Solutions Available

One of the key areas of development in Aotearoa NZ is the incorporation of Maori concepts that focus on place-based ¯ interconnections of ecosystems. Matauranga M ¯ aori offers a ¯ place-based understanding of environmental change derived from intergenerational observations and the transmission of that knowledge; this kind of information is necessary for managing CE. Meanwhile, the principles and values associated with the ki uta ki tai (mountain to sea) concept can provide a unifying metaphor that aligns with healthy ecosystems and can leverage both scientific and customary knowledge to support implementation (Jackson et al., 2018).

Crown obligations to Maori under the Treaty of Waitangi ¯ and as incorporated into contemporary legislation (such as the RMA) also provide Maori with a strong negotiating ¯ position (relative to many other Indigenous nations) (Bryan, 2017; Jackson, 2018). This means Maori values and rights ¯ are often positioned at the forefront of natural resource management negotiations and decision making, rather than being assumed as part of a process in which a single culture dominates. The prominence of Maori knowledge, ¯ culture and tikanga (ethical or appropriate ways of doing) has been further strengthened in Aotearoa NZ decision making through the ongoing settlements of Treaty claims and establishment of co-governance and co-management agreements between relevant councils and Iwi around the country (Mutu, 2012; Ruru, 2018; Te Aho, 2018). Treaty settlements have also resulted in substantial resource (natural and financial) (re)allocation to Iwi. The increasingly powerful position of Maori/Iwi can provide unique opportunities to explore ¯ alternative approaches to CE management. For example, a ki uta ki tai (mountains to sea) approach to CE management that draws from traditional and contemporary Maori approaches to natural ¯ resource management (Kainamu-Murchie et al., 2018) provides a proactive, holistic framing from which CE management can be undertaken.

Artificial divisions such as those between land and sea, regions/states, the EEZ and Territorial Sea, and activities/management authorities operating in the same physical space create jurisdictional boundaries that can be challenging, but not impossible to work around when it comes to CE management. The Reef 2050 Plan has achieved some improved coordination in this regard by linking GBRMPA, Queensland, and the Commonwealth (Commonwealth of Australia, 2018). While the Reef 2050 Plan recognises existing spatial boundaries and jurisdictions, it also includes an overarching set of principles and strategic approaches to assessing environmental health (e.g., the zone of influence approach) that support collective buy-in and responsibility both within and beyond these boundaries (Commonwealth of Australia, 2018). Through a long process of education, pressure, negotiation and compromise, GBRMPA has now been granted a substantial amount of authority to oversee the implementation of the Reef 2050 Plan, but how this works in practice (including questions related to funding, monitoring, and enforcement) will require further study, as efforts have only just begun.

Another approach to linking across scales in complex coupled human-environmental systems relies on models to simulate the interactions between the multitude of interacting subsystems (e.g., geophysical, ecological, climatological, social, political, economic, cultural). This type of model is rare, and generally heuristic in nature. Such heuristic whole-of-systems models are much better suited (and often developed) to improve understanding of the interactions within complex human-environmental systems rather than for use in management applications (Kelly et al., 2013; Allison et al., 2018). The use of DPSIR models in the context of the GBRMP is an example of a heuristic model that aims to enable managers to make better, more informed decisions about if, where and how activities under their control should take place, however these approaches are much more successful when applied at management-relevant local scales.

In recognising the complexities involved in cross-scale interactions, the door has been opened for the development of tools to assess how and what types of uncertainty are likely to affect environmental management outcomes. Examples include qualitative models (Anthony et al., 2013; Raoux et al., 2018), Bayesian belief networks (Mantyka-Pringle et al., 2017) and models (Schmelter et al., 2012), management strategy evaluations (Nuno et al., 2014), sensitivity analyses (Perz et al., 2013; Stock and Micheli, 2016), and Monte Carlo simulations (Stelzenmüller et al., 2018). These tools can be used to rank management decisions based on the likelihood that the intended results will be achieved given existing uncertainty.

Transparently assessing uncertainty is necessary for successfully implementing CE assessments, ecosystem-based management, adaptive management principles (Holling, 1973), and the precautionary approach. Tools, techniques and models are also needed to improve practices for assessing and managing CE (Clarke et al., 2016; Davies et al., 2018a). While there is no universally accepted framework, there are a number of tools to assess cumulative impacts, for example, carrying capacity analysis, impact or interaction matrices, modelling and expert opinion (Walker and Johnston, 1999; Natural England, 2014). However, each of these methods need to be developed and tested with a view to implementing them in line with CE management and decision-making processes (Halpern et al., 2008), which generally requires a substantial investment in collaborative and participatory processes to ground model development in the realities of practice (Voinov et al., 2018). If combined with efforts to take adequate time to collate and analyse existing data sets, this area of work holds great promise. Both case studies are testing the boundaries of how these relationships are established and maintained.

## Political Action

Although positive gains in terms of CE management may be attained through more localised and/or clustered efforts, an overarching policy to guide management activities and plans has long been considered the most straight forward way to rescue both sides of the Tasman Sea from ongoing challenges associated with fragmented legislative regimes; incomplete/inadequate ecological and social, political, economic and cultural data; and competition among stakeholders and Indigenous groups for scarce resources and power over decision-making processes. Despite historical attempts, both Australia and Aotearoa NZ have failed to pass an overarching oceans policy. Australia released an Oceans Policy in 1998 but was unsuccessful in its attempts to institute integrated policymaking (Vince, 2008) and EBM (Tsamenyi and Kenchington, 2012). Aotearoa NZ's oceans policy process came to a halt in 2003 after a significant effort of public consultation, including 71 hui and other meetings. The process was briefly restarted in 2005 but the political will that was driving the earlier effort had been lost with a change of government and priorities (McGinnis, 2012). The passage of the Reef 2050 Plan is a big step forward for GBRMP but is still only applicable to the particulars of the case study area.

A crucial subset of any ocean policy must address the rights of Indigenous peoples. In Aotearoa NZ, Treaty agreements and subsequent legislation provide some guidelines and protections regarding how co-management and co-governance practices should unfold, but in the GBRMP context, there is less of a unifying approach to involving the Traditional Owners of (sea) country than in Aotearoa NZ. The TUMRAs do set a good precedent in some areas (Nursey-Bray and Rist, 2009), but there are deep inequalities associated with TO inclusion in natural resource decision making in Australia in general (Davies et al., 2013; Day and Dobbs, 2013), and this also applies to many parts of GBRMP. The need to recognise TOs as partners, provide support for participation and capacity building and to include values and rights in CE management and decision-making remains as a gap even under the recently implemented Reef 2050 Plan.

## Navigating the Transition

The transition phase from the current system of CE management and governance to a more adaptive, resilient, and holistic one is likely to be unpredictable and turbulent, and therefore "can only be navigated, not planned" (Olsson et al., 2006, p11 ¶1). Successful transformations therefore require support from emergent shadow networks; these informal or semi-formal networks facilitate information flows, support social learning, and provide opportunities to experiment with alternative ways of doing governance and management (Schmidt, 2017). They can help to institutionalise the new normal during windows of opportunity, relying on a range of leverage points, including but not limited to economic incentives (Olsson et al., 2006).

Both case studies discussed the emergence of future-focussed, innovative and independent initiatives that brought some optimism to the CE management assembly (**Table 3**), but the initiatives from the GBRMP case (e.g., Reef Guardian Schools Programme – running since 2003) were relatively wellestablished and well-resourced in comparison with the Aotearoa NZ initiatives discussed (e.g., Navigating the Implementation Impasse research project – running since 2017). Either way, these efforts inject much-needed novelty into transitions (Chaffin et al., 2016), but the authority and reach of older and better resourced networks will clearly have a bigger impact when it comes to supporting CE management transitions. In Aotearoa NZ, the lack of shadow networks addressing CE management and governance has meant that changes tend to be fairly slow - approaches to CE assessment and management have thus far relied primarily on case law related to resource decisions made through the courts (Milne, 2008) and regional efforts by councils or community groups. Most efforts have yet to connect across scales.

Another strategy exhibited in both case studies that can be associated with the emergence of shadow networks is a commitment to participatory processes. Both case studies have leaned on participatory processes in a range of forms and forums to overcome some activity or rights-based conflicts and develop more collective identities and partnerships across vested interests. While there are some promising results arising from these efforts (e.g., the passing of the Reef 2050 Plan in the GBRMP case), the institutional structures, funding mechanisms, and social and cultural changes that are needed to support these developments long-term are still emerging.

Emerging leadership is also important in terms of supporting and maintaining CE governance and management transitions, especially in the institutionalisation of the "new normal." Leadership at GBRMPA has been highlighted for connecting with different sectors, reducing conflict, and spanning scales (Olsson et al., 2008). Strategies initiated by GBRMPA enabled the coordination of the scientific community, increased public awareness of environmental issues and problems, involved a broader set of stakeholders, and maneuvered the political system for support at critical times (Olsson et al., 2008). These efforts have enabled the organisation to build interagency and multiscalar support and have set a good precedent for the work that lies ahead in terms of implementing the Reef 2050 Plan.

## CONCLUSION

fmars-07-00025 February 17, 2020 Time: 14:36 # 15

By comparing Trans-Tasman cumulative effects management challenges and successes and considering these efforts against a transformative governance framework (Olsson et al., 2006), this paper has illuminated several features that are needed to enable effective CE management. The three key challenges associated with the implementation of improved CE management that are considered in this paper are: (1) fragmented legislative regimes; (2) a lack of standardised, long-term ecological-scale data; and (3) poor integration of socio-economic and cultural values, and Indigenous rights into management decision making. This paper addresses these challenges by drawing conclusions and identifying priority actions regarding how to mobilise resources and political will to address CE, how to deal with data scarcity and uncertainty, and how to promote comprehensive and inclusive CE management of coastal and marine areas. We believe these findings provide an important resource for future efforts to implement effective cross-scale CE management in Aotearoa NZ and GBR/Australia, while also being applicable to other international work on CE management.

## Mobilising Resources and Political Will

To implement effective CE management, there is an urgent need to mobilise diverse actors across scales, interests, institutions and cultures and promote collective actions. This research has revealed some key components that can help or hinder these efforts. First, having a uniting feature or notable value that is impacted by CE can mobilise leadership and generate action. In the case of the GBRMP, the reef itself is a distinct, charismatic feature which is internationally valued, and impacts of bleaching and cyclone damage are typically broadcast widely, encouraging motivated and innovative leaders to emerge from a range of networks to address an identifiable unifying cause.

In the Aotearoa NZ context, no clear value is threatened, and CE result in gradual change and shifting baselines. In cases such as this, it can be difficult to develop broad and effective action to address CE. The development of a unifying vision for management that appeals to disparate interest groups and actors across scales is crucial to the achievement of collective action under these circumstances. The identification of locally or regionally valued ecosystems and related social and cultural practices through participatory processes may help to promote this proactive agenda setting. In either case, institutional and individual leadership is needed to ensure that background knowledge and robust networks are in place so that when a window of opportunity emerges (or is created), leaders can act quickly to streamline fragmented legislative regimes in ways that will reduce stressors and ensure no further harm occurs.

Other key elements of successful CE management include investing in Indigenous partnerships and co-governance arrangements, as well as being generally inclusive of a range of partners – not just the most powerful – in order to ensure broader support through collaboratively defined long term plans for CE management and ecosystem sustainability. Ongoing resourcing and support of participation of diverse membership within interagency and transdisciplinary working groups (e.g., the range of interests represented on the GBRMPA Reef Advisory Committees) can help to ensure that these processes proceed somewhat independent of political whims. Resourcing these efforts may require flexible and creative funding; looking for opportunities to cluster efforts by region or seeking support from both higher and lower scales may help to assemble these resources.

## Dealing With Data Scarcity and Uncertainty

Addressing data scarcity and high levels of uncertainty in the context of CE governance and management is a notoriously difficult task, but in some cases data or information may exist that can be repurposed for CE management decision making. By partnering with indigenous and local knowledge holders, whose knowledge of an area may go back decades or even generations, it may be possible to gain some of the long-term data needed to improve CE management and decision making. Mining existing data sets for CE data may also prove to be a valuable option. Although most of these data sets are unlikely to be appropriate for use across large scales, they may be useful for local and/or regional CE management purposes.

In addition to promoting better use of existing data, it is important to collect new data to help reduce uncertainties where possible. Today the GBRMP has over 90 monitoring programms operating at a variety of spatial and temporal scales (Reef 2050 IMRP, 2018). Comprehensive, strategic social-ecological system assessments can provide a scientific baseline from which policy can evaluate ongoing changes to determine whether action is required to halt degradation; these kinds of assessments also provide clarity around the suite of social, cultural, and economic system components that should be included in CE assessments. Additionally, the DPSIR model is now being used by the GBRMPA as a unifying framework to describe integrated monitoring and management of the GBRMP. This EBM approach is designed to allow managers to understand connections between components of and processes acting upon the GBRMP to facilitate integrated monitoring and management. The choice of model is arguably not as important as its ability to provide a cohesive metaphor that transparently links human-nature interactions and types of knowledge and information for decision making purposes.

### Promoting Comprehensive and Inclusive CE Management of Coastal and Marine Areas

Effective CE management must be based on holistic systemsbased thinking that incorporates cross-scale interactions [e.g., ki uta ki tai (mountain to sea)], incorporates the multitude of overlapping, synergistic and antagonistic human and natural impacts, and accounts for the values and rights of current and future generations. A number of tactics support systems-based

thinking/acting, though these approaches are often contrary to existing conventional sector-based management and statutory regulations. For example, principles-based approaches such as those included in the Reef 2050 Plan and under development in the Aotearoa NZ context (Davies et al., 2019) provide a set of guidelines for CE management without being too prescriptive. Inter- and transdisciplinary research efforts also tend to enable holistic, interconnected approaches to problem resolution, rather than focusing on prescriptive limit setting for individual stressors (as these approaches tend to ignore synergistic or antagonistic interactions). While these approaches come with their own challenges, they are far better suited to CE management than anything that has yet been attempted.

## Priority Actions

This research indicates that collaborative approaches can generally improve the implementation and practice of CE management, but further prioritisation is needed to guide future efforts. The following priority actions are envisioned as being deployed in advance of a crisis; once a crisis is identified as occurring, circumstances may dictate that another sequence of events is needed. To improve CE management practices, we recommend that future work on CE:


## REFERENCES


## DATA AVAILABILITY STATEMENT

The datasets for this study will not be made publicly available because some data was obtained in confidence and therefore cannot be shared publicly. Identifiers have been removed from the data and key points and themes are included in the tables provided for publication.

## ETHICS STATEMENT

This study was carried out in accordance with the recommendations and protocol approval of the NIWA Human Research Approval Process with oral and/or written informed consent from all subjects.

## AUTHOR CONTRIBUTIONS

KD, KF, GC, and IP contributed to the conception, design and implementation of the research. KD, KF, and AA performed the initial data analysis, developed the tables and figures, and drafted the manuscript. GC, IP, JD, MF, and CL contributed to the interpretation of results, wrote sections of the manuscript, and provided critical revisions to the intellectual content of the manuscript. All authors contributed to manuscript revision, read and approved the submitted version.

## FUNDING

Funding for this research was provided by Ministry of Business, Innovation and Employment research contracts C01X1515 (Sustainable Seas) through NIWA Project Number SUSS18201 (Navigating the implementation impasse: enabling interagency collaboration on cumulative effects) and the Strategic Science Investment Fund NIWA research contract COME1903. Additional funding to support this work was provided by the New Zealand Coastal Society's Professional Development Award.

## ACKNOWLEDGMENTS

We would like to thank the focus group and workshop participants for their contributions to this research.

research. Environ. Model. Softw. 99, 147–155. doi: 10.1016/j.envsoft.2017.0 9.015

Anthony, K. R., Dambacher, J. M., Walshe, T. R., and And Beeden, R. (2013). A Framework for Understanding Cumulative Impacts, Supporting Environmental Decisions and Informing Resilience Based Management of the Great Barrier Reef World Heritage Area: Final Report to the Great Barrier Reef Marine Park Authority and Department of the Environment. Townsville: Australian Institute of Marine Science.



integrated environmental assessment and management. Environ. Model. Softw. 47, 159–181. doi: 10.1016/j.envsoft.2013.05.005



**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 © 2020 Davies, Fisher, Couzens, Allison, van Putten, Dambacher, Foley and Lundquist. 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.

# Exploring Balanced Harvesting by Using an Atlantis Ecosystem Model for the Nordic and Barents Seas

Ina Nilsen1,2 \*, Jeppe Kolding<sup>1</sup> , Cecilie Hansen<sup>2</sup> and Daniel Howell<sup>2</sup>

<sup>1</sup> Department of Biology, High Technology Centre, University of Bergen, Bergen, Norway, <sup>2</sup> Institute of Marine Research, Bergen, Norway

"Balanced Harvesting" (BH) has been suggested as a possible strategy to meet the objectives of the Ecosystem Approach to Fisheries, ensuring a high sustainable yield while maintaining ecosystem structure and function. BH proposes a moderate fishing mortality in proportion to productivity spread across the widest possible range of species, stocks, and sizes in an ecosystem producing a sustainable and overall non-selective harvest. The Norwegian and Barents Seas have been subjected to moderate fishing pressure on commercial species, and elements of an ecosystembased approach to management for many years, but not the fishing pattern proposed by BH. By using an Atlantis ecosystem model of the Nordic and Barents Seas, we investigated the effects of applying a BH regime to a region with existing successful fisheries management. This was done by running simulations with combinations of historic fishing pressure and fishing mortality rates proportional to 25% of the productivity of most species and sizes. The simulations were then compared to a control run where the historical fisheries were applied. The model results showed that implementing a BH regime in the Norwegian and Barents Seas would only produce marginal increases in total yields of currently commercially exploited stocks, likely because the Norwegian fisheries are already mostly well-managed. However, expanding the fishery to include species that are not commercially exploited today did produce higher yields, especially on lower trophic levels. This study represents the first attempted examination of implementing BH based on productivity using an Atlantis ecosystem model, as well as the first investigation of BH in the Norwegian and Barents Seas. We use this model as a case study to identify the gains that species-based BH can be expected to give over well-implemented traditional fisheries management rather than simply comparing to an over-exploited system.

Keywords: balanced harvesting, Atlantis, end-to-end modeling, management strategy evaluation, Ecosystem Approach to Fisheries

## INTRODUCTION

Fisheries today are generally considered to be in a scarce condition with little room for further expansion with some even proclaiming that there will be nothing left to fish within the next 50 years if current trends continue (Black, 2006). According to the FAO statistics global marine capture fisheries have been flat for over 30 years with an increasing number of the unassessed stocks

Edited by:

Jacob Carstensen, Aarhus University, Denmark

#### Reviewed by:

Marie Maar, Aarhus University, Denmark Jason Link, National Oceanic and Atmospheric Administration (NOAA), United States

> \*Correspondence: Ina Nilsen ina.nilsen@hi.no

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 31 May 2019 Accepted: 30 January 2020 Published: 05 March 2020

#### Citation:

Nilsen I, Kolding J, Hansen C and Howell D (2020) Exploring Balanced Harvesting by Using an Atlantis Ecosystem Model for the Nordic and Barents Seas. Front. Mar. Sci. 7:70. doi: 10.3389/fmars.2020.00070

regarded as overfished (FAO, 2018). One of the main challenges of modern fisheries management is to develop harvest strategies that ensure efficient and maximum sustainable utilization of marine production (UNCLOS, 1982), while also preserving the structure and functioning of harvested stocks and ecosystems (CBD, 1992). The concept of Ecosystem Approach to Fisheries (EAF), based on the 1998 Malawi Principles (UNEP/CBD, 1998), has been proposed as a holistic framework to deal with these objectives.

Norway is committed by law to implement an Ecosystem-Based Fisheries Management (EBFM) in the North Sea, Norwegian Sea, and Barents Sea (Miljøverndepartementet, 2006, 2009, 2011; Olsen et al., 2007). According to Pitcher et al. (2009), implementing ecosystem-based management in Norway, in line with the code of conduct of responsible fisheries (FAO, 1995), should be relatively straightforward. Although largely regulated by conventional single-species management, Norway already scores high on ecosystem-based principles. An example of ecosystem considerations is the management of Northeast Arctic cod and Barents Sea capelin where the importance of capelin as food for cod has been considered in the capelin fishery since 1991 (ICES, 2015a). Although the broad principles of EBFM are agreed, there are uncertainties in the specific implementation, for example how to find the balance between "exploitation" and "conservation" (Howell et al., 2016).

To operationalize the objectives of the EAF, "Balanced Harvesting" (hereafter BH) has been suggested as one possible strategy to ensure a high sustainable yields while maintaining ecosystem structure and function. Garcia et al. (2012) defined BH as "a moderate fishing pressure spread across the widest possible range of species, stocks, and sizes of an ecosystem, in proportion to their natural productivity so that the relative size and species composition is maintained." Clearly, BH is not a necessary part of an EAFs. Nor would it be a complete solution since any overall harvesting strategy would need to work hand in hand with, for example, strategies to protect vulnerable habitats. However, BH has been proposed as one possible component of full Ecosystem-Based Fishery Management, and we therefore seek to analyze the potential consequences of implementation of BH in a specific ecosystem.

We should note here that BH has been the subject of considerable debate in the scientific literature, with papers both supporting and opposing the idea. The concept of BH has received a number of criticisms on ethical and theoretical grounds (e.g., Burgess et al., 2016; Froese et al., 2016; Pauly et al., 2016). One issue that arises is that of practicality: to what extent is it practical to harvest across a wide range of the ecosystem (e.g., Howell et al., 2016). This objection encompasses the physical difficulty of harvesting some ecosystem components, the difficulty in providing scientific advice to support such harvesting, and economics of such harvesting which could render some of the fishery uneconomic.

The question we ask is "Given the model with the assumed best realism that we have, for an ecosystem which is already relatively well-managed, what would happen if BH were to be implemented in the Norwegian and Barents Sea ecosystem?" (**Figure 1**). Specifically, are there are gains to be made, and if so where do they come from, and what are the corresponding losses or structural changes to the system? We explicitly do not address issues of the practicality of such a fishery, nor do we attempt to model the economics or socio-economics of such a change. These are valuable questions but beyond the scope of the current work.

Balanced harvesting can be considered as one possible method to take fisheries management to the ecosystem level through exposing as many components of the ecosystem as possible to a fishing mortality proportional to their specific productivity. The idea has attracted broad interest worldwide and has been supported by both empirical studies in African lake ecosystems with small-scale fisheries (Kolding and van Zwieten, 2014; Kolding et al., 2015) and by modeling studies of marine systems (Garcia et al., 2012; Law et al., 2013). These studies suggest that a balanced harvest may increase the total sustainable yield while maintaining ecosystem structure compared to today's selective harvesting.

The concept of BH emerged from a widespread concern of the problems caused by conventional selective fishing management resulting in a stagnation in global catches (FAO, 2016), overfishing of target species (Costello et al., 2012; Sumaila et al., 2012), depletion of large predatory fish (Christensen et al., 2014) and potential fisheries-induced evolution that favors early maturation resulting in smaller fish (Heino and Godø, 2002; Law, 2007; Hsieh et al., 2010). Selectivity is deeply engrained in our fishery historically, and fishermen usually target the largest individuals and species for economic and ethical reasons (Kolding and van Zwieten, 2011). However, any kind of selective removal will inevitably alter the composition of a population and consequently the structure and biodiversity of the ecosystem – even at moderate fishing levels (Garcia et al., 2012).

It should be emphasized that BH does not call for unselective and indiscriminate fishing. In fact, it has been argued that BH fishing may actually require a higher level of selectivity (Reid et al., 2016). BH simply suggests a different type of selectivity at ecosystem level where the overall fishing pressure is spread over different species and body sizes in line with productivity in order to maintain the ecosystem structure (Garcia et al., 2015). If BH results in mimicking the natural mortality with predationlike fishing mortality, the evolutionary selection on life-history traits would be expected to be reduced. An implementation of BH would result in a more diverse fishing fleet with a wider range of fishing gears, and the risk of fishery induced selection on any trait is reduced (Zhou et al., 2019).

As productivity tends to decrease as a function of body size (Peters, 1986), moving toward a full implementation of BH would imply a reduced harvest of large fish and increased fishing on smaller species and individuals that are generally considered low-value and unusable in industrial countries. Although BH has been shown to be effective in giving high biomass yields with low impacts to the ecosystem size spectra in African small-scale subsistence fisheries, it is not clear that these results translate to large-scale modern commercial oceanic fisheries (Burgess et al., 2016; Howell et al., 2016). As a result it has been suggested that any implementation of BH would be a partial implementation (e.g., Howell et al., 2016), and we attempt to address this by running simulations to distinguish the effects

of BH in different parts of the system. Ethical issues also arise over the question of which fractions of the ecosystem should be considered as harvestable resources (e.g., Pauly et al., 2016). Finally, criticisms have arisen over the modeling techniques employed. Often, although not exclusively, the modeling studies have used simplified size-based model structures which do not well-resolve the species-specific dynamics of the ecosystem components. As noted above we do not intend to enter this theoretical debate here. We merely aim to investigate what might occur if BH were to be implemented in an Atlantis model of the Barents Sea, and hope that the results of our work give some more concrete input into the overall debate. This paper does not focus further on this discussion, which is covered in a recent review of Zhou et al. (2019), except to note that by using the Atlantis model we aim to include as much species realism in our analysis as is currently possible. Where species-specific details are poorly captured, we note this in the discussion.

The Norwegian fishery is currently considered as fairly wellmanaged with most commercial fish stocks harvested using harvest control rules (HCRs) with moderate fishing pressure. In these rules the fishing pressure is close to that which produces the maximum long-term yield without imposing an undue risk of over-fishing the stock, i.e., close to the maximum sustainable yield (MSY) as used in ICES fisheries management (ICES, 2018a). The Norwegian fishery scores high on the BH principle of targeting a range of species at different trophic levels, as it includes exploitation on lowlevel species like the copepod Calanus finmarchicus (Calanus, 2018) and higher-level species like bird eggs, seals and whales (Howell et al., 2016).

However, several relatively abundant stocks are either only lightly harvested (i.e., polar cod, Boreogadus saida) or completely unexploited (i.e., mesopelagic fish) (ICES, 2016). For all fisheries, minimum individual size restrictions apply, usually somewhat below the average size at maturation. Thus, the fishing intensity is not balanced between all the key species and harvesting within species is not balanced; rather a strong "traditional" size selectivity applies (Gullestad et al., 2014).

To investigate the implications of a BH fishery we will use an end-to-end Atlantis ecosystem model (Fulton et al., 2011) parameterized and tuned for the Nordic and Barents Seas (hereinafter the NoBa model) by Hansen et al. (2016, 2019). By running simulation scenarios of 50 years over the period 1980 to 2030 we study the interaction effects of components harvested with a fishing mortality rate relative to productivity. This is done by first exposing selected species (both commercial and non-commercial) one-at-a-time to a fishing mortality proportional to productivity, to investigate the ecosystem effects of harvesting individual species according to BH and identify those species which have a particular effect on the combined community. Subsequently, we progress gradually to a full implementation with combined runs with multiple species subjected to BH were set up in order to assess the cumulative effects of a BH regime. In addition, a gear selectivity option was applied to all age-structured groups to balance over age groups within species.

Balanced harvesting has been partly studied in multi-species models before (Bundy et al., 2005; Garcia et al., 2012; Kolding et al., 2016; Heath et al., 2017), but this study represents the first attempt of implementing a BH regime with fishery mortalities based on productivity in an Atlantis model. It is also the first

model application to study the ecosystem effects of BH in the Norwegian and Barents Seas.

## MATERIALS AND METHODS

Atlantis is currently considered one of the most advanced "what if "-scenario models of marine ecosystems (Plagányi, 2007). The model simulates spatial variation in both biogeochemical and socio-economic processes. The NoBa model domain is divided into 60 polygons covering the Nordic and Barents Sea of a total area of 4 million km<sup>2</sup> with up to seven depth layers depending on total depth (**Figure 1**) (Hansen et al., 2016). The Barents Sea is a relatively shallow shelf sea, with an average depth of 230 m located north of Norway and Russia, while the Norwegian Sea has a much deeper average depth of 2000 m and is located between Norway, Iceland and Svalbard (Sakshaug et al., 2009). The pelagic part of the Norwegian Sea has a relatively low biodiversity dominated by large stocks of migratory fish such as Norwegian springspawning herring (Clupea harengus), mackerel (Scomber scombrus), and blue whiting (Micromesistius poutassou). The Barents Sea, on the other hand, is relative diverse given its high-altitude location. It holds the largest cod stock in the world (Gadus morhua), in addition to other commercially important species such as haddock (Melanogrammus aeglefinnus), saithe (Pollachius virens), Greenland halibut (Reinhardtius hippoglossoides), capelin (Mallotus villosus), redfish (Sebastes spp. and Sebastes norvegicus), and prawns (Pandalus borealis).

Currently, the NoBa model contains 53 species and functional groups (**Table 1**) that are connected through a diet matrix. Most vertebrate species are age-structured while invertebrates are gathered into biomass pools. Atlantis does not calculate water fluxes between the polygons but uses outputs from oceanographic models. NoBa is forced bottom–up with time series on temperature, salinity, and currents from a Regional ocean modeling system (ROMS: Shchepetkin and McWilliams, 2005) covering the Northeast Atlantic (Skogen et al., 2007). The harvest sub-model deals with the human exploitation of the marine ecosystems, with a focus on the dynamics of fishing fleets. It allows for multiple fleets with its own set of characteristics like specific gear selectivity, target species and management structure (Fulton et al., 2011). The NoBa model includes 27 fisheries, so-called metiérs (Reid et al., 2016), with distinct characteristics and commercially targeted species (Hansen et al., 2019).

Implementation of BH requires information on the production or productivity of all species, but the literature does not provide a single clear answer of how these should be used to set fishing mortality. There is an ongoing debate (Heath et al., 2017; Zhou et al., 2019) on whether fishing mortality should be set in proportion to **productivity** (P/B with the unit 'per time,' as in this study) or in proportion to **production** (P with the unit 'mass per time'). The key difference between the two is that production (BH1) is density-dependent, while productivity

(BH2) can be density-independent (Eqs. 1 and 2):

$$\text{Production (BH1): } F(\mathbf{x}) = \mathbf{c} \cdot P(\mathbf{x}) = \mathbf{c} \cdot \mathbf{g}(\mathbf{x}) \cdot B(\mathbf{x}) \quad \text{(1)}$$

$$\text{Productivity (BH2): } F(\mathbf{x}) = \boldsymbol{c} \cdot \frac{P(\mathbf{x})}{B(\mathbf{x})} = \boldsymbol{c} \cdot \mathbf{g}(\mathbf{x}) \tag{2}$$

For both equations, the fishing mortality F, on species x, is determined by the magnitude of the exploitation constant, c, and the species-specific production, P(x), calculated from the biomass, B, and growth, g. Since fishing in proportion to BH1 is density-dependent, it tends to be low when the biomass is low and thereby protects species from collapse. Fishing according to BH2 on the other hand, is less sensitive to current biomass, and thereby allow species to be exploited to extinction, as the results of the current examination shows.

Heath et al. (2017) argued that since BH is an ecosystem approach to fishing with an explicit aim of maintaining the species richness of marine ecosystems, the density-dependent fishing mortality in BH1 should be applied and recommended. We followed the method described in Garcia et al. (2012) where gross production is described as individual growth plus recruitment, i.e., the amount of living material produced each year. The gross production was divided by the corresponding biomass to get a "per capita" productivity rate, often referred to as a P/B-ratio. This approach, setting fishing mortality proportional to the productivity or P/B ratio, is one of the alternatives suggested amongst BH-scientists (Jacobsen et al., 2014; Kolding et al., 2016; Zhou et al., 2019) and was the approach chosen for this study. It should be noted that there are other possible formulations (for a discussion of this, see Zhou et al., 2019).

Unlike the ECOPATH models, which uses the P/B ratio (or total mortality) as an input parameter, this is not included in the Atlantis model. Calculation of the P/B ratio was therefore done by using growth and production output generated by an initial run set up prior to this study with the intent of representing the historical fisheries. Atlantis has several ways of applying fishing mortality. For this purpose, the best option was a fishery-induced mortality rate where a proportion of biomass is set to be harvested each day. To capture yearly variations in productivity, the P/B ratio was calculated for each year. Information on growth, weight and numbers was extracted from model outputs to estimate the productivity and biomass of age-structured vertebrate groups. For invertebrates, the production was retrieved directly from the model outputs. Then the P/B-ratio of all selected components was calculated for each year and converted to a proportional fishing mortality by the following equation:

$$F\_{BH} = c \cdot \frac{P}{B} \tag{3}$$

The BH based fishing mortality, FBH, is then the productivity (P) given in tons pr. year over biomass (B) given in tons, multiplied with a dimensionless constant, c, determining the intensity of exploitation. Based on the Cadima estimator (Troadec, 1977) several values of exploitation intensity has been suggested, mainly ranging from 0.2 to 0.4 (Shepherd, 1982; Beddington and Cooke, 1983; Pauly, 1984; Garcia et al., 1989; Sparre and Venema, 1998). However, it was decided to follow

#### TABLE 1 | List of species and functional groups in the NoBa model, including which species the group is parameterized as.


Distribution in the Norwegian Sea (NS), the Barents Sea (BS) or both (NS + BS) is also included.

Kolding (1993) and use a relatively conservative constant of 0.25, corresponding to harvesting 25% of the stock's total annual production. However, during the simulation runs it became evident that, while sustainable for all commercial species, a harvest rate of 25% of production was too high for most of the non-commercial species. The FBH for these species were then halved to 12.5% to avoid immediate collapse (**Table 2**).

A size-specific selectivity was also applied to all age-structured groups (**Table 2**) based on the mean productivity of the age group throughout the simulated years (i.e., 1980–2030). A logistic length-based selectivity curve was chosen as the selectivity option, as it allows for different fishing pressures on age groups according to the productivity level within age-structured species. The selectivity curve usually follows the shape of a sigmoid curve ranging from 0 to 1, where the possibility of retention at lengths span from 0 to 100% (Sparre and Venema, 1998). The curve is given as

$$psel\_i = \frac{1}{1 + \exp(-slib \cdot (L-lsm))}\tag{4}$$

The selectivity curve (psel) of species i, is determined by the inflection point (lsm), i.e., the length at 50% selectivity where 50% escape and 50% are retained, selb which determines the steepness of the curve, and the lengths (L) in cm of the different age classes. Since the productivity typically decreases as a function of body size (Peters, 1986), the selection curve was expected to be descending with a negative selb to exert a greater pressure on young productive age classes. Atlantis uses the length-weight

TABLE 2 | List of all the species subjected to BH in the study, as well as the Fbh-level and whether the species were regarded as commercial in the model.


The selectivity option of the fishing gear was applied to all age-structured groups.

relationship to convert to length, as its cohorts are weight-based. To find the appropriate values for lsm and selb, the mean length of all species at different ages had to be calculated by solving the length-weight relationship equation (Hile, 1936; Martin, 1947) with respect to length:

$$W = a \cdot L^b \to L = \sqrt[b]{\frac{W}{a}}\tag{5}$$

The lengths (L) in cm were determined by two species-dependent parameters, (a) and (b), collected from literature and applied in the model (Hansen et al., 2016), and the weight by age (W) in kg was retrieved from the model outputs. A non-linear least square regression was used to find the best values for lsm and selb. By assuming some initial start values for lsm and selb and applying the selectivity curve equation (Eq. 4), the lsm and selb values giving the selectivity curve closest to the productivity levels were selected. This differs from the traditional gear selectivity curves, which aims to protect the young, i.e., smaller sizes, and target larger sizes.

The species and groups chosen to be subjected to BH are listed in **Table 2**. These were either species that were already commercially harvested (species 1–11) or non-commercial species selected on the basis of being relatively abundant, feasible to catch and a good source of food (species 12–21). The noncommercial components were not harvested in the control run and consisted of species that are either lightly harvested (e.g., Calanus, minke whale) or species that are completely unexploited in Norwegian fisheries (e.g., mesopelagic fish, jellyfish). Species like phytoplankton (impracticable to catch), corals (not edible) and polar bear (protected) were excluded in this study.

The runs were set up to first track the individual effects of BH on one species at the time (presented in the **Supplementary Material** and briefly described below), before the gradual full implementation where multiple species were subjected to BH simultaneously. This was done to investigate the isolated effect of BH on individual species, as well as the cumulative effect of harvesting multiple species within a BH regime. In addition, there are clear practical difficulties in extending BH to currently unharvested (or lightly harvested) species, so we have examined these separately from the main commercial species. The three combined runs were:


**Table 3** gives a complete list of the runs that will be presented for analysis in this paper. All runs were performed by modifying the control run through adjustments of fishing effort and by adding selection curves for all vertebrate groups. The commercial species that were not subjected to BH were harvested according to the fishing mortalities in the control run (FHisto in **Table 2**) with a

#### TABLE 3| List of all simulations selected for analysis.


Runs named as"BH on one" represents simulation when only one species was subjected to BH.

flat constant selectivity applied. The results are presented through changes in biomass (tons wet weight), catch (tons wet weight) and a set of indicators based on Fay et al. (2019). Price per kilo catch was taken from Sildesalgslaget (Norges Sildesalgslag, 2019) for the pelagic components and Råfisklaget (Norsk Råfisklag, 2019) for the demersal components. Trophic levels were based on the values applied in Coll et al. (2016). All plotting was carried out through "R studio" (RStudio Team, 2015) under version 3.5.2.

## RESULTS

The results first focus on changes in biomass as the result of implementing BH on both commercial and non-commercial species. The next part concentrates on the effects on the total catch under different BH regimes, while the last part investigates the effects on ecosystem structure and economy through chosen indicators.

## The Effects of BH on Biomass

The commercial fishery in the Norwegian Sea is dominated by large pelagic stocks of mackerel, blue whiting and Norwegian Spring Spawning herring (hereafter herring). In the Barents Sea, the main commercial species are the Northeast Arctic cod (hereafter cod), capelin, haddock, saithe, Greenland halibut, beaked redfish, golden,redfish and prawns. The mean of the calculated fishing mortalities proportional to production (FBH) for all commercial species are shown in **Table 4** together with the mean historic fishing levels in the control run (FHisto) and the fishing level based on MSY (FMSY).

To evaluate both the direct and indirect effects of a BH regime on each species, we used simulations where only one species was subjected to BH, and subsequently when all chosen species were fished at FBH levels. **Figure 2** shows the biomass of commercial species, when (i) only one species was subjected to BH (light green line) and when (ii) all selected species were subjected to BH (dark green line). The black line represents the biomass in the control run where the traditional fishery was applied.

As expected, **Figure 2** reflected the effects on biomass to changes in fishing mortalities (**Table 4**). Species with lower fishing mortality (like cod and golden redfish) showed an increase in biomass, while the biomasses of mackerel, Greenland halibut, and prawns were greatly reduced due to a much higher fishing pressure.

Yet, a decrease in biomass due to a higher fishing pressure is not necessarily critical for the stock as long as it does not result in recruitment overfishing. To evaluate this we plotted the spawning stock biomass (SSB) of the species and the bpa which is the precautionary reference point (ICES, 2017, 2018a) at which the stock runs the risk of recruitment overfishing. **Figure 3** showed that although the biomass of blue whiting, capelin and beaked redfish were greatly reduced under the BH scenarios, the SSBs were above the bpa. In contrast, the SSBs of mackerel and herring were driven to a level below the bpa under a BH regime. Golden redfish represents an example where the traditional fishing regime in the control run resulted in a critically



The commercial species are compared with the historical fishery mortalities (FHisto) applied in the control run, as well as the estimated fishing mortality of maximum sustainable yield (FMSY) (ICES, 2018d,e,f,g,h, 2019a,b,c,d,e,f). The fishing mortalities for some of the non-commercial species were modified to half of the 25% fishing to prevent collapse in the model (see Table 2). \*Capelin stock is managed by escapement rule strategy, not FMSY. \*\*Greenland halibut and saithe has no defined fisheries reference points. \*\*\*FMSY for prawns are estimated directly from the assessment model and changes when the assessment is updated.

low SSB below bpa, while the BH regimes increased the SSB to a safe level above the precautionary reference point.

The study also included harvesting of species that were considered "non-commercial", and not targeted in the control run. The calculated fishing mortalities (FBH) for these species are listed in **Table 4**. **Figure 4** showed that all of the noncommercial species experienced a decrease in biomass when being subjected to a BH regime, which is what one would expect when subjected to fishing. However, the magnitude of the reduction varied greatly among the species. Mesopelagic fish was driven to a near collapse, and both of the demersal fish groups (including ling, tusk and wolffish), as well as skates and rays, were reduced by close to 75%. Mesozooplankton and small pelagic fish experienced less decrease of around 50%, while gelatinous zooplankton, polar cod, benthic filter feeders and minke whale were even less affected.

## Effects of BH on Catch

Balanced harvesting aims to provide higher yields while better preserving ecosystem structure and functioning than conventional selective fishing. In the second part of the analysis we explored the effects on catches under various BH scenarios. The estimated annual total catch was represented as an average from the 20 last years of the simulations (year 2010–2030) to

to a BH regime, while the dark green line shows the biomass when all species in the study were subjected to BH. The black line displays the control run where a traditional fishing regime was applied. The biomass is given as an index.

FIGURE 3 | Spawning stock biomass (SSB) of all commercial species in the study compared to the precautionary reference point (Bpa) which is marked as a red line. The SSB is given as an index for the simulations when only one species is harvested according to BH (light green), all species are subjected to BH (dark green) and a control run where traditional fishing mortalities were applied (black line).

avoid bias from unsustainable short-term spikes in catches during the first years after implementation, as well as any other shortterm dynamics imposed by the change of fishing regime.

**Figure 5** displays the catch of commercial species when subjected to BH individually and when all species together were subjected to BH, compared to the control

run where a traditional fishing regime was applied to the commercial species and the non-commercial species were not harvested. The biomass is given as an index.

run. The results showed higher catches for some species (e.g., mackerel, capelin, beaked redfish) and lower catches of others (e.g., cod, golden redfish). Although the catches were initially unstable during the first years after BH implementation, the catches of most of the commercial species seemed to become more stable under a BH regime

compared to the traditional fishing a few years into the simulations.

The next question was which of the BH combinations (i.e., "BH on commercial," "BH on non-commercial," or "BH on all") that would give the highest total yields. The three combined runs were compared to the control run in which the commercial species were harvested at historic levels and non-commercial species were unharvested. The average total catch of the last 20 years of the simulations (year 2010–2030), when catches had stabilized, was used to evaluate the long term yields.

**Figure 6** shows the total catch composition when subjected to different variations of BH. Total catch increased when more species were included and harvested by a BH regime, but this increase was mainly caused by higher catches of capelin, prawns, and non-commercial species. The catch of the non-commercial species is lumped together in **Figure 6A** and shows nearly 80 mill tons additional yields. However, nearly all this catch (98%) consisted of mesozooplankton, which was excluded in **Figure 6B** for easier comparison of the remaining species.

When exposing only non-commercial species to a BH regime the total catch of the commercial species decreased by 24,000 tons (**Figure 6**, which is a relatively small reduction compared to the added 80 mill tons (of mainly low trophic level species). **Figure 7** shows that the reduction in catches of commercial species was primarily caused by blue whiting which was reduced by 200,000 tons, but this was partly compensated for by an increase in catches of herring and cod.

The added catch of the non-commercial species when excluding mesozooplankton, was approximately 1,6 mill tons. Most of this new catch consisted of mesopelagic fish and benthic filter feeders, as well as some smaller contributions of demersal fish, jellyfish, skates and small pelagic fish (**Figure 8**).

## Effects of BH on the Ecosystem Structure

**Figure 9** illustrates how the whole ecosystem responded through changes in biomass for each guild. The individual species and functional groups belonging to each guild can be found in **Table 1** and were represented as triangles in the figure. The figure showed that guilds with species subjected to BH had the strongest responses. However, the group of primary producers seemed most affected when non-commercial species were subjected to BH.

To better understand the full effects of implementing a BH regime we needed to include more aspects than catch and biomass. **Figure 10** compares five additional indicators (i) the mean trophic level of the catch (MTLCatch); (ii) the mean trophic level of the biomass (MTLBiom); (iii) the relationship between zooplankton and pelagic fish (ZooPel); (iv) the relationship between pelagic fish and demersal fish (DemPel); and (v) the value of the commercial catch only, by the three combined scenarios as well as the control run. The fully balanced run "BH on all" was the scenario which gave the highest value for all indicators except the mean trophic level of the catch. Conversely, the historic control run gave the opposite result with the lowest values for all indicators except MTLCatch. The scenarios where either the commercial or the non-commercial species were harvested balanced ended up somewhere in between the other two, with "BH on commercial" giving a higher value of the catch, while "BH on non-commercial" gave a higher mean trophic level of the biomass in the system. The "BH on all" scenario also scored the highest on the ratio of zooplankton to pelagic fish, as well as demersal fish to pelagic fish.

## DISCUSSION

The first attempt of testing the BH regime within the Nordic and Barents Sea system was performed by applying a NoBa Atlantis model. The results were studied through changes in biomass, levels of spawning stock biomass, catch estimates and some indicators considering value and trophic structure.

## BH Effects on Individual Species

When making wide-ranging changes to the fishery across the ecosystem, it is not easy to distinguish between direct and indirect effects. In specifying fishing according to productivity BH effectively proposes changes to both the selectivity for each species and to the relative fishing pressure between species. Given that these changes could be implemented separately it is of benefit to investigate how much of any change comes from the changes to the fishery on a given species, and how much comes from the indirect ecosystem effects of changing the between species balance in fishing pressure.

The results of the BH on individual species are presented in the **Supplementary Material**, and a few key findings are highlighted here.

Nearly all commercial species are presently fished close to their respective estimated MSY, with the exception of Greenland halibut, golden redfish and capelin (ICES, 2018b,c). For Greenland halibut, the current assessment model is tuned only to length data, and estimates of FMSY are uncertain (ICES, 2015b). It is therefore difficult to make a direct comparison. The long term catches predicted from the BH run were slightly lower than under the historical fishing scenario, and with a much lower stock biomass. Golden redfish is presently overfished (ICES, 2018b), and the BH fishing pressure was much lower than the historical fishing levels. Applying this reduced fishing pressure led to stock recovery and eventually higher catches. For cod, applying BH to cod alone resulted in lower long term catches than under historical fishing. However, applying BH to all components of the ecosystem increased the cod catch to only slightly below that under historical fishing in the long term. Modeled BH on capelin suggested a higher fishing mortality (**Table 4**) resulting in up to 3 million tons extra yield. However, the capelin fishery within Atlantis is modeled as a constant fishing mortality which is known to be a poor fishing strategy for short-lived stocks with large fluctuations in biomass, such as capelin. This does not match the actual management of this stock as the HCR of capelin is a so-called escapement strategy, in which a certain amount is allowed to spawn and only the surplus may be caught, which allows for large interannual fluctuations in yield. This dynamic

fishing regime is not well-replicated in the current Atlantis model, and therefore comparisons to the actual fishery are problematic for this stock.

These four stocks highlight several points of caution with the modeling conducted here, as well as the importance of considering the dynamics of the individual species. Firstly, where the stock status and reference points are unclear, the modeling and comparisons become uncertain. Secondly, where a stock is currently overfished (such as the golden redfish), then any reduction in fishing pressure is likely to be beneficial. In this case BH would aid stock recovery, but reductions in fishing pressure could equally be achieved without employing BH. In general, this is an indication that simulation testing of the merits of BH compared against good practice traditional management should compare to well-managed fisheries rather than to a current depleted stock status. For a predator such as cod, examining the differences between changing the fishing only on cod and on the whole ecosystem highlights the possibility for ecosystem level effects to partially compensate (or potentially exacerbate) for the expected catch losses. Finally, capelin represents an example where the fisheries management is not well-captured in the Atlantis model and where the results should therefore be treated with caution. Furthermore, capelin represents an example of highly variable short-lived species where a fixed fishing pressure is a poor fishing strategy, and therefore an example where a BH strategy would need to be extended to encompass these dynamics, such as the density-dependent BH1 (Eq. 1), which has not been studied in this analysis.

More single species details are presented in the **Supplementary Material**, but this overview serves to highlight that individual species dynamics are critical to the outcome of applying BH to an ecosystem as a whole, and that it is important to use modeling tools which are capable of resolving such detail.

## BH Effects on Total Catch

When comparing the control run with combined runs of BH on multiple species, the results indicated that more species being subjected to BH resulted in overall higher catches. However, the main increase among the commercial species came from capelin and prawns, which are two species that should be treated with caution. Modeling prawns in Atlantis models appears to be a well-known problem (B. Fulton, personal communication), and even in the stock assessments there are great uncertainties around biomass estimates (ICES, 2013). As mentioned, there are also difficulties in modeling the capelin fishery in the NoBa model, which cannot accommodate an escapement rule strategy. Capelin also has a complicated life-history strategy with a very high post-spawning mortality, which requires carefulness in the interpretation of the results.

When mesozooplankton was included in the combined balanced harvest regime, it completely dominated the potential total catch (**Figure 6A**) with nearly 80 million tons per year, which is 20 times more than the current total Norwegian catch (Fiskeridirektoratet, 2016). The fishing pressure on mesozooplankton was set to 25% of productivity, which resulted in a 50% decrease of the biomass (**Figure 4**), but this huge extraction of mesozooplankton had surprisingly small effects on other species. A harvest of nearly 80 million tons would not be feasible in the real world, but considering that the current quota is set to 165,000 tons (Fiskeridirektoratet, 2016) of a stock with a standing biomass of 30 million tons with an estimated annual production of 290 million tons, then an increased quota would likely have negligible direct effects. This study sets fishing pressure directly on each modeled species and does not account for bycatch of other species resulting from any changes in catches. We therefore do not account for any potential effects of increased bycatch of eggs and larvae on other species that could be expected from such a large increase in mesozooplankton catches.

Another interesting result was that most of the commercial species had less variable catches from year to year under a BH regime compared to the traditional fishing regime, suggesting that BH would produce steadier yields. This reflected the variations in fishing mortality, as the FBH were more stable compared to FHisto in the control run.

The simulations suggest that the gains from BH in the well-managed Barents and Norwegian Sea on already commercially exploited stocks are rather limited according to the current model. Although the higher fishing levels proposed by the BH regime produced higher catches for many species, it came at the cost of significantly reducing the standing biomass (**Figure 2**) and subsequent

of the commercial catch only, by the three combined scenarios as well as the control run.

decreasing catch per unit effort, and increasing risks of recruitment overfishing (4), with a few exceptions (e.g., beaked redfish).

However, this does not necessarily imply that BH is a bad idea. Howell et al. (2016) investigated the relationship between yield and production of 28 harvested species in the Norwegian and Barents sea based on an Ecopath model from Skaret and Pitcher (2016). They concluded that the current harvesting regime of the Norwegian and Barents Seas is already reasonably balanced, and more than most other marine systems (Kolding et al., 2016). This supports the finding that any extra yields would be expected to come largely from currently unexploited or underexploited species.

## BH Effects on Ecosystem Structure and Value of Catch

When considering the total value of the catch, the "BH on all" and "BH on commercial" scenarios gave the highest value. This came mainly from prawns, which were high in value and catches, in addition to capelin and beaked redfish. Beaked redfish is a relatively high-value species, so the increase in catches by almost

three times made a significant impact on the total value of the catch. Capelin, on the other hand, had a lower value compared to the other species but the catches in the BH runs were more than 10 times higher under the BH-scenarios, which increased the total value. However, the results from prawns and capelin should as mentioned be treated with caution. The potential economic benefits of the harvest of the non-commercial species should also be kept in mind, although such estimations were not performed here, as the prices and markets are unknown.

The trophic level of the catches was lowest under all the BH regimes, as would be expected since one of the main ideas behind BH is to increase harvest at lower trophic levels. However, this could be problematic as today's market generally value higher trophic level species. The size-specific selectivity applied to each species in proportion to productivity, generally suggest a stronger fishing pressure on younger fish compared to a conventional fishing regime. Obviously, catch of smaller fish in the NoBa area would also fetch lower prices in the current market, adding an economic cost (that we have not considered in this study). Ecologically, the increased F on young fish is just an adjustment of the selectivity curve to be better aligned with natural mortality (and productivity), than the current traditional gear selectivity curves, and therefore we would expect a reduced effect in terms of potential fisheries induced evolution (Law and Plank, 2018). The mean trophic level in the total biomass was correspondingly much higher under the BH regimes, which included harvest of non-commercial species, due to the increased removal of low trophic species, such as mesozooplankton from the system.

The ratios of ZooPel and DemPel were chosen to study how the structure of the ecosystem might change under BH scenarios. The relative amount of zooplankton to pelagic fish was biggest in the "BH on all" scenario. This was a bit surprising as this run included an enormous harvest of mesozooplankton. However, when looking at **Figure 9**, the removal of mesozooplankton was partly compensated for by an increase of other zooplankton species, mainly small zooplankton, which might have kept the overall zooplankton biomass stable while pelagic fish were harvested more intensely. The DemPel ratio was also greatest for the full BH scenario, and also here was this mainly caused by the decrease in pelagic fish rather than an increase of demersals. For both of these indicators the ratios doubled in the "BH on all" scenario compared to the control run, indicating that the amount of pelagic fish was halved compared to zooplankton and demersal fish. Removing such a large part of the "middle" trophic level could change the structure of the ecosystem over time with unknown consequences that should be considered. The results therefore indicate that implementing a full BH regime in the Norwegian and Barents Seas with an exploitation level of 25% of total production could cause more changes to the already exploited species in the form of reduced biomass, than would be gained in total yields. On the other hand, expanding the fisheries to target species that are not commercially exploited today, especially on lower trophic levels, could provide considerable extra yields, and in particular from mesozooplankton. Yet, lower trophic level species tend to be less economically beneficial, and large removal of certain trophic levels could pose unknown structural changes to the ecosystem which needs to be considered.

## Uncertainties and Future Studies

The results must be evaluated in terms of the assumptions and limitations of the applied model. Even though the Atlantis model is able to capture a wide range of the variability inherent in the ecosystems, increased uncertainty follows such increased complexity (Howell et al., 2016). Several assumptions and "guesstimates" had to be made to accommodate the lack of knowledge about processes and the absence of relevant data (Fulton, 2010; Hansen et al., 2016). Generality, precision and realism are three desired features in a model, but unfortunately highly complex models, with a multitude of parameters and high resolution, are generally not able to attain all three, and has a tendency to de-emphasizes one quality to optimize the other two (Olsen et al., 2016). Being an end-to-end model, Atlantis is designed to provide an overall context, but clearly some weaknesses and inexplicabilities have been discovered in this study.

One weakness is that the non-commercial functional groups are more uncertain during the parameterization and tuning of the model, both because the focus is on the "important" commercial species, and because there is less information on the noncommercial groups (Cecilie Hansen, personal communication). The first step in improving these results would be to do a comprehensive re-tuning of the non-commercial species to allow for a constant BH fishing mortality across all groups without the need for ad hoc adjustments as in this study (**Table 2**). The chosen constant exploitation level of 25% of estimated productivity is considered cautious and conservative since a level of up to 40% is usually considered sustainable even for forage species (Patterson, 1992; Pikitch et al., 2012). Thus, the need to reduce the exploitation level to 12.5% for most noncommercial species to avoid collapse in the present model would indicate that the parameterization and model tuning of these species may not be as accurate and robust as for the commercial species.

The method for calculating production and productivity levels for invertebrates should also be reviewed, as this resulted in very high FBH for both prawns and mesozooplankton. The production calculations for invertebrates was done differently than for vertebrates, as invertebrates are gathered into biomass pools with a given productivity, that was extrapolated by the total area.

As well as expanding this study on other ecosystems, the investigation should also be expanded to include various types of BH. As indicated, there are 2 types of BH suggested (Eqs 1 and 2), and it would be interesting to compare in a model like Atlantis how a BH based on production (BH1) would compare to a BH based on productivity (BH2), to see whether the assumed BH1 protection of species from extinction would be validated.

## Summary

Through scenarios with varying fishing pressure and balanced fishing patterns in proportion to calculated productivities, we investigated the interaction effects of harvesting different components in the ecosystem. The conclusions from these simulations were that:


We should acknowledge that the appropriate fishing levels and model parametrization for non-commercial species need further validation as these species in the current model were surprisingly much more vulnerable to collapse than the commercial, which does not make logical sense. Several other weaknesses and somewhat inexplicable results were identified in the current model, which illustrates the enormous amount of synoptic data needed in order to build robust end-toend models. In addition, biological production figures are not presently an output in Atlantis, and the methods used in this study to estimate productivity levels have not been previously tried or tested.

This is the first time a balanced fishing simulation on most living components of an aquatic ecosystem has been done using an Atlantis model, and all results should be considered tentative only. As usual, when endeavoring into uncharted and untested territory, we end up with more questions than answers. However, the broad outlines of the results are probably both robust and generic. Our findings indicate that while comparing BH with an overfished ecosystem may indicate that there are gains to be made, comparing BH with a relatively well-managed ecosystem gives a much more nuanced view. On currently exploited species, there were minimal economic gains in a BH pattern over the existing relatively well-managed fisheries regime, although the adverse biological side effects of a highly selective fishery may benefit more from a BH regime. Most of the gains identified, were a result of reducing fishing pressure on overfished species and by extending exploitation to currently unor lightly fished species.

While BH calls for harvesting "across the widest possible range of species, stocks, and sizes of an ecosystem" (Garcia et al., 2012), attention should be drawn to word "possible." There is no inherent requirement for BH to be applied to absolutely every component in an ecosystem. For example, it may be that, even under BH, societies might choose to exclude charismatic megafauna from the harvest. Or that certain species and size categories may be uneconomic to harvest. In examining the possible impacts of BH it is therefore important to employ models with a sufficient level of detail (such as Atlantis) in order to examine the detailed outcomes of different potential implementations of BH. We hope that this paper has demonstrated that this level of analysis is now achievable.

We would conclude by noting that with a rapidly growing human population, likely approaching 9 billion by 2050 (United Nations, 2019), the need for healthy food is one of our world's great challenges. The United Nations Sustainable Development Goals (SDGs) addresses in SDG2 the zero-hunger goal, in SDG3 good health, and in SDG14 conserving and sustainable use of life below water, as three of the 17 most important issues in the world (United Nations, 2015). The demand for nutritious and healthy food has never been more important, and there is an urgent need for developing new sustainable harvesting strategies that ensure increased food production without depleting the ecosystem. Today only 3% of the food is harvested from the oceans (Field et al., 1998) which suggest an untapped potential. This study demonstrates the potential of expanding our harvest pattern to unexploited species without dramatic disruption of the system, which is one of the primary objectives of BH.

## DATA AVAILABILITY STATEMENT

The datasets analyzed in this article are not publicly available. Requests to access the datasets should be directed to IN.

## AUTHOR CONTRIBUTIONS

This article is based on the Master thesis of IN (Nilsen, 2018), jointly supervised by JK, CH, and DH. In the following preparation for the manuscript, IN was the lead author, while JK, CH, and DH contributed by editing and revision.

## FUNDING

This work was carried out as part and funded by the Institute of Marine Research (IMR) Strategic Project 445 'Reduced Uncertainty in Stock Assessment' (2016–2020), project number 3680\_14809. It was also funded by the Research Council of Norway through the project The Nansen Legacy (project number 276730) which is a contribution to the Barents Sea Ecosystem Program at IMR. In addition, we are grateful to Isaac Kaplan, Gavin Fay, and Kelli Johnson for providing R-scripts to produce **Figure 9**.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2020.00070/full#supplementary-material

## REFERENCES

fmars-07-00070 March 4, 2020 Time: 15:2 # 17


end-to-end ecosystem model to parameter perturbations of key species. PLoS One 14:e0210419. doi: 10.1371/journal.pone.0210419


in subareas 1–8 and 14, and in Division 9.a (the Northeast Atlantic and adjacent waters) Rep. ICES Advis. Comm, (Copenhagen: ICES).


**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 © 2020 Nilsen, Kolding, Hansen and Howell. 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.

# Ensemble Projections of Future Climate Change Impacts on the Eastern Bering Sea Food Web Using a Multispecies Size Spectrum Model

Jonathan C. P. Reum1,2,3 \*, Julia L. Blanchard<sup>2</sup> , Kirstin K. Holsman<sup>1</sup> , Kerim Aydin<sup>1</sup> , Anne B. Hollowed<sup>1</sup> , Albert J. Hermann4,5, Wei Cheng4,5, Amanda Faig1,3, Alan C. Haynie<sup>1</sup> and André E. Punt<sup>3</sup>

<sup>1</sup> Alaska Fisheries Science Center, National Marine Fisheries Service, NOAA, Seattle, WA, United States, <sup>2</sup> Institute for Marine and Antarctic Studies and Centre for Marine Socioecology, University of Tasmania, Hobart, TAS, Australia, <sup>3</sup> School of Aquatic and Fishery Sciences, University of Washington, Seattle, WA, United States, <sup>4</sup> Joint Institute for the Study of the Atmosphere and Ocean, University of Washington, Seattle, WA, United States, <sup>5</sup> Pacific Marine Environmental Laboratory, Office of Oceanic and Atmospheric Research, NOAA, Seattle, WA, United States

#### Edited by:

Erik Olsen, Norwegian Institute of Marine Research (IMR), Norway

#### Reviewed by:

Martina Hanneliese Stiasny, Norwegian Institute of Marine Research (IMR), Norway Morgane Travers-Trolet, Institut Français de Recherche pour l'Exploitation de la Mer (IFREMER), France

> \*Correspondence: Jonathan C. P. Reum Jonathan.Reum@noaa.gov

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 19 August 2019 Accepted: 17 February 2020 Published: 17 March 2020

#### Citation:

Reum JCP, Blanchard JL, Holsman KK, Aydin K, Hollowed AB, Hermann AJ, Cheng W, Faig A, Haynie AC and Punt AE (2020) Ensemble Projections of Future Climate Change Impacts on the Eastern Bering Sea Food Web Using a Multispecies Size Spectrum Model. Front. Mar. Sci. 7:124. doi: 10.3389/fmars.2020.00124 Characterization of uncertainty (variance) in ecosystem projections under climate change is still rare despite its importance for informing decision-making and prioritizing research. We developed an ensemble modeling framework to evaluate the relative importance of different uncertainty sources for food web projections of the eastern Bering Sea (EBS). Specifically, dynamically downscaled projections from Earth System Models (ESM) under different greenhouse gas emission scenarios (GHG) were used to force a multispecies size spectrum model (MSSM) of the EBS food web. In addition to ESM and GHG uncertainty, we incorporated uncertainty from different plausible fisheries management scenarios reflecting shifts in the total allowable catch of flatfish and gadids and different assumptions regarding temperature-dependencies on biological rates in the MSSM. Relative to historical averages (1994–2014), end-of-century (2080–2100 average) ensemble projections of community spawner stock biomass, catches, and mean body size (±standard deviation) decreased by 36% (±21%), 61% (±27%), and 38% (±25%), respectively. Long-term trends were, on average, also negative for the majority of species, but the level of trend consistency between ensemble projections was low for most species. Projection uncertainty for model outputs from ∼2020 to 2040 was driven by inter-annual climate variability for 85% of species and the community as a whole. Thereafter, structural uncertainty (different ESMs, temperaturedependency assumptions) dominated projection uncertainty. Fishery management and GHG emissions scenarios contributed little (<10%) to projection uncertainty, with the exception of catches for a subset of flatfishes which were dominated by fishery management scenarios. Long-term outcomes were improved in most cases under a moderate "mitigation" relative to a high "business-as-usual" GHG emissions scenario and we show how inclusion of temperature-dependencies on processes related to body growth and intrinsic (non-predation) natural mortality can strongly influence projections in potentially non-additive ways. Narrowing the spread of long-term projections in future ensemble simulations will depend primarily on whether the set of ESMs and food web models considered behave more or less similarly to one another relative to the present models sets. Further model skill assessment and data integration are needed to aid in the reduction and quantification of uncertainties if we are to advance predictive ecology.

Keywords: uncertainty partitioning, predictive ecology, Arrhenius factor, body size, size-based food web, cumulative effects, commonality analysis

## INTRODUCTION

fmars-07-00124 March 14, 2020 Time: 17:27 # 2

Anthropogenic climate change is expected to have significant impacts on ocean biogeochemistry, primary and secondary production, and the distribution and productivity of higher trophic level species (Doney et al., 2012; Mora et al., 2013; Pecl et al., 2017). Given the complexity and large spatiotemporal scales at which marine ecosystems operate, modeling approaches are necessary for inferring possible outcomes and tradeoffs due to climate change. For large marine ecosystems, models of varying complexity have been used to project potential impacts on community structure, size composition, and fishery catches, and to evaluate management strategies under climate change (e.g. Niiranen et al., 2013; Barange et al., 2014; Marshall et al., 2017). However, efforts to quantify uncertainty in climate-forced ecological projections have lagged, which limits their utility for informing ecosystem approaches to management and decisionmaking (Payne et al., 2015; Cheung et al., 2016).

Ensemble modeling enables representation of multiple sources of uncertainty. The approach entails developing a set of models, with each model member representing different working hypotheses or alternative formulations of uncertain processes. For instance, regional studies have evaluated structural uncertainty using model ensembles that consist of different formulations of species interactions (Gårdmark et al., 2013) or biogeochemical processes (MacKenzie et al., 2012; Meier et al., 2012; Niiranen et al., 2013). However, climate-forced ecological projections also depend on assumptions regarding future socioeconomic policies, markets, or technological developments that result in different greenhouse gas (GHG) emissions scenarios (Payne et al., 2015; Cheung et al., 2016). For specific marine ecosystems, scenario uncertainty could also encompass implementations of different policies, for instance, that impact fisheries regulations or coastal land use patterns. The distribution of projected outcomes conveys the level of confidence conditional on the set of alternative future scenarios and the model ensemble. Moreover, the uncertainty can be partitioned according to source which helps characterize their relative influence on the projection spread and informs where gains in precision may be made, for instance, through model refinement, additional research and observations, or advances in theory (Hawkins and Sutton, 2009; Cheung et al., 2016). Ensemble modeling is now widely used in weather and climate forecasting (e.g. Murphy et al., 2004; Berliner and Kim, 2008), but remains underutilized with respect to climate-driven ecosystem projections (Cheung et al., 2016).

Mechanistic food web models offer a powerful framework for exploring potential tradeoffs and uncertainties under different environmental or management scenarios (Persson et al., 2014). In marine and freshwater ecosystems, predation interactions are strongly size-structured and size-based food web models offer a relatively simple way to capture key aspects of system dynamics (Kerr and Dickie, 2001; Andersen et al., 2016; Guiet et al., 2016; Blanchard et al., 2017). Size spectra depict the abundance of individuals as a continuous function of body mass, and the first dynamic size spectrum models were developed to explain regularities observed in the scaling of abundance with body mass in lake and ocean ecosystems [reviewed in Blanchard et al. (2017)]. In size spectra models, system dynamics emerge from rules regarding the prey size preference of predators and the allocation of ingested energy toward maintenance costs, growth, and reproduction. The models are effective at capturing largescale patterns in fisheries production despite their simplicity, and can be forced with Earth System Model (ESM) outputs to project future community size structure and bulk fisheries production (Blanchard et al., 2012; Woodworth-Jefcoats et al., 2013; Barange et al., 2014; Lefort et al., 2015). Recent extensions to the modeling framework, however, permit explicit representation of multiple interacting species and their fisheries (Andersen et al., 2016; Blanchard et al., 2017). Species can be distinguished according to life history and prey size and species preference and predation, growth, and reproduction are represented at the individuallevel using a dynamic energy budget framework (Hartvig et al., 2011; Blanchard et al., 2014; Andersen et al., 2016). This latter feature makes multispecies size spectrum models (MSSMs) strong candidates for evaluating climate impacts because the hypothesized effects of climate variables (e.g. temperature) on animal energy budgets can be modeled in a more mechanistic fashion and scaled up to the population and community levels (Maury and Poggiale, 2013; Lefort et al., 2015; Guiet et al., 2016; Woodworth-Jefcoats et al., 2019).

Here, we evaluated future climate impacts on the eastern Bering Sea (EBS) food web using an MSSM and ensemble modeling approach (**Figure 1**). The EBS is a highly productive, semi-enclosed subpolar sea that overlays a broad continental shelf (average width ∼500 km). Although physical and biological conditions in the EBS are characterized by high interannual variation (Stabeno et al., 2001), climate change is expected to have multiple impacts. Warmer conditions are expected to reduce the southern spatial extent and duration of seasonal sea ice cover, advance the spring transition, and increase water column stratification, which may negatively impact phytoplankton, zooplankton, and benthic production (Hermann et al., 2013, 2016, 2019). Among global-scale simulation studies of climate change impacts, projections for the EBS are inconsistent and include forecasts of increased total catch potential (Cheung et al., 2010), negligible shifts in pelagic fish biomass (Lefort et al., 2015),

and moderate reductions in total fish biomass (Lotze et al., 2018). However, out of practical necessity these studies lack taxonomic detail, make simplifying assumptions regarding growth potential and trophic structure, and are forced with ESM projections with coarse spatial resolution. To better understand the implications of climate change for higher trophic level species and their fisheries in the EBS, the interdisciplinary Alaska Climate Integrated Modeling project (ACLIM) was initiated by the NOAA Alaska Fisheries Science Center (Hollowed et al., 2019). As a component of ACLIM, capacity to dynamically downscale ESM projections to the EBS was expanded (Hermann et al., 2019) and an MSSM was developed for and calibrated to the EBS (Reum et al., 2019).

In this study, we built upon these advances and produced ensemble projections of the EBS food web that incorporated multiple sources of uncertainty (**Figure 1**). Specifically, we included two sources of structural uncertainty (**Figure 1**). First, down-scaled climate projections for the EBS differ across ESMs (Hermann et al., 2019). We therefore used downscaled climate projections from multiple ESMs. Second, we addressed structural uncertainty related to possible temperature-dependences in biological rates. Temperature influences metabolism, which may impact multiple processes including physiological rates that affect body growth (Kooijman, 2000; Brown et al., 2004) as well as "intrinsic" or non-predation natural mortality (i.e. disease, senescence rates; Munch and Salinas, 2009; Keil et al., 2015). Previous size-based studies have included temperaturedependencies on both body growth-related and intrinsic natural mortality rates (Blanchard et al., 2012; Woodworth-Jefcoats et al., 2013; Lefort et al., 2015), but biological rates can exhibit different scaling relationships with temperature (e.g. Englund et al., 2011; Brown et al., 2004; Rall et al., 2012) and it remains unclear to what degree these two processes influence emergent features of the food web. To account for this uncertainty, we considered multiple MSSM variants that differed with regard to how temperature affects key biological rates (**Figure 1**). We also included two sources of scenario uncertainty. The first relates to different scenarios of future GHG emissions (Payne et al., 2015; Cheung et al., 2016). The second corresponds to different fisheries management scenarios that, relative to status quo, prioritize fishing on different components of the EBS food web and reflect trade-offs fishery managers are confronted with in setting total allowable catches for each stock (**Figure 1**; Hollowed et al., 2019).

Our main objectives were to: (1) develop ensemble projections of fish and invertebrate community composition and size structure and partition projection uncertainties according to source over various time horizons and (2) evaluate how different sources of uncertainty interact. For the latter objective, we specifically sought to clarify how policy and decisionmaking at regional and intergovernmental scales interact by comparing catches and community composition under alternative fishery management scenarios and either businessas-usual or mitigation GHG emissions scenarios. Further, we evaluated how temperature-dependencies operating on individual-level processes related to body growth and intrinsic natural mortality influenced emergent community- and specieslevel projections and whether their combined effects were additive or not (e.g. Crain et al., 2008; Kaplan et al., 2013).

## MATERIALS AND METHODS

fmars-07-00124 March 14, 2020 Time: 17:27 # 4

## Overview of Modeling Approach

Our modeling framework included three components, each of which supplied outputs that flowed unidirectionally to the next component (**Figure 1**). The first component (A) consisted of IPCC-class ESMs forced using various GHG emission scenarios based on IPCC Representative Concentration Pathways (R). In the second component (B), ESM projections were dynamically downscaled to the EBS. Specifically, the ESM projections provided boundary conditions for a 10 km spatial resolution regional biophysical model (Hermann et al., 2016, 2019). In the third component (C), a MSSM was forced using dynamically downscaled temperatures and values for phytoplankton, zooplankton, and benthos standing stock from B. Four versions of the MSSM were used in the ensemble to evaluate alternative assumptions regarding temperature-dependencies on biological processes. Aspects of structural uncertainty were accounted for at the ESM and MSSM components (**Figure 1**), and the model ensemble consisted of all possible combinations of ESMs and MSSMs. Next, we provide an overview of the climate downscaling approach and MSSM implementation. Additional details of the MSSM equations, parameterization, and calibration are available in Reum et al. (2019) and **Supplementary Appendix I**; details regarding the biophysical model are available in Hermann et al. (2016) and Hermann et al. (2019).

## A and B: Earth System Models and Dynamic Downscaling of Climate

Due to the computational demands of dynamically downscaling regional climate projections, we assembled an "ensemble of opportunity" (e.g. Tebaldi and Knutti, 2007) that consisted of two sets of previously published downscaled projections from ESMs and GHG emissions scenarios used in the Intergovernmental Panel on Climate Change Assessment Reports (IPCC AR4 and AR5) and archived by the Coupled Model Inter-comparison Project (CMIP) (Hermann et al., 2016, 2019).

The first set of ESMs included downscaled projections from three ESMs. These included the Coupled Global Climate Model, T47 grid (CGCM3-t4) from the Canadian Centre for Climate Modeling and Analysis, the Hamburg Atmosphere-Ocean Coupled Circulation Model (ECHO-G), and the Model for Interdisciplinary Research on Climate, medium-resolution version (MIROC3.2-Medres; Hermann et al., 2016). We refer to these ESMs as CGCM, ECHO, and MIROC, respectively. All ESM outputs from this set were archived by CMIP3 (Meehl et al., 2007). Projections were obtained through 2040 for IPCC Special Report on Emissions Scenarios (SRES) A1B, which corresponds to a future scenario with moderate GHG abatement (Hermann et al., 2016; **Table 1**).

The second set of ESMs included: MIROC, the Community Earth System Model (CESM), and the Geophysical Fluid Dynamic Laboratory ESM2M model (GFDL-ESM2M, herein simply GFDL; **Table 1**). These ESMs were archived by CMIP5 and span CMIP5 member variability (Taylor et al., 2012). Under this set of ESMs, downscaling of projections were performed through 2100 when possible for IPCC Representative Concentration Pathways (RCPs) 4.5 and 8.5, which correspond to futures with moderate and "business as usual" GHG emissions, respectively (Hermann et al., 2019; **Table 1**). ESMs in both sets were originally selected based on performance for the Bering Sea under present conditions and the availability of physical and biogeochemical output (Hermann et al., 2016, 2019).

ESM projections were dynamically downscaled using a biophysical model for the EBS (Hermann et al., 2019). Briefly, daily atmospheric and monthly oceanic outputs from the ESMs were interpolated in space and time for use in the surface forcing and boundary conditions for the regional model (Hermann et al., 2013). The model was implemented at ∼10 km spatial resolution with ten vertical layers and spans the entire Bering Sea (Hermann et al., 2013). The biological component of the model consists of a Nutrient-Phytoplankton-Zooplankton model (NPZ) developed by Gibson and Spitz (2011) with modifications by Hermann et al. (2016). Biological groups in the biophysical model include small phytoplankton, large phytoplankton, microzooplankton, small copepods, large copepods, krill (euphausiids), jellyfish, and slow and fast sinking detritus, benthic detritus, and benthic infauna. In addition to the ESM projections, a hindcast simulation for the EBS was generated for year 1970–2015 using historical reanalysis atmospheric forcing and ocean lateral boundary conditions (Hermann et al., 2016).

## C: Multispecies Size Spectrum Model

The MSSM is based on source code for the R package "mizer" (Scott et al., 2014), as modified by Reum et al. (2019) and with additional updates (**Supplementary Appendix I**). The MSSM captures predator-prey interactions between fish and crab species and includes a submodel to represent the catch allocation process for EBS fisheries. In total, the model includes

TABLE 1 | Overview of the temporal extent of projections from Earth System Models (ESMs) that were used to generate ensemble predictions of the eastern Bering Sea food web.


Projections were available for GHG emissions Special Report on Emissions Scenarios (SRES) A1B and Representative Concentration Pathways (RCPs) 4.5 and 8.5. SRES A1B and RCP 4.5 correspond to moderate and strong mitigation, respectively. RCP 8.5 corresponds to an unmitigated future greenhouse gas emissions scenario. All climate projections start in 2007.

nine fish species, three crab species, and three fish functional groups (**Supplementary Table 1**). The included species support economically significant fisheries or are important prey items for other predators in the EBS, and combined, accounted for ∼95% of the community biomass based on estimates from annual bottom trawl surveys. The species are able to feed on each other, as well as two background spectra that represent additional pelagic and benthic prey resources. Predator species in the MSSMs are distinguished by several traits including maturation and maximum sizes, feeding and growth rates, and preferences for prey species and sizes (**Supplementary Table 1**). Additional details of the core model structure and prey selection parameterizations are available in Reum et al. (2019).

The submodel describing catch allocation in the EBS was incorporated into the MSSM to represent fishery management scenarios (**Supplementary Appendix I**). The aggregate total allowable catch (TAC) for several finfish and a few invertebrate fisheries is capped at 2 million metric tons for the larger Bering Sea-Aleutian Islands fisheries management zone (Livingston et al., 2011). Given this constraint, the North Pacific Fishery Management Council (a regional body that provides management recommendations for fisheries within the United States Economic Exclusive Zone surrounding Alaska) sets TACs by species based on stock assessment estimates of acceptable biological catch (ABC) and consideration of other factors such as market capacity, bycatch constraints, and fleet interests. A model describing TAC allocation for EBS fisheries, and that specifically used historical Council and fishery data to translate ABC to TAC and TAC to catches was adapted to generate catch predictions for each species depending on the fishery scenario (**Supplementary Appendix I**). ABCs were calculated for each species, based on current sloped harvest control rules that are intended to provide conservative catch recommendations, and the submodel returned realized catches that were used to calculate fishing mortality rates and total mortality calculations. The fishery submodel and ABC calculations are described further in **Supplementary Appendix I**.

At the MSSM level we sought to incorporate uncertainty into ensemble projections related to assumptions regarding temperature-dependencies on biological rates and, specifically, to evaluate the individual and interactive effects of temperaturedependencies on rates that influence body growth and intrinsic natural mortality. To do so, four MSSM variants were developed (**Figure 1**). The "baseline" model (M1) lacked temperature effects altogether, but background pelagic and benthic spectra were forced using downscaled projections from the biophysical model (**Figure 1**). The remaining models shared the same structure and forcings as M1, but differed in regard to whether temperaturedependencies were applied to the two categories of rates. The body growth category included maximum consumption, prey encounter, and metabolism rates and the intrinsic natural mortality category consisted solely of the intrinsic natural mortality rate which represents all mortality not explicitly captured by predation or fisheries in the model (Andersen et al., 2016). In mizer, intrinsic natural mortality is constant across body mass classes within species and calculated as an allometric function of species maximum body size (Scott et al., 2014) such that smaller species experience higher intrinsic natural mortality rates relative to larger species (Hartvig et al., 2011). The three additional MSSM variants (**Figure 1**) included temperaturedependencies in body growth-related rates (M2), intrinsic natural mortality rates (M3), and both body growth-related and intrinsic natural mortality rates (M4).

In models M2-4, Arrhenius temperature-dependent correction factors (Brown et al., 2004) were applied to biological rates. Originally intended for describing temperature effects on chemical reaction rates, the Arrhenius function is also appropriate for approximating temperature effects on metabolism and other biological rates at the individual, population, and community levels over environmentally plausible temperature ranges (Kooijman, 2000; Brown et al., 2004). For a given rate τ, the Arrhenius-corrected value at temperature T (in Kelvin) was obtained following Eq. (1):

$$\pi(T) = \pi(T\_{ref})e^{\frac{\pi}{\hbar}\left(\frac{1}{T\_{ref}} - \frac{1}{T}\right)}\tag{1}$$

where E is the activation energy of heterotrophic metabolism (0.63 eV), k is the Boltzmann constant, 8.62 × 10−<sup>5</sup> eV K−<sup>1</sup> , and Tref is the reference temperature (Brown et al., 2004). Temperature forcing was based on downscaled depth-averaged temperature projections that were averaged spatially and within 3 month intervals starting in January, in accordance with the time step of the MSSM. A seasonal Tref was therefore used and was obtained from averaging downscaled hindcast of depth-averaged temperatures over the model calibration period (1982–1991). All downscaled time series of temperature, benthos, and pelagic prey used in projections were bias-corrected relative to mean seasonal differences with the hindcast for the overlapping period 2002– 2014 (**Supplementary Figure 1**). Details of the bias-correction calculation are presented in **Supplementary Appendix I**.

We calibrated the MSSM using a multistep process that included the estimation of parameters that scale species abundances and tuning of prey species preferences. The model was calibrated to time-averaged estimates of SSBs, catches, and diets from the 1980s (1982–1991; **Supplementary Appendix I**). Additional post-calibration modifications were made to the baseline natural (non-predation) mortality rates of several species to improve correspondence between projected SSBs and stock assessment estimates and ensure that predators exhibited levels of density-dependent recruitment that were commensurate with levels implied by time series of recruitment and SSB from stock assessments (**Supplementary Appendix I**). To validate the final calibrated models, all four variants of the MSSM were forced with historical fishing mortality rates (Fs) and hindcast time series of temperature and benthic and pelagic resource spectra from 1982 to 2014. Four validation criteria were evaluated: (1) correspondence of diet projections to data from outside the calibration time period (2005–2014); (2) correspondence between observed and predicted weight-at-age relationships; (3) overlap in the 95% confidence intervals for long-term linear trends between projected and observed SSBs; and (4) continued persistence of stocks when the models were projected forward

assuming average historical climate conditions and status quo fisheries management from 2014 through 2100.

We focused on matches between long-term trends rather than simple correlation because population dynamics are partly controlled by stochastic recruitment events and these processes are not represented in the current class of MSSMs. This issue extends to other types of marine food web models where emphasis in model tuning has commonly been placed on matching averages and trends (e.g. Kaplan and Marshall, 2016). The last criteria was based on the observation that no finfish stocks have been overfished in the EBS and that the healthy status of EBS stocks is attributable in part to current (status quo) management practices (Livingston et al., 2011). We provide a thorough overview of the calibration and validation procedure in **Supplementary Appendix I**. The final post-calibrated MSSMs (all four variants) met the validation criteria and produced longterm trends that were similar to those from stock assessments (**Supplementary Appendix I**).

## Fishery Management Scenarios

We evaluated three fishery management scenarios based on current policies for setting total allowable catch (TAC). This procedure is essential because the Bering Sea-Aleutian Islands ecosystem cap requires that individual species TACs be reduced so that the sum of all species is at or below the 2 million metric ton ecosystem cap (Hollowed et al., 2019). We considered scenarios in which: (1) TAC was allocated based on recent historical patterns ("status quo"); (2) pollock and Pacific cod TAC is increased up to 10% relative to status quo at the cost of lower flatfish TAC, and (3) flatfish TAC was increased up to 10% relative to status quo at the cost of lower pollock and Pacific cod TAC. For brevity, we herein refer to the scenarios as the "status quo," "gadid," and "flatfish" scenarios, respectively. The gadid and flatfish scenarios have been developed through examinations of historical fishing data and extensive conversations with members of the North Pacific Fishery Management Council and other stakeholders about the key decisions in the TAC-setting process. The scenarios represent realistic shifts in management and harvest behavior along what managers have identified as a key axis of decision-making. The shift could be motivated by combinations of economic factors, more stringent bycatch limits in different fisheries, or technological improvements that reduce the cost of bycatch avoidance.

## Ensemble Projections

In total, seven ESMs were included in the model ensemble, and projections of these models under multiple GHG emissions scenarios were obtained, resulting in 11 unique ESM and GHG emission scenario projections that were downscaled to the EBS (**Table 1**). In turn, each unique downscaled projection was used to force M1-4 under the three catch allocation scenarios, resulting in 11 · 4 · 3 = 132 ensemble projections. All simulations were initiated in 1982 and forced with historical fishing mortality rates through 2014 (e.g. Blanchard et al., 2014) and thereafter downscaled bias-corrected projections of plankton and benthos prey (M1-4) and depth-averaged temperatures (M2-4). Catches and fishing mortality rates after 2014 were obtained from the catch allocation submodel.

## Partitioning Uncertainty

We partitioned uncertainty (variance) in the ensemble projections into five distinct factors that were categorized as scenario (GHG and fishery management) and structural (ESM and MSSM) uncertainty and internal variability. Internal variability in climate projections on annual to decadal time scales includes phenomena such as the El Nino Southern Oscillation (ENSO), Pacific Decadal Oscillation (PDO), or North Atlantic Oscillation (NAO). In addition, internal variability can emerge within biological systems at similar time scales due to predatorprey cycles or other density-dependent growth, recruitment or mortality processes (Cheung et al., 2016). Internal variability is emergent in many types of complex systems, and outcomes are typically sensitive to initial starting conditions. If multiple realizations based on different starting conditions are available, the variance component associated with internal variability at a given time slice can be estimated along with other uncertainty sources using Analysis of Variance (ANOVA) models (e.g. Yip et al., 2011; Bosshard et al., 2013). However, similar to other climate and ecosystem simulation studies (e.g. Hawkins and Sutton, 2009; Gårdmark et al., 2013), our ensemble projections lacked multiple realizations that differ in initial conditions only. Instead, we used an alternative approach based on Hawkins and Sutton (2009) and Cheung et al. (2016).

First, the raw projection outputs y for each ESM m, MSSM variant v, GHG scenario g, fishing scenario f, and year t are written as:

$$\begin{aligned} &\chi\left(m,\nu,\mathbf{g},f,t\right) \\ &= z\left(m,\nu,\mathbf{g},f,t\right) + \left(\mu\_{ref}\left(m,\nu,\mathbf{g},f\right) + \varepsilon(m,\nu,\mathbf{g},f,t)\right) \end{aligned} \tag{2}$$

where a reference level (invariant in time) for each unique ensemble member is denoted by µref , the long-term trend with y is represented by a smooth spline function z, and the regression residual error (due to internal variability) is ε. For each ensemble realization, the reference level is the 1995–2014 mean state. The variance of y (Vy) is described as a function of time t following Eq. (3):

$$V\_{\mathcal{Y}}(t) = V\_{\mathcal{z}}(t) + V\_{\mathcal{e}} \tag{3}$$

The estimate of internal variability is the variance (Vε) of the residual regression error ε(m,v,r,f,t):

$$V\_{\varepsilon} = \frac{1}{N\_m N\_\nu N\_\emptyset N\_f T} \sum\_{m=1}^{N\_m} \sum\_{\nu=1}^{N\_\nu} \sum\_{g=1}^{N\_\xi} \sum\_{f=1}^{N\_f} \sum\_{t=1}^{T} [\varepsilon\left(m, \nu, g, f, t\right)]^2 \quad \text{(4)}$$

and is considered to have constant variance over time. In their original formulation, Hawkins and Sutton (2009) assumed Vε was constant across the complete projection time span. We instead calculate a Vε for each of three time blocks (2015– 2040, 2041–2080, and 2081–2100) to account for changes in the representation of different ESMs and GHG scenarios in the ensemble projections (**Table 1**).

Vz, the variance associated with z(m,v,g,f,t), is calculated as:

$$V\_z(t) = \frac{1}{N\_m N\_\nu N\_\mathcal{g} N\_\mathcal{f}} \sum\_{m=1}^{N\_m} \sum\_{\nu=1}^{N\_\nu} \sum\_{\xi=1}^{N\_\xi} \sum\_{f=1}^{N\_f} [z\left(\cdot,\cdot,\cdot,\cdot\right,t\right) $$

$$-z\left(m,\nu,\mathfrak{g},f,t\right) \text{l}^2 \quad \text{(5)}$$

where z (·, ·, ·, ·, t) is the overall mean at time-step t of the smooth spline trends; therefore, it measures the spread of ensemble simulations trends around the ensemble mean trend.

We used commonality analysis (Whittaker, 1984; Ray-Mukherjee et al., 2014) to decompose V<sup>z</sup> into components that were uniquely and jointly associated with the four structural and scenario factors at time-step t. The approach entails performing multiple regression on the response variable (z), estimating the proportion of variance "explained" (R 2 ) by the four factors, and decomposing R 2 into unique and shared components (Ray-Mukherjee et al., 2014). The method proceeds as follows. For factor x1, the proportion of variance uniquely explained by x<sup>1</sup> is obtained by first regressing z on the full set of factors (x1, x2, x3, and x4) and the proportion of variance explained by the model (R 2 1,2,3,4 ) is calculated. Note, only main effects are included in the regression model. A second regression model is then applied, but excluding x1. The proportion of variance uniquely explained by x<sup>1</sup> (R 2 1|2,3,4 ) is obtained by subtracting R 2 2,3,4 from R 2 1,2,3,4 . Variance jointly explained by x<sup>1</sup> and the remaining factors is found by regressing z on x1, obtaining the corresponding explained variance (R 2 1 ), and subtracting R 2 1|2,3,4 from R 2 1 .

The absence of replicates (multiple realizations based on different initial conditions) meant that a fully saturated regression model with second, third and fourth order interaction terms would have zero degrees of freedom and no residual error. Consequently, we ascribed the "unexplained variance" associated with a model consisting only of main effect terms (that is, 1 – R 2 1,2,3,4 ) to variance associated with higher order interactions. We calculated the total variance associated with interaction terms for comparative purposes, but did not decompose it further since these components can be minor relative to those associated with the main effects and difficult to reliably estimate when replication at the lowest levels is limited (e.g. Yip et al., 2011).

We partitioned uncertainty in projections of catch, SSB, and mean weight for the community in aggregate and for individual species. We grouped species according to similarities in the decomposition of their catch, SSB, and mean weight projection uncertainties over time using a hierarchical cluster analysis which was based on a Euclidean distance matrix of the partitioned uncertainties and using Ward's minimum variance criteria. For a given species, year, and variable, the partitioned uncertainties were expressed as proportions of the total uncertainty.

## Interactions Between Fishery Management and GHG Mitigation Scenarios

We evaluated 2090 (mean for years 2081–2100) ensemble projections of abundance size spectra and catches, SSBs, and mean body weights to identify (1) differences between fishing scenarios in a warmer future (RCP 8.5) and (2) potential improvement in outcomes if GHG mitigation (RCP 4.5) is pursued. Specifically, we calculated average changes in 2090 projections relative to historical (1995–2014) levels for each fishing scenario under RCP 8.5. The effect of GHG mitigation was calculated as the difference in 2090 outcomes under scenario RCP 4.5 from those under RCP 8.5.

We characterized the reliability of ensemble projections in terms of the level of agreement in projecting positive or negative changes in relative values (e.g. Meehl et al., 2007; Bopp et al., 2013; Bryndum-Buchholz et al., 2019). Percent sign agreement (SA) was calculated as:

$$SA = 100 \times |P - N| \, / n \tag{6}$$

where P and N are the total number of positive and negative projections, respectively, and n is the total number of projections in the ensemble. If 50% of the ensemble projections are positive and 50% are negative the resulting SA is zero because every positive projection is matched by an opposing negative projection and vice versa. We focused on SA to emphasize qualitative differences in long-term ensemble projections. We considered SA of projections "high" and "low" when values were ≥80 and <80%, respectively.

## Temperature Effects

We compared 2090 projections of relative change in abundance size spectra, catches, SSB, and mean body size under the different temperature models (M1-4). To simplify comparisons, the calculation was limited to projections made under the status quo fishing scenario and RCP 8.5 which included the largest changes in temperature. The ensemble mean and SA were calculated for each model output.

In addition, we calculated the cumulative effects of the temperature-dependency assumptions. Previous studies of cumulative ecological impacts have proposed methods for classifying interactions between stressors as synergistic, additive, or antagonistic (e.g. Crain et al., 2008; Griffith et al., 2011). However, if the individual effects of two "treatments" have opposing signs, assigning the interaction of the two treatments into these categories is not straightforward (Kaplan et al., 2013). Instead, we calculated whether the sum of the individual effects of temperature assumptions represented in M2 and M3 were above, below or similar to outcomes predicted under the combined model, M4 (Kaplan et al., 2013). The deviation (d1,2) of the interaction from the value expected if the individual effects were additive was obtained following (Kaplan et al., 2013):

$$d\_{1,2} = Y\_{\rm AB} + Y\_{\rm CT} - Y\_{\rm A} - Y\_{\rm B} \tag{7}$$

where YAB is the ensemble mean value (catch, SSB, or mean weight) predicted under M4 (the subscript AB denotes both "treatments" are included), YCT is the ensemble mean value predicted under M1 (the "control"), and Y<sup>A</sup> and Y<sup>B</sup> are the ensemble mean values under M2 and M3. All Y values are expressed as a percentage of the control value (YCT is always 100%). We considered values of d1,<sup>2</sup> from −5 to 5% as additive,

and values below and above the range non-additive negative and non-additive positive, respectively (Kaplan et al. (2013).

## RESULTS

## Community Ensemble Projections and Uncertainty

Projected changes in aggregate community catch, SSB, and mean body weight for the full ensemble trended negatively on average through the end of the century (see **Figure 2A**). Projected model outputs through 2060 spanned both positive and negative values but thereafter projected values were solely negative, that is, SA was 100% (**Figure 2A**; **Table 2**). Projection uncertainties for aggregate catch, SSB, and mean body mass were dominated by internal variability through ∼2040 (**Figure 2A**) but at longer time horizons (2040– 2100) structural uncertainties (i.e. ESM and temperaturedependencies) dominated (**Figure 2A**). For catches, temperaturedependencies composed the largest uncertainty source whereas ESM was the greatest source of uncertainty for SSB and mean body weight (**Figure 2A**).

These general patterns were also evidenced by the larger spread of projected values when averaged according to ESM and MSSM variants as opposed to the GHG emission or fishery management scenario (**Figure 2B**). Future catches were substantially lower (∼30%) when body growth-related rates depended on temperature (M2 and M4 vs M1 and M3; **Figure 2B**). For SSB and mean body weight, the spread of projected values averaged according to ESMs were larger, and projected values under MIROC (the ESM with the warmest projections; **Supplementary Figure 1**) were considerably lower than under the remaining models (**Figure 2B**).

Although GHG scenarios accounted for only a small proportion of ensemble projection uncertainty over time (**Figure 2A**), average catches and SSB after 2060 were somewhat higher under mitigation scenario RCP 4.5 relative to the businessas-usual RCP 8.5 (**Figure 2B**). Among the fishery scenarios, total catches under the flatfish scenario were consistently higher than under the two alternatives after 2060, but differences for SSB and mean body weight among fishing scenarios were small (**Figure 2B**). In general, uncertainty due to interactions among the various sources of uncertainty increased over time and was larger in magnitude to the proportion directly associated with scenario uncertainty by the end of the century (both GHG and fishery; **Figure 2A**). Uncertainty explained by multiple sources (overlap) was minor (<1%, **Figure 2B**) for all variables.

## Species Ensemble Projections and Uncertainty

Average full ensemble projections of relative change in SSB, catches, and mean body sizes for 2090 were negative for 66, 33, and 86% of species, respectively (**Table 2**), and for the majority of species, projections were more uncertain (as measured by SDs) than those for the aggregate community (**Table 2**). Overall, SDs ranged from 7 to 138% among model outputs (**Table 2**). Across model outputs, SA for projections were high for only four to five species and only pollock, the species with the largest biomass, had an SA value of 100% for all three model outputs (**Table 2**).

Species clustered into three groups based on similarity in the decomposition of projection uncertainty (**Figure 3A**). For the first group (yellowfin sole, Alaska plaice, other flatfish, and Alaska skate), internal variability and fishing scenario accounted for ∼10 to 50% of uncertainty in model outputs through 2040, but thereafter projection uncertainty was increasingly dominated by temperature assumptions (**Figure 3B**). Fishing scenario was the dominant source of uncertainty for catch projections in the second group (flathead sole and arrowtooth flounder) over time, but structural uncertainty sources (both ESMs and temperature assumptions) were important (>25%) after ∼2060 for SSB and mean body size (**Figure 3B**). For the third group, projection uncertainties were initially dominated by internal variability (∼50 to 75%), but after 2040, structural uncertainties became increasingly important (**Figure 3B**). For all groups and model outputs, uncertainty related to interactions between variables increased over time, and accounted for between ∼5 and 20% of uncertainty; GHG scenario uncertainty was a relatively minor contributor to uncertainty (<10%) for all groups and model outputs (**Figure 3B**).

## Fishery Management and GHG Emissions

Overall, the largest differences in 2090 projections across fishery management scenarios included catches for flatfishes (flathead sole, other flatfish), which were ∼25 to 50% higher under the flatfish (F3) relative to status quo (F1) and gadid scenarios (F2) (**Figure 4A**). Reductions in total community catch were also ∼25% less severe under F3 relative to F1 and F3 (**Figure 4A**). For the remaining species, differences between model scenarios were smaller (**Figure 4A**).

Across all fishing management scenarios, SSB reductions were projected for the aggregate community and for 11 of the 15 species (**Figure 4A**). Reductions of ∼25% or more were projected for 6 species (flathead sole, Northern rock sole, walleye pollock, tanner and snow crab, and Alaska skate) with high SA. Smaller reductions (less than ∼25%) with low SA were projected for other species (yellowfin sole, Pacific halibut, red king crab, foragefish, and sculpin; **Figure 4A**). Both increases and decreases were projected across fishery management scenarios for the remaining species, with the notable exception of arrowtooth flounder which was projected to increase ∼50% across all fishery management scenarios (**Figure 4A**). These general patterns were similar to those observed for mean body weight for most species (**Figure 4A**), and were reflected in relative changes in abundance size spectra (**Figure 5A**). For each fishing scenario, reductions in abundance were projected across most body masses except for the interval dominated by arrowtooth flounder (∼103.7–104.<sup>0</sup> g; **Figure 5A**).

Under the GHG mitigation RCP 4.5 scenario, projections of SSB increased relative to RCP 8.5 across fishery scenarios for the aggregate community and 11 individual species (**Figure 4B**). The level of increase for individual species ranged up to ∼50% (sculpin) but for most species and model outputs, the increase was closer to ∼25%. Species that decreased in terms

plots that indicate the proportion of total variance in ensemble projections explained by scenario [greenhouse gas emissions (GHG) and fishing] and structural [Earth system model (ESM) and temperature effect] uncertainty and internal variability. The proportions explained by interactions between factors, and variance mutually explained by multiple factors (overlap) are also indicated. The vertical white lines demarcate time periods that differ with respect to the number of ESM members. For (B), solid colored lines correspond to projection averages within levels of the uncertainty source. For reference, maximum and minimum ensemble projections are noted (dotted lines).

TABLE 2 | Mean ensemble projections of average (2081–2100) relative SSB, catches and average body size for eastern Bering Sea food web members.


Ensemble projections include members forced by output from ESMs under RCP 4.5 and 8.5. Projections are relative to 1994–2014 levels. The standard deviation of relative values and the% sign agreement (SA) of the ensemble projections are indicated.

of SSB, included red king crab, Alaska plaice, and Alaska skate and reductions were <12.5% (**Figure 4B**). Overall, community abundance levels of individuals under RCP 4.5 increased ∼10 to 40% across size classes (**Figure 5B**). Patterns of net change in mean body weight and catches for most species were similar to those for SSB between scenarios (**Figure 4B**).

across species within the three clusters. Species name abbreviations in panel (B): N. Northern; AT, arrowtooth.

## Temperature Sensitivity

At the community level, model outputs from MSSM variants that included temperature-dependencies on body growth (M2 and M4) were ∼25% lower than those that did not (**Figure 6**). This general pattern also extended to the abundance size spectra: in size classes >10<sup>2</sup> g reductions in abundance were consistently largest under M2 and M4 (**Figure 7A**). For individual species, model outputs under M4 (body growth and intrinsic natural mortality) where usually lower than those projected under M1 (status quo), but the difference in model outputs between M1 and M2 and M3 was variable across species (**Figure 6**). Roughly a third of species exhibited cumulative temperature effects that were additive, a third that were positive non-additive, and a third that were negative non-additive for each model output (**Figure 6**). A mixture of cumulative responses was also observed for the abundance size spectrum: positive responses were observed for body sizes near ∼101.<sup>8</sup> and ∼10<sup>4</sup> and negative responses dominated from between ∼10<sup>2</sup> and 103.<sup>8</sup> g (**Figure 7B**).

## DISCUSSION

Our ensemble projections for the EBS food web lead to at least four significant insights. First, we show that aggregate community SSB, catches, and mean body weight (which are weighted toward pollock and which declines overtime), are likely to decrease by 2090 but ensemble projections for the majority of individual species were a mixture of increasing and decreasing trends. Second, structural uncertainty (both ESM and temperaturedependencies) dominated long-term (2060–2100) projections for many aggregate and species-level variables, which contrasts with global climate model ensemble projections of physical variables. In those studies, GHG emissions scenarios typically dominate end-of-century projection uncertainty (e.g. Hawkins and Sutton, 2009). Third, we show that temperature-dependencies on individual-level processes can impact emergent communityand species-level variables in complex and often non-additive ways. This highlights a critical aspect of structural uncertainty in climate-driven food web projections and the importance of frameworks such as MSSMs for scaling temperaturedependencies in individual-level processes to populations and communities. Last, while contributing less to long-term projection uncertainty, the moderate GHG mitigation scenario RCP4.5 also decreased the severity of projected long-term reductions in SSB, catches, and mean body weight for the

majority of species relative to the business-as-usual emissions scenario across the different fisheries management scenarios. These outcomes demonstrate how policies and decision-making related to global GHG emissions may filter down to impact the trajectory of regional systems.

The results suggest future reductions in EBS benthic and pelagic prey resource spectra will decrease aggregate community biomass and fisheries yield. Overall, pollock composes ∼60% of the total fish biomass in the EBS and drove reductions in aggregate community biomass, catches, and mean body weight. Generally considered a forage species, pollock feed primarily on pelagic resource spectra prey and as they grow fish and benthic invertebrates comprise larger proportions of their diet. In downscaled projections, average pelagic and benthic resource spectrum prey densities decline ∼25% and 35% and 18 and 29% under RCP 4.5 and 8.5, respectively, by 2090. This largely caused the reductions in pollock productivity and aggregate community variables across fishery management scenarios. Interestingly, the negative trend is similar to projections from EBS pollock studies that estimated environmental stock-recruit relationships and forced recruitment with sea surface temperature projections using both single-species (Ianelli et al., 2011; Mueter et al., 2011) and age-structured multispecies models (Holsman et al., 2016; Ianelli et al., 2016; Spencer et al., 2016). The agreement in

between 2090 ensemble projections under GHG mitigation scenario RCP 4.5 and RCP 8.5. Species names abbreviations: AT, arrowtooth; A.K., Alaska; N., Northern.

pollock trends across the different regional modeling studies is encouraging in terms of establishing confidence in projections, and contrasts with inconsistent total fish biomass projections from global-scale simulation studies (Cheung et al., 2010; Lefort et al., 2015; Lotze et al., 2018). That said, the directions of long-term trends in the ensemble were mixed for most other species and, with a few exceptions (e.g. Wilderbuer et al., 2002; Hollowed et al., 2009; Szuwalski and Punt, 2012), other regional projection studies are unavailable for these species. The ambiguity indicates heightened caution is warranted in drawing conclusions regarding the absolute value of potential net effects of climate change on the majority of EBS species using only

the present model ensemble and that considerable room for improvement exists.

Our analysis of projection uncertainty provides descriptive summaries of key sources of variation, clustered species with similar sensitivities, and provides a basis for setting research priorities for refining ensemble projections (Evans et al., 2015; Cheung et al., 2016). Importantly, we show that structural uncertainties dominate intermediate- and long-term ensemble projections and, for the majority of species, ESMs were the largest uncertainty source. ESM climate projections are more variable for high latitude seas relative to other locations in part due to seasonal sea-ice cover dynamics that strongly impact other physical properties and the seasonal production cycle (Hawkins and Sutton, 2009). The number of ESMs used in the current study is small, but were selected to span CMIP5 member variability for EBS projections (Hermann et al., 2019). Decreasing this uncertainty source may be possible by applying more stringent EBS-specific validation criteria to further limit the ESM suite (Stock et al., 2011). Alternatively, including a larger set of IPCC-class ESMs in the ensemble could also help characterize the central tendency and spread of projections. Methods for expanding the number of ESMs in the ensemble, such as the development of statistical models to generate predictions of downscaled forcing variables based on relationships estimated from smaller subsets of dynamically downscaled ESMs, may prove valuable in this regard (e.g. Hermann et al., 2019).

The strong influence of temperature-dependencies on model outputs reinforces findings from sensitivity analyses performed on other size-based food web models (Maury et al., 2007) and highlights an important consideration when interpreting climate-driven projections from food web models forced with only primary production (e.g. Brown et al., 2010; Howell et al., 2013). At the community-level, temperature-dependencies on both categories of biological rates lowered catches, SSB, mean body weight, and abundance across size classes, but the level of decrease was highly variable across species and model outputs, in part because each species relies on different prey and is vulnerable to different predators. Consequently, the indirect effects of temperature that propagate through the food web may oppose or amplify direct temperature effects depending on the species and result in net outcomes that are difficult to anticipate. This complexity is exemplified in part by the mixture of additive, non-additive negative, and non-additive positive cumulative effects observed across body size intervals in the size spectrum and for individual species and model outputs. Ultimately, identifying how climate change impacts will manifest in ecological communities requires accounting for species interactions and our findings underscore the value of mechanistic models such as MSSMs for linking individual-level climate impacts to population and community-level outcomes.

Structural uncertainty related to temperature assumptions was also important for most long-term projections, but relative importance can also easily be change based on which models are represented in the ensemble. For instance, the baseline variant M1 represents an extreme endpoint and was included to bracket the range of model structures with regard to temperature and to evaluate potential non-additivity between different temperature-dependencies. Removing M1 from the ensemble would reduce projection uncertainty and could be justified based on the pervasive influence of temperature on biological rates (Brown et al., 2004). That said, in the absence of detailed species-specific information, the model set could also be expanded to represent other general but more nuanced hypotheses regarding temperature effects. For instance, the scaling of temperature-dependencies may change with ontogeny (Lindmark et al., 2018), differ across biological rates (Englund et al., 2011; Rall et al., 2012), or scale with temperature in a manner different from that described by the Arrhenius correction factor (Woodworth-Jefcoats et al., 2019). The latter may occur if species are currently at or near their thermal maximum. If so, additional warming could reduce rates ultimately controlling body growth, for

example (Woodworth-Jefcoats et al., 2019). In the EBS, this issue may emerge for some species, particularly those with restricted northern distributions (e.g. snow crab, Northern rock sole) if the warmest future scenarios are realized. As our understanding of temperature effects on individuals and communities evolves, the ensemble members can be updated to formalize other possibilities, for instance, the effects of temperature-driven changes in phenology or distribution, and help identify the most consequential assumptions to projecting future system states.

Despite uncertainty in the absolute value of climate change impacts, we show that pursing GHG mitigation scenario RCP 4.5 ameliorated reductions in catches, SSB, mean body size, and abundance relative to business-as-usual RCP 8.5. These findings add to a growing body of research that demonstrate potential benefits to advancing coordinated, global-scale policies that abate GHG emission rates (Barange et al., 2014; e.g. Bryndum-Buchholz et al., 2019; Lotze et al., 2018). Importantly, we also show that these benefits were realized across the different fishery management scenarios for the majority of species over the long-term and that community-level catches were highest from after ∼2060 under the flatfish scenario relative to the other two scenarios. This latter observation suggests fisheries on currently underutilized species such as Arrowtooth flounder, flathead sole, Alaska plaice, and other flatfishes may partly offset future losses in pollock catches owing to climate change. Realization of the fishery management scenarios, however, is based on additional contingencies such as the opening of new markets and improvement in fishing gear technology and therefore suggests a direction in which to steer the larger EBS socioecological system.

The fishery management scenarios we considered are merely a subset of potential options and are premised on current fishery management polices remaining intact into the future. However, the framework can easily be adapted to evaluate a wider range of fishery management strategies including the effects of significant policy changes, for instance, modification or elimination of the 2 million m ton cap on TAC or the use of versions of multispecies maximum sustainable yield (Collie and Gislason, 2001; Moffitt et al., 2016) for setting ABCs rather than the currently used singlespecies version. To further increase the realism of different fishery management scenarios, methods for updating the reference SSBs that are used to calculate ABCs on annual or semi-annual time-scales would also be desirable to more closely simulate the management decision-making process. While room for improvement exists, the representation of the complex TAC setting process in the EBS is a major strength of our modeling framework because different fishery management strategies and scenarios can be compared to status quo management, and will be useful for exploring futures based on different regionalized socioeconomic pathways (Maury et al., 2017).

Due to computational demands, we were unable to evaluate several additional uncertainty sources. For instance, we did not address parametric uncertainty, but we note that uncertainty in parameters controlling allometric relationships and life history traits can strongly influence MSSM projections (Zhang et al., 2015). Outputs from the biophysical model, such as primary production, are also sensitive to biological parameterization uncertainty (Gibson and Spitz, 2011) and the issue also extends to ESMs. We did not incorporate parameter uncertainty because of practical computational constraints and because we sought to focus on uncertainty sources that have received less treatment in the ecological literature (Cheung et al., 2016). That said, methods for efficiently sampling parameter space to represent this uncertainty in the ensemble are available (e.g. Gibson and Spitz, 2011; Thorpe et al., 2015) and it is an important area for future research. Stochasticity in the stock-recruit relationships was also not represented in the MSSM. Consequently, the projections are based on the assumption that average recruitment relationships hold over time. Stochasticity in recruitment (or in parameters that directly control recruitment), can be a major uncertainty source in MSSM predictions (Blanchard et al., 2014; Zhang et al., 2016) and quantifying this uncertainty would help frame the importance of improving basic understanding of recruitment processes relative to other aspects of system structure. Last, we focused on two major climate forcings (shifts in basal prey resources, temperature), but other climate effects including ocean acidification, deoxygenation, or distributional shifts due to changes in habitat may also be important future drivers on fish and crab dynamics. As projections of additional variables become available for the EBS (e.g. Pilcher et al., 2018) and our understanding of their biological impacts improves, the model ensemble can be updated to consider a larger array of climate drivers.

Modeling frameworks that link global climate processes to regional ecological systems are vital test beds for evaluating management strategies under climate change (e.g. Weijerman et al., 2016; Hollowed et al., 2019). The framework presented here makes significant inroads in this regard and offers a template for other systems. Overall, we show that communitylevel catches, SSB, and mean body size are likely to decline for the EBS over the following century, but the level of decline is dominated by structural uncertainty. For many individual species, structural uncertainty also dominated projections, but for a subset (e.g. Arrowtooth flounder, flathead sole) fishery management scenario was instead important. This information can help inform and prioritize development of more concerted research programs based on both the species and objective. While we have partly focused on one facet of structural uncertainty at the MSSM level, we note that other single and multispecies models may also offer plausible representations of EBS fish and crab species dynamics and a major goal of ACLIM is to bracket the possible range of ecological effects of climate change by including models that differ in terms of their strengths and weaknesses (Hollowed et al., 2019). We expect ensemble projections that include a broader set of structurally distinct higher trophic level models will increase projection uncertainty. The estimates in the current study should therefore be viewed as conservative.

The method we propose for decomposing projection uncertainty can easily be adapted to account for additional categories of uncertainty represented in future ACLIM model suites and could be applied to other varieties of ensemble projections retroactively to glean further insight. Our modeling framework allows evaluation of different management and policy options and, like other ecosystem models, is best viewed as a strategic rather than tactical tool for supporting decision-making (e.g. Fulton et al., 2011; Andersen et al., 2016). In this vein, our efforts to characterize uncertainty in projections should facilitate uptake of results by resource managers and policy-makers alike (Addison et al., 2013; Cheung et al., 2016).

## DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request to the corresponding author.

## AUTHOR CONTRIBUTIONS

JR conceived of the study, developed the model code, performed the simulations and analysis, and drafted the manuscript. JB, KH, KA, and ABH helped conceive the study and aided with writing. AF and ACH provided code and text describing the eastern Bering Sea fishery system. WC and AJH provided downscaled climate projections for forcing the food web model. AP provided input on simulation design and analysis and contributed to writing.

## REFERENCES


## FUNDING

This study was partially by the Alaska Climate Change Integrated Modeling project (ACLIM) and was supported by grants through the NOAA National Marine Fisheries Service Fisheries and the Environment (FATE) program, the Stock Assessment Analytical Methods (SAAM) program, the Climate Regimes & Ecosystem Productivity (CREP), the Economics and Human Dimensions Program, the NOAA Alaska Fisheries Science Center, the NOAA Integrated Ecosystem Assessment Program (IEA), and the NOAA Research Transition Acceleration Program (RTAP). This publication is partially funded by the Joint Institute for the Study of the Atmosphere and Ocean (JISAO) under NOAA Cooperative Agreement NA15OAR4320063, Contribution No. 2020-1053. We also acknowledge support from the Australian Research Council Discovery Project DP170104240 ("Rewiring Marine Foodwebs").

## ACKNOWLEDGMENTS

Valuable comments and suggestions were provided by P. Woodworth-Jefcoats, G. Whitehouse, and K. Murphy on earlier versions of the manuscript.

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2020.00124/full#supplementary-material

with CMIP5 models. Biogeosciences 10, 6225–6245. doi: 10.5194/bg-10-6225- 2013



(Theragra chalcogramma) in a changing environment. ICES J. Mar. Sci. 68, 1297–1304. doi: 10.1093/icesjms/fsr010


parameterizing multi-species biological reference points. Deep Sea Res. II 134, 350–359. doi: 10.1016/j.dsr2.2015.08.002


**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 © 2020 Reum, Blanchard, Holsman, Aydin, Hollowed, Hermann, Cheng, Faig, Haynie and Punt. 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.

# Life Cycle Dynamics of a Key Marine Species Under Multiple Stressors

Saskia A. Otto<sup>1</sup> \*, Susa Niiranen<sup>2</sup> , Thorsten Blenckner<sup>2</sup> , Maciej T. Tomczak<sup>3</sup> , Bärbel Müller-Karulis<sup>3</sup> , Gunta Rubene<sup>4</sup> and Christian Möllmann<sup>1</sup>

1 Institute of Marine Ecosystem and Fishery Science, Center for Earth System Research and Sustainability, University of Hamburg, Hamburg, Germany, <sup>2</sup> Stockholm Resilience Centre, Stockholm University, Stockholm, Sweden, <sup>3</sup> Baltic Sea Centre, Stockholm University, Stockholm, Sweden, <sup>4</sup> Department of Fish Resources Research, Institute of Food Safety, Animal Health and Environment, Riga, Latvia

Identifying key indicator species, their life cycle dynamics and the multiple driving forces they are affected by is an important step in ecosystem-based management. Similarly important is understanding how environmental changes and trophic interactions shape future trajectories of key species with potential implications for ecosystem state and service provision. We here present a statistical modeling framework to assess and quantify cumulative effects on the long-term dynamics of the copepod Pseudocalanus acuspes, a key species in the Baltic Sea. Our model integrates linear and non-linear responses to changes in life stage density, climate and predation pressure as well as stochastic processes. We use the integrated life cycle model to simulate copepod dynamics under a combination of stressor scenarios and to identify conditions under which population responses are potentially mitigated or magnified. Our novel modeling approach reliably captures the historical P. acuspes population dynamics and allows us to identify females in spring and younger copepodites in summer as stages most sensitive to direct and indirect effects of the main environmental stressors, salinity and temperature. Our model simulations furthermore demonstrate that population responses to stressors are dampened through density effects. Multiple stressor interactions were mostly additive except when acting on the same life stage. Here, negative synergistic and positive dampening effects lead to a lower total population size than expected under additive interactions. As a consequence, we found that a favorable increase of oxygen and phosphate conditions together with a reduction in predation pressure by 50% each could counteract the negative effect of a 25% decrease in salinity by only 6%. Ultimately, our simulations suggest that P. acuspes will most certainly decline under a potential freshening of the Baltic Sea and increasing temperatures, which is conditional on the extent of the assumed climate change. Also the planned nutrient reduction strategy and fishery management plan will not necessarily benefit the temporal development of P. acuspes. Moving forward, there is a growing opportunity for using population modeling in cumulative effects assessments. Our modeling framework can help here as simple tool for species with a discrete life cycle to explore stressor interactions and the safe operating space under future climate change.

Keywords: Pseudocalanus acuspes, Baltic Sea, stochastic life cycle model, polynomial regression, cumulative and cascading effects, model simulations, density dependence, vulnerable life stage

#### Edited by:

Susana Agusti, King Abdullah University of Science and Technology, Saudi Arabia

#### Reviewed by:

Rene Friedland, Joint Research Centre, Italy Samuli Korpinen, Finnish Environment Institute, Finland

> \*Correspondence: Saskia A. Otto saskia.otto@uni-hamburg.de

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 16 October 2019 Accepted: 14 April 2020 Published: 15 May 2020

#### Citation:

Otto SA, Niiranen S, Blenckner T, Tomczak MT, Müller-Karulis B, Rubene G and Möllmann C (2020) Life Cycle Dynamics of a Key Marine Species Under Multiple Stressors. Front. Mar. Sci. 7:296. doi: 10.3389/fmars.2020.00296

## INTRODUCTION

fmars-07-00296 May 14, 2020 Time: 20:6 # 2

With a growing human population and advances in technology comes an increase in interconnected threats to marine ecosystems at a rate unprecedented in human history (Halpern et al., 2019). In the recent past, the cumulative effects of a changing climate and anthropogenic pressures, such as habitat degradation, overfishing, bioinvasions or pollution, have led to increased environmental stress and ultimately to persistent changes in ecosystem structure and function. These include shifts from coral dominance to less desirable, degraded ecosystems (Jouffray et al., 2015), global fisheries collapses (Pinsky et al., 2011; Sguotti et al., 2019b) or re-organizations of entire food webs (e.g., Hare and Mantua, 2000; Daskalov et al., 2007; Möllmann et al., 2009; Conversi et al., 2010; Törnroos et al., 2019). As a consequence, research on cumulative effects and ecological futures has become increasingly important for improving planning and decisionmaking processes (Halpern and Fujita, 2013; Korpinen and Andersen, 2016; Dietze et al., 2018). This is particularly true for key species, which may exert a dominant role in the cascading of predator-prey interactions due to a unique combination of physiological performance, metabolic demands and life history strategies (Verity and Smetacek, 1996). By dominating local ecosystem functions they have a proportionally large influence on the community and its stability (Power et al., 1996; Hooper et al., 2005). The more sensitive key species are to disturbances and stressors the less resilient and more prone to regime shifts communities become (Thrush et al., 2009).

Determining key species dynamics as a response to a single stressor is generally a relatively easy task as the relationship between the species' traits and the specific stressor can often be well described by a function defining a linear, exponential, asymptotic or bell-shaped curve (Ruel and Ayres, 1999; Brown et al., 2004; Schulte, 2015). The extrinsic stressor can be any natural or human-induced change in environmental conditions that places stress on the health and functioning of organisms and ultimately a population such as oxygen deficiency or fishing pressure. Most often, stressors occur in combination with joint effects that are less predictable because of their additive, antagonistic or synergistic interactive natures (Crain et al., 2008; Piggott et al., 2015; Sguotti et al., 2019a). Especially synergistic and antagonistic interactions are considered to be the dominant mode in animal populations (Crain et al., 2008; Jackson et al., 2016) and by definition greater than the additive effects produced by multiple stressors acting in isolation (Folt et al., 1999). Moreover, stressors might act simultaneously, at a rate faster than the rate of recovery (Paine et al., 1998), or at different time periods, e.g., due to cyclic patterns of the drivers (Becks and Arndt, 2013), which then could be dampened or amplified through density-dependence effects (Hodgson et al., 2017).

An important, yet underrepresented task in assessing cumulative effects and in projecting animal populations is the consideration of life cycle dynamics (Russell et al., 2012). Life stages may differ in their environmental niches, and hence vary in their adaptive ability to changes in the stressors (Catalán et al., 2019). Stage-specific differences in thermal tolerance have been demonstrated for instance among eggs, larvae, and pupal of moth species (Kingsolver et al., 2011) or juveniles and adults of water snakes (Winne and Keck, 2005). Seasonal components in the stage-specific response to environmental changes can be additionally relevant as shown for ungulates (Coulson et al., 2001) or birds (Dybala et al., 2013). Population dynamics are even more complex as species are not only driven by their abiotic environment but strongly interact with other species in the community. Species interactions such as competition or predation can have a great influence on the spatial and temporal dynamics of species (Guisan and Thuiller, 2005; Poloczanska et al., 2008; Albouy et al., 2019), being sometimes the key driver as observed in trophic cascades (Pace et al., 1999; Frank et al., 2007). In the light of this complexity, Russell et al. (2012) have advocated a more integrated approach in ecosystem predictions, including the identification of key species, critical life stages as well as their main interactions.

One ubiquitous group with multiple distinct life stages are copepods, a species rich taxon which dominates the marine zooplankton community by numbers and biomass (Miller, 2005). Copepods have a complex life cycle developing through 11 larval stages (6 nauplii and 5 copepodite stages) to the adult stage with one to 12 generations per year in temperate and high latitude regions (Mauchline, 1998). Because of their rapid reproduction with wide dispersal ability (Cowen and Sponaugle, 2009), their high sensitivity to environmental changes (Reid et al., 1998; Richardson, 2008) and ability to integrate and transfer environmental signals over generation time (Goberville et al., 2014) they represent suitable indicators of the effect of climate change on marine ecosystems. In addition, copepods are essential for larval fish recruitment (Llopiz, 2013) and as such they could be used to evaluate planktivorous fish status (Beaugrand, 2005). But despite this undeniable indicator potential, their omnipresence, great abundance and vital role in mediating the energy flow from primary producers to consumers (Turner, 2004; Richardson, 2008) copepods have been in contrast to higher trophic levels (DeMaster et al., 2001; Lindegren et al., 2010; Barbraud et al., 2011; Cheung et al., 2011) or primary production (Litchman et al., 2006; Jang et al., 2011) underrepresented in scenario modeling. A few studies have now emerged that apply life cycle or bioclimate envelope models to project future geographic distributions of copepod species or communities (e.g., Helaouët et al., 2011; Reygondeau and Beaugrand, 2011; Maar et al., 2013; Beaugrand et al., 2019). However, projections of zooplankton dynamics using simulation models that incorporate biotic interactions and non-linearity are sparse.

In the Baltic Sea, the largest brackish water system worldwide, the study of copepod dynamics is especially challenging as most species live here at their physiological limits. This is particularly true for the calanoid copepod Pseudocalanus acuspes, which is considered a marine, glacial relict species in this system (Holmborn et al., 2011). As one of the dominant zooplankton species in the Central Baltic Sea it serves as important food source for the planktivorous Atlantic herring (Clupea harengus) and European sprat (Sprattus sprattus) (Möllmann et al., 2004) but also for Atlantic cod (Gadus morhua) larvae during the spawning season in spring (Voss et al., 2003). P. acuspes also played a key role in the regime shift from a cod- to a sprat-dominated system

in the late 1980s/early 1990s (Möllmann et al., 2009). The climateinduced decrease in salinity and increase in temperature caused a change in dominance from P. acuspes to the copepod Acartia spp. (Otto et al., 2014a), which had negative implications for cod larval survival. In addition, the change in hydrography had a direct negative effect on the reproductive success of cod. Together with stabilizing fisheries-induced feedback loops this resulted in the dominance of the planktivorous sprat (Lade et al., 2015).

In contrast to other Pseudocalanus spp. congeners, the life cycle of Baltic P. acuspes is characterized by an annual generation with a reproductive peak in spring, usually around May (Renz and Hirche, 2006). Although a small fraction of the eggcarrying population is able to reproduce year round (Renz et al., 2007), a stable progression of the stage structure is observed with older copepodite stages accumulating in autumn and overwintering before they mature in spring. In addition to the seasonal stage structure, P. acuspes shows also an ontogenetic vertical distribution with nauplii living near the surface, younger copepodites in mid-depth water and older stages in the deep water near the permanent halocline (Hansen et al., 2006; Renz and Hirche, 2006). While experimental studies on the life cycle dynamic of the Baltic populations have been challenging, statistical analysis of seasonal stage dynamics in the Central Baltic Sea revealed a complex interplay of linear density and predation effects and non-linear hydro-climatic effects (Otto et al., 2014b). In general, younger stages (i.e., nauplii and copepodites I-III) were more affected by water temperature and regional climate conditions, while older stages (i.e., copepodites IV-V and females) were mainly influenced by salinity and predation.

While statistical analyses can help identify key stressors and sensitive life stages, quantifying the effect size of changes in individual stressors and their cumulative effects at the population level requires model simulations. For populations with nonoverlapping generations such as P. acuspes in the Baltic, discrete time models are a useful tool and easy to apply (Turchin, 2003). In zooplankton ecology discrete population models are typically based on the Leslie matrix (Leslie, 1945) or a modification of it (Caswell, 2001). These approaches explicitly address the population structure (e.g., Davis, 1984b; Miller and Tande, 1993; Twombly et al., 2007; Maar et al., 2013) but model mainly vital rates instead of abundances. Such models have mainly been used to reconstruct past dynamics rather than to develop scenariobased projections. In this study, we develop and expand an alternative approach (see Otto, 2012) that has been so far used for food web simulations only (Blenckner et al., 2015; Kadin et al., 2019). Specifically, we link empirically derived statistical relationships of individual life stages with their environment into an integrative, stochastic stage-resolved population model. Using our novel modeling approach we aim to reproduce the observed long-term life cycle dynamics of P. acuspes in the Central Baltic Sea and to identify the cumulative effects and interaction types under single and multiple stressors on the population size. In our model all external predictors are considered as potential extrinsic stressors (climatic or biological) as they all can place stress on the organisms once outside the optimal range.

## MATERIALS AND METHODS

## Data

### Zooplankton Time Series

Abundance (N m−<sup>3</sup> ) data for P. acuspes are derived from a database of a zooplankton monitoring program of the Institute of Food Safety, Animal Health and Environment (BIOR) in Riga, Latvia. The sampling is conducted seasonally since the 1960s, usually in February, May, August and October, with a variable number of stations in the Eastern Gotland Basin, i.e., within the boundaries of the International Council for the Exploration of the Sea (ICES) subdivision 28 (see Otto et al., 2014b). These stations are irregularly sampled with a Juday Net as sampling gear (UNESCO, 1979), which has a mesh size of 160 µm and an opening diameter of 0.36 m. It is operated vertically and considered to quantitatively catch all copepodite stages as well as adult copepods, whereas nauplii may be slightly underestimated. Individual hauls were carried out in vertical steps, resulting in a full coverage of the water column to a maximum depth of 150 m. During analysis, abundance of nauplii (N), early copepodites 1–3 (C13), later copepodites 4– 5 (C45) as well as adult females (F) were enumerated. For our simulation study, stage-specific abundance data were averaged across stations and season (i.e., annual quarters) and covered the period 1960 – 2017, except for winter where sampling stopped after 2008. Hence, the training data for the model fitting procedures ended in 2008. Only abundant stages of P. acuspes were considered for each season and all abundance data were ln(X + 1) – transformed to reduce intrinsic mean-variance relationships (see **Supplementary Table 1**). When modeling C13 stages in spring and summer we removed the outlier year 1983 as in Otto et al. (2014b).

### Observed Environmental Data

For model construction and retrospective simulation we used observations from the period 1960 – 2017 as forcing, derived from the ICES<sup>1</sup> database (**Supplementary Table 1** and **Supplementary Figure 1**). Spring and summer temperature time series (Tempspr and Tempsum) were computed for the midwater (20–60 m) shown to affect early copepodite stages, while salinity in the deepwater layer (70–100 m) affects late copepodites (i.e., C45) and females stages from spring to autumn (Salspr - Salaut) (Otto et al., 2014b). Oxygen concentrations were also calculated for the deepwater layer and aggregated for all four seasons (Oxywin-Oxyaut). All hydrographic time series were compiled to match the zooplankton sampling period, i.e., January–February for winter, April–May for spring, July–August for summer, and October–November for autumn.

The relative role of bottom–up control was tested using phosphate concentrations (PO4win; 0–10 m) during December – February as indicator for nutrient enrichment in the system and primary production in spring (Jurgensone et al., 2011; HELCOM, 2018). We included here December for the winter mean to ensure a consistent sampling design, since Phosphate was less frequently sampled than the hydrography. Top–down control

<sup>1</sup>http://www.ices.dk

was tested using annual sprat spawning stock biomass (SSB) as proxy for predation pressure. Late copepodite stages and females of P. acuspes represent major prey items for both Baltic planktivorous fish species herring (Clupea harengus) and sprat (Sprattus sprattus) (Möllmann and Köster, 2002). However, the population size of P. acuspes is mainly controlled by sprat (Casini et al., 2008). To cover the full study period we used SSB estimates for sprat from the latest official stock assessment that date back to 1974 (ICES, 2018) and extended the time series back to 1960 using estimates by Eero (2012). The recent time series reconstruction by Eero (2012) allowed us to include sprat as direct forcing in the model instead of using an indirect predation index based on Atlantic cod (Gadus morhua) as in Otto et al. (2014b).

## Modeling and Simulation Strategy

We developed an integrative, stochastic life cycle model of P. acuspes by coupling individual, i.e., stage- and season-specific, statistical models through density effects between successive life stages (**Figure 1**). Compared to an earlier study (Otto et al., 2014b), we exchanged Generalized Additive Models (GAM; Hastie and Tibshirani, 1990) for linear and polynomial regression approximations to facilitate integration into numerical food web and end-to-end models. All statistical modeling was performed in R 3.5.0 (R Core Team, 2018) and using the 'mgcv' package, version 1.8-26 (Wood, 2006) for the GAMs. Below we present our modeling and simulation approach.

### Step 1: Stage-Specific Statistical Modeling

First, we split the zooplankton and environmental data into a training and test set covering the periods 1960–2008 and 2009–2017, respectively. These periods were chosen to have a minimum number of training observations (∼50) required to robustly estimate coefficients in our complex multiple polynomial regression model for which we had also abundance data for all stages. At the same time the test period ensured a sufficient number of observations (∼15%) for model validation (**Figure 1**, step 1). Stage- and season-specific models were then built using GAMs in which seasonal stage abundances of the training period were modeled by a combination of internal density effects and external conditions:

$$X\_{\circ \circ} = \alpha + \sum\_{i} f\_i(D^i) + \sum\_{j} g\_j(l\_{\circ}^j) + \varepsilon\_{\circ \circ} \tag{1}$$

where Xsy is the natural logarithm [ln(X + 1)] of the abundance of a particular life stage group of P. acuspes during a particular season s in training year y. D i represents a vector of density effects, i.e., ln(X + 1)-transformed population abundances of different stage groups in the same or previous season of year y or y-1. Density terms that showed a significant albeit weak effect but could not be directly modeled due to low abundances in a specific season (e.g., the effect of summer females on summer C13) were excluded. Environmental variables of the same year y (and in the case of hydrographical conditions also the same season) are summarized in the row vector [I j <sup>y</sup>]. The superscripts i and j identify the single components of both vectors. α is the intercept and εsy random noise term assumed to be normally distributed with zero mean and finite variance. f<sup>i</sup> and g<sup>j</sup> are thin plate regression spline functions describing the effect of internal and external processes, respectively. Effective degrees of freedom (edf) were restricted to a maximum of 4 to avoid overfitting. The optimal model was selected based on the AIC in a backward-selection approach (Akaike, 1973).

Once the optimal GAM and final explanatory variables were found, we replaced the GAM with polynomial regression models where all non-linear smoothing terms (i.e., edf > 1) were substituted with polynomials of different order. After identifying the most parsimonious model with the best fitting polynomial order using the AIC, we additionally applied a robust coefficient estimation based on a 5-fold cross-validation (James et al., 2013). For this, we split the training data into five subsets and in each of the five iterations we fitted the final model to 4 of the 5 subsets (i.e., to 80%). The final coefficient estimates were then obtained by averaging across the estimates from each iteration, which gave us estimates less sensitive to outliers.

For the C45sum model we applied in addition an alternative model selection strategy to better capture the observed decline in the test period. Here, we applied a backward selection process on the GAMs using the full time series, which lead to a slightly more complex model including salinity and oxygen conditions. The model was then converted into the most optimal polynomial regression model. The coefficient estimation based on the k-fold cross-validation, however, was again conducted on the training dataset to allow for an evaluation of the model performance.

## Step 2: Model Coupling and Hindcast Simulations

The fitted models were first used to simulate the observed intraand inter-annual population dynamics of P. acuspes after linking the models through the detected stage dependencies (**Figure 1**, step 2). Specifically, we initialized our simulations in year 1960 with the first observation of older copepodites in winter (C45win). Subsequent life-stages were then predicted. Predictions of C45 in autumn eventually served as input for predicting C45win the following year hence closing the annual life cycle.

We performed time series reconstructions of individual stage abundances together for the training and test periods to compare model performances. We repeated the reconstructions 1000 times adding in each iteration the same level of random noise as observed in our models (see eq. 1) to the predicted abundances by re-sampling (with replacement) the residuals from the fitted models in step 1. For all model-specific predictions in a given year the residual samples were drawn from the same original year in order to preserve the contemporaneous correlation of errors. For model validation, we averaged the abundances of each seasonal life stage over the 1000 Monte Carlo simulations and calculated 95% confidence intervals. We used these mean abundances to calculate the error rates for the training and test period. For better comparison across life stages, we computed the root mean square error normalized by the standard deviation (NRMSE). Since the inter-annual variability in P. acuspes stage abundances is relatively high and the explanatory power of most models range between 50 and 60%, the remaining noise from which we re-sampled lead to rather high uncertainty levels. **Supplementary Figure 2** demonstrates how narrow the

range of prediction uncertainty becomes if the magnitude of random noise would be a quarter of the one observed and added to our simulation.

#### Step 3: Simulation of Single Stressor Changes

To estimate the effect size of changes in a single target stressor on the total population abundance of P. acuspes, we ran 11-year simulations using the mean C45win abundance for 1960–2017 as starting value. Instead of using past observed environmental conditions as in step 2, we forced the model with constant environmental conditions, i.e., using the time series (1960–2017) mean of each stressor, except for the target stressor. Here, we increased or decreased the stressor by a constant value that amounted to a ± 10%, ± 25% or ± 50% change of the time series average (**Figure 1**, step 3). The same levels across all stressors were chosen to allow a quantitative comparison. But since the variability differed greatly between stressors (see **Supplementary Figure 3**) we also tested an increasing or decreasing trend equal to the maximum observed change from the mean). For each stressor and scenario we again performed 1000 Monte Carlo simulations and calculated average population sizes across all simulations as well as the difference between the second year and the last year in percentages. The first year served as spin up period in our simulation.

#### Step 4: Simulation of Multiple Stressor Changes and Types of Cumulative Effects

The effect size of changes in multiple stressors was estimated for 15 scenarios (see **Table 1**) using the same simulation set up as in step 3. Specifically, we estimated the combined effects under same or opposing trends of climate-driven or nutrient load-driven stressors. We then contrasted manageable stressors, i.e., nutrient


TABLE 1 | Summary of scenarios and effect sizes under multiple environmental changes.

Each scenario is a combination of increasing or decreasing trends in two or more stressors that had the same positive, negative or opposing effects on Pseudocalanus acuspes abundances. The combined effect, expressed as mean percentage change (±st. err.) of seasonal abundances over the course of 10 years, is contrasted with the sum of individual (i.e., additive) effects calculated from the single stressor simulations (shown in Supplementary Table 3). Percent changes are calculated for each of the 1000 Monte Carlo simulations and then averaged. The difference between combined and additive effect is expressed in absolute and relative values. Negative effect size values are highlighted in red. For the two climate-nutrient scenarios [Representative Concentration Pathway (RCP) 4.5 and 8.5 combined with nutrient loads according to the Baltic Sea Action Plan (BSAP)] individual effect estimates were not available.

load-driven stressors (oxygen and phosphate concentrations) and fisheries-related stressors (sprat SSB), with each other. While major trends in sprat biomass in past decades have also been linked to climatic changes and nutrient enrichment, key drivers were still fishing and particularly the release of predation pressure through cod overfishing (Möllmann et al., 2009; Eero et al., 2016). Hence, we regard in this study sprat SSB as mainly fisheriesrelated stressor, either directly or indirectly.

We tested furthermore to what extent favorable conditions of manageable stressors could mitigate a decrease in salinity conditions, a key stressor for P. acuspes. For comparability, we kept the trend magnitude of ±50% for all stressors constant, except for salinity (±25%) for which we found a much stronger individual effect on the overall abundance of P. acuspes. Ultimately, we combined changes in temperature, salinity and phosphate concentrations as projected by Saraiva et al. (2019) under the proposed nutrient load abatement strategy, i.e., the Baltic Sea Action Plan, and two greenhouse gas emission scenarios corresponding to the Representative Concentration Pathways (R), RCP 4.5 and RCP 8.5 (IPCC, 2014). Using a coupled physical-biogeochemical circulation model Saraiva and co-authors provide in their study long-term projections for temperature, salinity and biogeochemical variables for the entire Baltic Sea as well as for individual subsystems. From this study, we estimated the percent change of each stressors in the respective water depth and used these as trends in our simulations. According to Saraiva et al. (2019), oxygen concentrations under the BSAP are not expected to change, hence, we did not add a trend in these two scenarios. Similarly, we did not include a trend in sprat SSB. While sprat SSB is estimated to increase under the EU multiannual plan (ICES, 2019) the lower productivity under the BSAP might counteract such increase as projected by Niiranen et al. (2013). These two RCP4.5-BSAP and RCP8.5- BSAP scenarios are less comparable with the single stressor simulations but represent more realistic scenarios to evaluate potential trajectories of P. acuspes under expected climate change and management scenarios.

To identify the type of cumulative effects and potential interactions resulting from multiple stressor changes, we contrast the mean percent change in seasonal abundances under each scenario (1N comb sc ) with the sum of individual effects (1N ind sc ) obtained from the single stressor simulation. Following Piggott et al. (2015) and Fu et al. (2018) we consider a 1:1 ratio as additive interaction, while a ratio ≷ 1 falls into the following category:


Consequently, any interaction type where 1N comb sc < 1N ind sc , i.e., where abundances are lower under the combined effect than expected under additive effects, is posing an additional risk for


TABLE 2 | Overview of the final linear and polynomial regression models we selected for our simulations.

py, previous year. The fitted models were coupled in the presented order to simulate the observed intra- and inter-annual population dynamics of P. acuspes. Estimated coefficients and adjusted R<sup>2</sup> based on the training dataset are presented. For C45 in summer we present 2 alternative models, i.e., the most parsimonious model including only 2 predictors with poor test performance (in gray) and the more complex model we finally selected for our life cycle model.

the population. This applies to neg. antagonism and synergism as well as to positive dampened interactions (Fu et al., 2018).

We used the same approach to identify how effects of single stressors on multiple stages accumulate within the life cycle and whether effects are magnified or dampened on the total population level. Here, we contrasted the mean percent change in total seasonal abundances under each of the single stressor scenarios with the sum of mean percent change in individual seasonal stage abundances, averaged across seasons.

## RESULTS

## Key Predictors and Types of Relationships

Positive linear density effects were found for all seasonal stage abundances we tested for, which allowed us to couple the individual regression models (**Table 2**). In contrast, responses to external conditions were highly stage- as well as season-specific and in terms of the hydrography often better characterized by polynomial equations. Midwater temperature was the key predictor for younger copepodites (C13) in spring and summer where the relationship was best described by a dome-shaped response curve with an optimum around 4 and 5◦C respectively (**Supplementary Figure 4A**). In contrast, deepwater salinity had a strong positive effect on females and older copepodites (C45) from spring till autumn. This effect saturated around 10 psu for females (**Supplementary Figure 4B**). Oxygen concentrations in the deepwater layer had a positive linear effect on the nauplii in spring and, hence, recruitment success, but also a highly non-linear effect on C45 summer abundances (**Supplementary Figure 4C**). The most parsimonious model for C45 in summer in fact did not include oxygen and salinity as explanatory variables (see **Table 2**), but failed to predict the observed decline during the test period (see **Supplementary Figure 5F**). Hence, we selected a better performing but slightly more complex model, which we trained based on the full time series. When including the last 9 years from the test period (i.e., 2009–2017), salinity and oxygen revealed also as important predictors of C45 summer abundances in addition to C13 summer abundances and sprat SSB.

Combinations of winter phosphate concentrations and sprat SSB as indicator for bottom–up and top–down control, respectively, were found for both nauplii and the younger copepodites in summer. On the other hand, older copepodite abundances in winter were better explained by phosphate alone, while in summer C45 were more affected by predation pressure from sprat. Similar to the density effects, all phosphate and sprat effects, which we consider here as indirect or direct biotic stressors, were linear in our final models.

## Performance of the Coupled Life Cycle Model

Our coupled model reproduced well the past dynamics of seasonal stage abundances in the training period (**Figure 2**). The increase in P. acuspes population size in the 1970s and the following strong decline around the mid-1980s, observed mainly for older copepodites, females and nauplii in spring, is adequately captured. Model results furthermore agreed with observations of earlier copepodite stages, particularly in spring, that exhibited stationary dynamics without a clear temporal trend. For most time series the normalized root mean square error (NRMSE) of the training data is fairly similar and ranges between 0.70 and 1 indicating that the deviation of the predicted values is smaller than the standard deviation of the observations. The lowest training error rates were obtained for spring females and copepodites in summer (**Figures 2B,E,F**). A good test set performance was also obtained for the younger stages, i.e., the nauplii and C13. Here, the NRMSE for the test data was at similar levels or even lower than the NRMSE for the training data (**Figure 2D**).

We have unfortunately no data on C45 abundance since 2008 and hence could not derive information on the test set performance. But a less optimal test set performance was obtained for females in spring and C45 in summer and autumn, where the NRMSE ranged between 1.4 and 2.2 (**Figures 2B,F,H**). However, when producing an alternative hindcast based on a life cycle model that includes the more parsimonious C45 summer model with only C13 summer abundances and sprat SSB as covariates (**Table 2**), NRMSE test values are even higher for these stages (see **Supplementary Figures 5B,F,H**) but not for the others

abundances within this and the following years were predicted (from A–H). The red line represents the mean of the 1000 Monte Carlo simulations, the red shaded area the 95% confidence interval. The gray line with circles represents the observed time series. The normalized root mean square error (NRMSE) for both the training and test period is provided as measure of model performance.

(**Supplementary Figures 5A,C–E,G**). This is particularly notable for C45 in autumn where the NRMSE increases from 2.15 to 2.48 under the simpler model. The fact that modifications of a single stage- and season-specific model cause changes in the predictive performance of subsequent stages and seasons highlights the strong internal coupling and importance of cascading effects.

## Individual Stressor Effects

The analysis of net effects of changes in single stressors on the total population size of P. acuspes revealed complex dynamics that were not simply the sum of individual stage effects. We found changes in individual stressors that have linear effects on individual stages (i.e., phosphate and sprat), leading to proportional changes in the total population size, while trends in hydrographical variables caused less self-evident population responses. When keeping all external stressors constant but increasing temperature (by ±10%, ±25%, ±50%, −60%, and +65%) from the observed time series mean, we found overall changes in the total population size of P. acuspes ranging between −11 and 0.2% (see **Figure 3** and **Supplementary Table 2**). The direction of population responses was independent of the direction in temperature changes as P. acuspes showed major declines both when temperature decreased and increased by more than 50%. This net effect resembles largely the stage-specific dome-shaped effect found for younger copepodites. Changes in deepwater salinity had by far the greatest impact on the total population size in our 10-year simulation. We found decreases of 10–50% leading to a drop in the overall population size

by 7–86%, which was mainly driven by severe declines of females in spring (**Supplementary Table 3** and **Supplementary Figure 6**). On the other hand, positive salinity trends caused little changes at the stage and population level except for unobserved high increases of 50%. Here, the population starts declining again triggered by the drop in spring females. An explanation for this is the saturating salinity effect on spring females we characterized by a polynomial approximation, which tends to become also dome-shaped and predict lower abundances at both side of the salinity range when extrapolating. However, the actual scale of salinity fluctuations observed in the past decades was at about −20 to +10% around the mean. Hence, potential changes of the total population size under more realistic salinity trends are not much greater than under observed temperature changes. In contrast, changes in oxygen, phosphate and sprat SSB caused only minor changes in the population size, which hardly exceeded 5% in our simulation (**Supplementary Table 2**). In addition, uncertainty levels were comparably greater, also in terms of the direction of change (**Figure 3**). Only at pronounced increases of oxygen concentrations by 160% as seen in the past, the overall stock size will most likely show a significant increase.

We found in our single-stressor simulations that stressor effects were clearly positively and negatively dampened at the total population level (**Figure 4**). While individual life stages directly affected showed a strong response to the stressor, effects mostly cascaded only to 1 or 2 subsequent stages and eventually faded over the course of the year (**Supplementary Figure 6**). For overwintering copepodites C45 and females in spring we even found opposite trends to previous stages indicating a dampening mechanisms through density dependence being greatest during the maturation period.

## Combined Stressor Effects and Dominant Mode of Interaction

In contrast to single stressors, multiple stressors had rather an additive effect on the total abundance, although we did find indications of positive dampening and negative synergistic interactions (**Figure 4** and **Table 1**). Both types of interactions lead to lower abundances than expected and were mainly found when multiple stressors affected the same life stage, e.g., oxygen conditions, phosphate concentrations, and sprat SSB affecting all nauplii in spring. The individual response direction, i.e., whether individual stressor effects on the population size are all positive, negative, or opposing, had no impact on the interaction type.

Given the strong individual response to climatic stressors and an additive interaction, the strongest decline in the P. acuspes abundances was observed under decreasing salinity conditions and changes in temperature. This decline could be mitigated by favorable changes in manageable stressors. In our multistressor simulation, we found that a 50% increase of oxygen and phosphate conditions together with a 50% reduction in predation pressure could counteract the negative effect of a 25% decrease in salinity by 6% (see **Table 1** and **Supplementary Table 2**). This is overall less than expected under a purely additive interaction indicating a negative synergistic interaction. Under constant

FIGURE 3 | Effect size of single environmental changes. The percent change of mean seasonal abundances over the course of 10 years under eight different trend scenarios is presented for each stressor. Percent changes are calculated for each of the 1000 Monte Carlo simulations and then averaged. The error bars represent the standard deviation. The numbers next to the stressor names represent the maximum observed negative and positive deviations from the time series mean.

combined increasing or decreasing trends in two or more stressors that had the same positive, negative or opposing effects on Pseudocalanus acuspes abundances. Effects are presented as percent change of mean seasonal abundances over the course of 10 years, averaged across 1000 Monte Carlo simulations. The additive effects are calculated as the sum of mean percent change under single environmental changes (shown in Figure 3). For single stressor simulations the combined effect represents the net effect on the total population in comparison to the sum of individual seasonal stage effects. Population responses to single changes in salinity by 25 and 50% are not shown as they lied outside the selected X- and Y-range. The gray and yellow areas indicate the type of interaction present, i.e., positive and negative synergism, antagonism and dampening effects. All interactions below the diagonal lines are considered as risk zones where combined effects lead to lower abundances than expected under additive interactions.

climate conditions the same stressor changes lead to notable but slightly dampened increases in the total population size.

Current projections of long-term salinity reduction in the Baltic Sea and specifically Eastern Gotland basins under the intermediate and high RCP scenarios greatly diverge depending on the model setup but are generally below the 25% simulated here. Given a salinity decline by only 7% (RCP4.5) and 6% (RCP8.5) together with an increase in temperature and phosphate concentrations between 30-60% we still observe a decrease in the population size of P. acuspes but only up to 7% (see **Table 1**).

## DISCUSSION

## Suitability of the Modeling Approach

We here developed a novel stochastic stage-structured life cycle model for a zooplankton population that plays a key role in Baltic Sea ecosystem dynamics. Discrete stage- or agestructured population models in zooplankton ecology usually belong to the group of matrix models (Carlotti et al., 2000), which theoretically can be linear or non-linear, deterministic or stochastic (Caswell, 2001), the latter achieved by randomization of model parameters (Turchin, 2003). However, unlike our approach, in zooplankton ecology population models including stochasticity is seldomly considered (Carlotti et al., 2000) and applied matrix models are usually built on linear relationships only (e.g., Torres-Sorando et al., 2003; Twombly et al., 2007). The great advantage of our modeling approach lies hence in the explicit consideration of non-linear relationships between discrete life-stages. Furthermore, adding stochasticity in the population dynamics allowed us to provide robust uncertainty estimates associated with the direction and magnitude of stressor effects for a species with high interannual variability. Our statistical model is however not directly comparable with mechanistic stage-resolved copepod models that can be coupled to a 3-dimensional ecosystem model (see e.g., Fennel, 2001; Fennel and Neumann, 2003; Dzierzbicka-Głowacka et al., 2013) as it is constrained to the sampling design and does not resolve the spatial dynamics, both horizontal and vertical, at any desired resolution. However, being a more simple and data-driven approach that requires less prior knowledge and supercomputing power, the presented modeling framework can serve as ideal tool for management strategy evaluation and cumulative effect exploration.

## Relative Importance and Magnitude of Biotic and Abiotic Stressors

Long-term dynamics of P. acuspes in our model were the result of multiple stressor effects (Otto et al., 2014b). These acted either in concert or asynchronously and then accumulated during the life cycle. Moreover, our model simulations demonstrate that climate-driven stressors, particularly salinity, are by far the most important players at the stage as well as population level. Females in spring and younger copepodites in summer were identified as the most vulnerable stages toward climate-driven changes in hydrography, while bottom-up and top-down processes were comparably negligible for all stages.

Previous studies on long-term monitoring data showed that zooplankton species are highly sensitive to environmental conditions, which, in turn, can be influenced by large-scale hydroclimatic processes (Roemmich and McGowan, 1995; Taylor, 1995; Stephens et al., 1998; Greene and Pershing, 2000; Möllmann et al., 2000; Beare et al., 2002; Beaugrand, 2003; Molinero et al., 2005; Gislason et al., 2009). In the Baltic Sea, reproduction and population expansions of endemic as well as invading marine species are generally limited by the strong salinity gradient (Jaspers et al., 2011; Vuorinen et al., 2015). Similarly, P. acuspes, a marine arctic relict species in the Baltic Sea (Grabbert et al., 2010), is generally perceived as being driven by salinity since the brackish water may display suboptimal reproductive conditions (Möllmann et al., 2003). Also in our analysis, salinity had the strongest and most significant impact on female abundance in spring, in particular when

salinity declined. Renz and Hirche (2006) suggested a direct physiological effect of low salinity on growth and reproduction but also considered a potential role of co-occurring low oxygen levels for the recruitment success. Indeed, our study suggests that while salinity is an important predictor for abundance of females, oxygen conditions determine the offspring survival. At low salinities reproducing females might be forced to move into anoxic or low oxygen deeper waters potentially affecting egg production, the attached eggs or hatched nauplii stages (Möller et al., 2015). On the other hand, temperature has not been identified as a key stressor of Baltic P. acuspes in previous longterm studies when disregarding stage-specific responses (Dippner et al., 2000; Möllmann et al., 2000), except for the spring season (Otto et al., 2014a). Here, we show that temperature plays a key role in summer when C13 stages are most abundant. Particularly juvenile copepodites are known to depend on temperature, affecting growth rates, development times and molting rates (Vidal, 1980a,b; Hirst and Bunker, 2003; Dzierzbicka-Glowacka, 2004) and ultimately abundances as shown in our simulation for C13 in spring and even more in summer.

While studies on long-term effects of biotic interactions on zooplankton species are sparse, we here show for the first time that climate effects have the potential to impact copepod species by nearly 2 orders of magnitude in comparison to biotic interactions, which supports recent findings of low ecotrophic efficiency in deeper basins of the Central Baltic Sea (Bernreuther et al., 2018). However, we did find stage-and season-specific bottom–up and top–down effects which we tested indirectly by winter phosphate concentrations and directly by sprat SSB. The genus Pseudocalanus sp. has been generally considered as not food-limited because of their low food requirements for maximal ingestions rates, development, and maximum body size and sufficient ambient concentrations of phytoplankton (Paffenhöfer and Harris, 1976; Corkett and McLaren, 1978; Vidal, 1980a,b; Davis, 1984a; Ohman, 1985). But in the Baltic Sea, P. acuspes stages inhabiting deeper water layers are dependent on sinking algae, microzooplankton or detritus (Peters et al., 2006; Möller et al., 2012), which can strongly decrease after the degradation of the spring bloom (van Beusekom et al., 2009). This could negatively affect summer C13 abundances but also the overwintering stages. A negative top–down control is mainly exerted on C45 abundances in summer, when the estimated total predation of sprat is highest (Arrhenius and Hansson, 1993) with larval and young-of-the-year sprat feeding greatly on older P. acuspes stages (Dickmann et al., 2007). In spring, when sprat recruitment starts to peak (Karasiova, 2002), first-feeding larvae select particularly nauplii stages (Voss et al., 2003), which could explain the negative predation effect found for nauplii.

## Cumulative Effects of Multiple Stressors

Predicting cumulative effects is challenging due to potential higher order interactions, but is key to devising appropriate management and conservation strategies under multiple changing stressors. The magnitude of climate and human impacts on the total size of a population can be attributed to a combination of the stressors' effect size on individual stages, the level of exposure of the stages to the stressor, the interaction type between stressors and cascading effects between life-stages. The complex interplay of changes in abiotic and biotic ecosystem components can act multiplicatively on animal's population dynamics, e.g., leading to an increase or decrease of the population or generating stable limit cycles (Shimada and Tuda, 1996). Identifying the type of interaction when both stressors operate in the same direction is generally much more straightforward than studying opposing individual effects (Piggott et al., 2015). More so, if the response curve for a stressor is a bell-shaped curve as in the typical thermal performance curve (Schulte, 2015) found also in our study, the effect direction might even be not constant. By applying an integrated life cycle modeling framework for P. acuspes, we revealed that population responses to stressors that act on distinct life stages are mitigated through density effects, particularly during the overwintering and maturation phase. Density-dependence is generally considered as a mechanism stabilizing animal population dynamics (Sinclair and Pech, 1996; Bjørnstad and Grenfell, 2001). The importance of density-dependence in population dynamics has been controversially debated for decades (Nicholson, 1933; Andrewartha and Birch, 1954), and only recently its role in zooplankton long-term dynamics has been acknowledged. Ohman et al. (2002) and Plourde et al. (2009) observed that mortality rates of eggs and nauplii of Calanus finmarchicus depended on the number of females during the growth period. Cannibalism and food supply are generally major sources for density-dependence in zooplankton, although both mechanisms are not well investigated for P. acuspes in the Baltic Sea.

Although density dependence has been shown to govern population responses to multiple effects (Hodgson et al., 2017), we did not find strong amplifications of the positive and negative dampening effect across life stages in our multi-stressor simulations. In fact, additive interactions were more prevalent than non-additive interactions when changing two stressors suggesting that stressors are operating independently of each other. Despite growing evidence for the importance of nonadditive interactions between multiple stressors (e.g., Folt et al., 1999; Crain et al., 2008; Darling and Côté, 2008; Piggott et al., 2015; Fu et al., 2018) additive effects are still the dominant interaction mode on the individual and population level (see review in Côté et al., 2016). But when changing more than two stressors in our simulations interactions tended toward negative synergism and positive dampening effects. An explanation for this could be the cumulative effect on the same life stage. Impacts of two stressors are more likely to directly affect life stages such as in the case of temperature (affecting younger copepodites) and salinity (affecting older copepodites and females). In contrast, when changing manageable stressors altogether to mitigate the salinity decrease, nauplii in spring and older copepodite in summer were affected by two or three stressors directly leading to a slightly stronger decline than expected under additive effects. Thus, stressors acting at different life stages or asynchronously may be additive, while those acting jointly at the same stage may be synergistic. Similarly, Fu et al. (2018) found a relatively high risk of negative synergism for lower trophic levels in a multi-model simulation across regional seas. As a consequence,

increases of predation pressure under low primary productivity would cause larger than expected biomass declines.

## Implications for Management Under Climate Change

Cumulative effects assessments have become an integral part of environmental impact assessments. Stage resolved population models offer a key tool for providing a mechanistic understanding of the impacts of multiple stressors and help predict future responses under changing human impacts. Simulations based on these models have the advantage to assess a broad range of potential scenarios and to explore possibilities in ecosystems with important uncertainties rather than to predict a unique future (Carpenter, 2002). Projections have gained increasing appreciation in recent years, particularly in conservation science and the decision-making process with a focus on socio-ecological scenarios (Tansey et al., 2002; Bohensky et al., 2006; Langmead et al., 2009). Our model simulations show that P. acuspes will most certainly decline under a potential freshening of the Baltic Sea and increasing temperatures, which is conditional on the extent of the assumed climate change (BACC II Author Team, 2015). A reduction in salinity conditions is likely to cause shifts in other marine species living at their physiological limits in the Central Baltic Sea (Vuorinen et al., 2015). Recent projections of water temperature based on dynamical downscaling of global model results using regional climate models show all an increase in time as a direct consequence of the increase in air temperature, which could be nearly twice as high under the RCP8.5 scenario (Saraiva et al., 2019). In terms of salinity projections, however, the differences between RCP 4.5 and RCP 8.5 are much smaller. Instead, the greatest differences were found between the individual climate models ranging from >25% to no salinity change until the end of the century. Given this high uncertainty in future salinity conditions, any long-term projection of P. acuspes is consequently difficult to make.

In the case of this Baltic key species, the observed prevalence of additive effects suggest that stressors are largely operating independently of each other, so mitigation of any of the individual stressors will yield predictable benefits. At the same time, the effectiveness of management strategies targeting eutrophication and commercial fish stocks depends highly on the climate development, due to the strong influence of temperature and salinity on P. acuspes. To mitigate the negative effect of the projected temperature and salinity changes on this key species, oxygen and phosphate concentrations would need to increase substantially while predation pressure would need to lessen by 50%. The implementation of the nutrient load abatement strategy under the Baltic Sea Action Plan is indeed expected to counteract the decrease in mean oxygen concentrations caused by increasing water temperature, but not to levels > 50% (Saraiva et al., 2019). At the same time, phosphate concentrations will rather be reduced under this strategy. A decrease in the sprat stock size is also not expected in the near future under the EU multiannual plan for Baltic cod, herring and sprat and the fisheries exploiting those stock (ICES, 2019), rather the opposite. Similarly, long-term projections of sprat under nutrient load reductions and reduced cod fishing mortality show a consistent stock size over the next decades. As a consequence, the planned management strategies do not necessarily benefit P. acuspes. Potential cascading effects in the ecosystem structure and functioning and ultimately ecosystem services could be extensive and underpin the importance of an ecosystem approach to management.

Our modeling approach demonstrates that zooplankton has the hitherto unused potential as an integrative indicator of ecosystem change under multiple global change drivers (i.e., climate, fisheries, eutrophication). Furthermore, our results provide a general framework for investigating how population consequences will be magnified or dampened under multiple stressors. For instance, management strategies to counteract climate-related stressors might be more successful and particularly more predictable when targeting life stages not affected by climate. Moving forward, there is a growing opportunity for using population modeling in cumulative effects assessments. Our modeling framework offers a simple tool for any species with a discrete life cycle to explore stressor interactions and the safe operating space under future climate change.

## DATA AVAILABILITY STATEMENT

The datasets generated for this study are available on request to the corresponding author.

## AUTHOR CONTRIBUTIONS

SO and CM conceived the original idea. SO, TB, and CM designed the research. SO, SN, MT, BM-K, and GR prepared the data. SO analyzed the data. SO led the writing with contributions from all co-authors.

## FUNDING

This study is a contribution to the BONUS AMBER project supported by BONUS (Art 185). SO received financial support from Stockholm University's Strategic Marine Environmental Research Funds through the Baltic Ecosystem Adaptive Management Program (BEAM). SO, TB, MT, BM-K, SN, and CM were partially financed by the BONUS BLUEWEBS project supported by BONUS (Art 185); BM-K and MT received support from EU Horizon 2020 project ClimeFish (grant agreement 677039).

## ACKNOWLEDGMENTS

We are grateful to the colleagues from the Department of Fish Resources Research, Institute of Food Safety, Animal Health and Environment (BIOR) in Riga, Latvia, for maintaining this extensive long-term monitoring program and providing such detailed data on P. acuspes. We would like to thank the International Council of the Exploration of the Sea (ICES) for providing access to their comprehensive oceanographic database. Moreover, we acknowledge that this manuscript is an expansion of SO's Ph.D. thesis.

## REFERENCES

fmars-07-00296 May 14, 2020 Time: 20:6 # 13


## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2020.00296/full#supplementary-material


and anthropogenic pressure: a discontinuous regime shift in the Central Baltic Sea. Glob. Change Biol. 15, 1377–1393.


UNESCO (1979). Zooplankton Sampling. Paris, France: UNESCO Press.


environmental monitoring. Ecol. Indic. 50, 196–205. doi: 10.1016/j.ecolind. 2014.10.019


**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 © 2020 Otto, Niiranen, Blenckner, Tomczak, Müller-Karulis, Rubene and Möllmann. 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.

# Characterizing Exposure to and Sharing Knowledge of Drivers of Environmental Change in the St. Lawrence System in Canada

David Beauchesne1,2 \*, Rémi M. Daigle<sup>2</sup> , Steve Vissault <sup>3</sup> , Dominique Gravel <sup>3</sup> , Andréane Bastien<sup>4</sup> , Simon Bélanger <sup>5</sup> , Pascal Bernatchez <sup>5</sup> , Marjolaine Blais <sup>6</sup> , Hugo Bourdages <sup>6</sup> , Clément Chion<sup>7</sup> , Peter S. Galbraith<sup>6</sup> , Benjamin S. Halpern8,9 , Camille Lavoie<sup>2</sup> , Christopher W. McKindsey <sup>6</sup> , Alfonso Mucci <sup>10</sup>, Simon Pineault <sup>11</sup> , Michel Starr <sup>6</sup> , Anne-Sophie Ste-Marie<sup>4</sup> and Philippe Archambault <sup>2</sup>

#### Edited by:

Saskia Anna Otto, University of Hamburg, Germany

#### Reviewed by:

Caihong Fu, Department of Fisheries and Oceans, Canada Sean Lucey, Northeast Fisheries Science Center (NOAA), United States

> \*Correspondence: David Beauchesne

david.beauchesne@uqar.ca

#### Specialty section:

This article was submitted to Global Change and the Future Ocean, a section of the journal Frontiers in Marine Science

> Received: 01 May 2019 Accepted: 05 May 2020 Published: 24 June 2020

#### Citation:

Beauchesne D, Daigle RM, Vissault S, Gravel D, Bastien A, Bélanger S, Bernatchez P, Blais M, Bourdages H, Chion C, Galbraith PS, Halpern BS, Lavoie C, McKindsey CW, Mucci A, Pineault S, Starr M, Ste-Marie A-S and Archambault P (2020) Characterizing Exposure to and Sharing Knowledge of Drivers of Environmental Change in the St. Lawrence System in Canada. Front. Mar. Sci. 7:383. doi: 10.3389/fmars.2020.00383 1 Institut des Sciences de la mer, Université du Québec á Rimouski, Rimouski, QC, Canada, <sup>2</sup> ArcticNet, Québec Océan, Département de Biologie, Université Laval, Quebec, QC, Canada, <sup>3</sup> Département de Biologie, Université de Sherbrooke, Sherbrooke, QC, Canada, <sup>4</sup> St. Lawrence Global Observatory, Rimouski, QC, Canada, <sup>5</sup> Département de Biologie, Chimie et Géographie, Université du Québec à Rimouski, Rimouski, QC, Canada, <sup>6</sup> Fisheries and Oceans Canada, Maurice Lamontagne Institute, Mont-Joli, QC, Canada, <sup>7</sup> Département des Sciences Naturelles, Université du Québec en Outaouais, Gatineau, QC, Canada, <sup>8</sup> National Center for Ecological Analysis and Synthesis, University of California, Santa Barbara, Santa Barbara, CA, United States, <sup>9</sup> Bren School of Environmental Science and Management, University of California, Santa Barbara, Santa Barbara, CA, United States, <sup>10</sup> Department of Earth & Planetary Sciences, McGill University, Montreal, QC, Canada, <sup>11</sup> Ministère de l'Environnement et de la Lutte contre les Changements climatiques, Québec, QC, Canada

The St. Lawrence is a vast and complex socio-ecological system providing a wealth of services that sustain numerous economic sectors. This ecosystem is subject to significant human pressures that overlap and potentially interact with climate-driven environmental changes. Our objective in this paper was to systematically characterize the distribution and intensity of drivers of environmental change (hereafter, drivers) in the St. Lawrence System. We gathered data-based indicators for 22 coastal, climate, fisheries, and marine traffic drivers through collaborations, existing environmental initiatives and open data portals. We show that few areas of the St. Lawrence are free of cumulative exposure. The Estuary, Anticosti Gyre, and coastal areas are particularly exposed, especially in the vicinity of urban centers. We identified six distinct clusters with similar suites of co-occurring drivers and show that certain driver combinations are inherent to different regions of the St. Lawrence and that coastal areas are exposed to all driver types. Of particular concern are two clusters capturing most exposure hotspots and that show the convergence of contrasting cumulative exposure profiles at the head of the Laurentian Channel. Sharing knowledge of drivers emerged as a priority to facilitate future environmental assessment efforts. We thus launch eDrivers, an open knowledge platform gathering experts committed to structuring, standardizing and sharing knowledge on drivers of environmental change in support of holistic science and management. eDrivers was built on a series of guiding principles upholding existing data management and open science standards. We therefore expect it to evolve through time to address knowledge gaps and refine current driver layers. Ultimately, we believe that eDrivers represents a much needed solution that could radically influence broad scale research and management practices by increasing knowledge accessibility and interoperability.

Keywords: ocean observing systems, St. Lawrence, cumulative exposure, multiple stressors, global change

## 1. INTRODUCTION

The St. Lawrence System, formed by one of the largest estuaries in the world and a vast interior sea, is a complex socialecological system characterized by highly variable environmental conditions and oceanographic processes (El-Sabh and Silverberg, 1990; White and Johns, 1997; Dufour and Ouellet, 2007). It constitutes a unique and heterogeneous array of habitats suited for the establishment of diverse and productive ecological communities (Savenkoff et al., 2000). As a result, the St. Lawrence System has benefited the Canadian economy. It sustains a rich fisheries industry targeting more than 50 species, serves as the gateway to eastern North-America by granting access to more than 40 ports and is the most densely populated Canadian region, hosts a booming tourism industry and an expanding aquaculture production, fosters emerging activities, and boasts a yet untapped hydrocarbon potential (Beauchesne et al., 2016; Archambault et al., 2017; Schloss et al., 2017). With major investments recently made and more forthcoming in economic and infrastructure development and research (e.g., Government of Québec, 2015; RQM, 2018), an intensification of the human footprint is likely in the St. Lawrence System. Consequently, the St. Lawrence System is exposed to an increasing number of drivers of environmental change, as is observed across ecosystems globally (see Halpern et al., 2015b, 2019). We broadly define drivers of environmental change (hereafter, drivers) as any externality that affects environmental processes and disturbs natural systems. Drivers may originate from natural or human-induced biophysical processes (e.g., sea surfacewater temperature anomalies and hypoxia) or directly from anthropogenic activities (e.g., fisheries and marine pollution). The potential for complex interactions between co-occurring drivers is the largest uncertainty when studying or predicting environmental impacts (Darling and Côté, 2008; Côté et al., 2016). Multiple drivers can combine non-additively and result in effects that are greater (synergistic effect) or lower (antagonistic effect) than the sum of individual effects (Crain et al., 2008; Darling and Côté, 2008; Côté et al., 2016).

Increasing exposure and the experiences of past ecological tragedies in the St. Lawrence System such as the collapse of cod fisheries (Frank et al., 2005; Dempsey et al., 2018) and the decline of the beluga and right whale populations (Plourde et al., 2014) together urge the need to characterize the distribution, intensity and co-occurrence of drivers in the system. Research on the effects of drivers in marine environments, nonetheless, remains overwhelmingly focused on single driver assessments (O'Brien et al., 2019). Whereas, co-occurring drivers may not interact, driver co-occurrence is a requirement for interactions to exist. Knowledge of their co-distribution can therefore identify areas where driver interactions are most likely observed.

Characterizing drivers is also a necessary step for the application of holistic management approaches. Holistic approaches typically involve, but are not limited to, selecting and describing valued ecosystem components (e.g. habitats and species) and drivers (e.g., marine traffic and ocean acidification), assessing the exposure and vulnerability of valued components to drivers, selecting a proper spatio-temporal scale, monitoring, and public and stakeholder participation (Dubé and Munkittrick, 2001). Gathering environmental knowledge for holistic initiatives can, however, be a very challenging and time consuming—not to say painful—process. On one hand, there is an overwhelming and expanding wealth of data available. Such information overload may inhibit our ability to make decisions based on scientific information, promote massive duplication of effort, disproportionately appropriate research funds to certain sectors, and obscure knowledge gaps amid a sea of information (Eppler and Mengis, 2004). On the other hand, crucial data are lacking and remain largely unavailable or inaccessible for a variety of reasons, including proprietary rights, lack of organizational time, capacity and training, and, in some cases, an unwillingness to share; this curtails our ability for appropriate decision-making.

Current initiatives facilitate the data gathering process by assembling, organizing and sharing environmental knowledge, such as the Ocean Biogeographic Information System (OBIS; OBIS, 2019) for biotic data and Bio-ORACLE (Assis et al., 2018) for abiotic data. However, equivalent platforms for drivers have largely focused on single drivers (e.g., Global Fishing Watch) and platforms collating data-based indicators and knowledge on multiple drivers in a comparable and interoperable way remain conspicuously missing (but see Halpern et al., 2015a). This is in spite of integrated management and assessment approaches requiring efficient data reporting, standardized data management practices, and tools tailored to the study of the effects of multiple drivers (Dafforn et al., 2016; Stock et al., 2018).

The main goal of this study is to characterize the distribution and intensity of drivers in the St. Lawrence System. More specifically, our objectives are to: (1) identify areas of high cumulative exposure to drivers and (2) characterize areas with similar cumulative exposure profiles, i.e., areas exposed to similar suites of co-occurring drivers. An additional objective emerged while addressing the main goal of this manuscript: sharing information about the distribution and intensity of drivers of environmental change in the St.Lawrence. We achieve this through the development of an open knowledge platform, eDrivers, that was designed to facilitate the widespread availability of driver characterization for holistic assessments and management approaches. Here, we present its guiding principles and accompanying tools.

## 2. MATERIALS AND METHODS

## 2.1. St. Lawrence System

The St. Lawrence System is composed of the St. Lawrence Estuary and the Gulf of St. Lawrence (**Figure 1**). The Estuary is defined by the limit of seawater intrusion, close to Île d'Orléans, to the west and by its connection to the Gulf near Pointe-des-Monts. The surface layer is composed of freshwater flowing seaward, primarily from the Great Lakes Basin through the St. Lawrence River. Atlantic waters intrude landwards at depth into the Gulf and Estuary from Cabot Strait, but as well as from the Strait of Belle Isle (see below).

The topology of the Northern Gulf is characterized by three deep channels (250–500 m). The Laurentian Channel is the main channel connecting the Estuary to the Atlantic through Cabot Strait. The Esquiman and Anticosti channels are two secondary channels that branch off from the Laurentian Channel to the north toward the Strait of Belle Isle and the Labrador and north of Anticosti Island, respectively. The Southern Gulf hosts the Magdalen Shallows, a vast area with an average depth of ~50 m. The water column in the Gulf and St. Lawrence Estuary includes a seasonal cold intermediate layer that separates the surface and deep layers. Seasonal sea ice affects circulation in the St. Lawrence. Finally, three islands impact the physical dynamics of the St. Lawrence: the Anticosti Island to the north, the Îles de la Madeleine in the middle of the Magdalen Shallows and Prince Edward Island to the south. See Saucier et al. (2003) and Galbraith et al. (2018) for more information on the physical oceanography of the St. Lawrence.

The St. Lawrence drains over 25% of global freshwater reserves through its connection to the Great Lakes Basin, which is home to over 45 million North Americans, i.e., 15 and 30 million in Canada and the United States, respectively. The coasts of St. Lawrence System, as delimited by our study area (**Figure 1**), boast a much lower population of approximately 1 million Canadians living within 10 km of the coast, with populations mainly located in a few coastal cities in the Estuary and the Southern Gulf (Statistics-Canada, 2017).

## 2.2. Drivers

Drivers, as broadly defined in this study, are data-based indicators of environmental conditions and human activities that are often referred to as driving forces, stressors, pressures, or states in the scientific and environmental assessment literature (e.g., Kristensen, 2004; Halpern et al., 2019). Defining such categories, however, can be difficult and is often context- and ecosystem-specific (Gari et al., 2015; Dempsey et al., 2018). As

FIGURE 1 | Description of the St. Lawrence System in Eastern Canada, composed of the St. Lawrence Estuary and the Gulf of St. Lawrence. The Estuary is defined by the limit of seawater intrusion, close to Île d'Orléans, to the west and by its connection to the Gulf near Pointe-des-Monts. The Gulf is an interior sea connected to the Atlantic by Cabot Strait and the Strait of Belle Isle to the south and north of Newfoundland, respectively.

such, we refrain from articulating our work around a specific framework or imposing categories on data-based products that may change with a user's objective. Instead, we focus on available data-based indicators that contribute to evaluate the ecosystem's cumulative exposure to multiple threats.

Drivers selection was informed by a global cumulative impact assessment initiative (Halpern et al., 2008, 2015b, 2019) and available from the National Center for Ecological Analysis and Synthesis (NCEAS) online data repository (Halpern et al., 2015a), regional holistic evaluations of the state of the St. Lawrence (Dufour and Ouellet, 2007; Benoît et al., 2012), and communications with regional experts (**Table 1**). Where regional data on drivers were unavailable, available global data at a resolution adequate for the scale of the St. Lawrence System were used instead (e.g., marine pollution).

We characterized the intensity and distribution of 22 drivers (**Table 1**). Drivers incorporated in the analyses are varied in origin, i.e., from terrestrial (e.g., nutrient input) to marine (e.g., shipping), and from large scale biophysical processes (e.g., temperature anomalies) to localized anthropogenic activities (e.g., fisheries). Drivers were divided into four groups: coastal, climate, fisheries, and marine traffic (**Table 1**). All data layers and methodologies are described in the **Supplementary Materials**.

As in Halpern et al. (2019), drivers with non-normal frequency distributions were log-transformed to avoid underestimating intermediate driver values. All drivers were scaled between 0 and 1 to allow comparisons. The 99th quantile of individual driver distribution was used as the upper limit for scaling to control for extreme values that may or may not represent real observations. The St. Lawrence System was divided into a regular grid of 1 km<sup>2</sup> cells into which all drivers were integrated (**Figure S2**).

## 2.3. Cumulative Exposure

We begin by providing a simplified two-driver example that focuses on the co-occurrence of hypoxia and demersal destructive fisheries, two drivers that mostly occur in deeper St. Lawrence waters. Driver co-occurrence was evaluated spatially by summing the scaled intensity of drivers in each grid cell. The intensity at which pairs of drivers co-occur was evaluated using a two-dimensional kernel density. This example demonstrates how driver co-occurrence was evaluated and serves as a stepping stone to the integrative indicators used hereafter, i.e., cumulative exposure and cumulative hotspots (objective 1).

We evaluated cumulative exposure (EC) for each grid cell as the sum of scaled driver intensities:

$$E\_{C\_\chi} = \sum\_{i=1}^n D\_{i,\chi}$$

where x is a grid cell, i is a driver, and D is the scaled intensity of driver i. A grid cell with a high E<sup>C</sup> value is either characterized by multiple drivers at low relative intensity, a few drivers at high relative intensity, or both.

We also identified cumulative hotspots (HC)—i.e., areas where drivers co-occur at high relative intensities—as the number of drivers in each grid cell with scaled intensity contained over their respective 80th percentile:

$$H\_{\mathbb{C}\_{\mathbf{x}}} = \sum\_{i=1}^{n} \mathbb{1} \{ D\_{i, \mathbf{x}} \in P\_{\mathbb{S}0, D\_i} \}$$

where, x is a grid cell, i is a driver and D is the scaled intensity of driver i and P80,D<sup>i</sup> is the 80th percentile of driver i.

## 2.4. Cumulative Exposure Profiles 2.4.1. Clustering

We identified areas with similar cumulative exposure profiles (objective 2) using a clustering approach (Bowler et al., 2019). We used a partional k-medoids clustering algorithm, CLARA (CLustering for Large Applications; Kaufman and Rousseeuw, 1990), which was designed for large datasets. The CLARA algorithm uses the PAM (Partition Around Medoids) algorithm on a sample from the original dataset to identify a set of k objects that are representative of all other objects, i.e., medoids and that are central to the cluster they represent. The goal of the algorithm is to iteratively minimize intra-cluster dissimilarity. Iterations are compared on the basis of the average dissimilarity between cluster objects and representative medoid to select the optimal set of k medoids that minimizes average dissimilarity. We used the clustering algorithm with the Manhattan distance since this measure is less affected by extreme values (Legendre and Legendre, 2012), as is the k-medoids clustering algorithm (Kaufman and Rousseeuw, 1990). We used 100 iterations using samples of 10,000 observations (i.e., ~5% of observations) to identify clusters. Analyses were performed using the cluster R package (Maechler et al., 2018). Partitional clustering algorithms require a user-defined number of clusters. Values of k ranging from 2 to 10 were tested and validated by selecting the number of clusters that maximized the average silhouette width (Kaufman and Rousseeuw, 1990) and minimized the total within-cluster sum of squares (**Figure S4**).

### 2.4.2. Inter-cluster Dissimilarity

Differences between clusters were explored by measuring the total inter-cluster dissimilarity and the contribution of each driver to the total inter-cluster dissimilarity using a similarity percentage analysis (SIMPER) with Manhattan distance (**Figure S5**; Clarke, 1993). The Manhattan distance was again preferred for continuity with the clustering analysis and to ensure that outliers did not have a strong influence on the analysis. As the drivers dataset is large (~ 250,000 observations), we used a bootstrap procedure for the SIMPER analysis, randomly selecting 5% of each cluster to run the analysis and repeating the process over 300 iterations. We also compared the mean intensity of each driver within each cluster to better capture the inter-cluster dissimilarity.

### 2.4.3. Intra-cluster Similarity

Intra-cluster similarity was evaluated calculating the intracluster Manhattan distance and by transforming the mean contribution to distance (Mc) of each driver by 0.1/(0.1 + Mc) TABLE 1 | List of drivers available on eDrivers and used for the analyses presented in this paper. Further details on methods and data are available in the Supplementary Materials.


to obtain a similarity measure for each driver (Sd). The total similarity is the sum of all S<sup>d</sup> (**Figure S6**). As with the intercluster dissimilarity, we used a bootstrap procedure for the intra-cluster similarity, randomly selecting 25% of each cluster observation to run the analysis and repeating the process over 50 iterations. We did not use the bootstrap procedure for clusters with less than 40,000 observations since computation time was manageable.

## 3. RESULTS AND DISCUSSION

## 3.1. Cumulative Exposure

We first present the simplified hypoxia-fisheries example to demonstrate how driver co-occurrence was evaluated. Hypoxic bottom waters area mainly found at the head of the Laurentian, Anticosti, and Esquiman channels (**Figure 2A**). Demersal destructive fisheries are concentrated along the Laurentian Channel, the heads of the Anticosti and Esquiman channels and around the Îles de la Madeleine (**Figure 2B**). By combining both drivers, we observe that hypoxia and demersal destructive fisheries co-occur mostly at high relative intensity (**Figure 2D**) in the vicinity of the Anticosti Gyre and the heads of the Esquiman and Anticosti channels (**Figure 2C**); these are the areas where we might expect interactions between these drivers to be more likely.

We now focus on the integrative exposure indicators. Apart from the northeastern Gulf, cumulative exposure is ubiquitous in the St. Lawrence (**Figure 3**). Cumulative exposure is generally highest along coasts (**Figure 3**), with hotspots located in the vicinity of coastal cities (**Figure 4**). In general, offshore areas are less exposed to cumulative drivers, with the Estuary and the Anticosti Gyre being notable exceptions (**Figures 3**, **4**). This is not to say that offshore areas are free from exposure, as most of the St. Lawrence is exposed to multiple overlapping drivers (**Figures 3**, **4**). For example, the heads of the Anticosti and Esquiman channels are highly exposed to cumulative drivers (**Figure 3**).

These results are consistent with observations elsewhere in the world, where cumulative exposure conspicuously arises from and markedly intensifies close to coastal cities and at the mouth of rivers draining highly populated areas (e.g., Halpern et al., 2015b; Feist and Levin, 2016; Mach et al., 2017; Stock et al., 2018). These are areas where human activities (e.g., coastal development and shipping) and footprints (e.g., pollution runoff) are most intense (Feist and Levin, 2016), and on which is overlaid a background of natural disturbances (Micheli et al., 2016). They are also the areas in which the most dramatic increases in exposure are expected, with populations increasing more rapidly along coasts than inland (Feist and Levin, 2016).

FIGURE 2 | Simplified 2-driver example of driver co-occurrence between hypoxia and demersal destructive fisheries in the St. Lawrence. An index of hypoxia (A) was created using bottom-water dissolved oxygen between 2013 and 2017 (Blais et al., 2019). Demersal destructive fisheries (i.e., trawl and dredges) (B) intensity was evaluated from fisheries catch data collected between 2010 and 2015 used to measure annual area weighted total biomass (kg) in 1 km<sup>2</sup> grid cells (DFO, 2016b). See Supplementary Materials for more information on specific methodologies. Relative hypoxic stress and demersal destructive fisheries intensity was summed (C) to visualize their combined spatial distribution and intensity. Finally, individual density and the co-intensity of hypoxia and demersal destructive fisheries was investigated with a two-dimensional kernel analysis (D).

In the St. Lawrence, large coastal cities are mostly located along the Estuary and the southwestern Gulf, whereas the northeastern Gulf is largely uninhabited or home to small coastal communities. Certain smaller coastal communities with high cumulative exposure are characterized by large industries (e.g., Sept-Îles and Charlottetown).

As for offshore exposure, the Estuary, along with the St. Lawrence River, provide access to and serves as the primary drainage outflow of the Great Lakes Basin, which is home to over 45 million North Americans and is the most densely populated region in Canada (Statistics-Canada, 2017). Most marine traffic thus converges into the Estuary.

Whereas, we cannot ascertain that high exposure areas are the most impacted, we can safely predict that these are the areas where studying ecosystem state will be the most complex due to the uncertainty associated with driver co-occurrence, an uncertainty bound to increase rapidly with the number of co-occurring drivers (Côté et al., 2016).

## 3.2. Cumulative Exposure Profiles

While informative, the hypoxia-fisheries example focuses on a single pair of drivers and falls short of the number of drivers typically overlapping at high intensities throughout the St. Lawrence (**Figure 4**). The number of drivers overlapping in the St. Lawrence increases with cumulative exposure (**Figure S3**). Areas with high exposure such as the Estuary, the Anticosti Gyre, and the southwestern Gulf coastline (**Figures 3**, **4**) are thus areas where driver interactions are most likely and where they can arise between a host of different drivers. Identifying areas with similar cumulative exposure profiles provides a crucial tool to simplify the multi-dimensional complexity of overlapping drivers (Bowler et al., 2019). This could facilitate assessments of the state of species, habitats, and ecosystems located within or moving through areas exposed to similar suites of drivers.

Six distinct clusters were identified in the St. Lawrence (**Figure 5**, **Figures S4**, **S5**). Based on their distribution and representative drivers, clusters can be divided into three offshore and three coastal clusters (**Figure 5**, **Figures S6**, **S7**). Coastal clusters (1–3; **Figure 5**) include all types of drivers other than hypoxia; they are also the most exposed clusters, both in terms of driver overlap and intensity. Cluster 1 encompasses the coastline and is characterized by higher direct human impact (i.e., population density). Cluster 2 is differentiated from other clusters by the presence of aquaculture sites. Cluster 3 is the most exposed and has a distribution similar to the most exposed coastal hotspots (**Figure 4**). This cluster is characterized by high intensities of land-based drivers (e.g., nutrient input), demersal non-destructive high bycatch fisheries (e.g., trap fishing), climate drivers and marine traffic drivers in the vicinity of ports.

Offshore clusters (4–6; **Figure 5**) are generally characterized by high intensity climate and marine traffic drivers. Cluster 4 is differentiated by demersal non-destructive high bycatch fisheries, higher marine traffic drivers compared to cluster 5, and generally corresponds to the whole Southern Gulf. Cluster 5 is characterized by more fisheries types (i.e., demersal destructive and pelagic high bycatch), generally lower intensity marine traffic drivers, and is located almost exclusively in the Northern Gulf. Finally, cluster 6 is the most exposed offshore cluster and includes all offshore hotspots (**Figure 4**). It is characterized by high intensity hypoxia, marine traffic and pollution, as well as demersal destructive and pelagic high bycatch fisheries. This cluster corresponds primarily to the Laurentian Channel and incorporates parts of the Esquiman and Anticosti channels.

Clusters 3 and 6 capture most coastal and offshore hotspots and are the two most exposed clusters in the St. Lawrence. They offer some insight into the potential importance of considering spatial dynamics in areas intersecting multiple clusters. For example, clusters 3 and 6 meet at the mouth of the Saguenay River. This area is particularly dynamic, with deep Atlantic waters advected through estuarine circulation mixing with surface waters from the St. Lawrence and Saguenay rivers (Dufour and Ouellet, 2007). This results in the convergence of climate drivers from the bottom of the Laurentian Channel and marine traffic drivers (cluster 6) with terrestrial run-off from river outflows and direct human impacts (i.e., population density; cluster 3).

## 4. OPEN KNOWLEDGE PLATFORM: eDrivers

Sharing the knowledge acquired through the description of drivers in the St. Lawrence emerged as a priority to curtail the need to contact dozens of experts across multiple organizations and over extensive periods of time to assemble the data needed for integrated research and management. It is also a requirement to ensure that this manuscript will not quickly become an outdated snapshot of drivers distribution and intensity in the St. Lawrence System, but rather serve as a stepping stone toward an adaptive and ever-improving collection of knowledge.

As such, we are launching eDrivers, an open knowledge platform focused on sharing knowledge on the distribution and intensity of drivers and on gathering a community of experts committed to structuring, standardizing and sharing knowledge on drivers in support of science and management. In launching this initiative, our objective is to uphold the highest existing standards of data management and open science. We identified four guiding principles (section 5.1) to meet this objective and structure of the initiative (**Figure 6**).

## 4.1. Guiding Principles

## 4.1.1. Unity and Inclusiveness **Why**

Operating over such large scales in time, space, and subject matter requires a vast and diverse expertise that cannot possibly be possessed by any one individual or organization. Consequently, we envision an initiative that seeks to mobilize all individuals and entities with relevant expertise.

### **How**

By promoting, consolidating, and working with experts involved in existing and highly valuable environmental initiatives already in place in the St. Lawrence. Notable examples of environmental initiatives are the annual review of physical

(Galbraith et al., 2018), chemical, and biological (Blais et al., 2019) oceanographic conditions in the St. Lawrence, the fisheries monitoring program (DFO, 2016b), the annual groundfish and shrimp multidisciplinary survey (Bourdages et al., 2018), the characterization of benthic (Dutil et al., 2011), epipelagic and coastal (Dutil et al., 2012) habitats of the St. Lawrence, and Canada's shoreline classification (ECCC, 2018). There are also nascent efforts to share information on several human activities in the St. Lawrence such as the Marine Spatial Data Infrastructure portal, which provides data on zoning, shipping, port activities, and other human activities in Canadian waters, including the St. Lawrence system (Government of Canada, 2018).

By working with existing data portals whose objective is to share environmental data. We are thus collaborating actively with the St. Lawrence Global Observatory (SLGO) to develop the initiative and to host the platform on their web portal. The mission of SLGO is to promote and facilitate the accessibility, dissemination, and exchange of official and quality data and information on the St. Lawrence ecosystem through

FIGURE 6 | Diagram of the platform structure. Community input in the form of raw data is accessed through the St. Lawrence Global Observatory (SLGO; https://ogsl.ca/en) repository—the platform host—or through open access repositories (e.g., NASA data). The raw data are then processed through a workflow hosted on the eDrivers GitHub organization (https://github.com/ eDrivers). Data processing may be as simple as data rescaling or make use of more complex methodologies. All data is then hosted on SLGO's geoserver and accessible through their API. We developed a R package called eDrivers to access the driver layers through R and we are actively developing a second R package called eDriversEx that will include analytical tools to explore drivers data. Finally, we have developed a Shiny application, eDrivers app, that allows users to explore drivers data interactively (https://david-beauchesne. shinyapps.io/eDriversApp/). All R components of the project are hosted and available on the eDrivers GitHub organization. Iterative and adaptive processes are identified by circular arrows.

the networking of organizations and data holders to meet their needs and those of users, to improve knowledge, and to assist decision-making in areas such as public safety, climate change, transportation, resources, and biodiversity conservation. SLGO is also one of three regional associations spearheading the Canadian Integrated Ocean Observing System (CIOOS<sup>1</sup> ), which will focus on integrating oceanographic data from multiple sources to make them accessible to end-users and to enable the national coordination of ocean observing efforts by integrating isolated or inaccessible data, and by identifying gaps or duplications in observations and research efforts. We are also developing collaborations with the Portal on water knowledge<sup>2</sup> , an initiative from the Québec provincial government. The aim of this portal is to collect and share accurate, complete, and up-to-date resources on water and aquatic ecosystems to support the mandate of relevant actors and stakeholders working in water and aquatic ecosystems management.

By actively inviting, seeking, and developing collaborations as well as encouraging constructive criticism from the inception and throughout the lifetime of the platform.

By inviting external community contributions (**Figure 6**). External researchers or entities wishing to submit marine data will be able to do so through SLGO web portal. Submissions through other data portals will also be accepted either through the development of data sharing agreements or with the caveat that shared data are under an open source license and that they adhere to the platform data standards.

## 4.1.2. Findability, Accessibility, Interoperability, and Reusability

**Why**

Open knowledge has been propelled to the forefront of scientific research in an era of open, collaborative, and reproducible science. By moving toward large scale, cross-disciplinary research and management projects, there is a growing need to increase the efficiency of data discovery, access, interoperability, and analysis (Reichman et al., 2011; Wilkinson et al., 2016). Our goal is to foster efficient and functional open science by creating a fully open, transparent and replicable open knowledge platform.

### **How**

By building an infrastructure adhering to the FAIR Data Principles, which states that data and metadata must be Findable, Accessible, Interoperable, and Reusable. These (sub)principles focus on the ability of humans and machines to automatically find and (re)use data and knowledge (Wilkinson et al., 2016). As the FAIR Data Principles already exist as a unified set of principles, we adopt them as a set of guiding subprinciples to our initiative.

By making data and associated tools accessible through a variety of ways: the SLGO web portal, two R packages called eDrivers<sup>3</sup> and eDriversEx<sup>4</sup> to access the data through SLGO's API and to provide analytical tools to explore data, respectively, and a Shiny application<sup>5</sup> to explore drivers data interactively (**Figure 6**). Note that the data are currently contained within and accessible through the eDrivers R package only, as we are actively working to allow users to download selected layers from SLGO's web portal and geoserver. The functions available in eDrivers to

<sup>1</sup>https://cioos.ca

<sup>2</sup>http://www.environnement.gouv.qc.ca/eau/portail/

<sup>3</sup>https://github.com/eDrivers/eDrivers

<sup>4</sup>https://github.com/eDrivers/eDriversEx

<sup>5</sup>https://david-beauchesne.shinyapps.io/eDriversApp/

access the data have however been developed to ensure forward compatibility once the data are migrated to SLGO's geoserver.

By defining clear data and metadata standards and specifications to support the regional standardization of current and future protocols and practices and to favor interoperability with national and international initiatives like the Essential Ocean Variables (EOV) identified by the Global Ocean Observing System (GOOS). As such, we will adopt the metadata standard currently targeted for CIOOS, i.e., the North American Profile of ISO 19115:2014—Geographic information— Metadata, a schema favored for geospatial data in Canada and the United States.

By providing version control and code access to the workflows set up to generate driver layers from raw data, the R packages and the Shiny application through a GitHub organization called eDrivers<sup>6</sup> .

### 4.1.3. Adaptiveness

## **Why**

In the face of uncertainty and in an effort to address impending environmental changes, adaptive management has been identified as the chief strategy to guide efficient decisionmaking (e.g., Margules and Pressey, 2000; Keith et al., 2011; Jones, 2016; Chion et al., 2018) and has already been discussed in the context of multi-drivers and cumulative impact assessments (Halpern et al., 2015b; Beauchesne et al., 2016; Côté et al., 2016; Schloss et al., 2017). Adaptive management can only be truly achieved through a commitment to adaptive monitoring and data reporting (Margules and Pressey, 2000; Halpern et al., 2012; Lubchenco and Grorud-Colvert, 2015). We further contend that adaptive management requires the development of adaptive monitoring tools and infrastructures, which we seek to address through a continuously-evolving platform.

## **How**

By setting up mechanisms structuring cyclic reviews of platform content, for the integration of new material (e.g., data and methods) as it becomes available or accessible, and by striving to provide time-series data that are crucial to assess temporal trends and potentially early-warning signals of ecosystem change (**Figure 6**).

### 4.1.4. Recognition

## **Why**

Like peer-reviewed publications, data must also be given its due importance in scientific endeavors and thus be considered as legitimate citable products contributing to the overall scientific output of data providers (Task Group on Data Citation Standards PractOut of Cite and PractOut of Mind: The Current Sices, 2013; FORCE11, 2014). Appropriate citations should therefore be provided for all data layers used and shared by the platform.

### **How**

By adhering to the Data Citation Principles (FORCE11, 2014), which focus on citation practices that provide appropriate credit to data products.

## 4.2. Using eDrivers

Using eDrivers is simplified through the tools already in place and will be increasingly accessible as the initiative evolves and other tools are developed to ease user experience. We provide an example of the ease with which the data can be accessed and used with the eDrivers R package to reproduce **Figure 2** (**Box 1**). The code to reproduce all the analyses and figures in this manuscript is also available on GitHub<sup>7</sup> .

Box 1 | Code snippet demonstrating how to use the eDrivers R package to reproduce Figure 2.

```
# Install and load eDrivers package
devtools::install_github('eDrivers/eDrivers')
library(eDrivers)
# Load data
drivers <- fetchDrivers(drivers
= c('Hypoxia','FisheriesDD'), brick = T)[[1]]
# Transform data
drivers <- log(drivers + 1)
drivers <- drivers / cellStats(drivers, 'max')
# Visualize data and combination
plot(drivers)
plot(sum(drivers, na.rm = T))
# Identify values > 0 and not NAs
drivers$FisheriesDD[drivers$FisheriesDD < 0] <- NA
drivers$Hypoxia[drivers$Hypoxia < 0] <- NA
id0 <- !is.na(values(drivers$FisheriesDD))
      & !is.na(values(drivers$Hypoxia))
# 2D kernel for driver co-intensity
library(MASS)
coInt <- kde2d(x = values(drivers[[1]])
[id0],
                y = values(drivers[[2]])
                [id0],
                n = 500, lims = c(0, 1, 0, 1))
image(coInt, zlim = c(0,max(coInt$z)))
# Driver density distribution
plot(density(drivers$FisheriesDD[id0]))
plot(density(drivers$Hypoxia[id0]))
```
## 5. PERSPECTIVES

Understanding how ecosystem state will be affected by global change requires a comprehensive understanding of how threats are distributed and interact in space and time, which in turn hinges on appropriate data tailored to multi-driver studies (Dafforn et al., 2016; Stock et al., 2018; Bowler et al., 2019). In the St. Lawrence, we found that few areas are free from cumulative exposure and that the whole Estuary, the Anticosti Gyre, and coastal southwestern Gulf are particularly exposed to cumulative drivers, especially close to urban areas. We

<sup>6</sup>https://github.com/eDrivers

<sup>7</sup>https://github.com/eDrivers/eDriversMS

also identified six geographically distinct areas that display similar cumulative exposure profiles; these reveal that coastal areas are particularly exposed to all types of drivers and that certain driver combinations are inherent to certain regions of the St. Lawrence. These results allow us to efficiently identify areas in need of heightened scrutiny from a science and management perspective.

Through eDrivers, these observations will be iteratively improved toward an increasingly robust assessment of cumulative exposure and areas with similar cumulative exposure profiles as gaps in knowledge are addressed or approaches to describe drivers are refined. Arguably, the most meaningful benefit anticipated from eDrivers will be the gain in efficient access to comparable data-based knowledge on the exposure of ecosystems to multiple threats. This could pay quick scientific and management dividends by efficiently drawing on the knowledge and efforts of a wide range of contributors, by expanding avenues of scientific inquiry, by decreasing overall effort duplication and research costs, and by increasing research efficiency (Franzoni and Sauermann, 2014).

Critically, eDrivers will allow the scientific and governmental communities to identify key knowledge gaps that will assist in prioritizing and optimizing research efforts. Ultimately, we believe that eDrivers will operationalize evidence-based decisionmaking by streamlining data management and research, allowing science output to be available and interpretable on a time scale relevant to management (see Sutherland et al., 2004; Reichman et al., 2011). The platform will thus greatly facilitate the application of broad scale, holistic research and management approaches such as marine spatial planning, ecosystem-based management, and strategic environmental assessments (e.g., Rice, 2011; Halpern et al., 2015b; Jones, 2016).

The next step will be the inclusion of other types of knowledge to our initiative. Our focus has been on a single element required for fully operational impact assessments. Data that provide knowledge on the exposure of ecosystems to drivers are called stressor-based indicators (Dubé and Munkittrick, 2001; Dubé, 2003). These indicators efficiently identify potential local impacts and can be proactively linked to decision-making, yet assume complete knowledge of drivers and fail to diagnose impacts on valued components or non-additive effects. In contrast, effect-based indicators are direct measurements of valued components (e.g., species abundance and biodiversity) and inherently capture the effects of multiple drivers (Dubé and Munkittrick, 2001; Dubé, 2003). Whereas, effect-based indicators are considered superior to stressor-based indicators, they fail to ascribe observed effects to specific drivers. Stressorbased and effect-based indicators are, therefore, both required to diagnose causes of ecosystem change (Jones, 2016). As a collection of knowledge on stressor-based indicators, eDrivers should be weaved with other, comparable, collections of knowledge describing valued ecosystem components that can be linked to drivers and allow for a better understanding of cumulative impacts. Ultimately, interdisciplinary collections of knowledge could be weaved together through social-ecological meta-networks analyses (Dee et al., 2017). In turn, these could be used in conceptual frameworks to help to establish causal relationships between drivers and valued ecosystem components such as the DPSIR (Driving forces–Pressure– State–Impact–Response) framework (Kristensen, 2004; Gari et al., 2015). Within such frameworks, data-based indicators provided through eDrivers could be categorized as driving forces, pressures or states, depending on the objective and context of a study.

Significant effort is still needed to bring our vision to fruition. Foremost is to maintain our efforts to foster collaborations, develop platform content and identify key knowledge gaps. A fair and efficient organizational structure will be developed in order to manage eDrivers as a community and appropriate funding must be secured to continue building this community and ensure the longterm viability of the initiative, although the partnership with SLGO partly addresses this issue. We also wish to provide users with enhanced capabilities and flexibility in using the interactive tool and R package. This could include creating automatic reports and more flexibility for user-defined driver-based indicators.

Finally, terrestrial and coastal environments must be incorporated, as sources of stress within those habitats extend to the marine environments. Moreover, despite coastal areas being recognized as the most exposed to environmental threats, we continue to delineate terrestrial and marine realms, considering coastlines as an impermeable barrier. Whereas, there is a sensible rationale for this division, we must strive to eliminate it if we are to appropriately study and predict the impacts of global change (e.g., see Bowler et al., 2019).

Despite the challenges and work ahead, we are hopeful that this initiative will be very successful. Ultimately, eDrivers represents a much needed solution to address important issues in data management that could radically shift broad scale research and management practices toward efficient, adaptive and holistic ecosystem-based management in the St. Lawrence and elsewhere in the world. All it requires to be successful is for the scientific and political communities to fully commit to open knowledge, adaptive monitoring and, most of all, an integrated vision of ecosystem management.

## DATA AVAILABILITY STATEMENT

The datasets generated for this study are openly available. Requests to corresponding author should be made for access to raw data.

## AUTHOR CONTRIBUTIONS

DB, RD, DG, and PA conceived the manuscript and the underlying objectives. DB prepared/formatted the data, performed the analyses, was in charge of technical developments and lead the drafting of the manuscript. All co-authors contributed to data, analyses, and writing based on their respective expertise and contributed to the revision of the manuscript.

## ACKNOWLEDGMENTS

We thank the Fond de Recherche Québécois Nature et Technologie (FRQNT) and the Natural Science and Engineering Council of Canada (CRSNG) for financial support. This project is also supported by Québec Océan, the Quebec Centre for Biodiversity Science (QCBS), Takuvik, and the Notre Golfe networks. This research is also sponsored by the NSERC Canadian Healthy Oceans Network and its

## REFERENCES


Partners: Department of Fisheries and Oceans Canada and INREST (representing the Port of Sept-Îles and City of Sept-Îles).

## SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmars. 2020.00383/full#supplementary-material


**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 © 2020 Beauchesne, Daigle, Vissault, Gravel, Bastien, Bélanger, Bernatchez, Blais, Bourdages, Chion, Galbraith, Halpern, Lavoie, McKindsey, Mucci, Pineault, Starr, Ste-Marie and Archambault. 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.