# TOWARDS IMPROVED FORECASTING OF VOLCANIC ERUPTIONS

EDITED BY : Corentin Caudron, Lauriane Chardot, Társilo Girona, Yosuke Aoki, and Nico Fournier PUBLISHED IN : Frontiers in Earth Science

#### Frontiers eBook Copyright Statement

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

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

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

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

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

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

ISSN 1664-8714 ISBN 978-2-88963-624-2 DOI 10.3389/978-2-88963-624-2

#### 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

## TOWARDS IMPROVED FORECASTING OF VOLCANIC ERUPTIONS

Topic Editors:

Corentin Caudron, UMR5275 Institut des Sciences de la Terre (ISTERRE), France Lauriane Chardot, Earth Observatory of Singapore, Nanyang Technological University, Singapore Társilo Girona, Jet Propulsion Laboratory, California Institute of Technology, United States Yosuke Aoki, The University of Tokyo, Japan Nico Fournier, GNS Science, New Zealand

Citation: Caudron, C., Chardot, L., Girona, T., Aoki, Y., Fournier, N., eds. (2020). Towards Improved Forecasting of Volcanic Eruptions. Lausanne: Frontiers Media SA. doi: 10.3389/978-2-88963-624-2

## Table of Contents


Julien Barrière, Nicolas d'Oreye, Adrien Oth, Halldor Geirsson, Niche Mashagiro, Jeffrey B. Johnson, Benoît Smets, Sergey Samsonov and François Kervyn

*61 Analysis of the Alaska Volcano Observatory's Response Time to Volcanic Explosions-1989 to 2016*

John A. Power and Cheryl E. Cameron

*72 Foundations for Forecasting: Defining Baseline Seismicity at Fuego Volcano, Guatemala*

Kyle A. Brill, Gregory P. Waite and Gustavo Chigna


Jeremy D. Pesicek, John J. Wellik II, Stephanie G. Prejean and Sarah E. Ogburn


Jérémie Vasseur, Fabian B. Wadsworth and Donald B. Dingwell

*154 Short-Term Forecasting and Detection of Explosions During the 2016–2017 Eruption of Bogoslof Volcano, Alaska*

Michelle L. Coombs, Aaron G. Wech, Matthew M. Haney, John J. Lyons, David J. Schneider, Hans F. Schwaiger, Kristi L. Wallace, David Fee, Jeff T. Freymueller, Janet R. Schaefer and Gabrielle Tepp


Daniel Dzurisin

*228 Renewed Explosive Phreatomagmatic Activity at Poás Volcano, Costa Rica in April 2017*

Rebecca O. Salvage, Geoffroy Avard, J. Maarten de Moor, Javier F. Pacheco, Jorge Brenes Marin, Monserrat Cascante, Cyril Muller and María Martinez Cruz


John Makario Londono and Beatriz Galvis


## Editorial: Towards Improved Forecasting of Volcanic Eruptions

Corentin Caudron<sup>1</sup> \*, Lauriane Chardot <sup>2</sup> , Társilo Girona<sup>3</sup> , Yosuke Aoki <sup>4</sup> and Nico Fournier <sup>5</sup>

<sup>1</sup> Univ. Grenoble Alpes, Univ. Savoie Mont Blanc, CNRS, IRD, IFSTTAR, ISTerre, Gières, France, <sup>2</sup> Earth Observatory of Singapore, Nanyang Technological Institute, Singapore, Singapore, <sup>3</sup> Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA, United States, <sup>4</sup> Earthquake Research Institute, University of Tokyo, Tokyo, Japan, <sup>5</sup> GNS Science, Lower Hutt, New Zealand

Keywords: volcanology, monitoring, forecasting, earth science, volcano

**Editorial on the Research Topic**

#### **Towards Improved Forecasting of Volcanic Eruptions**

Forecasting volcanic eruptions and their potential impacts are primary goals in Natural Hazards research. Active volcanoes are nowadays monitored by different ground and space-based instruments providing a wealth of seismic, geodetic, and chemical data for academic volcanologists and monitoring agencies. We have better insights into volcanic systems thanks to steady improvements in research tools and data processing techniques. The integration of these data into physics-based models allows us to constrain magma migration at depth and to derive the pressure evolution inside volcanic conduits and reservoirs, with the aim of ultimately forecasting volcanic eruptions.

Yet, it remains a challenge to answer the most crucial questions when the threat of an eruption looms over us: When will it occur? What will be its style and will it switch during the course of the eruption? How long will the eruption last? Most importantly: will we have enough time to alert and evacuate population? Addressing these questions is crucial to reduce the social and economic impact of volcanic eruptions, both at the local and global scales. For example, whilst the 2014 eruption at Ontake (Japan) impacted a relatively small surface area, dozens of hikers were killed due to their proximity to the eruptive vent; in contrast, the 2010 Eyjafjallajökull eruption (Iceland) did not cause any human loss yet paralyzed the European air space for weeks due to the resulting presence of ash in the atmosphere.

Several limitations arise when approaching the questions above. For instance, short-term eruption forecasts and models that relate changes in monitoring parameters to the probability, timing, and nature of future activity still bear a very high level of uncertainty. More reliable and useful quantitative forecasting requires substantial improvements both from a monitoring data accuracy and relevance standpoint, and their interpretation and modeling; only by developing optimized and integrated monitoring networks, along with statistical methods and models, the complexity of volcanic processes and system dynamics will be better captured and deciphered.

This Research Topic investigates these questions using multi-disciplinary approaches, challenges existing models and proposes some alternatives with the aim of improving the forecasting of volcanic eruptions and support decision making of local authorities. Below is the list of the 20 contributions to this volume.

Roman and Cashman challenge classic conduit formation models by reviewing the petrology and seismicity at six well-constrained case studies of arc volcanoes. They found that initial precursory seismicity is consistently several kilometers shallower than the magma reservoir and propose a new 3-phase model. In their model, precursory magma ascent could be detected well before the onset of seismic activity by continuous monitoring of the state of stress in the mid to shallow crust.

Edited and reviewed by: Valerio Acocella, Roma Tre University, Italy

\*Correspondence: Corentin Caudron corentin.caudron@univ-smb.fr

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 01 October 2019 Accepted: 12 February 2020 Published: 25 February 2020

#### Citation:

Caudron C, Chardot L, Girona T, Aoki Y and Fournier N (2020) Editorial: Towards Improved Forecasting of Volcanic Eruptions. Front. Earth Sci. 8:45. doi: 10.3389/feart.2020.00045

**5**

De Plaen et al. explore autocorrelations of ambient seismic noise at Mt. Etna volcano. They observed seismic velocity decreases accompanying paroxysmal eruptions, suggesting a significant pressurization within the plumbing system that caused some extensional strain and subsequent crack openings.

Lesage et al. investigate the existence of detectable precursors (surface displacements and seismic velocity variations) before the large dome collapse event of Volcán de Colima in July 2015. Their results show that no significant surface deformation or velocity change could be observed.

Barrière et al. used seismic and infrasound signals generated by the Nyiragongo lava lake to quantify lava lake dynamic using a single sensor and SAR data to constrain lava lake levels. Drop of lake levels are reflected in changes in the seismo-acoustic signals, with the appearance of long period events probably resulting from deep lateral magma intrusions beneath Nyiragongo.

Salvage et al. analyse the seismo-volcanic events recorded prior to the 23 April 2017 at Poás in Costa Rica. Hindsight analysis revealed an acceleration within the dominant family of LF (low frequency) waveforms. However, no confidence could be placed in the forecast using the Failure Forecast Method (FFM), reiterating that not all accelerating trends are suitable for analysis using the FFM.

Brill et al. show a description of seismicity observed at Fuego volcano in Guatemala during January of 2012 and compute a 1- D velocity model to locate earthquakes. This work establishes a baseline of activity at this volcano.

Einarsson compiled seismic crises preceding 21 eruptions between 1973 and 2014. All eruptions were preceded by detectable precursor with lead time between 15 min and 13 days. These observations indicate that, under favorable conditions, seismic crises may be used for pre-eruption warnings.

Pesicek et al. apply the β-statistic statistical tool to seismically monitored eruptions in Alaska of various styles to investigate seismic rate. Their results confirm that seismic rate increases are common prior to larger eruptions at long dormant, "closedsystem" volcanoes, but uncommon preceding smaller eruptions at more frequently active, "open-system" volcanoes, with more mafic magmas.

Londono and Galvis compared the volume of ash emission and seismic records of eruptions in Nevado del Ruiz volcano, Colombia, between 1985 and 2017. They found that the radiated seismic energy and reduced displacements can estimate the minimum ash loading, suggesting that seismic records may be used to monitor volcanic activity.

Campion et al. recorded SO<sup>2</sup> emissions from the pit crater of Popocatépetl volcano between 2013 and 2016. The authors found that >95% of the time the volcano is releasing gas passively at average rates of 45 kg/s and with a dominant periodicity of ∼5 min. Passive degassing was interspersed by small explosions with rapid return to pre-explosive levels and by strongest explosions which were preceded by rapid decreases of SO<sup>2</sup> flux. Strong explosions are proposed to be triggered by the accumulation of gases beneath the lava dome of Popocatépetl.

Kilburn proposes a physical model relating faulting and elastic deformation as a function of loading rate, to explain precursory time series to eruptions. This elastic-brittle model is parent to the popular material Failure Forecast Method and may allow better forecasts of crustal failure.

Christophersen et al. explores the potential of Bayesian Networks (BNs) in modeling multiple data streams for eruption forecasting and volcanic hazard assessment. They show that BN modeling techniques require to accommodate continuous variables and link latent processes with observables and with eruptive patterns, and to model dynamic processes.

Vasseur et al. demonstrate that mechanical heterogeneities control the accuracy of failure forecasts and that a minimum amount of data is required to reach convergence in the forecasts. They also propose a simple method to scale laboratory results to natural systems, using rupture length and frequency of seismic events.

Dzurisin explores the trigger mechanisms of the 1980–2006 and 2004–2008 Mount St. Helens eruptions. Both eruptions could have resulted from "second boiling" during crystallization of magma long-resident in an upper crustal reservoir, rather than from injection of fresh magma.

Peltier et al. shows changes in intensity, duration and time of appearance of long-term precursors, i.e., ground displacements and seismicity, for 43 eruptions at Piton de la Fournaise (France). These findings ultimately improve the alerting chain and communication with decision-makers.

Zhan et al. assimilate InSAR and GPS data using the Ensemble Kalman Filter. They provide a hindcast and derive the triggering mechanisms of the 2009 explosive eruption of Kerinci volcano (Indonesia) using a two-source analytical model.

Guldstrand et al. show how surface deformation induced by ascending eruptive feeders can be used to forecast the eruption location through a simple geometrical analysis. Their work builds on the results of 33 scaled laboratory experiments simulating magma intrusions in a brittle crust.

Coombs et al. describe a multidisciplinary approach to forecast, rapidly detect, and characterize explosive events during the 2016–2017 eruption of Bogoslof volcano (Alaska). An effective strategy for hazard mitigation in remote areas is described.

Cameron et al. evaluate the timeliness and accuracy of eruption forecasts for 53 eruptions at 20 volcanoes in Alaska. They suggest that volcano-specific characteristics should be considered when designing monitoring programs and evaluating forecasting success.

Power and Cameron focus on the Alaska Volcano Observatory (AVO)'s response time, to identify seismic signals associated with large ash-producing volcanic explosions and initiate public warnings. While shorter response times were achieved during sequences of explosive events, longer response times are recorded for unexpected or surprise explosions.

This Research Topic covered a wide range of methodologies and approaches which provide fundamental insights into (pre-)eruptive processes. Retrospective analyses of existing data and multi-disciplinary approaches may provide fundamental advances within the next few years. An obvious avenue of research concerns the tight link between observations and numerical modeling, ideally in near-real time using simple but efficient models. We also foresee an important improvement of forecasting thanks to satellite remote sensing data (e.g., Tropomi).

#### AUTHOR CONTRIBUTIONS

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

**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 Caudron, Chardot, Girona, Aoki and Fournier. 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.

## Sequential Assimilation of Volcanic Monitoring Data to Quantify Eruption Potential: Application to Kerinci Volcano, Sumatra

#### Yan Zhan<sup>1</sup> \*, Patricia M. Gregg<sup>1</sup> , Estelle Chaussard<sup>2</sup> and Yosuke Aoki <sup>3</sup>

<sup>1</sup> Department of Geology, University of Illinois—Urbana-Champaign, Urbana, IL, United States, <sup>2</sup> Department of Geology, State University of New York at Buffalo, Buffalo, NY, United States, <sup>3</sup> Earthquake Research Institute, University of Tokyo, Tokyo, Japan

Quantifying the eruption potential of a restless volcano requires the ability to model parameters such as overpressure and calculate the host rock stress state as the system evolves. A critical challenge is developing a model-data fusion framework to take advantage of observational data and provide updates of the volcanic system through time. The Ensemble Kalman Filter (EnKF) uses a Monte Carlo approach to assimilate volcanic monitoring data and update models of volcanic unrest, providing time-varying estimates of overpressure and stress. Although the EnKF has been proven effective to forecast volcanic deformation using synthetic InSAR and GPS data, until now, it has not been applied to assimilate data from an active volcanic system. In this investigation, the EnKF is used to provide a "hindcast" of the 2009 explosive eruption of Kerinci volcano, Indonesia. A two-sources analytical model is used to simulate the surface deformation of Kerinci volcano observed by InSAR time-series data and to predict the system evolution. A deep, deflating dike-like source reproduces the subsiding signal on the flanks of the volcano, and a shallow spherical McTigue source reproduces the central uplift. EnKF predicted parameters are used in finite element models to calculate the host-rock stress state prior to the 2009 eruption. Mohr-Coulomb failure models reveal that the host rock around the shallow magma reservoir is trending toward tensile failure prior to 2009, which may be the catalyst for the 2009 eruption. Our results illustrate that the EnKF shows significant promise for future applications to forecasting the eruption potential of restless volcanoes and hind-cast the triggering mechanisms of observed eruptions.

Keywords: ensemble kalman filter, InSAR, magma storage, eruption, kerinci volcano

#### INTRODUCTION

Volcanic unrest observations including surface deformation, seismicity, gas emissions, or fumarole activity may or may not indicate that a system is trending toward eruption (Biggs et al., 2014). Understanding the dynamics of the underlying magma reservoirs is crucial for volcanologists to link volcanic unrest signals to eruption potential. A key challenge is to take full advantage of monitoring data to update and optimize dynamic models of the magma storage systems (Mogi, 1958; McTigue, 1987; Yang et al., 1988; Battaglia et al., 2003; Currenti et al., 2007; Nooner and Chadwick, 2009; Cianetti et al., 2012; Gregg et al., 2012, 2013; Newman et al., 2012; Ronchin et al., 2013; Cannavò et al., 2015; Parks et al., 2015). Model-data fusion techniques are necessary to

#### Edited by:

Zhong Lu, Southern Methodist University, United States

#### Reviewed by:

Alessandro Tibaldi, Università Degli Studi di Milano Bicocca, Italy Carolina Pagli, University of Pisa, Italy

> \*Correspondence: Yan Zhan yanzhan3@illinois.edu

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 28 September 2017 Accepted: 05 December 2017 Published: 19 December 2017

#### Citation:

Zhan Y, Gregg PM, Chaussard E and Aoki Y (2017) Sequential Assimilation of Volcanic Monitoring Data to Quantify Eruption Potential: Application to Kerinci Volcano, Sumatra. Front. Earth Sci. 5:108. doi: 10.3389/feart.2017.00108

**8**

provide statistically robust estimations of volcano evolution during periods of unrest. Classically, volcanic activity has been evaluated using static inversions (Battaglia et al., 2003; Newman et al., 2012; Parks et al., 2015), finite element model (FEM) optimizations (Hickey et al., 2015), model-data comparison (Le Mével et al., 2016). Most inversion techniques provide an important snap shot into the state of volcanic unrest, but are limited in their forecasting ability. Fewer studies use timeevolving inversions from the InSAR data, which successfully provide a quantitative model to explain the dynamics of the magma storage system (e.g., Pagli et al., 2012). However, this approach is limited to regions where SAR data is widely available and consistent and continuous acquisitions are guaranteed. Furthermore, this method requires separated steps to determine the chamber geometry and the time-dependent loading, which requires that the storage geometry is relatively stable. More recently, Kalman filter statistical data assimilation approaches such as the Extended Kalman Filter (EKF) (Schmidt, 1966; Julier et al., 2000) and unscented Kalman filter (UKF) (Fournier et al., 2009) have been used to provide temporal models of volcanic evolution. However, EKF and UKF are computationally expensive and intractable for use with FEMs.

The Monte Carlo based Ensemble Kalman Filter (EnKF) successfully circumvents linearization issues and computational costs inherent to other Kalman filter approaches (Evensen, 1994). The EnKF has been widely applied and has proven effective for multi-data stream data assimilation in hydrology, physical oceanography, and climatology (vanLeeuwen and Evensen, 1996; Allen et al., 2003; Bertino et al., 2003; Evensen, 2003; Lisaeter et al., 2007; Skjervheim et al., 2007; Wilson et al., 2010). Gregg and Pettijohn (2016) first applied the EnKF in volcanology by conducting a series of 2D elliptical magma chamber tests to assimilate synthetic InSAR (Interferometric Synthetic Aperture Radar) and/or GPS data into a thermomechanical FEM. Zhan and Gregg (2017) further establishes a 3D EnKF workflow to update a Mogi source (Mogi, 1958) using synthetic data and illustrates that the EnKF is a robust method even where data are limited. Bato et al. (2017) provides an additional synthetic test of the EnKF to track the migration of magma between two sources based on synthetic InSAR and/or GNSS data. Although these three synthetic tests indicate great potential, until now the EnKF has not been utilized to analyze volcano deformation from a natural system.

In this study, the EnKF is used to assimilate InSAR time series data (Chaussard and Amelung, 2012; Chaussard et al., 2013) from Kerinci volcano in Indonesia to investigate the surface deformation associated with the evolution of an upper crustal magma storage system leading up to its 2009 eruption. Kerinci volcano, located in Central Sumatra along the Sunda (Indonesia) Arc (**Figure 1**), has had 32 confirmed eruptions (VEI = 1–2) since 1838, and three recent eruptions, including the 2009/04/01–2009/06/19 eruption, the 2016/03/31–2016/08/09 eruption, and the 2016/11/15–2016/11/21 eruption (Global Volcanism Program, 2009), but also has more than 50,000 people living within 20 km distance around it. Previously analyzed 2007–2011 InSAR time series data from the ALOS-1 satellite (Chaussard and Amelung, 2012; Chaussard et al., 2013) provide an excellent opportunity to test the application of the EnKF in tracking the dynamics of a shallow magma storage system before and after an eruption. We apply a two-step EnKF analysis with a two-source magma storage system which models a deflating, dike-like spheroid feeding an inflating shallow, spherical magma chamber. InSAR data are assimilated as they would have become available if distributed in semi-real time following acquisition and provide model parameter updates. Finally, best-fit model parameters are used to calculate the predicted stress state of the system leading up to the 2009 eruption.

### METHODS

#### InSAR Data

SAR data were acquired between 2008/1 and 2011/11 by the Japanese Space Exploration Agency ALOS-1 satellite (Chaussard and Amelung, 2012; Chaussard et al., 2013). The displacement time series with 14 epochs is calculated using the small baseline subset (SBAS) from the InSAR data (Chaussard and Amelung, 2012; Chaussard et al., 2013) (**Figure 1**). To reduce the random atmospheric noise (Hanssen, 2001; Li et al., 2005), we filter the time series data spatially with a low pass median filter. The InSAR time-series dataset with a 15-m resolution contains more than 150,000 pixels for each time slice when we set the study area as a 6 by 6 km square centered on the volcano. It is therefore computationally prohibitive to assimilate data from the entire InSAR database. A QuadTree algorithm based on root-meansquare-error of the displacement values is applied to reduce the number of samples for each epoch of InSAR data from ∼150,000 to ∼800 (Jónsson et al., 2002; Simons et al., 2002; Lohman and Simons, 2005; Zhan and Gregg, 2017; **Figure 2A**; Figure S1), further reducing the short wavelength random atmospheric noise.

We use the EnKF data assimilation method to find the bestfit storage model for the Kerinci volcano. The EnKF uses a Markov chain of Monte Carlo (MCMC) approach to estimate the covariance matrix in the Kalman filter. The EnKF overcomes the limitations of the Kalman Filter and EKF methods, such as computational expense, storage issues, and poor performance with highly nonlinear problems (Evensen, 2009). We follow the EnKF analysis scheme described by Zhan and Gregg (2017) to obtain the magma storage models. The initial ensemble of models is constructed according to the initial guess of the parameters (**Table 1**), based on which the forecast ensemble is obtained. At time t<sup>k</sup> when new data (InSAR time series data) is available, an EnKF analysis is conducted to update the model parameters and change the trajectory of the model. The updated model parameters are then used to create a new forecast ensemble, which will be assimilated at tk+<sup>1</sup> when another epoch of InSAR time series data is available. Effectively, the EnKF provides a temporal inversion that captures the system's dynamics through time. The final output of the EnKF can be used to investigate the system state at the time of the last observation, and can be propagated forward in time to provide a system forecast. The EnKF dynamic inversion strategy has proved robust, even when the InSAR data have a topographic shadow masking the flank of the volcano edifice (Zhan and Gregg, 2017). The ensemble

parameters in this implementation of the EnKF analysis is chosen based on previous synthetic tests (Zhan and Gregg, 2017; **Table 1**).

#### Magma Storage Model

The InSAR time series reveals uplift entered at the summit of the volcano and two subsiding areas located on the NE and SW flanks. To simulate both the uplift and subsidence signals, we combine an inflating spherical source with a deflating dike-like source located at an angle beneath the chamber to form an upper crustal magma storage system (Gudmundsson, 2006; Chaussard et al., 2013) beneath Kerinci (**Figure 1**) and calculate the elastic response of the country rock (**Table 2**). The deformation pattern can also be created by other sources. For example, the twopeak pattern of the subsiding signal can be approximated using two deflating sills beneath the NE and SW side of the volcano edifice. However, it is unlikely that three magma sources would develop so close to each other while remaining separate and stable thermally. On the contrast, the single deflating dike-like source at depth feeding a shallow magma reservoir is more reasonable.

We use McTigue's analytical approach (McTigue, 1987) to produce the displacement at the center due to a shallow inflating sphere. We reproduce the two-peak subsidence with a deflating near vertical oblate spheroid (Yang et al., 1988; Dzurisin, 2006), with a high ratio of its long and short axes (∼10), acting as a dike-like source. The center of the Kerinci volcano is located on the dilatational Siulak segment of the Great Sumatra Fault (Bellier and Sébrier, 1994; Sieh and Natawidjaja, 2000). Therefore, we assume a near vertical, NW-SE striking dike-like source guided by the preexisting stress field of the Great Sumatra Fault (Pasquare and Tibaldi, 2003; Gudmundsson, 2006; Tibaldi, 2015).

#### Two-Step Data Assimilation

Tracking both the upper spherical and lower dike-like source introduces too many parameters for the EnKF to obtain unique solutions. Thus, a two-step EnKF analysis is used to track the

TABLE 1 | Parameters of the Ensemble Kalman Filter.


Parameters with \*are constant during the EnKF analysis.

two sources separately. First, the EnKF estimates the subsidence generated by a deflating dike using the Yang et al.'s model (Yang et al., 1988) (**Figure 2B**; Figure S2). During this step, the uplift signal is masked from the InSAR data, and is treated as



missing data. The Yang et al.'s model (Yang et al., 1988) requires eight independent parameters beside the Young's Modulus and Poisson's ratio, including the x, y, and z coordinates, long, and short axis, plunging direction and dipping angle, the overpressure of the spheroid. As many parameters may cause strong non-uniqueness during the EnKF analysis (Zhan and Gregg, 2017), we assume that the location of the center of the dike is 5 km beneath the summit of the volcano, and it is striking NW-SE aligned with the Great Sumatran Fault system (**Table 1**). The residual displacement is calculated by subtracting the predicted subsidence from the corresponding InSAR data time step (**Figure 2D**; Figure S3). At the second step, the EnKF analysis (initial parameters listed in **Table 1**) tracks the inflating spherical source using the McTigue's model (**Figure 2E**; Figure S4) to reproduce the uplift signal in the residual displacement obtained from Step 1. A combination of both the deeper deflating dike-like spheroid model and the shallower inflating spherical model produces a modeled displacement, which closely matches the observed pattern of central uplift and flank subsidence (**Figure 2C**; Figure S5). Finally, the two models are combined to produce the total displacement. The misfit is the difference between the modeled displacement from the combined model and the measured displacement from the InSAR time series (**Figure 2F**; Figure S6).

#### Stress and Coulomb Failure Calculation

To calculate the stress field of the country rock around the magma storage, we follow the benchmarked strategy of Zhan and Gregg (2017). Elastic FEMs are established with the parameters estimated by the two-step EnKF analysis and then solved for by COMSOL5.2. The maximum and minimum principle stresses at the top of the chamber are calculated for failure determination. The application of the Coulomb failure criterion follows the same strategy as previous studies (Grosfils, 2007; Gregg et al., 2012; **Table 2**).

#### RESULTS

#### Volcanic Deformation

Down-sampled InSAR time series data (**Figures 3A,D**) reveal two deformation signals at Kerinci Volcano, an uplifting signal centered on the summit and a subsiding signal on the NE and SW flanks. Both signals are consistent in temporal and spatial domains, suggesting they are not associated with atmospheric delay and should be treated as deformation (**Figures 3**, **4**). Prior to the April 2009 eruption, the volcano experienced a continuous uplift at a maximum rate of ∼4 cm/yr (**Figure 4a**), while the NE and SW flanks subsided at a much lower rate of <1 cm/yr (**Figures 4b,c**). At the time of the eruption, the center and flanks of the volcano went through a rapid subsidence (**Figures 3A,D**, **4**), reflecting withdrawal of magma from the storage system. After the eruption, central uplift recommenced while deformation of the NE and SW flanks ceased (**Figure 4**). The deformation centered on the summit has a short wavelength, indicating a shallow source, while a deep source is more likely to create a long wavelength subsidence deformation signal. The symmetrical shape of the central uplift strongly suggests an inflating spherical source, while the two-peak pattern of the subsidence suggests a deflating dike-liked source.

A two-step data assimilation approach (**Figure 2**) is implemented to track the surface deformation created by a shallow, inflating spherical source (McTigue, 1987) and a deeper, deflating dike-like (oblate spheroid) source (Yang et al., 1988). The two-sources model reproduces both the observed central uplift signal and the subsidence signal on the flank (**Figure 3E**). Model errors are <1 cm in most regions and are <0.5 cm at the center of the volcano (**Figures 3C,F**). A comparison of the deformation time-series and the model predictions confirms that the two-step model is able to track the observed deformation within uncertainty (**Figure 4**). We further calculate the L<sup>2</sup> norm of the displacement to illustrate the total misfit between the forecast models and the data. The L<sup>2</sup> norm error estimation is the sum of the square of the differences between the model values and the measurement values, which also considers the size of the quad created during QuadTree down-sampling. Normalized L<sup>2</sup> norms of both the spherical source and tilted dike-like source decrease through time as more InSAR data are assimilated (**Figure 5**). The L<sup>2</sup> norm of the spherical source is

FIGURE 3 | Comparison between the QuadTree down-sampled InSAR time series (A,D) and the EnKF data assimilation results (B,E), before (top row) and after (bottom row) the 2009 eruption. (B,E) show the best fit two-sources model obtained from the EnKF data assimilation and (C,F) show the misfit between the EnKF prediction and the InSAR data.

overall significantly lower than for the tilted dike-like source, indicating that the spherical chamber model reproduces more accurately the uplift at the center of the volcano compared to the flank subsidence. The L<sup>2</sup> norms of the spherical source decreases rapidly after several InSAR assimilations and become static until the eruption (**Figure 5B**). The low displacement errors (**Figures 3**, **4**) and observed convergence in the L<sup>2</sup> norms (**Figure 5**) suggest that the two-sources combination of a shallower inflating chamber and a deeper deflating dike-like source is a good representation of the storage system at Kerinci and their volume changes due to the magma transport explains the deformation associated with the 2009 eruption.

#### Magma Source Parameters

The EnKF provides evolving estimates of the model parameters for the dike-like spheroid and shallow spherical chamber as new

the L2 norms of the 200 models in the ensembles.

SAR observations are assimilated (**Figure 6**; the detailed values of the parameter estimation are listed in the Supplementary Tables 1, 2). We focus on EnKF's predictions of the evolution of over-pressurization and volume of the shallow chamber to investigate eruption precursors. The negative overpressure of the dike-like source is consistent with deflation of this deep source, but its rapid change is suspicious (**Figure 6e**). It is difficult to constrain the depth of the center of the dike-like spheroid with the InSAR subsiding signal alone. To constrain the depth of the dike, we conduct a series of tests to model the deflating signal. Results indicate that a dike deeper than 7 km cannot produce the deformation signal revealed by the InSAR data (Figure S7). Alternatively, a dike shallower than 3 km may overlap with the inflating magma chamber. As such, the depth to the center of the dike should be in a range of 3–7 km. Therefore, we assume the dike center is at a depth of 5 km. Furthermore, a 2-km uncertainty in the depth will not affect the result significantly for a near vertical dike. Due to the uncertainty in the deeper deflating source, this study instead focuses on the host rock stress evolution surrounding the shallow inflation source.

The second step of the data assimilation estimates the evolution of the shallow inflating source. The model converges after two to three time steps (05/2008–07/2008), when the standard deviation of the parameters and the L<sup>2</sup> norms of the model significantly decrease (**Figures 5B**, **6**). After the model parameters stabilize at July 2008, the EnKF estimates that the shallow inflating source shoaled from 4.43 (±0.19) km to 3.99 (±0.05) km (depth-to-center) beneath the summit prior to the eruption, and after the eruption the shallow source migrated northward 0.46 (±0.2) km and shoaled to ∼1.12 (±0.1) km depth (**Figures 6a–c**). While this outcome provides a robust estimation of the migration of the pressure source, the variation through the time likely indicates magma migration in the magma storage system (through dikes or conduits), rather than the movement of a void chamber. However, the spherical chamber model provides a first-order approximation of the deformation source location through time. The EnKF predicts that the radius of the shallow

FIGURE 6 | EnKF parameter estimations for the spherical (red lines) and dike-shaped (blue lines) sources. The EnKF predictions start to converge after several three epochs of InSAR time series data are assimilated. The x- and y-location in (a) and (b) are the horizontal distances between the deformation sources and the center of the volcano. (c) Is the depth to the center of the source. (d,e,f) Are the radius, overpressure and volume change of the source. The colored solid lines and shaded areas indicate the ensemble means and standard deviations respectively. The colored circle symbols indicate the estimated parameters of the best-fit model from each ensemble at each time step. The black dashed lines indicate time steps when InSAR data were assimilated. Notice the upper and lower part of (e) have different scales of the Y-axis.

chamber is 2.27 (±0.01) km (**Figure 6d**), which is likely too large given its shallow depth (see Supplemental Tables). However, trade-offs exists between overpressure and radius due to the nonuniqueness of the analytical model (Zhan and Gregg, 2017). To account for the non-uniqueness of the model, overpressure and radius are combined to calculate the volume evolution (**Figure 6f**). The volume of the shallow chamber increases at a rate of 5.33 (±0.10) ×10<sup>5</sup> (±0.04) m<sup>3</sup> /yr (**Table 3**) throughout the pre-eruptive time period, reaching its maximum just prior to the eruption. The volume of the dike-like source decreases [1.22 (±0.04) ×10<sup>5</sup> m<sup>3</sup> /yr] during the same time period (**Figure 6f**). During the eruption period, both sources experience substantial deflation resulting in a strong subsiding signal observed in the InSAR data (**Figure 1**). After the eruption, the deeper source returns to a steady state, while the shallower chamber continues to inflate at a much smaller rate [0.57 (±0.15) ×10<sup>5</sup> m<sup>3</sup> /yr] than prior to the eruption. Given that the shape of the post-eruption inflation curve does not mimic a typical viscoelastic roll-off, it likely indicates a slow recharge stage of the next eruption cycle which culminated in the 2016 eruption.

#### DISCUSSION

#### Magmatic System Evolution at Kerinci

Based on the converged parameter estimation and the displacement agreement between the EnKF predictions and InSAR observations, we propose that the upper crustal magma transport-storage system of Kerinci is comprised of a shallow, spherical chamber at a depth of ∼4 km connected by a dike system below to a possible lower crustal reservoir (**Figures 1**, **8a,b**). The dike-like source may have developed aligned with the Great Sumatran Fault (Pasquare and Tibaldi, 2003; Gudmundsson, 2006; Tibaldi, 2015). Alternatively, other source combinations can also create the displacement pattern shown by the InSAR, such as inflating and deflating sills, and connected chambers. However, to model the two-peak pattern of the subsiding signal without an inclined feeder dike, at least two deflating chambers or sills would need to flank either side of the central inflating source, which is unlikely. Additionally, the preexisting faults beneath the volcano may provide an ideal path for magma transport (Tibaldi, 2015).



The coincident volume changes of the dike-like source and the chamber imply magma migration between these sources. Prior to the 2009 eruption, the volume of the shallow chamber continuously increased indicating possible magma injection (Mogi, 1958; Lister and Kerr, 1991) and/or differentiation (Tait et al., 1989). In the meantime, the volume of the dike-like spheroid decreased (**Figure 6f**), indicating that the dike-like source may only act as a pathway for magma to ascent from a lower reservoir (e.g., MASH zone; Hildreth and Moorbath, 1988), as suggested by seismic tomography (Koulakov et al., 2007; Collings et al., 2012). Following the eruption, the volumes of both the dike-like spheroid and the chamber decrease drastically, but because of the lack of data in the first months after the eruption, we cannot determine how fast this subsidence occurred. The volume loss is most likely related to the erupting steam-, ash-, and cinder-bearing plumes recorded in April 2009 (Global Volcanism Program, 2009). The total volume loss of the two-sources system is ∼1.6 × 10<sup>6</sup> m<sup>3</sup> , which is consistent with volume estimates for the April 2009 eruption (VEI = 1) (Global Volcanism Program, 2009).

Although the misfits between the surface displacement model and the InSAR data are small (**Figure 3**), some locations show higher misfits (up to 1.5 cm), especially in the subsiding areas to the SW. The minimal misfit at the volcano center confirms that the model accurately captures the parameters of the shallower spherical chamber. On the other hand, the misfit in the subsiding areas suggests a bias that could be due to lithospheric heterogenesis (Zhan et al., 2016) or could be associated with atmospheric noise in the data. We focus our discussion on the dynamics of the shallower chamber, as it is better constrained and the eruption is largely controlled by overpressure and failure of the rock surrounding it.

#### Overpressure and Stress Evolution Prior to the 2009 Eruption

A central paradigm in volcanology is that eruption is triggered when the overpressure within an expanding magma chamber exceeds the strength of the surrounding rock. Unfortunately, analytical models such as Mogi (1958) and McTigue (1987) are limited in their ability to provide reliable overpressure predictions, because the calculations are inherently non-unique. As previously discussed, this non-uniqueness makes it difficult for the EnKF to reconcile estimations of radius and overpressure. The magma system parameters estimated by the EnKF are used in combination with a series of FEMs with different combinations of radius and overpressure to predict the stress field of the country rock prior to and directly following the 2009 eruption. Calculations of stress evolution are focused at the top of the magma chamber where confining pressures are lowest and tensile failure is most likely (Grosfils, 2007). Additionally, the 2009 eruption fed from a central vent further indicating failure at the top of the magma reservoir.

We utilize the benchmarked COMSOL FEM approach for a pressurized sphere in 3D (Del Negro et al., 2009; Gregg et al., 2012; Zhan and Gregg, 2017) to perform a series of tests for magma chamber radius values of 100–2,500 m and their corresponding overpressures (**Figure 7**). Of particular interest is whether the magma chamber is in a stable configuration or in a state of tensile failure, potentially indicating imminent eruption.

**Figure 7** illustrates the tradeoff between overpressure and radius required to produce the same surface deformation given the optimal EnKF magma chamber depth-to-center estimation. Model configurations that result in either tensile failure or Mohr-Coulomb failure are shown. The most striking outcome of these tests is the clear correlation between chamber radius and failure. As the radius increases, the minimum principal stresses also increase, while the maximum shear stresses are significantly reduced due to decreasing overpressure (**Figures 7**, **8c,d**). This indicates that systems with smaller magma chamber radii are more likely to fail, given the same volume change. This finding has been previously indicated by other researchers (Grosfils, 2007; Gregg et al., 2012) and further indicates the need for an independent assessment of magma reservoir size.

The predicted overpressure prior to the eruption is at least two times higher than during and after the eruption (**Figure 7**) due to the depressurization of the system during the eruption. The models predict that a magma chamber with a radius of 500 m will experience tensile failure (**Figure 8c**), potentially

leading to an eruption. The model also predicts no tensile failure after the eruption if the chamber size is not greatly reduced (**Figure 8d**); the total estimated volume loss of the chamber is <1%. Similarly, Mohr-Coulomb failure calculated in the host rock prior to eruption is more extensive than after the eruption (**Figures 7**, **8**); however, while failure is predicted in both instances, the orientation and mode of failure may not be optimal for catalyzing eruption (Grosfils, 2007). Due to the nonuniqueness issue, radius estimates may be unreasonably large (**Figure 6**) and an analysis of the system's stress state assuming a variety of radii-overpressure combinations is necessary to investigate the possibility of an eruption. Future work using data assimilation with displacement and seismicity data may provide stronger constraints on the stress evolution, helping to decipher the dynamics of the magma storage system.

The L<sup>2</sup> norm evolution provides additional insights to aid eruption prediction. Since the EnKF analysis updates the model based on the previous time steps, a sudden increase of the L<sup>2</sup> norm (**Figure 5B**) means that the pre-eruption model is no longer able to reproduce the observed deformation, suggesting a sudden change of the magma storage system. Volume change due to magma withdrawal, opening of fractures and dikes (Lister and Kerr, 1991), and alterations of country rock's rheology due to temperature evolution (Annen and Sparks, 2002) could explain this change. Some of these transitions may occur just prior to eruptions and are captured by InSAR and/or GPS. Therefore, the L<sup>2</sup> norms provides useful information for characterization of unrest.

#### Near Real-Time Data Assimilation with InSAR Data

The advantage of SAR observations is that they offer a high spatial resolution, which provides a broad view of the region surrounding the magma system. The EnKF analysis is able to efficiently track surface deformation from the down-sampled InSAR time series of Kerinci (**Figure 3**). Prior to the 2009 eruption, the InSAR-ALOS time-series repeat interval is 46 days, providing observations of continuous uplift. The models become unconstrained just prior to and immediately following the eruption (gray shaded area in **Figure 5**) due to the gap in acquisitions. As EnKF is able to update deformation models in near real-time, getting access to SAR data in near-real time could lead to usage of these data to provide early warning of eruption. Additionally, higher temporal repeatability of the SAR systems could lead to improved constraints of the magmatic systems worldwide and of their temporal evolution.

In this EnKF study, 200 models are used in the forecasting ensembles adding up to more than 1,000 iterations. However, the computational expense is <3 min to finish the calculation on a workstation (3.2 GHz Intel Core i5). Although the EnKF is slightly longer than other inversion techniques (e.g., Pagli et al., 2012), it provides huge flexibility for incorporating a wide range of observations from deformation to seismicity, and from heat flow to geochemistry. The primary limitation of this study is the analytical models used. The analytical approach is ideal for decreasing the computational expense of calculating a large population of ensembles; however the models are oversimplified. In the future, more realistic physics-based models and FEMs will take the place of the analytical models to allow researchers to explore more realistic deformation based on other geophysical observations from tomography, gravity, and/or magnetotellurics. Coupling solid mechanics with the fluid dynamics (Le Mével et al., 2016), the evolution of the magma storage systems will be closely related to the magma flux inferred from geological records, instead of the enigmatic and oversimplified overpressure. In the case of a finite element approach, the computational expense is far more significant and a supercomputer is necessary to conduct the Monte Carlo suites for the data assimilation. Future efforts are necessary to optimize the EnKF approach for the use of more sophisticated and computationally expensive models (Gregg and Pettijohn, 2016).

#### CONCLUSION

A two-step EnKF data assimilation provides a shallow chamber connect to a deep dike-like source as the most likely model to explain the surface displacement around the 2009 eruption of the Kerinci volcano revealed by the InSAR data. The Yang

#### REFERENCES


et al. (1988) model is used to mimic a deep, deflating dike-like source, which can explain the subsiding signal on the flanks of the volcano. At the meantime, a shallow spherical source (McTigue, 1987) is built to reproduce the central uplift. The parameters with highest likelihood are applied to reconstruct the stresses around the magma chamber utilizing a benchmarked FEM. The stress model suggests that the shallow magma reservoir is most likely to fail prior to 2009, which may explain the eruption. Our results illustrate the great potential of the EnKF data assimilation as a technique to explore the dynamic evolution of the magma storage system, giving insight into the eruption forecasting of restless volcanoes.

#### AUTHOR CONTRIBUTIONS

YZ and PG conceived the study and YZ wrote the paper with input from all authors. EC and YA contributed to the InSAR data set.

#### ACKNOWLEDGMENTS

We would like to acknowledge helpful discussions with Dr. F. Amelung, J. Albright, Dr. J. C. Pettijohn, Dr. J. Freymeuller, Dr. Z. Lu, Dr. L. Liu, Dr. J. Biggs, Dr. G. Hou and the UIUC Dynamics Group. We would also like to thank Dr. V. Acocella, Dr. Z. Lu, Dr. A. Tibaldi, and Dr. C. Pagli for their comments which greatly improved our manuscript. Development of Ensemble Kalman Filter approach for modeling active volcanic unrests using InSAR data is funded by NASA (13-ESI13-0034).

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart. 2017.00108/full#supplementary-material


traveltime tomography. J. Geophys. Res. Solid Earth 117, B01312. doi: 10.1029/2011JB008469


2007–2014 unrest episode at Laguna del Maule volcanic field, Chile. J. Geophys. Res. Solid Earth 121, 6092–6108. doi: 10.1002/2016JB013066


Zhan, Y., Hou, G., Kusky, T., and Gregg, P. M. (2016). Stress development in heterogenetic lithosphere: insights into earthquake processes in the New Madrid Seismic Zone. Tectonophysics 671, 56–62. doi: 10.1016/j.tecto.2016. 01.016

**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 © 2017 Zhan, Gregg, Chaussard and Aoki. 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) or licensor 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.

## Experimental Constraints on Forecasting the Location of Volcanic Eruptions from Pre-eruptive Surface Deformation

#### Frank Guldstrand<sup>1</sup> \*, Olivier Galland<sup>1</sup> , Erwan Hallot <sup>2</sup> and Steffi Burchardt 3,4

<sup>1</sup> Physics of Geological Processes (PGP), The NJORD Centre, Department of Geosciences, University of Oslo, Oslo, Norway, <sup>2</sup> Univ Rennes, CNRS, Géosciences Rennes - UMR 6118, Rennes, France, <sup>3</sup> Department of Earth Sciences, Uppsala University, Uppsala, Sweden, <sup>4</sup> Centre for Natural Hazards and Disaster Science, Department of Earth Sciences, Uppsala University, Uppsala, Sweden

Volcanic eruptions pose a threat to lives and property when volcano flanks and surroundings are densely populated. The local impact of an eruption depends firstly on its location, whether it occurs near a volcano summit, or down on the flanks. Then forecasting, with a defined accuracy, the location of a potential, imminent eruption would significantly improve the assessment and mitigation of volcanic hazards. Currently, the conventional volcano monitoring methods based on the analysis of surface deformation assesses whether a volcano may erupt but are not implemented to locate imminent eruptions in real time. Here we show how surface deformation induced by ascending eruptive feeders can be used to forecast the eruption location through a simple geometrical analysis. Our analysis builds on the results of 33 scaled laboratory experiments simulating the emplacement of viscous magma intrusions in a brittle, cohesive Coulomb crust under lithostatic stress conditions. The intrusion-induced surface deformation was systematically monitored at high spatial and temporal resolution. In all the experiments, surface deformation preceding the eruptions resulted in systematic uplift, regardless of the intrusion shape. The analysis of the surface deformation patterns leads to the definition of a vector between the center of the uplifted area and the point of maximum uplift, which systematically acted as a precursor to the eruption's location. The temporal evolution of this vector indicated the direction in which the subsequent eruption would occur and ultimately the location itself, irrespective of the feeder shapes. Our findings represent a new approach on how surface deformation on active volcanoes that are not in active rifts could be analysed and used prior to an eruption with a real potential to improve hazard mitigation.

#### Keywords: surface deformation, laboratory modeling, cone sheets, dykes, eruption forecasting

### KEY POINTS


#### Edited by:

Nicolas Fournier, GNS Science, New Zealand

#### Reviewed by:

Alessandro Bonforte, Istituto Nazionale di Geofisica e Vulcanologia, Italy Luca Caricchi, Université de Genève, Switzerland

> \*Correspondence: Frank Guldstrand f.b.b.guldstrand@geo.uio.no

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 09 October 2017 Accepted: 22 January 2018 Published: 20 February 2018

#### Citation:

Guldstrand F, Galland O, Hallot E and Burchardt S (2018) Experimental Constraints on Forecasting the Location of Volcanic Eruptions from Pre-eruptive Surface Deformation . Front. Earth Sci. 6:7. doi: 10.3389/feart.2018.00007

### INTRODUCTION

Preceding any potential volcanic eruption, the propagation of magma at shallow depth induces deformation of the Earth's surface (Dzurisin, 2007). During the last decade, geodetic measurements of ground deformation due to such magmatic intrusions have become a standard tool in monitoring active volcanic systems (Amelung et al., 2000; Wright et al., 2006; Sigmundsson et al., 2010, 2015). These tools commonly include tiltmeters (Toutain et al., 1992), Global Navigation Satellite Systems (GNSS; Bonforte and Guglielmino, 2015; Lee et al., 2015), Interferometry Synthetic Aperture Radar (InSAR; Massonnet et al., 1995; Lanari et al., 1998; Fukushima et al., 2005), and photogrammetry (Cayol and Cornet, 1998; Hollingsworth et al., 2013; Hibert et al., 2015).

Surface deformation is one of the several routinely monitored observables from active volcanic regions that are used to assess the volcano's behavior and eruption probability. However, preeruptive geodetic data from active volcanoes that were acquired at high frequency suggest that they contain some precursory signals useful to track the pre-eruptive propagation of volcanic feeders (e.g., Toutain et al., 1992; Cannavò et al., 2015). Thus, they may have the potential to be used to forecast the location of subsurface magma in real time. Nevertheless, systematic and robust surface deformation precursors for volcanic eruption locations have not been identified yet.

Here we present results from 33 scaled laboratory models of shallow intrusions that ultimately fed eruptions (**Figure 1**). During each experiment we periodically monitored the surface deformation caused by the subsurface propagation of the feeder. A simple geometrical analysis of the surface deformation data reveals that the eruption locations were systematically forecastable without performing any geodetic modeling. We also observe that distinct shapes of the eruptive feeders, here dykes and cone sheets, exhibit distinct, characteristic surface deformation signatures. We conclude that geodetic surface deformation data, if acquired at high enough spatial and temporal resolutions, do have the potential to be used to follow magma pathways at shallow depth and to forecast the locations of imminent volcanic eruptions without any modeling.

#### METHOD

#### Experimental Protocol

All the experiments were performed in the experimental apparatus of, and using the experimental protocol described by Galland et al. (2009), Galland (2012), Galland et al. (2014), and Guldstrand et al. (2017). Galland et al. (2006) describe in detail the mechanical properties of the model materials and the scaling of the models. Below, we briefly summarize the experimental materials and protocol.

The model materials are fine-grained silica flour and molten vegetable oil, to simulate brittle rocks and magma, respectively. The flour consists of fine (∼15µm), angular grains of crystalline silica flour. It has a cohesive strength of 369 ± 44 Pa, a friction coefficient of 0.81 ± 0.06 (corresponding to an angle of internal friction of ∼39◦ ) and a tensile strength of 100 Pa (Galland et al., 2006, 2009). As 1 cm in the models represents 100–1,000 m in nature, the resulting stress ratio indicates that the model crust should be 13 × 10<sup>3</sup> -250 × 10<sup>3</sup> times weaker than its geological prototype (Abdelmalak et al., 2016). The silica flour fulfills this criterion. It reproduces the brittle Coulomb behavior of the Earth's crust (Abdelmalak et al., 2016). Additionally, the flour is cohesive and has the ability to stand non-negligible elastic stresses along stable vertical walls (Abdelmalak et al., 2016). However, the elastic properties of silica flour remain poorly constrained, as is the case for granular materials in general. It is therefore challenging to address how the elastic stresses in our models scale with those in geological systems (Galland et al., 2017).

The model magma consists of a vegetable oil that is solid at room temperature and melts at ∼31◦C (Galland et al., 2006). Molten, it is a Newtonian fluid with a weak temperaturedependent viscosity (Galland et al., 2006). Using these materials, a generic experiment consists in injecting hot oil into the flour at room temperature to generate an intrusion. At the injection temperature of ∼50◦C, the oil exhibits a viscosity of 2 × 10−<sup>2</sup> Pa s and a density of 890 kg m−<sup>3</sup> . Oil percolation within the flour during injection is inhibited as silica is chemically incompatible with the oil and an oil intrusion is dominantly accommodated by deformation of the flour (Galland et al., 2006). During an experiment, the effects of cooling of the oil against the flour can be neglected, as intrusion durations are shorter than conductive cooling timescales. Our model scales through assuming that the ratio of viscous stresses in the oil/magma to the cohesion of the flour/host rock are identical in the model and nature (Galland et al., 2014). In nature, magma velocities can be of the order of 1–10−<sup>2</sup> m s−<sup>1</sup> (Toutain et al., 1992). The experimental device allows for oil velocities of 10−<sup>3</sup> -10−<sup>1</sup> m s−<sup>1</sup> . As magma viscosities cover a wide range (10–10<sup>7</sup> Pa s), relevant model viscosities fall in the range 4 × 10−<sup>9</sup> -75 Pa s, which the oil fulfills. To simplify, the oil at 50◦C dominantly represents a rather viscous magma of intermediate to felsic composition.

For a generic experiment, the experimental setup consists of a 40 cm wide square box with a circular inlet pipe at its center, into which a known mass of silica flour is poured. Then a high-frequency vibrator shakes the box to compact the flour until a bulk density of 1,050 kg m−<sup>3</sup> is reached. A flat metal plate is placed onto the model surface during compaction to ensure repeatable experiment preparation and an initial flat and horizontal surface of the models; the metal plate is removed after compaction. A volumetric pump injects the molten oil at constant flow rate through the circular inlet. With such a setup, it is possible to vary, among other parameters, the injection depth, the diameter of the inlet, and the flow rate. Depending on these parameter settings, the models systematically produce various geometries of intrusions, such as vertical sheet intrusions (dykes) and cone sheets (**Figure 2**; Galland et al., 2014). The vertical sheet intrusions initiated at the inlet and propagated to the surface. They often split to form a hull-shaped termination or turned into inclined sheets before reaching the surface (Galland et al., 2014).

#### Surface Data

The surface deformation data used in the present study were acquired during 33 out of the 51 experiments from Galland et al.

2012), and fed eruptions (black star). At each time step of a given experiment, simple geometrical parameters (see legend box) were calculated from the surface uplift map. This study shows that their evolution with time represents a precursor for the location of the next eruption.

(2014). Note that although surface deformation was monitored during all their experiments, Galland et al. (2014) focused on the dynamics of the intrusion processes at depth and on the resulting intrusion shapes, only. The resulting surface deformation dataset has subsequently been analyzed by Guldstrand et al. (2017), who focused on mechanical interpretations associated with the intrusion mechanisms at depth. The present analysis of the dataset is different and discusses the implications for volcanic hazards assessment. The 33 experiments considered here correspond to those for which enough surface deformation data were available during the entire duration of the experiments. They are representative of the full ranges of the parameters explored by Galland et al. (2014).

During the experiments, surface data were monitored using a moiré projection apparatus. The moiré monitoring (Bréque et al., 2004; Galland, 2012) was performed through projecting sets of illuminated straight fringes onto the model surface. The fringes remain straight on a flat surface but deform when projected on a surface with topography, producing curved fringe patterns. A video camera perpendicular to the surface captured the evolving fringe patterns on the model surface periodically (by successive scans starting at time step intervals of 1.5 s), which were subsequently analyzed to compute time series of digital elevation models (DEMs; Bréque et al., 2004). The duration of a scan for the acquisition of an individual DEM was ∼1 s and we chose to set the time of each DEM at the beginning of each scan.

Focussing on surface deformation induced by the intrusions, we have analyzed differential digital elevation models (1DEMs) obtained from the difference between the DEMs at given time steps and the DEM of the initial model surface. To limit noise effects, 1DEM data were smoothed. The lateral resolution of the 1DEMs is <1 mm, and the vertical precision of the smoothed 1DEMs is ∼0.1 mm (Guldstrand et al., 2017). As only uplifts were observed for both dyke and cone sheet experiments, for each 1DEM we have defined the group of pixels corresponding to the uplifted area using an uplift threshold criterion of 0.1 mm. We have then calculated the location of the mean center (C) of the uplifted area by averaging the positions of each pixel in the uplifted area, giving the same weight to each pixel (**Figure 1**). The locations of the centers of the uplifted areas were then known for each time step of each experiment in a consistent way. The uplifted areas never extended further than about 15 cm from the box walls, so that sidewall effects are assumed to be negligible. This is confirmed by the random location of the eruption sites in our experimental series.

The experiments lasted between a few seconds up to about 1 min, from the time at which the injection started up to the time at which the oil erupted. The second and the last scans of moiré projections started at about the same times, within errors of 1.5 s, as the injection started and the eruption occurred, respectively. To compare experiments of varying durations, we have normalized the time t at a given time step by the experiment duration, te. Therefore, for each experiment, the dimensionless time, t/te, which varied from 0 to 1, approximately represents the relative duration of the intrusion up to the eruption.

#### RESULTS

The 33 experiments that produced suitable surface deformation data lasted from ∼6 to ∼53 s. They produced 16 dykes and 17 cone sheets depending on the values of depth and diameter of the injection inlet, as well as the injection velocity of the oil (Galland et al., 2014; Guldstrand et al., 2017).

All the experiments, i.e., both those producing dykes and cone sheets, displayed an initial symmetrical bell-shaped uplift of the surface followed by the development of an uplift asymmetry that grew until the oil erupted in the immediate vicinity of the point of maximum uplift (**Figure 3**; Guldstrand et al., 2017). The dykes systematically triggered uplift, regardless of their final shapes, i.e., vertical sheets with or without, split or inclined terminations. To quantify the uplift asymmetry, we have calculated the positions of (1) the center of the uplifted area and (2) the point of maximum uplift at each time step (points C and M, respectively, **Figure 1**). We defined a vector, −→<sup>V</sup> MC, connecting these points.

During the early stages of uplift, in all the experiments, points C and M closely clustered, as illustrated by the short vectors −→<sup>V</sup> MC (**Figures 4C,D**), the orientation of which strongly varied with time. The points of maximum uplift (M) then migrated away from the center (C), as shown by the lengthening of −→<sup>V</sup> MC (**Figures 4C,D**). Concomitantly, the orientation of −→<sup>V</sup> MC focused and stabilized in azimuth with time. Importantly, in all the experiments, −→<sup>V</sup> MC ultimately pointed toward the subsequent eruption location (**Figures 4C,D**). The eruptions systematically initiated at the intersection between the ultimate −→<sup>V</sup> MC direction and the marginal border-zone of the uplifted area.

We also calculated (1) the evolution of | −→<sup>V</sup> MC<sup>|</sup> scaled by the injection depth (d) and (2) the rotation angle (θMC) of the vectors −→<sup>V</sup> MC between two successive time steps (**Figures 5**, **<sup>6</sup>**). For each experiment, the evolution of −→<sup>V</sup> MC  /<sup>d</sup> quantifies how point M moved away from C, and θMC indicates the stability of the direction of VMC. We arbitrarily consider that θMC was stable once it remained <20◦ .

There are systematic differences in the evolution of −→<sup>V</sup> MC  /d and θMC for dykes and cone sheets (**Figures 5**, **6**). During dyke experiments, on average, −→<sup>V</sup> MC  /<sup>d</sup> remained small until t/te∼0.4, from which −→<sup>V</sup> MC  /<sup>d</sup> increased rapidly before stabilizing again at t/te∼0.8 (**Figure 5A**), displaying an overall stepwise or two-phase evolution. In detail for each individual experiment, the rapid −→<sup>V</sup> MC  /<sup>d</sup> increase started at different times (t/te∼0.2 to 0.8; **Figure 5A**) and was often relatively short in

FIGURE 4 | (A,B) 1DEM before eruption for a representative dyke (A) and cone sheet (B) experiment (uplift in mm). White and black crosses show the successive locations of the centers (C) of the uplifted area, and of the maximum uplifts (M), respectively. (C,D) Plots of the successive vectors −→<sup>V</sup> MC computed from the respective maps, (A,B), from the early (dark blue) to the final stages (dark red). A black star locates the eruption points. Final points of maximum uplifts almost locate the eruptions.

time. In contrast, for cone sheets, −→<sup>V</sup> MC  /<sup>d</sup> exhibited a gradual, progressive, quasi-linear increase (**Figure 5B**). In addition, for most of the dykes, θMC was highly variable for more than half of the experiment durations (up to t/te∼0.6; **Figure 6A**) before decreasing and stabilizing, whereas for cone sheets, θMC generally stabilized earlier (t/t<sup>e</sup> ∼ 0.3; **Figure 6B**).

1DEM is displayed as fringes, each fringe series corresponding to an uplift of 0.5 mm.

#### INTERPRETATION AND DISCUSSION

During the 33 experiments, the vector −→<sup>V</sup> MC systematically pointed toward the location of the subsequent eruption once approximately stabilized in azimuth (± 20◦ ; **Figures 4**, **6**). As −→<sup>V</sup> MC is a parameter that was directly extracted from surface deformation data using only minimal calculations, real-time

measurements of −→<sup>V</sup> MC are potentially achievable in natural systems. Therefore, the evolution of −→<sup>V</sup> MC represents a robust geometrical precursor that could be useful in forecasting where a real eruption should occur, with substantial implications for hazard mitigation in active volcanic areas.

Consistent with our observations, previous two-dimensional (Abdelmalak et al., 2012) and three-dimensional (Galland, 2012) experiments, as well as theoretical models of surface uplift due to sheet intrusions (Pollard and Holzhausen, 1979; Okada, 1985), have also shown that the points of maximum uplift roughly locate the shallowest parts of intrusive feeders, such as dyke tips, at depth. Hence, the migration of a point of maximum uplift at the Earth's surface in volcanic areas likely represents a relevant geometric proxy to locate where magma is the shallowest and is ascending underground.

The distinct surface deformation signatures associated with the experimental dykes and cone sheets likely reflect contrasting emplacement dynamics (cf. Guldstrand et al., 2017). The progressive increase of −→<sup>V</sup> MC  /<sup>d</sup> from the earliest stages of subsurface propagation reflects the gradual asymmetrical propagation of a cone sheet (**Figure 5B**). Conversely, the stepwise or two-phase increase of −→<sup>V</sup> MC  /<sup>d</sup> is interpreted to indicate a two-stage evolution with (1) an initial vertical ascent of a dyke at depth, followed by (2) the interaction with the free surface and possible splitting of the dyke tip or oblique propagation toward the surface from a shallower depth (Mathieu et al., 2008; Abdelmalak et al., 2012; Galland et al., 2014). The stabilization of the orientation of −→<sup>V</sup> MC (**Figure 6B**) may coincide with this second phase. In addition, the contrasting signatures of

the experimental dykes and cone sheets suggest that real-time analysis of the deformation of natural surfaces can be useful to infer the geometry of a propagating intrusion prior to an eruption.

Our model uses an initial flat surface and does not include the effect of an initial topography or slope, often relevant for volcanic systems. Additionally, our model crust material is homogeneous and does not account for any heterogeneity that may also influence surface deformation signatures due to intrusions. Whether or not our method applies for shallow intrusions that develop elsewhere than under flat volcanic fields or calderas and in stratified and/or fractured crusts has not been tested. However, we expect that any magma-induced surface deformation will reflect the underlying developing asymmetry of the intrusion, in which case the method proposed here should still be applicable from a non-flat initial surface and a heterogeneous crust.

The surface deformation above our experimental dykes differs from that associated with dykes emplaced in rifts (e.g., Wright et al., 2006; Biggs et al., 2009; Sigmundsson et al., 2015) and the expected deformation predicted by static elastic analytical models of dykes (e.g., Okada, 1985). The latter display two prominent lobes of uplift separated by a trough aligned above the dyke apex. In contrast, our experimental dykes only triggered surface uplift, regardless of whether the intrusions propagated vertically up to the surface or deviated into inclined sheets. Guldstrand et al. (2017) attributed the difference with the static elastic models to the use of a weakly elastic, cohesive Mohr-Coulomb flour, in which the experimental dykes likely propagated as viscous indenters instead of resulting in pure elastic tensile fractures. In addition, the experiments account for magma flow and intrusion propagation, whereas elastic models are static. They are thus likely relevant for volcanic systems where the shallow crust is weak (e.g., Thun et al., 2016) and/or in which the intruding magma is relatively viscous (Galland et al., 2014; Guldstrand et al., 2017). Guldstrand et al. (2017) also attributed the difference with surface deformation measured in rifts to the absence of farfield tectonic extension in the experiments, thus making them relevant for volcanic systems that are not located in rifts.

Uplifting in the form of doming is commonly measured in active volcanic areas and models of inflating/pressurized spherical sources or horizontal sheet-intrusions generally fits such uplifts (e.g., Pedersen and Sigmundsson, 2006; Walter and Motagh, 2014). From our results, an alternative interpretation may consist in propagating vertical sheet intrusions through a Mohr-Coulomb crust (Guldstrand et al., 2017). Moreover, as our experiments produced inclined sheets on top of some vertical dykes, and cone sheets, our analysis may also be relevant for interpreting surface deformation in volcanic areas prone to forming inclined sheets and cone sheets (e.g., Bagnardi et al., 2013).

As mentioned above, the relevance of using points of maximum uplift has been proposed earlier. Such points have been recorded among geodetic data measured on active volcanoes, e.g., at Piton de la Fournaise, Réunion Island (Toutain et al., 1992). The data and interpretation of Toutain et al. (1992) satisfactorily compare to those from our experiments. Indeed, the correlation between the zone of maximum uplift and the eruption location, as well as the two-stage behavior of the surface deformation due to an intrusive feeder that was interpreted as a dyke, exhibit encouraging similarities with our experimental results. Another famous example was the prominent asymmetrical bulging preceding the 1980 eruption of Mount Saint Helens (Dzurisin, 2007, and references therein). The bulging flank of the volcano happened to be the location of the 1980 explosion, and laboratory experiments demonstrated that the asymmetry of the bulging reflected the asymmetrical shallow growth of the underlying cryptodome (Donnadieu and Merle, 1998; Merle and Donnadieu, 2000). These examples suggest that the precursors identified in the laboratory may also be applied to active volcanoes. Consequently, monitoring surface deformation on active volcanoes with both high temporal and spatial resolution has the potential to constrain, in real-time, simple geometrical parameters, such as | −→<sup>V</sup> MC<sup>|</sup> and <sup>θ</sup>MC, to forecast the location of both shallow intrusions and imminent eruptions. To make such forecasts possible requires implementing high frequency monitoring methods, such as GNSS and/or tiltmeter, and fast data processing. However, the lack in spatial resolution does not ensure accurate identification of the locations of uplift center and maximum, which conversely can easily be identified using InSAR data.

Notably, our results show that the location of most of the experimental eruptions could have been accurately predicted to occur within an angular sector of about 20◦ from approximately half of the experiment duration (**Figure 6**). Transposed to nature, where enough time is required to take suitable societal measures before an eruption occurs, such a forecast could be achieved up to several weeks to days before the eruptions. Indeed, the very first signs of pre-eruptive deformation on volcanoes have been documented to occur approximately up to 3 months prior to the eruptions (Froger et al., 2004; Peltier et al., 2006; Poland et al., 2008; Chadwick et al., 2012; Langmann et al., 2012). As some intrusions may also propagate underground over shorter timescales (dykes may propagate as fast as several tens of cm/s; Toutain et al., 1992), the predictions would be accurate enough within just a few hours before a potential eruption, which may be inadequate for hazard mitigation. Nonetheless, in adequate situations, our results indicate that the accuracy in predicting the location of an imminent eruption increases as time proceeds and that the first predictions could be given earlier when the feeder is a cone sheet. Moreover, our analysis allows for excluding a large part of the deforming area depending on the early direction −→<sup>V</sup> MC. Efforts can then be made to focus analysis on the area highlighted by −→<sup>V</sup> MC.

Our modeling approach and results highlight the dynamic nature of surface deformation associated with shallow magma emplacement. Resolving surface deformation both at high spatial and temporal resolutions is relevant to follow the evolution of simple geometric parameters, such as the point of maximum uplift, which constitute proxies for the location of on-going magma ascent. In addition, as long as changes in the evolution of parameters, such as the focus in azimuth of the points of maximum uplift, develop a significant time prior to an eruption, they have the potential to be used as precursors, indicative of the approximate location of an imminent volcanic eruption. Extracted only from the direct observation of surface data, these precursors are purely geometrical and are not derived from any mechanical criteria or hypothesis. Yet they are relevant for various magma feeder geometries. Our analysis illustrates that time-consuming computational surface data modeling, as commonly used to analyse geodetic data, may not be necessary for the purpose of forecasting eruption locations.

#### CONCLUSION

In this study, we analyse the surface deformation monitored during 33 scaled laboratory experiments simulating magma emplacement in a brittle crust under lithostatic stress conditions, i.e., not subjected to regional or local extensional tectonic stresses. Depending on the parameter sets, the experiments simulated the emplacement of dykes or cone sheets (Galland et al., 2014); the associated surface deformation systematically exhibit surface uplift. Our main results are the following:


evolution of the vector −→<sup>V</sup> MC is a good proxy for identifying the nature of the sub-surface volcanic feeder.


#### AUTHOR CONTRIBUTIONS

FG produced the data analysis, figures and wrote most of the text. OG contributed with the concept and scaling of the laboratory model, experiments, discussions and interpretations. EH performed experiments and contributed with discussion and analysis. SB performed experiments and provided comments and discussion.

#### ACKNOWLEDGMENTS

Guldstrand's doctoral position is funded by the Norwegian Research Council (DIPS project, grant no. 240467). Part of the work was performed in the MeMoVolc Networking Programme from the European Science Foundation (exchange grant no. 4251). F. Bjugger and A. Souche are acknowledged for assistance and discussions with the Volcano Plumbing Systems group at PGP. Burchardt acknowledges financial support Uppsala University and the Swedish Research Council for a research visit to PGP. Hallot acknowledges his welcome to PGP thanks to a half-year sabbatical from Université de Rennes 1. The authors acknowledge the constructive reviews of two reviewers, the associate editor, Nicolas Fournier and chief editor, Valerio Acocella, whose comments improved the manuscript.

#### REFERENCES


into Mount St. Helens deformation. Geology 26, 79–82. doi: 10.1130/0091-7613(1998)026<0079:EOTIPD>2.3.CO;2


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

Copyright © 2018 Guldstrand, Galland, Hallot and Burchardt. 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 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.

## Short-Term Seismic Precursors to Icelandic Eruptions 1973–2014

#### Páll Einarsson\*

Institute of Earth Sciences, University of Iceland, Reykjavík, Iceland

Networks of seismographs of high sensitivity have been in use in the vicinity of active volcanoes in Iceland since 1973. During this time, 21 confirmed eruptions have occurred and several intrusions where magma did not reach the surface. All these events have been accompanied by characteristic seismic activity. Long-term precursory activity is characterized by low-level, persistent seismicity (months-years), clustered around an inflating magma body. Whether or not a magma accumulation is accompanied by seismicity depends on the tectonic setting, interplate or intraplate, the depth of magma accumulation, the previous history and the state of stress. All eruptions during the time of observation had a detectable short-term seismic precursor marking the time of dike propagation toward the surface. The precursor times varied between 15 min and 13 days. In half of the cases the precursor time was <2 h. Three eruptions stand out for their unusually long duration of the immediate seismic precursory activity, Heimaey 1973 with 30 h, Gjálp 1996 with 34 h, and Bárðarbunga 2014 with 13 days. In the case of Heimaey the long time is most likely the consequence of the great depth of the magma source, 15–25 km. The Gjálp eruption had a prelude that was unusual in many respects. The long propagation time may have resulted from a complicated triggering scenario involving more than one magma chamber. The Bárðarbunga eruption at Holuhraun issued from the distal end of a dike that took 13 days to propagate laterally for 48 km before it opened to the surface. Out of the 21 detected precursors 14 were noticed soon enough to lead to a public warning of the coming eruption. In four additional cases the precursory signal was noticed before the eruption was seen. In only three cases was the eruption seen or detected before the seismic precursor was verified. In general, eruptions are preceded by identifyable short-term seismic precursors that, under favorable conditions, may be used for pre-eruption warnings. In some cases, however, the time may be too short to be useful. The Hekla volcano stands out for its short precursory times.

Keywords: seismic precursors, eruption precursors, volcanoes in Iceland, pre-eruption warning, eruption forecasting, precursor time

#### INTRODUCTION

The interaction of the Iceland hotspot with the mid-Atlantic plate boundary leads to volcanism of unusually wide variety. The high rate of volcanism associated with the hotspot produces basaltic crust that is 15–40 km thick, 3–8 times thicker than normal oceanic crust (e.g., Bjarnason, 2008; Brandsdóttir and Menke, 2008). Extensional tectonism therefore occurs in an environment that is different from that of the oceanic parts of the plate boundary. In addition to the basaltic

#### Edited by:

Yosuke Aoki, The University of Tokyo, Japan

#### Reviewed by:

Diana Roman, Carnegie Institution for Science (CIS), United States Alessandro Tibaldi, Università degli Studi di Milano Bicocca, Italy

> \*Correspondence: Páll Einarsson palli@hi.is

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 31 January 2018 Accepted: 13 April 2018 Published: 08 May 2018

#### Citation:

Einarsson P (2018) Short-Term Seismic Precursors to Icelandic Eruptions 1973–2014. Front. Earth Sci. 6:45. doi: 10.3389/feart.2018.00045

**29**

magmatism of the extensional environment a significant component of silicic magmatism is also present (Jakobsson et al., 2008; Sigmarsson et al., 2008). Furthermore, several of the volcanoes are sufficiently far away from the plate boundary rift zones to be termed intraplate volcanoes. Being above sea level and moderately populated Iceland therefore offers a suitable laboratory to study a wide range of volcanic phenomena at reasonably close range.

In the 1100 years history of cohabitation with the Icelandic volcanoes several cases have been noted and documented of seismicity immediately preceding volcanic eruptions (Thoroddsen, 1925). The bishop Hannes Finnsson (1739– 1796) even suggested, following the beginning of the Hekla eruption of 1766–1768, that instruments such as barometers and compass needles might be useful, also that paying attention to the intensity and direction of seismic shocks could be used to predict and warn of impending eruptions (Finnsson, 1767), an interesting suggestion more than a century before the invention of seismographs.

When sensitive seismographs became available and were installed in the vicinity of the highly active volcanoes in Iceland it became clear that their eruptions generally have detectable precursors. Seismogrphs have since become an integral part of a monitoring system with the objectives to detect precursory activity and warn against impending eruptions. The seismicity pattern may be divided into two categories, longterm and short-term. The long-term pattern, months to years, is characterized by low-level, persistent seismicity clustered at the volcano, often accompanied by, or caused by inflation of a magma body. The short-term precursory activity, tens of minutes to a few days, is characterized by an intense and growing swarm of small earthquakes, marking the time when increasing pressure in a magma body breaches its walls and a dike starts propagating through the crust. A rapid deflation of the magma body may result. Intermediate-term precursory activity also exists, but is more difficult to define. In this paper the experience of short-term seismic precursors to eruptions in Iceland is summarized. It is shown that all eruptions in the last decades have been preceded by swarms of microearthquakes with precursory times ranging between 15 min and 13 days. In twothirds of all cases the precursory activity has been identified soon enough to issue a warning prior to the outbreak of an eruption.

#### TECTONIC SETTING OF ICELANDIC VOLCANOES

Active volcanism in Iceland is limited to the divergent segments of the mid-Atlantic plate boundary that crosses the island from SW to NE, and a few flank zones that are not directly in the zones of plate divergence (Sæmundsson, 1978; Jakobsson, 1979a; Einarsson, 2008). Some of the volcanoes are therefore tholeiitic and intimately related to the process of plate separation, such as Krafla in the Northern Volcanic Zone (**Figure 1**) and Grímsvötn and Bárðarbunga in Central Iceland, located in the central area of the Iceland hotspot. Other active volcanoes, such as Hekla, Katla, and Eyjafjallajökull, are located in a flank zone, where transitional alcalibasalts are the main products. The Vestmannaeyjar volcanic system, with the eruptions of Surtsey 1963–1967 and Heimaey 1973, is alcalic.

Volcanological terms used in Iceland may in some instances deviate from the ones used in other areas. This results from the somewhat special circumstances of a subaerial divergent plate boundary. Volcanic structures are exposed on the surface that are usually submerged by ocean and the volcanism often is of areal extent. Walker (1993) used the term central volcano to describe areas of intense volcanism in the Tertiary lavas of Eastern Iceland and found them to be associated with dyke swarms. He used the term volcanic system for the structural unit consisting of a central volcano and a dyke swarm. Sæmundsson (1974) pointed out that the present-day equivalents of the dyke swarms were the fissure swarms commonly found in the neovolcanic zones and defined the active volcanic systems of the volcanic rift zone in Northern Iceland. Later, using the same criteria, he defined the volcanic systems of the whole neovolcanic zone of Iceland (Sæmundsson, 1978). Jakobsson (1979b) established petrological characterisics of most of the systems of South and Central Iceland and showed that they could be distinguished by the chemical composition of their products.

For clarification we define these terms as follows:

Central volcano is an area of high volcanic productivity. Central volcanoes are usually of basaltic composition but often contain a significant quantity of rocks with high silca content such as rhyolite and dacite. A central volcano may have one or more calderas and geothermal systems. A fissure swarm is a collection of many similar, parallel or subparallel fissures and normal faults occurring in a limited area. A volcanic system is a structural and petrological unit consisting of a central volcano and associated fissure swarms.

We note that there is overlap between these terms and terms used elsewhere. The Hawaiian term rift zone, e.g., is almost synonymous with our term fissure swarm. The rift zones of Kilauea, the SW- and E-rift, would be called fissure swarms in Icelandic terminology. An Icelandic rift zone consists of several volcanic systems and would contain several fissure swarms. Also, our term volcanic system is in some cases similar to the term ridge segment used for structural units on the mid-ocean ridges.

Different authors have used different definitions for volcanic systems of Iceland and there is considerable confusion in the literature regarding names. In this paper we follow the classification of Einarsson and Sæmundsson (1987) which is widely referenced.

### MONITORING SYSTEMS

The technical possibilities to detect premonitory changes to eruptions has changed greatly during the last half century and the threshold of detection is lowered throughout the period (Einarsson and Björnsson, 1987). The following periods can be defined:

Pre-1965: The first seismographs were installed and operated 1909–1914 and then reinstalled in the 1920ies. Although insensitive, they gave an indication of the activity. The sensitivity was greatly improved by the installation of the 6-component WWSSN-station of Akureyri (AKU) in 1964. By then four stations were operated in the country. They were analog stations, recording on photographic paper.

1965–1973: Experiments were done with temporary field seismographs, both to record aftershocks, earthquakes swarms and seismicity associated with volcanoes and geothermal areas. The first experiment was conducted on Surtsey island produced in the eruption of 1963–1967 off the south coast (Einarsson, 1974). Microearthquake surveys were done of the whole country (Ward et al., 1969; Ward and Björnsson, 1971). Temporary seismic networks were operated on the Reykjanes Peninsula (Björnsson et al., in review) and prototype instruments were installed at permanent locations in South Iceland. They detected seismicity associated with the Heimaey eruption of 1973. Parts of the Surtsey eruption and the eruption of Hekla in 1970 were monitored with local seismographs recording on FM magnetic tape.

1973–1991: The number of sensitive seismographs increased greatly in the late 1970ies and a network of 40–50 stations was in in operation throughout the active areas by 1979. These were one-component, analog seismographs, recording with pen on paper, designed at the Science Institute, University of Iceland. These instruments were used to monitor the progress of the Krafla magmato-tectonic events in North Iceland 1975–1984, the eruptions of Hekla 1980–1981, and Grímsvötn in 1983. This network was extended into the interior of Iceland in 1985, when telemetered seismographs were installed at some of the active volcanoes there.

1991-Present: A new generation of seismographs was introduced in 1991 with the South Iceland Lowland digital network (Stefánsson et al., 1993). This network was expanded to North Iceland in 1994 and at the time of writing more than 60 stations are in operation in Iceland. The network operates semi-automatically and provides preliminary locations and magnitude determination on-line within a few minutes, at http://www.vedur.is/. The SIL-system has detected seismicity associated with the Hekla eruptions of 1991 and 2000, the Gjálp eruption of 1996, the Grímsvötn eruptions of 1998, 2004, and 2011, the Eyjafjallajökull eruptions of 2010, and the Bárðarbunga activity of 2014 onwards. The network has been augmented by local seismograph networks operated in South and Central Iceland since 2005 in cooperation between Cambridge University, Uppsala University, and Icelandic institutions and have provided data for valuable studies of Askja, Eyjafjallajökull, Katla, and Bárðarbunga Volcanoes.

#### HISTORIC ACCOUNTS OF PRE-ERUPTION SEISMICITY

As mentioned above, it has been general knowledge in Iceland for centuries that eruptions are preceded by earthquakes. Bishop Hannes Finnsson compiled some of the historical, written documents and Thoroddsen (1925) continued his work. In the pre-instrumental era the known cases are limited to those where earthquakes were felt prior to the beginning eruptions. Out of 18 large, historic eruptions of Katla (Thorarinsson, 1975), nine were accompanied by felt earthquakes according to written documents. This does not mean that the other nine were not accompanied by earthquakes also, but they are not mentioned. In the eruptions of 1311, 1625, 1721, 1823, 1860, and 1918 it is specifically stated that the felt earthquakes occurred well before the eruption was seen, from a couple of hours to several days. Hekla shows a very different behavior. The generally low seismic activity associated with Hekla eruptions is remarkable. Exceptions are the eruptions in the Hekla volcanic system outside the main edifice, such as the eruptions of 1554, 1725, 1878, and 1913 (Thorarinsson, 1967b). The last two, at least, were lava eruptions issuing from eruptive fissures, and may have a strong tectonic component. The rifting episodes of Krafla 1724– 1746, the Grímsvötn system in 1783–1785 (Laki Fires), and Askja in 1874–1876 were accompanied by felt earthquakes, also the Öræfajökull eruptions of 1362 and 1727. The historical reports of these events are compiled by Einarsson (in preparation 2018).

#### INSTRUMENTAL OBSERVATIONS, CASE HISTORIES

#### Surtsey

The first seismic recording of a beginning of an eruption in Iceland at a close range was obtained during the Surtsey eruption of 1963–1967. The eruptive activity began in November 1963 on the ocean bottom at the southern tip of the Eastern Volcanic Zone (e.g., Thorarinsson, 1964, 1965, 1966, 1967a, 1968; Thorarinsson et al., 1964), and continued until June 1967. It was divided into phases, separated by short quiet intervals. Three islands were formed but two were eroded down below sealevel in a few months. One of the phases ended on August 10 1966. On August 19 a new eruptive fissure opened up on the island of Surtsey. Three lava craters were active in the beginning but a few days later only one remained. Lava was erupted from this crater until June 5, 1967, building up a flat lava shield and extending the Surtsey island to the east. A small array of seismometers was in operation on the island when this lava eruption started (Einarsson, 1974) recording on an FM-magnetic tape. It recorded a swarm of small, very local earthquakes that began at 09:30 h (**Figure 2**). None of these events were felt on the island. The swarm ended by 10:30 h and shortly thereafter the amplitude of the background noise increased slowly. The noise increased rapidly at 10:50 h and it is inferred that this marks the beginning of the eruption. The eruption was not discovered until about 2 h later by the watchman on the island. This sequence of events has all the characteristics of precursory seismicity, i. e. a swarm of small earthquakes that ends before or about the time of the beginning of the eruption, and followed by eruption tremor.

#### Heimaey

An eruption started on January 23, 1973, between 01:50 and 01:55 h on the island Heimaey in the Vestmannaeyjar volcanic system, only 200–300 m east of the town of 5,300 inhabitants (Thorarinsson et al., 1973). The eruption was preceded by an earthquake swarm 30–14 h before the outbreak (Björnsson and Einarsson, 1974). A few events that occurred immediately before the outbreak were felt in the town. Other precursors were not reported. Within a few minutes the NNE-striking eruptive fissure was 300–400 m long. A length of about 3 km was attained on January 25 and was extended to the north an additional half kilometer on February 6. The earthquake swarm preceding the outbreak was recorded by prototype seismographs on the mainland but could not be located accurately. Later, during the eruption, when more seismographs had been installed, both on the mainland and the island of Heimaey, earthquakes could be located with fairly good accuracy. They turned out to occur at larger depth than had been seen before in Iceland, 15–25 km (Einarsson, 1991a). The similar waveforms of these later events suggest that the precursory earthquakes were also located at this relatively large depth.

#### Krafla

The volcano-tectonic episode that took place in the Krafla volcanic system in 1974–1989 was a source of many data sets and observations on the relationship between crustal movements, seismic activity and magmatism. The episode included at least 20 deflation events of the Krafla volcano, when an inflating magma chamber at about 3 km depth beneath the caldera was breached and magma was injected into the adjacent fissure swarms (e.g., Björnsson et al., 1977; Tryggvason, 1984; Einarsson, 1991b; Brandsdóttir and Einarsson, 1992; Buck et al., 2006; Heimisson et al., 2015). The magnitude of these deflation events was quite variable, from being barely measureable to amounting to 2 m of subsidence in the center of the caldera. Nine of these dike injections found their way to the surface and produced lava eruptions in the rift zone (Sæmundsson, 1991). Warnings were issued priorto all the eruptions except the first one. The eruptions were:

1975, December 20: The first and largest of the deflation events began rather abruptly after a few months of elevated seismicity in the caldera. An intense earthquake swarm began at 10:17 merging into continuous vibrations of the ground. This activity was detected on seismic stations across the whole of Iceland. About 15 min later a small lava eruption broke out at the center of the caldera. Report of this activity came from various directions, both about the visible eruption and the detected seismicity. A warning came only after the eruption had been sighted, however. The eruption was very small and only lasted about 2 h. It stopped when dikes propagated laterally out of the caldera, as shown by propagating earthquakes. The dike intrusion lasted almost 3 months and the dike attained a length of 60 km.

1977, April 27: Following three diking events and reinflation a new and rapid deflation event began at 13:17 accompanied by tremor and an earthquake swarm. The earthquakes propagated along the southern fissure swarm of Krafla and were accompanied by large scale surface rifting. Small eruptions occurred in two locations during this swarm, one in the center of the caldera and a small lava patch at the northern caldera rim. Because of a snow storm this day the timing of the eruptions is not accurately known. Warnings were issued to the

local population almost instantly, so it was almost certainly issued prior to the eruption outbreak.

1977, September 8: A new deflation event began with rapid subsidence of the caldera and earthquakes at 15:47. A small eruption broke out at the northern caldera rim about 18 h. The eruption ended rather abruptly, however, when earthquakes began propagating out of the caldera into the southern fissure swarm (Brandsdóttir and Einarsson, 1979). Warnings about a possible eruption were issued well before the eruptions broke out. In addition, the dike led to a small magmatic eruption through a geothermal drillhole in the southern fissure swarm (Björnsson and Sigurðsson, 1978), the only known magmatic eruption through a man-made structure.

1980, March 16: Rapid deflation and tremor mixed with earthquakes began at 15:17. The tremor increased markedly half hour later, and an eruption began slightly north of the center of the caldera at 16:20. The eruptive fissure extended northwards and in 25 min it had attained a length of 4.5 km. Earthquakes continued propagating northwards and then also southwards out of the caldera. The diking activity led to a pressure drop and eventually to an end to the eruption, first at the southern end of the fissure. The eruption was over by 22:30. The seismic activity was noticed immediately by the attendant of the instruments and information about the progress of the activity was distributed to the local community and the media.

1980, July 10: Slow deflation began according to tiltmeters at about 8 h in the morning. One hour later weak tremor was detected on the seismographs. Small earthquakes indicated propagation to the north. Tremor amplitude and earthquakes became larger as time went on. A low-frequency event was detected at noon and an eruption was finally verified in the fissure swarm north of the caldera at 12:53. Both the tremor and the earthquake activity diminished significantly when the eruption broke out. This eruption continued for 8 days and was much larger than any of the previous eruptions. No lateral diking was observed after the beginning of the eruption. Subsequent three eruptions were similar in magnitude. The local population and the media were kept well informed during the whole course of these events.

1980, October 18: The course of events in this eruption was quite a bit faster than in the previous eruption. Deflation began at 20:42 and tremor was detected a few minutes later. The deflation rate and earthquakes increased rapidly and were higher than ever in previous events. A low-frequency event was detected at 21:45 and 19 min later an eruption was seen about 2 km north of the center of the caldera. The eruptive fissure extended quickly to the north and south until it attained a length of 7 km. The vigor of the eruption then slowly diminished until the eruption came to an end on October 23.

1981, January 30: Slow deflation began at 7 h and became gradually faster. Tremor was detected half hour later (**Figure 3**). Deflation rate and tremor amplitude culminated about 9 h and then slowly decreased. A low-frequency earthquake was detected at 13:13 and following that the earthquake activity slowly decreased. An eruption broke out on a 2 km long fissure in the fissure swarm north of the caldera at 14:10. This eruption ended on February 4.

1981, November 18: Rapid deflation began at 00:36 accompanied by tremor, both increasing fast, culminating between 01:15 and 01:35. The first low-frequency earthquake was detected at 01:38 and an eruption broke out at 01:52. The first eruption site was about a kilometer north of the center of the caldera, but the fissure quickly propagated both south- and northwards until it reached 8 km length at about 4 h. The eruption vigor then slowly decreased until the eruption came to an end on November 23.

1984, September 4: The inflation rate was high in the first month following the 1981 November eruption. It then slowed down and became irregular as the level of inflation exceeded the previous maximum. This condition remained until a new deflation event began at 20:25 on September 4, 1984. Tremor was detected at the local seismic stations at 20:40 and earthquake activity increased, mostly between 22 and 23 h (**Figure 4**). The

tremor, E and F increasing high-frequency earthquake activity, G low-frequency earthquake, H earthquake activity decreases, eruption begins at 14:10, I and J mark

low-frequency tremor. From Einarsson (1991b). Reproduced with the permission of the copyright holder, the Icelandic Natural History Society.

first low-frequency earthquake was detected at 23:40. Nine 9 min later (at 23:49) the flare of an eruption was seen. The eruption began almost simultaneously on two fissure segments on either side of the northern caldera rim. The segments grew in length and new segments became active until a continuous wall of fire was active, 8.5 km in length, extending northwards from the center of the caldera well into the northern fissure swarm. The eruption followed similar course as previous eruptions for the first 3 days, slowly diminishing until only one crater was active. On September 7 there was a change in course. The lava flow rate from this crater began increasing day by day, accompanied by increasing deflation rate of the caldera. This increase continued until September 18, when the eruption came to a sudden end. This eruption was by volume the largest eruption of the whole series.

The volcano began inflating after this last eruption of the episode and eventually reached the pre-eruption inflation level after several years. No inflation has been detected since 1989. The total volume of erupted lava is difficult to estimate because of the extensive changes in topography associated with the rifting. Values in the range 250–350 Mm<sup>3</sup> seem reasonable. Volume of intruded magma during the whole rifting episode is probably larger, possibly much larger.

#### Hekla

Seismic observations are available for the 1970, 1980–1981, 1991, and 2000 eruptions of Hekla. The precursory seismic activity for all these eruptions is similar in many ways. No long-term changes have been identified (Soosalu and Einarsson, 2005). Short-term changes were recorded 23–79 min before the outbreak of the eruptions. This short precursor time appears to be inconsistent with the large depth of 15–20 km to the feeding magma reservoir (Soosalu and Einarsson, 2004; Ofeigsson et al., 2011; Geirsson et al., 2012). Sturkell et al. (2013) discuss this in connection with their finding that the surface eruptive fissure only extends

to shallow levels in the crust, below which a pipe-like conduit extends downwards to the reservoir. The conduit remains fluid between eruptions. The short precursor time then represents the travel time of magma from the top of the conduit to the surface and not the propagation of a dike from a deep reservoir.

1970, May 5: The eruption began at 20:23 according to eyewitnesses in two areas on the lower flanks (Thorarinsson, 1970; Thorarinsson and Sigvaldason, 1972). The precursory earthquake swarm is exemplified by the seismogram of the AKU station (**Figure 5**) in North Iceland, at a distance of 200 km from Hekla, recording the beginning of the eruption. Small earthquakes become visible at 19:58, then gradually becoming larger and more frequent (Einarsson and Björnsson, 1976). This eruption lasted until July 5.

1980, August 17: The first earthquakes were recorded by a short-period seismograph at 22 km distance at 13:04 h on August 17 (Grönvold et al., 1983). The eruption outbreak was timed at 13:27, which was also the time when continuous, low-frequency tremor became visible on the seismograms. The eruption was relatively intense in the first day but declined rapidly and was over by August 20. Activity was renewed on April 9, 1981, and low-level activity continued for about a week. This activity was accompanied by continuous tremor of low amplitude, but only a few small earthquakes.

1991, January 17: The precursory activity to the 1991 eruption (Gudmundsson et al., 1992) was recorded by many seismographs, both analog, telemetered stations in the highlands to the E and NE of the volcano and a new, digital network, the SIL-network, W of Hekla. The earthquake swarm began at 16:30 (**Figure 6**) with small events, which quickly became larger until the eruption broke out between 17:00 and 17:02 (Gudmundsson et al., 1992; Linde et al., 1993; Soosalu and Einarsson, 2002; Soosalu et al., 2003). The recorded seismic activity was detected by people before the eruption was seen, but only after the eruption had already started. The eruption was most intense during the first hours and then gradually diminished in intensity and ended on March 11.

2000, February 26: The beginning of the 2000 eruption is the best documented one. A short-period, telemetered, analog station operating on the flank of the volcano was being attended when the precursory swarm began with tiny earthquakes at 17:00 (Soosalu et al., 2005). Because of the unusual occurrence of earthquake swarms near Hekla, the activity was immediately taken as a possible precursory signal. Less than 20 min later the Civil Defense Authorities had been notified of a likely outbreak of a Hekla eruption. The beginning of the eruption was timed accurately at 18:19 h by an eyewitness interview over telephone, broadcast life on the National Radio. This eruption ended on March 8.

#### Grímsvötn

The mostly subglacial Grímsvötn volcano in Central Iceland is known as the most frequently erupting volcano of Iceland. Following a small eruption in 1934, however, it went into a quiet state for almost half a century (Björnsson and Einarsson, 1990; Guðmundsson and Björnsson, 1991; Guðmundsson et al.,

1995). Small eruptions of 1983, 1998, and 2004 were followed by a relatively large eruption of 2011. All these four recent eruptions were preceded by detected seismic precursory activity and warnings were issued prior to the outbreak of two of them.

1983, May 28: The eruption was preceded by a significant increase in seismic activity for about 3 months (Einarsson and Brandsdóttir, 1984). An intense swarm began with an M 2.9 event at 02:30 on May 28. The swarm lasted about 9 h and was followed by continuous tremor, first seen at about 12 h at the nearest seismograph, at a distance of 65 km from Grímsvötn. The eruption is inferred to have begun between the time of the last earthquake, at 11:47, and the appearance of the continuous tremor. It was not verified visually until the next day, due to the remoteness of Grímsvötn. The eruption lasted 5 days.

1998, December 18: The 1998 eruption was preceded by elevated seismicity for several months, apparently due to inflation of the volcano, consistent with GPS-measurements on the southern caldera rim (Sturkell et al., 2003a). A temporary study by Alfaro (2001) revealed vigorous seismic activity along the western and southern caldera rim during a recording period from late May to August 1998. A small earthquake swarm began on December 17 at about 22 h and a sharp increase in earthquake activity was recorded at 03:30 on December 18. Continuous tremor with volcanic characteristics was recorded at 09:20, marking the beginning of eruptive activity. The long-term restless state of Grímsvötn was recognized before the eruption, but the short-term precursory activity was first identified after the beginning of the eruption. The eruption column rose to 10 km height in 10 min and began to decline the following day and ended on December 28.

2004, November 1: The eruption of Grímsvötn volcano in 2004 was forecast on several different timescales. Inflation of the volcano began immediately after the 1998 eruption and was monitored by GPS measurements on the caldera rim. Seismic activity began increasing in July 2003 and by September 2004 the inflation level of the volcano had reached the 1998 pre-eruption level (Vogfjörd et al., 2005; Sturkell et al., 2006). It became public knowledge that a Grímsvötn eruption was imminent. At this time it also became known that the lake level in the Grímsvötn caldera was rising beyond the critical level for a jökulhlaup, a situation that could possibly trigger an eruption by sudden release of pressure on the magma system. The beginning of a jökulhlaup was detected on October 27 by high-frequency tremor on regional seismographs. The flood was detected on the lowland 2 days later. An announcement was given for increased probability of an eruption within a few days. A dense swarm of small earthquakes began in the early hours of November 1 signifying magma propagating toward the surface. The swarm intensified at 19:30 and by 20 h the activity was dominated by volcanic tremor (**Figure 7**), indicating the beginning of an eruption. This was a relatively small eruption. The maximum plume height of 12 km was measured a few hours after the beginning, and by November 3 the plume disappeared from radar. The eruption appears to have ended on November 6.

2011, May 21: This largest eruption of Grímsvötn for at least a century followed a period of inflation that began immediately after the 2004 eruption. Increasing seismic activity accompanied the inflation from 2009 onwards. A dense earthquake swarm began on May 21, 2011, at about 17:50 and an eruption began around 19 h (Hreinsdóttir et al., 2014). The eruption column reached a height of more than 20 km during the first hour, the tallest for several decades in Iceland. It occasionally reached height of 10–12 km during the following days, but then decreased fast. The eruption was over by May 28. A warning was issued well before the beginning of the eruption.

#### Gjálp

1996, September 30: The subglacial Gjálp eruption occurred on a 7-km long eruptive fissure between the calderas of Bárðarbunga and Grímsvötn and was preceded by several years of unrest, both earthquakes and jökulhlaups from glacial cauldrons (Einarsson et al., 1997). The immediate precursory activity was rather unusual, apparently triggered by an earthquake of magnitude 5.4 (Ms) at 10:48 on September 29 on the northern caldera rim of Bárðarbunga. Similar earthquakes had occurred 14 times before, in a sequence that began in 1974 (Einarsson, 1991a), and every event had been followed by seismic quiescence. This time, however, the earthquake was followed by an intense swarm of smaller events in the caldera that continued through September 29 and 30 and propagated southwards, out of the caldera and toward the Grímsvötn volcano (Einarsson and Brandsdóttir, 1997). A warning was issued that Bárðarbunga might be about to erupt, based on the intensity of the swarm and its unusual course of events. Tremor with characteristic low-frequency appeared on the seismograph at Grímsvötn in the evening of September 30 indicating the beginning of an eruption. Visual confirmation came the following morning when large cauldrons in the ice surface were seen from an overflying aircraft. The glacier in the area is 400–600 m thick and it took the eruption about 60 h to melt through the ice to form an eruption column. The eruption lasted until October 13 and formed about 0.5 km<sup>3</sup> of basaltic andesite (Guðmundsson et al., 1997; Gudmundsson et al., 2004).

#### Bárðarbunga

2014, August 29: In terms of erupted volume this was the largest eruption in Iceland since the Laki eruption in 1783. The eruption was preceded by a lateral dike propagation from the caldera of Bárðarbunga that began on August 16 and was identified that day (Sigmundsson et al., 2015). The progress of the dike was monitored by its associated earthquakes and crustal movements caused by volume changes of the dike. The volume loss beneath the caldera caused collapse of the caldera floor that could be monitored by a GPS-station on the glacier surface within the caldera (Gudmundsson et al., 2016). The dike propagated for more than 45 km at variable rate before a small eruption broke out at its distal end on August 29, in the Holuhraun area of Central Iceland, 13 days after the initial breach of the

magma storage beneath the caldera. This eruption lasted only 4 h. Another eruption broke out on August 31 on the same fissure. This eruption lasted until the end of February 2015. The eruptions were preceded by the formation of a narrow graben above the dike near the eruption site (Hjartardóttir et al., 2016). This long "short-term" precursor to the eruptions was detected and identified very quickly and aroused public attention and alert during the whole propagation process.

#### Eyjafjallajökull

2010, March 20: Following 18 years period of unrest, including 3 intrusions in 1994, 1999, and 2009 of several months' duration each (Pedersen and Sigmundsson, 2004, 2006), a small lava eruption broke out on the eastern flank of the volcano, at Fimmvörðuháls (Sigmundsson et al., 2010). The eruption immediately followed the fourth intrusion that began in early January 2010 accompanied by escalating seismicity, more intense than in any of the previous episodes. An eruption was anticipated but the time scale was uncertain. Earthquake hypocenters were propagating toward the eastern flank but the premonitory seismic signal was rather weak. When the eruption finally broke out it was first spotted visually by local inhabitants. The eruption continued with low-level fire fountains for 3 weeks, and produced a small lava field extending into the gullies on the NE flank. The eruption ended rather abruptly on April 12. The volcano did not deflate during the eruption, indicating that it was fed by magma directly from deep sources.

2010, April 14: The sudden end of the flank eruption suggested that the magma feeding channel had been blocked or breached, and further activity might be expected. This came sooner than anticipated. A new eruption broke out shortly after midnight on April 14 and this time in the summit region of the volcano (e.g., Guðmundsson et al., 2010). It was preceded by a distinct increase in earthquake activity beneath the summit caldera. The beginning of the eruption appears to have been very subtle and was only detected by low-amplitude volcanic tremor at 01:15. No further signs of the eruption were seen until 06:50. Then a pulse of tremor was recorded, a considerable body of water was released from the summit caldera and the eruption broke through the glacial cover. The eruptive column reached a height of 10 km in the first day.

This summit eruption lasted 39 days (Gudmundsson et al., 2012) and spread ash widely, including the European continent where it blocked air traffic for several days.

#### SEISMIC CRISES WITHOUT SUBSEQUENT ERUPTIONS

The most common short-term precursory activity to eruptions is due to a dike propagating away from an inflated magma chamber, the seismicity at the dike tip and the sudden pressure drop in the chamber. During the propagating phase it is usually uncertain whether or not the dike will reach the surface. A good proportion of all dikes does not, and yet the geophysical signals are identical to those preceding an eruption. The warning issued prior to eruptions therefore has to include the possibility that the dike might not lead to an eruption. This was the case in the Krafla rifting episode 1975–1984. At least 20 deflation events of varying size occurred, only nine of which included an eruption. Warnings were issued at the onset of all the events except the first one. A warning without eruptions can hardly be called "false alarm," however. Some of the intrusive events had more serious consequences than most of the eruptions, and a warning was of great importance to the population of the area. Significant diking events without eruptions occurred in September and October 1976, January 1977, January, July, and November 1978, May 1979, February and December1980 (Einarsson and Brandsdóttir, 1980; Einarsson, 1991b; Buck et al., 2006; Wright et al., 2012). In a few instances it was observed that the pressure drop associated with a dike propagation stopped an eruption that had already begun. This happened in December 1975, April and September 1977, and March 1980, see above. In all these cases an eruption began within the caldera shortly after the intial dike intrusion started, but stopped abruptly when the dike propagated out of the caldera.

The Eyjafallajökull eruptions of 2010 were preceded by 18 years of unrest. During this time three sill intrusions were detected without an eruption, in 1994, 1999, and 2009. These events were slow, lasted a few months each and were accompanied by uplift of the volcano and seismicity (Dahm and Brandsdóttir, 1997; Sturkell et al., 2003b; Pedersen and Sigmundsson, 2004, 2006; Hjaltadóttir et al., 2015). No intense swarms were observed, however, and there was never a question of whether a warning of an impending eruption should be issued. The fourth intrusion began toward the end of 2009 and was more intense than the previous three. It intensified greatly in February and March 2010 and finally culminated with the outbreak of the flank eruption on March 20, see above (Sigmundsson et al., 2010).

Historically among the most active volcanoes of Iceland, the Katla volcano requires a special mention. The caldera of the volcano, where all its eruptions have taken place in the last thousand years, is covered by a thick glacier. It takes a large eruption to melt through the ice to produce a subaerial eruption and an eruption column. Such eruptions have occurred about twice per century, each time accompanied by a catastrophic flood and destruction. The latest such eruption occurred in 1918 and the current quiet interval is the longest in historic times. At least three events have taken place, however, during which cauldrons have been formed in the ice cover of the volcano and floods have issued from the glacier edge. These occurred in 1955 (June 25), 1999 (July 18), and 2011 (July 8–9), and were accompanied by earthquakes and seismic signals that resemble volcanic tremor (Tryggvason, 1960; Einarsson, 2000; Gudmundsson et al., 2000; Sgattoni et al., 2017). No eruptive products were seen above the ice surface and therefore these events fall into the category of uncomfirmed eruptions. If these were eruptions, the times of their beginning are unknown. The precursory times can therefore not be determined. The 2011 event was studied in considerable detail. It was preceded by a peculiar sequence of small earthquakes that clustered on the southern flank of the volcano and continued for several years after the event (Sgattoni et al., 2016). Two kinds of contiuous tremor were identified, one associated with the water flood that issued from the glacier, the other originated near the two ice cauldrons where the flood originated (Sgattoni et al., 2017). The event was followed by increased earthquake activity in the caldera (Sgattoni et al., in review).

In addition to the confirmed eruptions listed above, there have been many sub-glacial events that resemble eruptions, but were not large enough to break through the ice cover, similar to the events in 1999 and 2011 at Katla. This includes a burst of tremor in the Grímsvötn caldera in 1984 (Einarsson and Brandsdóttir, 1984; Björnsson and Einarsson, 1990), tremor burst and a jökulhlaup at Hamarinn, NW of Grímsvötn in July 2011, and several tremor bursts following jökulhlaups from the Skaftá cauldrons NW of Grímsvötn (e.g., Einarsson et al., 1997; Soosalu et al., 2006).

Increased seismic activity and slow land uplift of the Hrómundartindur volcanic system at the Hengill Triple Junction in SW-Iceland indicated magma flow into the roots of the volcanic system at about 6 km depth (Sigmundsson et al., 1997; Feigl et al., 2000; Clifton et al., 2002; Pedersen et al., 2007). Slow uplift continued for more than 4 years, reaching 8 cm at the apex (Feigl et al., 2000). The volume of injected magma may be estimated 15 Mm<sup>3</sup> . Earthquakes of magnitude 5 occurred at the perifery of the uplifted area in June 1998 and 10 km farther south in November 1998 (Rögnvaldsson et al., 1998). The activity then faded away without any indication of propagating dikes.

Evidence of a dike injection in the lower crust was provided by elevated seismicity and surface uplift at the hyaloclastite mount Upptyppingar in the Northern Volcanic Zone in 2007– 2008 (e.g., Jakobsdóttir et al., 2008; Hooper et al., 2011; White et al., 2011). The distribution of hypocenters and the surface uplift field were consistent with the intrusion of an inclined dike or a sheet, at a depth of 15–25 km, with a dip of about 45◦ and striking transversely to the rift zone. The slow intrusion lasted about 1 year and ended without an extrusion to the surface.

#### DISCUSSION

Even though it has been general knowledge in Iceland for centuries that the outbreak of volcanic eruptions is commonly associated with earthquakes, it wasn't until sensitive seismographs were installed in Iceland and had been in operation for a few decades that it became clear that all eruptions are preceded by characteristic seismicity, eathquakes and continuous tremor. All cases compiled in the present study confirm this conclusion. The long-term pattern is characterized by low-level, persistent seismicity clustered at the volcano, often accompanied by, or caused by slow stress changes due to inflation of a magma body. The short-term precursory activity, on the other hand, is clearly distinguishable. It is characterized by an intense and growing swarm of small earthquakes, marking the time when a dike starts propagating through the crust. This general theme may have different expressions depending on many parameters, s.a. depth of the magma body in the crust, regional tectonic stress, viscosity of the magma, rate of prior pressure increase etc. We find that the precursor time of the 21 eruptions is highly variable, between 15 min and 13 days, see **Table 1**. Half of the observed precursor times were shorter than 2 h. The times are even quite variable for the same volcano. The nine Krafla eruptions during the rifting episode 1975–1984 had precursor times that varied between 15 min and 7 h. Precursor times for the four Grímsvötn eruptions varied between 90 min and 15 h. Hekla stands out for its short precursor times, 25, 23, 30, and 79 min, repectively.

Three eruptions stand out for their unusually long precursor times, Heimaey 1973 (30 h), Gjálp 1996 (34 h), and the Bárðarbunga 2014–2015 (13 days) eruptions. The circumstances for them are quite varied. The Heimaey eruption occurred within the off-rift Vestmannaeyjar volcanic system, and the eruption apparently was fed from unusually great depth, 15– 25 km, without the involvement of a magma chamber. The premonitory activity to the Gjálp eruption all took place within the Bárðarbunga volcano and apparently involved the caldera fault, propagating dikes and subsidiary magma bodies (Einarsson et al., 1997). The eruption of Bárðarbunga 2014–2015 was a part of a major regional rifting event, and occurred at the distal end of a laterally propagating dike, at the distance of 45 km from the feeding magma chamber (Sigmundsson et al., 2015; Gudmundsson et al., 2016).

The precursory signals of the 21 eruptions are quite varied, but they have common characteristics that can be used to identify them in real time and issue warnings to the local population. This has been practiced in Iceland since the time of the Krafla rifting episode in the seventies. The book-keeping is shown in **Table 1**. A pre-eruption warning was issued in 14 of the eruptions, a success rate of 67%. In four additional cases the precursory activity was detected on the instruments before the eruption was seen or verified, but after it began. In only three cases was the eruption seen before the instruments were checked.

The high detection rate is of great importance in a country where the observation conditions are not always perfect due to darkness and bad weather. It is, for example, common occurrence that a suspected eruption is reported, but subsequent checking of the monitoring networks shows no activity. The importance of such "negative warnings" should not be underestimated.

The term "false alarm" needs to be adressed. The precusory activity is generally ascribed to the propagation of a dike toward the surface. Quite frequently the dike does not reach the surface, and yet the signal is indistinguishable from that of a dike that does. The issued warning therefore has to include the possibility that the eruption may abort. Strictly speaking this is not a false alarm. A dike intrusion and its associated rifting may be just as destructive as an eruption. True false alarms have not been common in Iceland in the last 40 years. For example, the continuous inflation at Hrómundartindur in 1994–1998 and the year-long intrusion of Upptyppingar 2007–2008 never led to an alarm situation.

The short precursor times of Hekla eruptions are of particular concern. Two of the failure cases to issue warning are due to Hekla. In addition the beginning phase of the Hekla eruptions tends to be quite violent. Of the four most active Icelandic volcanoes Hekla is the only one that does not have a thick glacier cover. The rise of the eruptive column is therefore not delayed by the melting of the glacier cover and the column rises very fast.

#### TABLE 1 | Short-term precursory activity and precursory times.


W, warning issued before eruption.

D, detected first by instruments, then verified visually.

V, visual observation of eruption before instruments checked.

T, precursor time, from beginning of detected precursor until beginning of eruption.

∼, means that the timing of event is imprecise.

( ), means that the beginning of eruption is assumed.

This is of concern for overflying aircraft and groups of hikers on the flank of the volcano. There is heavy airtraffic over the summit of Hekla, mostly due to the unfortunate circumstance that the co-ordinate cross 64◦N and 20◦W, a commonly used waypoint for navigation, is immediately west of the volcano.

#### CONCLUSIONS


many respects. The long precursory time may have resulted from a complicated triggering scenario involving more than one magma chamber. The Bárðarbunga 2014–2015 eruption occurred at the distal end of a dike that took 13 days to propagate lateraly from its source until a suitable eruption site was reached.


and crustal deformation signals, but did not result in eruptions. These include inflation of the Hrómundartindur volcanic system in 1994–1998, intrusions into the roots of Eyjafjallajökull in 1994, 1999, and 2009, deep intrusion at Upptyppingar in 2007–2008, and several dike intrusions in the Krafla fissure swarms in 1975–1980.

9. Several cases of suspected but unconfirmed sub-glacial eruptions exist where no eruption is seen trough the glacial cover. Eruption-like seismic signals are then detected, often in connection with jökulhlaups, i.e., glacial outburst floods from volcanic areas. These include events at Katla volcano in 1999 and 2011, Grímsvötn in 1984, Hamarinn 2011, and several events at the Skaftá Cauldrons west of Grímsvötn.

#### AUTHOR CONTRIBUTIONS

The author confirms being the sole contributor of this work and approved it for publication.

#### REFERENCES


#### FUNDING

This study is the result of more than four decades of work on the seismicity of volcanoes and monitoring their activity. Funding for different aspects of the work has come from various sources, mostly the Icelandic government budget through the Science Institute, University of Iceland.

#### ACKNOWLEDGMENTS

This paper is a compilation of the work of many individuals involved with the monitoring of the Icelandic volcanoes and operation of the seismograph networks. The seismograms of **Figures 3**–**5** are from the stations at Reynihlíð, Skinnastaður, and Akureyri, stations attendants were Ármann Pétursson, Sigurvin Elíasson, and Gísli Ólafsson, respectively. Ásta Rut Hjartardóttir made **Figure 1**. The paper benefitted from the constructive comments by two reviewers and the editors.


episode at Eyjafjallajökull volcano, Iceland. Geophys. Res. Lett. 31:L14610. doi: 10.1029/2004GL020368


eruption and subsequent inflation. Geophys. Res. Lett. 30, 1182–1185. doi: 10.1029/2002GL016460


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

Copyright © 2018 Einarsson. 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 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.

## Single-Station Seismo-Acoustic Monitoring of Nyiragongo's Lava Lake Activity (D.R. Congo)

Julien Barrière<sup>1</sup> \*, Nicolas d'Oreye1,2, Adrien Oth<sup>1</sup> , Halldor Geirsson<sup>3</sup> , Niche Mashagiro<sup>4</sup> , Jeffrey B. Johnson<sup>5</sup> , Benoît Smets <sup>6</sup> , Sergey Samsonov <sup>7</sup> and François Kervyn<sup>6</sup>

<sup>1</sup> European Center for Geodynamics and Seismology, Walferdange, Luxembourg, <sup>2</sup> National Museum of Natural History, Walferdange, Luxembourg, <sup>3</sup> School of Engineering and Natural Sciences, University of Iceland, Reykjavik, Iceland, <sup>4</sup> Seismology Department, Goma Volcano Observatory, Goma, Democratic Republic of Congo, <sup>5</sup> Boise State University, Boise, ID, United States, <sup>6</sup> Royal Museum for Central Africa, Tervuren, Belgium, <sup>7</sup> Natural Resources Canada, Ottawa, ON, Canada

Since its last effusive eruption in 2002, Nyiragongo has been an open-vent volcano

#### Edited by:

Corentin Caudron, Ghent University, Belgium

#### Reviewed by:

Robin Matoza, University of California, Santa Barbara, United States David Fee, University of Alaska Fairbanks, United States

> \*Correspondence: Julien Barrière julien.barriere@ecgs.lu

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 28 February 2018 Accepted: 30 May 2018 Published: 19 June 2018

#### Citation:

Barrière J, d'Oreye N, Oth A, Geirsson H, Mashagiro N, Johnson JB, Smets B, Samsonov S and Kervyn F (2018) Single-Station Seismo-Acoustic Monitoring of Nyiragongo's Lava Lake Activity (D.R. Congo). Front. Earth Sci. 6:82. doi: 10.3389/feart.2018.00082 characterized by the world's largest persistent lava lake. This lava lake provides a unique opportunity to detect pressure change in the magmatic system by analyzing its level fluctuations. We demonstrate that this information is contained in the seismic and infrasound signals generated by the lava lake's activity. The continuous seismo-acoustic monitoring permits quantification of lava lake dynamics, which is analyzed retrospectively to identify periods of volcanic unrest. Synchronous, high-resolution satellite SAR (Synthetic Aperture Radar) images are used to constrain lava lake level by measuring the length of the SAR shadow cast by the rim of the pit crater where the lava lake is located. Seventy-two estimations of the lava lake level were obtained with this technique between August 2016 and November 2017. These sporadic measurements allow for a better interpretation of the continuous infrasound and seismic data recorded at the closest station (∼6 km from the crater). Jointly analyzed seismo-acoustic and SAR data reveal that slight changes in the spectral properties of the continuous cross-correlated low-frequency seismo-acoustic records (and not solely single events) can be used to track fluctuations of the lava lake level on a daily and hourly basis. We observe that drops of the lava lake and the appearance of significant long period (LP) "lava lake" events are a consequence of a probable deep lateral magma intrusion beneath Nyiragongo, which induces changes in its shallow plumbing system. In addition to contributing to understanding lava lake dynamics, this study highlights the potential to continuously monitor pressure fluctuations within the magmatic system using a single seismo-acoustic station located several kilometers from the vent.

Keywords: lava lake, Nyiragongo, infrasound, seismic, Synthetic Aperture Radar, single-station monitoring

### INTRODUCTION

Nyiragongo volcano in North Kivu (D.R. Congo) is among the most active volcanoes on Earth (Wright et al., 2015), with a persistent lava lake from at least between 1928 and 1977 and since 2002 (Smets, 2016). The two last effusive eruptions occurred in 1977 and 2002 and consisted of flank eruptions with high-velocity lava flows to the south toward the city of Goma (see **Figure 1A**).

**44**

During the eruption in 2002, which caused a major humanitarian crisis (Allard et al., 2002; Baxter and Ancia, 2002; Komorowski et al., 2004), the crater emptied of lava and its depth was evaluated between 600 and 800 m (e.g., Smets, 2016). About 4 months after the eruption, the crater filled up to reach a crater floor elevation of ∼3050 m above sea level. Since then, the lava lake remains at high level (i.e., level of the inner crater floor, ∼400 meters below the rim) with intermittent several decameter-scale rises or falls (see **Figure 1B**). As observed by Patrick et al. (2015) using tilt, GPS summit measurements and lava lake depth estimates from camera images for successive 2011 eruptive events on the East Rift Zone in Hawaii, Kilauea's lava lake acts as a piezometer for monitoring the magmatic reservoir. Inflation periods are associated with an increasing lake level while the deflation process due to magma escaping from the summit magma column (i.e., the eruptive event) is clearly visible through the drop of the lava lake level. A similar mechanism is thought to be responsible for the lateral drainage of magma from the lava lake at Nyiragongo during its last 2002 eruption (Wauthier et al., 2012). During short-term expeditions carried out since 2011, quantitative observations using close-range photogrammetry and a stereographic time-lapse camera (STLC) system have provided decisive insights into the recent lava lake level fluctuations of various amplitude and time scale (Smets et al., 2016). In order to use time-lapse camera observations in a monitoring context, the equipment requires a safe, permanent installation at the summit, adapted protections mostly against acid and humidity, and efficient data transfer capacity for realtime usage. At present, these requirements cannot be fulfilled at Nyiragongo and despite its strong potential, such a technique cannot yet be considered as a practicable monitoring tool for this volcano.

Monitoring seismicity is probably the most common geophysical application used at volcanoes and more recently, studying volcano infrasound has gained considerable interest (Johnson and Ripepe, 2011; Garcés et al., 2013 and references therein). Combining these two approaches could bring important constraints on the studied sources processes, i.e., the so-called seismo-acoustic sources generated in the vicinity of the boundary between the solid Earth and the atmosphere (Arrowsmith et al., 2010). It is now well-known that variations of (near-) surface activity at active volcanoes, either effusive or explosive, can be efficiently monitored by analyzing the coupled infrasound and seismic signals (e.g., Matoza et al., 2009, 2010; Ichihara et al., 2012; Richardson and Waite, 2013; Ulivieri et al., 2013). The seismicity characteristics of Nyiragongo and, more generally, of the entire Virunga Volcanic Province (VVP), are not very well known because of the lack of a permanent, local monitoring infrastructure until recently (Pagliuca et al., 2009; Oth et al., 2017). There is very little knowledge of the seismic activity preceding the two last eruptions in 1977 and 2002, such as sequences of tremor and volcano-tectonic activity recorded by only two analog seismometers in 2001–2002, thus preventing the understanding and the detection of similar pre-eruptive processes (Kavotha et al., 2002; Komorowski et al., 2004). The most significant seismic studies for the area in the past decade are: the analysis of teleseismic events associated to Nyiragongo after the 2002 eruption (Shuler and Ekström, 2009); a first multi-method approach encompassing crustal seismic velocity structure, volcano seismicity and seismic hazard assessment in the Kivu rift (Mavonga et al., 2010); a review article on multidisciplinary monitoring of pre- and co-eruptive processes at Nyamulagira, a neighboring highly active volcano, including data from one short-period digital seismic station and one analog seismometer (Smets et al., 2014); and passive and active seismic surveys during temporary experiments for studying the eastern Kivu rift structure on the Rwandan side (Wood et al., 2015).

The progressive deployment of a new local broadband seismic network, mostly between 2015 and 2017, has already provided some important and new insights into seismicity patterns accompanying Nyiragongo's activity (Barrière et al., 2017; Oth et al., 2017). These first studies have notably allowed for the detection and location of a continuous background, lowfrequency (<1 Hz) seismic tremor source at Nyiragongo volcano, most likely related to spattering activity of the persistent lava lake such as observed at Kilauea volcano (Patrick et al., 2016). For the specific case of lava lake volcanoes, variations of the lake level should influence the acoustic resonance in the upper air column within the crater. This means that a deeper lava lake would decrease the infrasonic dominant frequency, as suggested for instance at Kilauea by Fee et al. (2010) using an acoustic Helmholtz resonator model at Halema'uma'u crater. Other type of flow-induced resonance was described at the neighboring Pu'u O'o crater complex by Matoza et al. (2010). Recently, Richardson et al. (2014) showed that the lava lake fluctuations at Villarrica volcano may be inferred by the joint analysis of long-period seismo-acoustic single events and continuous tremor. This observation is therefore particularly relevant in the case of Nyiragongo for studying the activity of its lava lake. Since late 2016, the increased number of highly similar, discrete, long-period (LP) seismic events and their infrasound counterparts have accompanied decameter-scale oscillations of the lava lake level, which relate to major pressure changes in the upper magmatic system and at least one probable intrusion event. In this study, we will exploit seismic and infrasound records collected at a single station (KBTI, see **Figure 1A**) in order to extract a reliable seismo-acoustic signature of the lava lake activity, the latter being mostly characterized by its level fluctuations estimated using a SAR-based method (**Figure 1C**). In addition to lava lake estimates on a daily basis, we show that the presented approach also allows for a rapid quantification of strong variations of the lava lake level on an hourly time scale.

#### MATERIALS AND METHODS

#### Nyiragongo's Lava Lake Level Derived From High-Resolution Satellite Observations

No continuous and reliable terrestrial observations of the lava lake are available from the rim of the Nyiragongo main crater; however, satellite radar data may be used to recover the lava lake level. The idea behind this method consists of measuring

the length of the SAR (Synthetic Aperture Radar) shadow cast on the lava lake surface by the edge of the pit crater hosting the lake (d'Oreye et al. in prep), similar to studies focusing on the retrieval of building shapes using SAR shadow (e.g., Bolter, 2000; Tison et al., 2004). The shadow length is measured in pixels on the amplitude (module) of the SAR image in radar geometry and converted to distance in meters, based on the spatial resolution of the image. The level of the lake with respect to the rim is then obtained by multiplying the shadow dimension and the cosine of the incidence angle for the SAR sensor's orbital path.

The intensity of the SAR amplitude may vary significantly depending on the surface reflectance and geometry. Activity at the surface of the lava lake, when animated by fountaining or intense convections, may blur the contrast between illuminated and shadow zones. To minimize that effect, the length of the shadows is measured along an average of at least 10 amplitude profiles in the range direction. The detection of the edges of the shadow is made by fitting a 3 steps Heaviside function to that average amplitude profile. The method has been validated with in situ measurements (e.g., Smets, 2016) and results are consistent with ground based observations (d'Oreye et al., 2017).

In the present study, we measured lava lake levels using RADARSAT images acquired in Fine (RS-F2F) and Ultra Fine (RS-UF) mode along descending and ascending orbits respectively and COSMO-SkyMed (CSK) images acquired in spotlight mode along ascending orbits (**Figure 1C**). The range resolution of these images is 4.73, 1.33, and 1.25 meters per pixel and look angles are 41.75, 37.38, and 38.34 degrees at zone of interest, respectively. We used a total of 72 SAR images spanning 2 August 2016 to 20 November 2017.

The uncertainty of the method is typically 1 pixel in normal conditions. However, an eruptive cone, which has opened in March 2016 at the bottom of the Nyiragongo main crater, 190 m east of the lava lake (see **Figure 1B**), has intermittently emitted lavas that flow from the cone back into the pit. This has two effects: the bottom of the main crater is higher on the side of the cone, and edges of the rim where lavas flow into the pit are rounded and smoothed instead of sharp. As a consequence, the asymmetry of the platform induces a 10-m lake level offset between measurements in ascending and descending orbits. A second effect is decreased contrast between illuminated and shadow areas, which leads to an underestimation of the length of shadow because it is cast from a lower point. This effect only concerns RS-F2F images acquired along descending orbit and cannot be mitigated. The comparison with ground observations shows that these measurements (dark diamonds in **Figure 1C**) are probably slightly underestimated by a few meters from the second half of 2017 (June-July to November).

Due to the SAR reflectance characteristics, or layover effects (Massonnet and Feigl, 1998; Pinel et al., 2014), it is conceivable that signal backscatter from the opposite crater wall can be received prior to that from the lava lake surface. Given the diameter of the pit crater hosting the lava lake and the average incidence angle of the SAR signals, the layover will prevent the measurements of depths larger that roughly 110 m below the bottom of the crater. So far, we have never observed such a depth, although it nearly reached it in November and December 2016.

#### Co-located Seismic and Infrasound Equipment Near Nyiragongo Volcano

The seismological station KBTI used in this study (**Figure 1A**) is part of the cross-border (D.R. Congo, Rwanda, Burundi) regional network KivuSNet (see Oth et al., 2017 for more details about the network). This site is equipped with a broadband seismometer (Nanometrics Trillium Compact 120 s−100 Hz) and a smallaperture array (20 m) of three infrasound sensors manufactured by Boise State University (e.g., Johnson and Ronan, 2015). These infrasound sensors use MEMS pressure transducers and capillary filters providing an approximate in-band frequency response that is flat between 0.04 Hz and Nyquist (Marcillo et al., 2012). In this study, we use both locally archived and transmitted seismic and infrasound data at sample rate of 50 Hz. Data are transmitted in real time by cellular data network, then acquired and archived using the SeisComP3 software, which is installed at local and international partner institutions (Goma Volcano Observatory— GVO, R.D Congo and European Center for Geodynamics and Seismology—ECGS, Luxembourg).

KBTI is the closest station to the volcano summit crater (ground distance of about 6 km to the lava lake) where a seismometer was installed in September 2015. The closest site prior to this date was located about 17 km away at the volcano observatory in Goma, the most populated city in the region with about 800,000 inhabitants (Michellier, 2017). The station KBTI is located at a ranger station of the Virunga National Park, which is the start of the track into the jungle leading to the summit. KBTI is the closest permanent safe site to the volcano summit where geophysical instruments can be deployed. Following 10 months of successful seismic data recording and real-time transmission between September 2015 and July 2016, an array of three infrasound sensors was further installed in August 2016 on the same datalogger (Nanometrics Centaur). Permanent telemetered equipment at the summit of Nyiragongo is under consideration, but technical and safety requirements remain a challenge.

#### Identification of Repetitive Long-Period (LP) Seismo-Acoustic Events

Since the full deployment (seismometer and infrasound sensors) of the station KBTI, we have detected a family of long-period seismic events generally associated with detectable infrasound transients, such as the example in **Figure 2**. As observed on the time-frequency representation, this seismic LP event and its associated infrasound signal have a dominant frequency around 0.5 Hz. This coupled seismic-acoustic signal suggests a common source near the surface, potentially coming from Nyiragongo's crater and attributable to the lava lake activity. This is supported by the retrograde ellipse motion of the seismic signals, which is typical of surface Rayleigh waves (**Figure 3**). **Figure 4** shows two examples of back-azimuth determination of infrasound events at KBTI using the array of three sensors. We use a 2- D time-domain Fisher analysis as beamforming technique (e.g., Assink et al., 2008) consisting of a grid-search over the slowness vector of infrasound waves propagating across the array. This standard approach allows for the retrieval of the back-azimuth and apparent velocity of the event for the highest signal-tonoise power ratio, defined in this context as SNR<sup>2</sup> = (F-1)/N, N being the number of sensors and F the Fisher ratio. The LP seismo-acoustic event depicted in **Figures 4A,B**, which shares similar characteristics as the one displayed in **Figure 2** (i.e., dominant frequency around 0.5 Hz, general waveforms shapes, delay time between seismic and infrasound arrivals of about 10– 15 s), occurred in the direction of Nyiragongo. For comparison

Nyiragongo's lava lake. (B) Particle motion for the time window highlighted by the color scale blue-to-red in (A).

FIGURE 4 | (A) LP seismic event and (B) associated LP infrasound event (best beam) in the dominant frequency band (0.3–0.9 Hz). Each round marker corresponds to a sliding window of 5 s where time domain Fisher analysis is performed. Overlapping is set to 90% and the size/color of the marker is proportional to the SNR<sup>2</sup> . Signals from the three channels (zoom in the coherent part for IS1, IS2 and IS3) and the corresponding polar plot to the infrasound best beam (back-azimuth and apparent velocity) are depicted below. (C,D) Same as (A,B), respectively, for an atmospheric explosion of human origin. Here, the duration of the sliding windows is 1 s. A high cutoff frequency of 5 Hz was chosen because it generally improved the detection capability of such an event.

with other coherent signals detected at KBTI, **Figures 4C,D** shows the analysis of an atmospheric explosion occurring at a different azimuth, which is characterized by a more impulsive waveform and very attenuated or weak seismic coupling. This last event, which was also audible and signalized by some of the co-authors during a descent from Nyiragongo's summit on 20 June 2017, was certainly of human origin while the previous one (**Figures 4A,B**) was originating from Nyiragongo's activity. We know from field observations that none of numerous fractures and adventive cones located along the direction of Nyiragongo are active (Smets, 2016). Although other sources cannot yet be definitively excluded based on this beamforming analysis only, such seismo-acoustic events are most probably related to Nyiragongo's lava lake. The three infrasound sensors at KBTI have only worked simultaneously on rare occasions due to several technical issues (e.g., lighting strikes, dysfunctional digitizer channels or sensors, destructions by monkeys, invasion by insects), a fact that prevented the continuous detection and back-azimuth determination of LP infrasound events since the deployment of the sensors. In the following analysis, we will use one in three channels (IS3 in **Figures 2B**, **4B,D**), which has almost always been fully functional and can be analyzed in association with the seismic record.

Similarly to Matoza et al. (2009) at Mount St. Helens or Richardson and Waite (2013) at Villarica volcano, we look for similar events using a 3-components template matching procedure with the LP seismic event presented in **Figures 2A**, **3** as master event because it is visually representative of other LPs and has high signal-to-noise ratio. We select a window length of 24 s (corresponding to the interval [6:30] s in **Figure 3A**) and a threshold of 1.8 for the normalized correlation coefficient (i.e., 0.6 on average per channel). Almost 2500 highly reproducible seismic events are detected with this technique (**Figure 5A**). It is important to note that the first 500 events are detected during the first year (14 Sept. 2015 to 13 Nov. 2016), while about four times more occurred during the following year (13 Nov. 2016 to 20 Nov. 2017). Since the master event is associated with an infrasound counterpart, we also display the infrasound records for each detected time window in order to verify whether an infrasound event is consistently associated with such typical seismic LP events (**Figure 5B**). Single events with clear maxima are detected for the first hundred traces, then the infrasound LPs are less detectable with respect to the noise floor but still noticeable at a roughly constant timing about 10 s after the seismic arrival.

The seismic and acoustic signals propagate in different media (solid Earth vs. atmosphere) and hence arrive at different times at KBTI. The changing activity of the lava lake, manifested by lake level variations (discussed later), is a possible explanation for the timing variation between seismic and infrasound phases, however it is unlikely to be solely responsible for the ±0.5 s observed variation for an estimated propagation time of about 20 s (see next section Identification of Repetitive Long-Period (LP) Seismo-Acoustic Events). The atmosphere between Nyiragongo and KBTI is prone to strong wind and temperature variations, similar to the situation at other volcanoes worldwide (e.g., Matoza et al., 2009; Lacanna et al., 2014), and reasonably expectable effective sound speed variations of about 2.5% (i.e., 330–340 m/s) would be in accordance with observations by Johnson et al. (2012) at Villarica volcano 8 km away from the crater. In the following section, we demonstrate that the mean delay time observed between the seismic and the infrasound arrivals is consistent with a source at Nyiragongo's crater by performing a 2-D numerical simulation of the elastic and acoustic wave propagation for a transect passing through the station KBTI.

#### 2-D Numerical Simulation of the Seismo-Acoustic Propagation From a Point Source at Nyiragongo's Lava Lake

Seismo-acoustic source mechanisms at volcanoes are complex and numerical modeling can help to better understand such processes, as done for instance by Matoza et al. (2009) while studying the shallow source of coupled seismo-acoustic signals at

for the waveforms and the right ordinate is the event number index (2,447 in total). (B) Associated infrasound LP records for each time window where a seismic LP is detected in (A). Infrasound signals have clear maximum amplitude for the first hundred events, then the infrasound LP events are less detectable with respect to the noise floor but still noticeable at a roughly constant timing (between 20 and 25 s). For both panels, amplitudes are normalized trace by trace and the stack of all traces (band-pass filtered in the dominant frequency band 0.3–0.9 Hz) is indicated on top.

Mount St. Helens. The aim of the wave propagation simulations presented below is not to create a model as realistic as possible of the acoustic and elastic wave fields from a point source at Nyiragongo's crater (i.e., a bubble explosion). This would notably require a more complex 3-D simulation with a detailed meshing of the volcano edifice, differentiating the lava lake from the surrounding material, which is far beyond the scope of this article. The simulation presented here should rather be viewed as an illustrative validation of a seismo-acoustic delay time consistent with the location of the source in Nyiragongo's crater (**Figure 6**).

We used the 2-D Specfem2D numerical code based on the spectral element method (Komatitsch and Vilotte, 1998; Tromp et al., 2008). A 20 × 20 km computational domain is gridded using 200 × 200 elements (641601 grid points) and PML (Perfect Matched Layers) absorbing boundary conditions. The 2-D topography (except inside the crater) is estimated from the NASA SRTM-3 DEM (Digital Elevation Model) with a resolution of about 90 m (Farr et al., 2007). The crater geometry is roughly modeled considering the mean diameter (1.2 km) and the depths of the two main inner platforms (see **Figure 1B**). Homogenous media are assumed for the solid (elastic) Earth

(V<sup>P</sup> <sup>=</sup> 1,800 m.s−<sup>1</sup> , V<sup>S</sup> <sup>=</sup> 1,050 m.s−<sup>1</sup> , ρ = 2,000 kg.m−<sup>3</sup> ) and for a non-advecting atmosphere (V<sup>P</sup> <sup>=</sup> 340 m.s−<sup>1</sup> , V<sup>S</sup> <sup>=</sup> 0 m.s−<sup>1</sup> , ρ = 1.2 kg.m−<sup>3</sup> ). The rather low velocity values chosen for the Earth correspond to typical subsurface soft material and are in accordance with the travel-time curve obtained by Barrière et al. (2017) for surface waves generated by the continuous tremor at Nyiragongo. We define a line of 12 acousticseismic receivers spaced 500 meters along the topographic profile between the summit and KBTI. Each synthetic acoustic and seismic sensor is positioned 10 m above and below the free surface, inside an acoustic and an elastic element respectively.

We assume that the infrasound and seismic sources are co-located and conjoint in time. We perform two simulations for recording acoustic pressure and vertical seismic velocity using a source 50 m above the free surface (infrasound explosion) and 50 m below (seismic coupling to the lava lake's inner wall), respectively (**Figure 6B**). Decoupling both acoustic and elastic sources with two distinct simulations mitigates effects of epicentral or local air-ground conversions, but their close proximity maintains the timing and waveform shapes of the maximum amplitude seismic and acoustic arrivals (Rayleigh waves and direct sound waves, respectively). For both cases, the source time function is the time derivative of a Gaussian pulse with a central frequency of 0.5 Hz and located at the center of the lava lake. The sources correspond to a monopole pressure (infrasound) or a single force horizontally oriented (seismic).

Despite the fact that no clear onsets are detectable for the arrivals of the stacked (observed) seismo-acoustic LPs (**Figure 6A**), it seems that the delay time relative to the simulation is slightly longer (about 1–2 s) at KBTI. An efficient method for calculating the time shift between two related signals (here seismic and infrasound transients) is the computation of the cross-correlation function (CCF), which gives an estimation of the delay time at its maximum value. The shift of about 1–2 s is indeed confirmed by comparing the CCFs of the observed and simulated seismo-acoustic LPs (**Figure 6C**). However, considering the observed timing uncertainty as well as the simplified 2-D model (e.g., error in presumed sound or seismic wave speeds), the synthetic records provide a good estimation of the mean time shift between seismic and infrasound arrivals and confirm the origin of these seismo-acoustic LP events at Nyiragongo's lava lake. We will now compare their characteristics with the lava lake activity, which can be primarily described by its level fluctuations as estimated with the SAR shadow technique.

#### RESULTS

#### Occurrence and Amplitude Properties of the Seismo-Acoustic LP Events

LP events occurrence, as detected at station KBTI, is correlated with the lava lake level estimations obtained with the SAR processing (**Figures 7A,B**). A strong increase in LPs is observed in November 2016 and correlates well with the biggest drop (about 80 meters) of the lava lake level. This lake level drop occurs very rapidly, i.e., in few hours (see section Discussion: a Relevant Short-Term Unrest Indicator). Sustained LP seismicity then accompanies the rise of the lava lake over several months. The last months between October and November 2017 are characterized by low LP seismicity and high lava lake level, which suggests similar behavior to the activity observed more than 1 year prior to the main November 2016 drop. Because the seismic LP events are repetitively associated with infrasound transients (**Figure 5**) and their occurrence exhibits a relationship with the lava lake level fluctuations (**Figures 7A,B**), we assume the same non-destructive surface source process as inferred by Richardson and Waite (2013) for Villarrica volcano, which also hosts a permanent lava lake. At Villarrica, large bubble bursting (i.e., the infrasound explosion) is thought to induce drag forces at the lava lake's walls responsible for seismic LP events. Such typical surface activity has been observed several times at Nyiragongo's lava lake during field expeditions (Smets, 2016; Smets et al., 2016).

Histograms of event statistics, including peak seismic and infrasound amplitudes for the detected LP events, show significant trends (**Figures 7C–E**). The seismic amplitudes tend to slightly increase after the November-2016 drop, but the opposite trend is even more clear with the infrasound amplitudes. In other words, the swarm of seismo-acoustic LPs in November 2016 (more than 200 events between 13 and 14 November 2016) presents the highest acoustic-to-seismic (AS) amplitude ratio of the entire time period and then a decrease of this ratio is observed.

We interpret this changing AS amplitude ratio as analogous to the volcano acoustic-seismic ratio (VASR) parameter introduced by Johnson and Aster (2005). This non-dimensional parameter allows for an efficient characterization of some seismo-acoustic source properties at volcanoes. Johnson and Aster (2005) obtained high and stable VASR values at Erebus volcano and attributed this observation to the lava lake activity associated with repetitive source processes at the surface of a stable lava lake. The observed decrease of the AS amplitude ratio in the case of Nyiragongo is not in agreement with such a constant trend and is better explained by variable VASR as observed at Villarrica volcano by Richardson et al. (2014) during fluctuating lava lake episodes.

The highest rate of LP events and the highest AS amplitude ratio during the November 2016 drop could thus be interpreted as larger and more frequent bubble bursts in the lava lake. The ensuing decrease of the AS amplitude ratio might be related to a reduction of the spattering surface activity. This could be due to the decrease of bubble size as well as the deepening of the seismo-acoustic LPs within the lake, which in turn explain a better seismic coupling and a less effective acoustic transmission through the atmosphere. For the same period, the rate of seismo-acoustic LP events remains moderate (>1 and <25 events/day); however, the occurrence rate of LP events seems to decrease at the time of writing (late 2017-early 2018) and could reflect the return to a more quiet lava lake activity and high level.

FIGURE 7 | (A) Estimated lava lake depth from SAR processing using COSMO-SkyMed (CSK) and RADARSAT (RS-F2F and RS-UF) images. "Asc." and "Desc." refer to ascending and descending orbits respectively. (B) Temporal distribution of LP events detected by template matching (see Figure 5). The time series is divided into three consecutive time windows (blue, green, red) where maximum (zero-to-peak) seismic and infrasound amplitudes are plotted as histograms in sub-panels (C–E) respectively. The gray shading on the histograms (C) (blue) and (E) (red) correspond to the beginning and the end of the timeline respectively, both characterized by low LP seismicity (see B).

**51**

The amplitude analysis provides some important insights into the variations of the lava lake activity. Two further attributes are of significance with respect to the lava lake level, namely the delay time of the infrasound arrival with respect to the seismic one and the frequency content of the infrasound events (Richardson et al., 2014; Johnson et al., 2018). Unfortunately, the delay time information, which is potentially useful for measuring the level of lava in a conduit (Ripepe et al., 2002; Johnson, 2007) is particularly affected by 6 km of propagation through the atmosphere (see section Identification of Repetitive Long-Period (LP) Seismo-Acoustic Events). Moderate variations of the lava lake depth on the order of 80 m are probably not resolvable with the current seismo-acoustic station.

A more promising means to measure lava levels is afforded by the study of the acoustic resonance of the air column above the lava lake (Johnson et al., 2018). Several acoustic resonance models have been applied to other lava lake volcanoes (e.g., Bessel horn resonator at Villarrica in Richardson et al., 2014; Helmholtz resonator at Kilauea in Fee et al., 2010), the simplest one being the cylindrical geometry, or organ pipe, that could eventually fit to the shape of Nyiragongo's lava lake. However, the 1.2 km wide Nyiragongo crater is significantly more complex than a pipe or other standard geometries of resonators (see **Figure 1B**). In the following analysis, we do not intend to propose a physics-based model for the infrasonic frequency content but rather aim to use its variations as a direct proxy for the lava lake level since our expectation is that a deeper lava lake will lead to a graver infrasound dominant frequency.

Notably, we do not find any clear evidence of changing spectral properties of the LP infrasound events with respect to the fluctuations of the lava lake, even for the marked drop of about 80 meters in November 2016. This may imply that the variations during the studied time period are potentially too weak to be detected or have negligible effect on the resulting infrasound signal recorded at KBTI. On the other hand, these discrete acoustic LP events are not fully representative of the lava lake activity and have generally a low signal-to-noise ratio (see **Figure 5B**). At Villarrica volcano, Richardson and Waite (2013) suggest that the tremor recorded at greater distance from the volcano results from the superposition of single LPs distorted by path effects, which we consider also as the process responsible for the Nyiragongo continuous tremor as it shares the same dominant frequency band below 1 Hz (Barrière et al., 2017). Therefore, a more reliable acoustic signature from Nyiragongo's lava lake is potentially the continuous low-frequency seismic tremor associated with its spattering activity. We now investigate the infrasonic component of this tremor, which should be detectable at KBTI as well.

#### The Continuous Nature of the Seismo-Acoustic "Lava Lake" Source

Examples of seismic and infrasonic time records and corresponding spectrograms for three successive days in April 2017 at KBTI are displayed in **Figure 8**. No LP event was detected by template matching during this period characterized by a high lava lake level estimated around 20–30 m deep (see **Figures 1C**, **7A**). On the seismic spectrogram (**Figure 8A**), the secondary microseismic peak is clearly dominant between 0.1 and 0.3 Hz. Above 0.3 Hz, the continuous harmonic volcanic tremor is well-detected, preferentially below 1 Hz among all seismic stations of the KivuSNet network (Barrière et al., 2017; Oth et al., 2017) but it could be also detectable until 1.5 to 2 Hz as observed here for KBTI. Higher frequency contents until 4–5 Hz are discernable during night periods. However, as stated previously, for security reason, KBTI station is located at a ranger station, close to a road in an inhabited area and only few km north of the heavily populated city of Goma. The effects of anthropogenic activity are clearly visible on both seismic and acoustic records. Such undesirable seismic noise dominates above 2 Hz during daytime (**Figure 8A**) and the infrasound records exhibits also high-amplitude human-related signals, which are broadband (i.e., cover the frequency band of interest below 1 Hz) and would strongly disturb the detection of any other signals of volcanic origin (**Figure 8B**). This is a critical drawback for the detection of potential infrasonic tremor from Nyiragongo's lava lake activity. Moreover, no clear patterns of

continuous signals are detectable on the infrasonic spectrogram, which means that the acoustic component of the volcanic tremor recorded at KBTI would be near the noise floor such as observed for most of the discrete LP infrasound events (**Figure 5B**).

Following these observations, we highlight the continuous nature of the seismo-acoustic signal by calculating the crosscorrelation function (CCF) between the co-located seismic and infrasound waveforms at KBTI. A similar coherence analysis was for instance performed by Cannata et al. (2013) using cross wavelet transform for detecting explosive activity at Etna volcano. The principal difference with the present study lies in the usage of stations much closer to the active vent (∼1–1.7 km) and the strong intensity of the explosive activity, the combination of which leads to high signal-to-noise ratio for both seismic and infrasonic records. In our case, even if the infrasound signal of interest is of very low amplitude, such a cross-correlation approach should allow to extract the acoustic component coherently associated with the seismic signals originating from the lava lake (**Figure 9**). The seismo-acoustic CCF is calculated for non-overlapping 1-h intervals and averaged over 1 day in the frequency band [0.1–10] Hz. Because the full records are crosscorrelated without discarding undesirable noise, this approach leads to very low normalized correlation maxima (order of 10−<sup>2</sup> on average); however, the daily averaging generally allows for the quantification of robust time lags (**Figure 9A**). The CCF daily maxima are found at a stable time shift of 10 s between seismic and infrasound arrivals in the dominant frequency band [0.3–0.9] Hz (**Figure 9B**). This is consistent with the simulation results (**Figure 6C**) and the good match between the CCF of the stacked seismo-acoustic LP events (LPs CCF) and the daily tremor CCF strongly suggests a similar source for LPs and tremor at Nyiragongo's lava lake (**Figure 9C**).

Using these cross-correlation functions, we indirectly analyzed the variations of the infrasonic tremor frequency through analysis of the variations of the cross-spectrum between the seismic and infrasonic signals. Considering the remarkable stability of the seismic tremor source over the years (Barrière et al., 2017), we make the assumption that the frequency content of the seismo-acoustic cross-spectrum will be primarily affected by the infrasound component of the fluctuating lava lake

acoustic arrival) are an indication of continuous tremor originating from Nyiragongo (left: amplitudes normalized trace by trace; right: no normalization and zoom in the maxima). (B) Wavelet transform of the stacked CCF for the entire period highlighting the maximum correlation at −10 s. The color scale corresponds to wavelet coefficients (magnitude), from low (white-blue) to high values (red). The high-frequency correlation at zero lag is most likely due to coherent noise sources and air-ground coupling at KBTI. (C) Stack of the band-pass filtered CCF (dark gray line) in the dominant frequency band (0.3–0.9) Hz inferred from the time-frequency representation in (B).

activity. It is important to note that the infrasonic spectrum might be potentially influenced by path effects, such as those due to changing atmospheric conditions between day and night in atmospheric boundary layers (Fee and Garcés, 2007). However, such diurnal variations that can affect the acoustic propagation appear to be significant only at greater distance (in the diffraction or shadow zone) than for our station at ∼6 km and the dominant lava lake signature contained in the infrasonic frequency spectrum would be preserved such as observed by Johnson et al. (2012) at Villarrica volcano. Nevertheless, we do note diurnal variations of the spectral amplitude, which are due to anthropogenic activity dominantly observed during the day (see **Figure 8**). The signature of this noise is visible on the time-frequency representation of the seismo-acoustic CCF displayed in **Figure 9B** at zero lag (>2 Hz, blue color). In this case, there is no delay time between the infrasound and the seismic coherent arrivals because of direct coupling of the acoustic waves to the ground. Similar coupling at the recording station has also been observed for explosive sources at volcanoes (Ichihara et al., 2012; Matoza and Fee, 2014). In the case of Nyiragongo, it is clear that the dominant low-frequency seismic signal, which precedes the correlated infrasound by ∼10 s, does not come from local air-ground coupling, but originates from the propagation of Rayleigh waves generated in Nyiragongo's crater. We next analyze the spectral content of seismo-acoustic CCF and relate it to lake depth estimated by SAR processing.

#### Calibration Model Between the Lava Lake Depth and the Seismo-Acoustic Cross-Spectrum

The SAR-inferred lava lake depths are irregularly sampled (separated by 1–31 days), so they are linearly interpolated onto a regular daily time frame to compare with CCF spectral content. SAR data are then averaged over a ±10 days centered time window (**Figure 10A**), which incorporates at least 3 SAR images for most depth measurements (see later **Figure 11A**). For the calculation of the seismo-acoustic parameter (mean frequency of the daily CCF), a similar moving average is applied. The power spectra of each CCF are then calculated using a fixedsize time window for all selected days (**Figure 10B**). We use a 10% cosine tapered window of width determined automatically by selecting 75% of the total signal energy of the stacked tremor CCF displayed in **Figure 9C** (start at 12.5%, end at 87.5%). A correlation between CCF spectral content (lower frequencies) and deeper SAR-derived lava lake levels is apparent (**Figure 10C**).

The comparison between SAR estimates of the lava lake depth and the CCF mean frequencies shows a systematic relationship permitting an empirical calibration model (**Figure 11A**). We derived two best-fit power laws (in the least square sense) for a subset of the data (subset 1 = the first 19 lava lake depths) and for the full dataset (subsets 1 + 2 = 72 measurements). The subset 1 was chosen to be as limited as possible but also able to sample the full range of depths measured with the SAR technique (**Figure 11B**). In this way, we verified that this first test model (i.e., subset 1) gives similar predictions to the complete one (i.e., subsets 1 + 2), and thus that the empirical relationship between the lava lake depth and the seismo-acoustic mean frequency is applicable to additional new data (**Figure 11B**). The uncertainties of both models lie into a similar interval and are expressed as the standard deviation error of about ±10 m (**Figures 11A,B**). Overall, the daily depth estimates from the seismo-acoustic CCF mean frequency are highly consistent with the SAR observations.

The main discrepancy occurs for less well constrained SAR measurements made since June-July 2017 (see also section Nyiragongo's Lava Lake Level Derived From High-Resolution Satellite Observations). These points notably contribute to the minor difference between the two calibration models previously obtained at the highest frequencies, i.e., the shallowest lava lake levels (see **Figure 11A**). The fact that few SAR estimates are available for this period and acquired with the same satellite (RADARSAT) could partly explained this bias by considering these measurements as potential outliers. However, both descending and ascending modes are used and show consistent estimations of the lava lake level. Therefore, we rather tend to explain this discrepancy due to weak infrasound amplitudes (see section Daily Monitoring of Nyiragongo's Lava Lake Activity below). These slightly biased measurements represent a weak proportion of the available depth estimates and, for this reason, we chose the best fit curve using the full dataset as final calibration model (i.e., subsets 1 + 2).

### Daily Monitoring of Nyiragongo's Lava Lake Activity

Nyiragongo's activity since August 2016 may be partly summarized in terms of lava lake depth estimated from its seismo-acoustic signature and the continuous AS (acoustic-toseismic) amplitude ratio recorded at KBTI (**Figure 12A**). The relevance of this last parameter for characterizing the varying lava lake activity was discussed earlier during the analysis of the single LP events (section Occurrence and Amplitude Properties of the Seismo-Acoustic LP Events and **Figure 7**). Because the continuous infrasonic records at KBTI are disturbed by undesirable noise during daytime, we apply a similar smoothing procedure as the one applied for the seismo-acoustic CCF mean frequency for calculating a robust daily AS ratio. One-hour seismic and infrasonic RMS amplitude values are computed (with 50% overlapping) and smoothed over ±10 days using a median filter. In **Figure 12A**, this smoothed AS ratio is plotted together with the lava lake depth estimates. The highest AS ratio (i.e., high infrasonic amplitudes relative to the seismic ones) is obtained after the November 2016 drop when the lava lake reached its lowest levels. This conveys a strong spattering activity associated with high amplitude infrasound tremors. Then, as observed with the analysis of LP events (**Figure 7**), the acoustic component of the volcanic tremor exhibits lower amplitudes through time, notably for the last period starting in July 2017, indicating a progressive return to a quieter behavior. The lowest AS ratio is obtained for the period where the disagreement between SAR and seismo-acoustic depth estimates is the most significant. This effect is most likely responsible for the poorer characterization of the resonant frequency information retrieved at the station KBTI. This observation points out that, for obvious

FIGURE 10 | (A) Lava lake depth estimates using three sets of satellite images CSK, RS-F2F and RS-UF (round markers and dotted line, see also Figure 1C or Figure 7A). The initial estimations are irregularly sampled and have been linearly interpolated onto a regular daily frame, and then final depth estimates are obtained by applying a centered averaging window of ±10 days (colored squares). The color scale indicates relative lava lake depth. (B) Seismo-acoustic CCF for each selected day associated with the squares in (A) and using the same color code. The cosine-tapered window used for the computation of power spectra (C) is depicted as a solid black line. (C) Corresponding power spectra of the selected CCF waveforms in (B) plotted using the same color code.

FIGURE 11 | (A) Relationship between the lava lake depth estimated from SAR processing and the seismo-acoustic CCF mean frequency. The two best-fit power laws using the first 19 depth estimates (subset 1, round markers) and the full dataset (subsets 1+2, round and square markers) are represented as dashed and solid thick black line, respectively; gray area and dashed thin lines indicate estimates of one standard deviation error, respectively. Both SAR depth estimates and seismo-acoustic CCF mean frequency are averaged using a ±10 days centered moving average. Since the initial sampling of the SAR images is irregular (i.e., not daily), the number of images actually used for the averaging is also indicated (see text for details). (B) Daily seismo-acoustic "depth estimates" of Nyiragongo's lava lake computed using the two best-fit models in (A). The gray shading indicates estimates of one standard deviation error for the final calibration model (solid line, subsets 1 + 2).

reasons, the seismo-acoustic method applied here needs a proper recording of the seismic and infrasonic tremors for estimating the lava lake level. In order to mitigate the effect of varying AS ratio (which conveys nonetheless some valuable information), this approach would perform best if these signals have high SNR, thus ideally recorded at a station as close to the crater as possible.

The monitoring of Nyiragongo's activity is also characterized by the occurrence of LP events associated to large bubble bursting (**Figure 12B**), and the number of seismic events located deeper than 5 km below sea level and within a 15 × 15 km area around the volcano (**Figure 12C**). This deep and sustained seismicity has been continuously detected since the complete installation of the KivuSNet seismic network in October 2015, notably a batch

estimates are plotted with round markers. The smoothed AS ratio (±10 days) is plotted as a black line (right y-axis). (B) Daily count of LP events detected at KBTI (extended time series presented in Figure 7B). The color scale corresponds to the number of events per day. (C) Daily count of deep events at Nyiragongo (deeper than 5 km b.s.l). The gray areas correspond to periods when the number of available stations in the Virunga Volcanic Province (VVP) is smaller than 8.

of seismicity on the southeast flank of Nyiragongo. Although no detailed event classification has yet been performed, these seismic events generally have characteristics of volcano-tectonic earthquakes (e.g., higher and broader frequency content than LP events, more clear P or S wave onsets). Volcano-tectonic locations are obtained using a "picking-free" cross-correlation based technique adapted to low SNR events (see Barrière et al., 2017) and a simple low-resolution 1-D velocity model with two layers between 0 and 30 km depth for the Virunga Volcanic Province (VVP) (Mavonga et al., 2010). Precise hypocenter determination is adversely impacted by the use of this rough velocity model and relatively poor station coverage around the volcano (see Oth et al., 2017). A daily count is tabulated when at least 8 stations located in the VVP are available.

The time series extending to January 2018 shows a rather stable high lava lake level (20 to 30 meters below the inner crater floor) and low LP seismicity during the last few months. As previously highlighted in **Figure 7**, the first significant correlation is the link between the November 2016 drop and the swarm of LP events. Since the lava lake level drop there has been sustained (∼1 year) elevated LP activity, which became the new normal background seismicity. More interestingly, this drop is also accompanied by strong seismic activity at depth. A similar observation can also be made for the two other most significant swarms, with drops of 15 to 20 meters occurring in March and end of May 2017 (see SAR estimates), but without clear changes in LP seismicity. Taken altogether, this evidence suggests that the swarms of deep volcano-tectonic events are likely associated with magma intrusion, which simultaneously results in the drawdown in the level of the lava lake. However, the time resolution is unfortunately insufficient to definitely resolve the process taking place at that time.

#### Hourly Estimates of Nyiragongo's Lava Lake Depth

A detailed inspection of the dynamics in November 2016 is required in order to better understand the link between the two seismic modes (LP and deep) and the lava lake level. From a monitoring perspective, the ability to rapidly quantify the lava lake level is also of crucial importance. We consider whether it is possible to implement a short-term (near realtime) estimation of the lake depth by computing seismo-acoustic CCFs every 6 min for hourly time windows and 90% overlap. **Figure 13A** shows the depth estimates based on this short-term

FIGURE 13 | (A) Hourly estimations of the lava lake depth around the main drop of the lava lake (8–15 November 2016), based on seismo-acoustic CCFs calculated each 6 min for a 1–h lagging time-window. The top panel represents the maximum values of the seismo-acoustic CCF (i.e., correlation coefficient between the seismic and the infrasound signals, values above 10−<sup>2</sup> are in red). Bottom panel: estimates of the lake level as well as the available SAR measurements for this period. Gray dots correspond to each single calculation (i.e., each 6 min). The solid line is the 6–h lagging average (i.e., 60 calculations). Estimates corresponding to the highest values of seismo-acoustic correlation (i.e., above 10−<sup>2</sup> ) are highlighted in red. The dashed black line "SAR limit" corresponds to the maximum depth detectable by the SAR method. (B) Same as (A) using a pre-processing step (temporal normalization) for the calculation of seismo-acoustic CCFs. Effects of low correlation between seismic and infrasound signals that affect the short-term estimation of the lake depth (see A) are mitigated during daytime.

calculation between 8 and 15 November 2016. The main drop occurring on 12–13 November is well identified but other major decreases are related to low correlations between seismic and infrasound signals during daytime. Indeed, even in the low frequency band [0.3–0.9] Hz, the correlation of seismic and infrasound signals can be strongly hampered during daytime due to a noisier local environment. This also explains why the AS ratio as computed for **Figure 12** cannot be considered as a robust monitoring attribute at this short time scale. Despite the use of an additional lagging 6-h averaging window, the diurnal artifacts on the seismo-acoustic CCFs are still noticeable. Therefore, we test a temporal normalization approach traditionally applied in ambient seismic noise studies for enhancing background coherent noise signals. This pre-processing step is called runningabsolute-mean normalization and consists in computing a normalization weight prior to the calculation of the crosscorrelation functions (Bensen et al., 2007). For a single data trace of n points dn(either seismic or infrasound), the new corrected data D<sup>n</sup> is defined as:

$$D\_n = \frac{d\_n}{\left[\frac{1}{2N+1} \sum\_{j=n-N}^{n+N} |d\_j|\right]} \tag{1}$$

Following Bensen et al. (2007), the width of the normalization window (2N + 1) can be defined as 1/(2fc1), where fc<sup>1</sup> is the low cutoff frequency of the band-pass filter (here 0.3 Hz). The effect of this correction procedure is clearly visible in **Figure 13B**. For the same time period (8–15 November 2016), the artificial drops of the lava lake due to the decrease of seismo-acoustic correlation during daytime are efficiently reduced. Hence this correction needs to be applied to get a more accurate short-term estimation of the lava lake depth. It is important to note, however, that this normalization does not improve the calibration model previously obtained because the large averaging window of several days (±10 days) sufficiently mitigates the diurnal variations. This way, meaningful CCF mean frequencies could be determined and compared with SAR measurements (see **Figure 11**). Such as observed for the daily estimates (**Figure 12A**), the short-term estimation of the lava lake depth will be less accurate when the volcanic tremor intensity is weak. A minimum threshold value for the correlation between seismic and infrasound signals should be also defined. Here, this value is arbitrarily set to 0.01 (see **Figure 13**), yet must be tested in the future for other periods of unrest.

#### DISCUSSION: A RELEVANT SHORT-TERM UNREST INDICATOR

The short-term approach described previously allows for the identification of the lava lake drop accompanying the deep seismic swarm and the LP "lake" seismicity with enhanced temporal resolution (**Figure 14**). The maximum occurrence of deep, volcano-tectonic events (12 November) occurs simultaneously with the main drop of the lava lake and precedes the substantial increase of surface LP activity (13–14 November). These observations confirm that the major fluctuations of the lava lake level reflect the dynamics of the magmatic system and that its monitoring is of great importance for evaluating the potential of magma intrusion and eruption at

Nyiragongo. This is notably consistent with the interpretation of Wauthier et al. (2012), who inferred a deep dike intrusion (2– 10 km deep below sea level) from InSAR analysis that triggered the drainage of magma from the lava lake to a shallower eruptive dike in January 2002.

No eruptive episode was analyzed in this study, but the large November 2016 drop could convey a similar mechanism of possible lateral magma intrusion. In Hawaii, before eruptive events leading to the drainage of the lava lake at Halema'uma'u crater, Patrick et al. (2015) pointed out systematic increases of the lava lake level associated with inflation episodes. Such pressurization processes cannot be evidenced from our measurements acquired at ∼6 km from the summit. After several months of sustained high level (∼20 m below the rim, see **Figures 11**, **12**), there was a small increase of the lava lake level about 1 month before the large November 2016 drop, but this variation almost entirely falls into the confidence interval of the lake depth estimates. In the case of the Hawaiian eruptions studied by Patrick et al. (2015), the periods of increasing pressure preceding the eruptions were also associated with an increase of earthquakes occurrence rate while, in our case, there was no precursory swarm of seismic events. As highlighted by the short time window "12–13 November" depicted in **Figure 14**, the sudden increase of deep seismicity is synchronous with the decrease of the lava lake level, and the drop of the lava lake level is the only surface evidence of this magma intrusion at great depth.

There are two main notable increases of the lake level on 16 and 20 November that are potential artifacts due to low correlation between seismic and infrasound signals (parts of the solid line not highlighted in red). However, these episodes of increasing levels could be also due to gas pistoning, a common phenomenon at Nyiragongo (Smets et al., 2016). The pistoning effect, which is due to the accumulation of gas in the upper layer of the lake, is associated with low surface activity and quiet seismicity, as observed at Kilauea volcano (Patrick et al., 2016). Therefore, these two short increases occurring on 16 and 20 November, which indeed correspond to periods of low LP seismicity (**Figure 14B**), might be real episodes of gas pistoning. In the future, the comparison with terrestrial observations at the summit appears essential in order to better constrain this short-term estimation of the lava lake level.

#### CONCLUSIONS

In this study we showed that information about the lava lake dynamics at Nyiragongo, D.R. Congo, is contained in the particular seismic and infrasound signals generated by its spattering activity, similar to what is observed at other lava lake volcanoes such as Kilauea, Hawaii, or Villarrica, Chile. In the absence of close-range observations, we used seismic and infrasound data from a single station (KBTI) located about 6 km from the lava lake. The particularity of the present work is the use of SAR images for calibrating seismo-acoustic ground measurements with lava lake depth. In a noisy anthropogenic environment, a cross-correlation analysis of this continuous seismic-acoustic tremor reveals changing frequency content directly related to the fluctuations of the lava lake level. The processing approach defined here in order to retrieve consistent lava lake depth estimates could eventually be useful for other lava lake volcanoes when no close-range observations are possible.

From a monitoring perspective, the reliable interpretation of the seismo-acoustic signature of the lava lake activity recorded at several kilometers from the summit is a significant step forward in order to better interpret periods of unrest at this volcano on a daily and hourly basis. The intensity of bubble bursting can be inferred through the occurrence of repetitive LP events and their acoustic-to-seismic (AS) amplitude ratios. The AS ratio is also applicable on a daily timeframe for characterizing the volcanic seismo-acoustic tremor. We particularly focused on the November 2016 time interval because we show that this type of drop and the following intense LP "lava lake" seismicity are a consequence of deep magma intrusion and could eventually lead to the next eruption at Nyiragongo. Telemetered near real-time information from the summit (e.g., seismometer, infrasound, SO2/thermal/visible cameras) would obviously be a great help, complementing these remote observations, and could help to refine the model between the lava lake depth and the spectral content of the distant seismo-acoustic records in the future. Other important tools, such as the automatic classification of volcano-seismic events, should also be implemented in near real-time using station KBTI in order to provide an efficient, single-station monitoring solution.

#### REFERENCES


### AUTHOR CONTRIBUTIONS

JB, NdO, and AO: carried out the data analysis; JB and AO: processed the seismic and infrasound data; NdO and HG: processed SAR data; NM, NdO, BS, JB, AO, HG, and FK: carried out field work; JB: wrote the manuscript and all authors commented and contributed to this original version.

### ACKNOWLEDGMENTS

The authors thank Robin Matoza and David Fee for their constructive reviews, which helped to improve the original manuscript, and Valerio Acocella for his final helpful comments on this work. This article is a contribution in the framework of the project Remote Sensing and In Situ Detection and Tracking of Geohazards (RESIST, http://resist.africamuseum.be), funded by the Belgian Science Policy (Belspo), Belgium, and the Fonds National de la Recherche (FNR), Luxembourg. We are grateful to Belspo for funding CosmoSkyMed images (Italian Space Agency) and to Natural Resources Canada (NRCan) for sharing RADARSAT images (Canadian Space Agency). We also wish to thank the Congolese Institute for Nature Preservation (ICCN) and the MONUSCO (UN stabilization mission in Congo) for their continuous support and allowing us to host the stations in their compounds, as well as the entire Goma Volcano Observatory team and the sentinels of the stations, without whom the operation of the seismic network would be impossible. KivuSNet data are underlying an embargo policy following the conditions of the Memoranda of Understanding between the partner institutions of RESIST. Beyond this embargo policy, data may be shared for collaboration purposes upon request with the approval of all RESIST partners. Data archiving and accessibility is ensured through the GEOFON program of the GFZ German Research Centre for Geosciences (http://dx.doi.org/doi:10.14470/XI058335) and KivuSNet is registered within the FDSN with network code KV (http://www.fdsn.org/networks/detail/KV/).


**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 Barrière, d'Oreye, Oth, Geirsson, Mashagiro, Johnson, Smets, Samsonov and Kervyn. 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 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.

## Analysis of the Alaska Volcano Observatory's Response Time to Volcanic Explosions-1989 to 2016

John A. Power <sup>1</sup> \* and Cheryl E. Cameron<sup>2</sup>

*<sup>1</sup> Alaska Volcano Observatory, U.S. Geological Survey, Anchorage, AK, United States, <sup>2</sup> Alaska Volcano Observatory, Alaska Division of Geological and Geophysical Surveys, Fairbanks, AK, United States*

A major goal of volcano monitoring is the rapid identification of volcanic explosions and subsequent warning of associated hazards. Between 1988 and 2016 the Alaska Volcano Observatory (AVO) responded to at least 54 separate volcanic eruptions. During this period, AVO's monitoring program relied principally on seismic and satellite remote sensing data, supplemented with geodetic, gas, and visual observations to track volcanic unrest. In this study we focus on AVO's response time, or the time required for AVO to (1) identify seismic signals associated with large ash-producing volcanic explosions and (2) initiate public warnings. We restrict this analysis to volcanoes monitored by a local seismic network and explosive in character. We focus on the 1989–90 eruption of Redoubt Volcano (VEI 3), the 1992 eruption of Mount Spurr (VEI 4), the 1999 eruption of Shishaldin Volcano (VEI 3), the 2006 eruption of Augustine Volcano (VEI 3) and the 2016 eruption of Pavlof Volcano (VEI 2) as detailed records of the timing of formal warnings are preserved. These eruption sequences allow us to evaluate AVO's response time under a number of monitoring scenarios, including both expected (those with recognized precursory unrest) and surprise eruptions (those without identified precursory unrest) as well as individual and repetitive sequences of explosive events. Recorded response time ranges from ∼1 to 86 min. The shorter response times (∼1–13 min) were achieved during sequences of explosive events at Redoubt (1989–90), Spurr (1992) and Augustine (2006). The longer response times (31– 86 min) are recorded for unexpected or surprise explosions such as Spurr (August 18, 1992) and Pavlof (2016) and the only or first explosions in an eruptive sequence such as Shishaldin (1999) and Augustine (2006).

#### Keywords: volcanic explosion, response time, Alaska, aleutian, warning

#### INTRODUCTION

The Alaska Volcano Observatory (AVO) is responsible for monitoring active volcanoes in Alaska and providing warnings of hazards associated with volcanic activity. The observatory was founded in 1988 and is a cooperative program of the U.S. Geological Survey, the University of Alaska Fairbanks Geophysical Institute, and the Alaska Division of Geological and Geophysical Surveys. Since 1988, AVO has responded to 54 eruptions in Alaska at 18 volcanic centers as well as numerous episodes of volcanic unrest that did not result in eruption. The principal hazard from explosive eruptions of Alaska volcanoes is airborne volcanic ash to aircraft, and ashfall on local communities (Casadevall, 1994b).

#### Edited by:

*Nicolas Fournier, GNS Science, New Zealand*

#### Reviewed by:

*Pablo Samaniego, UMR6524 Laboratoire Magmas et Volcans (LMV), France Eisuke Fujita, National Research Institute for Earth Science and Disaster Prevention, Japan*

> \*Correspondence: *John A. Power jpower@usgs.gov*

#### Specialty section:

*This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science*

Received: *14 February 2018* Accepted: *22 May 2018* Published: *20 June 2018*

#### Citation:

*Power JA and Cameron CE (2018) Analysis of the Alaska Volcano Observatory's Response Time to Volcanic Explosions-1989 to 2016. Front. Earth Sci. 6:72. doi: 10.3389/feart.2018.00072*

During this time, AVO principally relied on seismic observations for real-time monitoring and eruption forecasting. From 1988 to 1995, AVO seismic instrumentation was focused on just Mount Spurr, Redoubt Volcano, Iliamna Volcano, and Augustine Volcano. These four volcanoes are in the Cook Inlet region, and are closest to the major population centers in south-central Alaska. In 1995, AVO began to expand seismic monitoring to volcanic centers further to the west, and currently networks of 4–15 seismometers are in operation on 32 of the 54 historically active volcanoes in Alaska. Seismic monitoring is supplemented with geodetic measurements (Cervelli et al., 2010), measurements of gas emission such as SO<sup>2</sup> and CO<sup>2</sup> (Werner et al., 2013) and Satellite Radar Interferometry (InSAR) (Lu and Dzurisin, 2014). Additional data streams from infrasound, web cams, lightning, remote sensing, and visual observations serve principally for eruption detection and confirmation. Geologic studies are used to define eruptive history and character and identify hazardous areas.

AVO's response to volcanic unrest includes increasing surveillance of real-time monitoring of seismic activity, and nearreal-time remote sensing data, up to and including round-theclock staffing at offices in Anchorage and/or Fairbanks. AVO also increases other monitoring data streams if possible. When large ash-producing explosions are detected, AVO conducts a "call down" and communicates warnings immediately by telephone, using an ordered list. Entities on the list include relevant government agencies such as the Federal Aviation Administration (FAA), the Division of Homeland Security (DHS), the National Weather Service (NWS), the Alaska Governor's Office, and others. When sufficient staffing is available, warnings are transmitted simultaneously through a call down from both the Anchorage and Fairbanks offices of AVO. Either office could complete the entire call down if the other office was not staffed (Neal et al., 2010). During the later years of this study call downs could be made from remote locations as the monitoring data was more widely available through computer networks. The amount of time required to identify and warn of a volcanic explosion is critical as delays in warning increase the chance of an encounter between and aircraft and a volcanic ash cloud and decreases the amount of time that communities have to prepare for ash fall. Once other government agencies such as the FAA, NWS, and DHS receive information from AVO they produce and disseminate additional warning products specific to the current volcanic activity and hazard (Neal et al., 2010).

During the 1989–90 eruption of Redoubt Volcano, AVO developed a four-step Volcano Alert Level, called the "Color Code," that used four colors to reflect both the level of volcanic unrest and associated hazards and as well as expected activity (Brantley, 1990). In 2006 the U.S. Geological Survey expanded the Alert-Notification system used for volcanic hazards to a dual system to more effectively communicate hazards on the ground (lahars, pyroclastic flows, lava flows, ash fall, and others) and to aircraft (airborne volcanic ash). Ground based hazards are communicated through four alert levels: normal, advisory, watch, and warning, and the aviation Color Code remained four colors, green, yellow, orange, and red, but with slightly different definitions. The new alert-notification system (Gardner and Guffanti, 2006) communicates current hazards and unrest as well as forecasts anticipated activity. This standardized system is now in use at the five volcano observatories in the United States. Immediate warnings are still communicated by a telephone call down to affected government agencies and these are followed by several forms of written communication that include Volcanic Activity Notice (VAN), Volcano Observatory Notice for Aviation (VONA), Status Report, and Information Statement (Neal et al., 2010). These notifications are transmitted by email, web posting, Facebook, FAX, and shorter statements may be released through Twitter (Schaefer et al., 2012).

In this study we evaluate the time required for AVO to identify a large ash producing explosion and initiate a call down to appropriate authorities. We restrict this analysis to those volcanoes that were monitored by a local seismic network, were explosive in character and produced large high-altitude ash plumes. We use the call down initiation time as earliest warning that AVO issues. It often takes a longer period of time to formulate and distribute a VAN, VONA or Information Statement. For this analysis we focus on explosions that occurred at Redoubt (1989–90), Spurr (1992), Shishaldin (1999), Augustine (2006), and Pavlof (2016), as each of these eruptions were characterized by explosive activity that produced large ash plumes, and because we retained accurate records of call down timing. All times in this manuscript are reported in Coordinated Universal Time (UTC). We generally do not consider small explosive events that may have preceded larger eruptions unless those events played a role in changing AVO's level of monitoring. Fortunately, this suite of volcanoes covers a range in type and intensity of precursory unrest and eruptive style and allows us to evaluate AVO's response time in a variety of monitoring situations such as the expected onset of eruptive activity (Spurr, 1992, Shishaldin, 1999, Augustine, 2006), sequences of successive explosive events (Redoubt 1989–90, Spurr, 1992, Augustine, 2006) and unexpected or surprise eruptions (Pavlof, 2016). A review of AVO's efforts to forecast eruptions and provide advance warning of hazards during the 1988–2016 time period is given by Cameron et al. (2018).

#### REDOUBT VOLCANO, 1989–90

Redoubt Volcano is a stratovolcano located in south-central Alaska on the western side of Cook Inlet (**Figure 1**) that erupts andesitic magma (Coombs et al., 2013). The volcano has erupted four times in recorded history in 1902, 1966, 1989–90, and 2009. The 1989–90 and 2009 eruptions are characterized by explosive events that are often separated by periods of lava extrusion in the summit crater that form domes (Miller, 1994; Schaefer et al., 2012). Explosive events produce high altitude ash plumes that endanger overflying aircraft, regional aircraft transiting to other communities, and traffic landing at Anchorage International Airport (Casadevall, 1994a). A network of 5–12 seismic stations has monitored Redoubt since 1988. **Figure 2** is a map of the seismic stations operating at the end of the 1989–90 eruption.

A subtle increase in seismic events beneath the volcano and observations of increased steaming preceded the eruption by

several weeks and an intense 23-h swarm of long-period seismic events occurred immediately prior to the onset of eruptive activity on December 14, 1989 (Power et al., 1994). In response to the increase in seismicity, AVO began 24-h on site monitoring on December 13, 1989 and continued until June of 1990. The eruption began with a small explosive event at 18:47 on December 14, followed by a sequence of 24 additional explosive events. These explosive events were often separated by periods of lava extrusion, which formed short-lived domes in the volcano's summit crater. Dome growth ceased in June 1990. The explosions are considered to be Volcano Explosivity Index (VEI) 3 and were often associated with dome collapse and destruction (Miller, 1994). Many of the explosions during the 1989–90 eruption were preceded by short swarms of Long-Period seismic events that lasted from 1 to 170 h (Stephens et al., 1994) and were used to forecast the individual explosive events (Chouet et al., 1994). Seismic data were transmitted to AVO offices in Fairbanks and Anchorage in real-time, and during the 1989–90 eruption the principal method of real-time analysis was drum recorders. **Figure 3** shows the drum record from the February 21, 1990 explosion at Redoubt from station RED. The 1989–90 eruption of Redoubt allows us to evaluate the time required for AVO to identify and respond to individual explosive eruptions while 24-h staffing and analysis was in place and the volcano was generating explosive events that were similar in character and relatively repetitive.

For this eruption the response time was the difference between the explosion onset time and the recorded time of the call to the FAA, made from AVO offices in Fairbanks (**Table 1**). The onset time of the explosion was taken as the time when the seismic signal exceeded twice the normal background on seismic station SPU (located near Mount Spurr) and was rounded to the nearest minute (Power et al., 1994). The call down time in Fairbanks was manually recorded in a logbook. Unfortunately, records of call down times do not exist for the first 11 explosions that occurred before January 11, 1990, and the explosive event on January 17, 1990 (**Table 1**). The range of recorded response times was 1–9 min and the average was 3.5 min with a standard deviation of 2.0. For this calculation we have not considered smaller explosive events that either did not create a seismic signal detectible on station SPU or produce a large ash plume.

#### MOUNT SPURR, 1992

Mount Spurr is a large composite volcano on the western side of Cook Inlet about 80 km west of the city of Anchorage Alaska (**Figure 1**). Spurr has erupted twice in historic time from the Crater Peak vent, a single explosive event on July 9, 1953 (Juhle and Coulter, 1955) and a series of three sub-plinian explosions (VEI 4) in the summer of 1992. Eruptive products from the 1992 eruptions of Mount Spurr's Crater Peak vent are andesites (Nye et al., 1995). A network of 11 short-period seismic stations

TABLE 1 | Explosions onset, call down, and response time for explosive events that occurred at Redoubt Volcano in 1989 and 1990.


*NR, No Record.*

monitored the 1992 eruption (Power et al., 1995) and data were telemetered to Anchorage and Fairbanks in real-time. During the 1992 eruptions, the principal means of real-time review of seismic data was drum recorders. For response analysis during the 1992 Spurr eruption sequence we used the onset times of explosive events and the time call downs reported by Eichelberger et al. (1995).

The 1992 explosions occurred on June 27, August 18–19 and September 17, 1992 and each lasted 3–4 h and produced large plumes of volcanic ash that reached altitudes of 14,000–15,000 m above sea level (Eichelberger et al., 1995). The June 27 explosion was preceded by a long-term sequence of earthquakes that began in August of 1991 and increased in rate in November of 1991 and April 1992 (Power et al., 1995). In the weeks prior to the June 27 explosion, increased steaming from Crater Peak and changes in lake color attributed to SO<sup>2</sup> scrubbing, and geysering were observed in the lake area (Doukas and Gerlach, 1995). Strong tremor began at 20:04 on June 26 and a shallow swarm of earthquakes preceded the onset of explosive activity by just 4 h (McNutt et al., 1995; Power et al., 1995). In response to this activity AVO raised the Color Code to yellow on the evening of June 26 and began 24-h monitoring from both Fairbanks and Anchorage. The June 27 eruption began at 15:04 and the first warning that an eruption was in progress was issued 12 min later at 15:16, following receipt of a confirming report from an aircraft pilot. The June 27 explosion was followed by relative seismic quiescence at shallow depths, although the deeper portions of the magmatic system in the lower crust remained seismically active (Power et al., 2002). The Color Code was lowered to Green and 24-h monitoring was suspended on July 8.

The August 19, 1992 main explosion was preceded by a small "premonitory explosion" at 23:41 on August 18 that was observed by passing aircraft. It produced a small plume of steam and ash to 300–600 m above the Crater Peak vent. In response to this small explosion, AVO began a call down and raised the Color Code to Yellow at 00:25, 44 min after the premonitory explosion, and began close review of drum records in both Anchorage and Fairbanks. This small explosive event was unexpected, but occurred during the normal workday so the repose time was relatively short. The major explosion on August 19 began at 00:42 and AVO issued a warning 5 min after the onset at 00:47 raising the Color Code first to Orange and then to Red (at 00:58) as the eruption grew in strength (Eichelberger et al., 1995). **Figure 4**

shows drum records from 00:05 UTC August 18 to 02:25 UTC on August 19, 1992 for station CRP that show the onset of both the premonitory and main explosions as well as the time of the call down following these explosions. A map showing the location of station CRP can be seen in **Figure 1** of Power et al. (1995).

amplitude explosion signal from over writing the record on the drum.

The September 17 explosion was preceded by 3 h of increasing tremor and a small premonitory explosion at 06:36. AVO issued a warning for the premonitory explosion at 06:45, raising the Color Code to Red. A strong seismic signal beginning at 08:03 is attributed to the start of the main eruption on September 17. Eichelberger et al. (1995) did not record the time of a call down following the onset of this major explosion, however a warning was issued when the Color Code was raised to Red 9 min after the smaller premonitory explosion at 6:36.

The three major explosive events during the 1992 eruption sequence at Mount Spurr all exhibited strong precursory activity that was easily captured by the seismic network, complementary visual observations of the volcano, and gas measurements. The 1992 Spurr eruption sequence allows us to analyze AVO's response capability when the AVO offices were staffed 24-h per day and close monitoring was taking place as well as a surprise explosion while the office was fully staffed. The response times of the 1992 eruptions of Mount Spurr were 12 and 5 min for the major explosive events on June 27 and August 19, and 44 and 9 min for the premonitory explosions on August 18 and September 17 respectively (Eichelberger et al., 1995). These warnings result in an average response time of 17.5 min with a standard deviation of 17.8.

#### SHISHALDIN VOLCANO, 1999

Shishaldin Volcano, located in the central portion of Unimak Island (**Figure 1**), is highly active, with 31 recorded eruptions since 1775 (McGimsey et al., 2004). The volcano typically erupts products that are basaltic andesite in composition, and its eruptions are often characterized by activity ranging from extrusive with occasional strombolian fountaining to large explosive events that reach VEI 3 or 4 and produce large, high-altitude ash plumes (Nye et al., 2002). The VEI 4 1999 Shishaldin eruption sequence was monitored by a network of six seismometers surrounding the volcano that had been in operation since the summer of 1997 (Jolly et al., 2001). The primary medium for real-time recording and analysis during this eruption was drum recorders.

The 1999 eruption sequence was preceded by months of unrest that included Deep Long-Period events (Power et al., 2004), small repetitive seismic events that were first noted on February 2 (Moran et al., 2002), and increased steaming and thermal output from the volcanoes summit that began on February 12 (Nye et al., 2002). Volcanic tremor was first observed on February 18, prompting AVO to raise the Color Code to Yellow. A shallow magnitude 5.2 earthquake located 10–15 km west of Shishaldin occurred on March 4, followed by numerous aftershocks (Moran et al., 2002). On April 7 the strength of tremor increased further and AVO raised the level of Concern Color Code to Orange and began 24-h monitoring that would continue until June 18. Strombolian explosions were observed by AVO staff members from an aircraft using a Forward Looking Infrared Radiometer (FLIR) on April 17 (McGimsey et al., 2004). Rasmussen et al. (2018) provide a multidisciplinary analysis of unrest and volcanic processes that proceeded the 1999 Shishaldin eruption.

The 1999 Shishaldin eruption sequence had only one large explosive event that took place at 20:33 on April 19, 1999, and was associated with a large increase in seismic tremor. This explosion continued for 7 h, and eventually the ash cloud reached an estimated altitude of 17,000 m (Nye et al., 2002). The April 19 eruption of Shishaldin allows us to evaluate AVO's response time for a large explosive eruption at a volcano that was the subject of heightened monitoring 24-h per day. Records indicate that the call down from the AVO Fairbanks Office to the Alaska Governor's office occurred at 21:04 UTC. The Alaska Governor's office is the second notification in the Fairbanks call down and consequently we surmise that the first warning was issued to the FAA in just under 31 min after the onset of explosive activity.

#### AUGUSTINE VOLCANO, 2006

Augustine Volcano is a stratovolcano located on a small island in lower Cook Inlet (**Figure 1**). The volcano erupts frequently with recorded eruptions in 1812, 1882, 1902, 1935, 1963–64, 1971, 1976, 1986, and 2006. The Augustine volcanic cone has experienced frequent sector collapse events in the recent past and the 1882 eruption generated a small tsunami in Cook Inlet (Waitt and Beget, 2009). Augustine magmas are typically andesitic in composition (Larsen et al., 2010) and eruptions are characterized by an explosive onset followed by sequences of explosive events that are often separated by the extrusion of lava from the volcano's summit, forming lava domes and lava flows. Seismic monitoring of Augustine began in 1970 and the 1976, 1986, and 2006 eruptions have remarkably similar precursory seismic sequences that allowed AVO to very accurately forecast the onset and style of the 2006 eruption (Power et al., 2006). The VEI 3 2006 eruption consisted of 13 larger magmatic explosions that occurred between January 11 and January 28, separated by the extrusion of magma within the summit crater. Magma extrusion continued until mid-March 2006 (Coombs et al., 2010).

During the 2006 eruption of Augustine, the volcano had a network of 8 permanent seismometers (**Figure 5**) and seismic data were transmitted in real-time to both the Anchorage and Fairbanks offices where staff monitored the incoming seismic data 24 h per day. Seismic waveforms were displayed using a variety of computerized techniques that included the capability to view the seismic signals in both the time and frequency domain (Cervelli et al., 2004; Thompson and Reyes, 2017). The onset times of explosions were determined at seismic station OPT, located 16 km north of Augustine's summit, and were rounded to the nearest minute (Power and Lalla, 2010).

The 2006 eruption of Augustine provides a good opportunity to evaluate AVO's response time to both the initial onset of explosive activity for an eruption that had well-recognized precursory seismicity and 24-h staffing in place, as well as a sequence of similar explosive events. For this eruption, the recorded response time reflects the time at which the call down was initiated in Fairbanks to the Alaska State Division of Homeland Security, which is the first number in the Fairbanks call down (**Table 2**). Both the Anchorage and Fairbanks offices of AVO were staffed 24-h per day, so this time is comparable to the time the Anchorage office called the FAA (records for the Anchorage call downs have not been recovered).

Response times during the 2006 Augustine eruption ranged from ∼1 to 86 min. The average response time was 20.7 min with a standard deviation of 30.2. Unfortunately, records of the call times for the first two explosions on January 13, the

Volcano in 2006 from Power and Lalla (2010). Solid triangles represent short-period stations, the solid square is a permanent broadband station at AUL, open squares represent temporary broadband stations, and the circle is a strong motion instrument at AU20.

TABLE 2 | Times of explosions onset, call down time, and response time for explosive events that occurred at Augustine Volcano in 2006.


*NR, No Record.*

explosion on January 17, and the first explosion on January 28, have been lost (**Table 2**). For the remaining explosions the comparatively longer average response time is because the first two explosions in the sequence were small (durations of 3.5 and 7 min whereas later explosions had durations in excess of 10 min), with indistinct seismic waveforms, and were not immediately recognized (**Figure 6**). A formal call down was not initiated for these two explosions until 15:10 (86 and 56 min after the first and second explosions, respectively), when satellite data could confirm the presence of an ash plume (**Table 2**). If we do not consider the response times for these two explosions and consider only those later in the sequence where we have records, the average response time decreases to 6.0 min with a standard deviation of 4.6.

#### PAVLOF VOLCANO, 2016

Pavlof Volcano has had more than 40 eruptions since 1817 and is generally considered to be the most active volcano in North America (Waythomas et al., 2006). Eruptions at Pavlof are characterized by Strombolian activity that produces minor amounts of volcanic ash and small lava flows from the summit crater. Pavlof magmas are typically basaltic andesites (Waythomas et al., 2017). Occasionally Pavlof erupts more explosively, producing ash plumes reaching altitudes higher than 15,000 m and vigorous lava fountaining (McNutt et al., 1991; Waythomas et al., 2006). Eruptions at Pavlof typically occur with little observed precursory unrest and is often described as an open-conduit system (Lu and Dzurisin, 2014; Cameron et al., 2018). AVO has operated a six-station seismic network on the volcano since 1996 (Dixon et al., 2013), and an earlier seismic network monitored the volcano from 1973 to 1990 (McNutt and Beavan, 1987). During the March 2016 eruption of Pavlof, seismic data were transmitted to AVO offices in Anchorage and Fairbanks in real-time and a variety of computerized real-time displays allowed the data to be reviewed in both the time and frequency domain.

The March 2016 eruption of Pavlof began unexpectedly on the afternoon of Sunday March 27, with a relatively small increase in tremor at 23:55 (**Figure 7**). A pilot from a passing aircraft reported the volcano in eruption with an ash cloud reaching 6,000 m at 00:18 on March 28. AVO initiated a full call down at 00:55, raising the Color Code from Green to Red 60 min after the onset of tremor that marks the start of explosive activity. At 01:33, satellite imagery revealed an ash plume rising to an altitude of 9,000 m. The tremor signal intensified at about 02:23 (**Figure 7**) and strong ash emission continued on March 28 with a plume reaching as high as 11,000 m. Lava fountaining was observed from the town of Cold Bay. Lower altitude ash plumes were observed for several days and activity then subsided and AVO lowered the Color Code to Orange on March 28, Yellow on April 6 and Green on April 22.

The March 2016 eruption of Pavlof provides an opportunity to review AVO's response time at a volcano that had no recognized precursory activity and on-site monitoring of real-time seismic data was not in place at the onset of activity. For this eruption, the first warning phone call to the FAA from AVO Anchorage office was made 60 min following the onset of explosive activity at 00:55.

#### DISCUSSION

The review of AVO's response times from eruptions at Redoubt Volcano (1989–90), Mount Spurr (1992), Shishaldin (1999), Augustine (2006), and Pavlof (2016) allows us to evaluate the time required to produce a warning under a number of monitoring scenarios including: repetitive sequences of explosive events that occurred over months to days (Redoubt, 1989–90, Spurr, 1992, Augustine, 2006), single explosions with clear seismic precursors (Shishaldin, 1999), and unexpected explosions (Spurr, August 18, 1992; Pavlof, 2016). All of these volcanoes were monitored by local seismic networks of 5–12 individual stations that ranged in distance from <1 to 28 km in distance (Dixon et al., 2013). Response times at these five volcanoes range from ∼1 to 86 min for the 28 explosions where records of the call down times are preserved.

The shortest response times were from Redoubt and Augustine when the volcanoes were producing sequences of explosions with well-known similar seismic characteristics. For these sequences of repetitive explosive events, response times ranged from ∼1 to 13 min (**Tables 1**, **2**). Faster response times were also observed at volcanoes with strong precursory activity that prompted AVO to move to 24-h continuous monitoring, such as Spurr in 1992 where response times ranged from 5 to 12 min. Of the 28 volcanic explosions considered in this study 23 had response times of <13 min.

The longest response times occurred for either the first or opening explosions in an eruptive sequence or for unexpected or surprise explosions when heightened review of incoming seismic data was not in place and ranged from 31 to 86 min. Only five of the 28 response times considered in this study took longer than 13 min. Many times the desire to confirm the presence of an ash cloud using alternative monitoring data such as satellite imagery or reports from aircraft pilots contributed to the longer response times. The 41-min response time for the premonitory explosion on August 18, 1992 at Spurr results from a smaller and less distinct seismic signal and the time to receive a confirming pilot report (**Figure 4**). The 2016 Pavlof eruption was not expected and 24-h monitoring was not in place. The 60-min response time results from time spent to contact and organize AVO staff, evaluate the incoming seismic data, and review other monitoring data to confirm the explosive character of the event and the presence of an ash plume. The 86-min response time for Augustine in 2006 occurred because the explosion was relatively small and the seismic envelope resembled a larger earthquake (**Figure 6**). The single call down for the first two explosions in 2006 was not made until satellite data could confirm an ash cloud.

The style and amount of precursory unrest is also a determinant in the response time. Many of the explosions at Redoubt in 1989–90 were preceded by swarms of Long-Period events (Stephens et al., 1994) and response times were all <10 min. The June 27, August 19 and premonitory explosion on

in volcanic tremor at 23:55 on March 27 and the AVO call down occurred at 00:55 on March 28.

September 16 at Spurr in 1992 all had well recognized seismic precursors and notifications were issued 12, 5, and 9 min after the explosion's seismic signal began. Given that the April 19, 1999 explosion at Shishaldin had well-recognized precursory activity and 24-h monitoring, the response time of 31-min is long. However, eruptions at Shishaldin are often characterized by extrusive and strombolian activity, and the April 19th explosion was the only explosive event during the 1999 eruption sequence. Consequently, the longer response time for a single explosion, which is identical to the first explosion in an eruptive sequence, is somewhat expected.

Many of the longer response times at the start of an eruptive sequence or for single explosive events often result from a desire to confirm the presence of an ash cloud using monitoring data independent from seismic observations, such as satellite imagery or reports from aircraft pilots. During the period of this study, satellite imagery used to identify and track ash clouds (GOES) can have a lag time of 15–30 min (Dave Schneider, Pers. Comm, 2017). Once the seismic waveform character of explosive events is established, AVO's confidence in issuing warnings based only on a seismic interpretation increases, as demonstrated for Redoubt 1989–90 and Augustine 2006. It may be that the best method of reducing response time would be to develop automated software tools that would identify and characterize the seismic signature of explosions and the incorporation of additional data streams to assist with confirmation of explosive activity and identification of large ash plumes. These techniques include infrasound (DeAngelis et al., 2012), lightning (Behnke et al., 2013), and radar (Schneider and Hoblitt, 2013). AVO has recently made efforts to incorporate these data streams into real-time use in developing warnings (Coombs et al., in preparation) and this may lead to shorter response times.

The observed response times do not seem to change significantly through the 28-year period of this study. The response times from sequences of events during the 1989–90 eruption of Redoubt, displayed on drum recorders, are not significantly different than the response times during the 2006 eruption of Augustine when computerized display and analysis packages were used. This is a surprising result, as we would have expected response times to shorten, given the additional analytical capability in the computerized displays.

#### SUMMARY AND CONCLUSIONS

Based on our analysis of 40 explosive events with available historical records: Redoubt Volcano 1989–90, Mount Spurr 1992, Shishaldin 1999, Augustine 2006, and Pavlof Volcano 2016, we can offer the following observations on AVO's ability

#### REFERENCES

Behnke, S. A., Thomas, R. J., McNutt, S. R., Schneider, D. J., Krehbiel, P. R., Rison, W., et al. (2013). Observations of volcanic lightning during the 2009 eruption of Redoubt Volcano. J. Volcanol. Geotherm. Res. 259, 214–234. doi: 10.1016/j.jvolgeores.2011.12.010

to seismically recognize and provide warnings of volcanic explosions.

The response time for AVO to provide notification—here defined as a telephone call to the FAA—of a major volcanic explosion ranges from ∼1 to 86 min.

Shorter response times ranging from ∼1 to 13 min were achieved during eruptive sequences at Redoubt 1989–90, Spurr 1992, and Augustine 2006 that contained numerous explosive events with seismic waveforms that were similar in character and occurred when AVO offices were staffed 24-h per day.

The response times for unheralded volcanic explosions that lacked recognizable precursory unrest ranged from 44 to 60 min. Examples are the August 18, 1992 premonitory explosion at Spurr and that on March 27, 2016 at Pavlof.

Longer response times were also recorded for a single or first explosive event in an eruption sequence. The response time for the single explosive event at Shishaldin in 1999 was just under 31 min. The response time for the first two explosive events at Augustine in 2006 were 58 and 86 min, respectively.

The response times to volcanic explosions are similar throughout the period of study (1989–2016) and do not seem to be affected by a move from drum recorders to computerized seismic display and analysis tools that provide the capability to view waveform data in both the time and frequency domain.

#### DATA AVAILABILITY STATEMENT

The datasets generated and analyzed for this study can be found in text and the tables contained within the manuscript.

#### AUTHOR CONTRIBUTIONS

JP assembled data on explosion and response times. JP and CC worked jointly on text. JP prepared the figures.

#### FUNDING

Funding for this project was provided by the U.S. Geological Survey Volcano Hazards Program and cooperative agreements with the USGS Volcano Hazards Program Cooperative Agreements to ADGGS (grant numbers G16AC00054 and G16AC00165).

#### ACKNOWLEDGMENTS

We thank the many individuals who contributed to volcanic eruption response efforts in Alaska between 1989 and 2017. The text and figures were improved through formal reviews by Jacob Lowenstern, Christina Neal, and two reviewers.

Brantley, S. R. (ed.). (1990). The Eruption of Redoubt Volcano, Alaska, December 14, 1989-August 31, 1990. U.S. Geol. Surv. Circular. C, 33.

Cameron, C. E., Prejean, S. G., Coombs, M. L., Wallace, K. L., Power, J. A., and Roman, D. C. (2018). Alaska Volcano Observatory Alert and Forecasting Timeliness. doi: 10.3389/feart.2018.00086


Spurr Volcano, Alaska," in The 1992 Eruptions of Crater Peak Vent, Mount Spurr Volcano, Alaska, ed T. E. C. Keith (U.S. Geol. Surv. Bulletin), 161–178.


**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 Power and Cameron. 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 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.

## Foundations for Forecasting: Defining Baseline Seismicity at Fuego Volcano, Guatemala

Kyle A. Brill <sup>1</sup> \*, Gregory P. Waite<sup>1</sup> and Gustavo Chigna<sup>2</sup>

<sup>1</sup> Department of Geological and Mining Engineering and Sciences, Michigan Technological University, Houghton, MI, United States, <sup>2</sup> Instituto Nacional de Sismología, Vulcanología, Meteorología, e Hidrología, Ciudad de Guatemala, Guatemala

Accurate volcanic eruption forecasting is especially challenging at open vent volcanoes with persistent low levels of activity and relatively sparse permanent monitoring networks. We present a description of seismicity observed at Fuego volcano in Guatemala during January of 2012, a period representative of low-level, open-vent dynamics typical of the current eruptive period. We use this time to establish a baseline of activity from which to build more accurate forecasts. Seismicity consists of both harmonic and non-harmonic tremor, rockfalls, and a variety of signals associated with frequent small emissions from two vents. We categorize emissions into explosions and degassing events (each emitted from both vents); the seismic signatures from these two types of emissions are highly variable. We propose that both vents partially to fully seal between explosions. This model allows for the two types of emissions and accommodates the variety of seismic waveforms we recorded. In addition, there are many small discrete events not linked to eruptions that we examine in detail here. Of these events, 183 are classified into 5 families of repeating, pulse-like long period (0.5–5 Hz) events. Using arrival times from the 5 families and other high-quality events recorded on a temporary, nine-station network on the edifice of Fuego, we compute a 1-D velocity model and use it to locate earthquakes. The waveforms and shallow locations of the repeating families suggest that they are likely produced by rapid increases in gas pressure within a crack very near the surface, possibly within a sealed or partially sealed conduit. The framework from this study is a short but instrument intense observation period, activity description, seismic event detection, velocity modeling, and repose period analysis. This framework can act as a template for augmenting monitoring efforts at other under-studied volcanoes. Even relatively limited studies can at a minimum aid in drawing parallels between volcanic systems and improve comparisons.

Keywords: Fuego volcano, volcano monitoring, volcano seismology, velocity modeling, repeating events

#### INTRODUCTION

Increased seismic activity is often the most discernable indicator of volcanic unrest (Tilling, 2008), and seismic monitoring of volcanic environments is therefore an essential component of any volcano observation endeavor. In many cases, the ascent of magma from the mid crust is signaled by swarms of earthquakes weeks or months prior to an explosive eruption (White and McCausland, 2016). Over the last 30 years, advances in the field of volcano seismology have been crucial to aiding

Edited by:

Nicolas Fournier, GNS Science, New Zealand

#### Reviewed by:

Patrick J. Smith, Montserrat Volcano Observatory, Montserrat Luis E. Lara, Sernageomin, Chile

> \*Correspondence: Kyle A. Brill kabrill@mtu.edu

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 28 February 2018 Accepted: 12 June 2018 Published: 03 July 2018

#### Citation:

Brill KA, Waite GP and Chigna G (2018) Foundations for Forecasting: Defining Baseline Seismicity at Fuego Volcano, Guatemala. Front. Earth Sci. 6:87. doi: 10.3389/feart.2018.00087 the scientific understanding of the processes that precede largescale volcanic eruptions (Chouet and Matoza, 2013). Even a small number of broadband seismic stations can be one of the most cost effective means of basic volcano monitoring if the goal is to forecast large eruptions (White et al., 2011). Despite these successes and associated advances in the field, medium-term accuracy and precision of eruption forecasting still has much room for improvement.

Sometimes the beginning or ending of a volcanic eruption is not a discrete event. Eruptive episodes can persist over time scales from days to years, and in rare cases decades (Siebert et al., 2011). These "open-vent" volcanoes (Rose et al., 2013) where connections between a magma body and the atmosphere are already established, or "quiescently active" (Stix, 2007) or "persistently restless" (Rodgers et al., 2013) volcanoes—where those connections open and close due to seemingly small changes within a system, provide opportunities for understanding volcanism as a phenomenon, but also present unique challenges for hazard mitigation (see Rose et al., 2013 for a review). When a volcano already exhibits frequent explosive eruptions, nearly continuous gas emission, and abundant volcanic seismicity, indicators that precede a shift to more dangerous levels of activity may be subtle (Roman et al., 2016). In these open systems, it is important to understand more detail about the seismicity to recognize changes in complex, low-level signals. Establishing a long, detailed, and well understood baseline of eruptive activity levels is one way to facilitate more accurate medium term forecasts (Tilling, 2008), and can be especially valuable in open vent situations (National Academies of Sciences, Engineering, and Medicine, 2017).

Fuego volcano is one of the most persistently active vents in the Central American Volcanic front, and has represented the main center of activity for the approximately 80,000-yearold Fuego-Acatenango massif for the past 8,500 years (Vallance et al., 2001). Fuego lavas have been chiefly basaltic-andesitic in composition, in contrast to the mostly andesitic activity of previous eruptive centers (Basset, 1996). Fuego has had more than 60 documented historical eruptions since 1524 (Escobar Wolf, 2013). The current eruptive episode began in 1999 and has been marked by periods of basaltic lava flows, strombolian style explosions and degassing events, and occasional paroxysmal events with Volcano Explosivity Indexes (VEI) of 2 and below.

Constant activity and relatively easy access to the flanks of the volcano make Fuego an excellent location to study open vent volcanic behavior. A number of research groups have partnered with the Guatemalan Instituto Nacional de Sismologia, Vulcanologia, Meterologia, e Hidrologia (INSIVUMEH) during this current eruptive episode to study the activity and work toward mitigating volcanic risk, with a large focus on using seismic and complementary data to characterize the magmatic system. A relatively long-term study by Lyons et al. (2010) used daily visual observations, seismic data, and thermal satellite images to characterize quasi-cyclic activity that included weeks to months of low-level explosive eruptions between paroxysmal eruptions that last for 1–2 days. Several field campaigns have collected data from a variety of sensors including seismometers, tilt meters, infrasound microphones, thermal imaging cameras, and SO<sup>2</sup> cameras to study explosive activity in more detail. Among the findings of these groups is the strong association between seismicity and gas emission. This includes intra explosion non-harmonic tremor accompanying gas emissions (Nadeau et al., 2011) and three repeating very-long-period (VLP) event types associated with explosive ash-rich emissions from two separate vents and weaker puffing activity (Waite et al., 2013). The multi-instrumental work has led to a model for Fuego in which a seal in the uppermost conduit develops rapidly through microlite crystallization. Tilt data show that the sealed vent results in a pressurization and inflation of the summit beginning 20–30 min before most explosions (Lyons et al., 2012). Inversions of the seismic signals for the source of VLP events have produced a model for the uppermost conduit which dips slightly to the west below a pipe-like uppermost portion (Waite and Lanza, 2016).

This recent work has focused primarily on eruption-related seismic activity to shed light on the explosion processes, but no broad characterization of local volcano tectonic (VT) or long period (LP) seismicity has been undertaken since the eruptive episode of 1975–1977 (Mcnutt and Harlow, 1983; Yuan et al., 1984). In this study, we describe the seismic activity during January of 2012 with an emphasis on LP activity. Based on discussions with the INSIVUMEH staff and compared to activity observed before (i.e., Lyons, 2011) and since the field occupation (i.e., Chigna et al., 2012; Global Volcanism Program, 2013), the volcanic activity observed during this time represents a typical period between paroxysms and serves as a good example of "background activity (**Figure 5C**). This study describes the seismic activity during this time and highlights processes not previously investigated.

### METHODS

#### Instrumentation

We installed 9 broadband seismometers around Fuego volcano from 11 January to 29 January 2012 (**Figure 1**) at distances between about 800 m and 3 km from the summit. Sites were chosen to provide full azimuthal coverage at distances as close as possible to the vent without compromising safety. Due to the steep topography and nearly continuous rockfall, the southern side of the edifice is less accessible than the north. Data were recorded on RefTek 130 Data Acquisition Systems at 100 Hz from seven Guralp CMG-40T (50 Hz to 30 s flat response) and two Trillium Compact (100 Hz to 120 s) seismometers. One of the Trillium Compact instruments was initially located at the N station site due to time constraints in the field and moved the following day to the NW1 site the following day for the remainder of the occupation. Two stations (NE1 and NW1) had collocated tilt-meters and arrays of three low-frequency microphones. Two time-lapse cameras located at ∼800 and 1,000 m NNE of the summit recorded images of volcanic activity and weather conditions. One of these cameras, a PlotwatcherPro made by Day6Outdoors, hereafter referred to as Cam1 captured images (1280 × 720 pixels) at 1 s intervals during daylight hours. The other, a Canon PowerShot A480 with a firmware modification, hereafter referred to as Cam2 recorded images (2272 × 1704 pixels) at 5 s intervals continuously (day and night)

while battery power and storage space remained. Camera clocks were calibrated by hand, referencing hand-held GPS units, so the accuracy of image time stamps is assumed to be ± 1 s of true GPS

Contour intervals are 500 m. The operational times of the stations and time-lapse cameras are shown in (C).

#### Event Detection

time.

We employed several methods to identify discrete events in the combined seismic, infrasound, and imagery data. This meant that our definition of what constituted an event was a somewhat fluid concept during the different stages of analysis. Initially, events were emissions that could be clearly identified with the camera images. The associated seismic signals were then analyzed and upon further inspection of the seismic data, events with similar seismic signals were identified. Many of these other events did not have associated clear visual records, either because the summit was obscured by clouds or because they simply did not produce emissions. We also found seismic signals associated with activity such as rockfall that was not clearly visible in the imagery data. The rest of this methods section will explain our use of multiple detection methods which allowed us to classify the different types of events discussed in section Recorded Activity.

#### Visual Identification

To begin our description of the activity, we sought to identify events and event timing visually, defining events as visible emissions from the volcano as the INSIVUMEH observers would in their daily reports. We identified 571 events using the images acquired by Cam1 and 225 events using the images acquired by Cam2, classifying them based on which vent they were emitted from, their initial speed, and the color and opacity of plume emissions during the day and based on incandescence at or above a vent position and incandescence of ejected material during the night. While this captured a large number of events, camera downtime and lack of visibility due to weather meant that most of the time period of the deployment was not recorded visually. In addition, atmospheric conditions above and around the crater produced condensation and or dust clouds which could closely mimic weak degassing emissions. Although great care was taken to exclude this type of event from the record, it is possible that some non-events were falsely identified as emission events.

We used seismic data (vertical velocity traces and FFT spectrograms with 1,024 s windows) recorded at temporary station NE1 to verify the volcanic activity associated with each of the events in the catalog derived from the images. This allowed us to eliminate events picked visually from the images which did not also appear contemporaneously with seismic activity as well as to describe the events in terms of their seismic characteristics. Upon removing false identifications, combining the datasets, and removing duplicate events observed with both cameras, we are left with a total of 448 events observed during our field campaign, averaging 2–3 events per hour over 7 days of camera operation. However, this event rate is most likely a gross underestimate because during those 7 days of camera operation, visibility was often limited or blocked due to atmospheric conditions. Quantifying an exact amount of time that visibility was limited is impossible because night images can only be classified as cloud free if incandescence is visible, but a lack of incandescence could be due to cloud obstruction or a lack of activity. An inspection of most single hours of activity while the cameras were recording with full visibility suggests 6–10 events per hour would be a better estimate, especially if weak degassing events are considered.

#### STA/LTA Algorithm

We used the seismic data to create a consistent catalog of seismic events during the deployment. The initial processing was done with Boulder Real Time Technologies Antelope 5.7 software. The data were processed using a short-time average/long-time average (STA/LTA) triggering algorithm using all seismic stations in the network. The algorithm was calibrated by comparing the number and duration of events detected to the visual activity observed on the time-lapse cameras. Data were first filtered from 1 to 25 Hz using a four-pole bandpass filter, and Root Mean Square averages were taken over 1 s (STA) and 9 s (LTA) windows with threshold ratios for detection at 2.5 times the signal to noise ratio. When more than 5 stations in the network trigger on the same event, it is added to a catalog. The five-station threshold effectively limits false detection of rockfalls as emission events, which tend to be very localized and not detected by stations on opposite flanks. **Figure 2** summarizes the timing of the identified events by showing variation in events per hour, inter-event time, and event duration. The events with especially long durations, i.e., longer than 10 min are generated by volcanic tremor coinciding with other activity which prevents reaching the detection shutoff threshold of 2.2 times the signal to noise ratio (SNR).

Events in this catalog were then reviewed manually, resulting in a total of 1,032 events detected on 5 or more stations through the occupation, an increase of 584 events from visual observations alone. Most of the events detected by the STA/LTA algorithms had emergent onsets which were very hard to discern. SNR has been found to be the main source of pick error for individual analysts (Zeiler and Velasco, 2009). Four members of our research group picked P-wave arrivals and determined pick uncertainties for 10 separate events from the middle of the dataset, and although some events have clear, impulsive arrivals, many also have arrivals which are much more ambiguous and therefore might not be reliable for earthquake location or velocity modeling. These results informed our decision to assign arrival weights to picks to reflect the impulsiveness of onset based on the analyst assigned pick uncertainty, which range from 0 to 3 for values less than 0.06, 0.15, 0.30, and 0.60 s respectively, and 4 for values greater than 0.60 s. For a first order approximation of event locations, we located these events using Antelope's dbgrassoc program which returns a location only if a detected event can be located within a user-specified grid and relies on the IASPEI91 (Kennett, 1991) crustal model. IASPEI91 gives a P-wave velocity of 5.8 km/s for the first 20 km depth, and 6.5 km/s from 20 to 35 km depth. The locations were later refined with a local 1D model as described below.

#### REDPy

The initial catalog of seismic events served as a starting point for further analysis. Recurring events, seismic events that have a similar mechanism and occur in roughly the same location, are common beneath volcanoes. In order to identify classes of these events, we used the Repeating Earthquake Detector in Python (REDPy) tool (Hotovec-Ellis and Jeffries, 2016). This detector begins by using an STA/LTA algorithm to identify event arrivals on different channels across a seismic network and stores events in a series of tables, just as with typical detection algorithms. The difference between REDPy and other tools is in the event association step. When enough stations or channels are triggered at once, an event is run through a series of cross-correlations in the frequency domain for comparison with other events in the catalog and assignment based on cross-correlation coefficient values.

A user can choose to manually delete events prior to analysis based on some criteria, such as erroneous triggers. The system stores both true events and those that the user flags as false events. Each newly detected event is compared with both groups of events. If a new event matches one previously defined as a false event, REDPy skips to the next event. If the new event does not correlate with any of the deleted events, the system writes that event to an "orphan" table, or an event without a currently identified "family" of other similar events.

As the program continues, REDPy looks at each good event in comparison with events in the "orphan" table to determine a cross-correlation coefficient. If a new event correlates with an "orphan" event above a user defined threshold on enough stations, those correlating events are moved from the "orphan"

table and grouped as a "family." The system designates the first event as a "core" event and writes it to a representative events table, which becomes important in the next step. If a new event does not correlate with any events in the current "orphan" table, it is cross-correlated with all previously identified "core" events in the representative events table. If the new event then correlates with any "core" events above the threshold, it is added to that event's family. If not, the new event is appended to the "orphan" table.

Clusters are defined using the Ordering Points to Identify the Clustering Structure (OPTICS) algorithm (Ankerst et al., 1999) which, in this usage, relies on correlation coefficients. In this implementation, an event only needs to correlate with one other event in the family to be included in the cluster; it favors fewer clusters with greater numbers of related events within each cluster family. At the same time, this algorithm identifies the event most closely correlated to all other events in the family and updates the representative event table accordingly. If the new event happens to correlate with more than one family, those family tables are merged without breaking OPTICS rules. Events are aligned within families after each clustering routine is completed so that correlation windows remain consistent.

Along with setting correlation thresholds, STA/LTA parameters, and minimum numbers of station or channel detections necessary to trigger the REDPy system, the user can search different frequency bands and give events on the "orphan" table expiration times after which they will no longer be compared to new events. We used the same STA and LTA window length settings for the STA/LTA algorithm that were used in the Antelope analysis, although the REDPy bandpass filter was in the LP band (0.5–5 Hz). We experimented with multiple filter bandwidths and found that including signal below 0.5 Hz caused the algorithm to return events which were essentially correlated microseism, and including signal above 5 Hz returned almost no events due to scattering and attenuation of signals along the path, or minor variations in source processes evident only in the higher frequencies. The STA/LTA trigger ratio was 2.5 as before and a ratio of 2.2 triggered the end of the event. We restricted this analysis to only the six closest stations, excluding S, SE2, SW2 because including them again returned more correlated noise than events in the final REDPy catalog. This effective reduction of stations from 9 to 6 resulted in our choice to opt for only requiring that 4 of the 6 stations return concurrent detections to be considered for clustering. For an event to be associated with an event family required a correlation coefficient of 0.7 or greater on 3 or more stations.

The program detected 370 events in five cluster families which had more than five repeating events between January 15 and January 24. An additional 1,867 events were found with the STA/LTA detector but were not well correlated with other events (**Figure 3**). Many different station configurations and STA/LTA settings were tested to optimize the detection of "true" events, but most produced more correlated noise than true event clusters.

#### Phase-Weighted Stacking

To improve the signal to noise ratio for the event families detected by REDPy, we use the time-frequency phase-weighted stacking technique. This technique weights the stack by

instantaneous frequency determined by the S transform allowing frequency-dependent time windowing (Stockwell et al., 1996; Schimmel and Paulssen, 1997; Pinnegar, 2005; Schimmel et al., 2011; Thurber et al., 2014). and shows significant SNR improvement (**Figure 3**). We pick arrivals from the resulting stacks using Seis\_Pick (Verdon, 2012) and assign those arrivals to the origin time and location of each core event from REDPy. Subsequently, we remove the remaining family members from the Antelope catalog.

#### Velocity Modeling and Earthquake Location

To better constrain the Antelope locations, we derive a 1-D velocity model by relocating events using VELEST (Kissling et al., 1994). We fix velocities below 9 km to values reported by Franco et al. (2009), assuming most of the activity to be concentrated within the edifice and considering the limitations on event depths based on the ∼4 km aperture of the array we had deployed. Station corrections are set initially to zero but are inverted for in each iteration and carried through to subsequent steps. We only use a subset of high quality events for the velocity modeling procedure due to the complexities inherent in exploring the model space with the simultaneous inversion for velocity and location. VELEST is able to use P-wave and S-wave arrival times, but allows modeling with only P-wave arrivals. Due to the assumed proximity of sources to receivers and the challenges in obtaining precise P-wave arrivals, we restrict our velocity modeling to P-waves only.

Our subset firstly includes the five phase-weighted stacks of clustered events (PWSCE). Next, we select high quality events from the Antelope database. To be considered high quality, the events need to fulfill three criteria: first, be located by Antelope closer to the summit than the furthest station, S1; second, be located by Antelope no deeper than 10 km below sea level; and third, have at least 6 stations with arrivals weighted 1 or 0. We exclude events already within the phase weighted stacks and are left with 60 events plus the five PWSCEs. Thirteen of these 60 events exhibit contemporaneous spikes on the infrasound channels normally indicating explosions, and these events are given an initial location at the top of the summit crater. This results in a starting set of 47 events with initial locations as determined by the Antelope locations, 13 events with locations fixed to the summit, and 5 PWSCEs with initial locations set to the surface directly under one of four reference stations.

For 17 events in the initial Antelope catalog, the seismic events not only exhibit contemporaneous spikes on the infrasound, but also occur within 2 s of the onset of emissions from the vent detected visually on one or both time-lapse cameras. We treat these events as shots within VELEST which allows us to constrain their locations without giving them a known origin time. This assumption may contribute a small error, however because shots are subject to selection criteria later in the process, we deem the approach to be appropriate. To these shots we add regional earthquakes detected with our network and contained in the International Seismological Centre On-line Bulletin (International Seismological Centre, 2016), again fixing location but not origin time. After subjecting these shots to the same selection criteria as the other events, we had 25 shots, 17 of which were fixed to the location of the summit vent.

Following the "recipe" of Kissling et al. (1994) and using approaches similar to Clarke et al. (2009) and Hopp and Waite (2016), we explore the initial model space by trying 1,000 different random initial velocity models for four different reference stations. Each of these models consisted of 10 possibly separate velocity layers. Programmatic constraints of VELEST require all seismic stations be within the first layer of the velocity model, so our layer boundaries are −4, −2, −0.5, 1, 3, 5, 7, 9, 17, and 37 km depth. The final model has six layers spanning from 2 km above sea level to a depth of 9 km. The thickness and spacing of these layers represents the minimum number of layers able to accommodate the variations we observed in early trials that also minimizes the number layers with redundant velocities. Layers eight, nine, and 10 which are below 9 km are fixed to the regional velocity model of Franco et al. (2009) which report velocities of 6.55 km/s from depths of 9–17 km, 6.75 km/s from depths of 17–37 km, and 7.95 km/s below 37 km.

We constructed the random velocity models starting with layer 1 (uppermost) and layer 7. We first generated 1,000 random numbers uniformly distributed between 0 and 1 using MATLAB's rand function (MATLAB, 2015). We established upper and lower velocity limits for this top layer between 0.6 km/s and 4 km/s, so we multiplied that vector of random numbers between 0 and 1 by 3.4 km/s and added 0.6 km/s to enforce our upper and lower bounds for the layer. We repeated this process for layer 7 with bounds set between 2 and 6.55 km/s. We used the randomly determined velocities for layers 1 and 7 as a new constraint on layer 4, and for each model, generated another random number between 0 and 1 and multiplied it by the difference between velocities in layers 1 and 7. Layers 2 and 3 were then created by generating two random numbers, between 0 and 1, multiplying them by the difference between layers 1 and 4, and then sorted with the lower velocity always being assigned to the shallower layer. This last step was repeated for layers 5 and 6 using the difference between layers 4 and 7 to complete the seven varying layers of the new randomly generated velocity models.

Although Kissling and others recommend against using shots in the early stages of exploring the initial 1-D velocity model because of the large effect they can have on the results and because they only sample the shallowest velocity layer, we chose to include shots because we assume that most of the events we recorded are occurring in the shallowest two layers, especially due to the programmatic constraint that all stations must be located within the uppermost modeled layer. The models were run for 10 iterations, or until the program failed to find a better solution after four tries. The best resulting model is selected as the model which minimizes both RMS error and station correction range.

The events are further refined for the next step by removing picks with residuals greater than 0.25 s. If this results in less than 6 picks for an event, the event is removed from the working catalog. Individual events with RMS greater than 1.0 or with station gaps of greater than 270◦ are also removed from the set. These criteria left 7 events plus 5 PWSCEs and 16 shots, 10 of which are explosions at the summit vent. These events pass through to another 1,000 random velocity models, but incorporate the new station corrections, event locations, and event origin times. We again select the model with the lowest RMS error and minimum station corrections as the best resulting model. We then iteratively feed these results into VELEST simultaneous inversion mode until the velocity model and earthquake locations stabilize to have minimal changes from one iteration to the next. We run the events through VELEST single event mode to locate the earthquakes without simultaneously inverting for velocity. The final results in **Figure 4** contain 7 events plus 5 PWSCEs with 14 shots (8 at the summit vent) which still meet the initial selection criteria.

#### RECORDED ACTIVITY

Fuego volcano, aside from being located on an overriding plate of a subduction zone is also situated relatively near the triple junction between the North American, Cocos, and Caribbean plates, which provides a large amount of tectonic activity to

FIGURE 4 | (A) Map of final locations of 5 PWSCEs (boxes and numbers), 7 other events (dots) and 14 shots (asterisk), all based on 1-D P-wave velocity model. Dark triangles represent stations with positive corrections, light triangles represent stations with negative corrections. (B) North-South cross section through Fuego vent, sharing latitude coordinates with (A). (C) East-West cross section through Fuego vent, sharing longitude coordinates with (A). (D) 1-D velocity model (E) Histogram of RMS errors for the events (F) Corrections applied to each station in final 1D model.

separate out from the volcanogenic seismicity in any seismic dataset collected at the site. From the 17th to the 23rd of January 2012, there were 75 magnitude 3.0 earthquakes and above within 5 radial degrees of Fuego's summit in the USGS Preliminary Determination of Epicenters (PDE) Bulletin (U. S. Geological Survey, 2017). Eight produce large peaks in real-time seismic amplitude measurements (RSAM) (Endo and Murray, 1991), and of the 34 individual measurements above 50 µm/s on an RSAM plot produced from the vertical component of station NE1, only 13 are due to volcanic processes, 12 from two episodes of tremor, and one from an actual summit emission (**Figure 5A**). A similar plot of RSAM recorded on INSIVUMEH's permanent, short period station FG3 shows many of the same general features present in the record from the closer broadband instrument. Two differences stand out: first the much lower signal to noise ratio present at NE1, and second the much smaller contribution of tremor amplitudes (**Figure 5B**). Units are in counts for FG3 as we were not able to obtain an accurate instrument response for the permanent station.

FIGURE 5 | Real-time Seismic Amplitude Measurement (RSAM) from station NE1 (A) and FG3 (B,C). Each sample in a and b is the mean amplitude of a 60 s, non-overlapping window of data, high pass filtered at 1 Hz. Gray dashed lines are regional tectonic earthquakes, with associated numbers reporting RSAM values (in µm/s for NE1 and uncalibrated counts for FG3) and reported magnitudes of regional tectonic earthquakes. Dotted lines represent peaks generated by volcanic tremor and red line marks the largest observed explosion. Solid gray horizontal line represents an arbitrary cutoff value, below which individual peaks are not described in detail. (C) plots the daily averaged RSAM from FG3 from January to September 2012. The green vertical bar represents the time periods captured in (A,B).

#### Summit Emissions

Emissions from the summit of Fuego are the most obvious and captivating activity we observed during our field occupation. There were different types of emissions from two active vents during the campaign; a "summit vent" and a "flank vent." Each vent exhibits impulsive onset, ash filled plumes, as well as emergent onset, ash poor plumes of white color (**Figure 6**). Signals from summit emissions are recorded across the temporary network and on FG3. Larger explosions show similar-shaped waveforms on all stations, but many even larger degassing events do not register at FG3 due to a generally lower signal to noise ratio at the farther station.

Even though the locations of these emissions appear spatially constant throughout our time in the field, the seismic waveforms generated by these explosions are very diverse. Inter-event times are very sporadic and do not show any correlation with amplitudes in the seismic or acoustic records, nor do plume type, location, or height from visual records. Some emissions have been linked to distinct types of very long period waveforms in previous work (Lyons and Waite, 2011; Nadeau et al., 2011; Waite et al., 2013; Waite and Lanza, 2016) and we continue to see these types of events in 2012. Further examination of these types of very long period events will be discussed in a future publication and is outside the scope of this current study.

#### Tremor

Volcanic tremor has been observed at many volcanoes worldwide and is a broad term covering seismic signals of sustained amplitudes (Konstantinou and Schlindwein, 2003; Chouet and Matoza, 2013). As noted above, tremor at Fuego makes the largest contributions to RSAM measurements of the volcanic processes we observed during our field campaign. We identify two types of tremor during the period of observation. First, broad band tremor with energy between 0.5 and 8 Hz occurs at different intervals throughout the dataset, lasting anywhere from 2 min to over an hour. Second, narrow band harmonic tremor with a fundamental frequency somewhere between 0.5 and 2 Hz with anywhere from three to eight overtones (**Figure 7**). Short, less than 100 s duration episodes are common, as well as episodes lasting longer than 30 min. Both long and short duration harmonic tremor exhibits non-stationary fundamental frequencies shifting as much as 2 Hz, easily identifiable by the strong gliding of overtone frequencies over time. Tremor is visible at the FG3 station, but with lower amplitudes than signals generated by other activity when compared with temporary network stations. Additionally, all but the first two overtones in episodes of harmonic tremor were absent, presumably due to attenuation of higher frequencies along the longer path to the short period station from the summit.

The broadband and harmonic tremor episodes can happen immediately after an explosive event or emerge out of background signal independent of other activity, and other types of activity occur simultaneously with both classes of tremor as well. During several of the episodes of both types of tremor, we observe steady white, ash poor emissions from the summit vent. Flank vent emission is also possible, but not detectable due to the positioning of the cameras.

station. The spectrogram of the vertical component is plotted above the trace of the same time. The lowest trace is from a collocated infrasound sensor. Trace units are normalized. (A) Summit Impulsive (B) Flank Impulsive (C) Summit Emergent (D) Flank Emergent.

#### Rockfalls

Rockfalls are ubiquitous during the observation period, mostly originating near the crater rim and proceeding down the southern flanks to the barrancas below on the order of several episodes an hour. Due to the lack of an active lava flow front, most of the material is sourced from older cooling lava flow terminal edges near the summit or from precariously perched material from more recent explosive events from one of the two active summit vents. Smaller rockfalls initiate at seemingly random intervals due to instability inherent in the location of the source materials, but most of the largest rockfalls take place soon after explosive events being apparently dislodged. These rockfalls posed a significant hazard to personnel during our field campaign and aside from the terrain itself proved the second largest limiting factor in where stations were ultimately located during the occupation.

The fact that most rockfalls occurred to the southwest meant that our time lapse cameras did not capture any good examples of this type of activity. The rockfalls are easily distinguishable in the seismic records though due to the frequency content being

almost exclusively above 10 Hz which distinguishes rockfalls from tremor, the emergent onset of the events followed by a ringing coda, a lack of infrasound signal, and the much larger amplitudes of the events on the stations located to the south (**Figure 8**). The choice of requiring 5 stations to simultaneously trigger to add an event to the catalog was also determined specifically to avoid detecting large amplitude rockfalls. Rockfalls are visible on the FG3 station, but most of the rockfalls activity while the temporary network was operating took place down the southwest flank away from the FG3 station and is therefore difficult to distinguish at FG3 from background signals without a simultaneous examination of the network stations.

#### Phase-Weighted Stacks of Clustered Events

Repeating seismic events in volcanic settings can highlight important physical processes. Interestingly, each of the PWSCEs show distinct signal characteristics (**Figure 3**), and none of the events within the stacks correspond with any type of consistent concurrent observed surface activity. We demean, apply a cosine taper, and apply a two pole, 0.5–5 Hz Butterworth filter forward and backwards (effectively creating a four pole filter) to each signal, and then demean again each detected event in a cluster before creating the phase-weighted stack. In each of the events, the southern stations show markedly lower amplitude signals and later onsets when compared to signals recorded on the northern portion of the temporary network, which is consistent with the events occurring near the vent.

#### PWSCE1

The first cluster contains 96 separate events (**Figure 9A**). The stack shows a small amplitude positive vertical first motion and negative first motion in the radial direction on all stations where a clear first motion is observable. On the NE1 and NW1 stations, a larger amplitude pulse follows for two cycles, and these cycles are identifiable on the SE1 and SW1 stations with much weaker amplitudes. The event shows high energy from 0.5 to 4 Hz at the onset and tapers down to 1–3 Hz after the first 2 s.

#### PWSCE2

The second cluster contains 22 separate events (**Figure 9B**). Low amplitude positive vertical motions precede a strong negative vertical motion at the onset of the event along with pulses away from the vent on the horizontal channels. The NW1 and NE1 stations record three similar cycles which are obscured and have a longer duration and almost ringing coda on the southern stations. All stations record a 1 Hz signal immediately prior to the main large amplitude signal which then extends from 0.5 to 3 Hz lasting 30 s.

#### PWSCE3

The third cluster contains 44 separate events (**Figure 10A**). The event shows a small amplitude positive first motion on all vertical channels as well as first motions away from the vent on the horizontal channels with several higher amplitude cycles following on all channels. A 1 Hz signal persists through the event and leads the main body of the signal which is distributed from 1.5 to 3 Hz, with much less energy below 1.5 Hz.

#### PWSCE4

The fourth cluster contains 7 separate events (**Figure 10B**). The beginning of this signal stack is quite noisy on the NW1 vertical station, and the signal stack is more emergent in nature which makes selecting a clear first motion in any direction difficult. The clearest station is the NE1, and that first arrival is small amplitude positive in the vertical direction. This same pulse can be matched on the NW1 station, with the horizontal channels showing a direction away from the vent at the same time. The spectrograms of the event on the different network station show the main energy arriving several seconds later at the southern stations compared with the northern ones, despite similar onset times in general. The spectrograms also show signal energy from below 0.5 to over 5 Hz despite the bandpass filter having been applied to each constituent of the stack.

#### PWSCE5

The fifth cluster contains 14 separate events (**Figure 11**). This signal is the most emergent of all the stacks, and as such was also the hardest to pick a clear first arrival or true first motion polarity. The event appears to have a small amplitude packet of energy appearing on all the stations. Unfortunately, upon closer inspection of the member events, this early signal does not appear to be consistent across all the events but rather a contaminating feature from one event. The spectrogram shows a strong band of energy at about 0.5 Hz for the duration of the event, with energy distributed through 4 Hz, and showing another brief peak of energy near 2.5 Hz lasting only 3 or 4 s.

#### DISCUSSION

#### Velocity Modeling

We make full use of the user controls allowed by VELEST to ensure that we arrive at a true minimum 1-D velocity model as defined by Kissling et al. (1994). We vary model damping parameters systematically for station correction, velocity, and earthquake locations and find the results are comparable over a wide range of parameters. Changing the

reference station does not appreciably change the relative corrections between stations in the network, instead only shifting absolute values. For example, if we choose the station which observes most arrivals first or last as the reference station, all other network station corrections are positive or negative respectively. If we choose the reference station as a station observing arrivals somewhere after and before other stations in the network, the station corrections distribute more evenly between negative and positive. The variations between modeled velocities of the top model layer reflects the effect of these shifts.

The station corrections we see make sense in the geologic context of Fuego with the stations to the north generally showing negative corrections and therefore faster velocities. We expect this material to be older and to have survived at least one hypothesized edifice collapse (Chesner and Halsor, 1997), meaning it should be physically more compacted and coherent than material to the south. The NE1 station seems to be

particularly fast, but despite looking at various events for possible miss-picks, arrivals do seem to reach this station consistently earlier than others in the network. The variability of material that we had to dig through during station installation would lead us to expect some site effect differences, but the large variability between adjacent stations must be reflecting a very complex three-dimensional velocity structure that we can only hope to approximate with a minimum 1-D model.

In our early runs, we saw velocities in the upper layers consistently falling to nearly 300 ms−<sup>1</sup> whilst reporting station correction factors above 5 s. Adding shots as model inputs keeps velocities in the upper layers of the model higher, and more geologically plausible. The lack of consistently clear arrivals for input to VELEST is the greatest obstacle to minimizing error in event locations, as very few events have sufficiently clear arrivals to pick on all nine stations in the temporary network.

#### Event Locations and Network Resolution

The event locations we report above must be understood in the context of the errors propagated along with the modeling

procedure. Hypocenter location errors for each event located with VELEST single event mode are between 110 and 480 m despite selecting only the most reliable events. We therefore turn to different parts of the modeling process to give us information on how reliable the locations of each earthquake might be, such as tracking the earthquake locations throughout the modeling process.

Events show general trends throughout the modeling procedure. For instance, events which locate in the center of the network in the final model consistently end up in the center of the network with very little variation in the horizontal or the vertical directions. Events which show strong impulsive infrasound signals associated with explosive events but not captured on any of the time-lapse cameras are given a starting location directly below the summit vent, but at 0 m of elevation, which due to Antelope's lack of topography indicated the surface. These events migrated successively closer to the top of the topography at each step of the velocity modeling, indicating a trend to stable locations near the top of the cone.

The only events with an azimuthal gap greater than 180◦ that we did not eliminate from the data set through all phases of velocity modeling were the PWSCEs, which consistently locate closer to the north stations of the network. Because no consistent surface activity occurs in the time-lapse images 1 min before and 3 min after those arrival times, we believe that these repeating events are being generated by subsurface processes not previously observed. However, the lack of any observed activity in this area at Fuego in the years following our field campaign, along with the large arrival timing errors leads us to doubt the accuracy of these locations, which we infer to be restricted to Fuego's active cone.

We gain critical insight to the model space by selecting for our updated a-priori 1-D model one which minimizes both the RMS error of the run as well as the lowest average station corrections for all nine stations. In early runs, the events consistently locate much shallower in the cone, and stayed closer to the vent. Station corrections are much more reasonable with a full network spread around three tenths of a second as opposed to almost a whole second with the initial method of only minimizing RMS of the model. Solutions also stabilize more quickly and show less variability based on the initial reference station. This change greatly increases our confidence in the velocity model reported in **Figure 4**, even if the locations are still not accurate beyond restricting event locations to within the cone.

Two single events which were well constrained from Antelope still located north of the network, and despite the persistence of these locations, the temporary network could not confidently constrain them. The last reported activity in the Acatenango portion of the massif was a series of phreatic eruptions which occurred in 1972 (Vallance et al., 2001), so seismic activity would not be unthinkable. Given the level of tourist activity on Acatenango, even a minor episode would potentially pose a risk to the dozens of people hiking the volcano on any given day. Differentiating these signals from other events in the Fuego vicinity would be even more difficult given that the whole complex is only continuously monitored by one short period station operated by INSIVUMEH. These events may occur deeper in the system but the depths cannot be constrained due to limitations in the temporary stations network aperture, and unmodeled complexity in the true three-dimensional velocity structure.

The five groups of similar events are likely driven by similar sources, although their waveform characteristics are quite different. The particle motions of the main amplitude pulse from the dominant cluster, PWSCE1, shows distinctly retrograde motion (**Figure 12**) indicating a prominent Rayleigh wave arrival. The shallow location of this event, coupled with the

each trace at each station, showing the retrograde motion.

pulse-like signal which decays further from the vent and lower frequency content of the main pulse are remarkably similar to shallow events recorded at Mount Etna, Turrialba, and Ubinas (Bean et al., 2014; Chouet and Dawson, 2016). While we were unable to reliably model this event given the relatively large distance to the nearest stations south of summit, the similarity of the waveforms to those of the well-modeled events at Etna suggests these repeating events likely result from a similar mechanism. Given that Chouet and Dawson (2016) favor fluiddriven sources over a slow rupture dislocations due to better cross-correlation values between recorded data and generated synthetics, we interpret this as a rapid increase in gas pressure within a crack very near the surface. The other event clusters did not have the same dominant Rayleigh wave pulse but their locations within the cone suggest a gas or magma-driven process. However, the short observation period and lack of drastic changes in activity do not allow us to test for temporal evolution of events which would be predicted in a slow brittle failure of poorly consolidated volcanic rock (Bean et al., 2014), again limiting the strength of our conclusions. Figures for the remaining four phase-weighted stacks of clustered events are included in the Supplementary Material.

It should be noted that we tried to identify events from the time windows around the phase-weighted stacks of events on the FG3 short period station operated by INSIVUMEH (**Figure 1**), but the low signal to noise ratio at the recording site made positive event identification impossible even in the stacked data. Adding arrivals from this station would have significantly increased our network aperture and the accuracy of deep earthquake locations, but for the velocity modeling section of this study the recordings at FG3 did not provide any helpful information.

While the occurrence of families of small seismic events suggests repetitive processes, another result this investigation highlights is the complexity of the explosions themselves. As noted above, similar surficial expressions exhibit markedly different seismic signatures. One explanation for this scenario would be that the conduit seals or partially seals between eruptions. Differences in the structure of each seal, how the seal forms, and where and how dramatically the seal fails would all produce different waveforms despite similar locations and otherwise constant inputs from the broader system. Our observations support the eruption mechanism proposed by Nadeau et al. (2011) of a crystal rich mush solidifying and capping the vent, allowing pressures to build until the cap fails mechanically and allows material to escape.

Finally, we attempt to classify seismic events which are associated with explosions and differentiate them from those which are not. Interestingly, none of the events show distinguishing characteristics in frequency content, duration, or impulsiveness of onset; they only differ substantially in whether or not they have an accompanying infrasound signal. But unfortunately, even the presence of an infrasound signal is not always a reliable indicator of strictly subsurface activity as several observed events with varying plume volumes occurred without measured acoustic signals.

#### Repose Period Analysis

The details of individual events such as their locations and waveform characteristics can provide information about the processes responsible. Similarly, a detailed catalog of seismicity can be used to illuminate driving processes more broadly through relatively simple statistics. Varley et al. (2006) apply statistical time-series analysis to volcanic activity at Colima, Tungurahua, Karymsky, and Mt. Erebus volcanoes. The authors show that different periods of activity can be distinguished by the distributions of the repose periods between events and event types. Data are classified as stationary or showing periodicity, clustering, or a trend, which points to events governed by constant processes independent of time or the competition between different processes. If each interval is independent of the one preceding it, the distribution of interval times is exponential and the governing processes in Poissonian in nature. One way to test for Poisson processes is to calculate the coefficient of variation, which is the standard deviation of the between events σ<sup>τ</sup> , divided by the mean interevent time τ , or C<sup>v</sup> = στ τ (Equation 1).

The governing process is Poissonian if C<sup>v</sup> = 1 and clustered if C<sup>v</sup> > 1. We calculate these values from interevent times from several sources which can be found in **Table 1**. Most of the measures of C<sup>v</sup> are slightly greater than 1, and like Varley et al. (2006), we report lower coefficients of variation in subdivided event families. Differences in C<sup>v</sup> imply distinct processes driving the activity, we see the largest difference when separating events by vent of origin or type of event (explosive vs. degassing). Degassing events in our dataset appear Poissonian and explosive events appear clustered. This fits well with a model of constantly degassing magma (Andres et al., 1993; Rodríguez et al., 2004; Lyons, 2011; Nadeau et al., 2011; Waite et al., 2013). However, the relatively short observation period, and therefore small sample number limits our ability to report distributions with any



All times are from the start of first event to the start of next event. The top two rows are seismic arrival times for events detected on five or more stations in the network. Antelope Origins are event origin times as determined by Antelope's dbgrassoc program using the iasp91 velocity model. NE1 Arrivals are human picked event arrival times from station NE1. Remaining rows are seismic arrival times on station NE1 of visually observed events and subsets thereof.

confidence. Increasing the catalog size would help to provide more confident interpretations in the future.

Further investigation of the types of governing processes active at Fuego during periods of background activity through time lapse imagery and seismic event timing from computationally cheap algorithms can extend the analysis to periods of years. For example, recent work by Castro-Escobar (Castro-Escobar, 2017) showed that Fuego's paroxysmal eruptions are statistically independent in time, suggesting that the system recovers to a background state between each eruption. This makes understanding the characteristics of that background state all the more important.

#### A Foundation for Improved Eruption Forecasting

This analysis provides an example of the important information that is useful for starting the process of eruption forecasting. Many volcanoes throughout the world are monitored by one or fewer stations, and while monitoring agencies are adept at using minimal amounts of data to keep local populations safe, it is clear that a better understanding of the monitoring data should yield better forecasts. Ketner and Power (2013) show an example of how close examination of seismicity recorded on a single station during Redoubt volcano's 2009 eruption can provide a richer understanding of the progression of an eruptive event.

In the case of Fuego volcano, INSIVUMEH relies on observers who live on the volcano's flanks together with realtime seismic data from a short-period station about 6 km southeast of the summit. Fuego's larger "paroxysmal eruptions" can produce pyroclastic density currents that threaten nearby population centers and ash clouds that threaten aircraft. While INSIVUMEH has been successful using this approach, we sought to provide more detail that could be incorporated into a better understanding of the volcano in the future. Being able to compare contemporaneously recorded signals at FG3 and a network of stations closer to the vent clarifies the sources of some of the more striking features and increases confidence in classifying activity as an explosion or local rockfalls. It also sheds light on information missing from this record, which could aid in interpreting increases in activity prior to paroxysmal activity.

In cases where only a single station is responsible for monitoring an entire volcano, insights from temporary instrument deployments can shed light on signals recorded at the permanent station and clarify sources of ambiguous signals. Rodgers et al. (2015) provide an example at Telica volcano in Nicaragua of using seismic records and eruption observations to classify activity as belonging to either stable (permitting opensystem degassing) or unstable (where open-system degassing cannot be maintained) phases. This example highlights an instance where low levels of seismicity, normally associated with quiescence can in some cases portend more dangerous activity. In cases where no permanent monitoring happens, temporary deployments during periods of quiescence can provide a baseline for comparison if activity later increases and requires further study to determine if that increased activity could become hazardous.

### CONCLUSIONS

Our proposed template for a temporary monitoring network starts with selecting sites to ensure adequate radial coverage around a volcano. Visual observations recorded by time lapse cameras help aid later interpretation. Ideally, the observation period is as long as possible, but even a short time can be leveraged for deeper understanding. Data analysis should begin with classifying different types of emissions, if any, and identifying signals which do not manifest as surface activity. Utilizing a pattern identification algorithm, in our case, REDPy, and identifying a 1-D velocity model can be quickly and easily done following our methods.

Several results are reported. First, by classifying local seismic signals based on observed surface activity, we can be more confident in knowing what is happening on the volcano even when visibility is poor. Second, we have identified repeating events not directly tied to surface activity which is evidence that the volcanic plumbing system includes some level of complexity which should be further investigated. Third, despite the difficulties of constraining exact arrivals for most events in our catalog, we identify a reasonable 1-D velocity model which can itself serve as a starting point for future analysis, and we can be more confident in this model due to the exhaustive analysis done to produce it.

This work provides examples of analytical operations which can help to establish baseline levels of activity at open vent volcanic systems. The challenge with these systems from a monitoring standpoint is that precisely because of their relatively low levels of activity, forecasting changes in activity often comes down to paying attention to small details and how they relate to

one another. Without a baseline to compare to, forecasting can never be more helpful than simply guessing based on experiences at other volcanoes.

#### AUTHOR CONTRIBUTIONS

GW, KB, and GC contributed to the conception and design of this work. KB wrote the first draft of the manuscript, performed initial data processing for the temporary array, statistical analysis, and velocity modeling. GW supervised all work, advised KB on data processing methods and provided revisions for the manuscript. GC facilitated access to Guatemalan field sites, performed initial data processing for the permanent station, gave insights to monitoring challenges and needs, and provided revisions for the manuscript. All authors read and approved the submitted manuscript.

#### ACKNOWLEDGMENTS

The authors wish to thank Eddy Sanchez, GC, Almilcar Calderas, and Edgar Barrios of INSIVUMEH and OVFUEGO for institutional support, lodging, and field contacts. Jake Anderson, Luke Bowman, Lloyd Carothers, Rüdiger Escobar Wolf, Anthony Lemur, John Lyons, Armando Pineda, Lauren Schaefer, Cara Shonsey, Josh Richardson, and James Robinson aided in instrument deployment, maintenance, and collection. Juan Martinez and the rest of the porters from La Soledad, Chimaltenango carried most of our food, water, and shelter for the field deployment. DISETUR and the PNC from Escuintla provided logistical support and physical protection respectively. Returning from the field, Kenny Rodriguez helped identify activity based on images, and Federica Lanza, Hans Lechner,

#### REFERENCES


Simone Puel graciously picked events for the pick uncertainty analysis. Chet Hopp helped get VELEST compiled and running on local hardware and greatly aided in KB's understanding of the algorithm. Collaboration with Federica Lanza on the MATLAB code used to explore the VELEST results as well as discussions about those results made this project possible. Funding for this work was provided by the GMES department of Michigan Technological University and grants from the National Science Foundation (#1053794 and #0530109). The seismic instruments were provided by the Incorporated Research Institutions for Seismology (IRIS) through the PASSCAL Instrument Center at New Mexico Tech. Data collected will be available through the IRIS Data Management Center. The facilities of the IRIS Consortium are supported by the National Science Foundation under Cooperative Agreement EAR-1261681 and the DOE National Nuclear Security Administration. The facilities of IRIS Data Services, and specifically the IRIS Data Management Center, were used for access to waveforms, related metadata, and/or derived products used in this study. IRIS Data Services are funded through the Seismological Facilities for the Advancement of Geoscience and EarthScope (SAGE) Proposal of the National Science Foundation under Cooperative Agreement EAR-1261681. This manuscript was significantly improved thanks to the feedback received from Valerio Acocella, Nicolas Fournier, and two reviewers.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart. 2018.00087/full#supplementary-material


Redoubt Volcanoes - invited," in Seismological Society of America Annual Meeting (Reno, NV).


MATLAB (2015). Release 2015a. Natick, MA: The MathWorks, Inc.


**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 Brill, Waite and Chigna. 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.

## Alaska Volcano Observatory Alert and Forecasting Timeliness: 1989–2017

Cheryl E. Cameron<sup>1</sup> \*, Stephanie G. Prejean<sup>2</sup> , Michelle L. Coombs <sup>3</sup> , Kristi L. Wallace<sup>3</sup> , John A. Power <sup>3</sup> and Diana C. Roman<sup>4</sup>

 *Alaska Volcano Observatory, Alaska Division of Geological and Geophysical Surveys, Fairbanks, AK, United States, Volcano Disaster Assistance Program, Alaska Volcano Observatory, U.S. Geological Survey, Anchorage, AK, United States, Alaska Volcano Observatory, Volcano Science Center, U.S. Geological Survey, Anchorage, AK, United States, <sup>4</sup> Department of Terrestrial Magnetism, Carnegie Institution for Science, Washington, DC, United States*

The Alaska Volcano Observatory (AVO) monitors volcanoes in Alaska and issues notifications and warnings of volcanic unrest and eruption. We evaluate the timeliness and accuracy of eruption forecasts for 53 eruptions at 20 volcanoes, beginning with Mount Redoubt's 1989–1990 eruption. Successful forecasts are defined as those where AVO issued a formal warning before eruption onset. These warning notifications are now part of AVO's Aviation Color Code and Volcanic Alert Level. This analysis considers only the start of an eruption, although many eruptions have multiple phases of activity. For the 21 eruptions at volcanoes with functioning local seismic networks, AVO has high forecasting success at volcanoes with: >15 years repose intervals and magmatic eruptions (4 out of 4, 100%); or larger eruptions (Volcanic Explosivity Index (VEI) 3 or greater; 6 out of 10, 60%). Therefore, AVO successfully forecast all four monitored, longer-repose period, VEI 3+ eruptions: Redoubt 1989–1990 and 2009, Spurr 1992, and Augustine 2005–2006. For volcanoes with functioning seismic monitoring networks, success rates are lower for: volcanoes with shorter repose periods (3 out of 16, 19%); more mafic compositions (3 out of 18, 17%); or smaller eruption size (VEI 2 or less, 1 out of 11, 9%). These eruptions (Okmok, Pavlof, Veniaminof, and Shishaldin) often lack detectable precursory signals. For 32 eruptions at volcanoes without functioning local seismic networks, the forecasting success rate is much lower (2, 6%; Kasatochi 2008 and Shishaldin 2014). For remote volcanoes where the main hazard is to aviation, rapid detection is a goal in the absence of *in situ* monitoring. Eruption detection has improved in recent years, shown by a decrease in the time between eruption onset and notification. Even limited seismic monitoring can detect precursory activity at volcanoes with certain characteristics (intermediate composition, longer repose times, larger eruptions), but difficulty persists in detecting subtle precursory activity at frequently active volcanoes with more mafic compositions. This suggests that volcano-specific characteristics should be considered when designing monitoring programs and evaluating forecasting success. More proximally-located sensors and data types are likely needed to forecast eruptive activity at frequently-active, more mafic volcanoes that generally produce smaller eruptions.

#### Edited by:

*Nicolas Fournier, GNS Science, New Zealand*

#### Reviewed by:

*Luis E. Lara, Sernageomin, Chile Alessandro Bonforte, Istituto Nazionale di Geofisica e Vulcanologia, Italy*

> \*Correspondence: *Cheryl E. Cameron cheryl.cameron@alaska.gov*

#### Specialty section:

*This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science*

Received: *01 March 2018* Accepted: *04 June 2018* Published: *03 July 2018*

#### Citation:

*Cameron CE, Prejean SG, Coombs ML, Wallace KL, Power JA and Roman DC (2018) Alaska Volcano Observatory Alert and Forecasting Timeliness: 1989–2017. Front. Earth Sci. 6:86. doi: 10.3389/feart.2018.00086*

Keywords: eruption, volcano, forecasting, Alaska, volcano monitoring, volcanic alerts

### INTRODUCTION

Since its inception in 1988, the Alaska Volcano Observatory (AVO; a joint program of the U.S. Geological Survey, the Geophysical Institute of the University of Alaska Fairbanks, and the State of Alaska Division of Geological and Geophysical Surveys) has been responsible for providing timely and accurate information on volcanic hazards, and warnings of impending volcanic activity, to local, state, and federal officials and the public. In December, 1989, a passenger jet inadvertently flew through a cloud of volcanic ash from Mount Redoubt, causing the loss of power to all four engines, and forcing an emergency landing in Anchorage. The jet landed successfully with no loss of life, but the aircraft sustained \$80 million in damages (Casadevall, 1994). This incident dramatically demonstrated the vulnerability of jet aircraft to volcanic ash and prompted the ongoing effort to instrument Alaska's volcanoes, which present a constant threat to trans-Pacific aviation in addition to hazards posed to communities in the State. Alaska has 54 volcanoes considered historically active (since 1760) and about 100 volcanoes active in the past 11,000 years (Cameron and Nye, 2014; Cameron and Schaefer, 2016). Since 1989, there have been at least 53 eruptions at 20 volcanoes in Alaska (**Figure 1**). AVO has also provided formal notification on 23 episodes of unrest that did not lead to eruption. AVO employs several monitoring approaches including seismic stations at 32 volcanoes, continuous Global Positioning System (GPS) stations at 8 volcanoes, regional and local infrasound sensors, and web cameras. In addition to ground-based monitoring, AVO relies on satellite remote sensing data, lightning detection, annual gas measurements (only for Cook Inlet volcanoes), local observers, and pilot reports (Dixon et al., 2017).

In reviewing AVO's public warnings, we use the term "forecast" to describe statements issued prior to expected eruptive activity. AVO's forecasts are typically relatively imprecise statements about the timing and nature of expected activity and are based on a synthesis of available monitoring data that may include eruptive history as well as seismic, geodetic, infrasound and satellite remote sensing data, gas measurements, and visual observations (Power et al., 1995). Ideally, Observatory notifications extend beyond forecast and into predictions (Swanson et al., 1985) and encompass many additional pieces of information: what is the probability of volcanic eruption? If there is an eruption, when, where, and how long will it last; how big will the eruption likely be; and what hazardous effects will it have? For the numerous and variable eruptions in Alaska over the last 30 years, it is impossible to fully evaluate how well AVO provided all of these pieces of information to the public and other agencies, as we do not have long-standing repeat surveys with stakeholders. AVO forecasts typically do not prompt evacuations because most eruptions are remote and the main hazard is the impact of ash on aviation (Guffanti et al., 2010). The task of forecast evaluation is especially hampered by the fact that many eruptions were unmonitored by ground-based instruments, at remote volcanoes, and it is sometimes difficult to determine when or even if an eruption occurred. In order to look broadly across an entire arc's worth of eruptions, we focus on one important aspect of volcano monitoring: was a formal notice of warning issued

prior to the onset of the eruption? Complementing this paper, a detailed analysis of the time lag between individual explosions and AVO calldowns is found in Power and Cameron (2018), and an analysis of seismic rate anomalies preceding Alaska eruptions is in Pesicek et al. (2018).

In this paper, we report on the advance warnings issued by AVO, and investigate how timeliness varies with parameters such as monitoring capabilities, erupted magma composition, most recent repose interval (a broad proxy for whether a system is "closed" or "open"), and Volcanic Explosivity Index (VEI). We show that for larger eruptions of andesitic magmas at seismically monitored volcanoes, AVO consistently provides timely advance warnings (forecasts). For smaller eruptions of more mafic magma at frequently active volcanoes ("open systems") forecasting is less successful, and for volcanoes without any ground-based monitoring, not surprisingly, forecasting is extremely difficult. These results can help guide future monitoring strategies in Alaska, where only ∼32 of over 100 potentially active volcanoes have any ground based monitoring, as well as in other volcanically active regions.

#### METHODS

#### Monitoring Data

AVO, like other volcano observatories, monitors volcanoes to "detect and correctly interpret the geophysical phenomena that result from rising magma in the earth's crust, in order to provide early and accurate warnings of impending eruptions" (Moran et al., 2008). AVO's forecasting efforts have principally relied on change detection in one or more monitoring data streams. Of those, seismic data has been most relied upon to provide real-time knowledge of a volcanic system. Some volcanoes erupt with no or minimal detected pre-eruptive seismicity (e.g., Okmok, Johnson et al., 2010; Fee et al., 2017; Pavlof, Waythomas et al., 2017). However, in general, by tracking increases in the rate, magnitude, and frequency content of earthquakes located near the volcano along with other changes in the character of the seismicity, it is often possible to forecast volcanic eruptions (Minakami, 1961; Shimozuru, 1971; Chouet et al., 1994; Ewert, 2007; McNutt and Nishimura, 2008; White and McCausland, 2016; McCausland et al., 2017). This is accomplished with a network of seismic instruments around a volcano, and AVO has required a minimum of four operating instruments with stable telemetry at a volcano to consider it "seismically monitored" (McGimsey et al., 2008). In 1989, AVO only monitored Augustine, Redoubt and Spurr volcanoes with real-time seismic networks of four or more stations (Power et al., 1993). In 1995, AVO began a major expansion of its seismic monitoring program, eventually operating seismic networks on as many as 34 volcanoes (Dixon et al., 2012).

Seismic data are analyzed in concert with other data types when and where available. These data types include satellite imagery that detects geomorphic change and provides information about surface temperatures and volcanic clouds (Wessels et al., 2013; Ramsey et al., 2015); continuous GPS or interferometric synthetic-aperture radar data that reveals surface deformation (Cervelli et al., 2010; Lu and Dzurisin, 2014); gas measurements (Werner et al., 2012, 2017; Lopez et al., 2013); and on-site visual and thermal imaging that reveal changes to the edifice such as ice melt, heating, or increased fumarolic activity (Bleick et al., 2013; Wessels et al., 2013). Increasingly over the past several years, infrasound arrays have been used to detect atmospheric disturbances that result from volcanic activity (De Angelis et al., 2012; Fee and Matoza, 2013). For our current analysis, however, we evaluate AVO's effectiveness depending on the presence or absence of seismic monitoring, since this is the most likely real-time data stream in place at an Alaska volcano for the longest period of time, and has often proved reliable for forecasting imminent eruptive activity (e.g., Power et al., 1994, 1995; Power and Lalla, 2010; Ruppert et al., 2011; Buurman et al., 2013; McCausland et al., 2017).

#### Forecasting Approach

AVO conducts routine checks of all incoming data: seismic, satellite remote sensing, webcam, and community observations, and posts daily (or more frequently, as needed) reports to a common digital log system. In addition, automated alarms are in use for several data streams, including seismic, infrasound, lightning, ash clouds detected by satellite imagery, and emails sent to the website. AVO duty staffing consistently includes a (1) remote sensing scientist—who reviews and reports on satellite and webcam data and alarms, (2) seismologist—who reviews and reports on seismic and infrasound data, and (3) staff scientist who incorporates all such data together and issues formal notifications and makes and receives calls from interagency partners and the public. Staff who receive specific alarms assess their validity and report directly to duty staff. AVO may contact (or receive observations from) citizens, pilots, mariners, and others at remote sites for additional information as needed. In recent years, the USGS National Earthquake Information Center conducts scheduled checks of seismic data afterhours. Volcanoes at elevated color code levels or with unusual activity are monitored more closely—the frequency of checks on these volcanoes depends on the intensity of the activity, and the effectiveness of various alarms for that particular volcano, and extends up to 24/7 in-office staffing during larger eruptions with more major impacts. AVO collaboratively makes decisions about issuing color code/alert-level changes and forecasts through group discussion, and the ultimate responsibility rests with the Scientist-in-Charge (SIC). During times of imminent or ongoing eruption, responsibility is delegated to duty staff members who have authority to issue warnings without compromising timeliness. Although AVO has crafted event trees (Newhall and Hoblitt, 2002) for some unrest episodes, AVO more commonly solicits staff opinions and data during meetings and conference calls, if time allows. These meetings and calls sometimes involve non-AVO (usually USGS) scientists. Within the formal definitions of the color codes/alert levels (Gardner and Guffanti, 2006), assignment of color codes/alert levels are crafted for each volcano based on AVO's level of knowledge and past experience with eruptions at that volcano. AVO maintains close telephone contact with critical agency partners such as the National Weather Service (NWS), the Federal Aviation Administration (FAA), and the State of Alaska Department of Homeland Security

and Emergency Management (DHSEM) so that each agency may provide updates on critical information. After making a color code/alert-level decision, AVO begins a formal calldown process to agency partners with urgent notification needs, such as the FAA, followed by issuing a written notification. This paper deals only with the timing of written notifications, but detailed information on interagency communication is given in Neal et al. (2010) and details of the calldown process are available in Power and Cameron (2018).

#### Notification Schemes—Color Code and Volcano Alert Levels Through Time

AVO's notification history can be broadly grouped into three time periods.


Since 1989, AVO has issued ∼300 color code changes, in addition to many written notifications of unrest or eruption that were not associated with a color code change.

#### Event Classification and Evaluation of Forecasting Effectiveness

We examined all Alaska eruptions and AVO notifications since the onset of the 1989–1990 eruption of Redoubt. Prior to this date, records of public statements issued by AVO are incomplete. Notifications are classified on a modified scheme from Winson et al. (2014), and are further examined with respect to individual eruption and volcano characteristics (**Supplementary Data Sheet 1**). We classify each notification/eruption pair with respect to the definition used by AVO at the time of the alert, so non-color code notifications can qualify as warning and forecast of an eruption, or, more commonly, formal notification of an eruption already in progress.

#### Descriptions of Notification Classes


#### Volcano and Eruption Characteristics

There are several characteristics that make this dataset heterogeneous. We tested the following characteristics (summarized in **Table 1**) to see how specific differences among volcanoes and eruptions affect AVO's forecasting ability.

#### Seismic Monitoring

AVO currently considers 32 volcanoes to be seismically monitored. This number has increased from three (in 1989), and has fluctuated based on individual network health. For this analysis, events are considered seismically monitored if the volcano has a local seismic network, comprised of four or more seismometers, operating at the time of the event, and operating with sufficient time before the event in order to characterize background seismicity, following Buurman et al. (2014) and TABLE 1 | Notifications/eruption-unrest classification (modified scheme from Winson et al., 2014).


*(Continued)*

#### TABLE 1 | Continued


*Event start: Time eruption or unrest (for non-eruptive events) began. VEI from Smithsonian Global Volcanism Program (2013). See* Supplementary Data Sheet 2 *for compositional information; A, Andesite; BA, Basaltic andesite; B, Basalt. Notice time delta, in days* = *notification latency for "Detect Only" eruptions, with positive values, in days, negative values indicate forecast lead time for "Good" eruptions.*

Pesicek et al. (2018). Events with a local seismic network and not formally listed as monitored by AVO (often due to insufficient time to categorize background seismicity, e.g., Pavlof, 1996) are marked "YES-unofficial" in **Table 1**. Unrest events with a local, degraded seismic network (e.g., Veniaminof, 2015) retain a "YES" for seismic monitoring status if AVO did identify elevated seismicity and use it as the basis for an elevated color code. Although many events in our list barely meet this standard, other

volcanoes have more dense seismic networks with as many as 17 instruments [Spurr has 17; Okmok and Akutan follow with 13], allowing for better detection of subtle seismic precursors (Dixon et al., 2012). Seismic monitoring status at time of event is based on network health analyses by Pesicek et al. (2018) and Buurman, et al. (2014). For those events with seismic monitoring, we also briefly examine whether or not precursory seismicity was detected by our network, as do Pesicek et al. (2018).

#### Composition of Erupted Products

Events (unrest and eruptions) are grouped into andesitic (57– 63 wt. % SiO2), basaltic andesite (52–57 wt. % SiO2), and basaltic (<52 wt. % SiO2) categories, following LeMaitre et al. (2002). No recent eruptions or unrest have been associated with more evolved magmas. We give the composition of magma erupted during a particular eruption, if known. For periods of unrest with no eruption, or for eruptions when the erupted composition is unknown, we have compiled the compositions from latest Holocene or historical activity and use that as a proxy for composition of erupted products (**Supplementary Data Sheet 2**).

#### Volcanic Explosivity Index (VEI)

The VEI numbers used here are extracted from the Smithsonian Global Volcanism Program website<sup>1</sup> (2013, https://volcano. si.edu). In this dataset, those events that AVO classifies as "uncertain" or "unrest without eruption" have no assigned VEI, although the Smithsonian database may record a VEI for those events.

#### Determining Event Onset

For all instances of unrest or eruption, the "start date" given in the dataset is either the earliest known date of significant unrest (for UwE events), or the known date of first "explosive ejection of fragmental material, the effusion of liquid lava, or both," following the Smithsonian Global Volcanism Program's definition of eruption (Siebert et al., 2010; Global Volcanism Program, 2013), using eruption information from the Alaska Volcano Observatory online database<sup>2</sup> (https://avo.alaska.edu). This definition also requires that eruptive activity within 90 days of previous activity be counted as the same eruption as the previous eruptive activity. Hiatuses of 90 or more days begin a new eruption. While this arbitrary cut-off for breaking activity into discrete eruptions is easy to apply where eruptive dates are confidently known, the 90 day cut-off may not be appropriate for volcanoes like Cleveland and Veniaminof, which tend to have protracted eruptions that may have breaks longer than 90 days, resulting in numerous, near-identical "eruptions" over a period of years. Mount Cleveland is classified as not "seismically monitored" for the purposes of this analysis, and has been in near-continuous eruption since 2005—any apparent lulls greater than 90 days result in the creation of a new "eruption." As Cleveland is both frequently-erupting and considered not seismically monitored (Cleveland has just two seismometers and an infrasound network, as of 2015), AVO tends to keep Cleveland at an elevated color code—Yellow or Orange—for very long periods (month to years), resulting in Cleveland erupting at elevated color codes, although the color code elevation may have occurred months prior. To correct this problem, although Cleveland has 12 eruptions between 2005 and 2017, we characterize these as a single eruptive period. Thus, because of the nature of Cleveland's activity, and AVO's response to it, these results are presented with only five total eruptions for Cleveland—four eruptions between 1988 and 2004, and the extended eruptive period between 2005 and 2017.

This definition of event onset includes phreatic explosions, although for eruptions where an initial phreatic explosion is not well-verified, a later date of more certain activity is used. At unmonitored volcanoes, a phreatic explosion may be AVO's first notice of activity at the volcano. Unfortunately, for many unmonitored volcanoes, these initial events are poorly reported, creating an inability to discern whether the activity was phreatic or not. These initial events may also constitute the only activity for that event. To explore how these uncertainties in timing of phreatic eruptions impact our results, we examined the dataset twice—both considering initial phreatic explosions as the eruption onset and again, considering the initial phreatic explosion as precursory activity. There is no significant difference in our results whether or not phreatic explosions are used as the onset of "eruption," except in cases where a single phreatic event composes the entire eruption, as with Kanaga, 2012 and Fourpeaked, 2006, where those eruptions would simply be removed from the dataset. In no case does changing the eruption start date to include only known magmatic activity alter a "Detect Only" result to "Good," although it can significantly shorten the lag time between eruption onset and AVO notification.

#### Recurrence Interval (Proxy for Closed vs. Open System)

This value is the difference between the start date of an eruption and the start date of the most recent prior eruption, although some eruptions (notably those at Veniaminof, Cleveland, and Shishaldin) extend for years. Start dates are used rather than end dates because start dates are better characterized across the dataset; not all events have well-known end dates. Values are rounded to the nearest tenth of a year. When the event month or day is not known, the 15th of June is arbitrarily used. These values are listed as "00" in **Supplementary Data Sheet 1**; this uncertainty occurs for three unrest events and two uncertain events.

#### RESULTS

Overall classification results are shown in **Figure 2**. AVO successfully forecast seven of the 21 seismically monitored eruptions (33%; **Figure 3**), including eruptions at Augustine, Redoubt, Spurr, Veniaminof, and Shishaldin. AVO has successfully forecast all eruptions at Augustine, Redoubt, and Spurr, representing 3 of the 8 seismically monitored volcanoes that have erupted since 1989. The unforecasted eruptions at seismically monitored volcanoes are from five volcanoes: Okmok, Pavlof, Kanaga, Veniaminof, and Shishaldin. At unmonitored volcanoes, AVO successfully forecast 2 out of 32 (6%) eruptions.

yellow), 19% Uncertain (18 events, dark blue), and 1% False Alarm (1 event, pink).

FIGURE 3 | Event classifications with respect to presence of seismic monitoring. "SEIS YES" on left indicates events with seismic monitoring, "SEIS NO" on right for events without seismic monitoring. Detect Only shown in medium blue, Good in orange, Missed in gray, Uwe (Unrest without Eruption) in yellow, Uncertain in dark blue, and False Alarm in pink.

#### Results by Notification Class

• Good: The dataset contains nine eruptions where AVO elevated the color code prior to eruption. Most (7) of these occurred at seismically monitored volcanoes. Seismically monitored eruptions with "Good" notification include Augustine 2005–2006, Redoubt 1989–1990 and 2009, Shishaldin 1999, Spurr 1992, and Veniaminof 2002 and 2013. Seismically unmonitored eruptions with "Good" notification are Kastaochi 2008 and Shishaldin 2014.


#### Results by Volcano Characteristics Seismic Monitoring

Forty-seven events (out of 95) were seismically monitored; this includes 21 eruptions (**Figure 3**). For the 21 eruptions with seismic monitoring, seven (33%) were classed as "Good;" and 14 (67%) were classed as "Detect Only." No seismically monitored eruptions were "Missed." Forty-eight events were not seismically monitored, including 32 eruptions. For these 32 unmonitored eruptions, two (6%) were classed as "Good;" 20 (63%) were classed as "Detect Only," and 10 (31%) were "Missed." The following characteristics of repose time, composition, and VEI necessarily include results for both monitored and unmonitored volcanoes.

#### Repose Time

The interval between an eruption (or unrest event) and the start of the previous eruption at a volcano varies from 0.3 years to 11,000 years, with a median repose interval of 3.3 years. Repose time is not given for Takawangha and Little Sitkin, due to very sparse geologic knowledge. Neither of these volcanoes has erupted in historical time. Twenty-five events have a recent repose period greater than 15 years: 12 eruptions and 13 unrest episodes. Of the 12 eruptions, five (42%) were classed as "Good," six (50%) were classed as "Detect Only," and one (8%) was classed as "Missed" (**Figure 4**). Seventy events have recent repose periods less than 15 years, including 41 eruptions, 17 uncertain events, 11 unrest events, and one false alarm. For the 41 eruptions with repose periods less than 15 years, four (10%) were classed as "Good;" 28 (68%) were classed as "Detect Only;" and nine (22%) were classed as "Missed."

#### Erupted Product Composition

Twenty-four events in the dataset are classed as andesitic (A), including four of the nine eruptions classed with "Good" notification (**Figure 5**). Fifty-three events are basaltic andesite (BA; three "Good"). Sixteen events are basaltic (B;

intervals of less than 5 years, 5–10 years, 10–15 years, and greater than 15 years. Detect Only (medium blue), Good (orange), Missed (gray), UwE (Unrest without Eruption, yellow), Uncertain (dark blue), and False Alarm (pink).

two "Good") (**Supplementary Data Sheet 2**). There are no dominantly dacitic or rhyolitic recent eruptions. One volcano, Amukta, has no compositional data. A few volcanoes have recent eruptions with compositional data that span the basaltic andesite—andesite range: notably Cleveland, Kasatochi, and Kiska.

#### Volcanic Explosivity Index

VEI values range from not applicable (for unrest and uncertain events) to 4 (**Figure 6**). There are three VEI 4 eruptions in the dataset: Kasatochi, 2008; Okmok, 2008; and Spurr, 1992, and two are "Good" with one "Detect Only" (Okmok, 2008). There are 15 VEI 3 events, 21 VEI 2 events, and 13 VEI 1. Seven of 18 (39%) of the VEI 3+ eruptions have "Good" notification; the

4. Detect Only (blue), Good (orange), and Missed (gray).

others are "Detect Only" (61%). For VEI less than 3, two of 35 (6%) are "Good," 23 (66%) are "Detect Only" and 10 (29%) are "Missed."

#### Post-notification Analysis

We can also examine these results from a post-notification perspective to answer the question: how often does an initial elevation to Yellow result in an eruption? For those volcanoes with formally-declared seismic monitoring and post-dating the development of AVO's color code (1990), AVO initially raised the color code at a volcano from Green to Yellow 31 times; eight (26%) of these elevations represent an eruption already in progress; 17 Green to Yellow elevations (55%) resulted in either no eruption or an uncertain eruption. Only 6 of these initial Green to Yellow elevations (19%) were followed by an eruption (**Figure 7**).

Although all of the eruptions classed as "Good" have an initial color code change from Unassigned or Green to Yellow, eruptions with seismic monitoring classed as "Detect Only" include three jumps from Green to Orange (Pavlof, 2013; May 2014, and November 2014) and two elevations straight from Green to Red (Okmok, 2008 and Pavlof, 2016). The "False

monitored volcanoes. Elevations from Green to Orange or Red occurred at volcanoes where an eruption was already in progress. The elevation from Yellow to Red is the false alarm event at Spurr in October and November, 1992. For the 31 initial elevations from Green to Yellow, the pie chart shows the eventual outcome of eruption, eruption already in progress, and unrest or uncertain eruption.

Alarm" (two elevations, 1 month apart) jump from Yellow to Red.

#### DISCUSSION

We can compare our notification class results to those of Winson et al. (2014) and their global volcano alert classification results, although they use different definitions and include only 20 Alaska events. For Winson and others' global dataset, 14% of those volcanoes with Level 0 or Level 1 monitoring (comparable to this analysis' "seismically unmonitored") had "Timely" or "Almost" notifications (comparable to this analysis' "Good.") For volcanoes they describe as having Level 2 or higher monitoring, 21% of those eruptions had "Timely" or "Almost" notifications. Our study finds a bigger improvement in notification/forecasts between unmonitored and monitored, with only 6% success rates for unmonitored and 33% for monitored. The outcomes between these two studies are different for several reasons: the studies analyze different sets of events; they have different criteria for "monitored" status (including differentiating monitored status at the time of each event); and this study of an individual observatory is able to assess notification/forecasting success based on specific observatory alert levels, rationale, and procedures for use. For example, the USGS Volcano Alert Level System used by AVO and other U.S. Observatories states that Yellow indicates unrest behavior, while Orange is appropriate for either increased unrest OR lowlevel eruption (Guffanti and Miller, 2013). Historically, AVO has also included low-level ash emission at Yellow, and still has a tendency to call very low-level eruptive activity "Yellow" (Brantley, 1990). Therefore our evaluation of notification success only requires that a volcano be elevated to Yellow (or a noncolor code notification, for non-monitored volcanoes prior to 2007) to be included as appropriate color code elevation during unrest or prior to eruption. Another instance of assessing success within an institution's specific needs is the case of long unrest periods at remote volcanoes. Alaska may be more willing than other, more populated areas to use elevated color codes for long periods of unrest, as these remote volcanoes do not often need disruptive mitigation measures. It is important to analyze alert level use within the context of the agency using it and the local users, in order to ensure "apples to apples" comparisons.

We can more accurately resolve the characteristics that influence successful forecasting by grouping characteristics (eruption size, repose interval, erupted product composition, and presence/absence of seismic monitoring; **Figures 8**, **9**), as individual volcano and eruption characteristics do not fully explain the variation in forecasting success rates. Many of these factors are not independent of each other, and are instead highly correlative (e.g., Passarelli and Brodsky, 2012). However, analyzing the success rates in overlapping regions of two or more characteristics (**Figure 8**) shows additional insight into groups of factors that correlate with eruption forecasting success or failure, and, more importantly, guidance that could improve AVO's forecasting abilities.


### Efficacy of Seismic Monitoring

AVO has much better forecasting success at seismically monitored volcanoes (33%) compared to non-monitored volcanoes (6%). Clearly, in situ seismic monitoring, preferably in concert with other geophysical instrumentation, is essential to improving abilities to successfully forecast a volcanic eruption (Sparks, 2003, and many others; Ewert, 2007; McNutt, 2008; Tilling, 2008). The Moran et al. (2008) report suggests that part of the instrumentation for a well monitored volcano includes at least one seismometer within 5 km of the vent (also see White and McCausland, 2016). It is interesting to note that the forecasting successes at Augustine (closest seismic station AUP = 0.6 km), Redoubt (RSO = 2.5 km), and Spurr (CP2 = 0.3 km) have seismometers within 2 km of the vent, but the closest seismometers at volcanoes without steady forecasting success (Pavlof, PV6 = 4.4 km; Veniaminof, VNSS = 7.8 km; Shishaldin, SSLS = 5.4 km) have seismometers more than 4 km from the vent.

Seismic instrumentation at Alaska volcanoes does yield a decrease in the lag time between eruption onset and AVO notification of the eruption for those eruptions that were not successfully forecast. The average notification delay for nonseismically monitored "Detect Only" eruptions for eruptions older than 10 years ago is 6 days (delay periods grouped into 3-h bins for delays of less than 1 day; days counted as integers for delays greater than 1 day), but over the same time period, the average notification delay for seismically-monitored "Detect Only" eruptions is 2 days (excluding a 76-day delay outlier from the 2004 eruption of Shishaldin). Looking at the most recent 10 years, there is only one non-seismically monitored eruption (notification delay of 9 days; Bogoslof 2016–2017), and the average delay for seismically monitored eruptions drops to ¼ day. Seismic monitoring decreases notification delay, and this notification lag is also decreasing with time. Faster eruption notification is likely due to the concurrent use of infrasound, lighting detection, alarm algorithms, improved satellite data, and increased intra- and extra-agency communication.

#### Closed System Successes

AVO has successfully forecast all four seismically-monitored VEI 3+ eruptions for volcanoes with repose periods longer than 15 years, as well as the closed-system but unmonitored VEI 4 eruption of Kasatochi in 2008 (lower right quadrant of **Figures 8**, **9**). AVO provided 130 days advance warning of the 2009 eruption of Redoubt, 19 days for Spurr 1992, 3 days for Augustine 2005–2006 (3 days prior to the December 2 first phreatic explosion; the initial magmatic explosion occurred on January 11, 2006), and about 18 h for the Redoubt 1989–1990 eruption. These eruptions had precursory seismicity, recorded on monitoring networks (Pesicek et al., 2018), allowing for substantial forecast lead times; the precursory seismicity and AVO response is detailed in Power and Cameron (2018). The single longer-repose period and monitored event which was not successfully forecast is Kanaga, 2012. This event is an outlier in many ways: it is smaller than the others (VEI 2 instead of 3 or 4), phreatic (Herrick et al., 2014) rather than magmatic, located

with both seismic monitoring and repose times longer than 15 years, Kanaga 2012 (the single not-forecast event) is the only phreatic eruption of this group.

outside of Cook Inlet (Alaska's most populous region), and had a recent repose time of just 18.1 years.

A final closed system forecasting success occurred for the VEI 4 eruption of Kasatochi in 2008. Kasatochi has no local seismic stations. This eruption had precursory seismicity of **M** >2, large enough to be recorded on seismic stations operated by AVO on Great Sitkin, Korovin, and Kanaga, beginning about 1 month before the eruption. The largest earthquake was a magnitude 5.8. Felt earthquakes were noted by U.S. Fish and Wildlife Service employees stationed on Kasatochi for about a week before eruption onset (Neal et al., 2011; Ruppert et al., 2011; Nye et al., 2017). This substantial precursory seismicity, attributed in part to rapid magma ascent (Neill et al., 2015), enabled eruption forecasting despite the absence of a local seismic monitoring network (Waythomas et al., 2010).

#### Open System Challenges

Nearly all volcanoes in this dataset that erupt more frequently and could be considered "open system" produce basalt or basaltic andesite eruptive products. Cleveland and Makushin are both classed andesite and are thus exceptions to this generality.

All of the seismically monitored volcanoes with repose times < 15 years are basaltic or basaltic andesite in composition (upper right corner of **Figure 8**). These open system, lowviscosity, variably-sized eruptions with seismic monitoring are generally not forecast (three out of 16) due to their notable lack of detected precursory seismicity (Pesicek et al., 2018). Most (13 of 16; 81%) of these eruptions are from two persistently active volcanoes (Pavlof and Veniaminof, neither of which has a seismometer within 4 km of the vent) and most (11 of 16; 69%) are small eruptions of VEI 1–2. AVO failed to forecast VEI 3 eruptions for Okmok, 2008, and Pavlof 2013 and 2014, despite the presence of seismic monitoring at both and geodetic monitoring in the case of Okmok. Okmok 2008, for example, had less than 2 h of clear precursory seismicity despite a dense proximal seismic network (Larsen et al., 2009), and a subtle precursory change in long-term inflation was only clear in geodetic data in retrospect (Freymueller and Kaufman, 2010; Lu and Dzurisin, 2010). Pavlof is one of the most frequently active volcanoes in Alaska, and has erupted seven times since the founding of AVO. And, although some intra-eruptive explosions have been successfully forecast (Power et al., 2018), AVO has not successfully forecast eruptive onset at Pavlof. This likely reflects the fact that Pavlof's persistently hot, open conduit allows magma slugs, which ascend rapidly with little contamination from crustal rocks (Mangan et al., 2009), to freely degas without pressurizing the surrounding crust, possibly coupled with the lack of a proximal seismic station. The three successful eruption forecasts for this group (monitored, shorter repose period) were Shishaldin 1999 and Veniaminof 2002 and 2013. Shishaldin 1999's precursory seismicity and AVO response is covered in Power and Cameron (2018).

For the frequently-erupting, basaltic and basaltic andesite, seismically unmonitored volcanoes (upper left quadrant of **Figure 8**), just one event was forecast, out of 18; Shishaldin 2014. AVO maintains a seismic network on Unimak Island that monitors Shishaldin, but prior to and during this eruption, the network was substantially impaired due to equipment failures. AVO raised the color code to Yellow on January 30, on the basis of increased surface temperatures seen in satellite data and increased steam emissions observed in webcam images (Cameron et al., 2017). Eruptions in the short-repose period and unmonitored group that AVO failed to forecast include eruptions of Akutan, Amukta, Gareloi, Pavlof, Seguam, Shishaldin, Veniaminof, Okmok, Korovin, Westdahl, Cleveland, and Makushin. Notably, almost all of the "Missed" events occur in this quadrant of **Figure 8**, but there have been no "Missed" events in the past 10 years, suggesting that technologies like infrasound and increased communications, including interpersonal, interagency, and satellite coverage have improved volcano notification even in the absence of ground-based instrumentation.

#### Unrest Without Eruption (UwE)—How Often, and Why?

This dataset contains 23 UwE events. Unrest which has prompted formal notification by AVO falls into three broad categories. The first is characterized by dominant volcano-tectonic seismicity that may or may not be accompanied by other seismicity such as low-frequency events or tremor, as well as increased degassing, heating of the edifice or increased measured gas flux. Some of these events have been described as "failed eruptions," meaning magma intruded into but stalled within the shallow crust (Moran et al., 2011). Some examples include Akutan 1996 (Lu et al., 2000), Iliamna 1996 (Roman et al., 2004), Iliamna 2012 (Prejean et al., 2012), Martin 2006 (O'Brien et al., 2012), Little Sitkin 2012 (Haney et al., 2014), Semisopochnoi 2014 (Cameron et al., 2017), Tanaga 2005 (Lu and Dzurisin, 2014), and Spurr 2004–2006 (Coombs et al., 2006). Other UwE events have similar characteristics but the processes that led to the unrest are more equivocal. In general, we describe this category of event as "possible intrusion and/or activation of the hydrothermal system." The second category of UwE includes 4 incidents that occurred at mafic, open-system volcanoes and were characterized by increased low-frequency seismicity, infrasound signals, and/or thermal output. We characterize these as "intense" degassing episodes, although they may reflect intrusion and "failed eruption" as well. As a final category, four additional UwE events have occurred within 1 year of significant (VEI 3+) eruptions, such as the unrest at Augustine 2007 and Okmok 2009. These may result from additional time periods of intrusion not immediately associated with eruption, or they may reflect adjustment of the crust after evacuation of magma.

An obvious question that arises is: for monitored volcanoes, how often does unrest lead to eruption? To answer this question, we look at events at monitored volcanoes that are classified as either UwE, or Good. A Good classification implies that there was enough precursory activity to warrant issuing an alert. We ignore eruptions that did not exhibit detected precursors. When looking at the 24 events that fit these criteria (i.e., UwE or Good, and seismically monitored with eruption precursors), seven (29%) resulted in eruption, and 17 (71%) did not. Looking only at andesitic volcanoes, one in four unrest sequences resulted in eruption. These are useful numbers to keep in mind during future episodes of unrest, and assigning probabilities such as during an event tree development.

#### Advancements in Monitoring and Implications for Next Generation of Volcano Monitoring Systems

AVO's recent advancements in multi-disciplinary volcano monitoring yield improved eruption forecasting capabilities, even considering that many Alaska volcanoes have frequent, small eruptions without much, if any, precursory seismicity. "Missed" events—highly undesirable for a volcano observatory are not known to occur in AVO's record since 2002, largely due to improved monitoring and observations, including: seismic; geodetic; satellite; infrasound; local observers communicating with AVO via the Internet and telephone; and AVO's interagency coordination with the FAA to receive pilot reports. The last ten years (2008–present) have seen four successful forecasts, while the previous two decades held only five, suggesting that the slow and continual expansion from three monitored volcanoes (1989) to 32 (2017) has substantially increased AVO's ability to forecast eruptions.

When the USGS first introduced the concept of a National Volcano Early Warning System, or NVEWS, a threat assessment (Ewert, 2007) ranked all U.S. volcanoes into four threat categories: very high, high, moderate, and low, based on objective hazards and exposure to population and infrastructure. At the same time, Moran et al. (2008) made instrumentation recommendations at four levels that were directly linked to the volcanicthreat rankings of Ewert et al. (2005) and Ewert (2007). Thus, low-threat volcanoes should have basic monitoring capabilities, and higher threat volcanoes should have subsequently denser and more sophisticated monitoring networks.

We suggest that in addition to considering volcanic-threat levels, instrumentation strategies also take into account more granular details about volcanic systems in question. In particular, "open-system" volcanoes require either denser seismic data, including near-vent or borehole stations, to be forecast; or additional data, including near-summit tilt, gas, and reliable web camera data. Recent advances in gas monitoring are particularly exciting for open systems (e.g., de Moor et al., 2016). Denser multi-disciplinary networks are highly desirable for frequentlyactive, low-viscosity systems, and may enable improved forecasts and understanding of the underlying volcanic processes that drive eruptions. In contrast, "closed-system" volcanoes can often, but not always, be successfully forecast with more traditional monitoring networks dominated by seismic instrumentation. Alaska has ∼100 volcanoes active in the Holocene, although only 30 of them have erupted in historical time, and AVO currently seismically monitors a slightly different set of 32. This leaves about 70 unmonitored volcanoes that, if they were to erupt, would likely have substantial precursory seismicity. The best chance to forecast these potentially large eruptions would be to instrument these long-repose time volcanoes. This is similar to the recommendation of the National Academy's recent consensus study report recommending working toward "sparse ground-based monitoring of all potentially active volcanoes. . . and that monitoring strategies need to be tailored to the type of volcano in question" (ERUPT Report, National Academies of Sciences Engineering Medicine, 2017). In a region such as Alaska, with a large number of volcanoes over a vast swath of remote wilderness, added insights into the nature of the volcanoes, derived from dense multi-parametric monitoring networks and geologic study, can further assist in prioritization of instrumentation.

### CONCLUSIONS

Our analysis demonstrates that it is important that measures of success or failure at eruption forecasting take into account an observatory's rationale and procedure for use of alert levels and alert notifications, as well as monitoring and individual volcano and eruption characteristics. AVO has had the greatest success forecasting larger eruptions (VEI 3+) at seismically monitored volcanoes with longer repose times and relatively more silicic magma compositions. Because these larger eruptions have the greatest impact on aviation, they are also the most critical for successful monitoring and forecasting. Eruptions at volcanoes with short repose times, typically with basaltic or basaltic andesite composition, are generally poorly forecast, due to a lack of detectable precursory seismic activity. Seismic monitoring at these volcanoes significantly shortens the eruption detection time, however. Denser multi-parametric networks for these fluid and frequently-erupting volcanoes could improve AVO's ability to forecast an eruption. Successful forecasting could also be aided by the continued development of volcano-specific alarm algorithms to detect very subtle changes in seismicity and infrasound. Fortunately, AVO is well-calibrated at interpreting co-eruptive seismicity at these volcanoes (Haney et al., 2009), so we can detect and make accurate assessments of eruptions relatively quickly. Non-seismically monitored volcanoes pose significant forecasting challenges, but improved communication, satellite coverage, lightning data, and infrasound have reduced eruption detection time and eliminated "Missed" eruptions. Greater multi-disciplinary instrumentation of all volcanoes in Alaska would reduce the chance of unforecast volcanic eruptions.

#### DATA AVAILABILITY

All datasets analyzed for this study are included in the manuscript, supplementary files, and https://www.avo.alaska. edu/activity/avoreport.php.

#### AUTHOR CONTRIBUTIONS

CC, SP, MC, and DR designed and performed the study, and compiled data for the analysis. JP and KW contributed to the analysis of the results. All authors provided critical feedback and helped write the manuscript.

#### FUNDING

Funding for this study is from the USGS Volcano Hazards Program. This project was partially funded through State of Alaska Division of Geological and Geophysical Surveys cooperative agreements with the USGS Volcano Hazards Program, grant numbers G16AC00054 and G16AC00165.

#### REFERENCES


#### ACKNOWLEDGMENTS

This work began as abstracts presented at Cities on Volcanoes 9 by Prejean et al. (2016) and Cameron et al. CC expresses gratitude for conversations with Tina Neal and Marianne Guffanti, who helped encourage this work, to Jeremy Pesicek for assistance in determining network health and to Sarah Ogburn for her eagle eye in checking eruption chronologies. Many thanks also to John Pallister and Janet Schaefer, whose thorough reviews strengthened this manuscript, and to the two reviewers assigned by Frontiers, who each helped us present these thoughts with improved clarity.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart. 2018.00086/full#supplementary-material

Supplementary Data Sheet 1 | Table of events and notifications issued by the Alaska Volcano Observatory, 1989–October 2017. All dates are local date/time. Column descriptions are as follows: Volcano: volcano where the event took place. Event Start: time the eruption or unrest began. 00 is used for unknown days, or months. All events have known years. Notification Class: classes of AVO's notification about the event, see text for full descriptions. Composition: composition of erupted material, or, in the case of eruptions without analyzed products or unrest events, the composition of the most recently erupted products. Further information on compositions is provided in Supplementary Material B. Repose Time: the length of time since the start of the most recent eruption, rounded to a tenth of a year. In cases where the date or the month is not known, the 15th of the month and June are used. Seismic Monitoring: was there a local, seismic network, functioning at the time of the event, functioning for sufficient time before the event to characterize background seismicity? Detect Only, notification lag: for those events that were "Detect Only," what was the time lag between event onset and AVO's formal, written notification? Time given in days, rounded to 3 h for lags of less than a day, and given in integers for lags >1 day. Notice Date: date of AVO's official notification, where applicable. Forecast Lead Time: for eruptions where AVO's classification was "Good", what was the amount of time between forecast and eruption? Time given in days, rounded to 3 h for differences of less than 1 day, and integers for time differences >1 day. First color code change, monitored volcanoes only: for those volcanoes monitored by AVO at the time of the event, this column shows the initial color code change for the event, with U, Unassigned; G, Green; Y, Yellow; and R, Red. Event Description: short description of the event. Main Citation: primary reference for more information on the event.

Supplementary Data Sheet 2 | Listing of volcanoes and compositional data, including data source information.


during the 2006 eruption of Augustine Volcano," in The 2006 eruption of Augustine Volcano, Alaska, ed. J. A. Power, M. L. Coombs, and J. T. Freymueller (Reston, VA: U.S. Geological Survey Professional Paper 1769), 427–452.


**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 Cameron, Prejean, Coombs, Wallace, Power and Roman. 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.

## Absence of Detectable Precursory Deformation and Velocity Variation Before the Large Dome Collapse of July 2015 at Volcán de Colima, Mexico

#### Philippe Lesage<sup>1</sup> \*, Alexandre Carrara<sup>1</sup> , Virginie Pinel <sup>1</sup> and Raul Arámbula-Mendoza<sup>2</sup>

<sup>1</sup> Université Savoie Mont Blanc, Université Grenoble Alpes, CNRS, IRD, IFSTTAR, ISTerre, Grenoble, France, <sup>2</sup> Centro Universitario de Estudios e Investigaciones de Vulcanología, Universidad de Colima, Colima, Mexico

#### Edited by:

Nicolas Fournier, GNS Science, New Zealand

#### Reviewed by:

Alessandro Tibaldi, Università degli Studi di Milano Bicocca, Italy Mie Ichihara, The University of Tokyo, Japan

> \*Correspondence: Philippe Lesage lesage@univ-smb.fr

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 29 January 2018 Accepted: 19 June 2018 Published: 09 July 2018

#### Citation:

Lesage P, Carrara A, Pinel V and Arámbula-Mendoza R (2018) Absence of Detectable Precursory Deformation and Velocity Variation Before the Large Dome Collapse of July 2015 at Volcán de Colima, Mexico. Front. Earth Sci. 6:93. doi: 10.3389/feart.2018.00093 Improving the ability to foresee volcanic eruption is one of the main objectives of volcanologists. For this purpose, it is essential to better detect eruption forerunners and to understand their relationship with eruptive processes. The evaluation of the performance of the forecasting methods partly relies on the estimation of the frequency of occurrence of the various precursory phenomena. Possible lack of precursor before some events must also be carefully documented and analyzed. In this study, we check for the existence of detectable precursors before the large dome collapse event of Volcán de Colima, which occurred in July 2015, leading to the emplacement of more than 10 km long Pyroclastics Density Currents and the opening of a large breach in the crater. Based on volumes of emitted magma, the 2015 eruption is the largest event recorded at Volcán de Colima since the 1913 Plinian eruption. Surface displacements in the summit cone area are quantified over the period November 2014-June 2015 based on Synthetic Aperture Radar (SAR) images acquired by Sentinel-1 satellite. Velocity variations are investigated by coda wave interferometry. Daily cross-correlation functions of seismic noise recorded at 5 broadband stations are calculated for the period January 2013–April 2017 and apparent velocity variations are obtained by applying the stretching method. We show that no significant surface deformation can be measured by the SAR images over an area reaching 5 km from the summit, such that the volume of emitted magma cannot have been accommodated elastically in the 6 months preceding the eruption at a depth shallower than 5 km. The time series of apparent velocity variations display fluctuations of the order of 0.05% with characteristic time shorter than 1 month. Sharp velocity decreases of up to 0.2% are associated with strong regional tectonic earthquakes. However, no velocity change with amplitude larger than the noise level is observed before the July 2015 eruption. The behavior of the surface deformation and the velocity variation is consistent with the relative quiescence of the volcano-tectonic and

**106**

low-frequency seismic activities observed before this large eruptive event. This situation could be frequent in case of so called open systems, where additional magma input is directly transferred to the surface, producing dome modification, without significant pressurization of the plumbing system.

Keywords: eruption precursor, dome collapse, deformation, InSAR, seismic velocity variation, coda wave interferometry, Volcán de Colima, eruption forecasting

#### INTRODUCTION

Volcanic eruptions result from complex processes that include feeding of magma storage zone, magmatic intrusion, interaction with surrounding rock and hydrothermal system, or changes in the physical state, chemical and mineralogical content of reservoirs and conduits. Many of these processes produce phenomena that can be observed at the free surface of the edifice before an eruption. They are thus considered as precursors and volcanologists use them for forecasting volcanic eruptions. Because these phenomena are the basis of volcano monitoring, many volcanological research aims at detecting and interpreting them. The most widely used methods are the study of the seismic activity, the measurement of ground deformations and the analysis of gas flow and composition (Scarpa and Gasparini, 1996). Other approaches, such as magnetic and electric studies can complement the classical ones. For example perturbations of magnetic field and variations of self-potential anomalies have been observed prior to a few eruptions (Johnston, 1997; Zlotnicki et al., 2009).

Several types of seismic precursory phenomena can be observed (McNutt, 1996). They include the increase of the level and energy of seismic signals, the rise of the number of events, the migration of the seismogenic zones, variations of the focal mechanisms of volcano-tectonic (VT) events, changes in anisotropy of the seismic velocities, emergence of swarms of various types of event. By analyzing earthquakes with similar waveforms (or multiplets – Poupinet et al., 1984) or ambient noise correlation functions, variations of seismic velocity in the medium have been detected before some eruptions. In a pioneering work, Ratdomopurbo and Poupinet (1995) used seismic multiplets and a technique known as Coda Wave Interferometry (Grêt et al., 2005; Snieder, 2006), to detect a velocity increase of 1.2% several months before the 1992 Merapi eruption. At Piton de la Fournaise, La Reunion Island, Brenguier et al. (2008) demonstrated, using noise correlation functions, that the velocity decreased by about 0.05% a few weeks before several eruptions in 1999 and 2000. Since then, many studies of velocity variations at Piton de la Fournaise have been published, extending the period of analysis (Duputel et al., 2009; Clarke et al., 2013; Rivet et al., 2014), improving the technique (Clarke et al., 2013; De Plaen et al., 2016), and locating the source of perturbation (Obermann et al., 2013). Velocity variations have also been detected before some eruptions of stratovolcanoes, such as Ruapehu, New Zealand, (Mordret et al., 2010), Mt Asama, Japan, (Nagaoka et al., 2010), Miyakejima, Japan, (Anggono et al., 2012), Etna, Italy, (Cannata, 2012), Mt St Helens, USA, (Hotovec-Ellis et al., 2015) or Merapi, Indonesia, (Budi-Santoso and Lesage, 2016), and shield volcanoes, such as Kilauea, Hawaii (Donaldson et al., 2017). Sharp velocity decreases induced by large tectonic earthquakes have also been observed at some volcanoes (Nishimura et al., 2000, 2005; Battaglia et al., 2012; Brenguier et al., 2014; Lesage et al., 2014).

Deformation data have also proven useful to reveal magma accumulation inside crustal storage zones or emplacement at shallow depth before an eruptive event (Dvorak and Dzurisin, 1997; Dzurisin, 2003, 2007), such providing both long (months to years) and short (hours to days) term precursors of volcanic activity. Deformation before eruptions are observed both on basaltic volcanoes (Sturkell et al., 2006) and andesitic and rhyolitic systems (Swanson et al., 1983). In recent years, the number of volcanoes where deformation data are available have drastically increased thanks to satellite radar interferometry, which provides high spatial resolution surface displacement with a precision of a few millimeters to a few centimeters depending on the available dataset (Biggs et al., 2014; Pinel et al., 2014; Biggs and Pritchard, 2017). This large and worldwide amount of available data has enabled to statistically question the link between deformation and eruptive activity.

Because the quest for eruption forerunners is of upmost importance, many scientific papers in this field present observations of precursory phenomena, investigate their interpretation and relationship with magmatic and hydrothermal processes, and discuss their use for forecasting. However, volcanic eruptions are not always preceded by precursors. There are cases where no precursory, or even co-eruptive, deformation was observed (e.g., Moran et al., 2006; Chaussard et al., 2013; Ebmeier et al., 2013; Biggs et al., 2014) or, more generally, where the pre-eruptive phenomena were too small to be detected and interpreted correctly for predictions. Such cases have been described at Popocatepetl, Mexico (Quezada-Reyes et al., 2013); Ontake, Japan (Kato et al., 2015), Soufrière Hills, Montserrat (Calder et al., 2002), and many other places. Most of the eruptive crises characterized by lack of precursors are probably not reported in scientific journals, although this information may be partly found in bulletins of volcanological observatories. This prevents from evaluating on a large scale the proportion of eruptions that are preceded, or not, by precursory phenomena.

The forecast of volcanic eruptions can follow two complementary approaches. The probabilistic one aims at estimating the probability that an eruption occurs in a given time interval either at short or at long term (Marzocchi and Bebbington, 2012). The deterministic approach tries to estimate the time of occurrence of an eruption. The most widely used method for the later approach is the material Failure Forecast Method (FFM–Voight, 1988) which gave encouraging results in a few cases (e.g., Cornelius and Voight, 1995; Kilburn and Voight, 1998; De la Cruz-Reyna and Reyes-Dávila, 2001). Nevertheless, eruption forecasting is partly an empirical task based on the knowledge of the previous volcanic activity, on the observations produced by the monitoring system, and on the experience of the volcanologists in charge. Forecasting methods still require to be improved in order the predictions to be more reliable and precise. Their performance and success rate need to be evaluated in a large variety of cases, including different types of volcanoes and various kinds and amplitudes of eruptions. In this evaluation, it is important to take into account the eruptions that were not preceded by precursors, as their occurrence can produce hazardous situations for inhabitants and visitors. Thus, published papers should also document volcanic crises characterized by lack of precursors as well as observations of forerunners that are difficult to interpret and to use for forecast. The difficulty in estimating the performance of forecasting methods is also illustrated by the fact that no more than 20% of the eruptions were anticipated by alert level changes (Winson et al., 2014). Furthermore, there is a tendency in volcano observatories to underreport intrusive episodes not leading to eruptions (Moran et al., 2011), even if some recent databases are trying to include these cases (Ebmeier et al., 2018).

In the present paper, we look for precursors of the large dome collapse of Volcán de Colima in 2015 by carrying out two approaches. First we apply techniques of seismic ambient noise correlation to estimate velocity variations in the structure by coda wave interferometry. Then we track potential deformation of the summit part of the volcanic edifice using SAR images. We show that there were no significant precursory signals in deformation and velocity variation before this major eruption, we interpret these observations from a volcanological point of view and we discuss the significance of the lack of precursor in term of hazard management.

#### VOLCANOLOGICAL SETTING

Volcán de Colima is located in western Mexico (19.51◦N, 130.62◦ W, 3,860 m asl) and is currently one of the most active volcanoes in North America. It has produced at least three sub-Plinian to Plinian eruptions in 1576, 1818, and 1913, and an average of one large magnitude eruption per century (Robin et al., 1987; Luhr and Carmichael, 1990). The 1913 eruption included an opening phase producing pyroclastic flows and surges, a vent clearing phase which destroyed the summit dome, and a Plinian phase with a 23 km high column, the collapse of which generated a surge and 15 km long pyroclastic density currents (PDC). A Volcanic Explosivity Index (VEI) of 5 was calculated by Simkin and Siebert (1994) and Saucedo et al. (2010). The later authors estimated that a similar Plinian eruption would threaten more than 300, 000 inhabitants nowadays. Recent periods of moderate activity occurred in 1991, 1994, 1998–1999, 2001–2003, and 2004–2005 with alternation of dome buildings and destructions, Vulcanian eruptions, lava flows, and PDCs. The hazard associated with this activity and the observation of precursory phenomena triggered several evacuations of inhabitants of the most exposed localities (Macias, 1999).

The seismic activity of Volcán de Colima is monitored since the installation of a permanent network in 1989. Reyes-Davila and De la Cruz-Reyna (2002) analyzed the behavior of the Real-time Seismic Energy Measurement (RSEM) before several moderate eruptions during a period of about 10 years. They observed that, in most cases, no clear pattern of increasing seismicity can be detected. Before some events, the seismic activity presented clear acceleration only a few hours prior the eruption onset. The precursory activity of two eruptions in 1994 and 1998 was characterized by a clear acceleration of the energy release during several days. The authors applied a version of FFM to forecast the date of the eruption either in hindsight (1994 event) or in foresight (1998 event) (De la Cruz-Reyna and Reyes-Dávila, 2001). Boué et al. (2015, 2016) proposed a Bayesian approach of FFM which uses an automatic recognition system (Benítez et al., 2007; Cortés et al., 2014) to classify and separate different types of seismic event. They processed 13 years of continuous recording of Volcán de Colima. Among 36 mild explosions, 24 were preceded by accelerations of LP events rate, 7 by a linear increase, and 3 occurred without seismic precursors. Successful forecasts were obtained for one third of the cases, while this proportion reaches 83% when some reliability criteria are fulfilled. Lesage et al. (2014) calculated and analyzed the velocity variations for the whole period 1998–2013 using data from a pair of short-period stations. They did not find any clear relationship between velocity changes and the mild to large vulcanian eruptions that occurred during this interval. They only noted that most large eruptions coincided with periods of decreasing velocity.

Limited in situ measurements of ground deformation have been performed at Volcán de Colima volcano based on tiltmeters, precise leveling, EDM and GPS campaigns (Murray and Ramírez-Ruiz, 2002; Murray and Wooller, 2002; Zobin et al., 2013). In particular, an acceleration of the summit inflation measured by EDM before the 1998 eruption was interpreted based of the FFM to predict a posteriori the time of the eruption (Murray and Ramírez-Ruiz, 2002). The application of InSAR technique to retrieve and quantify the displacement field at Volcán de Colima is challenging due to the limited coherence on the volcano slope as well as tropospheric artifacts (Pinel et al., 2011). SAR studies have thus been restricted to the 5 kilometer wide summit coherent area. The time series analysis of ASAR-ENVISAT data recorded from mid-2002 to the end of 2006 evidenced a summit subsidence reaching a rate of around 1 cm/yr and centered on the summit but enhanced on recent lava flows, which was interpreted as due to eruptive deposits load effects associated to a shallow deflating source (Pinel et al., 2011). Using high spatial and temporal SAR data acquired by TerraSAR-X satellite, Salzer et al. (2014) were able to catch a localized preexplosive deformation induced by a transient pressurization of the shallow plumbing system before an explosion which occurred in January 2013.

After a period of quiescence since June 2011, a new phase of activity initiated in January 2013 when several moderate Vulcanian explosions destroyed the lava dome emplaced in 2007–2011. Then lava extrusion occurred with rate of 0.1– 0.2 m<sup>3</sup> s −1 , forming a new lava dome and producing rockfalls and explosions until February-March 2014. An increase of lava extrusion and rockfall activity was observed in September 2014 with extrusion rate of 1–2 m<sup>3</sup> s −1 . In January-February 2015, a series of explosions produced the destruction of the dome and the end of the rockfall activity. Another lava dome was observed on 20 May 2015 and continued to grow at low rate. On 3 July, a moderate explosion occurred and was followed by a decrease of the explosive activity and an acceleration of the rockfall activity. More information on the volcanic and seismic activity of Volcán de Colima in 2013–2015 can be found in Zobin et al. (2015), Capra et al. (2016), Reyes-Dávila et al. (2016), Macorps et al. (2018), and Arámbula-Mendoza et al. (2018). On 10 July 2015, after 2 days of increased extrusion rate, a partial collapse of the dome occurred accompanied by large PDCs that reached 9.1 km on the south flank. On 11 July, 16 hours after the first events, another series of PDCs of larger size reached distances of 10.3 km from the crater. No eruptive columns were produced during the whole sequence. The total volume of material including PDCs and ashfalls was estimated to 14.2 × 10<sup>6</sup> m<sup>3</sup> (Reyes-Dávila et al., 2016). A more recent study estimated a volume of block-and-ash flow material of 7.7 ± 1 × 10<sup>6</sup> m<sup>3</sup> , based on optical and field data (Macorps et al., 2018). This volume makes the July 2015 sequence the most important eruption since the Plinian event of 1913. In contrast with the common behavior of the volcano (Arámbula-Mendoza et al., 2011), the July 2015 eruption was preceded by a decrease of the rate of LP events and explosions. The main precursory phenomenon was a marked increase of the number and energy of rockfalls and PDCs that accompanied the rise of the extrusion rate (Reyes-Dávila et al., 2016; Arámbula-Mendoza et al., 2018). Although the deformations and the velocity variations associated with the previous eruptions were small or undetectable, we may expect to detect stronger forerunners in the case of the 2015 exceptional events.

#### APPARENT VELOCITY VARIATIONS

#### Network, Data, and Method

We used data from the Volcán de Colima monitoring network which is part of the State of Colima's Seismological network (RESCO). In a first stage, it included 4 stations equipped with vertical SS-1 Ranger short-period seismometers and analog transmission. It records continuous signals since 1998. This network was completed in 2001 and 2007–2008 by 6 Guralp CMG-6TD broadband stations (**Figure 1**).

We extracted Green functions between pairs of sites by crosscorrelating ambient seismic noise recorded at the corresponding two stations (Weaver and Lobkis, 2001; Campillo and Paul, 2003). We followed a standard procedure which consists of trend and mean removal, band-pass filtering in the range [0.125– 2] Hz, spectral whitening and one-bit amplitude normalization (Bensen et al., 2007; Lesage et al., 2014) and then calculating ambient noise cross-correlation functions (NCFs) for delays of ±150 s. For three Guralp stations (SOMA, WEST, INCA), we detected periods of several months where their clocks were not synchronized, with time lags of up to ±1.3 s. These periods are 12 March 2014 to 25 November 2014 for station SOMA, 8 October 2014 to 30 April 2016 for WEST, and 23 January 2015 to 30 June 2015 for INCA. They are indicated by dotted lines in **Figure 2**. These time drifts were corrected for stations SOMA and INCA by comparing the corresponding NCFs with those obtained when the clocks were well-synchronized (Stehly et al., 2007; Sens-Schönfelder, 2008). The time drift pattern for station WEST was more complex and corrections are not reliable enough. Thus, we did not use this station for the following analysis.

Daily NCFs were calculated for the vertical components of station pairs EFRE-SOMA, EFRE-INCA, MNGR-INCA, SOMA-MNGR, and SOMA-JUBA from January 2013 to April 2017. The paths between the stations belonging to these pairs go through the volcanic edifice (**Figure 1**). The corresponding NCFs should be thus affected by any perturbation in the volcano. Then we estimated the apparent velocity variation (AVV) by using the stretching method (Lobkis and Weaver, 2003; Sens-Schönfelder and Wegler, 2006). A NCF stacked over 2013 was used as a reference and compared to each daily NCF. The reference NCF is stretched or compressed in order to maximize the correlation coefficient between the coda of both NCFs in a delay range of [10, 80] s. Because the correlation functions obtained are asymmetrical, we used the side of the NCF which presented maximum amplitude. The corresponding stretching coefficient is equal to the negative of the apparent relative velocity variation (Lesage et al., 2014). Using this approach, we calculated time series of AVV for each station pair, as well as their average. In the Supplementary Material we compare this approach with two other methods and we show that similar results are obtained in all cases.

#### Results

**Figure 2** displays the apparent velocity variations obtained for the 5 pairs of broadband stations from 2013 to 2017, together with their average. All the AVVs present short term (<month) fluctuations with amplitudes of 0.05–0.1%. The averaging of the AVVs reduces to less than 0.05% the amplitude of these fluctuations that can be considered as the noise level. Several tectonic earthquake with magnitude larger than 6 and epicentral distance smaller than 500 km from Volcán de Colima (vertical green lines in **Figure 2**) induced sharp velocity decreases. For example, a velocity drop of about 0.2% in average occurred on April 18, 2014 during a M = 7.1 earthquake located at 350 km from the crater. This phenomenon was reported by Lesage et al. (2014) who demonstrated that the corresponding perturbation is localized in the shallow layers of the edifice.

No clear velocity variations with amplitude larger than the noise level appeared before, during nor after the July 2015 large eruptions. A small oscillation of approximately ±0.07%

can be observed during less than 2 months before the events, but it is poorly significant and its use as a precursory signal would not be reliable. A sequence of mild Vulcanian eruptions in January-February 2017 was neither associated with AVVs. In the Supplementary Material we present several arguments that support the reliability of the estimation of the velocity variations.

### SUMMIT DEFORMATION STUDIED BY InSAR

#### Data and Method

We used SAR images acquired in C-band by the European Satellite Sentinel-1A over Colima volcano since November 2014. The present study is based on 8 descending images of

Track 12 (subswath 1, VV polarization, look angle over the volcano summit of 34.1◦ ) recorded before the July 2015 eruption from November 27th 2014 until June 15th 2015 (**Figure 3**). Images were provided by the European Space Agency (ESA) as Single Look Complex (SLC) images and processed using the NSBAS chain (Doin et al., 2012) modified to integrate Sentinel-1 data acquired in TOPSAR mode as described in Grandin (2015). Topographic contribution was removed using the SRTM DEM at 30 m resolution. Tropospheric contributions were corrected using the ERA Interim global meteorological model provided by the European Center for Medium-Range Weather Forecast (ECMWF) as explained in Doin et al. (2009). Twenty-eight interferograms were calculated (see **Figure 3** for the network) and unwrapped using the ROI\_PAC branchcut unwrapping algorithm. Unwrapped interferograms were geocoded on the 30 m resolution DEM. The phase delays of unwrapped interferograms were then inverted pixel by pixel using a least square method to solve for the cumulative phase delay through time. Pixels characterized by a large RMS (above a threshold of 0.9 rad) were discarded such that our displacement detection threshold on the remaining pixels can be estimated around 0.4 cm over the 6 month period studied.

#### Results

The mean velocity of the ground surface in Line of Sight derived from the time series analysis is presented in **Figure 4**. The surface displacement can be retrieved from the phase of the radar signal only if the ground backscattering properties remain stable though time, insuring a good coherence. This condition is not fulfilled on the vegetated volcano slopes such that this SAR dataset only provides information in the summit area (at a distance smaller than 4 km from the volcano crater). Ash deposits resulting from explosions further limit the area with good coherence thus restricting the available information to the south-eastern part of the summit area at some distance from the dome. No information is available on the deformation of the dome itself. A subsidence signal (which appears in blue on **Figure 4**) is observed above and nearby the summit lava flows (Global Volcanism Program, 2015). Elsewhere no significant (superior to 6 ± 3 mm) Line of Sight displacement is evidenced over the period preceding the July 2015 eruption. Based on this observation, we can deduce that no vertical displacement larger than 1 cm can be seen over a distance of 4 km away from the summit crater. Considering a Mogi source (Mogi, 1958) localized beneath the crater, we can estimate a maximum value for the magma volume that can be stored elastically at depth below the crater without inducing detectable surface displacement. As shown in **Figure 5**, this value increases with the reservoir depth. We thus show that no significant volume of magma (above 3 million of m<sup>3</sup> ) can have been stored shallower that 5.5 km depth during the 6 months preceding the eruption.

#### DISCUSSION AND CONCLUSION

Variations of seismic velocities have been observed as a precursory phenomenon of volcanic eruptions relatively recently thanks to the development of the techniques of continuous recording, of seismic multiplet analysis and of ambient noise cross-correlation. It is thus important to evaluate the frequency of its occurrence before eruptions and to investigate its origins and relationships with other observations. Several physical processes have been proposed to explain the velocity variations in volcanoes. Changes in the ground water level, related with precipitation, modify the pore pressure and can induce detectable velocity changes. This process is the source of seasonal effects that can be corrected when sufficient observations are available (Sens-Schönfelder and Wegler, 2006; Hotovec-Ellis et al., 2015; Rivet et al., 2015). Strong topographical changes related to caldera formation at Miyakejima and Piton de la Fournaise volcanoes were also accompanied by velocity increases or decreases (Duputel et al., 2009; Anggono et al., 2012; Clarke et al., 2013). Sharp temporary decreases in velocity have been observed in relation with strong ground shaking due to the passing of the seismic waves generated by large tectonic earthquakes (Battaglia et al., 2012; Brenguier et al., 2014; Lesage et al., 2014; this study). This phenomenon has been associated with the presence of highly pressurized hydrothermal or magmatic fluids at depth (Brenguier et al., 2014) or with mechanical softening and nonlinear elastic behavior of granular material in the shallow layers of volcanoes induced by the ground shaking (Lesage et al., 2014). The velocity variations observed before numerous eruptions of Piton de la Fournaise volcano have been related to overpressure induced by magma intrusion through a model of dilatancy and empirical laws linking perturbations of porosity and volume with changes in shear velocity (Brenguier et al., 2008). During the last days before

the large 2010 eruption of Merapi volcano, rapid fluctuations of velocity have been interpreted as the consequence of the progressive fracturation and healing of the plug due to pulses of magma intrusion (Budi-Santoso and Lesage, 2016). However, in most studied cases, the velocity variations are interpreted as the result of the dependency of physical parameters of rocks to stress (Birch, 1960; Ratdomopurbo and Poupinet, 1995; Cannata, 2012; Hotovec-Ellis et al., 2015; Donaldson et al., 2017; Lamb et al., 2017). When an increasing effective pressure is applied to volcanic rocks, which are porous and pervasively microcracked, the most complient cracks and pores close yielding the elastic modulus and seismic velocities to increase (Vinciguerra et al., 2005; Stanchits et al., 2006; Nara et al., 2011; Heap et al., 2014). This process is involved in the increase of velocities with depth in volcanic structure (e.g., Lesage et al., 2018 for a review). When deviatoric stresses are applied, cracks normal to the axis of the maximum principal stress close, while those parallel to it remain almost unaffected (Nur, 1971). At high deviatoric stress level, a new population of cracks appears in the direction parallel to the maximum stress axis. This damaging process induces strong velocity decrease (Lamb et al., 2017). Thus there are complex relationships between changes of the stress field and the field of velocity variations.

This is supported by the coeval observations of ground deformations and velocity changes (Clarke et al., 2013; Rivet et al., 2014; Donaldson et al., 2017; Hirose et al., 2017). For example, Rivet et al. (2014) showed that the velocity decreases during longterm inflations of Piton de la Fournaise and increases during deflations of the edifice. Moreover a strong velocity decrease was observed at the time as a large movement of the East flank of this volcano before the major 2007 caldera collapse (Clarke et al., 2013). However, even in a simple elastic half-space, a source of increasing pressure produces both extensional strain in the region above the source and compressional strain in the surrounding volume (Pinel and Jaupart, 2003; Budi-Santoso et al., 2013; Got et al., 2013; Donaldson et al., 2017). Therefore, the relationship between pressure evolution in the magmatic system, ground deformations, and velocity variations may be relatively complex.

On the other hand, the amplitude of the velocity changes before eruptions are very small. In many studies, the published values, generally obtained by averaging the values estimated using many station pairs, are of the order of a few tenths of percent

(Brenguier et al., 2008, 2016; Mordret et al., 2010; Donaldson et al., 2017). Only in a few cases, the relative velocity variation calculated using single station pairs or using seismic multiplets reaches about one percent (Ratdomopurbo and Poupinet, 1995; Nagaoka et al., 2010; Anggono et al., 2012; Hotovec-Ellis et al., 2015; Budi-Santoso and Lesage, 2016). Thus, the ability in detecting such small variations depends on the ratio of the amplitude of the signal of interest and that of the spurious fluctuations due to non-volcanic phenomena and to the nonstationarity of the sources of seismic noise (Stehly et al., 2007). It depends also on the relationship between the duration of the windows used to calculate the correlation functions of seismic noise and the characteristic time of the processes producing velocity variations.

At Volcán de Colima, although the 2015 dome collapse was the most important event in term of volume of emplaced material since the Plinian eruption of 1913, no clear signals of deformation and velocity variation could be detected before and during the eruptions, besides a good level of reliability. The main precursory phenomenon was the increase of rockfall activity that was interpreted as a consequence of accelerated extrusion rate and that was probably accompanied by strong dome modifications. Unfortunately, cloudy conditions due to rainy season prevented visual and photographic observations from quantifying the dome evolution.

In the study period 2013–2017, the amplitude of the common velocity fluctuations is about 0.05% for characteristic times less than a few months. Some sharp decreases of velocity followed by slow recovery are related with tectonic earthquakes, especially that of April 18, 2014. Other variations, such as that occurred on mid-September 2013, are not related to any known phenomenon. In May-June 2015, a sequence of increase-decrease-increase of the velocity coincides with the extrusion of a dome that was first observed on May 20 (Arámbula-Mendoza et al., 2018). However, the velocity recovered the value corresponding to its general trend by the first days of July. It is thus not possible to consider this May-June sequence as a direct precursor of the July dome collapse. The Green functions extracted by noise correlation are predominantly surface waves (Shapiro and Campillo, 2004). In the frequency range used to estimate velocity variations (0.125–2 Hz), the corresponding sensitivity kernels indicate that they are sensitive to velocity perturbations of the medium up to 2–3 km below the surface, i.e., at depth were possible sources of deformation are expected (Salzer et al., 2014).

Emplacement of magma inside the crust is generally expected to induce surface displacements (Dzurisin, 2007; Tibaldi, 2015). A volume of the order of the one emitted during the July 2015 event (14 million of m<sup>3</sup> ) would require to have been emplaced either at some lateral distance from the crater (more than 5 km laterally away) or below the crater at more than 8.5 km depth in order to remain undetected in our dataset. If we instead consider the volume estimation for Block and Ash Flow deposits from Macorps et al. (2018), which is expected to be closer to the DRE volume, the minimum depth would rather be 7.5 km. This threshold depth is derived neglecting the magma compressibility, which may be large for bubble-rich magma and thus may reduce the surface displacement produced by magma emplacement (e.g., Johnson et al., 2000; Rivalta and Segall, 2008). Accounting for the effect of magma compressibility could increase the amount of magma possibly emplaced at shallow depth without significant surface displacement, such that the potential storage zone could be shallower but even so it would have to remain below 5.5 km depth in order to be consistent with the SAR dataset. Many eruptions occur without any detected surface deformation neither in the pre- or co-eruptive phase. For instance, Ebmeier et al. (2013) clearly evidenced a statistically significant lack of deformation for active volcanoes of the Central American Arc. The reason evoked for this absence of detectable surface deformation are eruptions fed directly by rapid magma ascent from deep magma storage zones, a diffuse and extended shallow storage system made of several vertically elongated cracks or large magma compressibility due to high volatile content. Chaussard et al. (2013) also described an absence of deformation during the eruptive activity at several volcanoes among which Volcán de Colima. They explained this behavior by an open system where the presence of a permanent conduit allows magma to rise toward the surface without pressurizing the reservoir.

The seismic activity of Volcán de Colima in 2015 is mainly composed of LP events, small high-frequency events only detected by the closest station to the summit, volcanic tremor, small explosions and numerous rockfalls (Reyes-Dávila et al., 2016; Arámbula-Mendoza et al., 2018). Almost no VT events are detected in this edifice. In the days preceding the dome collapse, a clear decrease of the number of explosions and LP events was observed as well as an accelerated rate of rockfalls generated by instabilities of the front of several lava flows. Thus, the main geophysical observations, including the seismic activity, the deformations and the velocity variations, are all consistent with an open magmatic system in which no marked pressurization occurred at shallow level before the July 2015 eruptions.

At Soufriere Hills volcano, Montserrat, in 1996–1998, most large dome collapses occurred during periods of high extrusion rate, intense hybrid seismic events activity and cyclic deformations (Calder et al., 2002). However some of the large collapses occurred while no magma was extruding and they were not preceded by seismic activity. These events were interpreted as structural failures of steep crater walls (Calder et al., 2002). The two dome collapses of July 2015 at Volcán de Colima appear to be intermediate cases, as they occurred during an episode of high rate of extrusion and were preceded by a decrease of seismic activity. These dome collapses probably resulted from a mechanical instability of the crater walls triggered by the strong magma flow through an open conduit. This type of eruptive event, which is not preceded by usually observed precursors, is thus difficult to forecast with classical monitoring methods. However, it would be important to identify this kind of situation in the future in order to manage better the corresponding hazards. The integration of new observations and analysis methods to the monitoring system may also help detecting forthcoming eruptions. For example, pixel offsets tracking methods applied to optical images acquired at small distance from the dome might bring useful information regarding the dome growth rate (Salzer et al., 2016) while high resolution SAR images can provide information on the dome deformation in quiescent periods (Salzer et al., 2017). Indeed, a hypothetical observation of both summit deformation and velocity variations, with amplitude larger than the usual fluctuations (i.e., >0.2%) and not related with strong tectonic earthquakes, could indicate

#### REFERENCES


a possible impending eruption and should be taken into account by the warning system of Volcán de Colima.

#### AUTHOR CONTRIBUTIONS

PL coordinated this study, carried out the calculation of the seismic noise correlation functions and of the velocity variations and wrote the main parts of the manuscript. AC processed the Sentinel-1 data under the supervision of VP. Together they interpreted Insar results and wrote the sections about deformations. RA-M provided the seismic data used in this study and information on the eruptive activity of the volcano. All the authors revised the whole text and figures.

#### ACKNOWLEDGMENTS

We thank the European Space Agency for providing Sentinel-1 data. This study was supported by CNES through the TOSCA project MerapiSAR. We are grateful to Gabriel Reyes-Dávila, Carlos Ariel Ramírez, Alejandro Velasquez Martinez, and Miguel González Amezcua, from RESCO, for their support in keeping the network working. The monitoring network of the Volcán de Colima is partially maintained by the CONACYT project Atención a Problemas Nacionales 2015-916. We thank the reviewers and editors for their suggestions that helped improving the manuscript.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart. 2018.00093/full#supplementary-material

broad-band surface wave dispersion measurements. Geophys. J. Int. 169, 1239–1260. doi: 10.1111/j.1365-246X.2007.03374.x


Merapi (Indonesia). Geophys. Res. Lett. 22, 775–778. doi: 10.1029/95GL 00302


and Miyake-jima (Japan) volcanoes. Phys. Chem. Earth 34, 394–408. doi: 10.1016/j.pce.2008.09.012


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

Copyright © 2018 Lesage, Carrara, Pinel and Arámbula-Mendoza. 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.

## Prevalence of Seismic Rate Anomalies Preceding Volcanic Eruptions in Alaska

Jeremy D. Pesicek <sup>1</sup> \*, John J. Wellik II <sup>1</sup> , Stephanie G. Prejean<sup>2</sup> and Sarah E. Ogburn<sup>1</sup>

<sup>1</sup> Volcano Disaster Assistance Program, Volcano Science Center, U.S. Geological Survey, Vancouver, WA, United States, <sup>2</sup> Volcano Disaster Assistance Program, Volcano Science Center, U.S. Geological Survey, Anchorage, AK, United States

#### Edited by:

Nicolas Fournier, GNS Science, New Zealand

#### Reviewed by: Lauriane Chardot,

Earth Observatory of Singapore, Singapore Laura Sandri, Istituto Nazionale di Geofisica e Vulcanologia (INGV), Italy Valerio Acocella, Università degli Studi Roma Tre, Italy

> \*Correspondence: Jeremy D. Pesicek jpesicek@usgs.gov

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 05 March 2018 Accepted: 02 July 2018 Published: 20 July 2018

#### Citation:

Pesicek JD, Wellik JJ II, Prejean SG and Ogburn SE (2018) Prevalence of Seismic Rate Anomalies Preceding Volcanic Eruptions in Alaska. Front. Earth Sci. 6:100. doi: 10.3389/feart.2018.00100 Seismic rate increases often precede eruptions at volcanoes worldwide. However, many eruptions occur without such precursors. Additionally, identifying seismic rate increases near volcanoes with high levels of background seismicity is non-trivial and many periods of elevated seismicity occur without ensuing eruptions, limiting their usefulness for forecasting in some cases. Although these issues are commonly known, efforts to quantify them are limited. In this study, we consistently apply a common statistical tool, the β-statistic, to seismically monitored eruptions in Alaska of various styles to determine the overall prevalence of seismic rate anomalies immediately preceding eruptions. We find that 6 out of 20 (30%) eruptions have statistically significant precursory seismic rate increases. Of these 6 eruptions, 3 of them occur at volcanoes with relatively felsic compositions, repose periods >15 years, and VEI ≥ 3. Overall, our results confirm that seismic rate increases are common prior to larger eruptions at long dormant, "closed-system" volcanoes, but uncommon preceding smaller eruptions at more frequently active, "open-system" volcanoes with more mafic magmas. We also explore the rate of other anomalies not precursory to eruptions and investigate their origins. Some of these non-eruptive anomalies can be explained by aftershocks of regional seismic events, magmatic activity that did not lead to eruption, or unrest at other nearby volcanoes. Some open-system volcanoes have high non-eruptive anomaly rates and low pre-eruptive anomaly rates and are thus not amenable to forecasting based on earthquake catalogs. In this study, we find that 31% of anomalies lead to eruption. With continued calibration at more volcanoes, the β-statistic that we apply may be used more broadly to analyze future periods of seismic unrest at other volcanoes, properly placing such episodes into the context of the long-term background rate. These results may be useful for informing future eruption forecasts around the world, and the statistical tool may aid volcano observatories in identifying future seismic rate anomalies under changing network conditions.

Keywords: volcano-tectonic, earthquake swarm, Alaska, eruption forecasting, seismicity, eruption, volcano monitoring

### INTRODUCTION

Earthquake swarms are common occurrences at volcanoes worldwide and are often associated with periods of increased volcanic unrest. The widespread prevalence of swarm activity preceding eruptions has led to their common use in eruption forecasting (e.g., Minakami, 1961; Shimozuru, 1971; McNutt, 1996; Chastin and Main, 2003; Kilburn, 2003; Boué et al., 2015), and volcano-tectonic (VT) swarms have been recognized as the earliest precursor to eruptions at long-dormant volcanoes in many cases (White and McCausland, 2016). However, not all eruptions are preceded by precursory seismicity and others have precursors that are too brief or subtle for effective warning (Larsen et al., 2009; Waythomas et al., 2014; van Eaton et al., 2016; Cameron et al., 2018). Furthermore, although earthquakes often precede eruptions, swarms also commonly occur due to stalled intrusions, high pressure fluid migration, or other non-eruptive activity (Benoit and McNutt, 1996; Farrell et al., 2009; Moran et al., 2011; Shelly et al., 2015). Therefore, some degree of VT swarm seismicity may be considered normal at many volcanoes. Thus, the onset of earthquake swarms near a volcano does not necessarily herald an oncoming eruption, nor does the lack of seismicity preclude an eruption from occurring. The presence or absence of earthquake swarms at a volcano is but one factor to be considered when forecasting eruptions, and the seismic signature of intrusion can vary widely by volcano.

Further complicating the use of VT swarms for forecasting is their similarity to tectonic seismicity unrelated to volcanism. Both types of earthquakes result from the same fundamental source process—brittle fracture of the crust—and VTs are often only differentiated from tectonic seismicity based on the proximity of an event to a volcano. However, location of an event alone is a poor measure of causal mechanism, as tectonically active faults are common in volcanic arcs (e.g., Ruppert et al., 2012), and VT swarms have been convincingly associated with magmatic activity as far as 45 km away from eruption sites (White and McCausland, 2016). These issues make truly magmatic VT seismicity difficult to distinguish from crustal tectonic seismicity unrelated to volcanism. This is especially true for events far from the volcanic center or in sparsely monitored yet tectonically active areas like Alaska. The challenge then becomes separating "normal," or background seismicity from anomalous seismicity potentially indicative of a coming eruption.

To help distinguish VT swarm seismicity from tectonic seismicity, simple tests for "swarminess" of earthquake sequences have been applied, yet none of them are completely diagnostic. Volcanic earthquake swarms often have large b-values and no clear mainshock (McNutt, 1996, 2005), potentially distinguishing them from tectonic mainshock-aftershock sequences. Yet many counter-examples exist (Mori et al., 1996; Roman et al., 2004; Pesicek et al., 2008; Garza-Giron et al., 2018). Other general characteristics of pre-eruptive VT swarms are (1) the number of events and average energy increases over time, (2) the largest events occur in the middle of the swarm, and (3) the swarm includes several events within <sup>1</sup> /2 magnitude unit of the largest event (White and McCausland, 2016). However, these rules of thumb might also be true for tectonic earthquake swarms unrelated to volcanism (e.g., Vidale et al., 2006; Holtkamp and Brudzinski, 2011). In addition, at the onset of an earthquake sequence, it may be difficult to know which of these traits will hold true, if any. Conclusively attributing crustal seismicity near volcanoes to magmatic processes remains difficult, especially early in an unrest sequence before gas, heat, deformation, or other anomalies are detectable at the surface.

Despite these challenges, seismic monitoring remains the cornerstone of eruption forecasting. Volcano observatories commonly must decide at what point a seismicity increase is "anomalous" in comparison to background seismicity and at what point it is concerning enough to notify authorities. These decisions are complicated by seismic network outages, leading to earthquake catalogs with time variable completeness thresholds. Automated tools assist in this effort, and additional insights may be gained by more formally investigating the relationship between seismicity and subsequent eruptions at many volcanoes.

In this paper, we search for anomalous seismicity (dominantly VT) preceding seismically monitored eruptions to identify the circumstances under which seismicity may be most useful for eruption forecasting. To do so, we detect statistically significant seismic rate anomalies above a volcano-specific, empiricallyderived threshold using a variation of a common statistical test (the β-statistic; Matthews and Reasenberg, 1988). In order to determine whether anomalous rate increases occur or do not occur prior to eruptions, we require a catalog of eruptions of various sizes and styles that were seismically monitored during both inter-eruptive and eruptive periods. Specifically, we require the ability to determine long-term background seismicity rates. Although seismic monitoring of eruptions is now routine worldwide, few places on Earth have had seismic monitoring operating consistently for long enough to properly estimate a volcano's long-term background seismicity rate. Even fewer have this type of monitoring consistently at many volcanoes with multiple, well documented eruptions of various styles and sizes. One place where this type of monitoring exists is Alaska.

We use earthquake catalog data from Alaska and the record of eruptions seismically monitored by the Alaska Volcano Observatory (AVO) to statistically identify periods of seismicity exceeding background levels, and then associate these anomalies spatially and temporally with subsequent volcanic eruptions. We attempt to determine how often and under what circumstances seismic rate anomalies occur before eruptions in Alaska. Once identified, we search for patterns among the results that might prove useful for eruption forecasting in the future. We also investigate other non-eruptive seismicity increases and their causes and compare those to pre-eruptive seismicity. We take a quantitative approach using the β-statistic and specific temporal and spatial parameters to identify seismic rate anomalies. We vary these parameters over reasonable ranges and then use these results as a whole to infer when such anomalies might be useful for forecasting in the future. The findings presented herein should prove valuable for interpreting future seismicity, both in Alaska and at analogous volcanoes worldwide.

#### DATA

The data used in this study are composed primarily of Alaska earthquake catalogs and eruptive chronologies beginning with the June 1992 eruption of Mt. Spurr volcano and ending with the March 2016 eruption of Pavlof volcano (see Cameron et al., 2018). The chronology data (Ogburn et al., 2016) come from a variety of sources, including specific eruption literature and the Geologic Database of Information on Volcanoes in Alaska (GeoDIVA, Cameron and AVO staff, 2014). To define eruptive periods (**Table 1**), we consider the start of an eruption to be the first magmatic or major phreatic explosion (distinct from normal, background fumarolic activity, or steam plumes). Ends of eruptions are more difficult to determine but are based on the return to background levels of activity, the cessation of unrest, and/or the lowering of alert levels, depending on the specific eruption. We analyze 20 eruptions with Volcanic Explosivity Indices (VEI; Newhall and Self, 1982) ranging from 1 to 4 at 8 volcanoes (**Figure 1**, **Table 1**).

Although AVO produces its own earthquake catalog for seismically monitored volcanoes (e.g., Dixon et al., 2013), we use the Advanced National Seismic System (ANSS) composite catalog, which contains those events located by the Alaska Earthquake Center (AEC) and the USGS National Earthquake Information Center (NEIC), in addition to those located by AVO. The combined catalog allows us to analyze distal earthquakes that are potentially outside of the AVO network and thus not directly linked to a particular volcano. However, the composite nature of the ANSS catalog results in the loss of some useful event attributes, such as location uncertainty and source type descriptions, which limits our analysis to some degree (see section Discussion).

The background time period for each volcano is defined by the number of days of seismic monitoring minus eruptive periods and network outages. To determine seismic network outages, we use the results of the AVO network health analysis by Buurman et al. (2014) for the period October 2002 through December 2011 and perform our own similar analysis of the continuous seismic data to define outages since 2012. Network outages are defined as periods when less than four stations within 30 km of a volcano were transmitting data. Thus, our definition of background seismicity rate (T) is limited to years where network health can be readily determined from consistently archived continuous seismic recordings, which corresponds to the time period beginning in 2002 through early 2016.

Finally, we construct specific background earthquake catalogs for each volcano spatially. The catalogs contain all shallow (<=30 km) crustal events from the ANSS catalog since 1990 within a specified search radius (R) from the volcanic center, minus events that occurred during eruptive periods. We further limit the catalogs to events with magnitudes greater than each network's magnitude of completeness (Mc) over T, which we approximate as M<sup>c</sup> = 0 for all eight volcanoes studied (Dixon et al., 2013).

#### METHODS

A primary goal of this study is to quantitatively identify seismic rate anomalies near Alaska volcanoes and to associate these anomalies temporally with subsequent eruptions or lack thereof.


TABLE 1 | Eruption data.

\*Defined as first magmatic or large phreatic explosion or eruption onset.

†Using parameters T<sup>a</sup> <sup>=</sup> 14 days and R <sup>=</sup> 20 km.

‡Phreatic eruption only (Herrick et al., 2014).

\*\*Demarcated by years in repose > 15.

Accordingly, we seek to determine when a particular period of seismicity is statistically above the background rate. To do so, we use the β-statistic (Matthews and Reasenberg, 1988), which detects changes in earthquake rates by comparing the difference between the number of events in a given time period to the expected number of events in that time period (assuming the seismicity is Poissonian), normalized by the standard deviation of the expected number. This common statistical test has been used successfully in many tectonic environments to identify subtle changes in seismicity rates, such as identification of dynamically triggered seismicity and stress shadows following large earthquakes (e.g., Reasenberg and Matthews, 1988; Gomberg et al., 2001). Following Aron and Hardebeck (2009) and Aiken and Peng (2014), the β-statistic is defined as

$$\beta = \frac{N\_a - NT\_a/T}{\sqrt{N(\frac{T\_a}{T})(1 - \frac{T\_a}{T})}} \tag{1}$$

where N is the number of events in the background time period (T) and N<sup>a</sup> is the number of events in a specific time period (Ta) of interest (**Table 2**). As the null distribution for β is approximately Gaussian (Matthews and Reasenberg, 1988), absolute values of the resultant β ≥ 2.57 (1.96, 1.64) are statistically significant at 99% (95, 90%) confidence (Aron and Hardebeck 2009), and positive (negative) β values denote seismicity increases (decreases).

We compare long-term background seismicity rates to shortterm windows of interest and search for statistically significant differences at a 95% confidence level. This should occur when β exceeds a threshold of 1.96. However, because volcanic seismicity may not be strictly Poissonian, we additionally seek an objective empirical threshold (βE) for β, following Prejean and Hill (2018), to determine if seismicity in the time and crustal volume of interest is truly anomalous compared to background rates. To define this threshold for each volcano, we calculate the β-statistic every day over T for specific values of T<sup>a</sup> and select a β<sup>E</sup> that is exceeded only 5% of the time. With few exceptions, the resulting β<sup>E</sup> values are larger than 1.96 (**Table 1**). Where β<sup>E</sup> < 1.96, we use the significance threshold at the 95% confidence level (1.96) rather than the lower, empirically-derived β<sup>E</sup> threshold.

In Alaska, we seek all β above β<sup>E</sup> (hereafter "anomalies") and examine whether or not these rate increases are temporally associated with subsequent eruptions. Specifically, we search for all β anomalies preceding Alaskan eruptions of VEI 1 or greater since 1990. Because the β-statistic was designed to detect subtle divergences in seismicity from background rates and because volcanoes often have non-eruptive swarms, we do not expect every β-statistic anomaly to result in eruption. Nonetheless, this

TABLE 2 | β-statistic symbols.


technique allows us to explore and quantify subtle precursors, including any that may have been missed previously for eruptions not forecast (see Cameron et al., 2018). In order to incorporate all events occurring prior to an eruption onset, we specify T<sup>a</sup> windows that end at the eruption start time. Thus, T<sup>a</sup> windows are necessarily defined backward in time based on the eruption onset. This retrospective approach is applied in order to maximize the identified anomalies. However, we also illustrate the potential use of the tool for real-time forecasting later in the discussion section using forward moving overlapping windows. **Figure 2** shows an example of the test as applied to Augustine volcano, which erupted most recently in 2006.

#### RESULTS AND SENSITIVITY

The identification of seismic rate anomalies preceding eruptions in Alaska depends on the particular parameters chosen for the test, in particular on T<sup>a</sup> and R (**Table 2**). We explored a range of reasonable values for these parameters based on our study goals and prior knowledge of seismic sequences preceding past eruptions. For the radial earthquake search, we used R-values from 10 to 50 km from the volcanic centers. Although seismicity directly preceding eruptions generally occurs close to the eruptive vent, VTs >45 km distal of volcanoes have been associated with subsequent eruptions (White and McCausland, 2016). Thus, we allow for the possibility that precursory VT seismicity may occur as far as 50 km distal of the volcano, outside the AVO local monitoring networks. However, searching out this far from the volcanoes likely incorporates more tectonic seismicity, which may result in the inclusion of anomalies unrelated to magmatic activity and also affects our measurement of "background" seismicity rate.

The parameter T<sup>a</sup> defines the length of the time window over which to define an anomaly, and its choice is guided by the goals of a particular study. In this study, the choice of T<sup>a</sup> should be based on typical time spans of precursory sequences of seismicity leading into eruptions. However, the choice of T<sup>a</sup> also affects the resulting β<sup>E</sup> threshold and the size and number of detected anomalies. The design of our β<sup>E</sup> empirical threshold is such that the largest 5% of all possible T<sup>a</sup> windows are by definition anomalous. As a result, larger T<sup>a</sup> windows generally produce fewer anomalies and lower β<sup>E</sup> values than smaller windows, and vice versa. In addition, the length of T<sup>a</sup> need not be directly related to the time duration of anomalous seismicity and does not necessarily relate to eruption run-up time. Brief but sufficiently intense periods of seismicity can produce anomalies even when the size of T<sup>a</sup> is much longer than the event sequence. In contrast, longer but less intense periods of variably elevated seismicity tend to create multiple separate anomalies for small values of T<sup>a</sup> and may not produce anomalies at all for larger choices of Ta. It is the overall number of events (Na) occurring in T<sup>a</sup> that is important. Thus, the value of T<sup>a</sup> should be sufficiently long so as to encompass significant rate increases but not so long as to minimize their significance. In this study, we search a range of values for the T<sup>a</sup> window length (3–60 days), initially seeking the value that will identify as many precursory anomalies as possible, then varying the value to explore its effects on the results.

Searching all combinations of T<sup>a</sup> and R, we have identified 6 seismic rate anomalies preceding seismically monitored eruptions in Alaska for eruptions at Spurr (1992), Shishaldin (1999), Veniaminof (2005), Augustine (2005), Okmok (2008), and Redoubt (2009) volcanoes (**Figure 3**, **Table 1**). Of these eruptions, only Veniaminof (2005) and Okmok (2008) were not forecast by AVO (Power et al., 1995; Power and Lalla, 2010; Buurman et al., 2013; Cameron et al., 2018). In the case of Okmok (2008), the pre-eruptive seismicity was too brief (∼2 h) for AVO to publish a notification (Larsen et al., 2009). In the case of Veniaminof (2005), the seismicity increase consisted of only 13 events none of which were near the summit (**Figure 3**). For the remaining 14 eruptions at 3 volcanoes, we do not identify precursory seismic rate anomalies for any combination of T<sup>a</sup> and R. **Figure 3** shows the 6 precursory anomalies using T<sup>a</sup> = 14 days and R = 20 km, which are the maximum values for these parameters over which all 6 anomalies are identified. However, these pre-eruptive rate anomalies are identified over various combinations of T<sup>a</sup> and R, depending on the volcano, and no overall optimal values are illuminated by our analysis. In **Figure 4**, we keep T<sup>a</sup> = 14 days and explore how variations in R affect the results. Conversely, in **Figure 5**, we vary T<sup>a</sup> while keeping R = 20 km. With few exceptions, these figures show that the identified precursory anomalies are generally stable with respect to these variations in T<sup>a</sup> and R. **Figure 4** shows that precursory anomalies are identified at Augustine, Okmok, Shishaldin, and Redoubt no matter the choice of radius (T<sup>a</sup> = 14), whereas anomalies at Spurr and Veniaminof are dependent on the specific choice of R. At Spurr, only radii ≤20 km produce a precursory anomaly, while at Veniaminof, only radii ≥10 km produce anomalies. **Figure 5** shows precursory anomalies identified at Okmok, Redoubt, and Augustine for all choices of Ta. However, the β value of the anomalies varies in relation to the specific timing of seismic rate peaks. For instance, at Augustine, β correlates with window length, reflecting the extended nature of the precursory ramp up in seismicity. In contrast, β is anti-correlated with window length at Okmok, reflecting the short duration of the ∼2 h precursory sequence (Larsen et al., 2009). Anomalies preceding the eruption at Spurr are only identified when T<sup>a</sup> = 3, 7, 14, and 60 days, but not 30 days. At Veniaminof, precursory anomalies are not identified when T<sup>a</sup> = 60 days. The variations in anomaly detection due to R and T<sup>a</sup> reflect the spatio-temporal variability of the seismic

catalogs that likely results from volcano- or intrusion-specific factors such as the local stress field and/or intrusion size and rate.

Although T<sup>a</sup> and R are the most influential parameters on the identification of anomalies, the results may also be affected by other factors, such as the filtering of the volcano catalogs in order to remove events during eruptive periods. Although most eruption start dates are well-defined, the subjective demarcation of eruptive versus pre- and post-eruptive seismicity in some cases may introduce uncertainty into our results. In rare cases, visual confirmation of eruption onsets may be lacking, introducing uncertainty into the estimated start times. In Alaska, onsets of small eruptions at remote but frequently active volcanoes are often difficult to determine (e.g., 2004 and 2005 Veniaminof eruptions). The prime example of eruption onset uncertainty is the 1999 eruption of Shishaldin volcano, for which considerable uncertainty exists in the start date. Herein we consider the eruption to have started on 18 April 1999 (UTC), when visual confirmation of magmatic eruption was received, 2 days prior to the large explosive paroxysm of 19 April (Moran et al., 2002; Nye et al., 2002). However, vigorous and anomalous steam venting with possible ash, tremor, and a hot spot were detected as early as 9 February, followed by a 2 month pause in activity (Nye et al., 2002; McGimsey et al., 2004b). The uncertainty in this start date is important because on 4 March, a M5.2 earthquake occurred on the west flank, over a month before the April eruption start date. Using a February start date, the M5.2 and its aftershocks would be excluded from the analyzed catalog, and the event would be considered syn-eruptive, occurring after the eruption onset. Thus, in this particular case, the choice of start date determines whether this unusually large event and its aftershocks are included or omitted. Preferring 18 April as our start date, we identify and include a precursory anomaly for this eruption in our results (**Figure 3**). In effect, there is a clear seismicity rate precursor before the eruption paroxysm, but no precursor for a subtle eruption onset that may have occurred earlier.

With the exception of the Shishaldin and Veniaminof eruptions, most of the remaining eruption start dates have visual confirmation of the onset and thus minimal uncertainty. In contrast, eruption end dates are not well defined in Alaska or globally, even for well-monitored volcanoes (Siebert et al., 2011). Changes to eruption end dates may affect the categorization of significant periods of seismicity as either syn- or post-eruptive, potentially modifying our results. For some eruptions, seismicity remained elevated in the weeks following the defined end of the eruption (e.g., 2008 Okmok). Extending the eruption end date to include these events would exclude them from the analyzed catalog, potentially altering the computed β<sup>E</sup> threshold. In order to assess this potential issue, we modified the eruption end dates (**Table 1**) such that the eruption durations would change by ±20% and recomputed the results. We find that varying the

14, 30, 60 (brown, yellow, light blue, dark blue, purple). R = 20 km for all results shown. X-axes show the 3 months prior to the eruption. Y-axes are scaled by the maximum anomaly seen over the entire time period analyzed (see Figure 7). Maps show events within Ta days of eruption. See Figure 2 and text for additional details.

Figure 2 and text for additional details.

(see Figure 7). Maps show earthquakes within T<sup>a</sup> = 14 days of eruption. See

eruption end dates in this manner does not change the number of identified anomalies. Thus, our results are stable with respect to small changes in eruption end dates.

#### DISCUSSION

Our results statistically identify seismic rate anomalies preceding the eruptions at Spurr (1992), Shishaldin (1999), Veniaminof (2005), Augustine (2005), Okmok (2008), and Redoubt (2009) volcanoes (**Figure 3**). Using this method, 30% (6/20) of all eruptions analyzed and 43% (6/14) of VEI ≥ 2 eruptions have pre-eruptive seismic rate anomalies (**Table 1**). All magmatic eruptions (3/3) at closed-system volcanoes (repose > 15 years) have seismic rate anomalies (the phreatic eruption at Kanaga was not preceded by an anomaly). Further, 56% (5/9) of VEI ≥ 3 eruptions are preceded by seismic rate anomalies, including all 3 eruptions analyzed at long dormant, closed-system volcanoes. In contrast, eruptions at open-system volcanoes were rarely (13%; 2/16) preceded by seismic rate anomalies. Overall, the results support the widely held view that seismic rate anomalies are more common preceding eruptions at long dormant, felsic, closedsystem volcanoes than at more frequently active, mafic, opensystem volcanoes. In fact, Cameron et al. (2018) found similar relationships between VEI, composition, and open vs. closed systems and the success rates of AVO in forecasting eruptions of different types (Cameron et al., 2018), which is not surprising given that these eruptions were forecast based primarily on seismicity (e.g., Power et al., 1994, 1995; Nye et al., 2002; Power and Lalla, 2010; Buurman et al., 2013).

In evaluating the success rate of any forecasting tool, it is also important to quantify the number of false-positives, or herein, non-eruptive anomalies. We cannot completely quantify the number of seismic rate anomalies that are not followed by eruptions because we cannot apply our test to all seismically monitored volcanoes in Alaska, rather only those that have erupted at least once since 1992. We can, however, evaluate the number of non-eruptive anomalies produced by this particular method at those volcanoes that have erupted since 1992 (**Table 1**). **Figure 6** shows the complete time series for all 8 volcanoes analyzed, using T<sup>a</sup> = 14 days and R = 20 km. In addition to the 6 pre-eruptive anomalies (**Figure 3**), we have also identified a number of non-eruptive anomalies (69 for T<sup>a</sup> = 14 days and R = 20 km). Most of these non-eruptive anomalies are short-lived and represent brief increases in seismicity at volcanoes with high background noise or small numbers of earthquakes at seismically quiet volcanoes. Many of these short-lived anomalies are also small in amplitude, implying lower confidence (**Figure 6**). Some of the anomalies, however, are more sustained, and have known origins. The long duration, non-eruptive anomaly at Shishaldin in 2002 may represent shallow proximal unrest related to ongoing phreatic activity (Neal et al., 2005). The long period of multiple anomalies at Spurr in 2004 has been clearly associated with a magmatic intrusion, or "failed eruption," where magma stalled before reaching the surface (Power et al., 2004; Moran et al., 2011). The 2008 anomaly at Kanaga is related to aftershocks of a distal M6.6 earthquake, that is presumably tectonically driven. The sustained non-eruptive anomaly at Pavlof in 2002 is actually related to unrest at nearby Mount Hague, part of the Emmons Lake caldera system (Neal et al., 2005), illustrating one difficulty with using large radii to search for seismicity in areas with multiple, closely-spaced volcanoes. Several other non-eruptive anomalies occur in the days to months following after eruptions. The non-eruptive anomaly at Spurr following the 1992 eruption could be considered as post-eruptive unrest related to continued intrusion or crustal adjustment (Cameron et al., 2018). In fact, many of the other non-eruptive anomalies (e.g., Augustine, 2007; Okmok, 2009) could also be considered post-eruptive unrest (Cameron et al., 2018). The Okmok, 2009 non-eruptive anomaly, for example, coincides with a thermal anomaly, tremor-like events, and shallow slope failure in late February to early March 2009, followed by tremor bursts in May 2009 (McGimsey et al., 2014). Applying a post-eruptive window of 1-year eliminates 16% of these non-eruptive anomalies. However, because of the close spacing of Veniaminof eruptions, application of such a post-eruptive window would eliminate the precursory anomaly before the 2005 eruption. Finally, other non-eruptive anomalies could also be related to post-eruptive mass-wasting processes; for example, the small non-eruptive anomaly at Augustine in 1998 is related to the collapse of the 1986 spine (McGimsey et al., 2004a).

Although applying a post-eruptive window reduces the number of non-eruptive anomalies, some volcanoes still have high numbers of non-eruptive anomalies. For example, many short-lived non-eruptive anomalies are identified at Veniaminof, yet only one of the 7 eruptions shows a precursory anomaly (**Figure 6**). Thus, although an anomaly was identified before the 2005 Veniaminof eruption, AVO could not have confidently forecast the eruption based solely on seismicity rates because of the high non-eruptive anomaly rate. In general, open-system volcanoes like Pavlof and Veniaminof have high numbers of noneruptive anomalies and low numbers of pre-eruptive anomalies and thus do not appear to be amenable to reliable eruption forecasting based only on seismicity rates. Other data streams, such as gas or deformation data may be necessary to improve forecasting at such open-system volcanoes (e.g., de Moor et al., 2016). In general, the lack of precursory seismic anomalies at these volcanoes, despite the choice of T<sup>a</sup> and R, confirms that AVO did not miss any subtle or distal pre-eruptive seismicity and could not have forecast them based on seismicity (see also Cameron et al., 2018).

Unlike the anomalies that precede the eruptions, the number of non-eruptive anomalies varies significantly depending on the choice of parameters used for detection, and it is useful to investigate which parameter values minimize the overall number of non-eruptive anomalies. By fixing T<sup>a</sup> = 14 and varying R as in **Figure 4**, we find R = 30 km to produce the fewest noneruptive anomalies. However, changes in counts of non-eruptive anomalies due to variations in radius are small, and there is no clear trend that would identify a clear optimal value for minimizing non-eruptive anomalies overall. In contrast, the size of the T<sup>a</sup> window has a clear impact on the non-eruptive anomaly count because the length of T<sup>a</sup> is directly correlated with the number of anomalies and the β<sup>E</sup> empirical threshold, as discussed above. For example, when using T<sup>a</sup> = 60 days, the non-eruptive

FIGURE 6 | Complete results for all nine volcanoes with eruptions analyzed for T<sup>a</sup> = 14 days and R = 20 km, similar to Figure 3. Note that the 1989 Redoubt and 2014 Veniaminof eruptions (light red) were not analyzed in this study due to lack of sufficient seismic monitoring. See Figure 2 for additional details.

anomaly rate is reduced from 69 to 16. Although a window of this length may be less useful for real-time forecasting purposes, it does allow us to more easily investigate how often the most significant periods of unrest detected by this method lead to eruption. Overall, when implementing a post-eruptive window of 1 year, using R =20 km and T<sup>a</sup> = 60 days, we find that 31% of anomalies lead to eruption. This rate is partially confounded by other factors mentioned earlier, (e.g., close proximity of other restless volcanoes), and the fact that we do not consider noneruptive anomalies at volcanoes with no eruptions since 1992. However, Cameron et al. (2018) find a similar rate of unrest without eruption when evaluating AVO color code changes (29% of unrest led to eruption, 71% did not). Some studies find roughly similar ratios (30–38% of unrest led to eruption) using a variety of methods and proxies for unrest (e.g., Newhall and Dzurisin, 1988; Gudmundsson, 2006; Biggs et al., 2014), while other studies find higher rates of unrest leading to eruption (Klein, 1982; Phillipson et al., 2013; Winson et al., 2014; 60–67%). Quantifying the probability that unrest will lead to an eruption is a crucial open question for forecasting, and in fact forms an early (often first) node in many event trees used for eruption forecasting (e.g., Newhall and Hoblitt, 2002; Wright et al., in press).

Toward the overriding goal of further improving eruption forecasting, the methods and results presented herein represent progress toward better understanding the relationships between precursory seismicity and eruptive activity. Our results confirm that we can expect, with a relatively high degree of confidence, anomalously high seismicity rates preceding large (VEI ≥ 3) explosive eruptions at closed-system volcanoes. However, this method confirms that seismicity rate changes have a relatively low predictive power for smaller eruptions at open system volcanoes. These results help to weigh the significance of seismic anomalies detected relative to other monitoring data streams when evaluating unrest and formulating a forecast at volcanoes of different types, and are already being used by VDAP and many volcano observatories for forecasting around the world. However, we can go further and apply a slightly modified βstatistic test to make the method more directly applicable to future forecasting in Alaska. **Figure 7** shows how we might apply this test to volcanoes included in this study in near realtime, when eruption times are unknown. For this figure, we re-computed the β-statistic every day using seismicity from the preceding T<sup>a</sup> days, in contrast to previous figures where T<sup>a</sup> windows ended at the eruption start times (see section Methods). Due to the dependency of the results on the particular choice of T<sup>a</sup> (**Figure 5**), we simultaneously computed the results for different values of Ta, defining anomalies once the predetermined β<sup>E</sup> threshold (which could also be regularly updated) is exceeded. Although such anomalous seismicity would likely be noted by observatory staff, this approach would automatically confirm the anomaly rather than relying solely on human recognition, quantify its significance, and quickly place it in context of previous seismicity at the volcano. A quantitative approach such as this can properly account for factors changing with time (e.g., network upgrades) that may be missed by more qualitative or ad hoc assessments of seismicity. In this way, seemingly anomalous seismicity can be better and more quickly

assessed with respect to the long-term background rate and previous episodes of unrest. Outside of Alaska, this tool could also be applied in near real-time to aid in quantifying anomalous seismicity in comparison to background. Even when no prior eruptive activity has been seismically observed (and thus no useful β<sup>E</sup> threshold can be computed), we can still apply the test to more quantitatively compare current periods of unrest to previous unrest, provided that long-term catalogs of seismicity (ideally before and after historical eruptions) and information about the long-term monitoring history is available.

Although we have shown them to be useful, this tool and our results are imperfect attempts to address complex physical phenomena in a consistent way. We are not attempting to model or explain any physical volcanic process but rather are searching for commonalities in eruptive behavior despite important differences between and within the various volcanic systems. With this statistical tool, we seek to aid volcano observatories in identifying seismic rate anomalies above background when seismic network health and earthquake detection rates fluctuate. Toward this goal, we have made specific decisions because they allow us to consistently apply the test despite known shortcomings. For example, in our analysis we have not considered two event attributes usually included in earthquake catalogs for volcanoes: magnitude and event type. Although precursory patterns in event magnitude and overall energy release are quite important for forecasting (e.g., Murray and Endo, 1992; Cornelius and Voight, 1994), they are beyond the scope of the statistical test presented herein, which is concerned only with event rate. Similarly, we have not considered event type because the ANSS catalog does not retain this attribute in their combined catalog. As a result, we are including LP events in our analysis in addition to VTs. Although these events are relatively infrequent in most of the volcano catalogs included in this study (e.g., 12% of 2012 AVO catalog overall; Dixon et al., 2013), they are likely contributing to the identified rate anomalies in some cases, particularly the frequently active open system volcanoes, like Pavlof, which has a higher than average % of LPs. Finally, we have not formally considered location uncertainty in our analysis, which is also unavailable from the ANSS catalog. In general, AVO volcano catalogs have average uncertainties ≤ ∼2 km (Dixon et al., 2013), while AEC and NEIC location uncertainties are generally larger. However, we have implicitly incorporated epicenter uncertainty by varying the radial (R) search and exploring its effects on our results (**Figure 4**). Similarly, we found that small changes (±5 km) in the depth threshold we applied (30 km) did not affect the number of pre-eruptive anomalies identified.

Despite these limitations, our work contributes to improving eruption forecasting in several ways. Although seismic rate anomalies are commonly observed globally, most previous work has been focused on increases in seismic activity in the immediate vicinity of volcanic vents (e.g., Kilburn, 2003). In fact, definitions of "volcanic earthquakes" are often limited to those within 10 km of the summit (Shimozuru, 1971; McNutt, 1996). In addition, many studiesrely only on LP events for forecasting instead of VTs (e.g., Chouet et al., 1994; Boué et al., 2015). Although effective, such efforts are focused proximally, and there is the potential to miss earlier distal precursors, which may occur long before runups in vent related seismicity at long-dormant volcanoes (White and McCausland, 2016). For example, at Shishaldin volcano, distal seismicity between 10 and 20 km from the summit peaked more than 2 months before the 1999 eruption (Rasmussen et al., 2018; **Figures 3**–**5**); and distal seismicity (20–40 km from the summit) occurred at Augustine roughly 2 months before the start of the 2005–2006 eruption (Fisher et al., 2010; **Figure 4**). Finally, while retrospective deterministic eruption forecasts based on near vent seismicity continue to show promise in forecasting, they depend critically on rigorous independent calibration at each new volcano where they are applied (e.g., Boué et al., 2015, 2016; Chardot et al., 2015; Salvage and Neuberg, 2016). In contrast, our approach is to seek temporal seismic patterns that apply broadly to a set of volcanoes or a particular type of volcanic activity. The set could be the global set of eruptions, an ideal but lofty goal, or some specific subset, such as "eruptions at longdormant volcanoes in Alaska." However, our approach leaves several outstanding questions unanswered regarding the extent of the utility of our work. For instance, what utility might the βstatistic have when applied to more frequently erupting volcanoes such as Veniaminof? We've shown that seismic rate anomalies are rare preceding such eruptions in Alaska (**Figure 6**, **Table 1**), but we have not investigated why they occur in some cases (e.g., Veniaminof 2005; Shishaldin, 1999) and not others (e.g., all other Veniaminof eruptions). There might still be some correlation between VT swarms and certain types of eruptive activity that we could decipher if we had a larger statistical population to analyze, or if we incorporate other factors beyond event rate. We have not investigated correlations between seismicity rate and other variables, such as magma composition, run up times, or energy release (e.g., Thelen et al., 2010; Passarelli and Brodsky, 2012). These and other important correlations may also exist, and future work is aimed at finding them by analyzing seismic and other volcanic data beyond Alaska (Ogburn et al., 2016; Pesicek et al., 2017).

#### SUMMARY AND CONCLUSIONS

There are many documented cases of seismic rate increases preceding eruptions and intrusions worldwide. However, there are also many eruptions where no such precursors were identified, even when sufficient monitoring existed. In this study, we used the β-statistic (Matthews and Reasenberg, 1988), and determined an objective β threshold to quantify the prevalence of seismic rate anomalies preceding eruptions in Alaska and investigate their reliability as a forecasting tool. We find that 6 out of 20 eruptions in Alaska show precursory rate increases, including all 3 eruptions at volcanoes that have been dormant for at least 15 years, and that erupted with a VEI of 3 or greater (**Figure 3**). Thus, we confirm that seismic rate increases may be expected preceding eruptions at similar closed-system volcanoes in the future. Perhaps more importantly, although 3 other precursory anomalies were identified at volcanoes with shorter repose times (Veniaminof, Shishaldin, Okmok; **Figure 3**, **Table 1**), many other similar eruptions lack them, despite the fact that we are using a relatively sensitive test to identify rate increases (**Figure 6**). From this, we infer that seismic rate increases preceding eruptions at frequently active open-system volcanoes are relatively uncommon. Furthermore, we show that at the closed-system volcanoes with longer repose times, preeruptive seismic anomalies are usually the most significant anomalies identified. At open-system volcanoes, however, there are often higher numbers of non-eruptive anomalies, and this method has lower predictive power for these systems. Many other non-eruptive anomalies can be attributed to unrest at nearby volcanoes, non-eruptive volcanic activity, and cases of shallow intrusion of magma without eruption. At the limited number of volcanoes that we analyze, we find that 31% of seismic anomalies identified using this method lead to eruption, while 69% do not, in broad agreement with some other studies that quantify rates of unrest at volcanoes. Finally, we presented a statistical tool that may be useful for future eruption forecasting purposes, particularly when evolving seismic networks lead to temporally variable earthquake detection capabilities. The βstatistic properly considers the long-term background rate when analyzing periods of seismicity and provides a way to quickly and more easily assess apparent rate changes in the context of previous activity. We expect that with more calibration from a wider dataset this tool could prove useful for future eruption forecasting at volcanoes worldwide.

#### DATA AVAILABILITY STATEMENT

Earthquake data used in this study are from the ANSS Composite Catalog: http://www.quake.geo.berkeley.edu/anss/ catalog-search.html. Eruption chronologies are compiled from public sources as described in the manuscript and can be

#### REFERENCES


obtained by request from the authors. Seismic waveform data are available at IRIS: http://www.iris.edu.

#### AUTHOR CONTRIBUTIONS

SP, JP, and JW designed the analysis, JP and JW performed the analysis, and SO compiled the eruption chronologies. All authors contributed to the writing, interpretations and discussion of the results.

#### FUNDING

Funding was provided by the U.S. Agency for International Development Office of U.S. Foreign Disaster Assistance, under VDAP's Eruption Forecasting Information System (EFIS) database project.

#### ACKNOWLEDGMENTS

We thank C. Cameron for help defining AVO eruption dates and R. White and W. McCausland for discussions and feedback, and H. Buurman for sharing the results of her analysis of network health for the AVO seismic monitoring networks. We also thank Aaron Wech and two other reviewers for critical reviews that strengthened the manuscript. Data used in this study were collected by the Alaska Volcano Observatory, a cooperation between the USGS, The University of Alaska, Fairbanks, and the Alaska Division of Geological and Geophysical Surveys. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.


Newhall and R. S. Punongbayan (Seattle, WA: University of Washington Press), 339–350.


**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 Pesicek, Wellik, Prejean and Ogburn. 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.

## Changes in the Long-Term Geophysical Eruptive Precursors at Piton de la Fournaise: Implications for the Response Management

Aline Peltier <sup>1</sup> \*, Nicolas Villeneuve1†, Valérie Ferrazzini <sup>1</sup> , Séverine Testud<sup>2</sup> , Theo Hassen Ali <sup>2</sup> , Patrice Boissier <sup>2</sup> and Philippe Catherine<sup>1</sup>

1 Institut de Physique du Globe de Paris, Sorbonne Paris Cité, Observatoire Volcanologique du Piton de la Fournaise, Université Paris Diderot, UMR 7154 CNRS, La Plaine des Cafres, France, <sup>2</sup> UFR Sciences et Techniques, Université Jean Monnet de Saint Etienne, Saint Etienne, France

Edited by: Nicolas Fournier, GNS Science, New Zealand

#### Reviewed by:

Alessandro Tibaldi, Università degli studi di Milano Bicocca, Italy Dmitri Rouwet, Istituto Nazionale di Geofisica e Vulcanologia (INGV), Italy

#### \*Correspondence:

Aline Peltier peltier@ipgp.fr

#### †Present Address:

Nicolas Villeneuve, Laboratoire GéoSciences Réunion, Université de La Réunion, Institut de Physique du Globe de Paris, Sorbonne Paris Cité, CNRS, Saint Denis, France

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 29 March 2018 Accepted: 05 July 2018 Published: 25 July 2018

#### Citation:

Peltier A, Villeneuve N, Ferrazzini V, Testud S, Hassen Ali T, Boissier P and Catherine P (2018) Changes in the Long-Term Geophysical Eruptive Precursors at Piton de la Fournaise: Implications for the Response Management. Front. Earth Sci. 6:104. doi: 10.3389/feart.2018.00104 Anticipating eruptions early enough to give warning to authorities is one of the main goals in volcanology. However, identifying and providing unequivocal identification on volcano reawakening remain challenging issues, mostly when unrests are sudden or undetectable. At the Piton de la Fournaise volcano, a clear increase in both the seismicity and the ground displacements are systematically observed a few days/weeks before eruptions, and appear as clear eruptive precursors. Here a systematic study of these long-term precursors demonstrates the changes in their intensity, duration, and time of appearance during 1998–2017 (43 eruptions), directly linked to the influence of the pre- and post-summit caldera formation (April 2007) and to changes in the deep magma refilling process since 2016. These changes in the precursors were not without consequence on the early alert to the authorities, with some false alerts and late alerts. It is thus of prime importance for crisis management to bear in mind the possibility of these rapid changes and that of sudden volcanic unrest with little warnings, to be able to take the most appropriate decisions, in particular raising the level of alert or lifting it totally. The findings of this study have enabled the relevant authorities to improve the alert chain protocol, and scientists to communicate more efficiently with the decision-makers.

Keywords: Eruptive precursors, Piton de la Fournaise, volcano hazards and risks, volcano monitoring, crisis management

#### INTRODUCTION

Volcanic eruptions threaten communities, given their potential to cause fatalities and economic loss. Therefore, the expectation of society in volcano science is to provide local emergency management authorities and decision makers with timely forecasts of imminent volcanic eruptions. However, forecasting a volcano's behavior and being able to unambiguously construe volcanic unrest as eruption precursor remains complex. This applies especially when signals of renewed volcanic activity arise suddenly or are not even detected. On September 27, 2014, Mount Ontake (Japan) erupted without evident long-term precursors, killing 63 persons (including 6 missing). No long-term ground deformation and only 2 weeks of unusual seismic activity had been recorded before the eruption (Kato et al., 2015).

On the Piton de La Fournaise volcano (a hot spot volcano on La Réunion Island, Indian Ocean; **Figures 1A,B**), eruptive precursors, such as increase in volcano-tectonic seismicity, edifice inflation, and degassing, are now well identified, occurring on two time scales, which enable the OVPF (Observatoire Volcanologique du Piton de la Fournaise: Piton de la Fournaise Volcano Observatory) to warn local authorities of (1) high probability of an eruption in the following weeks/months (signs of reservoir pressurization; "Vigilance" alert level), and (2) high probability of an eruption in the following minutes/hours (signs of final dike propagation toward the surface; "Alert level 1—Imminent eruption"). Warning levels are issued within the framework of the ORSEC (Organisation de la Réponse de SEcurité Civile: Civil Security Response) "Piton de la Fournaise Volcano" plan, which is an emergency plan set up by the department responsible for the protection of the population in the event of unrest or activity of Piton de la Fournaise. Within this plan, OVPF is the first actor in the response chain; it informs the Prefecture (decentralized administrative service of the French State) via the Etat Major de Zone et de Protection Civile de l'Ocean Indien: EMZPCOI (Indian Ocean Zone and Civil Protection Headquarters) in the event of any change in volcanic activity and proposes to change the alert levels (Harris et al., 2017). The final decision to change the alert levels provided by the plan is the sole responsibility of the Préfet (head of the Préfecture). The Préfecture then communicates with other actors (town councils, gendarmerie (local police), central authorities, institutions and the media). Early warning in the event of the threat of an eruption is possible thanks to both a dense monitoring network implemented in the field by OVPF and recent advances in our knowledge of the volcano, making Piton de la Fournaise one of the world's best-studied basaltic volcanoes. Recent studies have enabled a better understanding of its feeding system, with evidence of several variably connected reservoirs distributed over ca. 11 km below the summit, its dynamics and their associated signals (e.g, Battaglia et al., 2005; Peltier et al., 2009; Brenguier et al., 2012; Got et al., 2013; Di Muro et al., 2014; Lengliné et al., 2016).

In spite of well-established patterns of eruptive precursors, major changes have appeared in the activity of Piton de la Fournaise—and in the associated long-term eruptive precursors—during the two last decades, notably as regards the links with the major collapse of the main summit crater in April 2007. For instance in 2016–2017, eruptions were preceded by totally absent or weak long-term precursors, and even for one eruption (September 11–18, 2016), this led to late communication with the Prefecture on the volcanic unrest state, going directly from "No alert" to "Alert 1—Imminent eruption" and by passing the intermediate "Vigilance" alert level.

A good knowledge of the long-term eruptive precursors and their evolution over time is of prime importance to be able to issue early and efficient warning and react to volcanic unrest, giving reliable information to authorities. It is against this background that the present study was set. We systematically studied the long-term seismicity and volcanic edifice deformation (duration and intensity) preceding each eruption during the period 1998–2017 (43 eruptions) and we studied the possible influence of the pre- and post-caldera formation on these eruptive precursors.

#### METHODS

Since December 1979, the Piton de la Fournaise volcano has been monitored by the OVPF, which depends on the IPGP (Institut de Physique du Globe de Paris) and is now one of the most efficiently and closely monitored volcanoes in the world, thanks to 100 or so instruments (seismometers, GNSS receivers: Global Navigation Satellite System, tiltmeters, extensometers, gas stations, optic and infra-red cameras) implemented in the field. From 1980 to the end of 2017, 68 eruptions were anticipated and followed by the OVPF. All these networks made it possible to identify three main precursors: volcano-tectonic seismicity, volcano deformation, degassing. These precursors occur on two time-scales linked to two distinct in-depth processes: (1) in the long term (weeks/months): slow edifice inflation (less than 3 mm per day), 10–100 VT earthquakes per day, and CO<sup>2</sup> ground degassing, linked to the refilling of the shallow magma system and its pressurization; and (2) in the short term (tens of minutes/hours): strong rapid ground deformation and a swarm of shallow (above sea level; asl) volcano-tectonic events (referred to as "seismic crisis"), linked to the final dike injection toward the surface (e.g, Peltier et al., 2009; Roult et al., 2012; Schmid et al., 2012; Boudoire et al., 2017; **Figure 2**). The Piton de la Fournaise volcano is a low degassing volcano making it challenging to monitor volcanic degassing on its flanks. The first permanent gas stations were implemented only in 2007, explaining why the first systematic degassing precursors have been evidenced at the Piton de la Fournaise only recently (Boudoire et al., 2017). Because of the late installation of the gas stations (after the 2007 collapse), the gas precursors are disregarded in this study and we have only focused on the evolution of the long-term seismic and ground deformation precursors. We made a systematic analysis of the pre-eruptive seismicity and deformation for the time series of 100 and 15 days (in order for the findings not to be affected by eruptions occurring too close to each other), preceding eruption onsets over the period 1998–2017 (**Table 1**). Since we studied the late stages of eruption triggering, we took into consideration only the shallow (0–2 km asl) volcano-tectonic seismicity below the summit, shown as daily rates, and the summit deformation shown here as baseline (i.e., line length between two GNSS stations) variations; both being related to the shallow magma system pressurization (Peltier et al., 2009; Roult et al., 2012; **Figures 2**, **3**, **Table 1**). In these time series, we did not consider the "seismic crises" and the strong rapid ground deformation preceding the eruptions by a few minutes/hours and linked to the final dike propagation toward the surface from the shallowest magma reservoir, located at about 1.5–2.5 km in depth (e.g., Peltier et al., 2009; Di Muro et al., 2014). The GNSS network being implemented only since 2004, in **Figure 2b** we also indicate the evolution of the FORX extensometer-opening component, in order to have an overview of the global edifice deformation throughout the period considered.

#### ERUPTIVE ACTIVITY BETWEEN 1998 AND 2017

Our study starts in 1998, the date of renewed eruptive activity after almost 6 years of rest. This date corresponds to the starting

point of new intense eruptive activity, but also to a densification of the monitoring network, with, notably, new seismometers and the implementation of a dense permanent GNSS network in 2004. This high eruptive activity and the new monitoring tools allowed a better characterization of the eruptive precursors.

Except for a late eruptive fissure opening at the end of the 1998 eruption, all eruptions of the 1998–2017 period occurred inside the Enclos Fouqué caldera (**Figure 1C**), like 97% of the recent eruptive activity (Villeneuve and Bachèlery, 2006). This caldera is fully uninhabited but visited by more than 100 000 people by year [source: Office Nationale des Forêts (Forestry Commission)].

From 1998 to 2017, we distinguish 3 periods of eruptive activity: 1998–2007, 2008–2010, 2014–2017, separated by two key events, the Dolomieu summit caldera collapse (340 m depth) during the major March-May 2007 eruption (about 240 Mm<sup>3</sup> of emitted lavas, i.e., about 10 times that of an average Piton de la Fournaise eruption), and a rest period of 41 months in 2011–2014 (e.g., Peltier et al., 2009, 2010, 2016; Roult et al., 2012; Staudacher et al., 2016; **Figure 2**, **Table 1**): (1) 1998–2007: 451 Mm<sup>3</sup> of cumulated emitted lavas (868 days of activity; lava flow rate of about 6 m<sup>3</sup> /s), a period marked by continuous refilling of the shallow magma plumbing system, and by signs



**136**


of short-term eruptive cycles culminating in more important distal eruptions (e.g., Peltier et al., 2009; Staudacher et al., 2016); (2) 2008–2010: 7.3 Mm<sup>3</sup> of cumulated emitted lava (91 days of activity; lava flow rate of about 0.9 m<sup>3</sup> /s), a period marked by a change in the eruptive activity after the 2007 collapse with only low-volume summit or near-summit eruptions (eight) and many aborted magma intrusions (eight). Roult et al. (2012) attribute this change to stress changes in the volcano edifice after the collapse and a concomitant decrease in volumes in magma input; (3) 2014–2017: 57 Mm<sup>3</sup> of cumulated emitted lavas (166 days of activity; lava flow rate of about 4 m<sup>3</sup> /s), a period marked by a renewal of the eruptive activity, with flank eruptions only and numerous long-lasting eruptions (3 eruptions of a duration >20 days; **Table 1**).

#### RESULTS

On the whole, a clear increase in both the seismicity and ground displacements (global edifice inflation) is systematically observed a few days/weeks before the eruptions, and these appear as clear eruptive precursors (**Table 1**, **Figure 2**). The summit volcanotectonic seismicity is of low magnitude (90% of the events have magnitude < 1) and is mainly located below the summit craters, at a depth comprised between 500 and 2,000 m asl (Massin et al., 2011; Lengliné et al., 2016). The source at the origin of the ground deformation has been attributed to a pressurized magma reservoir located at a depth of around 0–1,500 m above sea level (Peltier et al., 2009, 2016), just below the summit volcano-tectonic seismicity. In spite of the similarities in the location of the seismicity and of the pressure sources from one pre-eruptive period to the other, the intensity and the time of occurrence of these long-term eruptive precursors vary according to the periods considered, namely the three periods of activity previously distinguished.

(1) 1998–2007: long-term seismicity and inflation appeared about 100 days before the eruption (**Table 1**; **Figures 2**, **3**); with an almost steady state of continuous pre-eruptive edifice inflation (only short-term deflation was recorded following the major distal eruptions; **Figure 2b**); and throughout the period, an evolution toward an increase in the long-term pre-eruptive seismicity was observed with an average of about 40 earthquakes in 1998–2000 vs. 550 in 2006–2007, during the 15 days preceding an eruption (**Figure 4**). The increase in seismicity and continuous inflation clearly begins in 2000 (**Figure 2b**; Peltier et al., 2009).

(2) 2007–2010: pre-eruptive long-term seismicity and inflation appeared later, compared to the 1998–2007 period, about 40–50 days before the eruption (**Figure 3**); the longterm pre-eruptive seismicity included some strong seismic swarms with more than 100 earthquakes per day not followed by eruptions (**Figure 3**). The long-term pre-eruptive edifice inflations were weaker than previously observed (low volume of magma involved at depth originating low volume of erupted lava during this period, **Table 1**) and between these phases a major continuous summit deflation was recorded (summit contraction of ∼3 cm/y; **Figure 2**).

Between 2011 and June 2014, a period of volcanic rest with no eruption occurred and was characterized by edifice deflation and weak summit seismicity (**Figure 2**).

(3) 2014–2017: less intense pre-eruptive long-term seismicity and inflation; the renewal of eruptive activity on June 20, 2014 was preceded by only 11 days of precursors. Apart from 2 days of strong seismicity, the level of seismicity and deformation remained low (1712 earthquakes including 360 and 687 during two seismic swarms on June 13 and 17, respectively; and 1 cm of summit elongation; **Table 1**).

For the following 2015 eruptions (four events), the almost continuous inflation was observed once more, but at a lower rate compared to the two first periods (summit elongation of 0.2–0.4 mm/d vs. 0.6–1 mm/d in 2004–2007; **Table 1**; Peltier et al., 2016). And in 2016–2017, pre-eruptive long-term inflation became discontinuous, and seismicity was lower (a few dozen events during the 15 days preceding an eruption; **Table 1**) and appeared only one (September 11, 2016 eruption) to ten (January 30, 2017; May 26, 2016) days before the eruption. We can note that 40–60 days before the eruptions of September 2016, January 2017 and July 2017, 1–2 weeks of seismicity increase were observed during stages of short-term inflation renewal (**Figure 3**).

In short, the long-term pre-eruptive seismicity and deformation (linked to the plumbing system pressurization) increased progressively from 1998 until the major eruption of March-May 2007, during which the Dolomieu crater collapsed, and then progressively decreased and appeared later and later (from about 100 days before the eruption in 1998–2007 to about 15 days in 2014–2017; **Figure 3**). After the collapse of the Dolomieu crater, in 2008–2010, even though the time of occurrence of the seismicity was shorter, seismicity remained high, taking the form of a large number of seismic crises not followed by eruptions (**Figure 3**).

### DISCUSSION

#### Influence of the Pre- and Post-caldera Formation on the Long-Term Pre-eruptive Precursors

Between 1998 and 2017, major changes occurred in the Piton de la Fournaise activity and in the associated long-term preeruptive precursor, reflecting major changes in the dynamism and stress state of the volcano. This is especially well visible on **Figure 4**, where the number of long-term pre-eruptive earthquakes increased before the collapse and then decreased after. The events of March-May 2007 (a major eruption and the Dolomieu crater collapse) appeared as key events in the recent history of the volcano, and seem to have influenced both the volcano activity and its eruptive precursors (Peltier et al., 2010; Staudacher, 2010; Massin et al., 2011; Roult et al., 2012).

From 1998 to 2007, the sustained activity of the volcano was maintained by a continuous filling up of the plumbing system, evidenced by near continuous summit inflation (Peltier et al., 2009). This led to a progressive weakening of the medium (Got et al., 2013) at the origin of the increase in seismicity throughout the period (**Figure 4**), even though part of the accumulated stress had been released during distal eruptions. Indeed, Got et al. (2013) showed that the eruptive cycles spanning the 2000–2007 period were linked to nonlinearity in the stress state of the edifice, as the strong eastern flank plastic displacement during distal eruptions enabled the stress accumulated during an eruptive cycle to be release and to enable the start of a new cycle. In spite of this release, the long-term pre-eruptive seismicity remained high and continued to increase throughout the period, and the continuous and long-term damage in the medium culminated with the April 2007 crater caldera collapse. The observed increase in seismicity between 2000 and 2007 could thus be a very long-term precursor of the in-depth collapse initiation due to the progressive weakening of the medium, notably linked with the continuous edifice inflation; and the triggering of the final collapse would occur due to the fast draining of the reservoir during the beginning of the March-April 2007 eruption (Michon et al., 2011).

After this major eruption, important changes occurred in the volcano and on its surface, consequences of the draining of the shallow reservoir and damage to the shallow plumbing system and its surrounding medium, due to the crater caldera collapse. Between 2008 and 2011, 8 low-volume eruptions and 8 aborted magma intrusions occurred (**Figure 2**), mostly in the upper part of the volcano. The long-term pre-eruptive seismicity remained high, but mostly occurred during seismic swarms accompanied by ground deformation (magma intrusions) or not (stress release in the medium). The seismic swarms not accompanied by ground deformation were not linked to magma migrations in depth but rather to stress changes and readjustments in the volcanic edifice following the collapse. The large number of aborted intrusions and the location of the eruptions inside or very close to the summit were related to the low magma volume and overpressure involved in depth, sometimes too low for the magma to reach the surface. The main continuous summit deflation observed between eruptive phases confirms that no deep magma refilling occurred during these periods, and that the main contribution of the deformation field during this period was the edifice destabilization and the large withdrawal of the magma reservoir in April 2007 (**Figure 3b**). This deflation continued between 2011 and 2014, also a period during which no sign of deep refilling was observed.

It was not until 2014 that a more "typical" activity, close to the one observed in 1998–2000, was again observed at Piton de la Fournaise, with increasingly late appearance of long-term eruptive precursors (from about 100 days before the eruption in 1998–2007 to about 15 days in 2014–2017; **Figure 3**), and less and less long-term pre-eruptive seismicity reaching a level similar in 2017 to the one observed in 1998–2000 (**Figure 4**). This shows a progressive decrease in the influence of damage and stress readjustment linked to the collapse.

#### Influence of the Deep Magma Refilling on the Long-Term Pre-eruptive Precursors

Since 2016, the well-established long-term pre-eruptive continuous inflation pattern previously observed—both before the 1998–2007 eruptions and the 2008–2010 eruptions—has disappeared, replaced by a discontinuous inflation trend (**Figure 3**) probably reflecting discontinuous deep magma pulses entering into the shallow plumbing system.

This new deforming process of the volcano makes it more difficult to anticipate eruptions in the long-term, as the pressure building up and the stress accumulation inside the reservoir occurred in steps (discontinuous magma accumulation), and not continuously as previously observed (**Figures 2b**, **3**). The time between two stages can be long and the last stages can be very short in duration, as observed in September 2016, when the final dike propagation was preceded by only limited warning signs (**Figure 3**). After two periods of edifice inflation not followed by an eruption (**Figure 3**; around May 27—June 13 and July 9–27, 2016), 1 month of summit deflation led the Prefecture to change the alert level, replacing "Vigilance" by "No alert" on September 5, 2016, only 6 days before the onset of the September 11, 2016 eruption. Fortunately, the short-term precursors (seismic crisis and strong deformation) associated with the final dike propagation toward the surface led the observatory to alert the Prefecture 56 min before the onset of the eruption, and the alert level passed directly from "No alert" to "Alert 1—Imminent eruption," bypassing the intermediate "Vigilance" alert level.

### Implications for the Communication System

Volcanoes are complex and highly non-linear natural systems. Their unrest and activity, as well as their associated signals, can thus quickly change. These quick changes can have consequences on the efficiency of the alert chain for the authorities, with some false alerts and late alerts. Indeed, whatever the natural phenomena studied (geoscience or meteorology), early warnings, but also caution, are necessary in the communication chain. This is even truer when the first communication comes from an operational research center, such as an observatory. Within the ORSEC "Piton de la Fournaise Volcano" plan, the role of the Observatory is to communicate to the authorities any changes (increase or decrease) in volcanic activity and in the number and intensity of eruptive precursors, via real-time and 24/7 volcano monitoring. The authorities (via decisions taken by the Préfet) decide any changes to be made in the alert level issued and transmit this information to the public institutions directly concerned and to the media. Any change in the alert level triggers the ORSEC "Piton de la Fournaise Volcano" plan that applies a dedicated protocol for the services involved (gendarmerie (local police), Office Nationale des Forêts (Forestry Commission), civil protection), and can have consequences on access to the volcano, which can thus, notably, affect tourism in this sector. As specified in the ORSEC "Piton de la Fournaise Volcano" plan, during the "Vigilance" alert level, access to the volcano summit is permitted but restricted to the official track (this is particularly restrictive for commercial guided tours that cannot take other paths), whereas during "Alert level 1—Imminent eruption," access is prohibited and visitors (and, possibly, the threatened population) are quickly evacuated on foot or by helicopter, weather conditions permitting. In view of these elements and the difficulties of access to the volcano and evacuations related to the site morphology (mainly in terms of intervention time),

in the event of "Alert level 1—Imminent eruption," any change in the alert level is an important decision, which must be carefully assessed. This requires an efficient communication chain between each actor (the observatory, authorities, and media), and necessitates sending reliable information from the observatory. Recent changes in the Piton de la Fournaise plumbing system feeding pattern and the associated precursors (time of appearance (**Figure 3**), number (**Figure 4**, **Table 1**), and intensity) have made this communication more complicated. On the one hand, this has led to over-evaluations (with, notably, false alerts), and, on the other hand, to late issue of information (e.g., the September 11, 2016 eruption, see section Influence of the Deep Magma Refilling on the Long-Term Pre-Eruptive Precursors), with a risk of loss of credibility for the observatory. One example is the February 4, 2015 eruption, for which the first subtle precursors appeared in November 2014. OVPF alerted the authorities, who issued the "Vigilance" alert level. On December 1, in the absence of any eruption, the alert was lifted by the Prefecture. On December 4, following several observations released by the OVPF, the Prefecture again issued the alert level, before lifting it again in early January 2015. The February 4, 2015 eruption began after a 1-month period of relative calm (which followed magma and stress accumulation during November and December) and a short-term seismic crisis. Because of the bad weather, only a few people were at the volcano at the time of the eruption, and only one 70-year-old man was in difficulty as a result of the eruption. He was on one of the official tracks and less than 100 m from the lava flow front, which blocked his way back. Fortunately, he was rescued by helicopter just before nightfall, after walking 7 h in very difficult conditions. His presence was authorized because no alert had been issued at the time of his departure to the volcano (a few hours before the onset of the eruption). With operational feedback, notably concerning crisis management in February 2015 and September 2016, the authorities, and even the population, maintain a very high level of confidence in the observatory. They are now aware of the complexity of the volcano's periods of reawakening and of the impossibility of accurate assessment as regards the precise timing of a coming eruption, and of the necessity of being cautious during any level alert changes, as well as during press releases.

Piton de la Fournaise is not the only volcano where behavior change has led to delay or failure in communication with local authorities. Before the eruption of September 27, 2014 at Mount Ontake, the volcano warning level had not been changed, because the number of low-frequency earthquakes detected was much lower than those recorded for the 2007 eruption (Yamaoka et al., 2016), leading to a catastrophic loss of lives.

It is thus of prime importance for crisis management and decision-making that scientists and politicians bear in mind the possibility of these rapid changes in a well-established pattern for a specific volcano and of the possibility of sudden volcanic unrest with little warning; particularly since communication between the actors involved (scientists, politicians and then the general public) can be a long process.

### CONCLUSIONS

Our detailed study shows that the intensity, duration and onset of long-term eruptive precursors changed quickly at Piton de la Fournaise between 1998 and 2017. These changes were linked to changes in the internal stress of the volcano before and after the April 2007 summit caldera, and after the period of calm of 2011–2014.


Volcanoes are highly non-linear systems, and these changes in the precursors were not without consequence on the early alert to the authorities, with some false alerts and late alerts. The findings of this study have enabled the relevant authorities to improve the alert chain protocol, and scientists to communicate more efficiently with the decision-makers.

#### AUTHOR CONTRIBUTIONS

AP and NV contributed to the design and implementation of the research. VF analyzed the seismic data. ST and TH contributed to the compilation of the database. AP, NV, PB, and PC contributed to the deformation data analyses.

#### ACKNOWLEDGMENTS

Data used in this paper were collected by Observatoire Volcanologique du Piton de la Fournaise/Institut de Physique du Globe de Paris (OVPF/IPGP), and can be found in this portal: http://volobsis.ipgp.fr. Part of this work was funded by Agence Nationale de la Recherche under contract ANR-16-CE04-0004-01 (SlideVOLC). We are grateful for helpful proof-reading from Nicole Richter, and thorough reviews by Dmitri Rouwet, Alessandro Tibaldi, the associate editor Nicolas Fournier and the chief editor Valerio Acocella. This is IPGP contribution number 3951.

#### REFERENCES


**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 Peltier, Villeneuve, Ferrazzini, Testud, Hassen Ali, Boissier and Catherine. 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.

## Forecasting Multiphase Magma Failure at the Laboratory Scale Using Acoustic Emission Data

Jérémie Vasseur\*, Fabian B. Wadsworth and Donald B. Dingwell

Earth and Environmental Sciences, Ludwig-Maximilians-Universität, Munich, Germany

Magmas fracture under high shear stresses, producing radiating elastic waves. At the volcano scale, eruption is often preceded by accelerating seismicity, while at the laboratory scales, sample failure appears to be preceded by similarly accelerating Acoustic Emission (AE). In both cases, empirical relationships between the acceleration and the time of the singular final event have offered tantalizing possibilities for forecast of eruptions and material failure. We explore the success of these tools in the laboratory by briefly reviewing datasets that have been presented previously and comparing the range of errors on forecast times with the range of errors associated with attempts to retrospectively forecast eruptions. We demonstrate that the heterogeneity of a system is crucial to making accurate forecasts on the sample scale, such that homogeneous systems are inherently unpredictable. We then analyse the effect of having an incomplete data sequence, as might be the case for real-time forecasting scenarios. We find that for heterogeneous systems, there is a critical proportion of the sequence that needs to have occurred before a forecast time converges on relatively low errors. As might be expected, the final portion of the sequence is the most important, while uncertainty on the start of the sequence is less important. Finally, we explore the simplest method for scaling the laboratory results to the volcano scenario.

Keywords: forecasting, porosity, acoustic emissions, precursors, inter-pore distance, porous magma, likelihood, probability density function

#### INTRODUCTION

Volcanic eruptions affect ∼600 million people worldwide (based on World Bank population data and the analyses of Small and Naumann, 2001; Auker et al., 2013), and yet the toolkit available for forecast of eruption times remains unreliable in many cases (see the analysis by Salvage and Neuberg, 2016). Most deterministic predictive tools are based on the observation that many geophysical signals (e.g., strain and seismicity) appear to accelerate toward a singular time, which coincides approximately with the onset of eruption (Voight, 1988; Voight and Cornelius, 1991; Kilburn and Voight, 1998; De la Cruz-Reyna and Reyes-Dávila, 2001; Kilburn, 2003; Ortiz et al., 2003; Smith et al., 2007; Smith and Kilburn, 2010; Bell and Kilburn, 2013; Boué et al., 2015; Salvage and Neuberg, 2016). This precursory phase of signal acceleration can last for minutes to years (Linde et al., 1993; Robertson and Kilburn, 2016). Accelerating signals can therefore be used to infer eruption timing ahead of the event itself, and in near real-time (Voight and Cornelius, 1991). In most cases, the feasibility of using these signals as predictors of eruption onsets has been assessed retrospectively, with variable success, such that real-time forecasting is not yet a useful reality.

#### Edited by:

Lauriane Chardot, Earth Observatory of Singapore, Singapore

#### Reviewed by:

Philippe Lesage, Université Savoie Mont Blanc, France Luca Caricchi, Université de Genève, Switzerland Marco Fazio, Georg-August-Universität Göttingen, Germany

#### \*Correspondence:

Jérémie Vasseur jeremie.vasseur@ min.uni-muenchen.de

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 08 March 2018 Accepted: 14 August 2018 Published: 03 September 2018

#### Citation:

Vasseur J, Wadsworth FB and Dingwell DB (2018) Forecasting Multiphase Magma Failure at the Laboratory Scale Using Acoustic Emission Data. Front. Earth Sci. 6:132. doi: 10.3389/feart.2018.00132

**143**

A step forward can potentially be made if we understand the physical underpinning of the acceleration of volcanic signals toward eruption. The accelerating nature of certain geophysical signals toward a time-singularity has been interpreted to represent the coalescence of multiscale fracturing processes (Kilburn, 2003) scalable down to rock-fracturing processes in the laboratory (Voight, 1989; Smith et al., 2007; Benson et al., 2008; Lavallée et al., 2008). This implies that the empirical powerlaw relationships that generally describe failure phenomena in the lab or seismicity approaching an eruption could emerge from physically-grounded models for nucleation, growth and coalescence of small-to-large nested fracture systems, which has been shown to be the case for some rupturing systems (Main et al., 2017). This also means that there is a class of selfsimilarity across a truly vast range of scales, from samples just a few centimeters long in the laboratory, to volcanic conduits for which fracturing depths begin at 1,500 m and propagate to the surface during ascent (Neuberg et al., 2006; Thomas and Neuberg, 2012).

The scalability of laboratory signals to volcanic signals remains uncertain in detail, but hinges on the assumption that the point at which a rock or magma fails to remain loadbearing on a small scale, is analogous to the point at which fractures in magmatic systems become pervasive over much longer lengthscales, and an eruption can proceed by material failure (c.f. Kilburn, 2003; Neuberg et al., 2006). The scaling laws proposed are therefore relatively simple (Benson et al., 2008; Tuffen et al., 2008), and are repeated herein. However, it is clear that more laboratory work could bridge this scale gap more rigorously. For instance, laboratory investigation of fault rupture velocities in viscoelastic magma would permit to constrain slip rates in volcanic conduits in nature and would help refine those scaling laws as well as volcanic eruption forecasting models.

In this contribution, we summarize the technical insights that have arisen from the campaigns of laboratory investigation, which may shed light on volcanic precursory signal evolution. We contrast these with some of the geophysical observations made at the volcano-scale and show where the most compelling links have been made. We provide new analyses of the failure of heterogeneous rocks, and contrast those with relatively homogeneous systems across a range of volcanically-relevant textural complexity.

#### HOW SUCCESSFUL HAVE RETROSPECTIVE OR REAL-TIME FORECASTING OF VOLCANIC ERUPTIONS BEEN?

A good starting point in assessing how successful mock-forecasts can be is when data have been acquired and can be analyzed retrospectively. We acknowledge that there may be a bias in the published work toward forecasts that are apparently successful, while less successful attempts are perhaps less likely to be reported. Marked exceptions to this are studies for which the central aim was to find methods of improvement of inaccurate forecasts such as Boué et al. (2015) and Salvage and Neuberg (2016).

If we take the time that an eruption has been forecast to have occurred as t<sup>p</sup> and the actual eruption time observed as te, then we can take the error on the forecast as t<sup>e</sup> − t<sup>p</sup>  /te. There are a few eruptions for which sufficient information exists that can be used to find this error magnitude on published retrospective forecast attempts, which are given in **Table 1**. We can see that the minimum error reported is as low as 0.002 for the Redoubt eruption in 1989–1990 (Voight and Cornelius, 1991) and as high as 0.36 for Pinatubo volcano erupting in 1991 (Bell et al., 2013). In these two cases the values refer to the minimum and maximum differences between the forecast and the eruption.

In general, there is little evidence in **Table 1** that a particular volcano type, eruption style, or magma composition, results in a better predictability when using all the forecasts made. Rather, it seems more likely that variations on the error of any forecast is dominantly dependent on the placement and quality of instrumentation, the numerical forecasting technique applied (c.f. Bell et al., 2011 for a discussion of techniques), and perhaps the nature of the seismicity (all data, or discriminated datasets from picking of specific event types).

#### QUANTITATIVE BACKGROUND

Here we aim to summarize the theoretical or empirical formulations that have been used to understand the phenomena of (1) magma or rock fracture, (2) empirical forecasting tools and probabilistic variations thereof, and (3) techniques to describe statistically heterogeneous materials. The latter is especially useful in linking the concepts in (1) and (2) as shown in part in Vasseur et al. (2017).

#### Rock and Magma Failure During Magma Ascent

Magmas ascend through the Earth's crust, during which both the country rock and the magma itself can break (Goto, 1999; Kilburn, 2003, 2012; Iverson et al., 2006; De Angelis and Henton, 2011; Thomas and Neuberg, 2012; Dmitrieva et al., 2013; Kendrick et al., 2014). In many cases, the seismic signals used to forecast eruptions are simply the entire aggregated number of events occurring in the vicinity of a volcano (as used in the original demonstration of Voight, 1988; **Table 1**). However, Neuberg et al. (2006) and Salvage and Neuberg (2016) demonstrate that low-frequency seismicity results from the repetitive fracturing events occurring at the same depth, interpreted to originate in the magma itself, and that these events are especially useful in retrospectively forecasting eruptions. Similarly, Kilburn (2003) points out that volcano-tectonic (VT) events resulting from rock fracture ahead of ascending magma must be the most useful seismic source for accelerating events that can be used to forecast the onset of a new eruption. Therefore, we expect that there is utility in low-frequency, magma-fracture events at established conduits exploited by fresh


<sup>a</sup>Reference(s) for the forecast window data.

<sup>b</sup>Reference(s) for the forecast error data.

magma repetitively (e.g., at Soufriere Hills volcano, from 1995 onward), and in volcano-tectonic events leading to eruptions and originating from rock fracture ahead of new magma (e.g., Pinatubo, 1991). In either case, it is important to quantify the stress magnitudes necessary for fracturing to occur, which is also a useful comparison between the volcano- and laboratory -scale.

Magma is a viscoelastic fluid or suspension, which can fail in a brittle manner when shear stresses reach a critical value τs . For pure liquids without suspended bubbles or crystals, this value has been empirically found to be of order τ<sup>s</sup> ∼ 10<sup>8</sup> Pa, and varies between 100 and 300 MPa (Simmons et al., 1982; Webb and Dingwell, 1990; Cordonnier et al., 2012b; Wadsworth et al., 2018). Assuming that the liquid phase originates the fractures (acknowledging that crystals can break during flow of two-phase or multiphase magmas; Cordonnier et al., 2009; Deubelbeiss et al., 2011), we can parameterize these breaking stresses in terms of the physics of fracturing viscoelastic fluids. Assuming that Maxwell's viscoelastic model is appropriate, the breaking point has been found to occur at a single Deborah number, De = 10−<sup>2</sup> (c.f. Webb and Dingwell, 1990), where De is the ratio of the stress relaxation time λ<sup>r</sup> and the stress accumulation time λ (Wadsworth et al., 2018). The stress relaxation time in Maxwell's model is λ<sup>r</sup> = µ/G∞, which contains the temperature- and composition-dependent liquid viscosity, µ, and the elastic shear modulus G∞. The threshold De = 10−<sup>2</sup> , implies that <sup>τ</sup><sup>s</sup> <sup>=</sup> <sup>10</sup>−2G∞, which is indeed <sup>τ</sup><sup>s</sup> <sup>∼</sup> 10<sup>8</sup> Pa, when G<sup>∞</sup> ∼ 10<sup>10</sup> Pa across most silicate magmatic liquid compositions, and independent of temperature (Dingwell and Webb, 1990). Additional scaling for this threshold has been made for heterogeneous magmas involving crystals (Cordonnier et al., 2012a) and bubbles (Kameda et al., 2008). This threshold provides a clear magma strength that has been shown to be met at depth during magma ascent and is the proposed origin of some of the accelerating seismicity approaching eruption (Goto, 1999).

The onset of solid rock fracture also occurs at threshold stresses, which in the simplest view, depend on the lithostatic "confining" pressure, the pressure of fluids in the pore spaces and the driving distribution of shear stresses (Jaeger et al., 2009). Additionally, this depends on the size and volume fraction of heterogeneity elements in the material (Baud et al., 2014). In detail, it is the distance between two pre-existing cracks, two crystals or two bubbles—between elements of heterogeneity that must be bridged in order for a system-spanning fracture to occur, and failure to ensue (Sammis and Ashby, 1986; Ashby and Sammis, 1990). As a leading example relevant to porous volcanic rocks, the unconfined compressive strength has the form τ<sup>s</sup> = aKIc/(φ b √ πR), where a and b are empirical constants, R is the radius of the heterogeneity element, and KIc is the fracture toughness (in Pa.m1/<sup>2</sup> ), which scale with the volume fraction of heterogeneity φ (Sammis and Ashby, 1986; Zhu et al., 2011; Heap et al., 2016). Vasseur et al. (2017) found that these distances between textural heterogeneity elements control the strength predictably when used in conjunction with a scaled static fracture-mechanics model. In both the volcanic rock failure and magma failure, the value of strength is therefore highly dependent on φ, and, in detail, on the pore size distribution (Sammis and Ashby, 1986; Vasseur et al., 2017; Wadsworth et al., 2018). However, the magnitude of strength is similar at τ<sup>s</sup> ∼ 10<sup>2</sup> MPa when φ → 0.

#### The Forecasting Toolbox

Voight (1988) proposed an empirical relationship between the acceleration of an observable ¨ and its rate ˙ , where we use dot-notation for time-derivatives. This has the general form

$$
\ddot{\Omega} = A \dot{\Omega}^{\alpha} \tag{1}
$$

where A and α are constants. Following Voight (1988), we can integrate (Equation 1) assuming that ˙ <sup>=</sup> ˙ <sup>0</sup> at t = 0 to find solutions for <sup>α</sup> <sup>=</sup> 1 and <sup>α</sup> 6= 1. Here ˙ <sup>0</sup> is the background event rate at <sup>t</sup> <sup>=</sup> 0. In most experimental scenarios, ˙ <sup>0</sup> = 0 at t = 0, but in the natural case, this may not be true (discussed later). Nevertheless, for <sup>α</sup> <sup>=</sup> 1, the result is an exponential increase of ˙ with t of the form

$$
\dot{\Omega}\left(t\right) = h \exp\left(qt\right) \tag{2}
$$

where <sup>h</sup> and <sup>q</sup> are constants. An exponential increase of ˙ does not reach a singularity and so is only predictive of an eruption or of material failure if we define a critical ˙ beyond which those critical events will occur. The more common case is that α > 1, which results in the commonly used Time-Reversed Omori Law (TROL; Kilburn, 2003; Bell et al., 2013, 2018)

$$
\dot{\Omega}\left(t\right) = k\left(t\_c - t\right)^{-\mathcal{P}}\tag{3}
$$

where k, t<sup>c</sup> and p are constants that are allowed to freely vary such that best-fit values can be found. The constant p is equivalent to 1/ (α − 1), used in previous work (Equation 1), and controls the non-linear shape of the approach of to a singularity at t<sup>c</sup> . This singularity represents a predictive quality of Equation (3) if we assume that a run-away to an infinite ˙ must coincide with a run-away of behavior to eruption.

These two methods, the exponential (Equation 2) and the power-law (Equation 3), have a suite of fit-parameters that are not known a priori and therefore must be acquired by algorithmic fitting to data. Bell et al. (2013) demonstrated that statistically reliable fits can be found when a "Maximum Likelihood" (ML) method is applied to Equations (2) and (3). The ML parameters are those resulting in a model that gives the observed data the greatest probability and those that maximize the likelihood function. The parameters are adjusted by minimizing the negative log-likelihood function using a downhill simplex algorithm. The fundamental advantage of the ML method given here is that it does not require binning of (t) data to get binned measures of ˙ , and can rather be used directly on the data themselves. Using this technique, for a time interval [t0, tn], the log-likelihood for Equation (2) can be written as (Bell, pers. comm.)

$$\ln\left(L\right) = q \sum\_{i=0}^{n} t\_i + n \ln\left(h\right) - \frac{h}{q} \left(\exp\left(qt\_n\right) - \exp\left(qt\_0\right)\right) \tag{4}$$

where L is the likelihood and n is the number of events. The same approach can be taken with the power-law method (Equation 3), for which the log-likelihood becomes (Bell et al., 2013)

$$\ln(L) = \sum\_{i=0}^{n} \ln\left(k\left(t\_i - t\_i\right)^{-p}\right) + K \tag{5}$$

where

$$K = \begin{cases} k\left( (t\_c - t\_n)^{1-p} - (t\_c - t\_0)^{1-p} \right) / \left( 1 - p \right) & \text{for } p \neq 1 \\ k\left( \ln\left( t\_c - t\_n \right) - \ln\left( t\_c - t\_0 \right) \right) & \text{for } p = 1 \end{cases}$$

Finally, for consistency, it may be useful to define a linear evolution of ˙ with <sup>t</sup> which is of ˙ <sup>=</sup> <sup>c</sup> where <sup>c</sup> is a constant. As with the exponential model, it is not clear what use a linear model can be for forecasting critical events, but it nonetheless may be a reasonable descriptor of some datasets. For this, the log-likelihood is as follows (Bell, pers. comm.)

$$
\ln(L) = n \ln\left(c\right) - c(t\_n - t\_0) \tag{6}
$$

The best-fit ML parameters for each acceleration model are established by minimizing the negative of Equations (4–6). The observable typically is an acoustic or seismic event count, such that it is a pure number. For this reason, here we do not present the non-cumulative best-fit models graphically because they are not informative—the clarity comes in cumulative form where the data are effectively stacked and elevated into a line with a given curvature. Moreover, as we have the advantage of working with non-binned data, there is no use in looking at a timeline of event timings, which is what non-cumulative data amount to. The cumulative form 3 of the exponential model (corresponding to Equations 2 and 4) is (Bell, pers. comm.)

$$
\Lambda\left(t\right) = \frac{h}{q} \left(\exp\left(qt\right) - \exp\left(qt\_0\right)\right) \tag{7}
$$

That of the TROL (corresponding to Equations 3 and 5) is (Bell et al., 2013):

$$
\Lambda\left(t\right) = \frac{k}{p-1}\left( (t\_c - t)^{1-p} - (t\_c - t\_0)^{1-p} \right) \tag{8}
$$

And that of the constant rate model (corresponding to Equation 6) is:

$$
\Lambda(t) = c(t - t\_0) \tag{9}
$$

However, other metrics can be used such as strain, in the case of a constant driving-pressure scenario, or pressure, in the case of a constant displacement rate scenario. Other metrics that accelerate toward failure may exist. However, here we focus on event number as . The energy of acoustic signals cannot be used in the same way because the ML method relies on the event timings in the log-likelihood function. In what follows we will test each of these approaches against a range of experimental datasets.

#### Describing Heterogeneous Magmas

Magmas may be heterogeneous in texture. While the most important distinguishing features might be identified on a volumetric basis, such as the gas volume fraction (or porosity) φ, or the crystal volume fraction φx, it may be important to understand the spatially defined properties. Examples are the frequency distribution of pore or crystal sizes, the distribution of

inter-pore or inter-crystal distances, and degrees of anisotropy. Here we give the method described in Vasseur et al. (2017).

(Torquato et al., 1990) describe the void nearest-neighbor density function P (R) for a system of random heterogeneous overlapping particles with a characteristic radius R<sup>p</sup> from which a pore-size density function can be derived (Torquato, 2013):

$$\overline{P}\left(\overline{R}\right) = \frac{3\eta\left(1+\overline{R}\right)^2}{\phi} \exp\left(-\eta\left(1+\overline{R}\right)^3\right) \tag{10}$$

where P R = P (R) Rp, R = R/R<sup>p</sup> and η = − ln (φ). The first moment of Equation (10) allows us to compute the characteristic mean pore radius between the particles as follows

$$
\langle \overline{\mathcal{R}} \rangle = \int\_0^\infty \overline{\mathcal{R}} \overline{\mathcal{P}} \left( \overline{\mathcal{R}} \right) d\overline{\mathcal{R}}.\tag{11}
$$

Similarly, a single analytical expression for other metrics such as an inter-pore and an inter-particle distance can be derived from the first moment of a nearest neighbor function (Torquato et al., 1990):

$$\tilde{l}\_i = \frac{\Gamma\left(4/3\right)}{\eta^{1/3}}\tag{12}$$

where Ŵ is the gamma function, and η = − ln (1 − φ) when i = 1 (the inter-pore distance; l<sup>1</sup> = l1/R) and η = − ln (φ) when i = 2 (the inter-particle distance; l<sup>2</sup> = l2/Rp). In porous volcanic rocks, Equations (10–12) result in typical inter-pore lengths 10−<sup>6</sup> < <sup>l</sup><sup>1</sup> <sup>&</sup>lt; <sup>10</sup>−<sup>4</sup> m. We will use this range later in our discussion of data scaling from the laboratory to nature.

#### EXPERIMENTAL METHODS

There is some commonality of technique among deformation testing equipment. First, most tests are performed on cylindrical samples in either uniaxial or triaxial deformation rigs (e.g., Vasseur et al., 2013; Heap et al., 2017). In AE studies of rock or magma fracture and failure, it is common to use piezoelectric transducers. In the case of high-temperature experiments, these can be in contact with the load frame or deformation pistons (Lavallée et al., 2008; Vasseur et al., 2015, 2017) or in direct contact with the sample or sample jacketing system via waveguides (Benson et al., 2008; Tuffen et al., 2008). Waveguides attenuate acoustic signals but do not alter frequency-amplitude ratios (Meredith and Atkinson, 1983).

The data presented herein (and from Vasseur et al., 2015, 2017) were collected using a uniaxial, high temperature, high load press built by Voggenreiter GmbH (Hess et al., 2007; see **Figure 1** for a schematic). A linear variable differential transducer (LVDT) with a 150 mm travel range and a 10−<sup>6</sup> m resolution is used to track displacement of the top piston. The force is monitored with a Lorenz Messtechnik GmbH K11 load cell with a range of 300 kN and an accuracy in either tension or compression of 0.05 % of the measured force. The rates of displacement are wellcontrolled in the range 8.3 × 10−<sup>7</sup> to 1 × 10−<sup>2</sup> m s−<sup>1</sup> . A split 3-zone, 12 kW furnace (GERO GmbH) covers approximately 10 times the length of the sample and both pistons and can heat up to 1,100◦C, accurate to within 2◦C. With appropriate insulation, the stable hot zone is 0.12 m long. At the ends of both pistons, with a direct path through the pistons to the sample,

are piezoelectric AE broadband transducers with 125 kHz central frequency. A 40 dB buffered preamplifier is used to transfer the AE signals to the Richter data acquisition system (Applied Seismology Consultants), recording an AE voltage continuously at 20 MHz sampling rate.

Some samples analyzed herein, and which appear in Vasseur et al. (2017), were deformed in a similar apparatus at the Laboratoire de Déformation des Roches (LDR) at the Université de Strasbourg. This device also used an LVDT transducer to measure displacement and piezoelectric transducers with central frequencies in the 100-1,000 kHz range to monitor AE signals. We refer the reader to Heap et al. (2015) for more details.

For all tests, cylindrical samples of ∼ 10 mm radius and ∼ 40 mm height were cored from blocks of (1) synthetic samples of welded glass beads (originally characterized in Vasseur et al., 2013; see **Figure 2** for example 3D textures), or (2) volcanic rocks from Mt Meager volcano (Canada; Heap et al., 2014, 2015). We define the piston velocity as v = dL/dt and keep this constant during any test. The strain rate in the axial direction is then v/L0, where L<sup>0</sup> is the starting sample length. The samples were deformed at (1) a strain rate of 10−<sup>3</sup> s −1 and a temperature of ∼550◦C (slightly above the glass transition onset in the viscoelastic regime) and (2) a strain rate of 10−<sup>5</sup> s −1 and under room temperature to ensure an elastic response. The temperature ∼550◦C is chosen to give an example condition typical of magma deformation, in which the sample is a relaxed liquid prior to deformation, but is driven to behave in a viscoelastic way by the application of a strain rate that is high compared with the relaxation time. At this temperature, the sample chosen has a viscosity of ∼10<sup>12</sup> Pa s, and a relaxation time of 100 s, making the Deborah number De ≈ 0.1, which is above the critical value to ensure failure will ensue. AE event onsets were triggered and recorded automatically from the continuous acoustic datastreams using an adaptation of an autoregressive-Akaike-Information-Criterion (AR-AIC) picker (Beyreuther et al., 2010; Vasseur et al., 2015). The AR-AIC picker follows this workflow: (i) detection of the onset of a waveform above the baseline using an STA-LTA detector, (ii) de-noising of the acoustic signal, and (iii) AIC computation where the minimum indicates the arrival time. The STA-LTA window was set to 1 and 20 ms, respectively and the STA/LTA threshold was 2. The amplitude in dB and energy (typically in nJ), of each event were computed based on a resistance reference standard of 10 k.

All 42 samples were driven at constant rate as described above, until failure occurred where mechanical failure is defined as the point after which the sample is no longer load-bearing and the force drops to zero. This force-drop is easily picked in each mechanical dataset and provides excellent resolution on the measured t<sup>c</sup> , which can then be compared with the predicted t<sup>p</sup> using predictive tools described in section Quantitative Background.

#### FORECASTING THE FAILURE OF MULTIPHASE MAGMAS

We take a staggered approach to data analysis. First, we consider that effects of material heterogeneity on forecast efficacy can be best determined by using the complete data set of acoustic emissions (section The Effect of Porosity). However, we acknowledge that a true "forecast" would only be useful if it can be made before the final critical failure event has been reached, and therefore using less than the complete dataset. Therefore, in a second step, we analyse the effect of taking an incomplete sequence of data on the efficacy of forecasts (section Hindcasting or Simulated Forecasting).

#### The Effect of Porosity

Using datasets produced for deformation of sintered, variable porosity, variable grainsize, soda-lime silica glass beads (Vasseur et al., 2013, 2015), and natural sintered Mt Meager volcanic rock (Heap et al., 2015; Vasseur et al., 2017), we can apply the techniques described above to test the efficacy of failure forecast tools.

First, if we use 100% of the AE sequence, we can use the log-likelihoods given in Equations (4–6) to fit for the unknown constants in a linear form Equation (6), an exponential form Equation (4) and a power-law form Equation (5). In **Figure 3** we

plot the cumulative AE event number with time for low-to-high porosity samples for both sample types (sintered glass beads and welded volcanic debris from Mt Meager). The data are compared with the three model forms (Equations 7–9).

The power-law (Equation 3) includes t<sup>c</sup> as a fit-parameter, interpreted to be the best-fit modeled failure time (analogous to t<sup>p</sup> described earlier). The value of t<sup>p</sup> is much greater than the observed failure time t<sup>c</sup> when the sample porosity is low (**Figures 3A,E**), creating a substantial time-deficit that equates to a poor predictability. However, as porosity is increased, t<sup>p</sup> systematically approaches t<sup>c</sup> (compare **Figures 3A,E** with **Figures 3D,H**) such that the time deficit is reduced and the potential for predictability is increased.

In the case of the linear and exponential forms, they fit the data better at low porosity than at high porosity. Therefore, it is clear that low-porosity samples do not deform with powerlaw precursory signals and rather the precursory signals follow exponential behavior. Indeed, at very low porosities, the data are almost linear (**Figure 3E**). This leads to the proposition that the power-law behavior in these critical mechanical systems is due to individually unpredictable events bridging gaps between textural flaws. And that the power-law predictability is an emergent property of a complex system, rather than intrinsic to material failure.

#### Hindcasting or Simulated Forecasting

In real forecasting scenarios at volcanoes, the beginning of the precursory sequence may not be detected, and similarly, by definition, a forecast requires that the end of the sequence is incomplete. Here we test these scenarios in which a data sequence may be partially incomplete and how such cases affect the efficacy of forecast times.

First, if we assume that we can rigorously define the beginning of the sequence, such that the initial time is well-defined, then we can test the effect of missing data at the end of the sequence. This is similar to real-time forecasting scenarios in which we might be acquiring new data in real time and adding it to the sequence and at each time-step, the fitting procedure would be repeated using (Equation 5). Examples of single low- and high-porosity data sets are given in **Figures 4A,B** (sintered glass beads) and **Figures 4E,F** (Mt Meager volcanic debris), in which fits to 100, 90, 80, and 70% of the time data are shown (given as fractions of the data sequence 0.7 ≤ f ≤ 1.0). The quality of the fits is similar from f = 1 down to f = f ′ (where f ′ is about 0.8 for the sintered glass beads and 0.9 for the welded volcanic debris), and the error on the forecast time is similar in that window. However, with sequences of less than f = f ′ , the forecast errors become larger for the high porosity samples. This indicates that the forecast efficacy is highly dependent on the amount of the sequence that has occurred, and that this dependence is stronger for high-porosity samples (see **Figures 4C,G**). Indeed, for f < f ′ , the dependence of forecast error on porosity is the inverse of the dependence found for f > f ′ , such that it would appear that high-porosity materials are less well-forecast than low-porosity materials. This also shows that the forecast error for high-porosity materials collapses to near zero as the sequence converges on t → t<sup>c</sup> . Conversely, for low-porosity samples, we note that the inverse trend is observed, albeit with a lower degree: the variability in

(ML-EXP; Equation 7), and linear models (ML-CR; Equation 9) are given. (A–D) are for synthetic samples of variably sintered soda-lime-silica glass beads (Vasseur et al., 2013, 2015). (E–H) are for variably welded natural samples of Mt Meager deposits (Heap et al., 2014, 2015). Each panel represents a different porosity sample with low porosity on the left and high porosity on the right. The gray shaded boxes represent the time difference between the observed failure time (left margin of the box) and the failure time predicted by the extrapolated singularity of the ML-TROL power-law model (right margin of the box), such that the box itself represents the time-deficit in the forecast. All panels contain information about the coefficient of determination r <sup>2</sup> obtained for each fit. Adapted from Vasseur et al. (2017).

variably welded Mt Meager datasets. Panels (A,B,E,F) are the data for two porosities, for which the ML-TROL power-law model is fit using increasing fractions of the data. f = 1 represents the full dataset, while f = 0.7, for example, would indicate that 70% of the total data set, measured from the beginning, has been used in the fitting. Panels (C,G) are the effect of taking increasing fractions of the data on the failure time accuracy. Contrastingly, in panels (D,H), the curves at f < 1 refers to a case when data at the beginning of the sequence is missing. A normalized failure forecast of unity (zero on this log axis), represents a perfect forecast. (A,B,E,F) contain information about the coefficient of determination r <sup>2</sup> obtained for each fit.

the forecast error grows as more and more of the sequence is acquired.

We also check the effect of missing the beginning of the sequence, analogous to missing low-amplitude events at the beginning of a precursory phase of activity at a volcano (especially problematic during long-duration precursory unrest phases; Robertson and Kilburn, 2016). In **Figures 4D,H** we show this effect is relatively independent of porosity and less important than the data accumulating at the end of the sequence.

#### Probability and Accuracy

For 0.7 ≤ f ≤ 1.0, we show in **Figure 5** the effect of taking different proportions of the sequence on the forecast error. A complete sequence f = 1.0 relates to the forecast error for a complete sequence, and therefore represents a limiting case where the entire dataset is known ahead of time. Any reports of the predicted failure time for f = 1.0 are therefore not forecasts and are instead useful for assessing the quality of the functional forms for ˙ (t) that could be used in forecasts. Here we see the strong dependence of the error on the sample porosity, with high-porosity materials being fully predictable with nearzero error. However, at f = 0.7, for which the uncertainty on the signal is higher, we note that the variability in the forecast error for high-porosity materials is larger than for materials with φ < 0.2, for which it becomes easier to forecast failure. We cast these as a Probability Density Function (PDF) of a given forecast error (**Figure 5**). For a given sample, we do this by sweeping over a range of initial guesses (using a reasonable initial value combined with a multiplicative factor varying between 1 and 10 every 0.05) on the fitted parameters in Equation (5), performing fits and computing the distribution of fitted forecast errors. The distribution is then converted to a PDF weighted by the coefficients of determination obtained from the fitting procedure. Displayed in **Figure 5** is thus the intensity of the PDF obtained for each sample as a color map. The points represent the results obtained from using a single reasonable initial guess for each parameter and do not necessarily coincide with the most probable outcome.

#### SIMPLE SCALING FROM THE LAB TO THE FIELD

Across all values of porosity φ, the frequency F of the acoustic events in the laboratory-scale experiments detailed here ranged between F = 3 × 10<sup>4</sup> and F = 1 × 10<sup>6</sup> Hz. In **Figure 6b** we see that an additional complexity associated with high-porosity samples lies in the clear discrete onset being slightly masked compared with **Figure 6a** because the coda from the previous waveform overlaps with the onset of the new waveform.

In other laboratory set-ups, events at much lower frequencies are detected; for example in Benson et al. (2008) and Tuffen et al. (2008), events as low as F = 10<sup>4</sup> Hz, are found to be associated with pore fluid movement associated with sudden fracture propagation. This is only possible in pore-pressure controlled, jacketed triaxial experiments.

The scaling ratio most commonly deployed compares the product of a fracture lengthscale L and event frequency F at scale 1 to the same product at scale 2. This assumes that the fracture lengthscale is associated with the event that produced the signal

frequency. If we use subscripts to denote the two scales, then this relation is L1F<sup>1</sup> = L2F<sup>2</sup> (Aki and Richards, 2002; Burlini et al., 2007). If scale 2 is the volcano scale, and scale 1 is the laboratory scale, then we can most easily place constraints on L1, F1, and F2, and use these to predict L2. If we stick to order-of-magnitude analysis, as shown above, 10<sup>4</sup> ≤ F<sup>1</sup> ≤ 10<sup>6</sup> Hz and does not appear to depend on φ. We might expect that L<sup>1</sup> depends on φ and is the inter-pore length given by Equation (12). In a porous system, such as the sintered system used herein, we see that L<sup>1</sup> depends on the grainsize R. In natural sintered systems in volcanic environments, the grainsize is typically 10−<sup>5</sup> < R < 10−<sup>3</sup> m (Saubin et al., 2016). In turn, across the full range of φ from the initial packing φ down to low sintered φ > 0.03, using (Equation 12), we find that 10−<sup>6</sup> <sup>≤</sup> <sup>L</sup><sup>1</sup> <sup>≤</sup> <sup>10</sup>−<sup>4</sup> m (see section Describing Heterogeneous Magmas). Finally, we know that VT events at volcanoes are typically 1 <sup>≤</sup> <sup>F</sup><sup>2</sup> <sup>≤</sup> 10 Hz. This renders 10−<sup>3</sup> <sup>&</sup>lt; L<sup>2</sup> < 10<sup>2</sup> m and gives insight into the fracture lengthscales between flaws on the volcano scale and is consistent with the pervasive fracture system lengthscales expected in some of the source-mechanism models for seismogenic eruptions (Neuberg et al., 2006). This also implies that while fracture lengths in the laboratory are typically related to the flaw or maximally, the sample lengths, at the volcano scale these would be much larger on the millimetric to hundred-meter scale. We work on the assumption that low-frequency magma-fracture events are damped events of an original VT-frequency content, congruent with the model of low-frequency events as magma rupture events (Neuberg et al., 2006; Thomas and Neuberg, 2012; Salvage and Neuberg, 2016).

A key difference between the laboratory cases presented here and natural cases is that laboratory experiments of this kind tend to be performed at a constant strain rate, allowing the stress to evolve in response, until failure. However, in nature, the magmatic conduit system may be more likely to be in a state of variable local strain rate (e.g., constant pressure at the conduit base or constant flux; c.f. Gonnermann and Manga, 2003). Future research should aim to explore scaling from laboratory to nature across a wide range of conditions and we identify this as a frontier topic.

#### CONCLUSIONS

We show that it is the heterogeneity of the system that most effects the efficacy of forecasts of material failure. Given this insight, we have presented the simplest scaling from the laboratory to the natural case on the basis of the relationship between rupture lengthscale and radiated frequency. On the laboratory scale, it is the inter-pore lengthscales that fail in each individual acoustic event, which leads to larger scale failure at the critical time. By scaling, we see that these events and the associated frequencies would be equivalent to seismic events at volcanoes with much larger rupture lengthscales. However, independent tests of the rupture lengthscales at active volcanoes are poorly known and would represent fruitful future work.

We explore the effect of having an incomplete dataset during a deformation episode. We find that the error on an attempted critical time forecast is substantially affected by missing data at the end of the sequence. The implication is that in any

#### REFERENCES


real-time scenario, the efficacy of the forecast will improve as the critical time approaches, especially for highly heterogeneous systems. Poor constraint on when the deformation episode began, however, is less important for effective forecasting.

We identify the specifics of scaling heterogeneities from the laboratory to nature as a frontier topic in need of attention. We propose that experimental work at larger scales could be used to validate the scale independence of forecast efficacies in highly heterogeneous systems, and explore the effect of system size on the forecasts possible in homogeneous systems. The ability to scale from laboratory findings to real crises in nature is critical.

#### AUTHOR CONTRIBUTIONS

JV performed the experiments and processed the data. JV and FW conceptualized the study and analyzed the data. DD supervised the analysis. All authors contributed to the manuscript.

#### ACKNOWLEDGMENTS

Thanks to three reviewers for thoughtful comments, and to Yan Lavallée, Ian G. Main, Andrew F. Bell, Caron Vossen, and Taylor Witcher for interesting discussions about forecasting rupture of geomaterials. We thank Mathieu Colombier for technical assistance with 3D data-rendering.


vein textures from the 2008–2009 Chaitén Eruption. Front. Earth Sci. 4:59. doi: 10.3389/feart.2016.00059


**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.

The reviewer, MF, declared a past co-authorship with one of the authors, DD, to the handling editor.

Copyright © 2018 Vasseur, Wadsworth and Dingwell. 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.

## Short-Term Forecasting and Detection of Explosions During the 2016–2017 Eruption of Bogoslof Volcano, Alaska

Michelle L. Coombs<sup>1</sup> \*, Aaron G. Wech<sup>1</sup> , Matthew M. Haney<sup>1</sup> , John J. Lyons<sup>1</sup> , David J. Schneider<sup>1</sup> , Hans F. Schwaiger<sup>1</sup> , Kristi L. Wallace<sup>1</sup> , David Fee<sup>2</sup> , Jeff T. Freymueller<sup>2</sup> , Janet R. Schaefer<sup>3</sup> and Gabrielle Tepp<sup>1</sup>

#### Edited by:

Nicolas Fournier, GNS Science, New Zealand

#### Reviewed by:

Stephen R. McNutt, University of South Florida, United States Jan Marie Lindsay, University of Auckland, New Zealand Silvio De Angelis, University of Liverpool, United Kingdom

\*Correspondence:

Michelle L. Coombs mcoombs@usgs.gov

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 19 March 2018 Accepted: 10 August 2018 Published: 03 September 2018

#### Citation:

Coombs ML, Wech AG, Haney MM, Lyons JJ, Schneider DJ, Schwaiger HF, Wallace KL, Fee D, Freymueller JT, Schaefer JR and Tepp G (2018) Short-Term Forecasting and Detection of Explosions During the 2016–2017 Eruption of Bogoslof Volcano, Alaska. Front. Earth Sci. 6:122. doi: 10.3389/feart.2018.00122 <sup>1</sup> Alaska Volcano Observatory, Volcano Science Center, United States Geological Survey, Anchorage, AK, United States, <sup>2</sup> Alaska Volcano Observatory, Geophysical Institute, University of Alaska Fairbanks, Fairbanks, AK, United States, <sup>3</sup> Alaska Volcano Observatory, Alaska Division of Geological and Geophysical Surveys, Fairbanks, AK, United States

We describe a multidisciplinary approach to forecast, rapidly detect, and characterize explosive events during the 2016–2017 eruption of Bogoslof volcano, a back-arc shallow submarine volcano in Alaska's Aleutian arc. The eruptive sequence began in December 2016 and included about 70 discrete explosive events. Because the volcano has no local monitoring stations, we used distant stations on the nearest volcanoes, Okmok (54 km) and Makushin (72 km), combined with regional infrasound sensors and lightning detection from the Worldwide Lightning Location Network (WWLLN). Preeruptive seismicity was detected for 12 events during the first half of the eruption; for all other events co-eruptive signals allowed for detection only. Monitoring of activity used a combination of scheduled checks combined with automated alarms. Alarms triggered on real-time data included real-time seismic amplitude measurement (RSAM); infrasound from several arrays, the closest being on Okmok; and lightning strokes detected from WWLLN within a 20-km radius of the volcano. During periods of unrest, a multidisciplinary response team of four people fulfilled specific roles to evaluate geophysical and remote-sensing data, run event-specific ash-cloud dispersion models, ensure interagency coordination, and develop and distribute of formalized warning products. Using this approach, for events that produced ash clouds ≥7.5 km above sea level, Alaska Volcano Observatory (AVO) called emergency response partners 15 min, and issued written notices 30 min, after event onset (mean times). Factors that affect timeliness of written warnings include event size and number of data streams available; bigger events and more data both decrease uncertainty and allow for faster warnings. In remote areas where airborne ash is the primary hazard, the approach used at Bogoslof is an effective strategy for hazard mitigation.

Keywords: eruption forecasting, Alaska, volcano monitoring, Bogoslof, volcanic infrasound, volcanic lightning, volcano seismology, hazard communication

#### INTRODUCTION

feart-06-00122 August 30, 2018 Time: 17:5 # 2

Eruption forecasting can include both long-term forecasting, which provides an overall probability of eruption at a given volcano or region over a time period of years using geologic and historical records, as well as short-term forecasting, which estimates the probability, timing, and magnitude of an impending eruption at a restless volcano (Marzocchi and Bebbington, 2012). The latter relies heavily on local instrumentation and on the interpretation and analysis of real-time or near real-time monitoring data from a volcano (Sparks, 2003). Ideally, successful short-term forecasting can allow volcano observatories to issue warnings of unrest and the possibility of a volcanic eruption hours to weeks in advance.

The Alaska Volcano Observatory (AVO) monitors volcanoes in Alaska and issues notifications and warnings of volcanic unrest and eruption. Of the over 100 volcanoes in Alaska that have been active in the Holocene, only 32 currently have geophysical monitoring networks, making short-term forecasts of volcanic activity extremely challenging (Cameron et al., 2018). In 2016, unmonitored Bogoslof volcano (**Figure 1**) began a 9-month-long eruptive sequence that included at least 70 explosions, each minutes to tens of minutes long, that sent ash clouds as high as 14 km above sea level (**Figure 2** and **Table 1**). AVO was not able to forecast the beginning of the eruption; retrospective analysis shows that we missed at least four explosions in December of 2016 before being a pilot reported that the volcano was erupting. Immediately after receiving notification of the ongoing eruption, AVO implemented ad hoc, near real-time procedures to detect and forecast future explosive events. The new workflow exploited data from distant (within 100 km) seismic stations and other geophysical data streams.

Because Bogoslof is remote and uninhabited, like many Alaskan volcanoes, the main hazards associated with the 2016– 2017 eruption were from airborne ash with potential impacts to regional and trans-Pacific aircraft and ashfall on moderately distant communities and ships navigating along local routes. Similarly to other volcanic crises, during the period of eruption at Bogoslof aviation and civil authorities require answers to some basic, but crucial, questions: When? How long? How big? Where will the ash cloud go, and will ash fall on local communities? Therefore, a timely and coordinated response, ideally providing accurate estimates of atmospheric volcanic ash transport and dispersal, is critical. AVO coordinated the determination of factors such as timing, cloud altitude and dispersal direction with the National Weather Service (NWS) Anchorage Volcanic Ash Advisory Center (VAAC), which issues volcanic ash warnings and forecasts to the aviation industry within the Alaska Flight Information Region. AVO also provides guidance about ashfall to the NWS Anchorage Forecast Office, which issues ashfall statements, advisories, and warnings for the public on the ground and marine communities.

This paper describes the geophysical data streams used to evaluate unrest, as well as the protocols to communicate information about volcanic activity and hazards during explosive activity at Bogoslof in 2016 and 2017. We focus on short-term forecasting in the hours to minutes prior to discrete explosive events, detection as soon as possible after onset (typically within minutes), and characterization in the minutes to tens of minutes after event onset (**Figure 3**). We highlight the combined use of a variety of automated alarms on seismic, infrasound, lightning, and remote sensing data that allowed us to respond to the sequence without 24/7 staffing at the observatory. We also present the timing of our information products relative to event onsets, and analyze the factors that can improve timeliness of warnings.

#### The 2016–2017 Eruption of Bogoslof Volcano

Bogoslof Island sits north of the Aleutian volcanic arc, about 100 km west of Unalaska/Dutch Harbor (**Figure 1**). It is the tip of a mostly submerged back-arc volcano that had last erupted in 1992, one of at least eight historical eruptions documented at Bogoslof (Waythomas and Cameron, 2018). The 1992 eruption lasted about 3 weeks, produced episodic ash emissions up to 8 km asl, and ended with extrusion of a lava dome (McGimsey et al., 1995). Previous historical eruptions lasted months to years, and were characterized by intermittent explosive and effusive activity (Waythomas and Cameron, 2018). Erupted compositions range from basalt through trachyandesite (Miller et al., 1998).

The most recent eruption of Bogoslof began in mid-December 2016. Between December 2016 and August 2017, activity at Bogoslof was dominated by a series of at least 70 explosive events that lasted minutes to tens of minutes, and lofted volcanic clouds as high as 14 km asl (**Figure 2**). During the first 2 months, AVO detected 30 such events, occurring every 1–4 days. The pace of explosions slowed in early February, and an eruptive pause from mid-March to mid-May suggested that the sequence may have ended. Activity resumed on May 17 with a series of explosive events and the first observed subaerial lava dome of the sequence. This dome was first observed on June 5 and subsequently destroyed by an explosion on June 10. In mid-August, a second lava dome formed, which was destroyed by the time of the final explosive event on August 30. This marked the apparent end of the eruption, as hot ground and water in the vent area slowly cooled, and the volcano returned to a quiescent state by the end of 2017.

For the largest part of the eruptive period, Bogoslof's vent was submerged in shallow seawater probably less than 100 m deep, though on several occasions a subaerial edifice grew and the vent migrated above sea level. Most volcanic clouds drifted north over the Bering Sea, but three events produced ashfall on nearby communities and mariners east and south of Bogoslof (January 31, March 8, and May 17). The eruption sequence resulted in dozens of regional flight cancelations and flight diversions around the volcano<sup>1</sup> .

<sup>1</sup>https://avo.alaska.edu/volcanoes/activity.php?volcname=Bogoslof&page= impact&eruptionid=1301

#### MATERIALS AND METHODS

In the following section, we describe the classification scheme used to identify explosions, the monitoring data used in realtime (no latency) or near-real-time (latency of up to tens of minutes) to detect and characterize the events, and the protocols that were developed by AVO during the eruption with regards to internal and external communications and warning products. Other data streams, such as high-resolution satellite imagery and SO<sup>2</sup> measurements from satellite, along with petrologic analyses of eruptive products, reveal much about the eruption but did not play a role in the shortterm forecasting or detection and thus are not discussed here.

#### Explosive Event Onset and Classification

Following AVO routine practices during volcanic crises (e.g., Coombs et al., 2010; Bull and Buurman, 2013), explosions were assigned sequential numbers (**Table 1**). The onset of each event was defined using a combination of seismic and infrasound data. Whereas infrasound is a more reliable indicator that material was injected into the atmosphere (Fee and Matoza, 2013), this data stream was not always available due to wind noise and/or prevailing wind directions (typically more northward in the winter months of December through February) that can carry infrasound signals away from sensors (**Figure 1**). For events for which infrasound data were not available, the onset of co-eruptive tremor was used as event onset time.

TABLE 1 | Explosive events during the 2016–2017 eruption of Bogoslof.


(Continued)

#### TABLE 1 | Continued

feart-06-00122 August 30, 2018 Time: 17:5 # 5


VAN, Volcanic Activity Notice; nd, not determined; na, not applicable. Cloud top heights are taken from Volcanic Activity Notices issued at time of event and may change upon reanalysis. 52 not issued because AVO had gone to RED shortly before for 51; status report issued after. 47, 64, 65 not issued because small events. 48 had a brief infrasound pulse 78 min prior to confirmed eruptive activity.


FIGURE 3 | Generalized timeline illustrating different components of short-term volcano forecasting. In this paper, we focus on the events that occur within the hours and minutes just prior to and after the onset of an explosive event (dashed box): very short-term eruption forecasting, event detection, and event and ash-cloud characterization.

#### Seismicity

feart-06-00122 August 30, 2018 Time: 17:5 # 6

Because of its small size and wilderness designation, Bogoslof is not monitored by a local, on-island geophysical network. In the absence of a local network, AVO used seismic sensors from Okmok (∼50 km) and Makushin (∼72 km) volcanoes on neighboring Umnak and Unalaska Islands (**Figure 1**) to monitor seismic activity associated with the Bogoslof eruption. Storms are common in the Aleutians, especially during the winter months, and seismic signals were often masked by wind noise. Furthermore, the relatively large distances between the active vent and the closest seismic stations meant that only the more energetic explosions were detected. The interpretation of data was also complicated by the submarine nature of the eruption. Seismograms recorded body (P and S) waves as well as energy that was transmitted acoustically through the water column before coupling back into the solid Earth (T waves; Okal, 2008), a path-dependent process that manifests differently at different stations. Finally, tectonic tremor is common in the region (Li and Ghosh, 2017) and was sometimes mistaken for co-eruptive tremor.

Explosive events were characterized by minutes to tens of minutes of co-eruptive seismic tremor on the neighboring island networks. Explosive events during the first few months of activity often exhibited precursory seismicity as well, which allowed AVO to issue warnings prior to event onset (events with negative latency for calls or written notices; **Table 1**). Precursory seismicity primarily consisted of repeating earthquakes, which would become more closely spaced in time over a period of hours, culminating in eruption (e.g., **Figure 4A**). Such events, commonly observed at volcanoes worldwide, are often considered a sign that an explosion may be imminent (e.g., Malone et al., 1983; Powell and Neuberg, 2003; Hotovec et al., 2013). Other explosive events during the Bogoslof eruption may have been preceded by similar precursory seismicity that was not detected by our distant networks. For these, the onset of co-eruptive tremor marked the first seismic indication of unrest for a particular event. Retrospective analysis of data from a campaign hydrophone, deployed in May 2017 near the submarine base of the Bogoslof cone, does indeed show that later events were preceded by seismicity too weak to be detected at the Okmok and Makushin stations.

#### Infrasound

Although intense seismic tremor is often strongly suggestive of explosive eruptive activity, the atmospheric pressure oscillations produced by violently expanding volcanic gases and recorded on low frequency acoustic (infrasound) sensors unambiguously confirm that explosive activity is occurring (Fee and Matoza, 2013). AVO operates multiple infrasound sensors or arrays along the Aleutian Arc in order to detect volcanic activity and constrain an azimuth to the source (**Figure 1**). Explosion infrasound was recorded at all AVO arrays over the course of the Bogoslof eruption, including stations more than 800 km from the volcano. The array closest to Bogoslof is located on Okmok volcano (59 km), and was used most frequently for monitoring because it detected a larger number of explosions with greater amplitude and lower latency (∼3 min) than the more distant arrays. As with seismic data, wind noise can mask explosion signals in infrasound data, and seasonal changes in the predominant tropospheric and stratospheric wind directions can affect infrasound propagation and detection at regional distances. These factors resulted in no single AVO array detecting all of the explosive events at Bogoslof.

#### Lightning

Volcanic eruption columns and drifting clouds from explosive eruptions often produce lightning (Behnke and McNutt, 2014), and lightning detection played an important role in volcano monitoring efforts during this eruption. The World Wide Lightning Location Network (WWLLN<sup>2</sup> ) provided near-realtime automated alerts within minutes of lightning strokes near Bogoslof, detecting lightning from 26 of the 62 events (∼40%). Detections typically occurred within minutes after initiation of the explosive seismic signal, and lasted minutes to tens of minutes. Global lightning networks only capture the most energetic lightning, with WWLLN detecting >50% of all strokes above 40 kA peak current, and only 10–30% of the weaker strokes (Hutchins et al., 2012). Despite this limitation, WWLLN provided important confirmation that significant explosive activity had occurred. Volcanic lightning can be generated by a variety of processes in the ash column and downwind cloud (Behnke et al., 2013; Van Eaton et al., 2016). The timing, location and intensity of the lightning is likely related to a number of factors, including the eruption rate, amount of water in the plume (liquid and ice), and atmospheric temperature gradient. In some instances, AVO also detected volcanic thunder, a previously undocumented phenomenon, in conjunction with lightning on nearby infrasound sensors (Haney et al., 2018).

#### Satellite Remote Sensing

Alaska Volcano Observatory uses a variety of near-real-time satellite data to monitor volcanic unrest, detect eruptive activity, characterize eruption style, and track drifting volcanic clouds. Data from the AVHRR, MODIS, and VIIRS sensors aboard polarorbiting satellites, and from the geostationary GOES-15 and Himiwari-8 satellites were used during the Bogoslof eruption. Visible, shortwave-infrared and thermal-infrared data from these operational satellites were used. Spatial resolution ranges from 375 m for VIIRS, 1 km for AVHRR and MODIS, to more than 8 km for geostationary data. Geostationary data are generally available at 15-min intervals, with a typical data latency of about 20–45 min after collection. Polar-orbiting satellites have higher spatial resolution, but data are available less frequently. Images are typically available within 15–20 min of data collection, but due to orbital constraints there are gaps of about 4 h that occur twice daily (middle of the day and middle of the night local time).

Once an explosive event was detected in seismic, infrasound, and possibly lightning data, we used near-real-time satellite data to determine whether a significant volcanic cloud had been generated, to estimate its altitude, and to track its dispersion.

<sup>2</sup>http://wwlln.net/

infrasound from Okmok (gray line), timing of alarms and data receipt (colored symbols), and timing of warnings issued (colored vertical lines). (A) Event 36, February 19, 2017. This event is an example where precursory data allowed us to issue warnings prior to explosive event onset. (B) Event 39, May 16, 2017. This event was detected in multiple real-time data streams and was an example of a significant ash-producing event that was fairly easy to characterize. (C) Event 14 on January 2, 2017, is an example of a short explosion that was detected in seismic and infrasound data only. The initial call to FAA was fairly rapid, but lack of data resulted in difficulty characterizing the event's magnitude and led to a delay in issuing a written warning.

Height estimates were made primarily by using the satellitederived cloud top temperature and comparing it the atmospheric temperature profile determined from the Global Forecast System data. Bogoslof clouds rose to altitudes of 3 to ∼14 km above sea level, and were often discernible in satellite images for hours after an event.

As is common for explosive eruptions that occur in oceanic, lacustrine, or glacial settings (Mastin and Witter, 2000), Bogoslof produced volcanic clouds that show evidence for entrainment of large amounts of water from the vent region. Eye-witness and satellite observations of the clouds indicate that they were darker at the base, due to ash content, but the upper, higher parts of the cloud were frequently white and ice-rich. These distinctive characteristics affected cloud properties in satellite images, fallout and dispersion, and generation of lightning.

One result of incorporation of seawater into the eruptive column is that the widely used thermal-IR brightness temperature difference technique (Prata, 1989) is poorly suited for discriminating volcanic ash in these clouds. This is likely due to ice formation on ash particles, changing the spectral properties of the cloud. Three explosive events that showed no ash signature in satellite data produced documented ashfall on land (the others dispersed over the ocean and remote islands), supporting the hypothesis that satellite-based discrimination of volcanic ash was masked by ice formation, such as was seen in the 1994 eruption of Rabaul (Rose et al., 1995). Because the typical ash signature was mostly lacking, we identified volcanic clouds during event response primarily by their sudden onset, growth, temperature (i.e., altitude), and location over the volcano.

#### Pilot and Observer Reports

feart-06-00122 August 30, 2018 Time: 17:5 # 8

Pilot reports (PIREPs) or other observer reports sometimes provided details on the eruption including cloud height, dispersal patterns, and simple confirmation of activity during times when satellite views were obscured by cloud cover or no imagery was available. It was initially a PIREP on December 20 that alerted AVO to the Bogoslof eruption. The Federal Aviation Administration (FAA) Anchorage Air Route Traffic Control Center (ARTCC) collects and disseminates all PIREPs, including those that describe volcanic activity, in a database that is accessible by AVO staff for use as notification/verification of eruptive activity. In the event of significant volcanic activity, the FAA, or the co-located NWS Center Weather Service Unit (CWSU), will call AVO directly and conversely, AVO may contact the FAA or CWSU to verify or solicit PIREPS during events. During the Bogoslof sequence, AVO reviewed 84 PIREPS during 27 of the approximately 70 explosions.

In addition to PIREPs, AVO received several mariner or citizen reports that proved useful to verifying and characterizing activity. These came via phone, email, or social media. During event 36 on February 19, AVO personnel were on Unalaska Island and observed the eruptive cloud directly.

#### Alarms and Alerts

Alaska Volcano Observatory is not typically staffed 24/7, which remained the case throughout the Bogoslof eruption. As a result, automated alarms based on geophysical and remote sensing data played a critical role in the eruption response. Within an hour of learning of the eruption, AVO implemented an automated alarm based on regional infrasound data to detect explosion pressure waves from Bogoslof. Seismic data, two additional infrasound arrays, and lightning data were all added to the alarm workflow over the next 48 h. Infrasound array data were processed every minute, with algorithms looking for waveform characteristics consistent with an acoustic wavefield propagating from the direction of Bogoslof during the previous 3 min. Similarly, real-time seismic amplitude measurements (RSAM; Endo and Murray, 1991) were computed for seismic data from neighboring islands every 5 min, and alerts were sent out when enough stations exceeded a designated amplitude threshold. In response to activity at Bogoslof, AVO changed how it processed nearreal-time WWLLN lightning data. Previous alerts were passively received via email from WWLLN, and AVO developed a method to actively download the latest data every minute and push alerts to a response team via text message. An algorithm for detecting repeating earthquake sequences was added in late March (Tepp, 2018) to help identify precursory earthquake swarms. AVO also relied on airwave detection alarms from a more distant infrasound array 825 km away in Dillingham (**Figure 1**). This alert, while less useful for rapid detection, often provided valuable corroborating evidence of emissions into the atmosphere.

All AVO-based alarms were both developed and implemented internally by AVO research staff. Algorithms were initially written in MATLAB programming language and eventually converted to Python code. With a couple of exceptions, alarms are centrally managed on a dedicated alarms server, which handles scheduling, data processing and message dissemination. Alerts are delivered to the AVO's internal chat tool (named AVO Chat; using commercial platform Mattermost) for observatory-wide access, as well as via text message to recipients included in a centrally managed distribution list, which changes based on staffing and duty rotations. Algorithms processing seismic and infrasound signals also generate images of recent data, which are included in the text messages to allow recipients to rapidly determine whether the alert represents a true or false positive.

Critical to this alarming strategy and AVO's ability to depend on alarm functionality is a method for ensuring that the alarm system is working. Daily test messages are sent to ensure operability. AVO also uses Icinga, an open-source network monitoring application, to monitor the individual alarm modules themselves and, effectively, alarm the alarms. Upon completion (every 1 or 5 min) and regardless of detection, each alarm algorithm sends a "heartbeat" message to Icinga, which resides on a separate computer system. Using a separate messaging system, Icinga then sends text messages to system managers if a certain number of heartbeats are missed. This approach provides the robustness and assurance required for AVO to rely on geophysical alarms for event detection and has successfully alerted AVO staff of system failures on multiple occasions.

In addition to alerts developed and/or distributed by AVO, alerts from the National Oceanic and Atmospheric Administration's Cooperative Institute for Meteorological Satellite Studies (NOAA-NESDIS/CIMSS) VOLcanic Cloud Analysis Toolkit (VOLCAT) were used by AVO throughout the eruption<sup>3</sup> . The VOLCAT system autonomously generates alerts of explosive activity worldwide using operational satellite data, identifies volcanic cloud objects, and retrieves cloud properties (height, column mass loading, and effective radii) of those objects. This system uses several different algorithms to identify volcanic cloud objects, but the most useful one during the Bogoslof eruption used anomalous cloud vertical growth rates as observed by geostationary satellites. SMS text, email, and web products were received by AVO staff within 15 min of satellite data collection, and provided a rapid estimate of volcanic cloud top altitude.

#### Ash Dispersion Models

The potential for ashfall on local communities or mariners depends on wind direction, eruption intensity, cloud height, and the mass of ash that is produced during each explosive event. The United States Geological Survey (USGS) provides forecasts of expected ash dispersion (ash clouds) and deposition (ashfall) from volcanic eruptions using the numerical atmospheric transport model Ash3D (Schwaiger et al., 2012). AVO uses Ash3D model outputs to predict ashfall and ash cloud information based on either hypothetical or actual eruption information (see below). The NWS Anchorage Forecast Office then issues ashfall statements, advisories, and warnings for the public and marine communities.

The 2016–2017 Bogoslof eruption was the first for which Ash3D was used commonly in response mode. AVO runs

<sup>3</sup>http://volcano.ssec.wisc.edu/

hypothetical simulations twice per day for each of the volcanoes at elevated color code. These include separate simulations for the anticipated proximal fallout as well as a regional forecast of the drifting volcanic cloud. Results from these hypothetical forecasts are posted to the AVO public website along with output from similar models (puff and hysplit). These are clearly labeled as 'hypothetical' and reflect the style of eruption that was observed in the past at a given volcano in terms of plume height and eruption duration.

During the Bogoslof eruption, when an explosive event was confirmed, we initiated an event-specific run with the event start time, duration, and cloud height. The satellite-derived cloud altitude was used to initialize the Ash3D model to forecast ash fallout on nearby communities; these forecasts were presented on simple maps (**Figure 5**). These event-specific simulations were initially run with only a known start time and estimated plume height and duration, but then run iteratively as new information became available. The most recent model output reflecting the best estimate of volcanic cloud and fallout hazards was posted on the AVO public website with a prominent note indicating that these output results correspond to an actual eruption. Output results from the actual events remained on the public website for at least 12 h before reverting to the hypothetical simulations.

#### Intra Observatory Roles and Communication

Alaska Volcano Observatory has established protocols for duty roles during routine operations and during eruptions. These were modified to adapt to the Bogoslof eruption in several ways. Generally, duty staff consist of the Scientist-in-Charge, Duty Scientist, Duty Remote Sensor, and Duty Seismologist. In addition, on-call staff are responsible for maintaining web servers and data acquisition systems. Duty roles typically rotate on a weekly basis through a group of 8–10 staff members. Duty science staff perform routine data checks and, when necessary, respond to activity.

During the Bogoslof eruption, it was necessary to develop a second, sometimes overlapping group to respond specifically to Bogoslof events. Called the Primary Response Team, this group consisted of a Response Geophysicist, who received alarms and analyzed infrasound and seismic data; a Response Remote Sensor, who received alarms, analyzed satellite data, and acted as primary liaison with the Anchorage VAAC; an Ash3D specialist, who ran event-specific ash dispersion models and, when necessary, was the primary liaison with the NWS Forecast office responsible for ashfall forecasts; and the Duty Scientist, who integrated all data streams and wrote and distributed formal warning products (typically, Volcanic Activity Notices, see below).

This team was assigned weekly, and team members were on call for that week. Much of the activity took place outside of normal working hours, so communication amongst the team initially occurred via phone and often text messaging. During the eruption, AVO implemented use of an internal chat tool, AVO Chat, accessible on computer or mobile device, which allowed the team to communicate easily while also allowing other observatory staff to remain aware by following the discussion. More formal communication took place over AVO's internal log system, where staff document event summaries, and post other more in-depth analyses.

#### External Communication

Internal communication is necessary to bring interdisciplinary data and expertise together to make informed assessments of volcanic activity and hazards, but this information must then be communicated quickly and clearly to interagency partners and the public. A main focus of this study is determining how effectively AVO was able to do this during the Bogoslof eruption. The protocols for external communication during ash-producing eruptions in Alaska is formalized in the Interagency Operating Plan for Volcanic Ash episodes and described in detail elsewhere (Neal et al., 2010). Below, we briefly summarize the USGS Alert Levels and Color Codes, formal Information Products, and calldown procedures. We then describe how formal policy was modified during the Bogoslof eruption to accommodate the high pace of activity and paucity of monitoring data.

#### Official Warning Products

United States Volcano Observatories utilize a dual system of alerts: an Aviation Color Code to address aviation hazards (Guffanti and Miller, 2013), and a Volcano Alert Level to indicate the overall hazard at a volcano (Gardner and Guffanti, 2006). Changing Aviation Color Codes (GREEN, YELLOW, ORANGE, and RED) and Volcano Alert Levels (NORMAL, ADVISORY, WATCH, and WARNING) indicate increasing severity and likelihood of potential impacts. Unmonitored volcanoes, like Bogoslof, are designated as Unassigned if they are at apparent background levels of activity (not GREEN); when they exhibit unrest, however, elevated color codes and alert levels may be assigned as activity warrants. During the eruption of Bogoslof and other remote volcanoes, the Aviation Color Code and Volcano Alert Levels are almost always coupled (for example, ADVISORY/YELLOW). In this paper, we refer only to the Aviation Color Code for brevity.

In conjunction with the alert systems described above, AVO and other United States observatories issue a number of formal warning products to notify the public and other partner agencies of volcano hazards or other important information. All messages are posted on the AVO website, pushed out via email to key partners, and also freely available to anyone via email by subscribing to the Volcano Notification System (VNS)<sup>4</sup> . All formal notifications are issued via the web-based USGS HAzard Notification System (HANS). HANS facilitates rapid dissemination of information by providing database- and web-form-driven formatted notifications preset with headers and footers, volcano information (ID, location, elevation, existing color codes), issuance time, and other guides (e.g., summary of activity, cloud height, recent observations) allowing duty staff to quickly create and release notifications.

Event-specific messages include the Volcanic Activity Notice (VAN), which we issue to announce alert-level changes or

<sup>4</sup>http://volcanoes.usgs.gov/vns/

significant volcanic activity. Additional VANs are released as needed, depending on changes in volcanic activity, alert levels, or hazards. VANs also are used to declare the 'all clear' when an eruption is waning or has ceased. The Volcano Observatory Notice for Aviation (VONA) is a derivative product of the VAN and contains information emphasizing ash emission hazards in a format specifically intended for aviation users (pilots, dispatchers, air-traffic managers, meteorologists).

Alaska Volcano Observatory typically issues a Current Status Report to provide an update about volcanic behavior or monitoring activities during ongoing events of unrest or eruption. A status report may be issued multiple times in a single day. Finally, AVO issues Information Statements that announce topical information such as new monitored volcanoes, significant operational or monitoring capacity changes, ash resuspension, explanation of non-volcanic events at a volcano, and expanded descriptions of volcanic unrest and likely outcomes.

During the Bogoslof eruption, AVO developed a protocol for information products to be released for each explosion to speed up decision making during an event and release information as quickly as possible. As appropriate, we would release some or all of the following categories of VAN:


the event, its impacts, and state future actions on the part of AVO.

(5) Within 24 h of the explosive event, AVO would typically lower the color code from RED to ORANGE (done via another VAN), if the event had prompted the color code to be raised to RED. The Aviation color code remained at ORANGE for most of the eruptive sequence (**Figure 2**).

To assess the timeliness of these formal notification products, we show the time of the first such VAN, as determined by the automated time stamp provided in the HANS system, with respect to each individual explosive event (**Table 1**). Depending on the presence or absence of precursory signals, this first VAN would either be of the "imminent" or "initial" variety, as described above.

#### Call Downs

Upon determination of a significant change in the status of a volcano, whether increased likelihood of eruption, detection of eruption, change in eruption status, end of eruption, or color code change, AVO initiates a formal call down. First on the call-down list is the Federal Aviation Administration (FAA) Air Traffic Control Facility, followed by NWS offices, and other state and federal agencies. Call-down messages are brief and include the following general information: name of caller; volcano name and location; nature of activity and source of information (seismicity, PIREP, etc.); Aviation Color Code and Volcano Alert Level status or change in status; start and stop time of event or activity (if known); height of eruption cloud, how determined, and direction of cloud motion (if known). When significant unrest or activity is detected, AVO will make "heads up" calls to the FAA and NWS offices prior to the official call down (e.g., **Figure 4**).

Duty personnel sometimes record call down times on a written sheet, and often record the time that the entire call down was completed in AVO's internal log system. Because the entire call down can take several minutes to complete, we wanted to investigate the time that the initial call to the FAA was conducted for each event. Using cell phone records of AVO duty personnel, we have determined the timing of the first call to the FAA either immediately prior to, or immediately after, each explosive event (**Table 1**). The time between the event onset and this call is defined as the call-down latency and is separate from the latency of the written information product described above.

#### Integration With Other Agencies

The responsibility of providing notifications about volcanic ash is distributed among several agencies in Alaska. AVO and its partners have created an Interagency Operating Plan—an overview of an integrated, multi-agency response to the threat of volcanic ash in Alaska. A description of the roles of all partners is beyond the scope of this paper, but can be found in Neal et al. (2010), and the current plan itself is available online<sup>5</sup> . Below, we briefly summarize those partners who, in conjunction with AVO, issue formal warning products about volcanic ash and hazards.

The NWS Alaska Aviation Weather Unit (AAWU) also serves as the Anchorage Volcanic Ash Advisory Center (VAAC). The AAWU/VAAC is responsible for issuing Volcanic Ash Advisories (VAAs), which provide information on the distribution and forecast movement of ash, and Significant Meteorological Information (SIGMETs), which serve as the primary warning product to the aviation community for volcanic ash. AVO works closely with the AAWU/VAAC before and during ash-producing events to coordinate on timing and distribution of explosive events, interpretation of ash clouds in satellite imagery, sharing of lightning data and the extent and timing of NWS formal products. The FAA may institute Temporary Flight Restrictions (TFRs), in consultation with AVO.

The NWS Weather Forecast Office (WFO) in Anchorage is responsible for issuing all warnings of ashfall for the public and marine communities in Alaska. If ashfall is expected based on model output, AVO coordinates with the Anchorage WFO on the details of where, when and how much ashfall is expected and NWS warning products are issued accordingly. The United States Coast Guard may issue notices to Mariners about hazards in the marine environment.

### RESULTS

#### Observatory Response to the Eruption

Below we present a brief chronology of the eruption and describe how the operational response evolved with time.

#### Precursory Phase and Initial, Undetected Explosive Events (September Through Mid-December, 2016)

The first five explosive events, which occurred on December 12, 14, 16, and 19, 2016 (**Figure 2**), were only detected retrospectively using lightning, infrasound, satellite, and/or seismic data. These events, 1–5, were missed by AVO's routine data checks and ongoing alarms and thus AVO did not issue any notifications at the time of the events, not did any other partner agency. Seismic signals from the time of Events 4 and 5 on December 16 and 19 were noted during routine seismic checks, as being detected at Akutan, Makushin, and Okmok, but were suspected to be either tectonic tremor or low-level activity at Okmok. Retrospective analysis of the earthquake catalog, combined with match filtering, revealed that precursory volcano-tectonic earthquakes had been occurring since at least September 2016 (Stephen Holtkamp, written communication, 2016).

#### Rapid Explosive Events, Common Precursors (December 20–March 13)

Event 6 on December 20, 2016, was the first event of which AVO was aware. We were notified by a call from the FAA/CWSU calling to inform us of a PIREP of an ash eruption coming up and out of the Bering Sea. After confirming that this was from Bogoslof, AVO raised the Aviation Color Code from Unassigned to RED. A second, similar event (7) occurred about 24 h later.

An RSAM alarm that focused on Bogoslof was implemented shortly after Event 7. During this first phase of the eruption, the pace of explosions was exceptionally high, with, on average, one event every 58 h (**Figure 2**). Through late January there were about 29 short-lived events that put ash clouds

<sup>5</sup>https://avo.alaska.edu/pdfs/cit3996\_2017.pdf

up to heights of 6–11 km asl. On the night of January 30–31, a longer, ash-rich event resulted in sufficient tephra accumulation to produce a subaerial edifice that raised the vent above sea level for the first time. On January 31, AVO issued an Information Statement that provided an overview of the eruption to that time, status of the monitoring capabilities, and prognosis for future activity (**Supplementary File S1**).

Throughout the period from December through March, 12 events had precursory seismic activity that was detected by the RSAM alarm, as indicated by negative alarm latency (**Table 1** and **Figure 6**). Event 36 on February 19 is an example of such an event (**Figure 4A**). A classic sequence of coalescing earthquakes served as a prelude to a series of energetic eruptive signals that began at 17:08 and lasted over half an hour. This activity was first recognized during a scheduled duty check at about 13:00. The sequence then kept up with a relatively low rate until about 15:55 when the rate suddenly increased to about 30 earthquakes per hour. The rate then progressively increased over the next hour until the quakes had almost merged to tremor by 17:00. The first RSAM alarm triggered on the quakes at 16:44 pm. Earthquakes ceased at 17:07 and after a 1-min break transitioned to tremor. The eruptive signals consisted of about nine blasts that were clearly captured on multiple infrasound arrays. The infrasound on the Okmok array triggered the airwave alarm several times during the eruption. Because of the relatively long run-up, AVO called the FAA 124 min prior to event onset and issued an "imminent" VAN 33 min prior to the event.

#### Hiatus in Explosive Activity (March 13–May 16)

Following event 38 on March 13, there was a 9-week hiatus in explosive activity at Bogoslof. The Aviation Color Code remained at ORANGE until April 5, at which time AVO lowered it to YELLOW. The only detected activity observed during the hiatus was a swarm of volcano-tectonic earthquakes on April 15, which prompted AVO to raise the color code to ORANGE. The swarm lasted for several hours, comprised 118 detected earthquakes with M between ∼0.8 and 2.2, and is interpreted to reflect magmatic intrusion in the mid to upper crust because of the earthquakes' weak T phases (Wech et al., 2018). Following this swarm, which lasted for several hours, the color code was once again lowered to YELLOW on April 19.

#### Renewed Explosive Activity and Dome Building (May 16–August 30)

Bogoslof erupted again without precursors on May 16 (event 39; **Figure 4B**). From May 16 through August 30, AVO detected 32 explosive events at the volcano. Unlike during December– March, none of the explosions in the later phase was preceded by detectable seismic precursors, meaning that AVO was always responding to the onset of explosions rather than issuing warnings of impending activity. An exception was event 48 on June 10, which did not have seismic precursors but did have a brief initial infrasound pulse about an hour before confirmed explosive activity (**Table 1**).

On June 7, satellite imagery confirmed the presence of a subaerial lava dome at the volcano. It was located in the northern portion of the vent lagoon, had breached sea level, and was about 110 m across. The lava dome was short-lived, as it was completely destroyed during a 2 h and 10-min pulsatory event on June 10 (event 48). A second lava dome was observed on August 18 in the enclosed crater. The exact timing of the destruction of this dome is unclear due to a lack of satellite imagery, although we can infer that it occurred during the final detected event of the entire sequence on August 30 (event 70). Following nearly a month without activity, AVO lowered the aviation color code to YELLOW and Alert Level to ADVISORY on September 27, and finally, to UNASSIGNED on December 2, 2017.

During the 2016–2017 eruption of Bogoslof, AVO raised the aviation color code to RED 32 times, and was at ORANGE for most of the sequence (**Figure 2**). For eruptive periods of hours or days that included multiple explosions or prolonged seismic or infrasound signals, AVO remained at RED throughout such sequences before downgrading to ORANGE.

#### Alarm Timeliness and Efficacy

Once all Bogoslof-specific alarms were implemented (after event 7), all but four subsequent events were caught using one or more alarms (**Table 1**; see exceptions below). RSAM was the first alarm 60% of the time, and infrasound at Okmok was the first alarm 35% of the time. The median latency between event onset and receipt of alarm by observatory staff was 5 min; the mean time was 0 min (**Figure 7A**).

Three events were detected initially by lightning instead of RSAM or infrasound: 9, 25, and 27 (**Table 1**), and event 25 was the only event that was detected in real time using only lightning. During this event, wind noise masked any infrasound signal, and telemetry dropouts affected the seismic data.

Events 34, 43, 47, and 64 were the only confirmed events that did not trigger any alarms after alarms were implemented. Event 34 produced a seismic signal and an ash cloud that reached approximately 7.5 km ASL. Two of six seismic stations that made up the alarm at that time exceeded the alarm threshold, but the alarm needed three stations to trigger. Following this event, alarm thresholds were adjusted. Because the event took place during office hours, staff saw the seismic signal and issued a VAN 48 min after the event occurred (**Table 1** and **Figure 6**). Event 43 was a short-lived, low amplitude event seen seismically but not in infrasound, satellite, or lightning. An observer aboard the R/V Tiglax noted a white plume rising only several thousand feet above sea level. The seismic amplitude for this event was too low to trigger the RSAM alarm. A VAN was issued 106 min after the event. Events 47 and 64 were both very short-lived, seen only in infrasound data during retrospective analysis, and no notifications were issued for either.

Of the 58 alarmed events, 12 had RSAM alarms detect precursory seismicity (**Table 1**). These all occurred in the first half of the eruptive sequence (December through March). Event 48 on June 10 was preceded by an infrasound alarm about an hour before the main explosion.

#### Timeliness of Partner Calls

The time between event onset and the call to FAA (the first partner on the formal call-down list) ranged from 124 min before

an event (for those for which precursory seismic signals were detected; shown as negative values in **Table 1** and **Figure 7**) to 120 min after event onset (for events which were detected only). Of the 60 events for which we issued notifications, we were able to make the first FAA call prior to event onset four times, and for six additional events the recorded call time was within 5 min of the start of the event.

The median and mean latency between event onset and call time for all events were 20 and 27 min, and did not trend appreciably through the eruptive sequence (**Figure 6**). For larger events that produced plumes >7.5 km asl and were alarmed, median and mean call times were both 15 min; for smaller alarmed events with clouds below 7.5 km asl, median and mean times were longer: 26 and 33 min (**Figure 8**).

#### Timeliness of Formal Warning Products

The time between event onset and the first VAN/VONA issued for that event ranged from 33 min before event (those for which precursory seismic signals were detected; shown as negative values in **Table 1** and **Figures 6**, **7**) to 127 min after event onset (for events which were detected only). Of the 60 events for which notices were issued, we were able to issue the first notice prior to event onset 4 times (7%).

The median and mean times between event onset and issuance of the first VAN/VONA for each event were 37 and 41 min, respectively. Looking only at events for which alarms were in place, these values drop to 35 and 37 min, respectively. And for events that generated plumes greater than 7.5 km asl, and had alarms in place, the median and mean times to VAN issuance were 32 and 30 min, respectively (**Figure 8**).

Looking only at events for which **no** precursory activity was observed, the VAN latency averaged 45 min. This time reflects reaction time when we are in "detect only" mode—typical for most Bogoslof events in this sequence as well eruptions at other unmonitored volcanoes in Alaska (notably Cleveland—see De Angelis et al., 2012). This time reflects the time between event onset and initial alarm, a scientist evaluating the validity of the alarm(s), contacting one or more other duty staff, assessing other data streams, drafting the notice in HANS, and releasing it (e.g., **Figure 4C**). VANs contain event start time, duration (if not ongoing), data streams used to confirm event, and any information about cloud height and movement. Because all seismic data used during this eruption were distant, increased uncertainty about the precise nature of individual signals led to the desire to use multiple data streams.

As the eruption progressed, AVO scientists became more adept at distinguishing co-eruptive tremor signals from other types of seismicity and became more confident in interpreting these distant signals. This would hopefully lead to decreased latency between event onset and VAN. As seen in **Figure 6**, however, some events later in the eruptive sequence still had latencies of over 30 min. This is due, in part, to the changing character of the explosive events themselves. Smaller events later in the sequence, such as 63 and 70, with more equivocal signals and fewer data streams were harder to interpret, leading to greater uncertainty and longer time between event onset and notice release.

#### DISCUSSION

The Bogoslof eruption's high number of explosive events allowed us develop new operational tools and protocols, and to put these developments into practice. The large number of events also allowed us to retrospectively analyze the factors that affect warnings. For more short-lived sequences, it is not possible to investigate these factors. Below we discuss the factors that impact warning timeliness, the particular hazards posed by Bogoslof and other remote volcanoes and how to cater warnings to those hazards, and finally, implications for future monitoring and forecasting in Alaska and other remote regions.

#### Factors That Impact Warning Timeliness

The primary factor that influenced our ability to provide timely warnings was, of course, whether precursory seismic activity was detected by the remote networks. For those events that were preceded by seismic precursors, we issued warnings and calls prior to event onset (**Figures 4A**, **6**). That this was possible at all at an unmonitored volcano was due to the relative proximity of the

Makushin and Okmok networks, and the significant seismicity of the eruptive sequence, at least for the first few weeks.

means. Outliers are not shown; for full distribution, see Figure 7.

Power and Cameron (2018) investigated the time between explosive event onset and initial call down for large ashproducing events at seismically monitored volcanoes in Alaska since 1989. They find that in these instances, reaction time (call time) ranged from <1 to 86 min. Shorter times are for intrasequence events at Redoubt, Spurr, and Augustine; longer times are for explosions without geophysical precursors.

For the Bogoslof events with no detectable precursors and for which only detection was possible, notifications were typically issued faster for larger explosions, because there was less uncertainty associated with these events (**Figures 7**, **8**). In general, larger events "lit up" more of the primary real-time and near-realtime data streams that we used to monitor Bogoslof—seismic, infrasound, lightning, satellite, and observer reports. Our latency improved (got smaller) with an increasing number of available data streams (**Figures 4B,C**, **9**). Whereas uncertainty has been discussed as playing a role in hindering accurate forecasts of the onset of impending activity (e.g., Marzocchi et al., 2012; Doyle et al., 2014), we also show that uncertainty can impact the ability to confirm and characterize activity after it has started. In general, decreasing the uncertainty in the character of the event gave us more confidence in our forecasts and allowed us to issue them sooner. For smaller events with fewer corroborating data, it took longer to (a) confirm an event and (b) determine its magnitude (**Figure 4C**).

In addition to the overall number of available data streams, some types of data were more impactful in issuing timely warnings. **Figure 10** shows the distribution of VAN/VONA latency with and without four main data streams (this analysis was not done for seismic data, which was available for all but two of the events). The biggest decreases in warning time

come when lightning and satellite data were available, which caused average decreases in notification time of 13 and 11 min, respectively. Because lightning is otherwise so rare in the area around Bogoslof, its presence was excellent confirmation of activity, greatly decreasing uncertainty. And because ash-cloud height, normally related to mass eruption rate, was perhaps the most important factor in evaluating the hazard posed by each event, having satellite confirmation of the event combined with estimates of cloud height also allowed us to issue notifications more quickly.

It is also important to point out that the timeliness we are evaluating here has a distinct human factor—a number of different scientists, at various times of day or night, were responsible for releasing notices and making calls to partners. There will be natural variability in the speed with which different scientists can perform these duties. Despite implementing standardized protocols, the poor quality of the data (due to eruption of an unmonitored volcano) and decision to not have full-time staffing but to instead rely on alarms, undoubtedly led to slightly increased and variable latencies.

#### Warnings to Match the Hazards

Bogoslof is a remote volcano and the primary hazard is posed to aviation by ash clouds generated during explosive eruptions. Unlike volcanoes that are near large populations and infrastructure, where warnings related to volcano hazards may initiate complex and costly evacuations and public concern, warnings about activity at Bogoslof led to fairly straightforward

actions to ensure that aircraft would divert around any ashbearing cloud. In this case, the use of straightforward color codes and warnings was effective for managing the crisis (Papale, 2017).

In addition to the specific-event-driven warnings that were issued by AVO and the Anchorage VAAC with respect to airborne ash, the Federal Aviation Administration (FAA) imposed a TFR around Bogoslof Island from January 9 to October 9, 2017, with a radius of 10 nautical miles that reached from sea level to 12.2 km asl.

Additionally, Marine Weather Statements (for ashfall on a marine environment) were issued by the Anchorage WFO and broadcast via United States Coast Guard (USCG) for most of the 64 explosive events due to the busy marine shipping lanes and proximity to Dutch Harbors, the nation's busiest marine fishing port. Describing the hazards local to Bogoslof during frequent explosions, AVO worked with the USCG to issue a Local Notice to Mariners (LNM) for a six nautical mile radius of the island beginning January 31, 2017, for the duration of the eruption; LNM's are issued weekly on the Coast Guard's website. Based on wind direction and intensity of the eruption, ashfall was only expected to make landfall on about six of the 64 explosive events and advisories for communities were issued for each. Depending on wind speed, dispersion models suggest that trace ashfall reached as far as ∼200 km from the volcano, and typical onset of ashfall in the nearby community of Dutch Harbor was anywhere from ∼2 to 5 h after the beginning of an explosion.

#### Toward Rapid Detection of Eruptions at Remote Volcanoes

Although our multidisciplinary approach to monitoring yielded a response that resulted in zero encounters between aircraft and ash clouds, Bogoslof also highlighted the challenges in volcano science and monitoring in Alaska and elsewhere, especially where ground-based monitoring is absent. The ideal in eruption forecasting is to be able to warn of future volcanic activity, and for much of the sequence we were limited to timely detection.

The USGS uses a threat assessment framework to define threat levels at each United States volcano based on exposure and hazard factors and to prioritize future efforts to expand the monitoring network to include more in situ monitoring (Ewert, 2007). Because of Bogoslof's remote location, it falls in the moderatethreat category, and therefore, will not be a priority for in situ monitoring in the short term.

The Bogoslof eruption showed, however, that even without in situ monitoring, new tools can allow us to rapidly detect the onset of explosive activity, characterize the resulting cloud, and forecast hazards associated with the eruption. In the past 10 years, new installation and analysis tools have expanded the use of infrasound to monitor activity in Alaska and elsewhere (De Angelis et al., 2012; Fee et al., 2013). While infrasound has primarily been used as a detection and not a forecasting tool, detection may be sufficient for remote volcanoes where the main hazard is posed by airborne ash clouds that may impact aviation or deposit ash on distant communities. In addition to infrasound, Bogoslof clearly identified the benefits of lightning as a rapid detection tool. Integrated alarm systems that look at these and other data streams in concert will allow observatories to decrease false positives and rapidly identify activity at unexpected volcanoes.

Finally, as part of the National Volcano Early Warning System, the USGS has proposed a 24/7 Volcano Watch Office that would take full advantage of real-time monitoring networks and improve delivery of hazard information to key users (Ewert et al., 2006; United States Geological Survey [USGS], 2007). Such a watch office might have caught the initial explosive events from Bogoslof in December 2016, especially if integrated alarms are developed further. A watch office would also likely decrease latencies in ongoing eruptions such as Bogoslof since it would eliminate the "activation energy" that is present when responding to alarms. Scientists who staff such a watch office would need to monitor multiple, interdisciplinary data streams and be prepared to rapidly take action in the event of explosive activity.

#### CONCLUSION

For eruptions of remote volcanoes whose explosive ash clouds pose hazards to aviation and downwind communities, short-term forecasting and detection using remote tools can provide the necessary information to mitigate risks. During the 2016–2017 eruption of Bogoslof, which lacks an in situ monitoring network, this type of response was accomplished using a multi-disciplinary approach that included seismic, infrasound, lightning, and remote sensing data combined with observer reports, automated alarms, observatory protocols and communication tools, and ash dispersion modeling. Information about the onset time and duration of explosive events, and the height and movement of resulting volcanic clouds, was conveyed using telephone calls to partner agencies as well as written warnings.

Of the 60 explosive events for which notifications were issued, aviation authorities were notified by phone an average of 22 min after event, and written notice was issued an average of 37 min after onset. For more significant events that produced clouds higher than 7.5 km asl, these averages drop to 15 and 30 min, respectively. This improvement in timeliness is because larger events are typically seen in more data types, decreasing uncertainty about the existence and character of the eruption.

Future advancements in short-term forecasting and detection at volcanoes such as Bogoslof would be possible by improved alarm integration, better regional networks of infrasound and lightning sensors, decreased latency to receipt of satellite imagery, and 24/7 staffing of volcano observatories.

#### DATA AVAILABILITY

All datasets analyzed for this study are included in the manuscript and the **Supplementary File S1**.

### AUTHOR CONTRIBUTIONS

All authors contributed to the development of the tools and approaches used during the eruption response. KW, AW, JL, DS, MH, and MC compiled the data on explosive events. MC and AW created the figures and compiled data on observatory alarm and warning times. MC analyzed the response data. MC, AW, MH, JL, DS, HS, and KW wrote the text.

### FUNDING

Funding for this study was provided by the USGS Volcano Hazards Program.

#### ACKNOWLEDGMENTS

The eruption response was performed by a large team at AVO, which is a cooperative program of the United States Geological Survey, the Alaska Division of Geological and Geophysical Surveys, and the University of Alaska Fairbanks Geophysical Institute.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart.2018. 00122/full#supplementary-material

FILE S1 | Text file of the Information Statement released by the Alaska Volcano Observatory on January 31, 2018.

#### REFERENCES

feart-06-00122 August 30, 2018 Time: 17:5 # 17


**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 Coombs, Wech, Haney, Lyons, Schneider, Schwaiger, Wallace, Fee, Freymueller, Schaefer and Tepp. 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.

## Top–Down Precursory Volcanic Seismicity: Implications for 'Stealth' Magma Ascent and Long-Term Eruption Forecasting

#### Diana C. Roman<sup>1</sup> \* and Katharine V. Cashman<sup>2</sup>

<sup>1</sup> Department of Terrestrial Magnetism, Carnegie Institution for Science, Washington, DC, United States, <sup>2</sup> School of Earth Sciences, University of Bristol, Bristol, United Kingdom

Volcanic eruptions occur when a conduit forms to connect a crustal magma reservoir to Earth's surface. Conduit formation is generally assumed to be a 'bottom–up' process and a major driver of precursory volcanic seismicity, which is the most commonly monitored parameter at volcanoes worldwide. If both assumptions are true, initial precursory seismicity should coincide spatially with petrologically-estimated magma reservoir depths. A review of six well-constrained case studies of arc volcanoes that erupt after repose intervals of decades indicates that, to the contrary, initial precursory seismicity is consistently several kilometers shallower than the magma reservoir. We propose a model involving a three-phase process of unrest and eruption: initial (partial) conduit formation occurs during a 'staging' phase, either aseismically or long before the onset of the immediate precursory run-up to eruption. Staging may involve slow ascent rates and/or small volumes. A destabilization phase then coincides with the onset of precursory seismicity, leading to a 'tapping' phase that involves additional magma ascent from the magma reservoir. This model implies that, most critically, it may be possible to detect precursory magma ascent well before the onset of seismic activity by continuous monitoring of the state of stress in the mid to shallow crust.

#### Edited by:

Corentin Caudron, Ghent University, Belgium

#### Reviewed by:

Stephanie Prejean, United States Geological Survey, United States Jurgen Neuberg, University of Leeds, United Kingdom

> \*Correspondence: Diana C. Roman droman@carnegiescience.edu

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 14 March 2018 Accepted: 10 August 2018 Published: 04 September 2018

#### Citation:

Roman DC and Cashman KV (2018) Top–Down Precursory Volcanic Seismicity: Implications for 'Stealth' Magma Ascent and Long-Term Eruption Forecasting. Front. Earth Sci. 6:124. doi: 10.3389/feart.2018.00124 Keywords: volcanic conduit formation, seismic precursors, VT seismicity, volcanic eruption, magma ascent

#### INTRODUCTION

Magma migrates upwards from its ultimate source region in the mantle, often stalling in the midcrust for an indeterminate period of time before erupting. Several signals detectable at Earth's surface are thought to reflect magma migration through the crust, and are thus routinely monitored as a basis for detecting magmatic unrest and forecasting eruptive activity. One reasonable firstorder assumption is thus that the time-depth progression of a monitored signal should reflect upward movement of magma before an eruption. Another is that the earliest instances of the precursory signal should coincide spatially with the location of the magma reservoir feeding the eruption, which may be constrained by various petrologic indicators. These two assumptions motivate our analysis of published studies documenting time-depth patterns of precursory seismic activity and petrologically-constrained source depths, with the aim of evaluating the spatiotemporal dynamics of magma ascent and developing paradigms that could extend forecasts of impending volcanic activity by months to years.

### BACKGROUND

feart-06-00124 September 1, 2018 Time: 11:2 # 2

All models of eruption triggering are built around the central concept of an upper-crustal reservoir [3–10 km below sea level (BSL)], where magmas are staged for some duration before eruption (**Figure 1A**). This concept is supported by numerous geophysical, geochemical and petrologic observations (e.g., Zimmer et al., 2007), including maximum dissolved H2O content in melt inclusions, experimentally-reproduced phase assemblages and compositions, geodetic inflation/deflation source depths, and locations/depths of seismicity during and following eruptions (Lowenstern, 2003; Dzurisin, 2006; Scaillet et al., 2008; Segall, 2010; Edmonds and Wallace, 2017). Reservoir depth is regularly determined for individual eruptions through analysis of eruptive products in combination with geophysical observations including the maximum observed depth of seismicity (Hammer and Rutherford, 2003; Blundy et al., 2008). An emerging school of thought, however, conceives this upper-crustal reservoir in arc volcanoes as often volumetrically small and representing only the top of a vertically-extensive column of melt lenses embedded within a largely crystallized mush (Cashman et al., 2017); the extent to which these melt lenses can rapidly coalesce to form large single magma bodies (e.g., Druitt et al., 2012), or be tapped sequentially during a single eruptive episode (e.g., Tarasewicz et al., 2012), is a fundamental question in volcanology (Sparks and Cashman, 2017).

From a different perspective, we know that before an eruption can occur, a conduit must form to connect the upper-crustal reservoir to Earth's surface (Scandone et al., 2007) (**Figures 1B,C**). Conduit formation underlies the concept of eruption triggering mechanisms, which are commonly attributed to the upper-crustal reservoir reaching (1) a critical volume of eruptible melt (Parks et al., 2012), (2) a critical buoyancy (Caricchi et al., 2014), or (3) a critical volatile overpressure (Tait et al., 1989). These states may be achieved by input of new magma and/or volatiles from a deeper region or by crystallization-induced vapor saturation ('second boiling') of magma within the upper-crustal reservoir. These triggering mechanisms have been suggested based on evidence of mafic inclusions, magma mixing and disequilibrium crystal textures in erupted magmas (e.g., Plail et al., 2018). The timing of mafic inputs is commonly constrained by diffusion chronometry, and suggests eruption-generating disturbances to large magmatic systems may occur decades (Druitt et al., 2012; Barker et al., 2016), years (Morgan et al., 2004; Saunders et al., 2012); or months (Kilgour et al., 2014; Till et al., 2015; Rasmussen et al., 2018) prior to the eruption onset. Attempts to link diffusion time scales directly to monitoring data are limited; to date, the best correlations appear to be with deep-seated seismicity (Saunders et al., 2012; Kilgour et al., 2014; Rasmussen et al., 2018).

The process of conduit formation is generally assumed to be seismogenic (e.g., Rubin et al., 1998; Kilburn, 2003) and reflected by the days to months of precursory seismic unrest observed at most well-monitored volcanoes. A recent analysis of reports of volcanic unrest worldwide occurring between 2000 and 2011 demonstrated that seismic unrest was the most commonly documented pre-eruptive unrest indicator (57 instances of pre-eruptive seismic unrest compared to 26 instances of pre-eruptive degassing, the next most common indicator) and had a mean duration of 192 days ( ± 525 days) (Phillipson et al., 2013). Similarly, a recent study of caldera unrest

encompasses deep vertical transfers of magma. (B) Prior to eruption, a conduit must be formed. This process is assumed to be characterized by a 'restless' state involving elevated rates of seismicity, deformation, and gas emissions. In some cases, conduit formation begins but is never completed, leading to a "failed eruption" (Moran et al., 2011). (C) If a through-going conduit is formed, magma ascends to intersect the surface and erupt.

worldwide between 1988 and 2014 indicated that 72% of preeruptive unrest episodes involved high seismicity and degassing (Sandri et al., 2017). Precursory seismic unrest may involve a range of signal types, including high-frequency or 'volcanotectonic' (VT) earthquakes and low-frequency, or 'long-period' (LP) events. VT earthquakes are generally thought to reflect shear failure of rock due to stress changes in the crust produced by magma migration, while LP events are considered to be a more direct consequence of various fluid flow processes (Chouet and Matoza, 2013 and references therein). A recent study further proposed that explosive eruptions are preceded by 'distal VT seismicity' (which is defined in the study as high-frequency seismicity that occurs in swarm-like distributions and originates at lateral distances of 1–45 km from the eventual eruption vent) and that, at volcanoes in repose for two decades or longer, distal VT seismicity is the earliest seismic precursor (White and McCausland, 2016). However, at most volcanoes (including all four case study volcanoes considered by White and McCausland, 2016) the local seismic monitoring network is non-existent or rudimentary at the start of seismic unrest, and earthquake location errors generally exceed several kilometers, making it nearly impossible to accurately assess the exact location of initial volcano-seismic unrest.

None of the above-described models for eruption triggering address the mechanical process of conduit formation, nor the timescale over which it occurs. One conceptual model for conduit formation involves a 'bottom–up' process, whereby overpressures in an upper-crustal reservoir exceed the tensile strength of overlying rock, which fractures to form a dike that propagates upwards to intersect the surface (Rivalta et al., 2017 and references therein). However, a long-noted problem with this mechanism is the rarity of observations of upward propagation of volcanic hypocenters suggestive of magma migration, even at densely instrumented volcanoes (e.g., Roman and Cashman, 2006; Scandone et al., 2007). Rare cases of precursory hypocenter propagation are limited to hot spots (Battaglia et al., 2005; Taisne et al., 2011) or are subtle and apparent only following high-precision relocation (Patane et al., 2002), and are more often lateral than upwards (e.g., Rubin et al., 1998; Sigmundsson et al., 2015). As an alternative, Scandone et al. (2007) proposed that conduit formation may precede introduction of magma into the fractures that comprise it, with magma ascent a passive response to mechanical failure of the overlying rock. The proposed mechanisms for this process are either external (e.g., tectonic) or internal (e.g., release of volatiles from the magma). Another possibility is that conduits develop from the 'top–down,' as documented in observations of seismicity and degassing at mafic volcanoes in both open system (e.g., Girona et al., 2015; Ripepe et al., 2015) and rift (e.g., Tarasewicz et al., 2012) settings. A similar mechanism has been suggested for some mafic caldera-forming eruptions (Cashman and Giordano, 2014). Such 'top–down' mechanisms, however, apply to situations where the conduit has already been formed – i.e., magma has already migrated upwards from the reservoir, and not to the stage of initial conduit formation.

In light of the difficulty reconciling mechanisms of conduit formation with observations of precursory seismicity, we hypothesize that precursory seismicity does not reflect initial conduit formation, nor, by implication, reservoir destabilization. Rather, we suggest that reservoir destabilization and initial conduit formation may occur either aseismically, or long before the immediate seismic run-up to eruption. In the following sections, we evaluate this hypothesis by comparing information on reservoir depth and suggested eruption triggering mechanisms derived from petrologic analyses with time-depth patterns of precursory seismicity during the immediate run-up to eruption at well-monitored and frequently-erupting volcanoes. We use these data to develop a conceptual model for the processes occurring during conduit formation. Finally, we explore the implications of this model for long- and intermediate-term eruption forecasting.

#### CASE STUDIES

We searched the peer-reviewed literature, Smithsonian Global Volcanism Program database, and reports published on observatory websites for well-constrained case studies that document the timing and depth of pre-eruption seismicity at volcanoes worldwide. We consider only case studies that meet three criteria. First, the volcano must have been monitored by a 'Level 3' seismic network [defined by Moran et al. (2008a) as the minimum seismic instrumentation for accurate hypocenter location] centered on the location of earliest seismicity for at least a year prior to eruption. Level 3 seismic monitoring requires a network with at least two seismometers located within 5 km of the vent, and six seismometers within 20 km of the vent. Second, because we are interested in examining the earliest stages of conduit formation and its associated geophysical signals, we consider only eruptions that have occurred after a minimum 10-year period of no surface activity (either phreatic or magmatic eruption). Third, we consider only eruptions that involve magma. We take reported observations for all case studies meeting the above criteria at face value, that is, we consider that the earliest seismic unrest was detected and that hypocenter locations are accurate. We recognize, at the same time, that factors such as temporary instrument failures, event detection protocols, background noise levels, and inaccuracies in velocity models, may limit the accuracy of the timing and location of the earliest precursory seismicity. To mitigate this issue, we filter all data to consider only "well-located" earthquakes, which we define as having an azimuthal gap < 180◦ , RMS < 0.20 s, and horizontal and vertical location error < 3 km. We combine the seismic data with all available petrologic, gas emission, and geodetic observations that help to constrain the depth(s) of pre-eruptive magma staging and timing of magmatic unrest.

We identified six case studies that meet the above criteria, including four eruptions in Alaska's Aleutian arc (Crater Peak/Spurr 1992, Redoubt 2009, Augustine 2006, and Okmok 2008; **Figure 2A**), one eruption in the Cascade arc (Mt. St. Helens, Washington, 2004–2008; **Figure 2B**), and one in the Izu arc (Miyake-jima, Japan, 2000; **Figure 2C**). While these six case studies represent a range of magma compositions (from basalt to dacite) and settings (four in continental arcs and two in island

volcanoes. (B) Map of the Pacific Northwest region (United States) showing the location of Mt. St. Helens. (C) Map of Japan showing the location of Miyake-jima

arcs), all six volcanoes have experienced at least two eruptive episodes in the past 40–50 years (the main reason they have good seismic monitoring networks). Thus, together they present the opportunity to examine eruptive precursors in frequently active (i.e., erupting every few decades) systems, which may or may not be different from those in long-dormant (i.e., no eruption for centuries to millenia) volcanoes.

We summarize key precursory observations of seismicity, gas emissions, and deformation, along with petrological constraints on the depth of magma genesis and storage, in the sections below. One important piece of 'housekeeping' necessary for accurate comparison of earthquake depths and petrologically-derived magma storage depths is reconciliation to the same reference depth. Volcano seismologists generally (though not always) report earthquake depths relative to sea level (e.g., 2 km BSL). A complication is that, for computational simplicity, the top of the seismic velocity model (the computational space in which the earthquakes are located) may be slightly above the actual summit height, resulting in occasional 'airquakes' located above the summit elevation. For example, most velocity models used for Alaskan volcanoes extend to 3 km ASL (e.g., Dixon et al., 2013) even though summit heights of Alaskan volcanoes range from 1 to 3 km ASL. Petrologists, in contrast, report depth by converting pressures (inferred from melt composition including maximum dissolved volatile contents, phase assemblages, and/or textural observations) through an assumed crustal density as P(z) = gρz (e.g., 100 MPa = 4 km depth for an average crustal density of 2500 kg/m<sup>3</sup> ). Therefore, reported petrologic depths are generally relative to the summit elevation (the total overburden in km). To reconcile the two sets of depths, we correct (either add to seismic depth or subtract from petrologic depth) by the elevation of the vent.

In the sections below, we summarize key precursory observations of seismicity, gas emissions, and deformation, along with petrological constraints on the depth of magma genesis and storage. We also review what is known about the penultimate eruption at each volcano, including when it occurred, the eruption precursors, the petrology of the eruptive products, and the accepted interpretation of the magma storage conditions and/or eruption triggers. Our goal in these case study reviews is to build a picture of the magma storage system at each volcano, which we can then compare, in hindsight, to the information about the system provided by the eruption precursors.

#### Crater Peak, Alaska – 1992 Eruption (Vent Elev. 2.2 km ASL)

Crater Peak (Mt. Spurr), Alaska, experienced a series of three relatively quick (∼4 h) subplinian eruptions on June 27, August 18, and September 16–17, 1992 (Keith, 1995 and papers within this volume)) (**Figure 3**). The previous eruption of Spurr occurred on July 9, 1956 and comprised two main explosive phases (to a height of 20 km ASL) followed by steam and ash emissions. The erupted material was basaltic andesite which deposited mostly as ash, although an associated debris flow temporarily blocked the Chakachatna River (Eichelberger et al., 1995). There is no geophysical information on this eruption. Crater Peak has had a Level 3 seismic network since August 1989. The first sign of unrest preceding the eruptions was a swarm of volcano-tectonic (VT) earthquakes in August 1991

directly beneath the Crater Peak vent (1–4 km BSL) (Power et al., 1995). The rate of shallow VT seismicity then increased in the 7 months preceding the first (June) eruption, with a distal VT swarm ∼8 km NE of the vent in March including earthquakes located as deep as 10 km, and two particularly intense shallow VT swarms on June 5 and 27 (Power et al., 1995). Boiling, evaporation, and chemical changes (as denoted by changing color) of a small crater lake began in June, and the lake disappeared before the first eruption. No precursory gas emission data were collected, but lake water sampled on June 8 was found to have an elevated SO<sup>4</sup> content (Keith et al., 1995). Syneruptive seismicity was relatively low (a maximum of 10 events/day), but the final (September) eruption was both accompanied and followed by a strong swarm of VT earthquakes at −2 to 11 km BSL (Power et al., 1995), and two non-eruptive swarms in November (−3 to 3 km BSL) and December (8–10 km BSL) of 1992 (Power et al., 1995). Several months of abundant deep-long period (DLP) seismicity followed the eruption. A major episode of caldera-wide seismic and fumarolic unrest in 2004–2005 did not culminate in an eruption and has been interpreted as the result of a magmatic intrusion (Coombs et al., 2006; Koulakov et al., 2013).

There is no evidence for a mafic injection trigger for the 1992 Crater Peak eruption. That Crater Peak magmas were stored at mid-crustal depths immediately prior to eruption is indicated by the presence of abundant pristine hornblende phenocrysts in the erupted products (Harbin et al., 1995). The absence of hornblende breakdown rims also requires water contents of several percent (Grove et al., 1997; Gardner et al., 1998). Combined petrologic and seismic observations, therefore, indicate pre-eruptive magma storage at ∼10 km BSL (Power et al., 2002).

#### Redoubt Volcano, Alaska – 2009 Eruption (Vent Elev. 2.5 km ASL)

Redoubt Volcano erupted on March 23, 2009 (Waythomas and Webley, 2013 and papers within this volume) (**Figure 4**). The eruption lasted several months and consisted of multiple explosions that produced andesitic lava and tephra

(Schaefer, 2012), culminating with the extrusion of a lava dome in the summit crater. It has had a Level 3 seismic network since September 1990.

Its previous eruption was in 1989–1990, and a few small LP events were recorded once a local seismic network became operational in October 1989, but strong precursory seismicity began only ∼24 h before the eruption onset on December 14 (Power et al., 1994). Pilots reported steam plumes in late November and early December (Miller and Chouet, 1994. No geodetic monitoring existed at the time and the occurrence of the eruption in the middle of Alaskan winter may have limited opportunities to visually observe earlier surficial changes at the summit, such as increased ice melting. Early erupted material included both andesite and dacite (58.2–63.4 wt% SiO2); later products showed abundant evidence of mixing and trended toward an intermediate composition (58.5–60.5 wt% SiO2), suggesting that individual eruptions in the 1989–1990 sequence were fed by separate magma pulses from 6 to 10 km rather than tapping of a single large reservoir (Wolf and Eichelberger, 1997).

The first observed sign of unrest prior to the 2009 eruption, in retrospect, was anomalous ground deformation starting in April or May of 2008 at the continuous GPS station AC17 operated by EarthScope roughly 26 km northwest of the summit (Grapenthin et al., 2013). In mid-July, and again in mid-September of 2008, a strong sulfur smell was reported downwind of the volcano (Schaefer, 2012). Anomalous snow melt was observed in the Redoubt crater in mid-September, and increased fumarolic activity was observed through late 2008 (Bull and Buurman, 2013). Measured CO<sup>2</sup> emissions in October and November 2008 (1220–1368 t/d CO2) were relatively high and noteworthy given there was little other indication of increasing unrest (aside from continuing subtle inflation) (Werner et al., 2012).

No unusual earthquake activity was observed in association with the earliest (April–July) signs of unrest on seismic stations close to the volcano (Ketner and Power, 2013). Redoubt seismometers recorded a tremor-like signal in late September, and a small number of LP events were identified in October and November of 2008, but were too small to be located. On December 12, 2008, AVO began to locate DLP events and VT earthquakes at 28–35 km depth beneath the volcano. Near-continuous shallow volcanic tremor began on January 24, 2009, accompanied by episodic swarms of both VT and LP earthquakes located from −3 to 3.8 km BSL (January 25), −3 to 4.9 km BSL (January 30–31), and −3 to 9.28 km BSL (February 26–27). Relative seismic quiescence began on March 1 and culminated in a phreatic explosion on March 15. The onset of the magmatic phase on March 23 was immediately preceded by a 58-h long swarm of seismic events located from −1 to 9.1 km BSL. Syneruptive seismicity was characterized by ongoing swarms of discrete events and gliding, high-frequency harmonic tremor (interpreted as the superposition of frequently-repeating stick-slip earthquakes) immediately preceding explosions (Hotovec et al., 2013). Following the end of the eruption, seismicity rates gradually declined to low background levels by July 2009.

Petrologic observations point to months of pre-eruptive staging of all magmas erupted in 2009 at a minimum of 4–6 km depth (100–160 MPa), or approximately 1.5–3.5 km BSL. Equilibrium phase assemblages, amphibole rim thicknesses, and plagioclase rim hygrometry suggest that the earliest erupted product, a low-silica andesite, ascended to a depth of 1.5–3.5 km BSL from an unknown depth (assumed to correspond to the cluster of DLP events located between 28 and 37 km BSL) in mid-2008 to early 2009; here it paused, equilibrated, and remobilized stagnant mushy magmas already present in the mid-crust to produce the higher-silica products of the later eruptive phases (Coombs et al., 2013). Petrologic depth estimates could, however, be increased by up to several km if the magmas contained appreciable CO<sup>2</sup> (Coombs et al., 2013), as suggested by relatively high CO<sup>2</sup> fluxes and high C/S ratios measured during the precursory phase (Werner et al., 2012, 2013). Additional evidence that the petrologic depth is underestimated comes from analytical models of observed precursory deformation, which suggest a deeper reservoir located at ∼13 km BSL, and models of observed syneruptive deflation, which suggest a reservoir at ∼9 km BSL (Grapenthin et al., 2013).

#### Augustine Volcano, Alaska – 2006 Eruption (Vent Elev. 1.1 km ASL)

Augustine Volcano, Alaska, began erupting on January 11, 2006 with a subplinian eruption followed by 2 months of andesitic dome-building activity (Power et al., 2010 and papers within this volume) (**Figure 5**). Its previous eruption was in 1986, and it has had a Level 3 seismic network since June 1988. The 1986 eruption was preceded by several distinct shallow VT swarms separated by periods of quiescence (Power, 1988) – gas and deformation were not monitored prior to the eruption. Material erupted in 1986 had a range of whole-rock compositions from basaltic andesite to dacite, with evidence for a mafic injection trigger (Roman et al., 2006). Lee et al. (2010) analyzed SAR data for the period 1992–2005 and found evidence for wholesale and steady uplift of the entire volcano during this 13-year period, which they model with two Mogi sources beneath the summit – an inflating source at 7–12 km BSL (interpreted as a long-term magma storage zone) and a contracting source at 2–4 km BSL (interpreted as a subsidiary reservoir tapped during the 2006 eruption).

The first sign of immediate unrest preceding the 2006 eruption was a slow, steady increase in the number of volcano-tectonic (VT) earthquakes beginning in May 2005 (Power and Lalla, 2010). Relocated hypocenters had depths between 0.1 and 0.6 km ASL (Power and Lalla, 2010). The rate of shallow VT seismicity increased through December. GPS-detected inflation began in mid-summer, with radial deformation in GPS baselines indicating an inflation source at approximately sea level (Cervelli et al., 2006). Steam explosions began in early December and continued through the precursory period, accompanied by strongly elevated rates of shallow seismicity (∼0–3 km ASL). Explosive activity began on January 11, 2006, immediately following a 13-h long swarm of shallow VT earthquakes with depths of 0.5–1 km ASL. The volcano began deflating on January 28 and continued through February 10, with the locus of deformation at approximately 3.5 km BSL (Cervelli et al., 2006). On February 3–4, during this deflation period, there was

a swarm of VT earthquakes with depths 2–4 km BSL. From mid-February through the end of the eruption, the volcano was seismicially quiet, except for rockfall signals. Drumbeat seismicity accompanying lava extrusion emerged on March 8 and merged into a near continuous signal until March 13. Drumbeats ceased to be detected by March 16. There was little seismic activity through the rest of 2006.

The products of the 2006 eruption of Augustine are heterogeneous and record evidence for significant mixing of low- and high-silica andesite prior to the eruption (Larsen et al., 2010). Phenocrysts in the low-silica end member further suggest involvement of an unerupted basaltic magma. Together these data suggest that the high-silica endmember was stored in a mush region with its top at ∼5 km (4 km BSL) and that it was intruded by a basaltic magma that mixed to form the low-silica andesitic eruptive products (Larsen et al., 2010). Similar analyses from previous eruptions (e.g., Roman et al., 2006) suggest that this scenario is common for Augustine.

#### Okmok Volcano, Alaska – 2008 Eruption (Vent Elev. 700 m ASL)

Okmok Volcano, Alaska, experienced a 5-week-long phreatomagmatic eruption beginning at Cone D on July 12, 2008 (**Figure 6**). Its previous eruption in 1997, and all other historic eruptions of Okmok occurred at Cone A, a vent approximately 5 km to the southwest of the 2008 vents (**Figure 6B**). The 1997 eruption was completely unmonitored by groundbased instrumentation and included an explosive phase and emplacement of a basaltic lava flow. Okmok has had a Level 3 seismic network since January 2003. The 2008 eruption is notable for its lack of long-intermediate term seismic precursors despite the presence of a dense seismic network on and around the vent, although caldera-wide inflation resumed in early 2008 following a 3-year pause. The only detected seismic precursor to the eruption was a 5-h period of low-magnitude VT earthquakes immediately preceding the eruption, and occurring at an increasing rate in the hour before eruption (Johnson et al., 2010). Catalog locations for these events have epicenters distributed throughout the northern half of the caldera and depths ranging from −3 to 14 km BSL, with no apparent time progression in depth (**Figure 6B**). During the 5-week long eruption, the rate of seismic events was elevated (an average of 6 events/day), with most events located between 0 and 6 km BSL.

Several studies have assessed the depth of magma storage at Okmok using tomographic approaches. An analysis of ambient noise tomography in combination with InSAR data indicates that Okmok is underlain by two low-velocity zones, one extending from the surface to 2 km BSL and a second located at 4–6 km BSL (Masterlark et al., 2010). Finite-element models show consistency with an inflating magma body in the deeper zone and patterns of deformation spanning the 1997 eruption imaged by InSAR (Masterlark et al., 2010). Double-difference tomography spanning the 2008 eruption also finds a low Vp and Vs anomaly directly under the caldera in a shallow zone at 0–2 km BSL which is connected to a larger deeper zone that extends to about 6 km BSL (Ohlendorf et al., 2014). Using a newly developed 3D velocity model, Ohlendorf et al. (2014) produced relative locations for a subset of earthquakes spanning the 2008 eruption and show that relocated precursory earthquakes occur within the depth range spanned by catalog locations.

The 2008 eruption produced a phenocryst-poor tholeiitic basaltic andesite that is compositionally distinct from basaltic lavas erupted in 1997. Based on the presence of disequilibrated olivine phenocrysts and melt inclusion volatile contents,

Larsen et al. (2013) conclude that the 2008 Okmok eruption was ultimately triggered by an influx of melt-rich basalt originating from a magma storage region at 3–6 km BSL into a shallower (∼2 km BSL) more evolved magma body located beneath Cone D and the 2008 vents.

### Mt. St. Helens, Washington – 2004 Eruption (Vent Elev. 2.1 km ASL)

Mt. St. Helens, Washington, began erupting on October 6, 2004 (Sherrod et al., 2008 and papers within this volume) (**Figure 7**) with a week-long explosive phase that then transitioned into a

period of dome growth and destruction that lasted until 2008. Mt. St. Helens has been Level 3 seismically monitored since June 1980 (Malone et al., 1981), shortly after the beginning of its previous eruption in 1980–1986.

The 1980 eruption was preceded by approximately 2 months of precursory seismic unrest and the growth of a conspicuous bulge on the north flank of the volcano. Although a Level 3 monitoring network did not exist during the precursory phase, initial seismicity and deformation appear to have been shallow (within and just below the edifice), and magma intrusion into the edifice appears to have occurred aseismically (Malone et al., 1981). Multiple phreatic explosions occurred during the precursory phase, and the eruption ultimately produced ∼0.2 km<sup>3</sup> DRE of dacite. The eruption was triggered by edifice failure but required 3.5 h to construct a throughgoing conduit which ultimately tapped magma stored over a large vertical depth range (Blundy et al., 2008). Magma erupted in 1980 was low-Si dacite with no evidence for pre-eruptive mafic input (Blundy et al., 2008). Between 1987 and 2002, several swarms of VT earthquakes were recorded beneath the volcano at depths

between 2 and 8 km BSL, and are interpreted as intrusions or pressurization of magma (Moran, 1994; Musumeci et al., 2002).

The first sign of unrest in 2004 was a swarm of shallow (<0 km BSL) VT earthquakes on September 23, 2004 (Moran et al., 2008b), approximately 1 km shallower than the location of long-term background microseismicity (Lehto et al., 2010). The swarm intensity increased and then declined over the next 48 h, leveling out at a steady rate by September 25. On September 25, LP seismicity joined continued VT seismicity, overtaking VT events as the dominant type of seismic event by October 5. The overall rate of seismicity continued to increase, with all events located above 2 km BSL. A small difference in P-wave arrival time differences between September 25 and 27 indicated either a decrease in shallow seismic velocities or a subtle shallowing of earthquake depths, which was followed by visible cracking in the crater glacier. Earthquake activity continued to the first phreatic explosion on October 1, which was immediately followed by seismic quiescence (Moran et al., 2008b). Shortly thereafter seismicity reintensified and was joined by strong tremor, leading to the onset of juvenile explosive activity on October 5. Syneruptive seismicity was dominated by less frequent and smaller shallow events, often comprising families of 'drumbeat' earthquakes (Moran et al., 2008b). VT events were relatively rare through the remainder of the eruption and those that were detected had shallow (<3 km BSL) depths, in contrast to syneruptive VTs during the 1980–1986 eruption which had depths to 8 km BSL.

The 2004–2008 eruption produced ∼0.1 km<sup>3</sup> of homogeneous, crystal-rich dacitic magma, largely in the form of domes and spines extruded onto the crater floor (Pallister et al., 2008). The earliest erupted samples are glassier and more vesicular than later samples, but all have low volatile contents indicative of extensive shallow degassing-driven crystallization at around the level of VT seismicity (∼1 km BSL). The apparent low pressure of the latest phenocryst growth suggests that the magma was derived from depths of ∼5 km below the vent (∼3 km BSL). Temperatures, oxygen fugacities, PH2O (850◦C, 10−12.<sup>4</sup> , and 130 MPa, respectively) also provide petrologic evidence for an origin depth of ∼3 km BSL (Rutherford and Devine, 2004), although the deep-sourced (∼5–12 km BSL) deflation of the volcano accompanying the precursory seismicity (Dzurisin et al., 2008; Lisowski et al., 2008) suggests a magma source at > ∼3 km BSL. The immediate triggering mechanism of the 2004 eruption is unclear – there is little evidence of a mafic intrusion beyond rare andesitic inclusions in some 2004 dacite samples indicative of mixing and quenching at some point in the dacite's history. Magma erupted in 2004 is compositionally related to the 1980 magma, but slightly cooler and was stored at the upper depth (low pressure end) of the 1980 range.

#### Miyake-jima Volcano, Japan – 2000 Submarine/Summit Eruption and Dike Intrusion (Submarine Vent – 0 BSL, Summit Vent – 800 m ASL)

A submarine eruption occurred off the coast of Miyake-jima Volcano, Japan, on June 27, 2000 (Nakada and Druitt, 2005 and papers within this volume) (**Figure 8**). This eruption was followed by a large offshore dike intrusion and a phreatomagmatic eruption accompanying caldera formation at Miyake-jima's summit in July and August 2000. Miyake-jima had not erupted since 1983, but was Level 3 monitored beginning in March 1999 (Ukawa et al., 2000). The 1983 eruption was preceded by a few hours of felt earthquakes (depth/location unknown) and

involved basaltic fire fountaining on the flank of Miyake-jima island (Aramaki et al., 1986). On June 26, 2000, seismic activity began beneath the southwest flank of the island of Miyake-jima at a depth of ∼2–3 km BSL, with about 4,300 earthquakes registered within the first 24 h (Uhira et al., 2005). At the same time, GPS stations detected displacement on Miyake-jima island. On June 27, a strong earthquake to the west of Miyake-jima, coupled with observations of an area of discolored seawater and steam rising from the ocean surface, indicated that a submarine eruption had occurred approximately 1 km to the west of the island. ROV investigations later identified several craters on the ocean floor consisting of fresh spatter and lapilli (Kaneko et al., 2005). From June 27-July 1, a strong earthquake swarm propagated downwards and northwestward from Miyake-jima island, accompanied by deformation patterns suggesting offshore dike emplacement. Depths of these events range from 5 to 25 km BSL, with earthquake depths increasing with time. The intruded magma is thought to have been tapped partly from a chamber below Miyake-jima (Geshi et al., 2002), with additional magma sourced from beneath the dike (Yamaoka et al., 2005). On July 8, Miyake-jima volcano began a summit eruption, after which the summit collapsed as a caldera formed over the next 40 days. Seismicity accompanied caldera formation and summit eruptions extended from the vent down to approximately 5 km BSL. Continued eruptions at the summit ultimately produced 0.02 km<sup>3</sup> of ejected tephra, and the eruption was followed by 4 years of elevated SO<sup>2</sup> emissions.

The 2000 eruption of Miyake-jima produced both basaltic andesite (June 2000 submarine eruption) and basaltic (August 2000 summit eruption) juvenile material (Amma-Miyasaka et al., 2005; Kaneko et al., 2005). Petrological analyses indicate that a reservior of basaltic andesite magma containing residual magma from the 1983 eruption was intruded by a basalt sourced from a deeper reservoir at 8–10 km depth (Amma-Miyasaka et al., 2005; Kaneko et al., 2005; Saito et al., 2005). The depth of this reservoir is unconstrained by petrology; however, GPS vectors indicate a deflation source beneath Miyake-jima at 4.2 km BSL during the swarm (Nishimura et al., 2001). The injection of basalt from the deep magma chamber into the shallow chamber is thought to have occurred before the submarine eruption (Kaneko et al., 2005).

#### DISCUSSION

#### Time-Depth Patterns of Precursory Unrest

Three of the six analyzed case studies demonstrate a clear 'top–down' pattern of precursory unrest, in which a shallow volume of crust hosts the earliest precursory seismicity, followed by seismicity over a wider range of depths. At Crater Peak, which was fed in 1992 by magma sourced from 10 km BSL, the earliest-detected precursory activity was a cluster of VT seismicity at 2–3 km BSL, with later precursory, syn- and post-eruptive seismic activity reaching depths of 10 km BSL. At Augustine, which was fed in 2006 by magma sourced from below 5 km BSL, the earliest-detected precursor was steadily-increasing VT seismicity at a depth of ∼1 km ASL, accompanied by shallow inflation. Seismic events with depths to 5 km were observed several weeks following the onset of eruptive activity. At Mt. St. Helens 2004–2008, the earliest-detected precursor was a swarm of shallow VTs at ∼1 km ASL, slightly above the depth of persistent background seismic events. No deep seismicity was observed during the 4-year-long eruption, but swarms in the 1986–2004 inter-eruptive period and 2008-present post-eruptive period have included earthquakes with depths down to 7 km BSL.

Observations at Redoubt and Miyake-jima also suggest a pattern of top–down precursory seismic activity associated with their most recent eruptions. Petrologic evidence suggests that magma erupted at Redoubt in 2009 migrated upward from an unknown depth to 1–4 km BSL approximately 6–8 months before the eruption, where it resided until erupting. The earliest seismic precursor for this eruption was a swarm of LP earthquakes in August–September 2008 that were too small to be located. Their occurrence coincided with the appearance of a thermal anomaly, suggesting activation of the shallow part of the plumbing system at this time. Later precursory and syneruptive seismicity reached 10 km BSL, below the depth of the initial LP swarm and the ultimate (inferred) staging region for the erupted magma. At Miyake-jima in 2000, both the precursory unrest and submarine eruption occurred within a 36-h period, followed by a months-long phase of dike intrusion and accompanying seismicity accompanied by caldera collapse and phreatic/phreatomagmatic eruptions. The eruption is thought to have been preceded by injection of basalt from a deep (8–10 km BSL) chamber into a shallower chamber containing residual magma before the submarine eruption. The initial seismic unrest constituted a cluster of VT earthquakes beneath the western flank of Miyake-jima at a depth of 2–3 km BSL (Uhira et al., 2005), with later seismicity beneath Miyake-jima reaching depths of ∼5 km BSL.

The 2008 eruption of Okmok shows no spatial pattern of seismic unrest. The few hours of seismicity that preceded this eruption spanned a depth range from the summit to 10 km BSL, well below the estimated 2–6 km BSL source region for this eruption. The epicenters of these events are located throughout the northern half of the caldera rather than clustered beneath the vent (**Figure 6B**), and it is possible that they represent slip on a ring fault rather than conduit formation, in which case conduit formation would have occurred aseismically. A study of precursory shear-wave splitting found no evidence for precursory aseismic magma ascent (Johnson et al., 2010), although additional work is required to understand the exact relationship of the precursory seismicity to magma transport. Regardless, while the Okmok example demonstrates that there is likely no single mechanism by which magma transport and eruption occurs, our survey indicates that a top–down pattern of precursory unrest may be predominant, at least at regularly erupting arc volcanoes. In the remainder of this paper we develop a model for this pattern and explore its implications for magma transport and eruption forecasting.

We note that, of our six case study eruptions, three of which (Crater Peak, Redoubt, and Augustine) had repose intervals of two decades or more, only two (Crater Peak and Augustine)

were preceded by obvious distal VT seismicity. Furthermore, in both of these cases the distal VT seismicity was not the earliest reported seismic precursor – at Crater Peak a distal VT swarm occurred in March 1992 (**Figure 3**), 6 months after the onset of seismic unrest beneath the Crater Peak vent, and at Augustine a distal VT swarm occurred contemporaneously with proximal precursory seismicity (Fisher et al., 2010). Although Miyake-jima, 2000 is listed by White and McCausland (2016) as having precursory distal VT seismicity, the events in question actually occurred within 1 km of the submarine vent for this eruption (Kaneko et al., 2005; Uhira et al., 2005), highlighting ambiguities in their definition of what constitutes 'distal' VT seismicity.

#### Conceptual Model of Magma Ascent and Eruption

A key question in interpreting the apparent top–down pattern of precursory seismicity is the spatial relationship between VT earthquakes and magma. VT earthquakes may result from an increase in stress in the host rock surrounding an intruding or pressurizing magma body (e.g., Bonafede and Danesi, 1997). VTs may also result from an increase in the pore pressure caused by the addition of gas and/or heat, which creates a lower slip threshold on faults in response to ambient stresses. These two mechanisms produce VTs with differently-oriented fault-plane solutions. As magmatic conduits inflate in the direction of regional minimum compression, the fault-plane solution for a VT caused by magma intrusion will have a P-axis aligned with the regional minimum compressive stress. An increase in pore pressure, however, produces VTs with fault-plane solutions consistent with the regional stress field (P-axes parallel to regional maximum compression).

Several lines of evidence indicate that initial shallow VT seismicity at our case study volcanoes is the result of shallow magma intrusion (i.e., VT earthquakes are proximal to intruding magma) rather than advection of gas and heat into a shallow hydrothermal system. First, none of the six case study volcanoes have extensive hydrothermal systems such as those at Yellowstone and Long Valley Caldera, where seismic swarms are inferred to result from perturbations to the hydrothermal system (Waite and Smith, 2002; Hill and Prejean, 2005). Second, initial shallow VTs at Spurr (Roman et al., 2004), MSH (Lehto et al., 2010), and Okmok (Ohlendorf et al., 2014) have fault-plane solutions with P-axes that are rotated by 90◦ with respect to regional maximum compression, a pattern that has been linked to proximal dike inflation (Roman and Cashman, 2006). An in-depth analysis of fault-plane solution orientations at Augustine does not yet exist – however, the onset of shallow seismicity in 2005 was accompanied by GPS-detected inflation with a modeled source located within the volcanic edifice (Cervelli et al., 2006). At Redoubt, early stress field reorientation was detected through analysis of shear-wave splitting (Roman and Gardine, 2013) but the depth of the stressed rock cannot easily be localized along the raypath. Additional work to constrain the depth of the stressed region, using a technique such as seismic interferometry, is thus warranted but beyond the scope of the present study. At Miyake-jima, initial fault-plane solutions are not rotated with respect to the regional stress field; however, stress field rotations are generally not observed preceding basaltic eruptions (Roman and Cashman, 2006); again, additional analysis to constrain the depth of pressurizing magma are warranted. Overall, however, geophysical evidence points to co-located shallow VT seismicity and pressurizing magma in the majority of our analyzed case studies.

Observations of top–down seismicity/magmatic unrest thus suggest a three-phase process of unrest and eruption, which we illustrate in **Figure 9**. Phase 1 (the 'Staging' phase) involves movement of magma from the mid- to shallow crust. Staging may be accomplished seismically or aseismically depending on its timing relative to eruption. One possibility is that upward magma movement is accompanied by seismic unrest either at the end of the previous eruption or during an intereruptive period, but not during the immediate run-up to eruption. There is strong evidence for both the former and the latter case at (1) Mt. St. Helens in 2004, where strong seismic swarms in 1990–1991 and 1998 with earthquake depths ranging from 0 to 7 km BSL may represent upward transfer of magma intro a shallow reservoir (Moran, 1994; Musumeci et al., 2002), and (2) Augustine, where a swarm in December 1996 includes earthquakes with depths to 15 km BSL. Alternatively, magma may be shallow-staged by moving up immediately before the onset of precursory unrest, but at a rate or volume that is too slow/small to produce detectable seismicity. Evidence for slow aseismic magma ascent prior to the 2009 Redoubt eruption includes rotated fast split S-wavelets indicative of mid-crustal dike inflation starting in mid-2008 (Roman and Gardine, 2013). We note that the above staging mechanisms are not mutually exclusive and one or more may be active at a given volcano prior to eruption.

Phase 2 (the 'Destabilization' phase) triggers shallow seismic unrest through a marked increase in pressure in the shallowlystaged magma body. Destabilization may occur through primary magma vesiculation or second boiling because of continued crystallization (e.g., Stock et al., 2016), continued slow (aseismic) intrusion of shallow chamber that eventually exceeds the strength of the shallow host rock, or even quiescent degassing (Girona et al., 2015). Again, these processes are not mutually exclusive. Furthermore, both staging and subsequent shallow destabilization could be accomplished by a relatively small 'quantum' of vanguard magma (i.e., the 'quanta' of Scandone et al., 2007), with the deeper source ultimately providing the bulk of the erupted products.

Phase 3 involves upper-crustal reservoir destabilization and ultimately 'tapping' of that reservoir, leading to seismic unrest over a wide depth range, as well as upward movement and eruption of additional magma, which may by volume represent the bulk of the erupted products. Upper-crustal reservoir destabilization may occur by stress transfer from the shallow volume of staged/destabilized magma through host rock to a largely-disconnected deeper reservoir, or by transmission of a pressure pulse through a continuous mush column (e.g., Cashman et al., 2017; Shapiro et al., 2017; Sparks and Cashman, 2017), and occur either before the eruption

which are not mutually exclusive and can act alone or in combination. The upper red bar indicates the time window during which eruption may begin (i.e., eruption may start any time after the onset of Phase 2).

onset or once the eruption has started (e.g., Tarasewicz et al., 2012).

As a caveat, we note that this conceptual model is based entirely on observations from volcanoes that are relatively frequently active (i.e., volcanoes that erupt every few decades) as this is the reason for the presence of a Level 3 seismic monitoring network prior to the onset of unrest. Thus, the proposed threephase process of magma ascent and eruption may be specific to volcanoes with an already mechanically weakened or remnant conduit and shallow-stored magma from a previous eruption, and not to long-dormant volcanoes awakening for the first time in centuries or millenia. That said, there is evidence from much larger, and more infrequent, eruptions that early migration of magma to shallow levels may have ultimately helped to trigger eruptive activity. For example, the 1980 eruption of Mount St. Helens was preceded by intrusion of magma into the volcanic edifice at least 2 months prior to the eruption (e.g., Scandone et al., 2007), and lacked sufficient seismicity for detection on the regional seismic network. Similarly, the 1991 eruption of Mt. Pinatubo was clearly preceded by upward movement of magma, which first appeared at the surface as a dome, 2 days before the start of explosive activity. Much larger eruptions may also require initial 'priming' by either upward migration of small, and often hotter and/or more primitive magma (e.g., Cerro Galan: Wright et al., 2011; Druitt, 2014) or by rifting-assisted lateral magma migration (e.g., Allan et al., 2012).

#### Implications for Volcano Monitoring and Eruption Forecasting

Our proposed three-phase model for magma ascent and eruption has two main implications for eruption forecasting at volcanoes that erupt relatively frequently. The first is that it may be possible to detect precursory unrest before the onset of seismic activity. We note that at Redoubt, the earliest detected precursors were not seismic unrest, but deformation and increased gas emission. Even at a volcano that is only seismically monitored, aseismic staging via slow initial magma ascent may be detectable with continuous local stress field monitoring around the volcano, which may be expressed as changes in seismic velocity detectable using seismic interferometry or changes in shear-wave splitting of regional earthquakes (e.g., Brenguier et al., 2008; Roman and Gardine, 2013; Rasmussen et al., 2018). Alternatively, where staging is accomplished during the inter-eruptive cycle, as at Mt. St. Helens, relatively small-scale earthquake swarms may be taken as an indicator of the intermediate-term (e.g., yearslong) eruption potential, as they may represent charging of a shallow reservoir. This implication is consistent with recent observations of precursory volcanic unrest worldwide, which indicate longer mean durations for precursory deformation (932 days) and degassing (282 days) than for precursory seismic unrest (197 days) (Phillipson et al., 2013). The second implication is that the onset of deeper seismicity during an unrest episode may function as a short-term indicator of impending eruption, although deeper seismicity preceded eruption at Redoubt and Spurr by almost 2 months, and occurred only after eruption onset at Augustine and MSH 1980, and not at all at Mt. St. Helens 2004. However, additional analysis is needed to fully explore this possibility.

A final point relates to attempts to reconcile geophysical and petrological constraints on pre-eruptive magma storage regions. First, as mentioned above, the shallow magma intrusion responsible for precursory activity may represent only a small fraction of the total erupted volume. Second, it is often preserved only in the early erupted units, and can be dense (degassed and partially crystalline); for this reason, such material is commonly overlooked, with more obvious juvenile vesicular pumice chosen for study. Third, growing evidence that many of the crystals found in magma are antecrysts (were entrained from different parts of the magmatic system) means that experimental preeruptive storage conditions should be designed to determine local, not total, chemical equilibrium (Pichavant et al., 2007). These points are illustrated by our case study volcanoes, where petrological estimates of bulk magma storage often agree with syn- and post-eruptive geophysical estimates from seismicity and deformation.

#### SUMMARY

We compare time-depth patterns of precursory seismicity and petrologically-constrained magma reservoir depths and find that in the majority of six examined cases, initial precursory seismicity is proximal to the vent and several km shallower than estimated reservoir depths. The implication is that precursory seismicity does not reflect initial conduit formation, nor, by implication, reservoir destabilization. Rather, we suggest that reservoir destabilization and initial conduit formation may occur either aseismically, or long before the immediate seismic runup to eruption. We propose a three-stage model of precursory magma ascent and eruption, involving a staging of magma at upper-crustal levels followed by a period of destabilization of the shallowly-staged magma, leading to tapping of deeper portions of the reservoir. Our model implies that the staging phase, if detectable, may provide long-term warning of an eruption compared to the onset of precursory seismicity. Furthermore, it is possible that precursory seismicity may be distinguishable from non-eruptive episodes of unrest by its characteristic shallow depth.

#### AUTHOR CONTRIBUTIONS

Both authors contributed extensively to the work presented in this paper.

#### ACKNOWLEDGMENTS

Thank you to John Power for insights into the Alaskan case studies and to Wes Thelen for providing seismic data for Mt. St. Helens. Thank you also to the two reviewers whose constructive comments greatly improved the manuscript.

### REFERENCES

feart-06-00124 September 1, 2018 Time: 11:2 # 16


volcano from diffusion chronometry and their comparison to monitoring data. J. Volcanol. Geotherm. Res. 288, 62–75. doi: 10.1016/j.jvolgeores.2014. 09.010


of Crater Peak vent, Mount Spurr volcano, Alaska. Bull. Seismol. Soc. Am. 94, 2366–2379. doi: 10.1785/0120030259


**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 Roman and Cashman. 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.

## Forecasting Volcanic Eruptions: Beyond the Failure Forecast Method

Christopher R. J. Kilburn\*

UCL Hazard Centre, Department of Earth Sciences, University College London, London, United Kingdom

Volcano-tectonic seismicity and ground movement are the most reliable precursors to eruptions after extended intervals of repose, as well as to flank eruptions from frequently active volcanoes. Their behavior is consistent with elastic-brittle failure of the crust before a new pathway is opened to allow magma ascent. A modified physical model shows that precursory time series are governed by a parent relation between faulting and elastic deformation in extension, subject to independent constraints on the rate of crustal loading with time. The results yield deterministic criteria that can be incorporated into existing operational procedures for evaluating the probability of crustal failure and, hence, levels of alert during an emergency. They also suggest that the popular failure forecast method for using precursory time series to forecast eruptions is a particular form of the parent elastic-brittle model when rates of stress supply are constant, and that magma transport and crustal fracturing during unrest tend toward conditions for minimizing rates of energy loss.

#### Edited by:

Lauriane Chardot, Earth Observatory of Singapore, Singapore

#### Reviewed by:

Jérémie Vasseur, Ludwig-Maximilians-Universität München, Germany Raffaello Cioni, Università degli Studi di Firenze, Italy Grasso J-r, UMR5275 Institut des Sciences de la Terre (ISTERRE), France

> \*Correspondence: Christopher R. J. Kilburn

#### c.kilburn@ucl.ac.uk Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 28 February 2018 Accepted: 14 August 2018 Published: 07 September 2018

#### Citation:

Kilburn CRJ (2018) Forecasting Volcanic Eruptions: Beyond the Failure Forecast Method. Front. Earth Sci. 6:133. doi: 10.3389/feart.2018.00133 Keywords: elastic-brittle failure, eruption forecasts, failure forecast method - FFM, volcano-tectonic seismicity, ground deformation, alert levels, probability of eruption, bulk failure

#### INTRODUCTION

Most of the world's active volcanoes are not regularly monitored (Tilling, 1995; Sparks et al., 2012), especially those that have been in repose for centuries or more. When such volcanoes reawaken, short-term forecasts of eruption are normally based on data from rudimentary monitoring networks installed after the start of unrest. Local microseismicity, or volcano-tectonic (VT) events, and ground deformation are the pre-eruptive phenomena most frequently detected (Zobin, 2003; McNutt, 2005; Dzurisin, 2007; White and McCausland, 2016) and forecasting strategies use accelerations in both signals to estimate when their rate will reach a critical pre-eruptive value. A popular approach follows the failure forecast method (FFM) developed by Voight (1988) and Cornelius and Voight (1994,1995,1996). However, the best-fit trends for individual time series are not unique and statistical fits can yield ambiguous results (Bell et al., 2011, 2013; Boué et al., 2015; Vasseur et al., 2015). Here we argue that the precursory time series are governed by a deterministic parent relation between seismicity and deformation, subject to independent constraints on how the crust is loaded with time. The FFM then emerges as a particular form of the parent relation when rates of stress supply are constant. Because the parent relation is generic, it can be applied in the absence of information about previous unrest and offers the prospect of enhancing the reliability of forecasts by integrating deterministic estimates of eruption time with existing probabilistic evaluations. The physical basis for the parent relation is reviewed and updated, before it is used to propose new operational procedures for emergency forecasts of eruptions.

repose.

feart-06-00133 September 7, 2018 Time: 13:48 # 2

### SEISMIC AND DEFORMATION PRECURSORS TO ERUPTION

Precursors to eruption describe how magma is able to overcome resistance to its ascent through the crust. When magma can utilize an existing open conduit, resistance is provided by the magma's rheology and friction along the conduit's walls (**Figure 1**). When an existing conduit is not used, resistance is dominated by the need to break open a new pathway for magma ascent; this so-called closed condition applies to long-dormant volcanoes, which have sealed previous conduits, and to volcanoes creating new pathways in addition to an existing conduit, such as flank eruptions through volcanic edifices that already have an open summit vent (**Figure 1**).

A crustal control on opening new magmatic pathways favors similar patterns of precursory behavior, independent of a volcano's tectonic setting, magma composition and style of eruption; similar patterns support the view that short-term forecasts of eruptions are feasible and can be quantified by rock mechanics; and the control of rock mechanics suggests that repeatable patterns are likely to be scale-independent. Voight (1988, 1989) was among the first to promote the importance of scale-independence by recognizing the similarities between precursors to eruptions and rock deformation in the laboratory. He argued that restricted ranges of precursory behavior are driven by a positive feedback between the rate (d/dt) and acceleration (d2/dt 2 ) of a precursory signal with time:

$$d^2\Omega/dt^2 = A(d\Omega/dt)^a\tag{1}$$

where A is a constant and α, which usually lies between 1 and 2, describes the strength of the feedback. As the rate increases, the acceleration becomes larger and increases the rate more quickly than before. The deterministic forecasting potential of the relation depends on estimating the time at which the rate tends to infinity. This singularity is interpreted to represent when a major change occurs in the structure of the deforming rock that ultimately favors an eruption, such as bulk failure and the formation of a pathway along which magma can reach the surface (Voight, 1988; Cornelius and Scott, 1993; Cornelius and Voight, 1994, 1995, 1996; Kilburn and Voight, 1998; De la Cruz-Reyna, 2001; Chastin and Main, 2003; Collombet et al., 2003; Kilburn, 2003; Bell and Kilburn, 2011).

Constrained by the limited test data available at the time, Voight (1988) proposed that individual precursory sequences are characterized by a single value of α, that α can take values from 1 to 2, and that the associated range of accelerations can be expected for any precursory signal related to ground deformation and fracturing. Subsequent studies, however, suggest that α can evolve from 1 to 2 during an individual precursory sequence (**Figure 2**; McGuire and Kilburn, 1997; Kilburn and Voight, 1998; Kilburn, 2003, 2012), and that different signals need not accelerate throughout a full sequence. For example, closed volcanoes and rock-physics experiments show that rates of fracturing can accelerate while the deformation rate remains constant (**Figure 3**; Kilburn, 2012; Robertson and Kilburn, 2016). Under such conditions, Eq. (1) may accommodate the VT event rate, but not the contemporaneous rate of deformation, and so describes only part of the precursory processes that lead to eruption.

Nevertheless, Eq. (1) has been fundamental in advancing the quality of short-term forecasts: it has revealed the control of rates of change in a precursory signal on determining the approach to eruption (rather than a threshold value of the precursor); it has showed that precursory mechanisms are expected to be scale-independent; and it has demonstrated that numerous, independent empirical relations for rock failure share common dynamic constraints (Main, 1999; Turcotte et al., 2003; Ojala et al., 2004; Davidsen et al., 2007; Schmid and Grasso, 2012). In addition, the preferred values of 1 and 2 for α correspond to exponential and hyperbolic increases with time. Both types of increase are common among self-accelerating processes in biological (Monod, 1949) and physico-chemical (Frank-Kamenetskii, 1939; Bowden and Yoffe, 1952; Gruntfest, 1963; Shaw, 1969) systems and so, in retrospect, ought not to be a surprising feature of accelerating rock failure.

#### PRECURSORS CONTROLLED BY AN ELASTIC-BRITTLE CRUST

Precursory unrest is usually detected ∼1–10 months before eruption at closed strato-volcanoes (Zobin, 2003; McNutt, 2005; Dzurisin, 2007; Bell and Kilburn, 2011; White and McCausland, 2016), but may continue for decades at closed large calderas (Robertson and Kilburn, 2016). At strato-volcanoes, it commonly involves as many as ∼10<sup>4</sup> VT events, which occur without systematic changes in hypocentral locations, and ground displacements of ∼0.01–1 m over distances of ∼1–10 km (Kilburn, 2003, 2012; Bell and Kilburn, 2011); at large calderas, the total number of detected VT events can be at least an order of magnitude greater (e.g., ∼10<sup>5</sup> events at Rabaul before its eruption in 1994; Robertson and Kilburn, 2016) and ground displacements reach ∼1–10 m over distances of kilometers (Bellucci et al., 2006; Acocella et al., 2015; Di Vito et al., 2016; Robertson and Kilburn, 2016). Most of the detected VT events have magnitudes of 0–2 and are triggered by the movement of faults ∼10−2–10−<sup>1</sup> km across, or ∼0.1–10% the size of deforming crust. VT events

since 01 January). The mean daily VT event rate (dashed curve) increased exponentially as 16 exp (t/55.6) [data from Shepherd et al. (2002)]. (B) 13 August–30 September (t = days 225–273). Constant mean VT rate at c. 17 events per day [line; data from Gardner and White (2002)]. (C) 01–14 November (t = days 305–318). The mean daily rate increased hyperbolically (curve) as (0.037–0.0024t) -<sup>1</sup> until the eruption [curve; data from Kilburn and Voight (1998) and Kilburn (2003)]. Additional stations were added to the monitoring network between 18 and 31 July [days 200–212; Gardner and White (2002)] Most VT events had magnitudes between 1.3 and 2.7 (Power et al., 1998; Gardner and White, 2002). (D). The three datasets suggest that the mean VT trends evolved with time from quasi-elastic (yellow), through steady inelastic (orange) to accelerating inelastic (magenta); dashed portions of the curve are qualitative interpolations.

and ground deformation can thus be viewed as proxies for the inelastic and total deformation of a crust that contains a dispersed population of small faults.

The trends described by Eq. (1) can be incorporated into a general model by recognizing that deformation and VT events are mutually dependent and that their time series are the result of a parent relation between inelastic and total deformation, constrained by specified changes with time in loading the crust. The parent relation describes the potential for rock to fail when supplied with a mean differential stress Ssup. For each increment of supplied stress, 1Ssup, a proportion 1S is used elastically to deform atomic bonds, while the remainder, 1Sloss, is lost inelastically by breaking bonds:

$$
\Delta \mathcal{S}\_{\text{sup}} = \Delta \mathcal{S} + \Delta \mathcal{S}\_{loss} \tag{2}
$$

The total elastic component determines the mean, or bulk, differential stress S that is established across the deforming volume V of crust and, hence, also the amount of elastic energy being stored. Under idealized conditions, the dominant mode of deformation is initially elastic (without seismicity; 1Ssup = 1S) and, once faulting has begun, evolves with increasing differential stress from quasi-elastic to inelastic (**Figure 4**), as the additional energy being supplied is at first mostly stored elastically (quasielastic, 1Ssup ≈ 1S) and, later, mostly consumed by faulting (inelastic, 1Ssup ≈ 1Sloss). In the quasi-elastic regime, the mean deformation approximately increases in proportion to S until it reaches its maximum value SF, when the amount of stored energy has reached its maximum capacity. Any additional stress (e.g., from a pressurizing magma body) is consumed in fracturing and fault movement and deformation continues under a constant maintained stress, which defines the inelastic regime. Bulk failure begins within one or more portions of the crust, where the stresses at fracture tips remain large enough to persistently overcome rock resistance and so allow fractures to grow and link together (Griffith, 1921).

Deformation, rather than stress, is measured in the field, so that the quasi-elastic and inelastic regimes have to be identified indirectly from relations between VT seismicity and total ground movement. In the quasi-elastic regime, the number of VT events (a measure of inelastic deformation) is expected to accelerate with the total amount of ground movement (a measure of total deformation) as the proportion of inelastic deformation becomes larger. In the inelastic regime, net deformation continues by faulting, so that the ground movement is expected to increase in proportion to the number of VT events. A complete failure sequence is thus expected to show an accelerating and then

1.17 exp (t/23.3); the mean daily VT event rate increases 0.05 exp (t/23.3). The combination of exponential VT rate and constant deformation rate is consistent with quasi-elastic brittle behavior under a constant rate of stress supply (Table 1). The eruption occurred almost immediately after the end of the quasi-elastic regime when t/23.3 ≈ 4.

linear increase in the number of fracturing events with ground movement (**Figure 4**).

**Figures 4**, **5** illustrate how similar failure sequences have been observed before eruptions under a wide range of closed conditions. The examples are from Kilauea, in Hawaii, El Hierro, in the Canary Islands, and Rabaul, in Papua New Guinea. The sequences continued for about 0.22–0.25 years (approximately 3 months) at Kilauea (Nakata, 2006; Bell and Kilburn, 2011) and El Hierro (Istituto Geográfica Nacional [IGN], 2011; Sagiya, 2011; Kilburn et al., 2017) and for 23 years at Rabaul (Johnson et al., 2010; Robertson and Kilburn, 2016); they include both strato-volcanoes (Kilauea and El Hierro) and large calderas (Rabaul); and the resulting eruptions ranged from basaltic effusive at Kilauea (Swanson et al., 1971) to andesitic phreato-Plinian at Rabaul (Rabaul Petrology Group, 1995). All three VT-deformation trends show the expected evolution from quasielastic to inelastic behavior, independent of the type of volcano, the duration of precursor, magma composition, style of eruption and, as discussed below, how VT event rate and deformation rate vary with time. Such similarity is compelling evidence that elasticbrittle failure of the crust determines the pattern of seismic and deformation precursors to eruptions at closed volcanoes.

Elastic-brittle failure additionally favors the scale-independent behavior implicit by Voight's analysis (Voight, 1988). VT events are produced by fault slip; slip is accommodated by extending the margin of a fault; and fault extension is controlled by cracking within the zone of stress concentration, or process zone, that develops around a fault's tip (Atkinson, 1984; Main and Meredith, 1991; Lockner, 1993; Cowie and Shipton, 1998). Smaller discontinuities (both cracks and slip surfaces) grow in the process zone until it is broken, so extending the parent fault. Similarly, the discontinuities grow by cracking at a still smaller scale around their own process zones. Whatever the scale of observation (e.g., in the field or laboratory), slip can be viewed as an hierarchical process, in which the larger movements are the result of cracking at successively smaller scales. It is therefore expected that large-scale movements can ultimately be described by processes operating contemporaneously at the smallest scale (Kilburn, 2003, 2012).

#### QUANTIFYING REGIMES OF DEFORMATION

#### Parent Relations for Bulk Failure

Bulk failure begins when fracture growth becomes selfperpetuating within part of the stressed crust. It may occur when the failure stress, SF, is first achieved, or after an extended interval of inelastic deformation under a stress maintained at SF, during which the externally supplied stress replaces the stress lost by fracturing and fault movement, so resembling rock creep (**Figure 4**).

Applying well-known methods from classical statistical mechanics (Reif, 1985; Ruhla, 1992; Guénault, 1995), the magnitudes of local stresses are expected to follow a Boltzmann distribution about the bulk value, for which inelastic deformation εin increases with supplied stress Ssup as (Kilburn, 2003, 2012):

$$d\varepsilon\_{in}/d\mathcal{S}\_{sup} = (d\varepsilon\_{in}/d\mathcal{S}\_{sup})\_a \exp(-\mathcal{S}\_{act}/\mathcal{S}\_{ch})\tag{3}$$

where the activation stress Sact is the additional stress required to initiate bulk failure, the characteristic stress Sch is the atomic free energy per volume available for deformation, (dεin/dSsup)<sup>a</sup> is the rate at which natural fluctuations in atomic configuration attempt to initiate failure, and exp (−Sact/Sch) measures the probability that an attempt is successful (Reif, 1985; Ruhla, 1992; Guénault, 1995).

The activation stress is SF−S, the difference between the failure and applied differential stresses. In the quasi-elastic regime, it decreases to zero as S<sup>F</sup> is approached and all local attempts to fracture are successful [dεin/dSsup = (dεin/dSsup)a]. Fracturing continues to accelerate in the inelastic regime, owing to the local redistribution of stress around fractures (Lawn, 1993; Valkó and Economides, 1995). The redistribution concentrates stress at fracture tips and, as long as the bulk stress is maintained at SF, increases the stress remaining after each increment of growth. A decreasing proportion of Ssup is required to continue fracture growth, so that the interval between growth steps persistently decreases, leading to runaway growth when the interval becomes infinitesimally small.

If 1S<sup>R</sup> is the mean amount of stress redistributed around fractures, then, from the viewpoint of fracture tips, the activation energy becomes (SF−1SR)−S = (SF−1SR)−S<sup>F</sup> = −1SR. The negative sign indicates the local surplus of stress at fracture tips (which is balanced by a local deficit where stress has been transferred from around the sides of a fracture). Although the failure strength is not changed, it appears to be decreasing by an amount equal to 1S<sup>R</sup> and so mimics fracturing promoted by progressive rock weakening. Interpretations that have qualitatively related accelerated fracturing to stress corrosion at fracture tips (Anderson and Grew, 1977; Atkinson, 1984; Main and Meredith, 1991; Lockner, 1993; McGuire and Kilburn, 1997; Kilburn and Voight, 1998; Kilburn, 2003) may thus alternatively be viewed in terms of increasing stress concentration around fractures in rock with unchanged strength. Setting εin proportional to the number N of VT events, Eq. (3) yields for the quasi-elastic regime:

$$dN/d\mathbb{S}\_{sup} = (dN/d\mathbb{S}\_{sup})\_a \exp\{ (\mathbb{S} - \mathbb{S}\_F)/\mathbb{S}\_{ch} \} \tag{4}$$

and for the inelastic regime:

feart-06-00133 September 7, 2018 Time: 13:48 # 5

$$dN/d\mathcal{S}\_{sup} = (dN/d\mathcal{S}\_{sup})\_a \exp(\Delta \mathcal{S}\_{\mathbb{R}}/\mathcal{S}\_{\text{ch}}) \tag{5}$$

Equations (4 and 5) are the basic expressions for deriving relations between VT event rate and both deformation and time. A first comparison with Eq. (3) suggests that the probability of bulk failure is proportional to exp [(S−SF)/Sch] and to exp (1SR/Sch) in the quasi-elastic and inelastic regimes, respectively. The exponential terms, however, do not represent the same type of probability. The quasi-elastic regime describes conditions before the start of bulk failure, so that exp [(S−SF)/Sch] measures the probability that bulk failure will begin under an applied differential stress S. As a result, the probability becomes one when S = SF. The inelastic regime, in contrast, describes when the failure process will be completed, so that exp (1SR/Sch) measures the probability that failure will occur within a given time interval.

#### The Characteristic Stress

The value of the characteristic stress in Eqs (4 and 5) changes with the conditions of loading. Independent of the applied stress, atoms deform naturally as they oscillate around their equilibrium positions. The oscillations may randomly yield configurations that assist stress-induced deformation and so promote failure at a stress smaller than SF. The total free energy per volume for oscillations is S <sup>∗</sup> = (3φT+Pc−Pp)/3, where T is absolute temperature (K), P<sup>c</sup> and P<sup>p</sup> are the confining and porefluid pressures, and the molecular energy per unit volume per temperature, φ, has a notional value of (7 ± 1) × 10<sup>4</sup> J m−<sup>3</sup> K −1 for common silicate rocks (Kilburn, 2012). How much of the total free energy is utilized determines the characteristic stress. Failure in compression is limited by shearing between atoms, but in tension by the tensile failure of bonds. Shearing requires the integrated deformation of all bonds and so utilizes the full amount of S ∗ . Tension, however, requires deformation only among bonds in a particular direction and so utilizes a fraction of S ∗ ; the amount utilized in tension defines the rock's tensile strength, σT.

The condition that Sch = S ∗ in compression has been verified against laboratory experiments (Kilburn, 2012). It was also applied to VT precursors of the 1991 eruption of Pinatubo, in the

FIGURE 4 | (A) The amount of deformation due to fault movement is given by the difference between the equilibrium (unbroken curve) and ideal elastic (broken line) deformation trends. As the differential stress increases to its failure value S<sup>F</sup> , behavior evolves from elastic (white), through quasi elastic (yellow) to inelastic (magenta). This yields (B) accelerating quasi-elastic and linear inelastic increases in VT event with deformation. (C) For a constant rate of supplied stress, the deformation rate remains constant until a hyperbolic increase during inelastic behavior. The hyperbolic increase may be preceded by an interval of steady inelastic VT rate (orange). (D) The contemporaneous VT event rate increases from exponentially to hyperbolically with time, again possibly separated by an interval at constant rate (see Figure 2). (E) The model VT-deformation trends are consistent with normalized pre-eruptive sequences observed at El Hierro, Canary Islands, 18 July–12 October 2011 [squares; data from Istituto Geográfica Nacional [IGN] (2011) and Sagiya (2011)], Rabaul, Papua New Guinea, 1971–1994 [circles; data from Robertson and Kilburn (2016)], and Mauna Ulu on Kilauea, Hawaii, 14 November 1971–04 February 1972 [triangles; data from Bell and Kilburn (2011)]. See also Figure 3. Measures of ground movement (λ; displacement and tilt) and their characteristic values (λch) are used as proxies for deformation. In all three cases, inelastic behavior (magenta) began when λ/λch ≈ 4 and eruptions occurred (violet symbols) when λ/λch ≥ 4. The trends were normalized using the total number of VT events at the end of the quasi-elastic regime, 6NQE,m, and the characteristic ground movement λch. See Figure 5 for the non-normalized trends. Typical magnitude ranges for VT events were 0.5–2.0 for Rabaul, 1.0–2.5 for El Hierro and 1.5–4.0 for Mauna Ulu.

FIGURE 5 | Changes in the total number of VT events with deformation for precursors to eruption at (A) Rabaul, Papua New Guinea, 1971–1994 [data from Robertson and Kilburn (2016)]: (B) El Hierro, Canary Islands, 18 July–12 October 2011 [data from Istituto Geográfica Nacional [IGN] (2011) and Sagiya (2011)]; and (C) Mauna Ulu, Kilauea, Hawaii, 14 November 1971–04 February 1972, [data from Bell and Kilburn (2011)]. The quasi-elastic regime shows exponential trends between VT number and the proxy for deformation, yielding characteristic values of 0.53 m for uplift at Rabaul [6N ∝ exp (λ/0.53)], 5 mm for displacement at El Hierro [6N ∝ exp (λ/5)] and 4.5 rad for tilt at Kilauea [6N ∝ exp (λ/4.5)]. The inelastic regime shows linear trends. The data are consistent with transition between regimes (star) when the ratio of total deformation to characteristic value (λ/λch) is about 4. See Figure 4 for the normalized trends.

Philippines, and 1995 eruption of Soufriere Hills, on Montserrat (Kilburn, 2003), implying the unlikely scenario of eruptions through crust in compression. Extensional stresses are instead anticipated in crust being deformed by a pressurizing magma body and, as demonstrated below, the precursory sequences at Pinatubo and Soufriere Hills can be explained more simply by equating Sch instead with the tensile strength.

#### APPLICATIONS TO FIELD DATA

#### Changes in Seismicity With Deformation

Equations (4 and 5) describe the change from an exponential to linear increase in VT number with deformation, corresponding to the evolution from quasi-elastic to inelastic behavior. In the quasi-elastic regime, 1Ssupp ≈ 1S [Eq. (2)] and both Ssup and S can be set to Eε, the product of Young's modulus and bulk deformation. If ground movement, λ, is proportional to the bulk deformation, Eq. (4) yields exponential increases with ground movement of both the change in VT seismicity and the total number of VT events [**Table 1**, Eqs (T1) and (T2)]. In the inelastic regime, the additional elastic strain supplied by Ssup is consumed in faulting (Section 3). Changes in N and λ now both measure VT events, so dN/dλ is constant and total VT number increases linearly with ground movement [**Figures 4**, **5** and **Table 1**, and Eq. (T3)].

The transition from quasi-elastic to inelastic behavior occurs when the applied differential stress reaches its failure value, SF. Since Sch = σ<sup>T</sup> for tensile deformation, the transitional value for S/Sch in Eq. (4) is expected to be SF/σ<sup>T</sup> for crust being stretched. Applying the Mohr–Coulomb–Griffith criterion for bulk failure, SF/σ<sup>T</sup> is four or less in tension, between 4 and 5.6 in extension (combined tension and shear) and 5.6 or more in compression (**Figure 6**; Secor, 1965; Shaw, 1980; Sibson, 1998). In the quasielastic regime S is proportional to ground movement, so that SF/σ<sup>T</sup> can be obtained from λ/λch at the end of the regime. The examples in **Figures 4**, **5** are consistent with values for λ/λch of about four and, hence, with bulk failure in tension.

#### Quasi-Elastic Changes in Seismicity and Deformation With Time

The VT-deformation trends describe how the proportion of inelastic movement varies with total deformation. To transform the trends into time series, independent constraints must be introduced to specify how loading changes with time. The simplest condition is a constant rate of stress supply. Under this condition, Eq. (4) for quasi-elastic behavior leads to exponential increases with time in VT rate and the total number of VT events [**Table 1**, Eqs (T4)-(T5)]. The contemporaneous rate of ground movement is constant because Ssup is proportional to deformation [**Table 1**, Eq. (T8)]. Quasi-elastic behavior at a constant deformation rate thus naturally yields the exponential VT-time trends frequently observed at closed volcanoes (**Figures 2**, **3**; Shepherd et al., 2002; Bell and Kilburn, 2011; Kilburn, 2012; Wall, 2014).

Under a constant deformation rate, the ratio λ/λch is measured by t/tch, where tch is the characteristic timescale of the exponential VT time series. If T is the duration of the full exponential sequence, then T/tch = SF/σT. Observations at closed volcanoes suggest preferred values for T/tch of 2–4 (**Figure 6**). The range is consistent with the limiting values for λ/λch obtained

from the VT-deformation trends (**Figures 4**, **5**), so reinforcing the interpretation of deformation in tension (**Figure 6**).

The particular value of SF/σ<sup>T</sup> in tension changes with the geometry of failure. The generic Mohr–Coulomb–Griffith criterion applies to failure along a plane (e.g., creating a new failure plane or pulling opening a sealed fault) and values for SF/σ<sup>T</sup> are ≤4. Pressurized bodies, in contrast, rupture their margins at values of SF/σ<sup>T</sup> that change according to their shape; for example, SF/σ<sup>T</sup> is three for a sphere, but two for a long vertical cylinder (Jaeger, 1969; Saada, 2009). The preferred field values of 2–4 are thus consistent with the onset of bulk failure at the margins of magma bodies, as well as the opening of healed faults in crust being stretched by those bodies. In all cases, tensile bulk failure requires the effective principal stresses (normal stress−pore-fluid pressure) to be less than three times the tensile strength (**Figure 6**). For a notional strength of 10 MPa, therefore, the effective principal stress cannot exceed about 30 MPa, which corresponds to maximum lithostatic depths of about 1.2 and 2 km in dry and water-saturated crust. Tensile failure at greater depths thus implies deformation of super-saturated crust with pore-fluid pressures greater than hydrostatic (Shaw, 1980).

#### Inelastic Changes in Seismicity and Deformation With Time

In the inelastic regime, a constant rate of stress supply generates hyperbolic increases in VT and deformation rates with time [**Table 1**; Eq (T9)]. After a time τ, the rates tend to an infinite value, which is taken to be the mathematical equivalent of continuous fracture coalescence and successful bulk failure (Voight, 1988; Kilburn, 2003); τ therefore defines the duration of accelerating inelastic deformation, which is analogous to tertiary brittle creep under a constant external stress.

The hyperbolic trends are equivalent to linear decreases with time in the inverse VT and deformation rates. Applied to VT seismicity, principal episodes of fracture coalescence coincide with local peak rates (dNp/dt) - or local minima in inverse-rates [(dNp/dt) −1 ] - during which the mean increase in stress concentration per event is B(S<sup>F</sup> 2 /E), where E is Young's modulus and the geometric constant B = π for straight-edged fractures (Kilburn, 2003). Hence 1S<sup>R</sup> = B(S<sup>F</sup> 2 /E)N<sup>p</sup> and so, after manipulation, the parent relation for inelastic deformation [Eq. (5)] yields dNp/dt = (dNp/dt)<sup>a</sup> exp [B(S<sup>F</sup> 2 /E)Np/σT] for a constant rate of stress supply in extension. This form of relation can be re-expressed to show changes in VT event rate with time (Voight, 1988), leading to:

$$(dN\_\mathcal{P}/dt)^{-1} = (dN\_\mathcal{P}/dt)\_{0,IN}^{-1} - \gamma^\*(t - t\_0, \boldsymbol{\chi}) \tag{6}$$

or, alternatively, to:

$$(dN\_p/dt)^{-1} = \left(dN\_p/dt\right)\_{0,IN}^{-1} \left[1 - (t - t\_0, \, \_{IN})/\mathfrak{x}\right] \tag{7}$$

where the subscript 0,IN denotes values at the start of the inelastic regime.

The gradient for the inverse-rate VT minima in Eq. (6) has a magnitude γ <sup>∗</sup> = B(SF/σT) 2 (σT/E) and units of "per VT event" (**Table 1**). For common crustal rocks, σT/E∼10−<sup>4</sup> (Mogi, 2007; Heap et al., 2009) and so, with B = π and SF/σ<sup>T</sup> between 2 and 4 (**Figure 6**), the gradient is expected to be ∼10−3−10−<sup>2</sup> per event, which is consistent with the values of 6.4 × 10−<sup>3</sup> and 2.4 × 10−<sup>3</sup>


B, Geometric factor for cracks growing in the inelastic regime; E, Young's modulus; L, Reference length for determining mean deformation; N, Number of VT events; Nd, Number of detected VT events; Np, Value of local peaks in the number of VT events; S, Mean maintained differential stress; Sch, Characteristic stress or free energy per unit volume (in extension, Sch is the tensile strength, σT); SF, Mean failure stress; Ssup, Mean supplied differential stress; 1SR, Mean stress redistributed around fractures; t, time; tch, characteristic time for exponential trends, Sch/(dSsup/dt); β, Mean ground movement per VT event; ε, Mean strain; γ ∗ , B(SF/σT) 2 (σT/E); λ, Measure of ground movement as proxy for mean deformation; λch, Characteristic λ for exponential trend = (Sch/E)L; λp, Value of local peaks in ground movement; τ, Duration in the inelastic regime of acceleration to infinite rates of seismicity and deformation, [(1SR/Sch)(dNp/dt)0,IN] −1 , [(1SR/Sch)(dλ/dt)0,IN(1/β)]−<sup>1</sup> . Subscripts. a, Attempt value; 0, Value at start of a trend; 0,IN, Value at start of acceleration in the inelastic regime.

per event found for inverse-rate minima at Pinatubo in 1991 and Soufriere Hills, Montserrat, in 1995 (**Figure 7**; Kilburn, 2003).

In Eq. (7), the duration of inelastic acceleration is given by τ = [γ ∗ (dNp/dt)]−<sup>1</sup> [so when (t−t0,IN) = τ, the inverse-VT rate tends to zero and the VT rate tends to infinite values]. At andesitic strato-volcanoes, observed values of dNp/dt are ∼10 events per day, which, with typical values of γ ∗ , give the observed durations on the order of 10 days (Kilburn, 2003). Hence, if about 30–50% of the total time is needed to confirm a hyperbolic increase in rate, realistic warning times at such volcanoes are on the order of days.

#### ELASTIC-BRITTLE FAILURE AND VOIGHT'S RELATION

Equations (4 and 5) quantitatively capture the essential features of VT-deformation trends before eruptions. Their good agreement with field data confirms the mutual dependence of VT and deformation precursors and suggests that (a) precursory deformation evolves from quasi-elastic to inelastic, (b) when the rate of stress supply is constant, the increase in VT event rate with time changes from exponential (quasi-elastic behavior) to hyperbolic (inelastic behavior), and (c) precursory behavior can be interpreted in terms of elastic-brittle deformation, without the need to invoke additional rheological responses in the crust (such as plastic flow); this does not preclude the operation of additional processes, only that they need not be invoked unless field data indicate otherwise.

For a constant rate of stress supply, Equations (4 and 5) yield time series (Eqs. T4 and T6) that can be re-expressed as

$$d^2N/dt^2 = (1/t\_{dh})(\chi^\*t\_{dh})^{\alpha-1}(dN/dt)^{\alpha} \tag{8}$$

which has the form of Voight's original FFM Relation [Eq. (1)].

The exponential and hyperbolic VT rates coincide with Voight's relation for α = 1 and 2, but it has yet to be confirmed whether intermediate values of α have any physical meaning. Indeed, it is possible that intermediate values are artifacts from seeking best-fit single trends to partial data sets that extend across the quasi-elastic and inelastic regimes. Comparison with Eq. (1) also shows that the term A in Voight's FFM Relation is not a constant, but changes from 1/tch for α = 1 to γ ∗ for α = 2; in other words, 1/A is a characteristic timescale when α = 1, but a characteristic number of VT events when α = 2.

Applied to contemporaneous changes in rates of ground movement with time, an expression similar to Eq. (8) is available only for inelastic behavior, when the rate increases hyperbolically (α = 2) to yield d2λ/dt <sup>2</sup> = (γ ∗ /β)dλ/dt (where β is the mean ground movement per VT event). Hence, the deformation and VT rates both favor α = 2 immediately before bulk failure, a feature noted in Voight's original analysis (Voight, 1988). During quasi-elastic unrest, however, the deformation rate remains constant while the VT event rate increases and so Voight's relation is not equally applicable to the two precursory signals.

When the rate of stress supply varies with time, the VT and deformation time series derived from Eqs (4 and 5) no longer follow Voight's relation. Hence, the fact that the FFM has

FIGURE 6 | (A) The Mohr–Coulomb–Griffith representation of bulk failure in terms of applied normal and shear stresses. The outer dashed curve shows the parabolic form of the failure envelope [for mathematical treatments, see Secor (1965); Shaw (1980), and Sibson (1998)]. The Mohr circles (continuous and dashed semi-circles) describe states of stress between the maximum and minimum effective normal stresses; positive values are in compression, negative values in tension. The diameter of a circle gives the applied differential stress, S. Failure occurs when the minimum normal stress meets the failure envelope. Tensile failure corresponds to where the circles meet the envelope at zero shear stress (hence intersect the horizontal axis) and the minimum effective normal stress equals the tensile strength (with a normalized value of –1). To satisfy this condition, the Mohr circle must have a diameter less than or equal to 4σ<sup>T</sup> (or normalized diameter of four; arrows beneath graph) The maximum value (larger continuous semi-circle) may represent the opening of a failure plane. For comparison, failure at the margins of a long vertical cylinder is shown with a normalized diameter of two (smaller continuous semi-circle). The dashed semi-circle (normalized diameter of 5.6) shows the Mohr circle at the transition between failure in extensional (tension and shear) and compression (shearing alone). (B) Values of T/tch for quasi-elastic pre-eruptive behavior at Etna, Sicily (blue squares; Wall, 2014), Kilauea, Hawaii (green triangles; Bell and Kilburn, 2011), Mauna Loa, Hawaii (magenta diamond; Lengliné et al., 2008), Piton de La Fournaise, Réunion (small dark red triangle; Lengliné et al., 2008) and Soufriere Hills, Montserrat (black diamond; Figure 2). Most of the values for T/tch lie between 2 and 4 (lower and upper dashed lines).

been found to describe numerous precursory sequences at closed stratovolcanoes (e.g., Cornelius and Voight, 1994, 1996; Kilburn and Voight, 1998; De la Cruz-Reyna, 2001; Kilburn, 2003; Bell and Kilburn, 2011; Wall, 2014; Boué et al., 2015) suggests that

such sequences may often develop under a constant rate of stress supply.

Pre-eruptive sequences at strato-volcanoes commonly have durations of years or less. Longer sequences provide a greater opportunity for significant variations in the rates of stress supply, in particular the decadal time intervals for precursory sequences at large calderas, such as Rabaul (Robertson and Kilburn, 2016) and Campi Flegrei (Kilburn et al., 2017). For example, the full 23-year sequence of unrest before Rabaul's 1994 eruption did not show changes with time in rates of VT event and of uplift consistent with Voight's relation - even though it followed the trend between VT events and deformation expected for an evolution from quasi-elastic to inelastic behavior (**Figures 4**, **8**). Thus the VT event rate and uplift rate both fluctuated by an order of magnitude or more (∼102−10<sup>4</sup> events per month and c. 0.015−0.4 m y−<sup>1</sup> ), and the peak rates occurred half-way through the sequence (**Figure 8**). During its final 2 years, however, the sequence did show changes in rates that were consistent with hyperbolic increases with time (**Figure 8**).

The peak rates in the Rabaul sequence have been associated with the arrival of magma at a depth of about 2 km, possibly intruding into an existing magma chamber (McKee et al., 1984).

FIGURE 8 | (A) VT event rate (columns) and uplift rate (gradient of dashed curve) both fluctuated unevenly with time during the 23-year approach to Rabaul's eruption in 1994. (B) The change in number of VT events with uplift shows the evolution from accelerating quasi-elastic (yellow) to linear inelastic (magenta) behavior (a normalized version is shown in Figure 4). The dashed green lines show conditions corresponding to the 1983–1985 crisis when rates reached their peak values (see A). Note how the expected VT-uplift trends persist in spite of significant variations with time in rates of seismicity and ground movement. (C) The inverse mean rate of uplift decreased linearly with time during 1992–1994. The best-fit trend (dashed) is [Inverse Mean Rate of Uplift] = 12784–14.38t (r <sup>2</sup> = 0.99) for inverse-rates in days m-<sup>1</sup> and time t in days; the trend corresponds to a hyperbolic increase in the mean rate of uplift. The contemporaneous inverse VT rates showed a less clearly defined trend (Robertson and Kilburn, 2016).

The final rates have instead been associated with segments of the ring fault being torn open to allow magma to erupt (Robertson and Kilburn, 2016). The full sequence thus represents the caldera being stretched to breaking point under a variable rate of increasing magmatic pressure. A constant rate of stress supply

was established during the final 2 years of deformation, when the crust had already entered the inelastic regime, so promoting the hyperbolic increase in rates with time (described by the Voight's relation with α = 2). However, rates of stress supply were not constant for the first 20 years of unrest, during which time Voight's relation could not be used to describe the precursory time series.

#### DEVIATIONS FROM MODEL ELASTIC-BRITTLE CONDITIONS

By describing common trends among VT and deformation precursors, Eqs (4 and 5) provide a starting point also for identifying when the model cannot be applied without adjusting its underlying assumptions. For example, the model assumes that the crust's mean behavior follows that of an elastic medium containing a dispersed population of small faults. It does not explicitly accommodate conditions when bulk deformation is controlled by movements of faults with lengths similar to that of the crust being deformed. Slip along such faults would favor the occurrence of a small number of large-magnitude VT earthquakes instead of a large number of small VT events; as a result, VT trends may be dominated by only a few earthquakes, whose total number is too small to yield repeatable mean behavior.

Even in the presence of large faults, VT seismicity may remain controlled by small earthquakes. For example, before Rabaul caldera's 1994 eruption, VT seismicity was concentrated in an annulus related to ring faults extending to depths of 2–4 km and produced by caldera collapse about 1,400 years beforehand (Nairn et al., 1995; Jones and Stewart, 1997). Most of the VT events had magnitudes of 0.5–2, triggered by the slip of small faults in the crust surrounding the ring faults, rather than movements along the ring faults themselves (Robertson and Kilburn, 2016). Similar behavior has been argued for VT seismicity in the vicinity of ring faults at the Campi Flegrei caldera in southern Italy (Troise et al., 2003). In both cases, the large ring faults appear to have constrained the size and geometry of the crust being deformed, rather than to have contributed directly to rates of VT seismicity.

The initial application of Eqs (4 and 5) implicitly assumes that self-accelerating crack growth begins at the start of the inelastic regime. In this case, precursory VT event rates are anticipated to accelerate with time across the transition from quasi-elastic to elastic regimes. Such behavior is indeed observed in the field (Bell and Kilburn, 2011) and laboratory (Kilburn, 2012). However, the transition may also be marked by an interval of constant VT event rate between the quasielastic and inelastic accelerations (**Figure 2**). Such steady rates may develop in the inelastic regime when the coalescence of fractures, which favors accelerating rates, is temporarily retarded by stress barriers in rock between non-interacting fractures (Main, 2000; Heap et al., 2009; Girard et al., 2010). An accelerating transition may thus represent the limiting case when fracture coalescence dominates from the start of the inelastic regime. Hence, should a steady rate appear after quasi-elastic acceleration, it cannot be taken to indicate an approach to stability, but must be viewed as a temporary lengthening of unrest until the onset of hyperbolic accelerations to bulk failure (**Figure 2**).

Although bulk failure in the crust is necessary before eruption at a closed volcano, it does not ensure that an eruption will occur (Kilburn, 2003; Grasso and Zaliapin, 2004). For example, similar precursory sequences have been identified before eruptions and intrusions at Kilauea (Bell and Kilburn, 2011) and at Krafla in Iceland (Blake and Cortés, 2018). Failure may be initiated at the margin of a magma body, after which an eruption requires a magmatic overpressure large enough to drive magma to the surface (Jellinek and DePaolo, 2003). It may also begin at a more distant location where stresses are concentrated (e.g., the tips of a major fault) or where the crust is locally weak (e.g., altered rocks in a hydrothermal system), in which case the propagating fracture must additionally extend to the magma body itself. The probability of eruption can thus be expressed more generally as the product of the probability of bulk failure, the probability that failure breaches a magma body, and the probability that magma can erupt through the new breach. The elastic-brittle model addresses only the first of these and so, by itself, is a necessary but insufficient condition for guaranteeing an eruption.

In common with most analyses of VT and deformation precursors, the elastic-brittle model does not specify conditions within the magmatic, or other, systems driving volcanic unrest. One exception is the model by Lengliné et al. (2008), who investigate how specific conditions of magma accumulation may control rates of precursory signal at basaltic volcanoes. The acceleration in VT number with time may also increase with the size of the following eruption (Schmid and Grasso, 2012). Future models that couple crustal stresses with magmatic processes promise to yield new insights for refining and enhancing forecasting procedures.

#### SCALE-INDEPENDENT FRACTURING

The elastic-brittle model is inherently scale-independent, because it expresses rates of VT seismicity and deformation in terms of the ratio of mean applied differential stress to characteristic stress. Applied stress measures changes in the free energy per volume available for atoms to do work (in this case to deform and break bonds); the characteristic stress measures the total free energy per volume already available for a specific type of deformation (e.g., tension or compression) before a differential stress is applied. Thus, if 1e is the change in free energy and e 0 the free energy available, then S/Sch = (1e/V)/(e 0 /V), where V is the volume of deformation. At the atomic scale, V is the volume of an atom itself. At the macroscopic scale, it refers to the size of rock being deformed, including the volume of discontinuities. At any given scale, the same V is used to calculate both stresses from their respective energy terms, so that S/Sch always measures the ratio of change in free energy to the reference free energy available. Hence, even though the values of individual stress terms may change with scale and amount of fracturing, their ratios remain the same.

The influence of fracturing is shown by the change from quasi-elastic to inelastic deformation regime, which appears to correspond to a change from weak to strong interactions among fractures (Vasseur et al., 2017) and favors the onset of self-accelerating crack growth. Exploring the interactions has generated an independent class of fracture models, most of which have focused on strong interactions immediately before bulk failure (Main, 1996, 2000; Molchan et al., 1997; Guarino et al., 1998; Amitrano et al., 2005; Girard et al., 2010). Applying statistical methodologies, they have shown that increases in seismic event rate with time in the inelastic regime are expected to follow a power-law trend of the form:

$$dN/dt \propto \left[1 - (t - t\_0)/\tau\_f\right]^{-\rho} \tag{9}$$

where t<sup>0</sup> is the starting time, τ<sup>f</sup> is the duration of the sequence before bulk failure and p is an undetermined empirical factor. The result is scale-independent because it depends on the interaction among fractures, regardless of size, and values of p between 0.5 and 2 have been obtained from field data (Amitrano et al., 2005; Schmid and Grasso, 2012). When p = 1, the result is identical to that obtained for the acceleration to failure in the inelastic regime [Eq. (7)] and corresponds to the limit of α = 2 in Voight's analysis (Voight, 1988). The similarity of Eqs (7) and (9) suggests agreement between the different modeling strategies. Their future integration may thus yield enhanced procedures for general forecasts of bulk failure.

#### PRECURSORS AS INDICATORS OF ENERGY LOSS BEFORE ERUPTIONS

In addition to demonstrating the elastic-brittle features of unrest, precursory time series indicate preferred energy states in magmatic systems before eruption. Common quasi-elastic behavior suggests a preference for steady rates of deformation, which are a natural consequence of minimizing rates of energy dissipation. Primary controls on energy dissipation are VT seismicity in the crust and frictional resistance during the transport of magma. Under a constant deformation rate, the VT event rate increases exponentially with time and so is not an obvious minimizing factor. A constant deformation rate, however, is favored by a constant rate of pressurization in the magmatic source and this, in turn, is promoted by a constant flux of magma from depth. Approximately steady deformation rates may thus reflect a preference to minimize rates of energy loss during magma transport.

probability that bulk failure begins in the quasi-elastic regime to the probability that failure will be complete before the time interval τ in the inelastic regime.

A second remarkable feature is the agreement between theory and observed inverse-VT rate gradients in the inelastic regime. The model VT expressions implicitly assume that the number of detected events is similar to the number of essential events for bulk failure. In principle, however, fault movement may be triggered in parts of the crust that do not directly affect the volume that will ultimately fail, so that the number of detected events is greater than the essential number. The difference will not affect expressions for VT events in the quasi-elastic regime, because it changes only pre-exponential terms and so cancels from both sides of the relevant equations. In the inelastic regime, however, N appears within the exponent itself [Eq. (T6)], so that the inverse-rate field gradient is given by γ <sup>∗</sup>Nes/N - that is, the model gradient multiplied by the ratio of the number of essential (Nes) to detected (N) VT events - which is ≤γ ∗ . At Pinatubo and Soufriere Hills, Montserrat, the observed gradients are within the expected theoretical range, suggesting that the proportion of non-essential VT events is small. Although such a condition may not be universal, the fact that it appears for these two examples implies that precursory deformation may tend to involve the minimum amount of crust possible for initiating bulk failure, which again indicates a preference for minimizing energy loss.

#### POTENTIAL APPLICATION TO OPERATIONAL FORECASTS

Event trees have become a popular aid for deciding levels of alert during a volcanic emergency (Newhall and Hoblitt, 2002; Aspinall, 2006; Sandri et al., 2009; Marzocchi and Bebbington, 2012; Selva et al., 2012; Sobradelo et al., 2013, 2014). They pose a sequence of interconnected questions about interpreting precursory signals, the combined answers to which are used to estimate the probability of eruption. Critical pre-eruptive values for individual signals (e.g., number of VT events and amount of ground deformation) are identified empirically from observations of previous eruptions at the same volcano or at apparently analogous volcanoes elsewhere (Sobradelo and Martí, 2015).

The progression from quasi-elastic to inelastic behavior yields objective pre-eruptive criteria that complement the existing empirical evaluations. This is especially important at volcanoes for which no data are available from previous unrest. An example of how additional procedures may be applied is shown in **Figure 9** and summarized below. It is based on the simplest conditions for a constant rate of stress supply and no steady VT rate between quasi-elastic and inelastic regimes.

1. Identify the regime of deformation. Quasi-elastic behavior will show an exponential increase in VT number with ground movement and yield a characteristic movement, λch. VT events and deformation will follow distinctly different time series. For a constant rate of stress supply, the deformation rate is constant and the VT rate will increase exponentially with time, yielding the characteristic timescale tch; following the VT rate here is crucial, because a constant deformation rate on its own may erroneously be interpreted as a sign of dynamic stability. Inelastic behavior will show a constant increase in VT number with ground movement and VT event rate maxima and deformation rate will both increase hyperbolically with time (equivalent to linear decreases with time in inverse event rate for VT minima and inverse deformation rate).


In all cases, observations must cover a minimum range of values to confirm that a trend is not spurious. For exponential trends, the minimum range is one characteristic interval (λch or tch), because shorter observations may yield trends statistically indistinguishable from linear; at least one quarter of a trend's total duration must therefore be used to identify its existence. For the inverse-rate trends, the smallest range is a decrease in inverserate minima by a factor of two (hence, at least half of a trend's total duration) before a linear trend can be proposed.

Under the ideal conditions when unrest data follow a complete sequence from lithostatic equilibrium (without a differential stress) to bulk failure, the elastic-brittle model provides estimates of the maximum ground movement (2λch–4λch) before the emergence of inelastic deformation and the onset of bulk failure (and, for a constant rate of stress supply, also the maximum time to the transition, 2tch–4tch). For a continuous acceleration in VT rate across regimes (without at intervening interval at steady rate), bulk failure is expected after an additional time τ at most.

The probability that quasi-elastic behavior will lead to bulk failure increases exponentially with the amount of ground movement (and with time for a constant deformation rate); once it has started, the probability that failure will be completed before an additional interval τ increases hyperbolically with time (**Figure 10**). The ends of the quasi-elastic and inelastic regimes thus coincide with critical changes in the potential for eruption. They are natural stages for step-like increases in alert (**Figure 10**) and can readily be incorporated into existing forecasting event trees (Selva et al., 2012; Sobradelo et al., 2013, 2014) or measures of unrest (Potter et al., 2015).

At volcanoes reawakening after long repose, emergency measurements are often gathered systematically a significant time after the start of unrest, so that the transition from quasielastic to elastic behavior will occur sooner than the ideal maximum time. In the least-favorable case, measurements may become available only when the crust is already in the inelastic regime.

#### CONCLUSION

The good agreement between field observation and parent model for changes in seismicity with deformation at closed volcanoes suggests that their behavior before eruptions is normally governed by the stretching of elastic-brittle crust. Immediate implications are that there is no starting requirement to invoke deformation mechanisms other than elastic-brittle, and that VT signals reflect how a population of small faults responds to changes in stress.

The trend between VT events and deformation persists independently from how seismicity and ground movement vary with time, as shown most clearly by the case for Rabaul (**Figure 8**). A corollary is that VT and deformation time series are mutually dependent and controlled by the rate at which the crust is being stressed. In particular, they reveal how a constant rate of stress (e.g., from magma pressurization) will yield the observed combinations of exponential VT event rate with constant deformation rate, and of hyperbolic VT event rate with hyperbolic deformation rate. These combinations show that the popular FFM model of Voight (1988) is a particular form of the parent VTdeformation model under a constant rate of stress supply. They also provide new criteria to be tested for integrating deterministic physical constraints into probabilistic forecasts of eruption.

The parent model, however, is still not a complete description of pre-eruptive conditions at closed volcanoes. It identifies conditions for bulk failure in the crust, which, although necessary to open a new pathway for magma ascent, do not guarantee that magma will reach the surface. It also focuses on conditions that favor continued acceleration in VT seismicity with time, without significant intervals of steady VT rate. The parent model thus provides a starting point for identifying additional precursory trends and preeruptive criteria. Outstanding goals include identifying whether precursory signals can distinguish between pre-eruptive and non-eruptive outcomes; whether VT rates will accelerate to bulk failure without an interval of steady behavior; and, indeed, whether final VT accelerations are inevitable a significant time before eruption. Achieving these goals will provide new constraints for coupling changes in crustal stresses with specific magmatic processes and, hence, yield greater insights for refining and enhancing current forecasting procedures.

#### AUTHOR CONTRIBUTIONS

CK developed the elastic-brittle model for precursors to eruption and prepared the manuscript.

#### ACKNOWLEDGMENTS

Careful reviews by Jérémie Vasseur, Jean-Robert Grasso, Raffaello Cioni and Valerio Acocella clarified and improved the original manuscript. The research was privately funded. Thanks are also due to Barry Voight, whose support and advice during the emergency at Soufriere Hills, Montserrat, in 1996 laid the foundations for the present work.

feart-06-00133 September 7, 2018 Time: 13:48 # 13

### REFERENCES

feart-06-00133 September 7, 2018 Time: 13:48 # 14


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

Copyright © 2018 Kilburn. 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.

feart-06-00133 September 7, 2018 Time: 13:48 # 15

## Mount St. Helens Retrospective: Lessons Learned Since 1980 and Remaining Challenges

#### Daniel Dzurisin\*

David A. Johnston Cascades Volcano Observatory, United States Geological Survey, Vancouver, WA, United States

Since awakening from a 123-year repose in 1980, Mount St. Helens has provided an opportunity to study changes in crustal magma storage at an active arc volcano—a process of fundamental importance to eruption forecasting and hazards mitigation. There has been considerable progress, but important questions remain unanswered. Was the 1980 eruption triggered by an injection of magma into an upper crustal reservoir? If so, when? How did magma rise into the edifice without producing detectable seismicity deeper than ∼2.5 km or measurable surface deformation beyond the volcano's north flank? Would precursory activity have been recognized earlier if current monitoring techniques had been available? Despite substantial improvements in monitoring capability, similar questions remain after the dome-forming eruption of 2004–2008. Did additional magma accumulate in the reservoir between the end of the 1980–1986 eruption and the start of the 2004–2008 eruption? If so, when? What is the significance of a relative lull in seismicity and surface deformation for several years prior to the 2004–2008 eruption onset? How did magma reach the surface without producing seismicity deeper than ∼2 km or measurable deformation more than a few hundred meters from the vent? Has the reservoir been replenished since the eruption ended, and is it now primed for the next eruption? What additional precursors, if any, should be expected? This paper addresses these questions, explores possible answers, and identifies unresolved issues in need of additional study. The 1980–1986 and 2004–2008 eruptions could have resulted from second boiling during crystallization of magma long-resident in an upper crustal reservoir, rather than from injection of fresh magma from below. If reservoir pressurization and magma ascent were slow enough, resulting strain might have been accommodated by viscoelastic deformation, without appreciable seismicity or surface deformation, until rising magma entered a brittle regime within 2–2.5 km of the surface. Given the remarkably gas-poor nature of the 2004–2008 dome lava, future eruptive activity might require a relatively long period of quiescence and reservoir pressurization or a large injection of fresh magma—an event that arguably has not occurred since the Kalama eruptive period (C.E. 1479–1720).

Keywords: Mount St. Helens, eruption forecasting, volcano monitoring, second boiling, magma reservoir, Pinatubo, Redoubt, Augustine

#### Edited by:

Nicolas Fournier, GNS Science, New Zealand

#### Reviewed by:

Olivier Bachmann, ETH Zürich, Switzerland Raffaello Cioni, Università degli Studi di Firenze, Italy

> \*Correspondence: Daniel Dzurisin dzurisin@usgs.gov

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 01 May 2018 Accepted: 10 September 2018 Published: 05 October 2018

#### Citation:

Dzurisin D (2018) Mount St. Helens Retrospective: Lessons Learned Since 1980 and Remaining Challenges. Front. Earth Sci. 6:142. doi: 10.3389/feart.2018.00142

**204**

#### INTRODUCTION

feart-06-00142 October 5, 2018 Time: 12:18 # 2

Of fundamental importance to the understanding of active magmatic systems and to assessments of volcano hazards is the process of magma accumulation in crustal reservoirs. Indicators can include anomalous seismicity, ground deformation, changes in flux or composition of aqueous or gas emissions, and residual gravity changes (Tilling, 2008). Given current understanding and recent experience at several volcanoes worldwide, we can expect at least subtle changes in those indicators to occur months to perhaps years before the onset of eruptive activity—especially at closed-vent volcanoes after long repose periods. However, detecting those changes continues to pose a difficult challenge at one of the most closely monitored volcanoes on Earth.

At Mount St. Helens, Washington, United States, two eruptive episodes in the past four decades have provided opportunities to study changes in crustal magma storage at an active arc volcano: (1) a sequence of explosive and dome-forming eruptions during 1980–1986, and (2) a continuous dome-forming eruption during 2004–2008. Much has been learned, but important questions remain unanswered. New insights and modeling approaches have revealed complexities that have yet to be fully addressed. Seemingly enigmatic observations can be explained on an ad hoc basis, but a lack of hard constraints limits most of those explanations to plausibility arguments.

This paper explores the current state of knowledge, highlights some unresolved issues, and offers suggestions for ways to move forward both at Mount St. Helens and elsewhere. The emphasis is on seismic and geodetic information, reflecting the author's background in geophysics; lesser amounts of geochemical and petrologic evidence are cited where pertinent. More information about geochemical signals to be expected at restless volcanoes is available in a recent review of that topic by Inguaggiato et al. (2018, and references therein). The principal focus of this paper is on the upper 15 km of Mount St. Helens' magmatic system, including a magma storage zone that fed the 1980–1986 and 2004–2008 eruptions. The entire system from upper crust to Moho has been imaged recently by the multidisciplinary Imaging Magma Under St. Helens (iMUSH) project<sup>1</sup> . An anomalous zone between about 4 and 14 km below sea level (BSL), characterized by high seismic Vp/Vs ratios and low Vp values, is inferred to represent a crustal magma reservoir; maximum melt fractions of 10–12% are inferred at 4–6 km BSL (Kiser et al., 2016, 2018). Henceforth, when citing depths, I have in some cases specified below the surface, below the vent, or below sea level; where not specified, below the surface is implied. At Mount St. Helens, surface elevation in the vent area (1980 crater floor) is ∼2 km above sea level.

#### MOUNT St. HELENS 1980: LINGERING QUESTIONS

The Goat Rocks eruptive period at Mount St. Helens ended in 1857 (Hoblitt et al., 1980), 123 years prior to the volcano's

<sup>1</sup>http://imush.org/

reawakening in 1980. Any signs that Mount St. Helens' century-long quiescence was nearing an end went unnoticed until less than 2 weeks before the first phreatic eruption. The earliest indication of unrest was a gradual increase in daily earthquake counts starting on March 15, 1980, followed by the first of many magnitude 4+ earthquakes on March 20. The ensuing swarm intensified further on March 25 and peaked on March 27, when the first phreatic eruption occurred (Endo et al., 1981). Observers on an overflight several hours later reported (Christiansen and Peterson, 1981, p. 18):

A new crater about 60–75 m across had formed in the northern part of the old 400-m-wide ice-filled summit crater, and snow on the southeast sector of the volcano was covered by dark ash emitted from the new crater. . . The summit area was bisected by an east-trending fracture nearly 1,500 m long that extended from high on the northwest flank, across the old crater, down the upper northeast flank. Another less continuous fracture system paralleled this master fracture just north of the old crater rim and bounded the south side of a newly uplifted block, or bulge, on the volcano's north flank. These changes clearly had occurred during the period of extremely high seismicity and initial eruption, between observations on the morning of March 25 and the afternoon of March 27.

Photogrammetric and geodetic measurements showed that the bulge grew outward at rates of 1.5–2.5 m/day until it failed catastrophically on May 18, disrupting a growing cryptodome, triggering a lateral blast, and producing a massive debris avalanche (Jordan and Kieffer, 1981; Lipman et al., 1981; Moore and Albee, 1981; Voight et al., 1981). Seismic, deformation, and petrologic evidence indicates that the main explosive phase of the eruption tapped a vertically elongate reservoir at least 7–8 km below vent level, extending to perhaps twice that depth (Scandone and Malone, 1985; Pallister et al., 1992 and references therein; Waite and Moran, 2009), consistent with recent seismic tomography results (Kiser et al., 2016, 2018).

Nearly all recorded seismicity during March 15–May 18, 1980, including more than 2,400 earthquakes with local magnitudes greater than 2.4, was confined to a small volume less than 2.5 km beneath the volcano's north flank (Endo et al., 1981). The appearance of a graben and north-flank bulge on March 27, combined with the fact that the May 18 landslide transected an active cryptodome, indicates that magma had already intruded the edifice, probably to within several hundred meters of the surface, only 7 days after the onset of intense seismicity on March 20 (12 days after the initial uptick in seismicity on March 15). At the time there was no local seismic network, only one short-period vertical-component seismometer that was installed on the volcano's west flank in 1972 (SHW, 3.5 km from the summit; Endo et al., 1981; Malone et al., 1981) (**Figure 1**). So, the lack of any deeper or longer-term seismic indication that the volcano's repose was coming to an end can be attributed in part to limited seismic monitoring. However, the same is true for geodetic precursors: If any occurred prior to the onset of shallow seismicity or beyond the north-flank bulge, they went unnoticed.

were installed or first measured in 1972, respectively; all other stations and lines were installed or first measured after the onset of intense shallow seismicity on March 20, 1980. In addition to the spirit-level triangle at Timberline campground (TBL), the surface of Spirit Lake was used as a water tiltmeter (see text). Some peripheral seismic stations are omitted to keep the map at a legible scale. Seismic station locations are from Malone et al. (1981). EDM station and target locations are from Lipman et al. (1981). Tiltmeter station locations are from Dvorak et al. (1981). See original publications for additional information. Hillshade is based on a USGS topographic map and depicts pre-1980 topography.

Three types of ground-tilt measurements were made on the lower flanks and around the base of Mount St. Helens starting in late March 1980 (**Figure 1**). The ice-covered surface of Spirit Lake, ∼8 km NNE of the summit, was used as a large water tiltmeter, and a spirit-level triangle was established at Timberline campground ∼5 km NE of the summit (**Figure 1**, TBL). No tilts larger than the estimated measurement precision of 2 microradians were observed at Spirit Lake. At TBL, small inflationary tilts were observed before some phreatic eruptions and deflation during them, but no long-term trend indicative of a deeper magmatic source was apparent (Lipman et al., 1981). Five electronic tiltmeters were installed during the latter half of April at radial distances of 3–15 km from the summit; none recorded a significant change in tilt prior to the May 18 eruption (Dvorak et al., 1981).

A 7.6-km EDM line from Smith Creek Butte to East Dome part of a trilateration network surrounding Mount St. Helens that was established in 1972—was re-measured for the first time on April 10, 1980, and again on April 25 (**Figure 1**, SCB–EDM). All line-length changes were within measurement uncertainty (Lipman et al., 1981). Dzurisin (2007, pp. 346–347) raised the possibility that the line's configuration made it insensitive to pressure changes in the reservoir-conduit system that fed the May 18, 1980, eruption. An alternative explanation, consistent with the absence of progressive changes in ground-tilt at Spirit Lake, TBL, and all five tiltmeter sites, is that little or no far-field deformation occurred during 2 months of seismicity, phreatic eruptions, and cryptodome growth leading up to the catastrophic events of May 18. The situation was dramatically different on the volcano's north flank, however.

A local geodetic network was established at Mount St. Helens in April 1980 (**Figure 1**). Repeated EDM and theodolite measurements revealed intense deformation of the volcano's north flank but only small changes elsewhere on the edifice (1.5–2.5 m/day versus a few mm/day). Lipman et al. (1981, p. 151) concluded: ". . .except for the bulge, Mount St. Helens was geodetically stable during this period" (April 10–May 18, 1980).

In summary: (1) recorded seismicity leading up to the May 18, 1980, eruption was confined to a small volume beneath the volcano's north flank within 2.5 km of the surface; (2) little or no surface deformation was detected outside the north-flank bulge after late March 1980; (3) magma had intruded the edifice by March 27, 1980, only 7 days after the onset of intense seismicity; and (4) seismic, geodetic, and petrologic evidence suggests the main phase of the eruption was fed from a magma reservoir at least 7–8 km deep. The apparent absence of deeper seismicity and far-field deformation can be attributed at least in part to lack of a local seismic network and sparse deformation measurements prior to March 1980. But if those precursors occurred, they were too early or too subtle to be detected by the monitoring efforts described above.

Was the 1980 eruption triggered by an injection of magma into the upper crustal reservoir beneath the volcano? If so, when? If after 1972, when seismic and geodetic monitoring began (albeit sparse), why was no evidence of magma injection detected? Following a 123-year repose, how did magma rise into the edifice without producing detectable seismicity deeper than ∼2.5 km or measurable surface deformation beyond the north-flank bulge? Would deeper or longer-term seismic precursors have been detected with a modern network of three-component broadband seismometers? Would far-field deformation indicative of a deep source have been detected if a continuous GPS (CGPS) network and InSAR monitoring had been in place? The answers to these questions are hypothetical, but nonetheless they are important for designing monitoring strategies at arc volcanoes, including Mount St. Helens and others in the Cascade arc. Before addressing issues raised by Mount St. Helens' 1980 eruption further, the next section summarizes pertinent aspects of the volcano's 2004–2008 dome forming eruption—and raises more questions.

#### MOUNT St. HELENS 2004–2008: QUESTIONS OLD AND NEW

The May 18, 1980, eruption was followed by five smaller explosive eruptions during the remainder of 1980 and by episodic growth of a lava dome that ended in October 1986 (Swanson et al., 1987; Swanson and Holcomb, 1990; Brantley and Myers, 2000). Dome growth was accompanied by hundreds of small explosive events, some of which sent ash to heights of several kilometers above the volcano, showered the crater floor with rocks, and generated small lahars. These events, many of which were documented by observers in the field, provided context for similar events that occurred later. During August 1989–October 1991, at least six small ash-producing explosions occurred from the 1980–1986 dome, part of a series of 28 explosion-like seismic events with signatures similar to those described above (Mastin, 1994).

Measurements of a high-precision trilateration network surrounding Mount St. Helens were made with a Geodolite in 1982 and 1991, and with GPS in 2000 and 2003 (**Figure 2**). A small amount of areal dilatation, consistent with repressurization of the magma reservoir-conduit system that fed the 1980–1986 eruption, occurred at an average rate of 144 ± 39 nanostrain/yr during 1982–1991. With one exception, no additional deformation was detected during 1991–2000 or 2000–2003. The exception was campaign GPS station DMSH on the 1980–1986 lava dome, which moved downward ∼87 mm/yr and east-northeastward ∼35 mm/yr during 2000–2003, presumably as a result of cooling and compaction of the dome. Lisowski et al. (2008, p. 329) concluded: "Remarkably little far-field volcanic deformation has occurred around Mount St. Helens since shortly after the crater-forming collapse and eruption in 1980. Data collected in 1982 and 1991 for surveys of a regional high-precision trilateration network provide the clearest evidence for recharge of the volcano's magma system."

The trilateration results are consistent with data from CGPS station JRO1, which was established 9 km north of the volcano in 1997. It recorded no anomalous movement until the onset of shallow seismicity a few days before the appearance of a welt on the south crater floor and subsequent dome-forming eruption starting in October 2004 (see below). Together, the trilateration and JRO1 CGPS data show that only a small amount of deep-seated inflation occurred during 1982–1991 (perhaps starting after 1986, when dome extrusion ceased), and little or none occurred during 7+ years immediately preceding the start of the 2004–2008 eruption.

Starting in March 1980, the Mount St. Helens network of short-period vertical-component seismometers was expanded to include 10 stations within 10 km of the summit; more were added after the 2004–2008 eruption began, including two broadband stations (Moran et al., 2008a) (**Figure 2**). In 1987, earthquakes began to occur beneath Mount St. Helens at depths greater than 3 km, which had not been the case earlier except in the months following the May 18, 1980, eruption and briefly in March 1982 (Weaver et al., 1983; Moran, 1994; Musumeci et al., 2002) (**Figure 3**). The latter sequence was associated with an explosive eruption on March 19, 1982, that melted glacier ice in the 1980 crater and produced a lahar that reached Spirit Lake (Waitt et al., 1983). Moran (1994) computed focal mechanisms for earthquakes during 1987–1992 in two depth ranges: 4–6.5 km and 6.5–10 km BSL. Earthquakes in the first group are tightly clustered, their focal mechanisms are primarily strike-slip, and many P and T axes point in directions ∼80◦ offset from the regional stress field. These characteristics led Moran (1994) to conclude that they were caused by interaction of the regional stress field with the volcano's magmatic plumbing system.

The second, deeper group of earthquakes surrounds an aseismic zone that widens with depth – an inferred upper crustal magma reservoir, presumably the same one that fed the May 18, 1980, eruption (**Figure 4**). Earthquakes in the second group also have mostly strike-slip focal mechanisms, but unlike the first group their P and T axes show no dominant orientations. Instead, according to Moran (1994, p. 4343) ". . .there is a striking geometrical pattern when either the P or T axes are plotted in map view. . . The compressional axes define a wheel-spoke pattern, pointing radially away from the center of the 1987–1992 aseismic zone, and the tensional axes are tangentially aligned." The pattern led Moran (1994) to conclude that the earthquakes occurred in response to pressurization of the magma storage system. Musumeci et al. (2002) came to the same conclusion based on

FIGURE 2 | Map showing seismic and GPS stations at Mount St. Helens prior to the 2004–8 eruption. Two broadband seismic stations, JRO and STD, and two continuous GPS stations, TWRI and TWIW, were deployed in October 2004 after the start of a shallow earthquake swarm on September 23, 2004. JRO1 was the only CGPS station operating when the swarm began; power to CGPS station DOM1 was lost in January 2004 and restored on September 27, 2004. The campaign GPS network was measured in 2000, 2003, and partially in late September–early October 2004 (see Lisowski et al., 2008). Some peripheral seismic stations are omitted to keep the map at a legible scale. Seismic station locations are from Moran et al. (2008a). GPS station locations are from Dzurisin et al. (2008). See original publications for additional information. Hillshade is an ESRI Image Service product from 2013 accessed through ArcGIS Online and depicts post-1980 topography.

are plotted regardless of the quality of the location. Depths are relative to sea level; surface elevations in the area are mostly 1–2 km above sea level.

precise relative hypocenter relocation of the earthquakes. Note that pressurization is not synonymous with replenishment, as the latter term is used here. By replenishment we mean addition of magma to a reservoir from below; in the case of a sealed reservoir or low gas-diffusivity magma, pressurization can occur without replenishment as a result of volatile exsolution during crystallization ("second boiling").

Seismicity continued in much the same way described by Moran (1994) through 1998. Thereafter, events deeper than ∼2 km became less common while seismicity near 2 km depth (0 km BSL in **Figure 3**) became more persistent. Also evident in **Figure 3** is the sporadic nature of seismicity during 1987–1998. Much of the activity occurred in years-long swarms of volcano-tectonic (VT) events, with hypocenters that align along a vertical pipe from ∼0–8 km BSL. Moran (1994) and Musumeci et al. (2002) attributed the concentration of events near 2 km depth (0 km BSL) to progressive development of a plug near the top of a magma conduit outlined by seismicity in the 0–8 km BSL depth range, a process attributable to crystallization of conduit magma beneath the 1980–1986 dome. The last notable swarm below 2 km depth prior to the onset of short-term seismic and geodetic precursors to the 2004–2008 dome-forming eruption occurred in 1998. Paradoxically, seismicity deeper than 2 km was less intense during the 5 years immediately preceding the start of the eruption than it was during the previous decade (**Figure 3**). Recall from above that surface deformation followed a similar pattern: A small amount of dilatation occurred sometime during 1982–1991 (presumably after the 1980–1986 eruption ended), but thereafter the area was stable for more than a decade leading up to the 2004–2008 eruption.

A swarm of small VT earthquakes beneath the 1980 crater floor and south rim began on September 23, 2004, intensified for ∼36 h, and then declined to a minimum early on September 25. All located events were less than 2 km deep, most had magnitudes less than 1 (max = 2.2), and fault plane solutions were mixed. At that point the activity resembled a shallow swarm that occurred on November 2–4, 2001 (Moran et al., 2008a).

An important difference, however, is that in this case CGPS station JRO1 abruptly began moving toward the volcano and down at an average rate of 0.5 mm/day, concurrently with the start of the earthquake swarm (Lisowski et al., 2008) (**Figure 5**). L1 CGPS station DOM1, located ∼300 m north of what became the 2004–2008 vent (collocated with campaign GPS benchmark DMSH) also moved, but the timing is poorly constrained. Power to the station was lost in January 2004 and restored on September 27, 2004 (LaHusen et al., 2008). Sometime between January and September 28, 2004, DOM1 moved 236 ± 16 mm north-northwest and 90 ± 40 mm up, in strong contrast to its east-northeast and down motion during 2000–2003. Its 2004 movement was up and away from a shallow intrusion under the south crater floor beginning in late September 2004 (see below), suggesting a causal relationship. More than a dozen campaign GPS stations on and around Mount St. Helens were occupied during late September–early October 2004. All offsets were small and no convincing pattern of deformation emerged (Lisowski et al., 2008). Only CGPS stations JRO1, 9 km from the 2004–2008 vent, and DOM1, 300 m away, moved appreciably in the days before extrusion began—toward and away from the vent, respectively.

The earthquake rate stabilized briefly but then increased over several days starting on September 26, while low-frequency (LF) and hybrid events began to accompany VT events (Moran et al., 2008a). Fresh crevasses in glacier ice on the south crater floor were recognized in hindsight in photographs taken during an observation flight on September 26. A photograph taken a day earlier by a hiker on the south crater rim showed no obvious

disturbance in the same area, indicating that intense surface deformation began simultaneously with the uptick in seismicity on September 25–26. That inference is consistent with the large offset of DOM1 measured between January 2004 and September 28, 2004, although the timing of that motion is unknown. During the next several days, a surface welt indicative of a shallow intrusion grew in the deforming area at rates as high as 8.9 m<sup>3</sup> /s. By October 10 it was more than 100 m high, and the following day the first of several lava spines emerged. Extrusion continued uninterrupted until January 2008, forming a composite dacite dome and severely disrupting Crater Glacier (Dzurisin et al., 2008; Vallance et al., 2008; Dzurisin et al., 2015).

The appearance of fresh cracks in glacier ice 2–3 days after the onset of a shallow earthquake swarm on September 23, 2004, and subsequent growth of a welt on the south crater floor are reminiscent of the scenario that unfolded in 1980. In the earlier instance, surface faulting and growth of a bulge on the volcano's north flank began 7 days after the onset of intense seismicity that might have signaled the start of magma's ascent into the edifice. In both cases, magma reached the surface without producing notable seismicity deeper than 2–3 km below the surface, and without producing a commensurate amount of far-field surface deformation—even though the 1980 eruption tapped a magma reservoir at least 7–8 km deep and the same reservoir was the pressure source responsible for surface deformation during the 2004–2008 eruption (Lisowski et al., 2008). Note that the pressure source deduced by modeling deformation data need not be the same as the magma storage zone that fed the 2004–2008 eruption. On the contrary, phenocryst assemblages in Mount St. Helens dacite erupted in 1980 and 2004 imply different equilibrium source depths of 8.6 ± 1 km and ∼5 km below the surface, respectively (Rutherford, 1993; Pallister et al., 2008; Rutherford and Devine, 2008). The apparent discrepancy is explained if the shallower source tapped during the 2004–2008 eruption was in pressure communication (presumably contiguous) with the deeper source tapped in May 1980 (**Figure 4**). A likely scenario is that both eruptions were fed by magma from the upper part of a reservoir that had experienced a complicated crystallization history, i.e., the 2004–2008 magma was "left over" from the 1980–1986 eruption. Thornber et al. (2008, p. 727) wrote:

The diverse range in composition of amphibole in all samples of 2004–2006 dacite, and the complex zonation observed in many phenocrysts, suggests a well-mixed source magma with components that are subjected to repeated heating and (or) pressurization within this pressure-temperature window [∼900◦C to ∼800◦C between 100 MPa and ∼350–400 MPa or ∼4-km and 13.5–15-km depth]. Amphibole textural and compositional diversity suggest dynamic conditions in the upperreservoir zone, which has been tapped steadily during ∼2 years of continuous and monotonous eruption. This well-mixed crystal mush is likely to have been subjected to repeated injection of hotter magma into cooler crystal-laden magma while simultaneously assimilating earlier generations of dacitic roof material and surrounding gabbroic rock.

The implications of these and other observations for several unanswered questions at Mount St. Helens are discussed below.

#### DISCUSSION

feart-06-00142 October 5, 2018 Time: 12:18 # 8

Nearly four decades of study at Mount St. Helens has greatly increased understanding of the volcano's eruptive history, magma plumbing system, and associated hazards (e.g., Sherrod et al., 2008 and references therein). However, progress has been accompanied by the realization that firm answers to several fundamental questions remain out of reach. Arguably, there are plausible answers to all of the questions posed above. The challenge is to move beyond plausibility arguments to a self-consistent explanation of what happened at Mount St. Helens since 1980, with an eye toward what the future might hold and how better to capture essential monitoring data in the future. The following sections address several unresolved issues, discuss one or more plausible explanations in each case, and suggest what might be done to move forward.

#### In Search of Answers: Mount St. Helens 1980–Present

#### Was the 1980 Eruption Triggered by an Injection of Magma Into an Upper Crustal Reservoir and, if so, When?

This question cannot be addressed with available seismic or geodetic observations, which did not reveal any unusual activity until the onset of shallow-seated unrest in March 1980. If one were to assume that a single short-period seismometer 3.5 km from the summit (SHW, **Figure 1**) would have detected seismicity produced by an injection of magma into the 7+ km-deep reservoir that fed the 1980 eruption, then the injection must have occurred prior to 1972. Likewise, if measurements of a single EDM line in 1972 and April 1980 (Lipman et al., 1981) (SCB–EDM, **Figure 1**) would have detected deformation caused by such an injection, then it must have occurred before 1972. Neither constraint is helpful, because the underlying assumptions are not likely to be valid. Neither a single shortperiod seismometer nor measurements of a single EDM line spanning 8 years is likely to have detected an injection of magma at depths ≥ 7 km, especially if the volume was small or the process was gradual. Whether a modern seismic or geodetic network would have detected such an event remains an open question (see section "Lessons From Other Volcanoes"). The seismic and geodetic monitoring networks were much better prior to the 2004–2008 eruption (**Figures 1**, 2), but again no activity deeper than ∼7 km that could be attributed to magma injection into the reservoir was detected. Possible explanations include: (1) there was none, (2) it had occurred prior to 1980, or (3) it was muted by the presence of mush zones in the magma column and by viscoelastic accommodation of strain in host rock.

Pallister et al. (1992) reviewed activity at Mount St. Helens during the past 500 years, including the Kalama (C.E. 1479–1720), Goat Rocks (C.E. 1800–1857), and 1980–1986 eruptive periods. The Kalama period began with two large-volume, highly explosive eruptions of dacite pumice, the first in C.E. 1479 (Wn tephra), the second in C.E. 1482 (We tephra). Activity during the ensuing ∼240 years included explosive eruption of dacitic and andesitic tephras and pyroclastic flows, relatively low explosivity eruption of andesite lava flows, and extrusion of a large dacitic summit dome. Erupted products gradually became more mafic: explosive dacitic eruptions were followed by weaker andesitic tephra eruptions and extrusion of andesitic lava flows. The compositional trend reversed near the end of the eruptive period with extrusion of the large dacitic summit dome (Mullineaux, 1996; Clynne et al., 2005). The Goats Rocks period began with a highly explosive, moderate-volume dacitic eruption in C.E. 1800 (T tephra), soon followed by extrusion of the andesitic Floating Island lava flow. Several small explosive eruptions were reported during C.E. 1831–1857, the largest in C.E. 1842. At about the same time, the dacitic Goat Rocks lava dome was extruded incrementally on the volcano's north flank (Mullineaux, 1996; Clynne et al., 2005).

Pallister et al. (1992) assessed geochemical evidence for an influx of new magma to the crust during each of the eruptive periods mentioned above. They reported several indicators of magma mixing during Kalama time, including banded pumice and scoria clasts with compositions ranging from basalt to dacite (Pallister et al., 1992, p. 133–138). They concluded that the Kalama products were produced by mixing of arc dacite with REE-enriched basalt, i.e., by injection of basalt into a preexisting upper crustal reservoir of dacite. Pallister et al. (1992, p. 138–139) also reported that compositional variations during Goat Rocks time and 1980–1986 were smaller than during Kalama time, and evidence of mixing prior to or during those eruptive periods is equivocal or lacking. They concluded (p. 139): ". . .products of the Goat Rocks and current [i.e., 1980–1986] periods represent two diminishing oscillations following the major perturbation of the magmatic system brought about by influx of basaltic magma during the Castle Creek [2.55–1.895 ka] and Kalama periods." A small additional influx of magma prior to Goat Rocks time or 1980–1986 cannot be ruled out, but Pallister et al. (1992) assessed that the geochemical data are consistent with fluid withdrawal from a zoned reservoir that might not have been replenished since Kalama time.

If the most recent injection of magma into Mount St. Helens' upper crustal reservoir occurred more than 500 years ago, the absence of longer-term or deeper-seated precursors to the 1980 activity is understandable. Given available information, it seems plausible and perhaps even likely that the 1980 eruption resulted from crystallization and "second boiling" of magma that had resided in the reservoir for centuries to millennia. Second boiling is the process in which volatiles are increasingly concentrated in the residual melt as a magma body crystallizes; eventually they become saturated in the residual melt and are exsolved, raising the magma pressure to the point of eruption. Theoretical studies have shown second boiling to be a viable eruption trigger for magmas ranging in composition from basalt to rhyolite

(e.g., Blake, 1984; Tait et al., 1989; Woods and Cardoso, 1997). Tait et al. (1989) showed that the time required is on the order of a few years for mafic melts and a few centuries for silicic melts. Accordingly, for the intermediate-composition reservoir at Mount St. Helens, the 123-year eruptive hiatus from 1857 to 1980 would be sufficient for second boiling to be a plausible eruption trigger. That possibility is reinforced by the reservoir's large volume relative to the volume erupted on May 18, 1980 (0.22 km<sup>3</sup> DRE, Nathenson, 2017). Barker and Malone (1991) estimated the reservoir volume to be 5–7 km<sup>3</sup> based on stress-field modeling of post-1980 eruption earthquakes. Mastin et al. (2008, p. 480) came to a similar conclusion (". . .the magma reservoir at Mount St. Helens is several to perhaps a few tens of cubic kilometers in size. . .") based on modeling of geodetic and dome-growth measurements for the first ∼1.5 years of the 2004–2008 eruption. An unknown and possibly large proportion of the reservoir probably consists of low-melt-fraction crystal mush (Bachmann and Huber, 2016; Cashman et al., 2017) that is not eruptible. On the other hand, 0.78–2.3 km<sup>3</sup> DRE of dacite magma erupted during the largest Holocene eruption at Mount St. Helens, the 3.4 ka Yn eruption (Nathenson, 2017), so a reasonable lower bound on the melt volume is ∼2 km<sup>3</sup> .

In the second-boiling scenario, high explosivity of the May 18, 1980, eruption derives from two factors. First, sometime prior to March 1980 a highly explosive and relatively buoyant body of magma rose from the gas-rich top of the primary reservoir, where volatiles had been accumulating since at least Goat Rocks time. Final ascent of the magma from 2–3 km depth starting in mid-March 1980 resulted in formation of the cryptodome, which was transected by the north-flank avalanche with explosive consequences. A second factor that contributed to high explosivity is that gravitational unloading of the magmatic system by the avalanche allowed a still gas-rich portion of the reservoir at least 7–8 km deep to be tapped directly, as suggested by a change in eruptive style and vigor around 1215 local time on May 18, nearly 4 h after the eruption began (Carey et al., 1990).

Second boiling of already-resident magma might explain the lack of longer-term precursors to the 1980 eruption, given that reservoir pressurization would have been gradual and might have spanned centuries, but what explains the lack of deeper-seated precursors?

#### How Did Magma Rise From a 7+ Km-Deep Reservoir Into the Edifice Without Producing Seismicity Deeper Than 2.5 Km or Measureable Deformation Beyond the Volcano's North Flank?

Aseismic or near aseismic ascent of magma from great depth has been documented at several arc volcanoes, including Soputan, Indonesia (Kushendratno et al., 2012); Pinatubo, Republic of the Philippines; Volcán de Colima, Mexico (White and McCausland, 2018); and also at Hekla, Iceland (Einarsson, 2018). So it is plausible that Mount St. Helens' 1980 eruption was triggered by magma injection to its upper crustal reservoir in the absence of detectable deep seismicity. However, as noted above, there is no geochemical evidence of magma mixing in the 1980 products to indicate such an occurrence.

Another plausible answer to the question posed above follows directly from the discussion in the preceding section. If no new magma was added to the upper crustal reservoir between Kalama time and 1980, and the 1980 eruption resulted from second boiling of long-resident magma, then the buildup to 1980 would have been gradual and any precursors spread across decades to centuries. Pressurization of the reservoir would have progressed at the rate magma crystallized and volatiles exsolved, a slow process in a large reservoir at 7+ km depth. Any resultant surface deformation would have accumulated at the same pace and likely would not have been detectable in the years or decades prior to 1980. Likewise, any seismic indications of second boiling might have been subtle and likely would have escaped detection by the minimal seismic network operating at the time.

The appearance of the north flank bulge on March 27 without attendant seismicity deeper than 3 km or surface deformation beyond the bulge itself seems more problematic. Hoblitt and Harmon (1993) studied dacite clasts incorporated in the May 18, 1980, blast deposit that are inferred to be fragments of the north-flank cryptodome (Hoblitt et al., 1981; Moore and Sisson, 1981). Based on the clasts' chemical composition and mineralogy, Hoblitt and Harmon (1993) concluded (p. 428): "The magma that produced the cryptodome dacite is apparently derived directly from the source magma chamber rather than from magma remaining in the conduit from the preceding eruptive episode (the Goat Rocks episode)." We might reasonably expect that the rise of 0.11 km<sup>3</sup> (Moore and Sisson, 1981) of dacite magma from the top of the reservoir at ∼7 km depth to within several hundred meters of the surface to form the cryptodome would have produced measureable seismicity over that depth range and surface deformation at distances commensurate with those source depths. The fact that neither was observed suggests the rise of cryptodome magma was accommodated by viscoelastic deformation of conduit fill and host rock. Presumably, that process occurred relatively slowly over some considerable time period prior to March 1980. From 2–3 km depth to the surface, the conduit fill and host rock were brittle enough to fracture at the strain rate imposed by the more rapidly intruding magma, accounting for the observed shallow seismicity. The fact no deformation was observed after the local geodetic network was established in April 1980 suggests the upper conduit was already decoupled from the edifice by that time, presumably as a result of the reaming process that began on March 20 with the first magnitude 4 earthquake and finished by March 27 when the north flank bulge was recognized and the first phreatic eruption occurred.

#### Did Additional Magma Accumulate in the Crust Between the End of the 1980–1986 Dome-Forming Eruption and the Start of the 2004–2008 Eruption? If So, When?

As described above, there is circumstantial evidence from both seismicity and surface deformation for repressurization of the upper crustal storage reservoir prior to the 2004–2008 eruption. A small amount of areal dilatation occurred between trilateration surveys in 1982 and 1991, and there were sporadic earthquake swarms at depths ranging from 2–8 km during

1987–1998. Both occurrences are indicative of magma system pressurization but neither is compelling evidence of magma influx; volatile exsolution during crystallization might account for both. Curiously, no additional deformation was detected during 1991–2003, and there was significantly less seismicity below 2 km depth during 1999–August 2004 (**Figure 3**).

Pallister et al. (2008) analyzed samples of 2004–2008 dome rocks and concluded (p. 648): "A relatively low pressure of last phenocryst growth suggests that the magma was derived from near the apex of the Mount St. Helens magma reservoir at a depth of about 5 km. Viewed in the context of seismic, deformation, and gas-emission data, the petrologic and geochemical data can be explained by ascent of a geochemically distinct batch of magma into the apex of the reservoir during the period 1987–1997, followed by upward movement of magma into a new conduit beginning in late September 2004." They acknowledged evidence permissive of reservoir replenishment during the eruption, but preferred the following scenario (p. 683): "Since 1980, the magma reservoir has seen minimal replenishment, such that convection has produced a batch of well-mixed, crystal-rich magma with 65 percent SiO<sup>2</sup> near the reservoir roof. A slow increase in pressure was brought about beginning in late 1987 by convection-driven exsolution of a water-rich volatile phase (which outpaced volatile losses through roof rocks). Magmatic pressure finally exceeded lithostatic load in late September 2004, and the eruption ensued." In this scenario: (1) the small amount of surface dilatation measured between 1982 and 1991 probably started sometime after dome-building ended in late 1986 and reflected exsolution-driven pressurization of the reservoir (second boiling), (2) second boiling was responsible for the series of small ash-producing explosions and explosion-like seismic events that occurred during August 1989–October 1991 (Mastin, 1994), and (3) progressive sealing and pressurization of the conduit induced sporadic swarm seismicity at 0–7 km depth BSL during 1987–98 (**Figure 3**).

#### What Is the Significance of a Relative Lull in Seismicity and Surface Deformation for Several Years Prior to the 2004–2008 Eruption Onset?

Lack of measurable deformation starting in 1991 and the relative lull in seismicity after 1998 suggest the system reached a metastable condition in which the rate of pressurization was low enough to be accommodated mostly by aseismic, viscoelastic deformation. Such a condition might arise if a pressure connection existed between the reservoir and that segment of the conduit between the top of the reservoir and the brittle carapace at 2–3 km depth. It seems reasonable that such a connection might have persisted during the 18-year hiatus between the end of the 1980–1986 eruption and start of the 2004–2008 eruption. In that situation, the slow increase in conduit pressure might progressively inhibit exsolution of water from reservoir magma, slowing the rate of pressurization to the extent that the system remained metastable for several years while the pressure in the upper part of the conduit slowly approached and eventually exceeded lithostatic load. The paucity of earthquake activity below 2 km depth is understandable if the conduit and its host rock were hot enough to accommodate strain by viscoelastic relaxation rather than brittle failure.

#### How Did Magma Reach the Surface in 2004 Without Producing Seismicity Deeper Than 2 Km or Measurable Deformation More Than a Few Hundred Meters From the Vent?

According to the scenario above, intense seismicity between 2 km depth and the surface, soon followed by appearance of a welt on the crater floor, marked magma's sudden breaching of the brittle carapace. Brittle-failure earthquakes did not occur along the conduit below the level of the carapace because at those depths the conduit and wall rocks were hot enough, and the strain rate imposed by slow ascent of gas-poor magma was low enough, that strain was accommodated by viscoelastic deformation. The same conditions explain the lack of surface deformation beyond a few hundred meters from the vent. The implication is that the conduit and its host rock between the surface and ∼2 km depth became cold and brittle from 1986 to 2004, while below that depth it remained hot enough for viscoelastic deformation to occur at low strain rates. Mount St. Helens receives average annual rainfall of 3.56 m and hosts numerous thermal springs (Tilling et al., 1990; Bergfeld et al., 2017). Perhaps the abundance of meteoric water and the fractured nature of the uppermost 2 km of the conduit facilitated water circulation and convective heat extraction, allowing a cold, brittle plug to form during the 18-year hiatus between eruptions. Hundreds of small explosive events during August 1989–October 1991 and sporadic earthquake swarms during 1987–1998 might have been a consequence of second boiling in the magma reservoir plus cooling of the upper part of the conduit (see section "Did Additional Magma Accumulate in the Crust Between the End of the 1980–1986 Dome-Forming Eruption and the Start of the 2004–2008 Eruption? If So, When?").

#### Has the Magma Storage System Been Replenished Since the 2004–2008 Eruption Ended, and Is It Now Primed for the Next Eruption?

Four lines of observational evidence bear on this question: seismicity, deformation, residual gravity, and hydrothermal-system chemistry. Seismicity of the type attributed by Moran (1994) to re-pressurization after the 1980–1986 eruption began soon after the 2004–2008 eruption ended and was continuing when this was written in July 2018 (Moran and Lisowski, 2014) (**Figure 3**). Individual swarms below 2 km depth have been less intense than during 1989–1991 and 1998, for example, but seismicity has occurred over a similar depth range. The concentration of earthquakes near 2 km depth that was attributed to formation of a conduit plug in the late 1980s has yet to develop (**Figure 3**, 0 km BSL). If the conduit cooled at about the same rate after 2008 as it did after 1986, we can infer that the rate of re-pressurization has been less since 2008, i.e., although enough time has elapsed since 2008 for a plug to form, magmatic pressure is not yet great enough to begin fracturing it. Lava extruded during 2004–2008 was considerably more crystal-rich and gas-poor than 1980–1986 dome lava (Pallister et al., 2008). For that reason, we might expect less gas to be exsolved from

magma remaining in the conduit after the more recent eruption, consistent with a lower rate of re-pressurization and the paucity of seismicity near 2 km depth relative to pre-2004. Alternatively, Pallister et al. (2008) inferred that conduit magma had solidified to a depth of 1–2 km below the vent as a consequence of decompression crystallization during the 2004–2008 eruption. They proposed the resulting plug was relatively porous along a marginal gouge zone, allowing escape of exsolved gasses (see also Gaunt et al., 2014). In either scenario, pressurization of the conduit occurred at lower rate after the 2004–2008 eruption than during 1986–2004, consistent with continued weak seismicity at 0–2 km depth.

A second line of evidence for magma-system re-pressurization since 2008 is a reversal in trend at several CGPS stations on or near the volcano that started while the 2004–2008 eruption was waning or soon after it ended. JRO1, for example, moved south (toward the volcano) and down during the eruption, but north and up during 2008–2012 (**Figure 5**). Since then, any motion has been below the detection threshold of ∼1 mm/yr. Dzurisin et al. (2015) attributed the reversal to partial re-pressurization of the magma system, but Segall (2016) showed that such a reversal could result from relaxation of a viscoelastic aureole surrounding the magma reservoir. The current situation is reminiscent of the small amount of areal dilatation that occurred between 1982 and 1991, followed by no additional deformation during 1991–2003 and then by the 2004–2008 eruption.

Repeated measurements of residual gravity at Mount St. Helens show progressive increases in subsurface mass during 2010–2016 which, combined with seismic and deformation observations described above, suggest replenishment of the upper crustal magma system (Battaglia et al., 2018). However, models of the gravity and deformation data are non-unique and no single source of mass increase explains all of the observed gravity increase. Battaglia et al. (2018) showed: (1) addition of ∼50 × 10<sup>6</sup> m<sup>3</sup> of magma to the 5–6 km-deep magma reservoir that fed the 2004–2008 eruption can explain ∼20% of the gravity increase; (2) addition of ∼30 × 10<sup>6</sup> m<sup>3</sup> of magma to the upper ∼2 km of the conduit can explain ∼60% of the increase, but the corresponding amount of surface deformation (assuming incompressible magma and elastic response of host rock) would be too great to have gone undetected; and (3) accumulation of groundwater in a shallow aquifer can explain ∼20–60% of the increase.

Chemical and isotopic analyses of water and gas collected at Mount St. Helens during 2002–2016 likewise provide suggestive, but not compelling, evidence for replenishment of the upper crustal magma system since 2008. Bergfeld et al. (2017) reported an increase in δ <sup>13</sup>C values of dissolved carbon in hot springs during 2006–2009 and a similar shift in δ <sup>13</sup>C-CO<sup>2</sup> in bubble gas emissions, which they attributed to CO<sup>2</sup> from an undegassed body of magma distinct from the 2004–2008 magma. They noted that the increases in δ <sup>13</sup> were accompanied by rising trends in <sup>3</sup>He/4He ratios in fumarole gases, another indicator of fresh magmatic input. Also, air-corrected helium isotope values (RC/RA) were distinctly higher after 2011 than during 1986–2002, which Bergfeld et al. (2017, p. 109) interpreted as ". . .additional evidence for some involvement of new magma as early as 2006, and possibly earlier, given the unknown time needed for CO<sup>2</sup> and He to traverse the system and arrive at the springs." What remains unclear is the depth of the fresh magma, i.e., whether below or within the upper crustal reservoir. As is the case for deformation signals, geochemical indicators of fresh magma input are likely to be affected by the presence and maturity of mush zones in the magma column (Parmigiani et al., 2017), making the recognition of an injection event more difficult.

In summary, seismicity provides the clearest evidence for repressurization of the upper crustal reservoir since the end of the 2004–2008 eruption. A small amount of surface inflation is consistent with that idea, but it could also be caused by relaxation of a viscoelastic aureole around the reservoir. Increases in residual gravity can be explained partly by addition of magma to the reservoir or conduit (replenishment), but also could be explained by groundwater accumulation. Finally, there is evidence from thermal springs for degassing of fresh magma since the 2004–2008 eruption, but it is not clear whether the source is the upper crustal reservoir or somewhere deeper in the system.

#### Why Has It Been Difficult to Detect Changes in Crustal Magma Storage at Mount St. Helens, Even Since It Has Become One of the Most Intensively Studied Volcanoes on Earth?

The preceding discussion suggests that several factors contribute to the difficulty in detecting changes in crustal magma storage at Mount St. Helens. First, the volcano's upper crustal reservoir is ≥7 km deep, several to 10 s of km<sup>3</sup> in volume (melt plus mush), and long-lived (>10<sup>4</sup> yr). Second, there is no compelling evidence that magma has been added to the upper part of the reservoir (≤7 km depth) in the past 500 years. All of Mount St. Helens' eruptions during historical time (1800–1857, 1980–86, 2004–2008) can be understood as outcomes of an injection and mixing event prior to or during Kalama time (C.E. 1479–1720) (Pallister et al., 1992, 2008). That being the case, the likelihood of detecting seismic or deformation precursors sourced in the reservoir prior to 1980 is greatly reduced – especially with sparse monitoring networks at the time. The reservoir might have pressurized gradually as a result of second boiling, but the process would have spanned centuries and any surface deformation would have been muted by the reservoir's depth, and perhaps also by compressibility of reservoir magma and viscoelastic accommodation by the reservoir's host rock. Portions of the reservoir are likely rich in volatiles exsolved during formation of crystal mush zones (Parmigiani et al., 2017), and volatile-rich portions might be highly compressible. Thermal conditioning over the reservoir's long lifespan would have facilitated viscoelastic behavior, further mitigating any surface deformation signal. Second boiling might have produced long-period seismicity indicative of fluid movement, but of what magnitude and at what rate is unknown. Nichols et al. (2011) examined 60 deep long-period (DLP) earthquakes that occurred during 1980–2009 under six of the ten Cascade volcanoes in Washington and Oregon, including Mount St. Helens. They proposed that DLPs are produced by movement of magma and/or magmatic fluids in the mid-to-lower crust, and that they might

be an early indicator of renewed eruptive activity. Unfortunately, the minimal seismic network at Mount St. Helens prior to 1980 might not have detected DLPs if they occurred, so it is unknown whether their numbers or magnitudes increased prior to the start of intense shallow seismicity on March 20, 1980. None were detected by a local network of portable seismometers operated during March 20–May 18 (Weaver et al., 1981).

Another circumstance that makes Mount St. Helens a difficult monitoring target is the behavior of the crust below 2–3 km depth in response to magmatic processes. Very few earthquakes below that depth were detected before the May 18, 1980, eruption, while the volcano's 123-year quiescent period was drawing to a close. On the other hand, earthquakes indicative of magma withdrawal and subsequent tectonic adjustment occurred to depths of 20 km beneath the volcano in the aftermath of the eruption (Weaver et al., 1981; Barker and Malone, 1991; Moran, 1994). We can infer that the crust below 2–3 km depth deformed in non-brittle fashion under the strain rates that prevailed prior to the morning of May 18, 1980. Only during and soon after the eruption were strain rates high enough to induce appreciable brittle failure below 3 km, making early (pre-March 1980) recognition of impending unrest difficult.

Even with a much-improved seismic network in place since 1980 (**Figure 2**), the most seismically productive zone has been between 2–3 km depth and the surface (**Figure 3**). Likewise, the amount of surface deformation prior to the eruptions in 1980–1986 (minor beyond the north-flank bulge) and 2004–2008 (minor during 1982–1991, none during 1991–2003) is disproportionately small relative to the volumes of erupted material if one were to assume (1) the crust beneath the volcano behaves as a homogenous, elastic half-space, (2) reservoir magma is incompressible, and (3) the reservoir and conduit walls are rigid. None of these conditions are likely to prevail beneath real-world volcanoes.

On a positive note, Moran (1994) interpreted focal mechanisms for earthquakes at 6.5–10 km depth during 1987–1992 as evidence for repressurization of the magmatic system, a conclusion borne out by the 2004–2008 eruption. Likewise, Moran and Lisowski (2014) reported seismic and geodetic evidence for repressurization since the end of the 2004–2008 eruption. Both studies point to the likelihood that future eruptive activity will be foreseen well in advance.

At Mount St. Helens, it seems, departures from idealized behavior are large enough, at least at the strain rates imposed by activity since 1980, to have strongly influenced patterns of precursory activity. At other volcanoes, recent eruptions have produced different patterns in response to different sets of conditions. The following sections describe a few examples in varying amounts of detail to illustrate specific points that might be instructive for understanding Mount St. Helens' past and future behavior.

#### Lessons From Other Volcanoes

Not all arc volcanoes are as inscrutable as Mount St. Helens vis-à-vis eruption precursors. There have been many hundreds of eruptions at dozens of volcanoes since 1980, and precursory information is being compiled in several global volcano databases [e.g., Smithsonian Institution Global Volcanism Program<sup>2</sup> , World Organization of Volcano Observatories WOVOdat<sup>3</sup> , Volcano Disaster Assistance Program (VDAP) Eruption Forecasting Information System (EFIS)]. A comprehensive examination of eruption precursors would no doubt be instructive, but is beyond the scope of this paper. Instead, the following three sections focus on volcanoes that in recent decades have telegraphed their intentions earlier or with more clarity than Mount St. Helens. In these cases, seismic or geodetic precursors differed substantially from those at Mount St. Helens prior to its 1980–1986 and 2004–2008 eruptions. The differences reflect a few of many factors that undoubtedly influence precursory behavior, a topic to be addressed in a later section.

#### Mount Pinatubo, Philippines, 1991 – Rapid Injection of Basalt Into a Silicic Reservoir at a Long-Dormant Volcano

The June 15, 1991 eruption of Mount Pinatubo, Philippines (VEI 6, ∼5 km<sup>3</sup> DRE) was the world's largest since the 1912 eruption of Novarupta, Alaska (VEI 6, 13 ± 3 km<sup>3</sup> DRE). Novarupta was essentially unmonitored prior to the eruption owing to its remote location; as a result, little is known about precursory activity. Pinatubo had been dormant for ∼500 years prior to 1991 and it, too, was not monitored before April 1991. However, more than 30,000 people were living on its flanks when residents began feeling local earthquakes in mid-March, 1991, followed by a series of steam explosions in early April. The high and immediate hazard prompted the Philippine Institute of Volcanology and Seismology (PHIVOLCS) to establish a local observatory and begin monitoring within a few days of the first reports (**Figure 6**).

As activity escalated through April and May, PHIVOLCS with assistance from the Volcano Disaster Assistance Program (VDAP) expanded the monitoring effort to include seismicity, ground deformation, and gas emission. The resulting record, together with subsequent analyses of erupted products, enables a remarkably detailed reconstruction to be made of events leading up to the climactic eruption (Newhall and Punongbayan, 1996). As summarized by Pallister et al. (2008, p. 682):

Rapid addition of new magma and fluids to a crustal reservoir can increase pressure and trigger an eruption from below. This was the case for the eruption of Pinatubo in 1991, in which new hydrous mafic magma from depth entered a crystal-rich dacite reservoir, vesiculated, mingled, and created a buoyant plume, which rose through the viscous and crystal-rich upper part of the reservoir, increased pressure in the hydrothermal system, fractured a new pathway to the surface, and triggered the eruption (Pallister et al., 1996). In the Pinatubo case, deep longperiod seismicity recorded and tracked ascent of basalt from 35 or 40 km deep to the crustal reservoir at depths of less than 14 km before the eruption (White, 1996).

That scenario is strikingly different from the one that played out at Mount St. Helens in 1980. The volcanoes are similar in

<sup>2</sup>https://volcano.si.edu/

<sup>3</sup>http://www.wovodat.org/

some respects and so, too, were their eruptions' precursors—but with one important exception.

Hillshade is an ESRI Image Service product from 2013 accessed through ArcGIS Online and depicts post-1991 topography.

The crustal magma reservoirs (magma plus mush) beneath Mount Pinatubo and Mount St. Helens are both primarily dacitic. Both are large (40–90 km<sup>3</sup> and several to 10 s of km<sup>3</sup> , respectively) (Mori et al., 1996; Mastin et al., 2008) and long-lived, with a top near 7 km and a bottom (mush zone) near twice that depth (Scandone and Malone, 1985; Pallister et al., 1992, 1996, 2008; White, 1996). Both volcanoes had been dormant for more than a century before their most recent climactic eruption. In both cases, strong precursors began 2–4 months beforehand and included intense shallow VT seismicity. Far-field deformation was not detected in either case, but that might be an artifact of inadequate monitoring. The primary difference is that the Pinatubo reservoir ingested what was probably a substantial volume of basaltic magma in a relatively short period of time. Pallister et al. (1996) suggested the process started on or before April 2, 1991, the date of the first phreatic explosions, and continued episodically until the climactic eruption on June 15, 1991. A similar scenario might have played out at Mount St. Helens during Kalama time, but not since. In terms of eruption precursors, the difference was manifest most clearly by the absence of detected seismicity deeper than 2.5 km at Mount St. Helens in 1980 and the preponderance of deep seismicity, especially long-period earthquakes, at Pinatubo in 1991. The distinction is clouded by the fact that only one short-period seismometer (SHW, 3.5 km from the summit) was operating at Mount St. Helens when intense seismicity began on March 20, 1980. However, four seismic stations were added within 28 h of the swarm onset, and by March 30 a total of 11 stations were operating within 35 km of the volcano (4 within 10 km) (Malone et al., 1981) (**Figure 1**). There remains a possibility that deep earthquakes occurred but went undetected at Mount St. Helens prior to its climactic eruption on May 18, 1980. If so, it seems likely they were considerably less energetic than at Pinatubo in 1991, and possibly also less than at Redoubt in 2009 and Augustine in 2006 (see below).

White (1996) began his report on precursory DLP earthquakes at Pinatubo with a remarkable statement (p. 307): "About 600 deep long-period (DLP) earthquakes occurred beneath Mount Pinatubo in late May and early June 1991. This number is higher than the combined total number of such earthquakes previously reported at all convergent-margin

volcanoes worldwide." He interpreted the DLP seismicity as ". . .the elastic manifestation of the injection of deep-seated basaltic fluids into the base of the magma chamber." The contrast could not be starker: 600 DLPs in a matter of weeks at Pinatubo versus no known seismicity of any kind below 3 km depth during 2 months of precursory seismicity at Mount St. Helens. Clearly, DLPs can be an important indicator of magma injection into the crust. However, they are not a "magic bullet" for eruption forecasting, for at least three reasons: (1) some DLPs occur far from volcanic areas and might have non-magmatic source mechanisms; (2) during 1980–2009, more DLPs occurred under Mount Baker than any other volcano in Washington or Oregon, including Mount St. Helens; and (3) no DLPs in the study were associated with volcanic activity, including the 1980–1986 and 2004–2008 eruptions at Mount St. Helens (Nichols et al., 2011).

We might suspect that crustal injection events also have distinctive deformation signatures. Unfortunately, that suspicion cannot be confirmed or rejected for Mount Pinatubo, where pre-eruption monitoring was not adequate to have detected a deep-seated deformation signal. However, such signals have been recognized at arc volcanoes elsewhere. Two cases are instructive here.

#### Redoubt Volcano, Alaska, 2009 – Monitoring Data Track Months-Long Ascent of Magma From Deep in the Crust

Prior to its 2009 eruption, Redoubt Volcano had last erupted in 1965–1968 and 1989–1990, each time following a hiatus comparable to that at Mount St. Helens from 1986 to 2004. The 2009 activity at Redoubt included dozens of explosive events, extrusion and destruction of at least three, and possibly four lava domes, ending with extrusion of a final dome with a bulk volume of 72 × 10<sup>6</sup> m<sup>3</sup> . Total volume erupted was 80–120 × 10<sup>6</sup> m<sup>3</sup> . Geodetic, gas-emission, seismic, and petrologic evidence support a complex eruption scenario in which a deeply sourced, upward-moving, low-silica andesite encountered and mixed with high-silica andesite in the mid- to upper crust, forming a third, intermediate-silica andesite. All three magma types erupted during the 2009 activity (Bull and Buurman, 2013).

Compared to Mount St. Helens in 1980 and 2004, precursors to the 2009 Redoubt eruption were early, conspicuous, and telling. The nearest permanent CGPS station, 28 km northeast of Redoubt (AC17, installed in 2006 as part of the Plate Boundary Observatory), began recording subtle radially-outward motion in May 2008 (Grapenthin et al., 2013) (**Figure 7**). An H2S odor was reported by field geologists working on the edifice in July 2008, and an aircraft pilot reported a sulfur odor and increased snow melt in September 2008. Airborne gas-emission measurements detected anomalous levels of H2S, SO2, and CO<sup>2</sup> during overflights starting in October 2008 (Werner et al., 2013). A pronounced increase in lower crustal DLP and VT earthquakes began in mid-December 2008: 30 DLP events were located at depths of 28–35 km beneath Redoubt from December 12, 2008, to December 31, 2010 (Power et al., 2013). Two months of discontinuous shallow volcanic tremor started on January 23, 2009, culminating in a phreatic explosion on March 15, 2009. Extrusion of the final dome ended by July 1, 2009 (Bull and Buurman, 2013).

As was the case at Pinatubo in 1991, pre-eruption monitoring data and post-eruption petrologic analyses enable detailed inferences to be made regarding Redoubt's magmatic system and the eruption scenario that played out in 2009. Elevated CO<sup>2</sup> fluxes and high CO2/SO<sup>2</sup> ratios detected in airborne samples starting 6 months prior to the eruption onset most likely resulted from degassing of magma at mid- to lower crustal depths or of magma that recently ascended from such depths (Werner et al., 2013). Grapenthin et al. (2013) modeled surface displacements at AC17 plus four temporary CGPS stations that were installed several weeks before the eruption onset at distances of 4–12 km from Redoubt. They identified two and possibly three deformation sources in the mid- to upper crust: (1) a Mogi source at 13.5 km depth, (2) a vertical prolate spheroid at 7–11 km depth, and (3) a secondary reservoir at 2–4.5 km depth that was inferred from petrologic evidence (Coombs et al., 2013) but poorly resolved by the geodetic data.

Uncertainties in the deformation models are such that sources 2 and 3 are compatible with the conceptual model proposed by Power et al. (2013, p. 42–43) based on seismic observations:

. . .the Redoubt magmatic system consists of a diffuse magmatic source zone at depths of 28 to 32 km depth, a mid-crustal magma storage area at depths of roughly 3 to 9 km and a conduit and system of interconnected cracks that extends from the mid-crustal storage zone to the Redoubt Crater floor. The deeper diffuse magmatic source area is defined by the hypocenters of DLP events and VT earthquakes that are observed at these depths before and after the 2009 eruption. The mid-crustal storage area is defined by hypocenters of VT earthquakes that increased in rate following the initiation of eruptive activity in December 1989 and April 2009 and are a persistent feature of the seismic record between 1989 and 2010. The shallow system of cracks is defined seismically by repetitive swarms of LP events, small VT earthquakes, and hybrid events that occur most prevalently during the eruptions in 1989–1990 and 2009.

The description by Power et al. (2013), in turn, is congruent with the 2009 eruption scenario that Coombs et al. (2013) inferred from geochemical and petrologic observations: (1) low-silica andesite (LSA) ascended from the lower crust, i.e., from the DLP and VT seismic source at 28–35 km depth, during or before precursory activity starting in May 2008; (2) ascending LSA encountered differentiated, high-silica andesite (HSA) remnant from pre-2009 intrusions at about 13 km depth (deformation source 1); (3) mixing of LSA and HSA produced intermediate-silica andesite (ISA) at depths of 7–11 km (deformation source 2); (4) ascent stalled temporarily at depths of 2–4.5 km (deformation source 3), where 2009 magmas staged for several months before eruption.

Redoubt Volcano shares several characteristics with Mount St. Helens. Both volcanoes have erupted a range of compositions from basalt to dacite during their histories (Mullineaux and Crandell, 1981; Till et al., 1994). Recent eruptive products are only slightly more mafic at Redoubt (∼58–63 wt.% SiO<sup>2</sup> in

accessed through ArcGIS Online.

1989–1990 and 2009) than at Mount St. Helens (∼62–65 wt.% SiO<sup>2</sup> in 1980–1986 and ∼65 wt.% SiO<sup>2</sup> in 2004–2008). There is evidence in both cases for crustal magma storage at depths below ∼7 km, although the size and longevity of the Redoubt reservoir are unknown. Eruption intervals at both volcanoes are broadly similar. Mount St. Helens erupted repeatedly during 1800–1857, 1980–1986, and 2004–2008. Redoubt erupted in 1902, 1966–1968, 1989–1990, and 2009; there are unconfirmed reports of eruptions in 1881 and 1933 (Alaska Volcano Observatory, 2018a). The character and size of recent eruptions are similar. Redoubt erupted explosively more than 20 times during 1989–1990. Fourteen lava domes were extruded, 13 of which were destroyed by explosive activity. Total bulk volume of erupted products was 0.1–0.2 km<sup>3</sup> (Miller, 1994; Miller and Chouet, 1994). At Redoubt in 2009 there were more than 19 explosive events, two and possibly three domes were extruded and later destroyed, and a fourth survived. Total erupted volume was ∼0.1 km<sup>3</sup> . At Mount St. Helens there were six explosive eruptions in 1980, two lava domes were extruded and destroyed, and a third dome grew episodically during 1980–1986. Total erupted volume was ∼0.3 km<sup>3</sup> DRE. The 2004–2008 eruption was dominantly extrusive; bulk volume of the resulting dome was ∼0.1 km<sup>3</sup> .

These similarities belie some striking differences in precursory activity, only some of which can be attributed to differences in monitoring capability, and in the nature of erupted products. As was the case at Pinatubo in 1991, both the 1989–1990 and 2009 eruptions at Redoubt produced banded pumices indicative of magma mixing (Nye et al., 1994; Coombs et al., 2013). Such evidence is lacking at Mount St. Helens in 1980–1986 and 2004–2008 (Pallister et al., 1992, 2008). At Redoubt, the earliest precursors (sulfur odor, increased snow melt, far-field deformation) were recognized, at least in hindsight, 8–10 months before the start of the 2009 eruption. DLP and deep VT seismicity increased about 3 months before the eruption onset. Power et al. (2013, p. 40) suggested that a similar increase might have accompanied the 1989–1990 eruption, noting that some events could have escaped detection. It is clear that no such increase in deep seismicity preceded the 2004–2008 eruption at Mount St. Helens. None was detected before the 1980–1986 eruption, although the seismic network at the time might have lacked sufficient sensitivity. The same is true for far-field deformation:

none was detected prior to 1980 and very little occurred prior to 2004. Any change in gas emission, if it occurred prior to the onset of shallow-seated unrest in 1980 or 2004, went unnoticed (Gerlach et al., 2008). Limited monitoring capability, especially at Mount St. Helens before 1980, might account for some of the differences in recognized precursory activity. However, the preponderance of evidence (DLP seismicity, banded pumices) suggests the recent Redoubt and Pinatubo eruptions share one characteristic that Mount St. Helens in 1980–1986 and 2004–2008 does not, i.e., magma ascent from the lower crust, mixing, and mobilization of magma stored in a crustal reservoir.

#### Augustine Volcano, Alaska – Long-Term Changes in Crustal Magma Reservoirs Tracked With InSAR and CGPS

Augustine Volcano in Cook Inlet, Alaska, erupted frequently during the 20th and early 21st centuries, most recently in 1976, 1986, and 2006 (Alaska Volcano Observatory, 2018b). Lee et al. (2010) examined more than 50 ERS-1 and ERS-2 radar interferograms (InSAR images) of Augustine Island that collectively span 1992–2005. Using a refined small baseline subset (SBAS) InSAR technique, they produced a time series of observations that showed 2–8 cm uplift of the entire volcano during the 13-year period of investigation – slightly less on the upper flanks than on the lower flanks. Lee et al. (2010) showed the InSAR deformation field was consistent with an expanding Mogi source 7–12 km BSL and a contracting Mogi source 2–4 km BSL, which they interpreted as magma storage zones.

Augustine erupted explosively on January 11, 2006, soon after the end of the InSAR time series analyzed by Lee et al. (2010). The eruption included: (1) an initial explosive phase (January 11–28, 2006) with 13 discrete explosive events and extrusion of one or perhaps two lava domes; (2) a continuous phase (January 28–February 10, 2006) with essentially continuous ash emission, pyroclastic flows, occasional seismic signals thought to represent more explosive events with associated ash clouds, and steady effusion of a lava dome and flow starting on February 3; and (3) an effusive phase (March 13–16, 2006) with growth of a larger summit dome, renewed growth of the earlier lava flow, and formation of a second flow – all accompanied by vigorous block-and-ash flows (Power et al., 2006; Coombs et al., 2013). Some of the larger explosive events during the first two phases of the eruption were accompanied by long-period earthquakes at 1–5 km depth BSL, which Syracuse et al. (2011) attributed to magma transport from a storage zone, through a conduit, to the surface. In light of experience at Pinatubo and Redoubt discussed above, and given clear evidence for magma mixing at Augustine prior to its 2006 eruption (see below), we suspect that deeper LPs also occurred at Augustine as a result of magma injection into a crustal reservoir. That no DLPs were detected might be an artifact of the small aperture (<10 km) of the Augustine seismic network, which is confined to Augustine Island (**Figure 8**).

Larsen et al. (2010) reported that the 2006 eruption produced five major lithologies: (1) low-silica andesite (LSA) scoria, mostly during the initial explosive phase; (2) high-silica andesite (HSA) pumice, prevalent during the continuous phase; (3) dense low-silica andesite, predominantly during the late effusive phase; (4) dense intermediate andesite, erupted during all phases but most prevalent in deposits of the continuous phase; and (5) banded clasts of any combination of the above, present throughout the eruption but most abundant in the continuous phase. Based on petrological and geochemical evidence, they concluded (p. 335):

. . .the HSA was stored as a crystal-rich mush with its top at ∼5-km depth. An influx of basalt remobilized and partially mixed with a portion of the mush, forming the hybrid LSA. The lower viscosity LSA ascended toward the surface as a dike, erupting during the explosive phase in mid-January 2006. In late January, a large explosion produced the first significant volumes of HSA, followed by several days of rapid HSA effusion during the eruption's continuous phase. After a 3-week hiatus marked by elevated gas output, signifying an open system, degassed LSA erupted during the final, effusive phase.

Cervelli et al. (2006; 2010) analyzed data from 11 permanent or semipermanent CGPS stations on the island that collectively span the eruption (**Figure 8**). Precursory inflation that began in mid-August 2005, 5 months before the eruption onset, can be modeled with a Mogi source located under the summit area and near sea level. Cervelli et al. (2010) attributed the inflation to pressurization by volatiles trapped near the impermeable base of the edifice—volatiles we can infer were released from the contracting magma reservoir 2–4 km BSL that Lee et al. (2010) identified. Cervelli et al. (2010) reported that deformation switched from inflation to deflation in late January 2006 during the period of continuous explosive activity. They modeled the source as a cylindrical magma body with a top depth of 3.5 ± 1.0 km and a bottom depth of 8.5 ± 2.0 km below Augustine's summit (1.2 ± 1.0 km BSL and 7.2 ± 2.0 km BSL, respectively)—perhaps the top of, or a conduit above, the expanding magma storage zone at 7–12 km BSL modeled by Lee et al. (2010), which in turn might be related to the crystal mush zone with a top at ∼5-km BSL identified by Larsen et al. (2010).

The salient points here are: (1) InSAR evidence for long-term pressurization of a magma storage zone in the mid- to upper crust, (2) several months of precursory deformation from a shallower source tracked by CGPS, (3) deflation of the crustal reservoir during the continuous phase of the eruption, also tracked with CGPS, and (4) clear evidence for magma mixing triggered by an injection of basaltic magma into a crustal reservoir. Inflation of the shallow source prior to the eruption followed by deflation of the deeper source during the eruption implies a pressure connection between the two, initially for volatiles and eventually for the magma erupted in 2006. The apparent absence of DLP seismicity prior to the eruption, even though there is evidence for an influx of basalt to the crustal storage zone, might be an artifact of the small-aperture island seismic network.

#### What Additional Precursors, if Any, Should Be Expected Before Mount St. Helens' Next Eruption?

Considering recent experience at Mount St. Helens and the three volcanoes profiled above, a case can be made that another dome-forming eruption at Mount St. Helens could begin with

product from 2013 accessed through ArcGIS Online.

little warning beyond a few days of shallow seismicity and localized ground deformation. If patterns that emerged prior to the start of the 2004–2008 eruption repeat, we should expect more seismicity near 2 km depth, indicating formation of a conduit plug, and perhaps a general lull in seismicity before the plug eventually is overcome by slowly increasing magma pressure. The likelihood of little advance warning is greater if the next event is another in a progression of dome-forming eruptions that began in 1980–1986 and continued during 2004–2008, i.e., episodic ascent of relatively gas-poor magma remnant from the 1980 explosive activity. The case for this scenario is stronger if the 1980 activity was initiated by an injection event 500 years ago during Kalama time and not by a more recent perturbation of the upper crustal reservoir, as suggested by petrologic evidence cited above. In fact, progressively smaller compositional oscillations in erupted products from Kalama time through 2004–2008 (Pallister et al., 1992, 2008), combined with the remarkably gas-poor and low-temperature character of 2004–2008 lava (Gerlach et al., 2008; Vallance et al., 2008), suggest the cycle that began in Kalama time might be coming to an end. If so, the next eruption might not occur until there is another injection of gas or magma into the upper crustal reservoir to initiate a new cycle.

If an injection event were to trigger the next eruption, the case studies cited above suggest the eruption would be heralded by detectable DLP seismicity, far-field ground deformation, and increased gas emission, likely for several months or more unless the injection were large enough or rapid enough to mobilize reservoir magma or "blow through" the reservoir more quickly.

#### Factors That Affect Precursory Activity

The preceding sections call attention to a spectrum of precursory activity at arc volcanoes and, by implication, to a multitude of factors that might influence a volcano's behavior prior to eruption. Following is a partial list of those factors, their likely effects, and implications for optimizing monitoring networks and strategies.

Pressurization of crustal magma reservoirs can result from ingestion of fluids (magma or volatiles) from below or exsolution of volatiles from already-resident magma during second boiling. With adequate monitoring, both processes are likely to produce

detectable seismicity, ground deformation, and gas emission. If the fluid is magma and the intrusion is large enough or shallow enough, there also will be measureable changes in residual gravity; less so if the fluid is composed of volatile species (e.g., H2O, SO2, and CO2). Experience at Pinatubo and Redoubt suggests the intensity of DLP seismicity might scale with the size of the injection/mixing event: large or rapid > small or gradual, although the sample size is small and many factors probably influence the intensity of DLP occurrence. The same is probably true for ground deformation, which might be more pronounced and easier to detect if: (1) the intrusion is large and rapid, (2) the reservoir is relatively shallow, (3) the magma involved (intruding and resident) is relatively incompressible, (4) the reservoir walls are relatively rigid, and (5) the host rock is elastic. Viscoelastic behavior will tend to mute both seismicity and deformation, meaning that long repose periods and small reservoir size (i.e., colder plumbing systems) are conducive to stronger precursory activity than short reposes or large, long-lived reservoirs. Ascending silicic magmas, by virtue of their greater viscosity and gas content, are expected to produce more precursory VT and LP seismicity than mafic magmas. On the other hand, the greater compressibility of gas-rich magmas means that more can be accommodated in a given volume without producing as much deformation of host rock as would be the case for gas-poor magma. For that reason, gravity measurements are especially important for detecting accumulation of gas-charged, potentially explosive magma that might produce LP seismicity but relatively little ground deformation. Unfortunately, with current instrumentation and techniques, it might not be possible to detect a change in residual gravity if the mass change is small or the accumulation zone is deeper than a few kilometers.

#### Implications for Monitoring Strategies and Modeling Studies Monitoring

If the next eruption at Mount St. Helens occurs relatively soon (within a few decades) and taps the same 1980-residual magma that fed the 1980–1986 and 2004–2008 dome-forming eruptions, we should expect precursors similar to those observed prior to the 2004–2008 event: (1) sporadic swarms of small VT earthquakes at depths of 2–10 km, indicating magma-system pressurization, (2) intensification of seismicity near 2 km depth, indicating incipient fracturing of a conduit plug, (3) a days-long swarm of intense shallow seismicity indicating cascading failure of the plug, accompanied by (4) localized, intense deformation of the crater floor or Crater Glacier (formation of a "welt"), and (5) slow extrusion of gas-poor dacite, perhaps preceded by phreatic eruptions when rising magma encounters shallow groundwater. A small increase in residual gravity might also be detected as magma accumulates aseismically beneath the conduit plug. If the current repose lasts more than a few decades and ends with another dome-forming eruption, precursors might be more protracted and intense because the conduit plug would have become more resilient. If there is an explosive onset, beyond phreatic explosions involving groundwater, magmatic

volatiles would be implicated. In that scenario, we might expect LP seismicity and subtle far-field deformation as the magma system pressurizes; some increase in magmatic gas emission also is possible if the system is not completely sealed. Gaunt et al. (2014) showed that this is likely to be the case. Their study of 2004–2008 dome rocks revealed a marginal shear zone along the conduit wall that greatly facilitates upward mobility of volatiles. Volatile escape would slow the pressurization process and thus delay the onset of shallow seismicity, as seems to be the case as this is written in June 2018.

these concepts provide a useful framework for this paper's discussion of

activity at Mount St. Helens since its reawakening in 1980.

An alternate scenario is that the next eruption does not occur until the upper crustal reservoir is replenished during an injection event. In that case, precursors including energetic DLP seismicity, pronounced far-field ground deformation, increased gas emission, and residual gravity

increases might herald the eruption several months or more in advance.

An effective monitoring strategy for Mount St. Helens must address the full range of plausible eruption scenarios in order to capture precursors that might range from subtle to obvious, localized to widespread, and protracted to sudden. Such a strategy was outlined by Ewert et al. (2005) as a framework for the National Volcano Early Warning System (NVEWS). A central tenet of the NVEWS approach is to avoid "playing catch-up" with a dangerous volcano, a situation that arises when a volcano awakens before a robust monitoring system is deployed, necessitating a suboptimal response. The proactive NVEWS approach is reminiscent of the conclusion by Dzurisin (2000, p. 1564) regarding the need for comprehensive deformation monitoring: ". . .to distinguish among the full range of possible source locations and geometries, especially if multiple sources might be present, it is necessary to make measurements virtually 'everywhere, all the time'. . ." In this context, "everywhere" refers to the area susceptible to surface deformation from known or plausible sources, under the assumption that deformation is elastic. For the upper crustal reservoir at Mount St. Helens (7–14 km depth), the potentially deforming area extends more than 20 km from the volcano. In order to account for tectonic deformation, a regional network of CGPS stations at even greater distances is necessary. "All the time" means continuously, because deformation events at volcanoes can be sudden and/or reversible, especially given the complex rheology of the subsurface.

Mount St. Helens was among 18 United States volcanoes that Ewert et al. (2005) identified as "very high threat" – an assessment based on a combination of hazards and exposure. Current monitoring networks at Mount St. Helens generally conform to the recommendations made by Moran et al. (2008b) for very high threat volcanoes. Monitoring networks for seismic activity,

ground deformation, and gas emission are capable of detecting the types and levels of precursory activity to be expected before the next eruption, at least a few days (renewed dome building) to several months (new explosive eruptive cycle triggered by magma or fluid injection) in advance. In light of experience in 2004, when preeruption deformation was confined to within a few hundred meters of the vent, it would be prudent to install two or more additional CGPS stations on the 1980–1986 dome or 2004–2008 dome (the only proximal areas not covered by Crater Glacier). Currently, only one CGPS station is operating this close to the vent, at the former site of DOM1, and the next closest station is ∼2 km away (**Figure 2**).

#### Modeling

A detailed conceptual model of Mount St. Helens' magmatic system and behavior has emerged from nearly four decades of monitoring information and related research. The model applies remarkably well to three other arc volcanoes profiled above, and shares essential features with a generalized volcano model discussed by McCausland et al. (2017) (**Figure 9**).

Essential features of the model are: (1) a deep source of magma generation in the upper mantle/lower crust; (2) a primary magma reservoir in the mid-to-upper crust (7–14 km depth); (3) subsidiary magma storage above the primary reservoir along a conduit system to the surface; (4) a plastic zone surrounding the magmatic system and extending to within a few kilometers of the surface, which serves to mute both brittle-failure seismicity and surface deformation except at high strain rates or large, relatively sudden pressure changes; and (5) crystal mush zones in the magma column that have a similar mitigating effect on geophysical and geochemical signals (Blake, 1984; Tait et al., 1989; Woods and Cardoso, 1997).

Concurrent with the evolution of conceptual volcano models has been the development of physics-based models, most recently in a Bayesian framework (e.g., Anderson and Segall, 2011, and references therein). As the name implies, physics-based models apply physical principles and observational constraints to infer the behavior of magmatic systems (**Figure 10**). This approach has been applied to the 2004–2008 eruption at Mount St. Helens (Anderson and Segall, 2013; Wong et al., 2017) and to the 1983–present eruption at K¯ılauea Volcano, Hawai'I (Anderson and Poland, 2016, 2017) with good success.

It would be useful to apply such a physics-based approach to eruption precursors in general and to some of the specific issues raised here. For example, it is plausible that there has been little or no input to Mount St. Helens' primary magma storage reservoir since Kalama time, and all subsequent eruptions are part of a declining cycle that might have come to an end in 2008? What conditions in the subsurface are necessary for that scenario to be viable? Is it reasonable that gas-poor, near-solidus-temperature magma rose from ∼5 km depth to the surface without causing brittle-fracture earthquakes deeper than 2 km? What are the implications for the temperature structure and rheology in country rock near the conduit? What fracture permeability is required for circulation of meteoric water to have cooled the upper 2 km of the conduit from 1986 to 2004? Is there sufficient gas left in 1980 magma to initiate another dome-forming episode, or is additional fluid input required from below? How far into the future will additional dome-building be possible without tapping a new batch of more gas-rich magma? Such questions would be difficult to answer with any certainty, given limited knowledge about subsurface conditions and even about the physical processes involved. But the Bayesian framework is well-suited to such a challenge, and physics-based models have the potential to evaluate ad hoc explanations and thus to improve conceptual volcano models.

#### CONCLUDING STATEMENT

Mount St. Helens since 1980 has provided an opportunity to study the behavior of an arc volcano from reawakening through a climactic eruption and its decades-long aftermath. Many questions have arisen and answers have been proposed. Behaviors that seemed enigmatic at the time (e.g., 1980 bulge, 2004 welt, and general lack of long-term or deep-seated precursors) now have plausible explanations in the form of a conceptual model supported by extensive research and modeling. Understanding Mount St. Helens has been difficult: in some respects, the volcano has been more inscrutable than anticipated. But with nearly four decades of experience, remaining questions have come into better focus and the tools necessary to answer them are at hand. We end where we began, by asking questions. The answers to those questions are both a challenge and an opportunity future research.

#### AUTHOR CONTRIBUTIONS

The author confirms being the sole contributor of this work and approved it for publication.

#### FUNDING

This research was supported by the USGS Volcano Hazards Program and USGS Volcano Science Center through the Cascades Volcano Observatory project WC00CKD.

#### ACKNOWLEDGMENTS

Volunteer Behnanz Hosseini used her considerable ingenuity and GIS skills to create **Figures 1**, **2**, **6**, **7**, and **8**. Mike Lisowski thoroughly analyzed the JRO1 time series data and produced **Figure 5**. Wes Thelen and Greg Waite provided original versions of **Figures 3** and **4**, respectively. John Pallister reviewed an early version of the manuscript, encouraged me to carry on, and managed to make petrology understandable to a geodesist. Thorough reviews by Raffaello Cioni and a reviewer improved the presentation considerably.

#### REFERENCES

feart-06-00142 October 5, 2018 Time: 12:18 # 21





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

This work is authored by Daniel Dzurisin on behalf of the U.S. Government and, as regards Dr. Dzurisin and the U.S. Government, is not subject to copyright protection in the United States. Foreign and other copyrights may apply. 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.

## Renewed Explosive Phreatomagmatic Activity at Poás Volcano, Costa Rica in April 2017

Rebecca O. Salvage1,2 \*, Geoffroy Avard<sup>1</sup> , J. Maarten de Moor <sup>1</sup> , Javier F. Pacheco<sup>1</sup> , Jorge Brenes Marin<sup>1</sup> , Monserrat Cascante1,3, Cyril Muller <sup>1</sup> and María Martinez Cruz <sup>1</sup>

<sup>1</sup> Observatorio Vulcanológico y Sismológico de Costa Rica, Universidad Nacional, Heredia, Costa Rica, <sup>2</sup> Department of Geoscience, University of Calgary, Calgary, AB, Canada, <sup>3</sup> Department of Earth Sciences, University of Oregon, Eugene, OR, United States

#### Edited by:

Corentin Caudron, Ghent University, Belgium

#### Reviewed by:

Wendy A. McCausland, Volcano Disaster Assistance Program (USGS), United States Lauriane Chardot, Earth Observatory of Singapore, Singapore Andrew Bell, University of Edinburgh, United Kingdom

> \*Correspondence: Rebecca O. Salvage beckysalvage@gmail.com

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 01 February 2018 Accepted: 24 September 2018 Published: 16 October 2018

#### Citation:

Salvage RO, Avard G, de Moor JM, Pacheco JF, Brenes Marin J, Cascante M, Muller C and Martinez Cruz M (2018) Renewed Explosive Phreatomagmatic Activity at Poás Volcano, Costa Rica in April 2017. Front. Earth Sci. 6:160. doi: 10.3389/feart.2018.00160 Phreatic and phreatomagmatic eruptions at volcanoes often present no short term precursory activity, making them a challenge to forecast. Poás volcano, Costa Rica, exhibits cyclic activity with phreatic and some phreatomagmatic eruptions separated by times of quiescence. The latest phreatomagmatic stage began in March 2017 with increases in crater lake temperatures, SO<sup>2</sup> flux, and the rate of seismicity, as well as accelerated ground inflation near the active crater. On 23 April 2017 at 04:12 UTC, a large phreatomagmatic eruption occurred at Poás, sending blocks up to 1 m in length to distances >1 km. Hindsight analysis revealed a precursory seismic sequence from 25 March to 22 April of similar seismic events (in terms of their frequency and waveform characteristics). Fourteen families of similar seismic events (containing ≥10 events per family) were identified during this precursory sequence, totaling over 1,300 events. An acceleration within the dominant family of LF (low frequency) waveforms was identified, suggesting that a forecast for the onset of the eruption may have been possible using the Failure Forecast Method (FFM). However, no confidence could be placed in the forecast generated, reiterating that not all accelerating trends are suitable for analysis using the FFM, in particular in conjunction with a least-squares linear regression. Our residual analysis further supports the concept that using a least-squares linear regression analysis is not appropriate with this dataset, and allows us to eliminate commonly used forecasting parameters for this scenario. However, the identification of different families of similar seismicity allows us to determine that magmatic fluid on its way to the surface initially became stalled beneath a chilled margin or hydrothermal seal, before catastrophically failing in a large phreatomagmatic eruption. Additionally, we note that 24 h prior to the large phreatomagmatic eruption, all LF families became inactive, which could have been falsely interpreted in real time as the waning of activity. Our results suggest that identifying families of seismicity offers unique opportunities to better understand ongoing processes at depth, and to challenge conventional forecasting techniques.

Keywords: Póas volcano, failure forecast method, similar seismicity, families, eruption forecasting, phreatomagmatic eruption

#### 1. INTRODUCTION

Phreatomagmatic eruptions at volcanoes occur when magmatic fluid or gases migrating toward the surface interact explosively with ground or surface water. At the point of direct contact between the two, the magmatic fluid or gases, which are of a higher temperature than the static water, rapidly cools, and the static water abruptly heats and expands due to a rapid phase change from liquid to gas, generating an explosive eruption (Büttner et al., 2002). The products of these types of eruptions are usually fine ash deposits (as a result of the energetic fragmentation process of rising magma and the subsequent fragmentation of the surrounding edifice; Büttner et al., 1999), hydrothermal fluids, and large blocks in the vicinity of the erupting vent. Lahars are also common, depending on the amount of water at the surface and the local topography (Barberi et al., 1992), meaning these eruptions can pose a serious hazard to local populations.

Precursors to phreatic eruptions are often short-lived, and are often difficult to distinguish from normal background levels. The September 2007 eruption of Mt. Ruapehu, New Zealand, produced a steam column up to 15,000 feet, ballistic and surge flows, and lahars without any usable precursors indicative of an impending eruption (Christenson et al., 2010; Jolly et al., 2010). Similarly, the September 2014 eruption of Mt. Ontake, Japan, which left 64 people dead or missing, showed very little precursory activity with only small changes in the amplitude of recorded tremor, a small migration in high frequency seismicity and anomalous tilt measurements <10 min before the onset of eruption (Kato et al., 2015; Oikawa et al., 2016), and one very long period earthquake in the preceding 25 s (Maeda et al., 2015). Poás volcano, Costa Rica, on which this study is focused, showed very short-lived precursory fluctuations in lake gas composition (between 24 and 36 h) prior to a number of phreatic eruptions between April and June 2014 (de Moor et al., 2016). In other cases, precursors are identifiable but may not be able to be processed and analyzed in an appropriate time frame to give an alert, for example, the 1990 eruption of Kelut volcano, Java, which was preceded by spasmodic tremor in the hours prior to a number of phreatic explosions (Lesage, 1995). These types of eruptions are therefore particularly difficult to forecast in terms of timing and intensity, especially when lacking precursory activity, principally as groundwater reservoir locations and heat flow within the volcano are often poorly constrained. In addition, phreatic eruptions may not always involve the movement of magma toward the surface (which usually generates precursory signals) and may occur simply due to changes in the shallow hydrothermal system or variations in gas input (Rouwet et al., 2014; de Moor et al., 2016).

Here we report on the largest phreatomagmatic explosive eruption to occur at Poás volcano during the latest phase of unrest, which was first identified in real-time by the staff of the Observatorio Vulcanológico y Sismológico de Costa Rica Universidad Nacional (OVSICORI-UNA) on 26 March 2017, and lasted several months. A number of geophysical and geochemical parameters rapidly increased from this date onwards, suggesting a sudden renewal of activity, and a number of smaller phreatomagmatic eruptions were indeed identified in early April 2017, which became more energetic with time. As unrest was already elevated, further activity suggesting a much larger phreatomagmatic eruption was imminent was not identified in real time, but did occur on 23 April at 04:12 UTC. We show that although the similar seismicity identified at Poás volcano produced an accelerating trend, it was not possible to use it in conjunction with the Failure Forecast Method to generate a successful forecast. However, the techniques employed allowed the identification of far greater numbers of events compared to a traditional amplitude based detection algorithm or manual identification by OVSICORI-UNA personnel. The identification of families of similar seismicity also allowed us to gain a more detailed insight of ongoing processes at depth, including a better understanding of the number of active sources beneath the volcano.

### 2. POÁS VOLCANO, COSTA RICA

Poás volcano, Costa Rica (N 10.1968; W 84.2305, ∼2,300 m a.s.l., **Figure 1**) is a large stratovolcanic complex located in the central valley of the country, 22 km north of the main international airport (SJO) and 35 km north-east of the capital city San José, where approximately one third of the countries' population live. Poás has three main volcanic centers: Botos cone, containing a cold, mildly acidic crater lake, which last erupted approximately 7,500 years ago (Prosser and Carr, 1987; Alvarado-Induni, 2005); the von Frantzius crater, an extinct volcanic cone (Casertano et al., 1983); and the principal crater, which is usually partially filled with the Laguna Caliente, a warm, very acidic crater lake (20–60◦C, pH ≤ 1.8), which experiences frequent phreatic and more occasional phreatomagmatic eruptions (de Moor et al., 2016). The Laguna Caliente is the result of meteoric water interacting with hydrothermal-magmatic gases and rocks. Prior to its closure in April 2017, Poás was the second most visited national park in the country, with more than 400,000 visitors per year. Eruptions at this volcano consequently pose a great threat to those in this vicinity, and in particular to tourism, both in terms of health and economy.

Historical activity at Poás is reported as early as 1828 (strong degassing with sporadic phreatic eruptions; Casertano et al., 1983), with the first phreatomagmatic eruption reported on 25 January 1910, when a large steam and ash cloud reached 4 to 8 km above the summit (Martínez et al., 2000). The last period of strong phreatomagmatic activity occurred between 1953 and 1955 and saw the emission of both ash and larger bombs from the principal crater, followed by the complete drying up of the Laguna Caliente (Casertano et al., 1983). This eruptive period resulted in the creation of a small dome-like feature at the edge of the Laguna Caliente, growing up to 45 m high at its greatest, partly made up of primary pyroclastic deposits (Rowe et al., 1992). The crater lake returned by 1961 with sporadic geyserlike phreatic eruptions over the coming decades, sometimes with associated ash and rock emissions (Martínez et al., 2000). More recent activity occurred in 2005–2008 and 2011, 2014, and 2016 as reported by the OVSICORI-UNA (e.g., Global Volcanism

Reference location of Costa Rica and Poás). The active crater (Laguna Caliente) is represented by the black diamond; and the dormant Botos crater is shown as a gray diamond to the south-east of the active crater. The Mirador outlook was the viewing station of the National Park for observation by tourists into the active crater, which was closed on 13 April 2017. Seismic and GPS stations are named as referenced in the text. CRPO, the Multi-GAS station (purple star) and the camera (green star) were all destroyed by eruptions in April 2017. (B) Photograph of Poás Laguna Caliente, 12 March 2017 from the Mirador before levels of unrest increased. Photo Credit: R. O. Salvage. (C) Example phreatomagmatic eruption at Poás crater, 19 April 2017 from the Mirador. Photo Credit: J. F. Pacheco.

Program, 2016). Intense fumarolic activity, in particular in and around the edges of Laguna Caliente, was reported between 1980 and 1985, 1986–1991, 1994, 1999, and 2005–2008. New fumaroles opened on the south-western inner crater walls between 1995 and 2000, and toward the east between 1999 and 2007, away from the Laguna Caliente (Martínez Cruz, 2008). Between 2009 and 2011 numerous phreatic eruptions occurred with column heights varying from 5 to 500 m. In the latter half of 2011, fumaroles around the dome reached temperatures up to 980◦C. Further smaller phreatic eruptions occurred in April, September, and October 2012. On 1 and 2 May 2013, a series of phreatic eruptions occurred, which were preceded by a clear increase in seismicity from 29 April onward (≥200 low frequency events, above a background level of 50 events/day), as well as an increase in the temperature of active fumaroles on the dome from 102◦C on 30 January to 380◦C on 16 April (Fischer et al., 2015). In 2014, 46 phreatic eruptions with varying plume heights up to several hundred meters were detected using the seismic network (Avard et al., 2015). In early 2015, incandescence was observed around the dome, with fumaroles registering temperatures of 625◦C, and mobile Multi-GAS measurements yielding high SO2/CO<sup>2</sup> gas ratios from the acid lake. Renewed phreatic activity was observed at Poás between June and August 2016.

#### 2.1. Monitoring Network

Systematic monitoring began at Poás volcano in 1978, conducted by the Universidad Nacional, presently known as the OVSICORI-UNA (Martínez Cruz, 2008). Prior to the volcanic crisis in April 2017, the volcano was monitored by 2 seismic stations in close proximity to the volcano, CRPO [Nanometrics Taurus digitizer with Trillium Compact sensor (sensitivity of 0.008–100 Hz)] at 550 m east of the dome, and VPVF [Guralp CMG-DM24- EAM digitizer with CMG-3ESPC sensor (sensitivity of 0.003– 50 Hz)] 900 m north of the dome, and one seismic station (VPTE, Guralp CMG-DM24-EAM digitizer with CMG-3ESPC sensor) at 4.5 km from the volcano; 2 GPS (VPCR at 900 m north, and VPEV 1,350 m south-west of the dome); 1 stationary Multi-GAS station (300 m north of the dome); and a web cam (470 m north of the dome) (**Figure 1A**). The dome (which was subsequently destroyed during this eruptive episode) was located on the very southern edge of Laguna Caliente. Two additional seismic stations, VPLC and VPEM (both Quanterra Q330HRS digitizers with Nanometrics Trillium Compact sensors), were installed following the eruptions in April 2017. All seismic stations are broadband, 3-component sensors with 24/26 bit digitizers. Regular visits to the volcano by staff of OVSICORI-UNA allowed mobile DOAS measurements on foot along the western crater rim (remote sensing of SO<sup>2</sup> in the atmosphere based on UV absorption spectra); in-situ FLIR (thermal imaging); and sampling of lake water, fumaroles and eruptive products from within the principal crater containing Laguna Caliente.

#### 2.2. Renewed Activity: April 2017

From January to March 2017, the number of seismic events identified at Poás volcano began to increase from an average of 30 events a day in January, to over 100 events a day from 12 February onward, and over 200 events a day from 26 March 2017 (**Figure 3A**). Events were identified manually from the incoming seismic record by OVSICORI-UNA personnel. The number of events per day was not considered unusual until the end of March 2017. No surface manifestations were observed until 1 April 2017 when a new water-rich degassing fumarole appeared 100 m west of the active fumarolic field on the edge of Laguna Caliente, that was active for ∼24 h. On 7 April, a boiling (∼90◦C) and highly acidic spring (with a diameter of 4 m and the same pH and chemical composition as the lake, known locally as a "Borbollón") opened 200 m south of the dome. Due to high concentrations of detected SO2, the administrators decided to evacuate the Poás National Park on 9 April and kept it closed for the rest of the day. On 13 April, at approximately 02:00 UTC, a new boiling vent opened on the dome in the south-eastern corner of Laguna Caliente, generating a small lahar which affected the village of Bajos del Toro, located 7.5 km west of the active crater; ∼10% of the dome was destroyed; and ballistics up to 25 cm in length were reported on the western crater rim at distances of 400 m. The park closed on this day, and has remained closed up to the time of writing (September 2018). A phreatomagmatic eruption on 13 April at 21:45 UTC generated a plume of ash, water and magmatic gas to 500 m above the crater, destroying the Multi-GAS station and web cam in the process (stars, **Figure 1A**). On 14 April, another phreatomagmatic eruption occurred at 13:57 UTC, generating a plume that rose up to 4 km high and was visible from the capital city, San José (35 km away) as well as the closer provinces of Alajuela and Heredia, and destroyed a further ∼80% of the dome. Ballistics up to 50 cm in length impacted the Mirador (national park tourist viewing platform) located 650 m from the main vent, at an elevation of 260 m above the crater. From 13-19 April, borbollón activity from the dome area, with sporadic energetic phreatomagmatic eruptions reaching elevations up to 500 m (**Figure 1C**) were observed, with erupted material progressively filling the lake around the active vent. On 21 April, this eruptive material reached the lake surface. The most energetic phreatomagmatic eruption of this eruptive episode occurred on 23 April at 04:12 UTC, launching ballistics up to 1.5 km away from the active vent. Spatter bombs up to 3 m in length were found 450 m away on the south-east crater rim. Approximately 80% of the forest within 60 m of the crater was completely destroyed, as well as a seismic station (CRPO, **Figure 1A**). Small, discontinuous eruptions continued until the end of May 2017, when the lake completely disappeared. On 6 June, Poás began a phase of passive ash emissions (i.e., nonexplosive) and continuous magmatic degassing. Since the closest seismic station had been destroyed in the 23 April eruption, 2 new seismic stations (VPEM and VPLC, **Figure 1A**) were installed on the flanks of Poás in June 2017 to strengthen the monitoring network, in addition to two new web cameras and two new Multi-GAS monitoring stations.

#### 2.2.1. Seismicity

The number of low frequency earthquakes (LFs, ≤ 5 Hz, **Figures 2A,B**), tremor and volcano-tectonic earthquakes (VTs, high frequency, 5–15Hz, **Figures 2C,D**) started to increase in January 2017. However, the values at the time were not seen as unusual based on Poás' previous eruptive episodes (**Figures 3A,B**). In hindsight, these events probably represent the renewal of activity at Poás. On 26 March, a total of 22 high frequency VT events registered over 11 h, although none of the events were locatable within reasonable errors. This was the first warning recognized in real time of renewed activity at Poás. The Real-time Seismic Amplitude Measurement (RSAM, Endo and Murray (1991)), was calculated by taking the average amplitude of seismic events every 10 min within two discrete frequency bands (1–5 Hz and 5.1–15 Hz) to analyse amplitudes related to both LF and VT events, respectively. Some authors use the term SSAM to describe the use of discrete frequency bands (e.g., Cornelius and Voight, 1994; Rogers and Stephens, 1995). We chose the more general RSAM whilst explicitly stating the frequency bands of computation. No significant change in the filtered high frequency RSAM 5.1–15 Hz (**Figure 3F**) was observed, suggesting that although there was an increase in VT event count at the end of March 2017, there was not a change in the energy released. A significant increase in VT seismicity (a three-fold increase in event count) was registered from 13 April onward (**Figure 3B**), which coincided with the onset of daily phreatic activity at Poás and an increase in high frequency seismic energy release (**Figures 3B,F**, light blue shaded areas).

On 28 March, a significant number of LFs and short duration tremor was registered, with more than 400 events a day following a systematic increase in LF events over the previous few days (**Figure 3A**). Significantly, on the same day, the first Very Low Frequency (VLF) earthquake was recorded for this eruptive phase (**Figures 2E,F**). LF events once again dropped below 300 events a day on 10 April, although this was still considered to be elevated. From the 12 April onward, although the number of LFs did not appear to increase until 21 April (**Figure 3A**), the energy release significantly increased (**Figure 3E**, shaded blue area). The maximum number of LFs between February and April 2017 was recorded on 22 April, with ≥750 events on a single day (**Figure 3B**, red dashed line). Since the end of April 2017, seismicity at Poás has gradually reduced to almost no registered activity by April 2018.

#### 2.2.2. Deformation

From January to February 2017, GPS sites did not register any significant deformation. A clear inflation of the edifice began in mid March, which inflated by ∼2 cm by the end of March, and by a further 3–4 cm by 23 April 2017. Following the large phreatomagmatic eruption on 23 April at 04:12 UTC, a significant deflation of 2 cm was registered over a 48 h period, probably related to the expulsion of mass during the eruption, after which the crater began to slowly inflate once

more (**Figure 3C**). In addition, from February to April 2017, the distance between VPEV and VPCR (approximately 2,160 m) increased by 18 mm, after 2 years of relative stability. Similar to observations at other volcanoes (e.g., Soufrière Hills volcano, Montserrat; Voight et al., 1998; Kilauea, Hawaii; Tilling, 2008; Volcán Santiaguito, Guatemala Johnson et al., 2014), a clear correlation is observed between the deformation signals and earthquake activity (**Figures 3A,C**), in particular during February and early March 2017, where a sudden increase in both the count of low frequency events (**Figure 3A**) and high frequency events (**Figure 3B**) coincides with an increase in GPS signal (**Figure 3C**), associated with the inflation of the edifice.

#### 2.2.3. Gas Measurements

The permanent Multi-GAS station located near Laguna Caliente registered a change in degassing behavior in late 2016. Prior to the end of November 2016 the SO2/CO<sup>2</sup> ratio had been decreasing over a period of more than a year. In early December 2016 the SO2/CO<sup>2</sup> gradually started increasing, which is considered a signal of increasing likelihood of eruption at Poás (de Moor et al., 2016). Dramatic changes were noted in gas composition immediately before the eruptive phase in April 2017. On 29 March, the H2S/SO<sup>2</sup> ratio measured with a stationary Mutli-GAS station dropped from an average value of ∼2.4 in March 2017 to ≤0.01 on 31 March. On the night of the 31 March, the SO2/CO<sup>2</sup> ratio increased from 0.04 to 0.16, before the first visible manifestation at the surface (new water rich degassing fumarole on the edge of Laguna Caliente). The increase of the SO2/CO<sup>2</sup>

ratio continued to 0.44 on 1 April and from then on increased linearly until 12 April, where the average value was 7.4, until destruction of the station on 13 April. The SO<sup>2</sup> flux appeared very low at ∼20 t/d on 28 March (at the lower limit of the detection threshold); at 180 t/d on 4 April; at ∼440 t/d on 10 April; at 1,500 t/d on 13 April; and 2,200–3,000 t/d after 24 April (measured with a DOAS instrument, on a drone as it was considered too dangerous for personnel to enter the crater).

#### 2.2.4. Petrology

An energetic geyser-like eruption at Poás in 2016 produced ash with ∼6% juvenile material content (modal analysis on stereoscopic microscope, phi = 1–2). SEM analysis of this ash sample showed irregular angular to sub angular shards, as well as a few rounded shards. The 12 April 2017 eruption that destroyed 10% of the dome presented 9% juvenile material in the ash deposit. The juvenile component increased to ∼30% on 14 April; to 63% on 20 April; and to 85% on 22 April, clearly showing the increasing influence of a magmatic component in the evolving eruption. SEM analysis of material collected from the 23 April eruption exhibited a sharp, glassy surface texture. After the 23 April 2017 eruption, the juvenile content of erupted ash remained constant at ∼80%, until it increased further on 16 June to 96% (Cascante, 2017). Bombs from the April eruptions were highly vesicular and porphyritic in texture, with a matrix abundant in phenocrysts such as plagioclase (∼30%), pyroxenes and a small number of altered olivines, suggesting an intermediate

2017; the light blue shaded areas indicate the timing of phreatic and phreatomagmatic eruptions from 12 April onward; the red dashed line indicates the timing of the large phreatomagmatic eruption on 23 April at 04:12 UTC. (A) Daily counts of LF (low frequency) seismicity. The daily counts in February and early March were not considered unusual. (B) Daily counts of VT (high frequency) seismicity. Very few seismic events were identified before March 2017. The VT swarm on 26 March is clearly visible (green dashed line). (C) GPS height change in meters (vertical component, station VPEV). A clear inflation of 4–5 cm can be observed from early March to 23 April. The large deflation signal in January 2017 is considered to be an outlier, as measurements returned to what was considered background levels the next day. (D) SO2 flux, measured by walking a transect along the western crater rim. (E) 10 min filtered RSAM (1–5 Hz) recorded at station VPVF, 1 March to 30 April 2017. (F) 10 min filtered (5.1–15 Hz) RSAM recorded at station VPVF, 1 March to 30 April 2017.

composition of basaltic-andesite to andesite (Martínez et al., 2017).

#### 2.2.5. Lake and Boiling Springs Geochemistry

The temperature of Laguna Caliente showed a two-fold increase between late March and mid April 2017, increasing from 35◦C on 28 March to 41◦C on 4 April; 54◦C on 10 April; and 64◦C on 13 April. After this, it was not possible to visit the lake for direct sampling and measurements due to safety concerns. At the same time, the lake level increased by 1 m between 28 March and 11 April, and by another 0.5 m by 13 April, despite a lack of rain during this time, suggesting the increasing addition of material and/or fluid to the base of the lake from active fumeroles and vents. A boiling spring appeared on the crater floor 180 m to the south of the acid lake issuing acidic and salty hot waters (93◦C) with chemical characteristics (pH, salinity, anion concentrations) similar to that of the Laguna Caliente brines, but over saturated in cristobalite, a high temperature silica polymorph. The pH of the lake remained stable passing from 1.44 on 28 March to 1.39 on 13 April. The SO<sup>4</sup> <sup>2</sup>−/F<sup>−</sup> ratios in the lake waters decreased during the first 2 weeks of April 2017, indicating input of halogen-rich fluids into the lake, supported by the increasing lake levels over this time.

Dissolved unreacted SO<sup>2</sup> increased in the acid lake from 5 ppm on 28 March to 400 ppm on 13 April. After the April 2017 eruptions, the ultra acidic crater lake of Poás shrunk rapidly until it disappeared in June 2017, allowing the subaqueous fumaroles to discharge directly into the atmosphere. Several ponds of molten sulfur and sulfur cones were observed at the dried bottom of the crater after June 2017. Some ponds contained bright yellow molten sulfur, but one contained molten pyrite-rich black sulfur, indicating boiling temperatures between 113◦C and 116◦C, depending on the type of impurities present. A new lake started to form by mid January 2018, due to the gradual slowing of magmatic activity throughout the second half of 2017, and the high levels of precipitation between August 2017 and January 2018 as a result of several tropical storms and low pressure atmospheric systems affecting the country. The new acid lake had a temperature of 60◦C and a pH of 0.60 in January 2018. This lake dried out during March 2018, and was absent for approximately 1 month.

#### 3. SIMILAR SEISMICITY AT PÓAS VOLCANO

Seismicity has remained a primary monitoring tool for the detection of volcanic unrest because it can be remotely analyzed in real-time, often by an automated system and it is considered a highly valuable tool for decision-makers. The characterization of seismicity in a volcanic environment is traditionally based upon the signals' time and frequency characteristics: different bands of frequency relate to different active source processes at depth, which can be distinguished from one another, although the frequency bands associated with each process may overlap (Lahr et al., 1994). Low frequency seismicity (LFs, **Figures 2A,B**) are characterized by frequencies between 0.5 and 5 Hz; show emergent P- and S- wave arrivals (Chouet and Matoza, 2013); and have been linked to resonance of seismic energy trapped at a solid-fluid interface within a crack (e.g., Chouet, 1988) or a volcanic conduit (e.g., Neuberg et al., 2000). The trigger mechanism of such seismic energy may be generated by a stickslip motion along conduit walls as magma ascends (e.g., Iverson et al., 2006; Kendrick et al., 2014); the brittle failure of the rising magma itself due to an increase in strain and viscosity rates (e.g. Lavallée et al., 2008; Thomas and Neuberg, 2012); interactions between the magmatic and hydrothermal system (e.g. Nakano and Kumagai, 2005); or through the slow rupture of unconsolidated material on volcanic slopes (Bean et al., 2014). High frequency events (VTs, **Figures 2C,D**), which are characterized by frequencies >5 Hz, have clear, impulsive P- and S- arrivals; and are generated when magmatic processes create enough elastic strain to force the surrounding edifice into brittle failure (Arciniega-Ceballos et al., 2003). Many volcanic seismic events fall between these two end-member categories and are classified as "hybrid" events. Very low frequency events (VLFs) typically occupy the frequency range below 1 Hz (**Figures 2E,F**). VLFs are typically attributed to the coalescence or ascent of gas slugs within a volcanic conduit during migration (e.g., Ripepe et al., 2001; Chouet et al., 2008), or to magmatic gas release (Jolly et al., 2010), although trigger mechanisms are still widely debated. Site and path effects, as well as the type of sensor deployed, significantly influence the waveform shape and its frequency recorded at a seismic station, meaning that the same event may be classified differently at two different stations. In addition, it is possible that differences in the location of the seismic event may influence the frequency content recorded. Consequently, it is essential to take these effects into full consideration when trying to classify volcano seismicity.

The further classification of seismic events into "families," which all have a similar waveform shape as well as the same frequency content, allows the depiction of temporal and spatial changes in the source mechanism and the source location on a much smaller scale (e.g., Thelen et al., 2011; Salvage and Neuberg, 2016). By definition, families of seismic events should be generated by the same source mechanism and at the same source location (estimated at between one quarter and one tenth of the wavelength; Geller and Mueller, 1980; Neuberg et al., 2006) in order for the detected waveforms to have the same recorded shape at the seismometer (Minakami et al., 1951), as long as site and path effects on the seismic wave are minimal. Families of similar seismicity were first identified at Usu volcano, Japan, during a dome building eruption in 1944 by Minakami et al. (1951) and during the 1955 eruption of Bezymianny, Kamchatka (Gorshkov, 1959), and have since been identified at a number of active volcanoes around the world, including Redoubt, Alaska (e.g., Buurman et al., 2013); Mt. St. Helens, USA (Thelen et al., 2011); Colima, Mexico (Arámbula-Mendoza et al., 2011); Merapi, Indonesia (Budi-Santoso and Lesage, 2016); and Soufrière Hills, Montserrat (e.g., Rowe et al., 2004; Salvage and Neuberg, 2016).

Waveform similarity in terms of shape and duration can be evaluated by a cross correlation procedure where identical signals will result in a maximum cross correlation coefficient of 1 or –1, dependent upon their relative polarity and signals with no correlation resulting in a cross correlation coefficient of 0. The choice of similarity threshold (above which events are considered similar) is important: if it is too low there is a risk of placing events that are not similar into the same family; if it is too high similar events can be missed (Salvage et al., 2017). We define a "family" of events as events which show similar waveform shape characteristics, defined by having a cross correlation coefficient >0.7. This is in agreement with Green and Neuberg (2006), Thelen et al. (2011), and Salvage and Neuberg (2016), who suggest a cross correlation coefficient threshold of 0.7 in andesitic volcanic environments, since this is significantly above the correlation coefficient that can be produced from random correlations between noise and a waveform (Salvage et al., 2017). Here we first identify families of seismicity using a simple amplitude ratio algorithm, in addition to high waveform similarity on a multiple station network

(≥ 0.7 cross correlation coefficient) using REDPy (Repeating Earthquake Detector, Hotovec-Ellis and Jeffries, 2016). Secondly, we take the core event of each family identified in the previous stage (an average of the stack of all events in that family aligned to the point of maximum cross correlation), in addition to all events identified during the study period that were identified using an amplitude ratio detection algorithm, and use these as a template to find more events within the continuous waveform data that may have previously been hidden by noise or simply not detected using a trigger algorithm [EQcorrscan, Calum Chamberlain, Victoria University of Wellington (Chamberlain et al., 2017)]. This template matching technique identifies a repeating event when the absolute cross correlation coefficient sum for a given template exceeds a threshold of 0.7 on each station, across a minimum of three stations. If repeating events occur within 0.5 s of one another, the strongest detection within this time period is taken, i.e., the event which exceeds the threshold the most. Both of these algorithms are written in Python and are open source, available through GitHub.

REDPy uses a simple amplitude ratio algorithm to detect seismic events from the continuous record on a multiple station network and then determines whether any events are similar by identifying all events over a given cross correlation threshold (0.7 in our case). This method, however, fails to recognize events hidden by noise, or events that are too closely spaced to retrigger the algorithm (events within ∼7.5 s of one another in this case). The core event is defined using OPTICS (Ankerst et al., 1999), a sorting-based clustering algorithm. The core event corresponds to the event with highest "reachability" within each family, meaning it correlates highly with many other members of that family. We took the core event of each family identified by REDPy (as well as all events which triggered the detection algorithm but appeared to have no similar events during the study period) and used these as input for EQcorrscan, which relies upon template matching rather than a trigger algorithm. Each core event was 10 s long in order to ensure that the entire waveform and coda were included and filtered between 0.5 and 15 Hz. In order for an event to be identified as similar as one of the core events, we again required a cross correlation coefficient >0.7. By this combined methodology, we found that the number of events identified in the dominant family (family containing the greatest number of events) went from 268 (using REDPy) to 771 events (using REDPy and EQcorrscan combined). EQcorrscan can be computationally intensive, however significantly more events were identified using this method (**Table 1**). The high discrepancy in number of events detected by these two different methodologies is a consequence of their detection algorithms. REDPy uses an amplitude ratio algorithm that only detects events above a given (user designed) threshold so events occurring within sections of high noise for example are not likely to be detected, and it can only detect one event per window length (in this case ∼7.5 s). EQcorrscan is a template matching algorithm which therefore can detect event during noisy periods as it is simply focused on finding matching waveform patterns. In addition, the minimum inter-event time (IET) is 0.5 s, meaning it can in theory TABLE 1 | Details of fourteen families of similar seismicity identified in this study.


Three families contained low frequency events (0.5–5 Hz); 11 contained high frequency events (5–15 Hz), as identified by the Frequency Index (FI). Using EQcorrscan in combination with REDPy significantly increases the event counts for each family.

detect 7 times more events than REDPy within the same time window.

One family of similar seismicity that contained >10 events was identified in hindsight at Poás using REDPy (containing a total of 278 events). Furthermore, 84 events were detected using the amplitude ratio detection algorithm, but were not classified into families as no similarity was detected between these events. Many more events (a total of 771 within this family) were identified from the continuous record using the template matching technique of EQcorrscan (**Table 1**). Fourteen families, each containing >10 events per family, were identified using the combined methodology of REDPy and EQCorrScan (**Figure 4**). We cross correlated the core event for each family (**Figure 4A**) with every other core event and found that none of the families were similar to one another (the maximum cross correlation coefficient determined was 0.49 between two families of waveforms), suggesting that no single event is likely to belong to more than one family. LF families contained many more events and were active for longer periods of time (**Figure 4B**, green) than VT families, although all families showed distinct temporal patterns, in particular in relation to the rate of events with time (**Figures 4E,F**). Significantly, LF families began much earlier in the sequence (around the end of March) compared to the VT families that were active at the time of the large phreatomagmatic event on 23 April. The dominant family of similar waveforms showed two phases of heightened activity (**Figure 4C)**: (1) an increase in daily events from 28 March to 2 April 2017, followed by a steady decline; and (2) a smaller, apparently less significant increase in event rate in the days prior to the large phreatomagmatic eruption on 23 April. A similar pattern of two phases of activity was observed for all other low frequency families identified. VT families showed a single period of heightened activity, starting in early

FIGURE 4 | (A) Core events (normalized) of each family identified by REDPy. VT events are shown in gray, LF events in green. (B) Duration of families, colored with respect to (A). (C) Daily count of dominant LF family (greatest number of events) corresponding to the first waveform in (A) identified by template matching technique (EQcorrscan). (D) Daily counts of dominant VT family (greatest number of events) identified by template matching technique (EQcorrscan). (E) Cumulative counts of earthquakes within each LF family. (F) Cumulative counts of earthquakes within VT families. The dashed black line indicates the timing of increased VT activity counted by OVSICORI-UNA personnel (26 March), the dashed blue line indicates the beginning of daily phreatic eruptions (12 April), and the dashed red line indicates the timing of the large phreatomagmatic eruption (23 April).

April (**Figure 4D**). Following the beginning of daily phreatic eruptions at Poás (12 April onwards), the dominant LF family registered a significant increase in the RMS (Root Mean Square) amplitude of events (e.g., **Figure 5A**, following the blue dashed line), which decreased again on 25 April, 2 days after the large phreatomagmatic eruption. Significantly, the dominant LF family

also showed a clear decrease in FI in the hours prior to the large phreatomagmatic eruption on 23 April (**Figure 5A**). The Frequency Index (FI), developed by Buurman and West (2010), is a proxy for the spectral content of the waveform, and is based upon the ratio of energy in low and high frequency windows, with a base-ten logarithm in order to reduce the index to a single value. A negative FI means the waveform is dominated by low frequency energy; a positive FI demonstrates a majority of energy in the high frequency band; and a FI of zero suggests that the waveform has equal amounts of high and low frequency energy. Here, we define a low frequency window between 0.5 and 5 Hz, and a high frequency window between 5.1 and 15 Hz. A decrease in FI would therefore suggest that the events began to contain a higher proportion of low frequency energy, potentially related to the increase ability for fluid to move through the volcanic system with ease, as the fracture network (responsible for the generation of higher frequency events) is thoroughly developed. This decrease in FI occurs coincidently with an increase in RMS amplitude, suggesting that the lower frequency events are larger in amplitude. Within the dominant LF family, the IET appears to evolve slowly with time, as events become more sporadic (**Figure 5A**). The minimum IET detected within the dominant LF family was 1.18 s, although ≥ 95% of the IETs detected were greater than the event duration of 10 s. Furthermore, in the 12 h prior to the large phreatomagmatic event on 23 April, no events from the dominant LF family were identified (**Figure 5A**). Since the dominant VT family continues to register a small number of events during this period (**Figure 5B**), this could be interpreted as either a sudden shut off in the conditions necessary to generate these types of events, or the possibility of aseismic magma movement (e.g., Neuberg et al., 2006). VT families appear to show no significant patterns in RMS amplitudes of events, IETs or changes in FI over the investigated study period (**Figure 5B**).

#### 4. HINDSIGHT FORECASTING OF 22 APRIL 2017 ERUPTION

The ability to forecast the timing, intensity and type of volcanic activity is one of the key issues facing volcanologists today. Since the time series analysis of families at Poás in April 2017 identified an accelerating trend within the dominant LF family, the Failure Forecast Method (FFM) may be applicable for identifying a forecasted timing of eruption. The FFM is based on an empirical power-law relationship relating the acceleration of a precursor (d <sup>2</sup>/dt<sup>2</sup> ) to the rate of that precursor (d/dt) (Voight, 1988) by:

$$\frac{d^2\Omega}{dt^2} = K\left(\frac{d\Omega}{dt}\right)^a\tag{1}$$

where K and α are empirical constants. can represent a number of different geophysical precursors, for example low frequency seismic event rate (Salvage and Neuberg, 2016), event rate of all recorded seismicity (Kilburn and Voight, 1998), or the amplitude of seismic events (Ortiz et al., 2003). The parameter α can range between 1 and 2 in volcanic environments (Voight, 1988), or may even evolve from 1 toward 2 as seismicity proceeds (Kilburn, 2003). An infinite d/dt suggests an uncontrolled rate of change and here is associated with an impending eruption. The inverse form of d/dt is linear if α=2, and therefore in this case the solution for the timing of failure is a linear regression of inverse rate against time, with the timing of failure relating to the point where the linear regression intersects the x-axis in graphical form (Voight, 1988). As a deterministic approach for forecasting the timing of volcanic eruptions, the FFM relies upon several assumptions, including that the acceleration can be described by a simple power law and that the time of the eruption is related to the time at which this power law reaches a singularity (Boué et al., 2016). The FFM has proved useful at accurately forecasting volcanic eruptions in both near real time (e.g., Cornelius and Voight, 1994) and in hindsight evaluation (e.g., Budi-Santoso et al., 2013; Salvage and Neuberg, 2016). However, Boué et al. (2016) have suggested from a study of data over a 13 year period at Volcán Colima, Mexico, and a 10 year period at Piton de la Fournaise, Reunion, that not all cases of accelerating seismicity are suitable for the analysis by the FFM, in particular if the acceleration does not follow a single power law increase. In fact, only 36% of eruptions at these volcanoes could be forecasted in real-time, although hindsight forecasting fared better, with ∼ 50% of eruptions being successfully forecast when utilizing the entire precursory seismic sequence in the forecasting model.

The FFM was first developed as a tool for forecasting the timing of slope failure using accelerating material creep; a cause and consequence of one single active system generating failure (Fukuzono, 1985). However, a volcanic system is inherently more complex, and accelerating magma ascent could be detected at several positions in the magma plumbing system with different phase delays and amplitudes. It may be advantageous to use the FFM in collaboration with a single family of seismicity since a single family originates from the same source location and is produced by the same source mechanism (e.g., Geller and Mueller, 1980; Petersen, 2007; Thelen et al., 2011), which may produce a more accurate forecast, potentially a consequence of focussing upon a single active system at depth without interference from other sources of error (Salvage and Neuberg, 2016). In other cases, using only a single family in conjunction with the FFM produces a less useful forecast than when using all the identified seismicity (e.g., Tungurahua, Ecuador Bell et al., 2017).

Here, we set α=2, to allow for simple implementation, and because it has proved to be a good choice for the value of α at a number of volcanoes (e.g., Cornelius and Voight, 1994; Chardot et al., 2015; Salvage and Neuberg, 2016). We only consider data from 18 April onward in this analysis and consider the acceleration over a 96 h period from 18 April until 22 April. The event count is dependent upon events not occurring within 0.5 s of one another (a user defined minimum threshold in EQcorrscan), however, in some volcanic environments individual events, in particular low frequency seismicity, may merge into tremor as their IET decreases (Neuberg et al., 2000). Consequently, the event count recorded in our case will not include tremor episodes if individual events cannot be constrained, despite this signal being associated with precursory activity at some volcanoes e.g., White Island, New Zealand (Chardot et al., 2015). We use the R 2 value of the leastsquares linear regression as an initial indication of the confidence of the forecast made. The closer the R 2 value is to one, the more confidence that can be placed in the forecast, although in this case we use it simply as an indicator of the fit of the regression before performing more detailed residual analysis. R <sup>2</sup> <0.65 are considered to represent a poor relationship between the observed data and the fitted FFM model (Barrett, 1974). Consequently, we define a successful forecast as one where the timing of the eruption is within 3 h of the known timing of the eruption, with an R <sup>2</sup> > 0.65.

Accelerations in the number of seismic events per hour prior to the large phreatomagmatic eruption on 23 April 2017 at 04:12 UTC (**Figure 6**) were only identified when using all families combined, and in the dominant LF family. No clear accelerations could be identified in VT families. We use the number of events per hour as an indicator of the activity, rather than events per day, as we consider this to be more useful if the process is to move into real time: the number of events occurring within a time period of <1 h would be difficult to process and understand quickly enough for decision makers; and the number of events per day may be too long between processing times to generate usable forecasts. The acceleration identified when using all families combined (**Figure 6A**) is subtle, and in fact a clear acceleration followed by a deceleration can be seen from 18 to 19 April, which contains higher event counts. Application of the FFM to the entire accelerating sequence produces a very poor forecast, with R 2 values much lower than what is deemed confident (R <sup>2</sup>≪0.65, **Figure 6B**), despite the forecasted time of eruption occurring within 5 h of the known timing of the eruption. **Figure 6C** shows the acceleration of seismic events per hour for the dominant LF family (30-03-2017 08:46). An acceleration can be identified from the 19 April onward (**Figure 6C**). The acceleration of this family appears to stop approximately 24 h before the large phreatomagmatic eruption, producing a period of quiescence where no repeating events are identified. Application of the FFM (**Figure 6D**) indicates a poor forecast, with an R 2 value of 0.13. The least-squares linear regression generates a forecast for 22 April at 06:48, approximately 22 h before the known timing of the eruption.

Both generated forecasts were extremely poor, with R <sup>2</sup> ≪0.65 and only one forecast was made within 5 h of the known eruption time, even though an observable acceleration in seismic event rate could be identified. Following Chardot et al. (2015), residuals were calculated through time in order to test the assumptions of using a least-squares linear regression with this data, and plotted as a histogram, where the number of bins is equal to 2n 1/3 (n is the number of samples), as according to the Rice rule, to verify the normality of the residual distribution (**Figure 7**). Residuals were defined as the difference between the observed hourly event count for each family, and the best fit model to this data. Using the residual analysis, the least-squares linear

regression model can be deemed appropriate if: (1) the residuals do not follow a trend; (2) the residuals do not increase or decrease as a function of time; and (3) the residual distribution follows a Gaussian distribution. Our residual analysis (**Figure 7**) suggests that the error structure of the data is inconsistent with a least-squares linear regression model when α is equal to 2, since the distribution is not Gaussian. This has been suggested previously by Bell et al. (2011) for seismic event rates. In addition, all of our residuals appeared to follow a trend, which suggests that a more complex model is needed to define the data, and residuals increase as a function of time, indicating they do not exhibit equal variance. Consequently, we can conclude that the least-squares linear regression applied in this instance is not the most appropriate model for describing the accelerating behavior observed.

### 5. DISCUSSION

#### 5.1. Characteristics of Families Identified

The identification of families of similar seismicity at Poás (events with a cross correlation coefficient >0.7) suggests: (1) a stable source process (non-destructive); (2) that the trigger mechanism of such events must be able to be recharged quickly (in this case in <2 s since this was the minimum IET observed); and (3) that these conditions must occur consistently at the same location (e.g., Green and Neuberg, 2006; Petersen, 2007). In particular, the identification of a number of families that are simultaneously active suggests that a number of distinguishable sources must be active at the same time beneath Poás. Furthermore, the identification of both LF and VT families acting simultaneously suggests a diversity in the ongoing physical processes at depth, once path effects have been accounted for in the classification. As discussed earlier, the source mechanisms for LF and VT seismicity are often disputed, but here we suggest that both VT and LF occurrence is related to the movement of magmatic fluid and gases toward the surface, with VT families suggesting the generation of fractures and the potential opening of new magmatic pathways due to increased stresses at depth as a result of the presence of magmatic fluid (e.g., Lahr et al., 1994; Kilburn, 2003), and LF families potentially suggesting the movement of this fluid through these fractures as a result of an increased strain rate in the magma (e.g., Chouet, 1988; Neuberg et al., 2000; Thomas and Neuberg, 2012). Since the VT families contained fewer events than families containing low frequency events (**Table 1**), we suggest that the dominant ongoing process during this time (once a new fracture pathway had been opened, demonstrated by the swarm of the VTs in March 2017) was magmatic fluid movement, most likely, a mixture of magma and gases. This is supported by the significant increase in RMS amplitude of LF events (in particular after 12 April, **Figure 5A**) in comparison to VT events (**Figure 5B**), suggesting that the movement of fluid dominated the processes occurring at depth

during April 2017. We note that this method does not allow for the characterization of volcanic tremor, but instead will either count individual events within a tremor episode if the IET is >0.5 s instead of a single tremor sequence, or miss the tremor episode entirely, affecting the final event count. Although there are instances where volcanic tremor has appeared as a precursor to eruptive episodes at some volcanoes (e.g., Chardot et al., 2015), here we are interested in families of repeating seismicity as a precursor, and not other signals. Further analysis is required to determine whether volcanic tremor played a significant role as a precursor to this eruptive event.

All identified families of low frequency events (including the dominant family) appear to return to fewer seismic counts immediately prior to the eruption after an acceleration (**Figure 6**). This could represent a decrease in magmatic flow rates due to a physical obstruction, the generation of a damage zone that is responsible for impeding fluid flow, or the intermittent advance of magmatic fluid through a fracture network. A sudden drop in seismic event rate could also represent a change in the system to more aseismic magma movement, as fluid pathways become fully open, meaning less seismic events are generated. In our case, since the seismicity in this instance can be classified as low frequency, either the physical conditions necessary for the generation of this family are gradually changing (e.g., Stephens and Chouet, 2001); a change in the fluid flow rate leading to the generation of systematically fewer seismic events is occurring; or the movement of magmatic fluid becomes aseismic. Hotovec et al. (2013) also noted a period of quiescence prior to explosive events at Redoubt volcano, Alaska during an eruptive period in 2009, although on the timescale of seconds, rather than hours. Rodgers et al. (2015) noted a systematic decrease in LF seismicity at the same time as an increase in VT seismicity, over a number of months prior to eruptive activity at Telica volcano, Nicaragua from 2010 to 2011. They attribute this changing seismicity to an increase in pressurization (indicated by the VT seismicity) and thus a sealing of an active hydrothermal system (indicated by the drop in LF seismicity), suggesting it may drive phreatic eruptions. At Poás this scenario appears unlikely since we see no increase in VT seismicity with a decrease in LF events, and the reappearance of LF seismicity at the time of the eruption suggests that this 12 h of quiescence does not mark the timing of a sealing of the hydrothermal system, as it is unlikely to occur on this timescale. A significant decrease in the cross correlation coefficient immediate prior to the period of quiescence suggests a small, but significant change in the conditions needed to generate the dominant LF events, which become less similar with time (**Figure 5A**). Petrological evidence suggests that changes in fluid flow rate in response to obstructions in the conduit may be a more plausible explanation. We interpret the increasing juvenile content of the ash sample over time as being related to a fresh batch of rising magma "cleaning" out a path to the surface. Initially the path is obstructed by older, altered material (such as a plug), which must first be ejected before fresh magmatic material can reach the surface. Consequently, earlier in the

eruptive period, the ash content is dominated by interactions of the hydrothermal system with the rising magma, and the opening of a pathway to the surface. Seismic events within the dominant family are no longer recorded in the final hours prior to the eruption since a clear pathway to the surface has been generated, and therefore fluid movement becomes aseismic.

#### 5.2. Forecasting Potential

At Soufrière Hills volcano, Montserrat, the use of similar seismicity in collaboration with the FFM appears to allow the generation of a more accurate forecast, since isolating a single system at depth avoids additional uncertainties introduced by averaging data over a number of different accelerating phenomena, and therefore reduces the misfit between the data and the forecast (Salvage and Neuberg, 2016). At Poás, the identification of similar seismicity and its use as a forecasting tool is promising, since our hindsight analysis highlights that identifying repeating seismicity allows a much more detailed interpretation of ongoing processes at depth, in addition to the identification of far greater numbers of events for analysis. Using the methodology described in this paper, it would have been possible to detect accelerations in families of similar seismicity in near real time, although the analysis of this data is unlikely to have produced confident forecasts. Firstly, R 2 values for the least-squares linear regressions identified were always considered to be unacceptable for a confident forecast (R <sup>2</sup>≪ 0.65). Secondly, a number of accelerations in these families of seismicity could have been identified earlier in April: some prior to eruptive events (e.g., 12 April), and others occurring with no associated surface manifestations. Thirdly, the identification of decelerating trends, or a sudden drop in the seismic event counts may have falsely suggested a decline in activity at Poás, and therefore reduced the likelihood of a confident forecast. Lastly, accelerations were not identified in all families. For example, no clear accelerations could be identified within the families of VT seismicity. These factors become particularly important when trying to forecast eruptive events in real-time, and therefore more research is required to determine how some or all of these issues can be accounted for in a forecasting model.

Previously, Bell et al. (2011) suggested that it is inappropriate to use the FFM with a least-squares linear regression, since this does not account for the correct error structure of earthquake count data. Our residual analysis for the least-squares linear regressions applied when α is set to 2 supports this, since the residual error structures are not Gaussian (**Figure 7**). They therefore suggest that using a Generalized Linear Model (GLM) and a Poissonian error distribution where α is equal to 1 may be more appropriate. However, using the FFM in conjunction with the GLM also violates certain assumptions which are associated with the volcanic environment. The GLM suggests that the system being modeled is memoryless and that past events do not affect the future. We agree with Hammer and Ohrnberger (2012) and suggest that this is not an appropriate assumption for seismicity occurring in volcanic settings. Calculation of α (e.g Boué et al., 2016) is an essential next step in the characterization of the accelerations observed at Poás. An alternative model which may be more appropriate is a maximum-likelihood methodology which utilizes observed data to determine a model result with the greatest probability to maximize the likelihood function (e.g., Bell et al., 2013). This methodology also does not require the binning of data into time windows, which is advantageous. Although there are many issues with using the FFM in conjunction with a least-squares linear regression analysis (setting α to 2), in particular in real time scenarios, other regression models violate different assumptions of the FFM, indicating the inherent complexity of this problem. Our research indicates that a leastsquares linear regression is not an appropriate tool to use for forecasting this eruptive event at Poás, consequently eliminating one of the methods that is currently popularly used. We consider these "negative" results a contribution to our understanding of the scenarios that can be forecast using this methodology and hope to expand further on this analysis by implementing different regression models in the future to this data set.

#### 5.3. Conceptual Model: Phreatomagmatic Explosive Eruptions at Poás

Investigations into the volcanic structure of Poás in the 1980s suggest that a high density cylindrical plug, approximately 1,000 m in radius, sits between 500 and 800 m beneath the crater floor (Rymer and Brown, 1986; Casertano et al., 1987), potentially connected to the surface by a (now solidified) vertical intrusion beneath the dome in the active crater (Fournier et al., 2004). Fischer et al. (2015) suggested that phreatic eruptions at Poás are therefore caused by a gas pressure build up beneath this sealed plug (which provides the surface crater lake with heat and volatiles and is likely a chilled margin resulting from the crystallization of an older magmatic body), which is catastrophically released through hydrofracturing. De Moor et al. (2016) showed that phreatic eruptions at Poás were accompanied by short-term increases in SO2/CO<sup>2</sup> and higher SO<sup>2</sup> fluxes, again suggesting that high temperature magmatic gas injection drives phreatic eruptions. A similar mechanism for the onset of phreatic eruptions has been suggested for Mt. Ruapehu, New Zealand (Christenson et al., 2010), where sulfur within the system creates an impermeable plug within the volcanic conduit and consequently leads to the accumulation of gases beneath it, the elevation of pore pressures, and the sudden catastrophic release when critical pressures are reached.

The earliest identified indication of renewed activity at Poás in 2017 was a swarm of VT (high frequency) seismicity on 26 March (**Figure 3B**). High frequency seismicity is associated with the generation of new fractures when critical stresses are reached, that potentially allows fluid movement (e.g., Kilburn, 2003). This VT swarm is likely to represent the brittle fracturing of a previously sealed chilled margin or hydrothermal seal, either as a new batch of magmatic fluid beneath it begins to push upwards, or as a result of partitioning volatiles into residual melt during the crystallization process of the upper portion of the cooling magma body. It is unclear, however, as to whether the magma itself is forcing its way up toward the surface due to volatile decompression, or whether the fracture (generated as a result of increased pressurization from the nearby magmatic fluid and gases) creates a vaccuum as gas escapes out its top, allowing the magma beneath to be "sucked" upwards, as the magma below decompresses and migrates toward the surface. Increased fracturing of the seal may also result from the increased local strain rate around the intruding magma body, forcing the surrounding rock into failure (e.g., Fournier, 1999). In hindsight, the counts of LF events in February and early March 2017 (**Figure 3A**) may be an early indication of movement of magmatic fluid to beneath the seal allowing the build up of pressure, although this was not noted at the time as the event rate was not considered unusual.

In July 1980, an increase in VT seismicity was first observed before an increase in low frequency events at Poás. Although no large explosive eruption occurred associated with this seismicity, a significant increase in gas emissions was noted, and was believed to be caused by crystallization-induced degassing beneath a water saturated chilled margin caused by an intrusive episode (Casertano et al., 1987; Rowe et al., 1992; Rymer et al., 2000). In March 2017, the swarm of VTs was short lived and contained relatively few events (a total of 22 earthquakes) unlike the swarm in July 1980, that contained hundreds of events (Casertano et al., 1987). This may be evidence of the "Kaiser effect," in which fracturing and seismicity cannot be generated unless the previous maximum stresses of the system are exceeded (Fredrich and Wong, 1986; Smith et al., 2009). Consequently, it would be expected that after the resealing of the chilled margin as volcanic activity wanes, greater stresses are needed in order to refracture that body in a new eruptive episode, even decades later. Tuffen et al. (2003) has suggested that the rehealing of small veins through which fluid and gases can travel in an andesitic volcanic environment may be able to occur on timescales of minutes to hours, meaning that Poás would have had the opportunity to reseal itself since the last activity in 1980. Other explanations could include a more ductile environment at depth in 2017, and/or changes to the magnitude distribution of events between the two time periods.

An increase in pressurization beneath a seal at depth is supported by GPS measurements from early March 2017, which registered inflation of the edifice (**Figure 3C**), suggesting that the VT swarm in late March represents a time when pressurization at depth (induced by a seal) reached a critical level and allowed fracturing to occur. Hypocentre locations may help to define the depth of this seal, but this is beyond the scope of this paper. Prior to the 23 April 2017 eruption, only 2 GPS stations were recording, leading to large uncertainties in the location of a pressurization source at depth. Simply determining the intersection of deformation vectors suggests the pressurization source lies between 1 km and 6 km beneath the volcano. Casertano et al. (1987) and Rymer and Brown (1986) have previously suggested that a dense plug, which may act as a seal, sits between 500 and 800 m beneath the crater surface, in agreement with our observations for a pressurization source. The accelerated inflation of the crater in April 2017 suggests magma reaching shallower levels within the crust as it moves through a (now generated) fracture network toward the surface, likely behind the seal through which fluid migration is still impeded, but not impossible. The movement of magma to shallow levels is also supported by the increased SO<sup>2</sup> flux at the surface, increases in the SO2/CO<sup>2</sup> ratio, increases in lake temperatures, as well as a significant increase in low frequency seismicity from 22 March onward (**Figure 3**). In particular, the significant change in the gas composition from 29 March 2017 onward suggests the injection of gas to shallower levels, probably facilitated by the developing fracture network on 26 March.

The identification of a number of families of seismicity indicates that the movement of fluid (LF families) and the generation of new fractures (VT families) occurred consistently in the same location and by the same source mechanisms, and occurred simultaneously during the precursory period. Therefore, it is likely that developing fracture networks occurred in a limited number of locations, and since low frequency events occurred more frequently and with greater amplitudes and energy release, that the movement of magmatic fluid was the principal ongoing process at this time. Similarly, from 1978 to 1990, counts of LF events were considerably higher than those for VT events, during a variety of phreatic and protoplasmatic activity at Poás(Martínez Cruz, 2008). Increased pressurization below the partially sealed chilled margin due to the build up of magmatic fluid and the degassing of this magmatic fluid in shallower reservoirs is likely to lead to the catastrophic failure of the seal, and therefore to phreatomagmatic eruption. As new magmatic fluid moved toward the surface through the newly fractured seal, it may have picked up some of the surrounding altered conduit material and breccia from the generation of fractures, meaning the first erupted material in early April contained lower percentages of juvenile material than later eruptions, as the conduit was not yet fully open. As the magmatic fluid reached very shallow levels, it came into contact with the active hydrothermal system at Poás including the crater lake at the surface, which is likely to further enhance explosive activity.

We therefore suggest that the large phreatomagmatic eruption at Poás on 23 April 2017 was the result of a fresh batch of magmatic fluid becoming initially stalled behind a sealed chilled margin or hydrothermal seal at approximately 1 km depth. Once pressures were critical, some fracturing of this seal occurred (VT swarm in March) allowing the movement of magmatic fluid, and in particular volatiles, to shallower levels (noted by the increase in low frequency seismicity, gas fluxes, inflation of the crater area and significant changes in the gas composition). As the magmatic fluid migrated through the system to shallower levels, rapid degassing occurred, and it picked up surrounding edifice material and material form the seal. When it came into contact with the active hydrothermal system and crater lake, an even more energetic explosive eruption was generated.

#### 6. CONCLUSIONS

Phreatic and phreatomagmatic eruptions at volcanoes are poorly understood, difficult to forecast, and often pose a serious threat to local populations and tourists visiting volcanic areas. A large phreatomagmatic eruption occurred at Poás volcano, Costa Rica, on 23 April 2017 at 04:12 UTC, following approximately one month of unrest (rapid inflation, increased seismicity, increased gas fluxes and smaller phreatic eruptions), potentially as a result of pressure build up beneath a partially sealed, chilled margin or hydrothermal seal at depth. The sudden fracturing of this seal as critical pressures were reached triggered an explosive eruption, which was further enhanced by interaction of the magmatic fluid with the active hydrothermal system at Poás, and the crater lake at the surface.

Fourteen families of similar seismicity were identified in the days prior to this eruption, which included both LF and VT families. We suggest that the VT families are indicative of further fracturing of the volcanic edifice and the seal, creating new fluid pathways, and that the LF seismicity reflects the movement of magmatic fluid through these pathways toward the surface. Detailed analysis of these families suggests that only the dominant LF family produced an accelerating trend in the hours prior to the catastrophic eruption on 23 April, which we have used in collaboration with the FFM to forecast the timing of the event. When using all families of seismicity to forecast the timing of the eruption on 23 April, an accelerating trend was also identifiable, although it did not produce an accurate forecast. The dominant LF family of events showed an accelerating trend which produced a forecast approximately 22 h from the known timing of the eruptive event. Analysis of each LF family individually suggested a cessation of events in the 12 h prior to the large phreatomagmatic eruption, which was most pronounced within the dominant family of LF events. This deceleration could have been misinterpreted in real time to signify a slowing of the activity at Poás, but we suggest that at this time fluid movement through the system became aseismic. No confident forecasts were generated, despite an obvious acceleration in seismic event rate. The lack of acceptable forecasts may result from the use of a least-squares linear regression with the FFM, which based on residual analysis of this data is not an appropriate model to use. Our analysis allows us to eliminate some common parameters which are commonly used for forecasting volcanic eruptions, to search for other more complex models to explain the accelerating seismicity in this scenario. Furthermore, the use of similar seismicity rather than simply defining seismic events according to their frequency content alone allows a more detailed analysis of time series trends to be carried out. For example, this methodology allowed us to identify far greater numbers of events for analysis, and identified that a number of distinct specific sources were generating seismicity at depth, as well as significant changes in the frequency content of waveforms with time. Further investigation is required to determine whether all large phreatomagmatic eruptions at Poás are preceded by accelerating families of seismicity and consequently whether this can be successfully used as a forecasting tool for future events in real time, if more suitable regression analyses can be determined for use at this volcano.

#### DATA AVAILABILITY

All monitoring data for the period of interest, and all data generated in analysis for this manuscript is available upon request, without undue reservation, to any qualified researcher. Please contact Rebecca O. Salvage with requests.

#### AUTHOR CONTRIBUTIONS

RS identified the seismic families, computed the family analysis and accelerations, calculated the forecasts using the FFM and is responsible for writing the manuscript. GA provided key information regarding the chronology of the eruption in April, as well as petrological analysis of ash samples and lake water samples. JdM conducted gas analysis, JP and JB were responsible for monitoring of seismicity and seismic counts. MC conducted SEM and microprobe analysis of ash and other collected samples. CM conducted deformation (GPS) analysis. MMC was responsible for the lake and boiling springs geochemistry analysis and SEM-EDS and ICP-MS analyses of some of the 2017 bombs of Poás. All authors provided critical feedback on the manuscript and the ideas which are formulated here, and approved the final version.

#### ACKNOWLEDGMENTS

We would like to thank the entire team at OVSICORI-UNA, but in particular the field technicians and volcanological personnel who tirelessly maintain the monitoring network for all the volcanoes in Costa Rica, including Poás. RS would like to particularly thank Lauriane Chardot for inviting her to contribute to this special issue; and Calum Chamberlain and Alicia Hotovec-Ellis for answering numerous questions related to their respective codes. JdM thanks the Deep Carbon Observatory (DCO), Istituto Nazionale di Geofisica e Vulcanologia (INGV), and the United States Geological Survey Volcano Disaster Assistance Program (USGS-VDAP) for contributions to gas monitoring efforts at Poás. We thank the United States Geological Survey Volcano Disaster Assistance Program (USGS-VDAP) for providing seismic equipment to replace damaged seismic stations following the eruption. Finally, we thank W. McCausland, Andy Bell, and an anonymous reviewer as well as Corentin Caudron, the editor, for helpful comments which greatly improved the manuscript.

#### REFERENCES


explosive eruptions during the 2004–2005 period at Volcán de Colima, Mexico. J. Volcanol. Geothermal Res. 205, 30–46. doi: 10.1016/j.jvolgeores.2011.02.009

Arciniega-Ceballos, A., Chouet, B., and Dawson, P. (2003). Long-period events and tremor at Popocatepetl Volcano (1994-2000) and their broadband characteristics. Bull. Volcanol. 65, 124–135. doi: 10.1007/s00445-002-0248-8

Avard, G., de Moor J. M., Martínez, M., Bracamontes, D., Müller, C., Pacheco, J., et al. (2015). Volcán poás: Actividad Freática en el 2014. Technical report, Observatorio Vulcanológico y Sismológico de Costa Rica.


**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 Salvage, Avard, de Moor, Pacheco, Brenes Marin, Cascante, Muller and Martinez Cruz. 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.

## Breathing and Coughing: The Extraordinarily High Degassing of Popocatépetl Volcano Investigated With an SO<sup>2</sup> Camera

Robin Campion<sup>1</sup> \*, Hugo Delgado-Granados <sup>1</sup> , Denis Legrand<sup>1</sup> , Noémie Taquet <sup>2</sup> , Thomas Boulesteix <sup>3</sup> , Salvador Pedraza-Espitía1,4 and Thomas Lecocq<sup>5</sup>

<sup>1</sup> Departamento de Vulcanología, Instituto de Geofísica, Universidad Nacional Autónoma de México, Mexico City, Mexico, <sup>2</sup> Posgrado en Ciencias de la Tierra, Centro de Ciencias de la Atmósfera, Universidad Nacional Autónoma de México, Mexico City, Mexico, <sup>3</sup> Laboratorio Nacional de Geoquímica y Mineralogía, Universidad Nacional Autónoma de México, Mexico City, Mexico, <sup>4</sup> Posgrado en Ciencias de la Tierra, Universidad Nacional Autónoma de México, Mexico City, Mexico, <sup>5</sup> Seismology-Gravimetry, Royal Observatory of Belgium, Brussels, Belgium

#### Edited by:

Társilo Girona, Jet Propulsion Laboratory, United States

#### Reviewed by:

Giancarlo Tamburello, Istituto Nazionale di Geofisica e Vulcanologia, Italy Alessandro Tibaldi, Università degli studi di Milano Bicocca, Italy

> \*Correspondence: Robin Campion robin@geofisica.unam.mx

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 02 March 2018 Accepted: 28 September 2018 Published: 19 October 2018

#### Citation:

Campion R, Delgado-Granados H, Legrand D, Taquet N, Boulesteix T, Pedraza-Espitía S and Lecocq T (2018) Breathing and Coughing: The Extraordinarily High Degassing of Popocatépetl Volcano Investigated With an SO2 Camera. Front. Earth Sci. 6:163. doi: 10.3389/feart.2018.00163 How do lava domes release volcanic gases? Studying this problem is crucial to understand, and potentially anticipate, the generation of the sudden and dangerous explosive eruptions that frequently accompany dome extrusions. Since its awakening in 1994, Popocatépetl volcano has produced more than 50 lava domes and has been consistently among the strongest permanent emitters of volcanic gases. In this work, we have characterized the passive and explosive degassing between 2013 and 2016 at a high time resolution using an SO<sup>2</sup> camera, to achieve a better understanding of the conduit processes. Our 4-year average SO<sup>2</sup> flux is 45 kg/s, in line with the long-term average of the whole current eruptive period. We show that Popocatépetl volcano is essentially an open system and that passive degassing, i.e., degassing with no associated emission of lava or ash, dominates >95% of the time. This passive degassing is continuous and sustained, whether the crater contains a lava dome or not. It shows most of the time a strong periodic component, with a pseudo-period of ∼5 min, and amplitudes of 30 to 60% of the average value. We could distinguish two types of explosions based on their SO<sup>2</sup> flux patterns. The first type (E1) occurs in the middle of the normal passive degassing and is followed by a rapid return of the SO<sup>2</sup> flux down to its pre-explosive level. The second type (E2), which corresponds to the strongest events, is anticipated by a rapid decrease of the SO<sup>2</sup> flux to abnormally low values and is followed by a return to its normal values. The E2 explosions are probably caused by the accumulation of gas below a rapidly compacting permeable dome. We suggest that transient episodes of gravitational compaction of the usually permeable dome and the upper conduit is the only mechanism that is fast enough to explain the sharp decrease of the SO<sup>2</sup> flux that anticipates the E2 explosions. Our model is potentially applicable to a large number of andesitic volcanoes that undergo passive degassing interspersed with short-lived explosions.

Keywords: volcanic degassing, SO2 camera, Popocatépetl, lava dome, permeability, explosions

### INTRODUCTION

Lava domes are structures that result from the extrusion and accumulation of extremely viscous, quasi solid, lava and that are commonly formed at andesitic stratovolcanoes. They are often affected by dangerous eruptive phases involving partial collapse and/or the sudden transition to highly explosive activity (e.g., Boudon et al., 2015) that result in potentially dangerous pyroclastic density currents. The generation of explosive eruptions from lava domes is thought to be caused by spatial and temporal changes of their permeability (e.g., Collinson and Neuberg, 2012) and of their ability to exsolve and release volatiles (e.g., Sparks, 1997; Stix et al., 1997), but in detail, the causes are still a matter of debate. Recent work has been done on measuring experimentally the porosity and permeability of lava dome samples (e.g., Gaunt et al., 2014; Farquharson et al., 2015), but relatively few studies have focused on field measurements of gas fluxes from lava dome eruptions. Most of these studies have reported SO<sup>2</sup> fluxes that were generally low (0.5–10 kg/s), highly variable, or even intermittent (e.g., Young et al., 2003; Holland et al., 2011; Smekens et al., 2015), indicating that the studied domes (Soufriere Hills in Montserrat, Santiaguito in Guatemala and Semeru in Indonesia, respectively) were relatively weak emitters of SO<sup>2</sup> and that cyclic extrusion processes were controlling the release of the gas. Two volcanoes with lava domes, Lascar (Matthews et al., 1997) and Popocatépetl (Delgado-Granados et al., 2001; Delgado-Granados, 2008) were found to emit significantly larger SO<sup>2</sup> fluxes in the 1990s. While the former is not erupting anymore nor is it emitting large quantities of gas, the latter has been carrying up with its eruptive period and strong gas emission as of 2018, and is the subject of the present study.

Popocatépetl volcano (5,452 m a.s.l.) is a large compound stratovolcano located in central Mexico (**Figure 1**), between the megacities of Mexico City (∼25 million inhabitants, distant 70 km) and Puebla (∼7 million inhabitants, distant 40 km). It has been active since ∼500,000 years, erupting lava that ranged from basaltic andesites to dacites belonging to the calc-alkaline series (e.g., Siebe and Macías, 2006). Its historical activity has consisted of small to medium-scale explosions accompanying or alternating with extrusions of viscous intracrateric lava flows or domes. However, large effusive eruptions (Espinasa-Pereña and Martín-Del Pozzo, 2006), five powerful plinian eruptions (Siebe and Macías, 2006) and one massive sector collapse (Siebe et al., 2017) have occurred at the volcano during the last 25,000 years. The high recurrence of such events, coupled with the extraordinary large population living around Popocatépetl, makes the volcano one of the most probable candidate for a large volcanic disaster in the future (Siebe et al., 1996; De la Cruz-Reyna and Tilling, 2008; Delgado Granados and Jenkins, 2016).

FIGURE 1 | (a) Location of Popocatépetl Volcano within the Trans-Mexican Volcanic Belt. (b) Satellite image of Popocatépetl Volcano showing the viewing points used in this study.

After several years of increasing seismic and fumarolic unrest, a new eruptive period started at Popocatépetl in December 1994, and is still going on at the time of writing. The activity initially consisted of vent-clearing explosions and ash emission of phreatic origin until 1996 when, for the first time, a flat shaped lava dome was observed in the crater (De la Cruz-Reyna and Siebe, 1997). Since then, cycles of dome building and destruction have characterized the activity of the volcano (Gómez-Vazquez et al., 2016), slowly filling its 850 × 600 m wide summit crater (**Figure 2**). The total volume of erupted lava has not exceeded 4 10<sup>7</sup> m<sup>3</sup> (Gómez-Vazquez et al., 2016). The domes usually grow relatively quickly (within a few days to weeks), stall in the crater without further growth during a period that can last between a few days to a few years, until an explosion or a series of explosions destroys them. Gómez-Vazquez et al. (2016) found a weak correlation between the size of the domes and the magnitude of the explosions that destroy them. The mechanism of these explosions has been postulated to be gas accumulation beneath (or within) the cooling and crystallizing lava dome (Stremme et al., 2011; Gómez-Vazquez et al., 2016). The peak of activity, in 2000–2003, was characterized by the rapid growth of large lava domes, tens of strong vulcanian explosions, SO<sup>2</sup> fluxes up to 1,700 kg/s and a powerful subplinian phase that sent an ash column up to 17 km a.s.l. and produced 5 km long pyroclastic flows (Martin-Del Pozzo et al., 2003; Delgado-Granados, 2008). Evacuation of the closest villages was ordered during this eruptive phase. Several other phases of strong activity have occurred since then, such as in March-June 2012, April-July 2013, or January-May 2015 and September-November 2017.

Arguably the most distinctive aspect of the whole eruptive period has been the extremely high emissions of volcanic gases, and the extreme disproportion between the emitted gas and the erupted lava. The SO<sup>2</sup> flux has been measured at Popocatépetl since 1994, first with a COSPEC instrument (Delgado-Granados et al., 2001; Delgado-Granados, 2008), then with a network of scanning DOAS spectrometers, and more recently using satellites. The long-term average of SO<sup>2</sup> emission rates over the 24-year (1994–2017) eruptive phase has been around 55 kg/s (∼4,800 tons/day), while the peak emission rate reached the extraordinary value of 1,700 kg/s in December 2000. The cumulative SO<sup>2</sup> release over these 24 years of activity reaches the extremely large value of 4 ± 1 10<sup>7</sup> tons. For comparison, this amounts to twice as much as what Pinatubo emitted during its large plinian eruption of 1991, which is the highest measurement of eruptive SO<sup>2</sup> release on record. If the amount of gas emitted mostly passively by Popocatépetl during these 24 years had escaped massively in a short lapse like at Pinatubo, it could have fueled a plinian eruption comparable to those that the volcano produced in the last 25,000 years. Based on melt inclusions data in scarce olivine crystals, Roberge et al. (2009) concluded that this amount of gas could have been produced by the degassing of at least 3 km<sup>3</sup> of volatile-rich basaltic magma, which intrudes at depth >10 km and has remained essentially unerupted. Here we investigate the conduit processes that allow such a high and sustained degassing using measurements acquired with an SO2-camera at a high time resolution.

#### METHODOLOGY

Our measurements were obtained with an ultraviolet SO<sup>2</sup> camera (Mori and Burton, 2006; Kern et al., 2010b, 2015) during

FIGURE 2 | Typical styles of activity occurring at Popocatépetl Volcano. (a) Weakly explosive activity and (b) continuous ash emission associated to the construction of a lava dome. (c) Lava dome filling the crater and degassing passively. (d) Passive Degassing without a lava dome. (e) Vulcanian explosion associated to dome destruction. All photos by R.C. except (c) by Ramon Espinaza.

punctual campaigns through 2013–2016. Our instrumental setup is composed of two co-aligned Alta U260 cameras equipped with Pentax BUV2528 silica lenses, and UV band pass filters centered at 310 and 330 nm, respectively (Asahi XBPA310 and XBPA330, of 10 nm FWHM, and 75% peak transmittance), located in front of the lens. An additional Hoya340 filter was placed in front of each bandpass filter to avoid longer wavelength radiation to reach the CCD sensors through the leaks of the filters' transmittance function off their main peak. The instrument was operated from one of the spots shown in **Figure 1**, which are located at distances of 4–7.5 km from the crater. This range of distance, given the large size of the Popocatépetl volcanic plume, is considered as the best compromise for limiting the effect of light dilution (Mori et al., 2006; Kern et al., 2010a; Campion et al., 2015) and having the plume well-framed within the instrument's field of view (23◦ ). Images were acquired at a sampling rate of 5–15 s, depending on the distance to the plume and on the wind speed. The images were processed using the methodology described in Campion et al. (2015). The scattering coefficient of the atmosphere was first retrieved based on the exponential attenuation of the contrast with respect to the distance in a scattering atmosphere. Then, the differential absorbance was calculated for every pair of images after having them corrected for the light dilution effect using the scattering coefficient retrieved earlier. Finally, the 2D distribution of the SO<sup>2</sup> in the plume was obtained by multiplying the absorbance images with a calibration coefficient that was obtained by imaging a series of 5 calibration cells containing known concentrations of SO<sup>2</sup> (0, 500, 1,000, 1,500, and 3,000 ppm.m). The SO<sup>2</sup> emission rate was calculated by integrating the column amount measured along a profile perpendicular to the plume and multiplying this quantity by the projected velocity of the plume, which was measured by autocorrelation on a profile that was drawn parallel to the plume direction. All the results presented in this study were obtained during optimal measurement conditions, i.e., no clouds or haze between the plume and the volcano, plume fully framed in the field of view, well defined plume transport direction, distances inferior to 7 km and optically thin plume. Therefore, we estimate the total error on the flux to be below 25% (Campion et al., 2015). However, a larger error likely affects the column amounts retrieved just above the vent and in the very proximal parts of the plume because the high aerosol optical densities and SO<sup>2</sup> columns, which often exceeded our highest calibration cell, make the plume nearly opaque at 310 nm. We avoided this problem by measuring the SO<sup>2</sup> flux downwind of the crater where the plume has already been diluted enough to be optically thin and have its SO<sup>2</sup> column in the range of our calibration cells. Ash in the plume causes a systematic underestimation of the retrieved SO<sup>2</sup> column, because in that case the light reaching the camera originates mostly from reflection of the sun light on the particles of the outer shell of the plume.

#### RESULTS

Over the 4-year period (2013–2016), we collected SO<sup>2</sup> camera data fulfilling the above-defined quality criteria over 20 days, amounting to ∼80 h of recordings (**Table 1**). Based on the visual observations during the measurements, we distinguished three types of activity: passive degassing, explosions and continuous ash emissions. Passive degassing, by far the most common form of activity at Popocatépetl, is defined as the continuous release of ash-free plume. Explosions are short-lasting energetic emissions of ash-laden plume that occasionally eject rock fragments outside of the crater. Dense juvenile material is by far the dominant component in the ashes produced by the explosions. Episodes of continuous ash venting, the less frequent form of activity, usually last from a few hours to a few days and are associated with the growth of lava domes. Abundant lava fragments are also ejected in -or outside of the crater during these episodes. The ash emitted during these episodes is a mixture of dense and vesiculated juvenile fragment, unlike the ash from explosions. Observation through four permanent webcams shows that the volcano has a rather monotonous behavior and that these three styles are enough to describe the whole activity of the volcano in the last 4 years (see also Gómez-Vazquez et al., 2016 and Centro Nacional de Prevención de Desastre, 1995). We acknowledge that the low number of measurements hours and days over the reporting period is insufficient for establishing the long-term evolution of the SO<sup>2</sup> flux, and emphasize that this study focuses on rapid fluctuations associated to conduit processes. However, we obtained SO<sup>2</sup> camera measurements of each eruptive style, so that, although the total duration of our measurements only amounts to 80 h, they can be considered as representative of the short-term volcano behavior.

#### Passive Degassing

Passive degassing at Popocatépetl is permanent and our measurements showed that typical SO<sup>2</sup> fluxes range between 20 and 80 kg/s, with a 4-year average of 45 kg/s (3,900 tons/day). SO<sup>2</sup> fluxes measured a few hours to weeks after a dome growth episode are similar to periods where no dome was present. This implies that the presence of a lava dome does not seem to decrease the overall permeability of the conduit system. A distinctive characteristic of Popocatépetl degassing is its puffing behavior, which is characterized by quasi-periodic oscillations of the SO<sup>2</sup> flux time series, whose relative amplitude is typically 30–50% of the mean value (**Figure 3** and **Video 1**). The SO<sup>2</sup> mass of individual puffs ranges between 0.5 and 10 tons. This is the first time that puffing is quantified at Popocatépetl volcano, although some hints of its existence had been previously obtained by visual observation and by flying with a COSPEC parallel to the plume axis (Delgado-Granados, unpublished data). Puffing at Popocatépetl is observed systematically every time the wind speed is below ∼15 m/s. At higher wind speeds, the plume is forced back into the crater by a strong vortex that develops downwind from the summit and subsequently bent down along the upper slope of the volcano. This homogenizes the plume and blurs the puffing signature. We applied a Fourier Transform to the time series of SO<sup>2</sup> flux to derive their power spectra. The spectra show a prominent peak corresponding to the periodic puffing (**Figures 3D,E**), whose fundamental period is systematically between 200 and 400 s. Longer period flux



The peak period was calculated by taking the Fourier Transform of the SO<sup>2</sup> flux time series. The relative puff amplitude is calculated as the average, for the whole series, of the (Fmax - Fmin)/(Fmax+Fmin) where Fmax and Fmin are the flux maxima and minima associated to each successive puff. The average and standard deviation of the peak periods were calculated excluding those days where a longer period component was present in the power spectra of the time series, which was usually associated to explosions.

variations dominate only on days where explosions occur, and are associated with the decrease of the SO<sup>2</sup> flux before them.

#### Explosions

Explosions occur rather frequently at Popocatépetl volcano (e.g., **Figure 2e**), varying in size from small ash puffs to strong vulcanian explosions showering the slopes of the volcano with ballistic fragments up to distances of 4 km. The high ash content of the explosions plumes induces a systematic underestimation of the SO<sup>2</sup> measurements, and can even completely hamper the retrieval if the plume is completely opaque. A total of 17 explosions were captured by the camera during the campaigns. Based on the evolution of the flux before, during and after each explosion, we could recognize two types of explosions. Explosions of the first type (hereafter called E1) produce a peak of SO<sup>2</sup> flux interrupting the normal passive degassing, and are followed by a rapid (a few minutes) return to the pre-explosion flux values (**Figure 4**). The E1 explosions usually produce low to moderate amounts of ashes. In some instances, once the plume was sufficiently diluted, we could obtain a lower constraint on the SO<sup>2</sup> mass released by each explosion by integrating the SO<sup>2</sup> flux peak above a baseline, defined as the SO<sup>2</sup> flux measured just before the leading edge of the explosion plume reaches the integration line. Several of these explosions appear as spikes emphasized with red arrows, in the graph of **Figure 4B**, which shows one of the longest time series we have been able to obtain so far on a day where explosions were occurring. The resulting values range between 2 and 12 tons of SO2. Assuming a standard subduction zone magmatic gas that contains 2 mol% SO2, 90 mol% H2O, and 8 mol% CO<sup>2</sup> (e.g., review by Taran and Zelenski, 2014), these SO<sup>2</sup> masses translate into total amounts of released gas is in the range of ∼30–200 tons.

The second type of explosions (E2) is characterized by a period of anomalously low SO<sup>2</sup> flux preceding the explosion and a return to the more typical high and sustained flux after the explosion has occurred (**Figure 5**). An animation made from SO<sup>2</sup> measurements during a moderate E2 explosion is provided as Supplementary Material (**Video 2**). The E2 explosions seem to be less common than the E1, as only four of them (compared to 13 E1) were recorded with the UV camera over the measurement

SO2 fluxes on a day where E1 explosions (shown as red arrows) were occurring frequently. The mass released by each explosion is written next to its arrow, when it was possible to calculate it.

period. They are also usually more energetic, produce larger quantities of ash and sometimes eject bombs outside of the crater. The SO<sup>2</sup> flux pattern associated with these explosions suggests that they are triggered by the accumulation of gas under a temporary plug or seal of the upper conduit, as is the case for the vulcanian explosions of Sakurajima (e.g., Iguchi et al., 2008; Kazahaya et al., 2016). The high ash content in E2 explosions unfortunately prevents quantifying their SO<sup>2</sup> content using an SO<sup>2</sup> camera, because of the complete opacity of the plume close to the crater. However, since the decrease of the flux preceding these explosions was on some occasions well characterized as a sharp drop of SO<sup>2</sup> emissions from their initial values, we could estimate the mass of accumulated gas by integrating the flux curve below its former baseline. This yields values of 10 to 50 tons of SO<sup>2</sup> per explosion, which, assuming the same gas composition as earlier, correspond to total amounts of accumulated gases in the range of ∼160 to 800 tons. However, it should be emphasized that the four E2 explosions that we have measured are by far not the largest (in terms of the number of ballistic fragments expelled, the distance they reach and the eruptive column height) that the volcano has produced over the reporting period. These stronger E2 explosions, observed both in the field and with the

FIGURE 5 | SO2 time series for the four E2-type explosions that we recorded so far. Except for the explosion of 29/04/2014, for which the SO2 measurements started just 20 min before the event and the flux decrease was probably missed, all the explosions share a common pattern, featuring a rapid decrease of the SO2 flux 20–60 min before the eruption, a period of low flux where gas accumulates and pressure builds up, and significantly higher post explosive flux values. The SO2 mass accumulated before the explosions (whose time is indicated as the red arrow) is calculated by integrating the curve below the background flux (highlighted by the red dashed line).

webcam images, have visually the same behavior as the small and moderate E2 explosions that we were able to measure. They are preceded by a period of reduced gas emissions, have an impulsive start and are followed by a prolonged period of stronger, pulsating gas emissions.

#### Sustained Ash Emissions

Due to their rarity, only one episode of sustained ash emission could be measured over the reported period, on the 26/01/2016, during an episode of dome growth that lasted for 3 days and emplaced ∼2 10<sup>6</sup> m<sup>3</sup> of lava (Centro Nacional de Prevención de Desastre, 1995). The average SO<sup>2</sup> emission rate for this day is 120 kg/s, which corresponds to the highest value measured with the camera over the reporting period. Yet, this value is probably still an underestimation because of the presence of ash in the plume. The processing of an image taken by the satellite-based Ozone Monitoring Instrument (OMI, Carn et al., 2008) on the next day, while the episode was still in progress, yields an emission rate of ∼250 kg/s (**Figure 6**). Interestingly, since the beginning of the current eruptive period, the highest SO<sup>2</sup> flux measured with COSPEC have also been systematically associated with episodes of lava dome growth.

#### DISCUSSION

#### Comparison With Previous Studies at Popocatépetl Volcano

In this section we compare our SO<sup>2</sup> measurements with the previous published studies, summarized in **Table 2**. Delgado-Granados et al. (2001) and Delgado-Granados (2008) have reported measurements of the SO<sup>2</sup> flux with a COSPEC during the earliest part of the current eruptive period. Their measurements had a long-term average of 100 kg/s (8,600 tons/day) and ranged from 10 to 1,500 kg/s during the most intense volcanic activity, in December 2000-January 2011. Grutter et al. (2008) reported results from a 3-week long multidisciplinary campaign in March 2006, involving Mobile DOAS traverses, COSPEC traverses and a fixed scanning DOAS instrument. The average SO<sup>2</sup> flux values resulting from their study was 28 kg/s (2,450 tons/day). Lübcke et al. (2013) measured the SO<sup>2</sup> flux using an SO<sup>2</sup> camera. They reported an average flux of 13 kg/s (1,120 tons/day), without light dilution correction. These last two studies were made when Popocatépetl was in a notably lower state of activity than during 1997–2003 or since 2012. Finally those two last studies and our results are systematically higher, by a factor of about two, than the corresponding yearly-averaged fluxes computed by Carn et al. (2017) using the images of OMI. We suspect that the cause of this discrepancy lies in the turbidity and thickness of the boundary layer (the three lowermost kilometers of the troposphere where most of the water vapor and aerosols reside) over central Mexico, which alters the radiative transfer and the air mass factor compared to the model parameters used in OMI retrievals.

#### Comparison With Other Volcanoes

The long-term average of our SO<sup>2</sup> emission rate measurements at Popocatépetl, 45 kg/s, places the volcano as one of the five strongest permanent emitters of volcanic SO<sup>2</sup> over 2013– 2016, together with Ambrym (100 kg/s; Allard et al., 2016),

compared to measurements obtained by processing the OMI image of the following day (B) using the traverse method (Campion, 2014). The SO2 flux, in kg/s, measured on each traverse downwind of the plume is annotated next to its corresponding traverse.


Reference numbers are as follows: (1), Delgado-Granados et al. (2001); (2), Delgado-Granados (2008); (3), Grutter et al. (2008); (4), Carn et al. (2017); (5), Lübcke et al. (2013).

Nevado del Ruiz (20–80 kg/s; Lübcke et al., 2014), Kilauea (10– 60 kg/s; Nadeau et al., 2014), and Nyamuragira (20–60 kg/s; Coppola et al., 2016). It should be noted that over the considered period, Popocatépetl has been a stronger SO<sup>2</sup> emitter than other volcanoes well known for their strong degassing such as Etna, Masaya, or Nyiragongo.

Relatively few studies are available on high time resolution SO<sup>2</sup> measurements at dome volcanoes with intermittent explosive activity. Fischer et al. (2002) investigated the degassing of Karymsky volcano and reported very low passive emission of SO2, interrupted by short period of higher flux (up to 3 kg/s) associated to mild explosive activity. They attributed this pattern to the pressurization of the upper conduit beneath a lava plug that quickly sealed after releasing its pressure through explosive degassing. They could distinguish two types of explosions, those followed by a rapid return of the SO<sup>2</sup> emission to their very low background level and those followed by a gradual, waxing and waning decay toward background. Smekens et al. (2015) reported the very same type of behavior at Semeru volcano (Indonesia). Holland et al. (2011) measured slightly higher interexplosive SO<sup>2</sup> emissions at Santiaguito (Guatemala) and gave a different interpretation of the degassing mechanism, calling forth enhanced exsolution and release of gas through ring fractures during the stick-slip upwards motion of the lava plug. The pattern reported in this study differs significantly from all those other dome-bearing volcanoes. The SO<sup>2</sup> emissions of Popocatépetl are two orders of magnitude higher than at Semeru, Karimsky and Santiaguito and unlike these volcanoes, are sustained permanently. Even the maximal SO<sup>2</sup> fluxes measured at the above-mentioned volcanoes during explosive activity is still largely inferior to the emission rates emitted by Popocatépetl between two puffs of purely passive degassing. Such high and persistent gas flux values imply that sealing of the conduit does usually not occur at Popocatépetl and the degassing models developed for the afore mentioned volcanoes may not be applied.

Measurements of SO<sup>2</sup> at the actively growing lava dome of Soufriere Hills volcano (e.g., Young et al., 2003) showed significantly higher SO<sup>2</sup> fluxes (typically 3–15 kg/s) than for those three former volcanoes. The fluxes were sometimes following well-defined periodic cycles of several hours, which were interpreted as being caused by the pressurization of the magma conduit and associated to changes of extrusion rate. The degassing of Popocatépetl, which is nearly one order of magnitude stronger than that of Soufriere Hills is also different in its behavior because the fluctuations of the SO<sup>2</sup> flux at Popocatépetl are much faster (a few minutes vs. several hours) and not associated to the growth of the lava dome.

SO<sup>2</sup> camera measurements at Sakura-jima volcano (Kazahaya et al., 2016) have shown a very similar pattern to the one reported here for Popocatépetl, with high, sustained flux and occasional short-lived drops preceding explosions. Although Sakura-jima does not often build volcanic domes, geophysical and gas measurements have been inferred to be modulated by the temporary formation and destruction of lava plugs (Iguchi et al., 2008; Kazahaya et al., 2016), which can be viewed as lava domes at embryonic stages. Visual observations suggest that a number of volcanoes in the world share a similar behavior to Popocatépetl, among which Tungurahua (Ecuador, Hall et al., 2015), Ubinas and Sabancaya (Peru, Author's observations), Dukono and Agung (Indonesia, Syahbana, pers. com.) and Nevado del Ruiz (Colombia, Chacón-Ortíz, pers. com.).

#### Origin of the Periodic Puffing

Periodicity in passive degassing has been observed in time series of SO<sup>2</sup> fluxes of many other volcanoes, such as Stromboli (period of around 1 s, Tamburello et al., 2012), Turrialba (period of about 100 s, Campion et al., 2012), Erebus (period of 500– 1,000 s, Boichu et al., 2010) and Etna (Tamburello et al., 2013). Two processes can be envisaged as a cause of the puffing: pulsating release of gas directly at the vent area or turbulent entrainment of atmospheric air when the hot gases mix with the colder atmosphere. Moussallam et al. (2016) have argued that since turbulence is a chaotic process, it should not produce periodical puffs. However, a chaotic behavior can include, time to time, intermittency that means periods of regular and/or periodic behavior. The pulsating behavior is already present when measured on a transect drawn very close (200 m) to the crater rim, where it is actually stronger and better-defined (**Figure 7**). This argues against the hypothesis that the puffing is a transport effect (Tamburello et al., 2013). In the case of Popocatépetl, we propose that the puffing likely has a volcanic origin because its regularity and its characteristic frequency are independent of the climatic conditions. A strong argument in favor of the volcanic origin, is provided by the higher altitude reached by the distinct

close and further downwind of the volcano. (A) Distribution map of the SO2 in the plume, with the proximal transect shown in black (traced ∼200 m downwind of the crater), and the distal transect shown in red (∼2.3 km from the crater). (B) Respective time series of SO2 measured on the two transects. (C) Respective power spectra computed from these time series. The spectra of the distal time series shows a less prominent peak at the puffing frequency, and higher power distribution at the lower frequencies, indicating a smoothing of the puffs with transport.

gas pulses (**Figure 3A**), which results from a higher thermal energy of the puffs. The more energetic release of the puffs is well perceptible in **Video 1**.

#### Permeable Dome and Gas Transfer Mechanisms

Similarly to the earlier (1994–2003) stages of the current eruptive period that started in 1994 and is ongoing as of 2018 (Delgado-Granados et al., 2001; Delgado-Granados, 2008), the SO<sup>2</sup> flux measured during the reporting period (2013–2016) is more than an order of magnitude too high to result solely from the degassing of the erupted magma. The estimated ∼10<sup>7</sup> m<sup>3</sup> of magma emitted over the 2013–2016 should have produced an average SO<sup>2</sup> flux of at least 0.8–1.8 kg/s, estimated assuming the complete degassing of a primitive magma having a density of 2.5 g/cm<sup>3</sup> and containing an initial S content of 2,500 ppm (Witter et al., 2005; Roberge et al., 2009). This is much smaller than the average value of 45 kg/s measured over the whole survey period. It is thus clear that the degassing of the sulfur from the magma is taking place in the deeper part of the magmatic system (>10 km according to Roberge et al., 2009) and that <2% of the intruded magma reaches the surface, while the gas that this magma produces is efficiently transferred through the conduit system. Our results show that the whole conduit system of Popocatépetl volcano is essentially permeable to this deep gas flow, whether being capped by a lava dome or not. This is supported by airborne observations that the domes are affected by numerous fractures that, together with the dome-conduit boundary, let the gas escape freely to the atmosphere. It is likely that this fracture-network permeability develops as early as the growth stage of the dome. The gas transfer mechanism within the deep conduit system is not known with certainty, magma convection, and gas fluxing within interconnected vesicles being the most likely candidates. These mechanisms are not mutually exclusive. Depending on the magma viscosity, vesicularity, and percentage of interconnected vesicles, each of them may dominate at certain depths or time.

Immediately beneath the dome and in the upper part of the magma column, the gas transfer is probably achieved through a network of highly connected vesicles (e.g., Burgisser and Gardner, 2005; Schipper et al., 2013) that pervades the highly viscous magma in prolongation of the fractures of the dome. At Popocatépetl volcano, the magma column in the upper conduit is stalled for much of the time, except during the relatively infrequent and short episodes of dome growth. Thus, magma shearing cannot be invoked as a factor that helps maintaining the bubble network connected, as it has been inferred from laboratory experiments (Okumura et al., 2006) and field/sample studies (Schipper et al., 2013). The absence of magma movement also excludes the stick-slip mechanisms and the associated repeated fracturing of the magma/conduit interface that is thought to foster relatively quiet degassing in lava dome eruption (e.g., Holland et al., 2011). Therefore, the gas flow pressure is the only mechanism that may explain that the fracture networks in the shallow dome and the vesicle networks in the upper magma column below the dome stay open and permeable. The continuous fluxing of pressurized gases from depth is maintaining the vesicles network and fracture network of the upper conduit, acting against lithostatic pressure that tends to compact and close the system. Pressure oscillations resulting of the opposition between these two forces may be the cause of the puffing behavior of the passive degassing. An increase in magma pressure or an upward movement of the underlying magma column could also theoretically promote the compaction of the upper conduit system, but we believe that the gas flow would increase accordingly, maintaining the permeable networks open. In addition, if the compaction of the upper conduit was due to an increase of the magma pressure, the explosions would be followed by an episode of magma emission at the surface, once the dome is destroyed and is no more an obstacle to the further rise of magma, which has not been observed. At higher depth, the magma should be less viscous and support bubble flow and/or magma convection to transport gas toward the surface, but more work on the depth-dependence of the magma viscosity is needed to identify the gas transfer mechanism in the deep conduit system.

#### The E1 Explosion: Percolating Slugs or Bigger Than Normal Puffs?

In this section, we discuss the possible origin of the E1 explosions. E1 explosions usually produce small plumes with relatively little ash, allowing sometimes to quantify the SO<sup>2</sup> with only a modest underestimation. These explosions occur in the midst of the normal degassing and are characterized in the SO<sup>2</sup> flux time series by a spike lasting a few minutes followed by a rapid return to the pre-explosion flux values. This SO<sup>2</sup> flux pattern associated with the E1 explosions is similar to the explosions of Stromboli volcano (Tamburello et al., 2012), which are thought to be caused by the bursting of gas slugs ascending through the conduit system (e.g., Vergniolles et al., 1996; Burton et al., 2007). Based on this similarity, E1 explosions could be caused by the ascent of slugs through the deep conduit and their successive percolation through the dome. Instead of bursting at a free magma interface like in Stromboli and other strombolian volcanoes, a gas slug reaching the upper part of Popocatépetl volcano would have to percolate through the interconnected vesicles zone and through the fractured dome. The increase of SO<sup>2</sup> associated with the E1 explosions is emergent rather than impulsive, which is consistent with the percolation of the gas slug rather than its bursting. The gas masses calculated for E1 explosions are about two orders of magnitudes larger than the SO<sup>2</sup> masses emitted by the typical strombolian explosions in Stromboli (Mori and Burton, 2009; Tamburello et al., 2012; Delle Donne et al., 2016) but this scales generally with the difference in the SO<sup>2</sup> flux of the two volcanoes.

The SO<sup>2</sup> masses released by the E1 explosions, although likely underestimated, are higher than those released by individual puffs, but not completely out of their range, as shown in the histogram (**Figure 8**). This suggests the alternative hypothesis that E1 explosions might share a common process of formation with the puffs, and be actually larger or more energetic puffs involving coalescence events and fragmentation in the interconnected vesicles zone.

#### A New Model for the Explosion Mechanism and Triggering

Vulcanian explosions at Popocatépetl have been proposed by various authors (Love et al., 2000; Schaaf et al., 2005; Stremme et al., 2011; Gómez-Vazquez et al., 2016) to be caused by the

FIGURE 8 | Histogram comparing the frequency distribution of the SO2 masses emitted by individual puffs of passive degassing and by E1 explosions, on a same day.

gas accumulation below a dome that is cooling until it plugs the conduit. Positive feedback between crystallization and degassing in the shallow magma column was also invoked (Stix et al., 1997; Schaaf et al., 2005) to produce the overpressure necessary for the strong vulcanian explosions. Arguments in favor of this model were:


$$2\text{SiO}\_2 + 4\text{ HF} < ->\text{SiF}\_4 + 2\text{ H}\_2\text{O} \tag{1}$$

whose equilibrium is displaced to the right at low temperature (Symonds and Reed, 1993). Love et al. (2000) and Stremme et al. (2011) interpreted the increase of SiF<sup>4</sup> to result from colder equilibrium temperature of the gas, and to record the cooling of the lava dome. However, the equilibrium temperatures calculated by these authors were unrealistically low (150–180◦C) and in contradiction with the continuous incandescence observed in the crater at night.

However, our results and other observations do not support the cooling and crystalizing model of the explosions generation. These are:


while cooling, crystallization and solidification of lava domes require periods of weeks to years depending on their volume (e.g., Hicks et al., 2009).


The model, illustrated in **Figure 9**, that we propose for the generation of the E2 explosions also assumes accumulation of gas before the explosions, but differs in the cause of this accumulation. It accounts for the above-mentioned observations as well as for those in favor of the old cooling and crystalizing model. In our model, the accumulation of the gas is due to a compaction of the permeable networks that normally allows the gas to flow through the upper conduit and dome. This lithostatic squeezing leads to a dramatic decrease of the upper conduit permeability, which promotes the accumulation of the deep gas until it reaches enough pressure to disrupt the blockage through an explosion. Laboratory experiments of uniaxial, gravitational compaction of rhyolitic magmas by Okumura and Sasaki (2014) have shown drastic decreases of permeability achieved in timescales of 100–1,000 s. These timescales are strikingly similar to the decrease in SO<sup>2</sup> flux that we observe preceding E2 explosions, and way faster than any other mechanisms able to decrease permeability, such as cooling or mineral deposition in fractures and pores. An additional argument in favor of our model of compacting-induced explosions is that inward sagging and deflation of the dome have often been observed at Popocatépetl during occasional surveillance overflights, although their infrequence hampers to establish a univocal systematic time correlation between these phenomena and the explosions (Centro Nacional de Prevención de Desastre, 1995). Since we have shown in section Origin of the Periodic Puffing that the high pressure of the gas flow is the main factor that maintains open the fracture network in the dome and the vesicle network in the upper magma column a slight reduction of the gas flux would leave these permeable networks unsupported and would allow the lithostatic pressure and the weight of the dome to compact them, initiating the gas accumulation. One of the key observations invoked to support the earlier cooling and crystalizing dome model was the increased SiF4/SO<sup>2</sup> ratio in the emissions before and during an explosion (Love et al., 2000; Stremme et al., 2011). However, thermodynamic data reported by De Hoog et al. (2005) for equation (1) show that the pressure dependence of this equilibrium is actually much stronger than its temperature dependence, especially at the pressures corresponding to a shallow magmatic column. More recently, Taquet et al. (2017) measured the SiF4/SO<sup>2</sup> ratio over a period of several months and reported increases so large and so fast associated with explosive events that they are explained much

more convincingly by an increase of the equilibrium pressure of the emitted gas than by a decrease of its equilibrium temperature. Our model readily explains this increase of equilibrium pressure by the pressurization of the gas rapidly accumulating below a gravity-compacted dome and underdome. A similar dynamics has been proposed to explain the eruptive behavior of Lascar volcano (Northern Chile) between 1984 and 1994, which was characterized by high gas fluxes, and cycles of building of low aspect ratio lava domes, decreasing of the degassing, subsidence of the dome and strong vulcanian explosions (Matthews et al., 1997). If our model is correct, then E2 explosions should be preceded by a small transient deflationary signal in the tilt accompanying the dome compaction, followed by a slow inflation corresponding to the phase of gas accumulation and finally a rapid deflation associated with the explosive decompression of the upper conduit system. Due to the relatively superficial origin inferred here for the E2 explosions, it would be important to place tiltmeters as high and close to the crater as possible.

form a permeable network for the gas flow at a shallower level.

#### CONCLUSION AND FUTURE WORK

SO<sup>2</sup> camera measurements at Popocatépetl confirm that this volcano emits extraordinarily high SO<sup>2</sup> fluxes despite having its crater occupied most of the time by a stalled lava dome. This implies that this lava dome and the underlying upper conduit are mostly permeable to the flux of gas coming from the deeper parts of the magmatic system. This high permeability is maintained for long periods (up to several months) despite the absence of magma motion in the conduit, which has been often invoked as a factor enhancing the permeability of the magma-conduit interface. We thus propose that the gas flux is maintaining open the fracture network in the dome and the interconnected vesicles network below it. These permeable networks, however, can close rapidly through compaction if the gas flux slightly decreases, causing gas accumulation and pressurization that eventually leads to an explosion. The puffing and the frequent E1 explosions maintain the upper conduit permeable, while the E2 explosions restore its permeability when it drops due to compaction. Future work toward a more complete understanding of the degassing dynamics should include the installation of a web camera on the crater rim, to investigate the distribution of the degassing vents inside the crater, and the time relationship between the inferred dome subsidence and explosions. Installation of closefield tiltmeters would also help to validate our new model for the generation of E2 explosions and to constrain the depth of the gas accumulation. Infrasound measurements would help to elucidate the origin of the puffing, which we tentatively attribute to pressure oscillations in the gas flow through the permeable networks. Measurements of the gas composition could be performed more systematically to elucidate the origin of the E1 explosions. The recognition of two different types of explosions and the hypothesis we formulate on their mechanism could form a process-based fundament for the seismic-based distinction between exhalation and explosion (De la Cruz-Reyna and Tilling, 2008). Finally, our model of explosion generation by rapid compaction of the upper magma column is applicable to

#### REFERENCES


other andesitic volcanoes that exhibit sustained gas emissions and undergo frequent, rapid transitions to explosive activity, such as Tungurahua, Ubinas, Sabancaya, Nevado del Ruiz, Sakura-jima and Dukono. We suggest that at those volcanoes, similarly to what happens at Popocatépetl, a decrease in the gas flux could actually foster the lithostatic compaction of the upper magma column and trigger a transition from passive degassing toward more intermittent and violent release of gas through the so-called vulcanian explosions.

#### AUTHOR CONTRIBUTIONS

RC make the measurements in the field, wrote the code for processing the data, interpreted the data, wrote the manuscript. HD-G impulsed this research, NT and SP-E took part to the fieldwork. SP-E contributed to the code for processing the data. TL wrote the code for operating the Camera. All discussed the data and their interpretation and revised the manuscript.

#### ACKNOWLEDGMENTS

This research was funded thanks to the proyect Papiit IA103816. We thank Servando de la Cruz-Reyna and Yuri Taran for fruitful discussions. Two reviewers as well as editors Tarsilo Girona and Valerio Acocella are thanked for helping to improve the manuscript.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart. 2018.00163/full#supplementary-material


central Mexico: past key to the future? Geology 24, 399–402. doi: 10.1130/0091-7613(1996)024<0399:RVDIPT>2.3.CO;2


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

Copyright © 2018 Campion, Delgado-Granados, Legrand, Taquet, Boulesteix, Pedraza-Espitía and Lecocq. 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.

## Seismic Data, Photographic Images and Physical Modeling of Volcanic Plumes as a Tool for Monitoring the Activity of Nevado del Ruiz Volcano, Colombia

#### John Makario Londono1,2 \* and Beatriz Galvis 1,3

<sup>1</sup> Servicio Geológico Colombiano, Bogotá, Colombia, <sup>2</sup> Faculty of Engineering and Architecture, Prevention, Reduction and Atention of Disasters Esp., Universidad Católica de Manizales, Manizales, Colombia, <sup>3</sup> Geology Department, Faculty of Exact and Natural Sciences, Universidad de Caldas, Manizales, Colombia

Edited by: Yosuke Aoki, The University of Tokyo, Japan

#### Reviewed by:

Edouard Kaminski, UMR7154 Institut de Physique du Globe de Paris (IPGP), France Luca De Siena, University of Aberdeen, United Kingdom Tobias Dürig, University of Otago, New Zealand

> \*Correspondence: John Makario Londono jmakario@sgc.gov.co

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 19 April 2018 Accepted: 25 September 2018 Published: 06 November 2018

#### Citation:

Londono JM and Galvis B (2018) Seismic Data, Photographic Images and Physical Modeling of Volcanic Plumes as a Tool for Monitoring the Activity of Nevado del Ruiz Volcano, Colombia. Front. Earth Sci. 6:162. doi: 10.3389/feart.2018.00162 Quantification of volcanic plume parameters is a fundamental task to characterize the behavior of an active volcano. The volcanic plume mass, flow rate and ash injection were determined from seismic data, in addition to photographic images and integration of scaling laws of several volcanic plume models, for the period from 1985 to 2017 for the Nevado del Ruiz Volcano (NRV), Colombia. With these parameters we quantified the ash volume emitted during this period and established a relationship between seismicity and the volcanic plume parameters. The results revealed a decrease of approximately two orders of magnitude in the volume of ash plumes from the November 13, 1985, eruption (0.12 km<sup>3</sup> ) to the September 1, 1989, eruption (1.43 × 10−<sup>3</sup> km<sup>3</sup> ). This pattern continued for the June 30, 2012, eruption and 2015–2017 eruptive cycle, with volumes five times smaller than that observed in 1989. The results also exhibited a correlation between the radiated seismic energy (RSE) of the volcanic tremor and ash load for higher (>1 km) and longer-duration (>240 s) plumes. It was possible to calculate a minimum value of ash load based on RSE release and reduced displacement (RD, a means of normalizing volcanic tremors to a common scale) of volcanic tremor signals associated with the eruptions for the period 2015-2017. Moreover, changes in the volume of the ash plume were correlated with changes in the RD and RSE associated with different stages of volcanic activity. These findings can be used as a tool for monitoring the NRV. The continuously decreasing ash plume volumes from 1985 to 2017 suggest a common volcanic cycle that is almost ending. On the other hand, the evidence of new magmatic input in 2007 might suggest that a new volcanic cycle started on that date and is still in the process of ascending magma. It is likely that in the near future surface evidence of the new cycle will be observed at the NRV.

Keywords: volcanic plume source, nevado del ruiz volcano, reduced displacement of volcanic tremor, volcano monitoring, radiated seismic ennergy, volcanic eruption

### INTRODUCTION

Volcanic ash plumes pose a serious threat to both the neighboring population (e.g., Sparks et al., 1997; Horwell and Baxter, 2006; Jenkins et al., 2015) and civil aviation (e.g., Guffanti et al., 2010; Guffanti and Tupper, 2015). Their formation and dynamic behavior are controlled by the generation processes of ash (e.g., Dürig et al., 2012; Dellino et al., 2014; Dioguardi et al., 2016) and the fluid dynamic conditions at the source, which are described by the eruption source parameters (e.g., Wilson and Walker, 1987; Woods, 1988; Carazzo et al., 2008; Dellino et al., 2014). Quantification of volcanic plume parameters is fundamental to characterize the behavior of an active volcano and therefore a crucial step toward assessing its risk. Different approaches have been proposed. Starting from the seminal work of Sparks (1986), who proposed a basic physical model to calculate plume height, various 0D (0D models refer to scaling laws that relate plume height to eruptive mass or mass flow rate with no spatial variables being considered, e.g., Mastin et al., 2009), 1D (e.g., Devenish, 2013; Mastin, 2014), and 3D (e.g., Cerminara et al., 2016) models have been developed to quantify the main parameters of a volcanic plume (Bonadonna et al., 2002b; Kaminski et al., 2011; Degruyter and Bonadonna, 2012; Woodhouse et al., 2013; Suzuki and Koyaguchi, 2015; see Costa et al., 2016 for a review).

On the other hand, the current availability of cameras, videos, and other devices is becoming useful to understand in realtime the dynamics of volcanic plumes. In addition, seismic data can supply important information about the inner activity of a volcano. By combining all these data together, it is possible to obtain a better image of the volcano behavior, helping to monitor and forecast increases in volcanic activity.

Various studies have used such data to obtain a picture of the eruptive history of different volcanoes. For example, at Mt. Etna (Italy), Andronico et al. (2013) related ash emissions and the seismo-acoustic signals associated with them. They conclude that it is possible to differentiate two ash emissions types from the seismo-acoustic signals; one related with closed and the other with open conduits. McNutt et al. (2013) studied the 2009 explosive eruptions of Redoubt Volcano (Alaska) using seismic and infrasound signals. They associated the seismic energy and pressure obtained from infrasound data with column height. They found a correlation of seismic energy with gas release (SO2) and pressure with column height. Dürig et al. (2015) analyzed the 2010 eruption of Eyjafjallajökull Volcano (Iceland) by using high-resolution video and aerial observations. They were able to observe the pulsatile activity of the volcano and to quantify the velocity of the pulses of ash emission. In addition, they estimated mass flux and plume height in good agreement with data from plume height and mass discharge models. De Angelis et al. (2016) studied the gas and ash explosions at Santiaguito Volcano (Guatemala) and the associated seismic signals (acoustic waveforms) and thermal infrared images with the aim to assess the bulk density of the eruptive plume, in addition to the fraction of ash and gas in the eruptive plume. They concluded that small to moderate explosions contained small fractions of ash. Romero et al. (2016) studied the eruption dynamics of Tungurahua volcano (Ecuador) based on fieldwork, thermal images and photographic images. They modeled the source parameters and suggested changes in the eruptive style. Fee et al. (2017) estimated the erupted mass of Sakurajima Volcano (Japan) by using infrasound waveforms, in combination with ash and gas data. They inverted infrasound waveforms associated with eruptions to quantify eruption flow rate and masses of 49 explosions. Their results agree with those obtained from groundbased ash collection and SO<sup>2</sup> data. Although far from complete, this brief list demonstrates the importance of using integrated approaches with multiple different data to assess key eruption source parameters.

Mastin et al. (2009) considered that one of the most exact methods to determine the duration of a volcanic eruption, even more precise than the direct observation, is the seismic signal associated with it. Such a seismic signal is called tremor, which is characterized by having a long duration and being sustained in time, with variations in seismic amplitude and frequency depending on the changes in the eruption dynamics (Denlinger and Moran, 2014). For this reason, in this study, the calculation of the duration of all the eruptions was based mainly on their associated seismic signals and checked against photographic records when available.

On the other hand, meteorological information such as atmospheric profiles has recently become a key factor to determine eruptive plume parameters more precisely (Woodhouse et al., 2013). A few decades ago, these factors were neglected or simplified. Therefore, to obtain more realistic source parameters, it is mandatory to use such information and the models that incorporate the information (Kaminski et al., 2011; Degruyter and Bonadonna, 2012). According to the method of Kaminski et al. (2011), the height of the eruptive plume for plinian (subplinian to ultraplinian) eruptions can be calculated by using a model that includes the following: the atmospheric conditions (atmospheric stratification), the magma temperature, the mass flow emitted, an entrainment coefficient that is a function of the plume buoyancy and environment temperature, and a partition factor related to the efficiency of magma fragmentation. For transient vulcanian eruptions, the Druitt et al. (2002) and Bonadonna et al. (2002a) models treat the plume as a discrete thermal and suggest that its height (H) can be obtained from the mass (M) emitted instantaneously into the atmosphere instead of the mass flow rate.

NRV is located in the center of Colombia and is well known for the deadly phreato-magmatic eruption on November 13, 1985. That eruption was cataloged as VEI = 3 based on some unclear estimations of duration and column height available at that time (Naranjo et al., 1986). On September 1, 1989, a VEI = 2 phreato-magmatic eruption occurred. On May 29 and June 30, 2012, two small phreato-magmatic eruptions were recorded. From 1985–1991 to 2015–2017, continuous small (vulcanian) phreatic eruptions occurred. Estimation of the mass flow rate, among other parameters, for most of those volcanic plumes has not been performed yet. In this study, we combine seismic records of the eruptions with groundbased photographic images and integral equations (0D models) or scaling laws derived from 1D plume models to estimate the source parameters for Nevado del Ruiz Volcano eruptions. With these results, we expect to increase the knowledge of the dynamics of the volcanic plumes of the NRV and contribute to volcano monitoring and volcanic hazard assessment. In addition, we expect a better quantification of plume parameters to enable comparisons between them and seismic parameters. Considerations regarding the 2010–2017 reactivation of the system are also presented.

#### SUMMARY OF THE RECENT ACTIVITY OF NEVADO DEL RUIZ VOLCANO

The NRV is a stratovolcano located in center Colombia, with a summit height of approximately 5,311 m above sea level (a.s.l.). It is well known for the deadly and catastrophic phreatomagmatic eruption of November 13, 1985, which killed more than 20,000 people. In September 1989, another phreatomagmatic eruption occurred, with no victims. From 1986 to 1993, frequent small vulcanian eruptions were common at the NRV. In 2002 and 2003, an increase in seismicity and phreatic activity occurred. After almost 8 years of quiescence, it exhibited signs of reactivation in October 2010 (Londoño, 2016). In May and June 2012, two small phreatomagmatic eruptions occurred. After the end of 2014 and during 2015, continuous small vulcanian eruptions were common at the NRV. In 2015, a small dome was emplaced at the bottom of the active crater. This new activity is similar to that of the 1985–1991 period, with the difference of a dome building. Currently (Sept. 2018), the activity of the NRV continues at high levels.

New injection of magma in 2007 in the NRV zone was evidenced by changes in seismicity, deformation, and 3D seismic tomographic images (Londoño, 2016). The onset of the new activity was marked by increases in seismicity and SO<sup>2</sup> emission in October 2010, followed by a small ash emission. Furthermore, in 2011, a deep deformation source was detected far away to the SW of the volcano (Lundgren et al., 2015). In April 2012, strong shallow seismicity was detected to the SW of the active crater, and then on May 29 and June 30, two small phreatomagmatic eruptions occurred. The chemical composition of the ash was basically the same as the compositions of the ash from the 1985 to 1989 eruptions (Martinez et al., 2012). After those eruptions, the spatial distribution of volcano seismicity changed; several new volcano-tectonic (VT) seismogenic zones appeared, some of them located at the intersection of fault systems crossing the volcano. Between 2013 and 2014, some VT earthquakes were felt in Manizales City (30 km away from the NRV), some of them reaching local magnitudes (ML) of between 4.7 and 5.0. In November 2014, small vulcanian eruptions characterized by ash emissions were common. Since July 2015, some changes in seismicity and deformation have been observed near the active crater in addition to a number of small vulcanian eruptions, which became more continuous. In September of the same year, a small dome was emplaced at the bottom of the active crater and is still growing (June 2018). After mid-2017, decreases in seismicity and the number of vulcanian eruptions compared to previous years (2015–2016) were observed. Recently, some authors have suggested that during from 2015 to 2016, new batches of fresh magma were injected into the plumbing system of the NRV (Londoño and Kumagai, 2018).

#### METHODS

In this work, we used several scaling laws of volcanic plumes to estimate the source parameters for the activity of the NRV. For those eruptions with durations lasting <4 min (240 s), we used the models of Druitt et al. (2002) and Bonadonna et al. (2002a), since they are established for an instantaneous release of mass (Bonadonna et al., 2002a). For eruptions longer than 4 min, we applied the models that are established for a steady plume of Kaminski et al. (2011), Degruyter and Bonadonna (2012) and Woodhouse et al. (2013).

In the Druitt et al. (2002) and Bonadonna et al. (2002a) models, the thermal plume height (H in m) is

$$H = 1.89 \text{ J}^4 \text{ m} \left(\varphi Mc (T - To) \right)^{0.25},\tag{1}$$

where M is the mass of the solid material (kg), ϕ is the fraction of particles that contribute to the thermal mass of the plume (0.8), T-To is the difference in temperature between the atmosphere and the crater (To) and the emitted solid (T), and c is the specific heat of the solid (approximately 1,100 J kg−<sup>1</sup> K −1 ). From Equation (1), it is possible to calculate H (in m) for a vulcanian plume as follows:

$$H = BM^{0.25} + H\nu,\tag{2}$$

where Hv is the height of the crater (in m) and B = 55 m kg−<sup>4</sup> for Montserrat conditions (ϕ = 0.8; c = 1,204 J kg−<sup>1</sup> K −1 ; T-To=800 K). Since Montserrat volcano is similar in composition and eruptive style to the NRV, we used this formulation to model the NRV's vulcanian plumes. For the case of the NRV, an average of magma temperature of 1,173 K (Melson et al., 1990; Sigurdsson et al., 1990) and an average atmospheric temperature of 263 K based on meteorological data from IDEAM were used, given a value of B = 57 m J−<sup>1</sup> .

In the model of Kaminski et al. (2011), which is valid for subplinian to ultraplinian eruptions, there is no wind considered, but there is a variable entrainment coefficient, whereas in the other models mentioned above, the entrainment coefficient is constant (0.1). In this model, the plume height (H in m) is

$$H = 300 \text{ m s}^{4} \text{ kg}^{-4} \text{s}^{Q} \text{s}^{\frac{1}{4}},\tag{3}$$

for H < 12,000 m and

$$H = 5530 \,\mathrm{m} + 160 \,\mathrm{m} \,\mathrm{s}^{4} \,\mathrm{kg}^{-4} \,\mathrm{Q}^{1/4},\tag{4}$$

for H > 12,000 m,

where Q<sup>f</sup> is the effective mass flow of the plume (ash + gas) (kg/s), which is given by

$$Q\_f = \left[n\_o + \varphi \left(1 - n\_o\right)\right] Q\_o \tag{5}$$

where n<sup>o</sup> is the gas fraction in magma, which ranges from 3 to 7% for silica magmas (Kaminski et al., 2011), and Q<sup>o</sup> is the mass flow (kg/s), φ =fp/100, where f<sup>p</sup> is the partition factor, which depends on magma fragmentation (fp= 100% for plinian eruptions and fp<10% for effusive basaltic eruptions). For those eruptions in which portions of the emitted mass are not ejected into the atmosphere through the buoyant and ascending eruptive plume, but through pyroclastic flows or lavas, it is necessary to calculate the partition factor (see Equations 19–22 of Kaminski et al., 2011). From the effective mass flow (Q<sup>f</sup> ), it is possible to calculate the ash flow (Qash):

$$Q\_{ash} = \varphi \left(1 - n\_o\right) Q\_f,\tag{6}$$

where φ =fp/100, and Q<sup>f</sup> can be obtained from 1D models of eruptive plumes, such as

$$Q\_f = aH^4 + b,\tag{7}$$

where a and b are coefficients that depend on atmospheric conditions. According to the works regarding plumes and turbulent jets by Carazzo et al. (2008) for tropical zones, such as that found in the NRV area, a = 7 × 10−<sup>11</sup> kg s−<sup>1</sup> m−<sup>4</sup> and b=0 kg s−<sup>1</sup> for H<18,000 m, and a=2.78 × 10−<sup>11</sup> kg s−<sup>1</sup> m−<sup>4</sup> and b = 2.5 × 10<sup>7</sup> kg s−<sup>1</sup> for 18,000<H<25,000 m. The partition factor, fp, was taken to be 100% for all the eruptions of the NRV for the period from 1985 to 2017, except for the November 13, 1985, eruption, in which pyroclastic flows were generated (Calvache, 1990), corresponding to a f<sup>p</sup> value of 98%. This is the only eruption with evidence of pyroclastic flows at the NRV in the studied period.

Kaminski et al. (2011) established a functional model for ash flows >1 × 10<sup>5</sup> kg/s to determine the amount of ash at the top of the plume for a variable entrainment coefficient (see **Figure 3** of Kaminski et al., 2011). Due to the possibility that for the NRV, several small phreatic eruptions had ash flow lower than 1 × 10<sup>5</sup> kg/s, we obtained a fit to the curve in **Figure 3** of Kaminski et al. (2011) with a power law, considering ash flows (Qash) lower than that value (E. Kaminski, pers. com. 2017). Consequently, the ash load (L given in mg m−<sup>3</sup> ) for lower values of ash flows for the NRV was defined as

$$L = 396.45 \text{ mg m}^{-3} \times \ln(Q\_{abs}/Q\_o) - 1,550.7 \text{ mg m}^{-3}, \quad \text{(8)}$$

where <sup>Q</sup>o=1 kg s−<sup>1</sup> , a normalization constant.

An alternative model to calculate the mass flow of the volcanic plume is the model of Degruyter and Bonadonna (2012), which considers the local atmospheric conditions instead of a general atmospheric model, that is, it considers the wind conditions at the moment of the eruption. In that model, the mass flow is defined as

$$Q\_o = \pi \frac{\rho\_{a0}}{g'} \left( \frac{2^{5/2} \alpha^2 \overline{N}^3}{z\_1^4} H^4 + \frac{\beta^2 \overline{N}^2 \overline{\nu}}{6} H^3 \right), \tag{9}$$

where

$$\mathcal{g}' = \mathcal{g}\left(\frac{\varepsilon T - \varepsilon\_{a0}\theta\_{a0}}{\varepsilon\_{a0}\theta\_{a0}}\right);\ \overline{N^2} = \frac{1}{H}\int\_{\rho}^{H} N^2\left(z\right)dz;$$

$$N^2\left(z\right) = \frac{\mathcal{g}^2}{C\_{a0}\theta\_{a0}}\left(1 + \frac{c\_{a0}}{\mathcal{g}}\frac{d\theta\_a(z)}{dz}\right),\tag{10}$$

$$
\overline{\nu} = \frac{1}{H} \int\_0^H \nu(z) \, dz,\tag{11}
$$

where ρa<sup>0</sup> is the reference density of the surrounding atmosphere, α is the radial entrainment coefficient (α = 0.1), β is the wind entrainment coefficient (β = 0.5), ca<sup>0</sup> is the heat capacity of the surrounding atmosphere, θa<sup>0</sup> is the temperature of the surrounding atmosphere, g is the gravitational acceleration, g' depends on g and is computed via Equation (10), θa(z) is a profile of environment temperature, v(z) is a wind profile, N is the average buoyancy frequency of the atmosphere, v is the wind velocity across the plume height, z<sup>1</sup> is the maximum nondimensional height, and z is the vertical coordinate above the source.

A model similar to that of Degruyter and Bonadonna (2012) is the model of Woodhouse et al. (2013), in which the mass flow is given by

$$Q\_o = \left[\frac{H}{0.318\left(\frac{1+1.373\tilde{W}\_s}{1+4.266\tilde{W}\_s+0.3527\tilde{W}\_s^2}\right)}\right]^{3.95},\tag{12}$$

where <sup>W</sup>˜ <sup>s</sup> = 1.44V1/NH1, V<sup>1</sup> is the wind velocity at the tropopause, H<sup>1</sup> is the local height of the tropopause (in m), and N is the atmospheric buoyancy frequency. Whereas, the model of Degruyter and Bonadonna (2012) considers the plume ascending in a calm atmosphere and bending over when it finds a strong wind field, the model of Woodhouse et al. (2013) considers the plume ascending in a linear shear crosswind in an intermediate regime, where the ascending plume speed and the wind speed are similar. For the NRV, we estimated the height of the tropopause and stratosphere as 12 and 20 km above the top of the volcano, respectively, from atmospheric models and radiosonde data from NOAA.

We calculated the density of the mixture contained in the volcanic plume (ρP) using the following formulation (Woods, 1988; Mastin, 2007):

$$\rho\_P = \left[ n\_o \rho\_\mathcal{g}^{-1} + \frac{(1 - n\_0)}{\rho\_m} \right]^{-1},\tag{13}$$

where ρ<sup>m</sup> is the magma density (ash) and ρ<sup>g</sup> is the gas density (vapor = 0.2 kg/m<sup>3</sup> at 5,500 m a.s.l.).

In the previous 0D models, H refers to the height of the centerline of the plume, which is not identical to the height of the plume top in a bent-over situation. The 0D model of Mastin et al. (2009) established the column height based on an average eruption rate for different eruptions around the world. The column height (H in m) was defined as

$$H = 2\mathbf{x}10^3 \text{ V}^{0.241},\tag{14}$$

where V is the volumetric flow rate (m<sup>3</sup> of dense rock equivalent, DRE per second). In this model, both the column height at the top of the plume and column height at the center of the umbrella were considered (Mastin et al., 2009).

The radiated seismic energy (RSE) for the seismic signals (tremor) associated with the eruptions can be calculated by using the seismic power (Dibble, 1974; Cristofolini et al., 1987; Alparone et al., 2003):

$$RSE = \left(\pi\rho\_s\nu\_s(r \ge 10^{-2})\right)^2 A^2 t,\tag{15}$$

where ρ<sup>s</sup> is the density of the upper part of the volcanic edifice (2,600 kg m−<sup>3</sup> , Londoño et al., 2014), v<sup>s</sup> is the average seismic velocity of the shallowest layer (2,500 m s−<sup>1</sup> , Londoño and Sudo, 2002), r is the source-station distance (in cm, assuming the active crater as the source), A is the average particle velocity (amplitude) filtered at the predominant frequency of the volcanic tremor, and t is the duration of the eruption.

Moreover, the instantaneous reduced displacement (RD, Fehler, 1983) for surface waves (Denlinger and Moran, 2014) for the volcanic tremor associated with the eruptions can be calculated using the following expression:

$$RD = \frac{A\_{pp}\sqrt{\lambda \cdot r}}{2\sqrt{2}M},\tag{16}$$

where App is the maximum amplitude peak-to-peak (in cm) of the tremor filtered at the dominant frequency f (in Hz), λ is the wavelength for surface waves (in cm; λ =vr/f, where v<sup>r</sup> is the average velocity of surface waves=2.16 km s−<sup>1</sup> ), and M is the magnification of the instrument. The factor 2<sup>√</sup> 2 is the correction of the mean square root of the amplitude. **Table 1** presents a list of the variables used in this study.

#### DATA AND PROCESSING

The seismic, photographic, and SO<sup>2</sup> data used in this work belong to Servicio Geológico Colombiano (SGC). We used the seismic signals associated with the eruptions for the period from 1985 to 2017. Unfortunately, not all the eruptions before 2012 have photographic records, but most of them have data of direct visual observation obtained by Volcanological Observatory of Manizales from SGC. After 2012, most of them have photographic records. We used meteorological data regarding the studied zone belonging to IDEAM (the Colombia meteorological institute) and the NOAA Satellite and Information Service from USA. Additionally, we used data for some eruptive plumes height from Washington VAAC (Volcano Ash Advisory Code).

We used the maximum wind speed in the tropopause for the NRV by using the radiosonde located at Bogota City, 130 km to the SE of the crater. In a few cases, we used data from radiosondes located in Panamá, Ecuador or Curacao. We also calculated temperature gradients for the stratosphere of the NRV area.

For the period from 1985 to 1991, we analyzed only the larger eruptions, since there is partial information about column height and seismic records for many small phreatic eruptions that occurred. Nevertheless, we consider that the larger eruptions allow us to have an idea about the mass and emitted volumes between 1985 and 1991, which we estimated to account for approximately 85 to 90% of the total mass. We analyzed the volcanic plumes of the eruptions that occurred on September 11, 1985, November 13, 1985, January 4, 1986, May 4, 1986, July 20 and 29, 1986, June 10, 1987, March 25, 1988, September 1, 1989, and Apr 9, 26 and 29, 1991. Between 1992 and 2011, no eruptions were detected at the NRV. For the period from 2012 to 2017, a complete dataset is available, and almost all of the eruptions were analyzed.

Four hundred and twenty eruptive plumes were analyzed (see **Appendix A**). Eruptions that occurred during nighttime were not considered or included in this study; fortunately, these were few, and all of them were of small size and duration. Therefore, the results are not affected very much by this exclusion.

To measure the volcanic plume heights from the groundbased photographic images, we established a calibration with several places of known height observed in the images, taking the average height from available images of different photographic cameras for each eruption. **Figure 1** shows an image of the location of the NRV and some examples of eruptive plumes. The error in the estimation of plume height was up to 10% (several 100 m); consequently, a higher error will occur for the parameters of the eruptive plume, such as the mass flow rate, mass, and volume. For each plume dataset, the source parameters were calculated using the different volcanic plume scaling laws mentioned above. Moreover, the values of some variables used in those models (**Table 1**) yield additional uncertainties of 10%; for instance, temperature can be affected by the presence of water for phreatic and phreato-magmatic eruptions. Thus, the total uncertainty of the source parameters calculated in this work can be estimated to be approximately 30% or even greater. Nevertheless, we are interested in the systematic evolution of the system as a function of time, which makes the absolute determination of the parameters less relevant.

To measure the duration of the eruption, we used seismic records of the eruption. The duration of the eruption was constrained by both the seismic signal associated with the eruption and the synchronized photographic images available. The decay of the amplitude of the seismic signal was the main parameter used to define the end of the eruption, which in most of the cases was consistent with the stopping of ash emission observed in the photographic cameras. In a few cases, the end of the emission could not be observed directly from the cameras, but the seismic record was still available (see the example in **Appendix B**).

Additionally, the radiated seismic energy (RSE) and reduced displacement (RD) were calculated for the seismic signals associated with the eruptions. **Figure 2** shows an example of a seismic signal associated with a small eruption and the parameters used to calculate the RD and RSE.

#### RESULTS

#### Plume Modeling

For the eruption of November 13, 1985, Naranjo et al. (1986) estimated a column height of 31 km, a duration of 20 min and a volume of dense material of 0.039 km<sup>3</sup> based on modeling

#### TABLE 1 | Description of the variables used in this study.


of the column height (Carey et al., 1986) and on eyewitness reports. We revisited that calculation and found that the data can be more precisely estimated. First, the duration of the eruption was corrected. Naranjo et al. (1986) estimated the duration of the eruption based on eyewitness observations, but the eruption occurred at nighttime, which makes the eyewitness technique not fully reliable. On the other hand, the signal recorded at several seismic stations around the volcano was used to calculate the duration of the eruption more accurately (**Appendix C**). By analyzing the recent high-resolution digitized analog seismic records available at SGC for the November 13, 1985, eruption carefully, we estimated that the eruption main pulse should have

FIGURE 1 | (A) The location of the NRV and seismic (circles), camera (rectangles) and SO2 (triangles) stations used for the analysis. (B) Examples of seismic records (digitized analog and digital) of some eruptions at the NRV (see Appendix C for details about the seismic signals). (C) Examples of typical eruptive plumes for the period from 1985 to 2017 (photographic images from the SGC database).

FIGURE 2 | Example of parameters used for calculation of DR and RSE from seismic signals associated with eruptions. Amplitude values are in nm/s already corrected by instrumental response (name of variables are listed in Table 1).

lasted between 90 and 105 min. After the main pulse, smaller pulses occurred and were recorded seismically. However, we only focus on the main phase of the eruption. Regarding the plume height, according to the work of Krueger et al. (1990) based on images of SO<sup>2</sup> from the Total Ozone Mapping Spectrometer (TOMS) instrument on the Nimbus 7 polar orbiting satellite available for the November 13 eruption, the estimated maximum column height reached up to 25 km and the main portion of the plume was between 7 and 14 km. If we remove the volcano height (approximately 5.3 km), we obtain a maximum eruptive plume height of approximately 20 km above the top of the volcano, which is at least 10 km less than the Naranjo et al. (1986) estimation, which was based on modeling. Krueger et al. (1990) used satellite images, which were possibly not available when Naranjo et al. (1986) and Carey et al. (1986) calculated the column height. Additionally, we estimated an atmospheric profile for the NRV region for the November 13, 1985, eruption based on data from NOAA from the Panamá radiosonde. We obtained a maximum wind speed at the troposphere for the Colombia, Panamá and Venezuela region (the region of plume dispersion) between 15 and 20 m/s and a temperature gradient of −2 ◦C/km for the stratosphere. The heights of the tropopause and stratosphere were estimated to be 12 km and 20 km above the top of the volcano, respectively.

With these new data, the estimated flow rate ranged from 3.55 ×5 10<sup>7</sup> ± 1.06 × 10<sup>7</sup> kg/s according to the Mastin et al. (2009) scaling law to 6.69 × 10<sup>7</sup> ± 2.01 × 10<sup>7</sup> kg/s according to the Kaminski et al. (2011) scaling law. Therefore, the estimated mass emitted by the November 13, 1985, eruption was 3.6 × 10<sup>11</sup> ± 10<sup>11</sup> kg, 2.9 × 10<sup>11</sup> ± 8.9 × 10<sup>10</sup> kg, 2.9 × 10<sup>11</sup> ± 8.7 × 10<sup>10</sup> kg, and 5.7 × 10<sup>10</sup> ± 1.7 × 10<sup>10</sup> kg according to the models of Kaminski et al. (2011), Degruyter and Bonadonna (2012), Woodhouse et al. (2013), and Mastin et al. (2009), respectively.

with duration < 240 s (yellow circles). Y-axes in log scale. Error bars are included.

For the September 1, 1985, eruption, we proceed in a similar manner. Méndez and Valencia (1991) calculated a volume of 1.6 × 10−<sup>3</sup> km<sup>3</sup> and a plume height of 6 to 8 km above the vent by using isopach data, in addition to a duration of 2 h 24 min. Based on analyzing in detail the digitized high-resolution analog seismic signal that was recently made available, the eruption lasted approximately 90 min for the main phase and approximately 6 h in total (**Appendix C**). We obtained an atmospheric profile for that day from radiosonde data from the BOGOTA station of the NOAA, and we computed a maximum wind speed in the troposphere of 9.7 km/s and a temperature gradient of −2 ◦C/km for the stratosphere. The height of tropopause and stratosphere was estimated to be 12 and 20 km above the top of the volcano, respectively. We assumed a plume height of 8 km (Méndez and Valencia, 1991).

With the new data, the September 1, 1989, eruption emitted a mass of 2.7 × 10<sup>9</sup> ± 8.2 × 10<sup>8</sup> kg, 3.8 × 10<sup>9</sup> ± 1.2 × 10<sup>9</sup> kg, 5.2 × 10<sup>9</sup> ± 1.5 × 10<sup>9</sup> kg, and 4.3 × 10<sup>9</sup> ± 1.3 × 10<sup>9</sup> kg according to the models of Kaminski et al. (2011), Degruyter and Bonadonna (2012), Woodhouse et al. (2013), and Mastin et al. (2009), respectively.

For the minor eruptions, we proceeded in a similar manner, also. **Figures 3**, **4** show the mass flow rate (kg/s) and volume (m<sup>3</sup> ), respectively, of eruptive plumes of the NRV for the period from 1985 to 2017 using different models. **Figure 5** shows in detail those parameters for the period from 2015 to 2017. As it can be observed from **Figure 3**, the flow rate was variable during the studied period, with the highest values in 1985 and decreasing over time. A similar tendency was observed for the volume. For the period from 2015 to 2017 (**Figure 5**), those parameters exhibited a variable tendency, increasing and decreasing randomly. The total mass for the period from 2010 to 2017 for thermals (duration <4 min) was 1.6 × 10<sup>8</sup> ± 4.8 × 10<sup>7</sup> kg, and for the other plumes, it ranged from 2.5 × 10<sup>8</sup> kg to 5.39 × 10<sup>8</sup> kg. The total plume volume was 1.8 × 10<sup>5</sup> m<sup>3</sup> , using a mixture density of 3.3 kg/m<sup>3</sup> (Ripepe et al., 2013), average air density of 0.7 kg/m<sup>3</sup> at a height of 6 km, gas fraction between 3 and 6% for dacitic-andesitic rocks, and density of solid rock (without pores) between 2,400 and 2,600 kg/m<sup>3</sup> for the NRV ash (Londoño et al., 2014; L. Martínez pers. com., 2017). **Table 2** presents a summary of the source parameters obtained for the most relevant eruptions for the period from 1985 to 2017 for the NRV using different scaling laws.

Additionally, we estimated the ash load (mg/m<sup>3</sup> ) at the top of the eruptive plume (Kaminski et al., 2011) for the period from 1985 to 2017 using Equation (8). **Figure 6** shows the results. In general, the ash load was bigger for the November 13, 1985, eruption (3,587 mg/m<sup>3</sup> ). For the other eruptions, the ash load values ranged from approximately 500–3,400 mg/m<sup>3</sup> .

#### Relation Between Seismic Data and Eruption Plume Parameters

**Figure 7** shows a plot of H vs. RD and H vs. RSE and a contour plot of H as a function of RD and RSE for the period from 2015 to 2017. In general, there was no correlation between H and RD or between H RSE for small values of H (< 3 km). In contrast, values of H >5 km corresponded to the highest values of RSE and RD, although they were not plotted. For small plumes (H <3 km) with duration >240 s, there is a tendency of H as a function of RD and RSE (**Figure 7C**), that is, higher values of RD and RSE corresponded to high values of H, although there were several eruptions with H >1.5 km with relatively low RSE values.

**Figures 8**, **9** show details of the time series of volcanic plume parameters vs. RSE and RD, respectively, of tremors for the period from 2015 to 2017. From these figures, it is possible to observe that, in general, RSE and RD did not exhibit any clear

relationship with plume parameters during that period. On the other hand, almost all the higher values of RD were registered during 2015, whereas higher values of RSE were distributed from 2015 to 2016.

We related the RSE with RD and the duration of the eruption with respect to the date of the eruption (**Figure 10**). There was a different pattern distribution of RSE and DR with respect to the duration of the eruption, depending on the date. On the other hand, the higher values of RD corresponded to the shortest durations of eruptions, which is not useful to forecast eruptions or volcanic behavior. This is a first indication that RSE is the preferable parameter when quantifying the volcanic tremor and its plume parameters compared to RD (see below).

### DISCUSSION

According to Degruyter and Bonadonna (2012) and Woodhouse et al. (2013), if wind conditions are not accurately estimated or are neglected in the modeling of volcanic plumes, the mass flow rate can be underestimated by up to an order of magnitude. On the other hand, the duration of an eruption is another key parameter that must be determined as accurately as possible, since mass and volume parameters depend on it. Seismic records are a powerful tool to obtain eruption durations accurately (Mastin et al., 2009). In this study, we have used wind profiles and photographic and seismic records to calculate the eruptive plume parameters for the most recent period of activity of the NRV (1985–2017). It is noteworthy to mention that temperature is another factor that affects the eruption mass flow rate, as we pointed out previously. For phreatomagmatic pulses, it is possible that the magmatic temperature is less than that used in this study, due to presence of water; in addition, other factors, such as the initial thermal energy and mass of surface water, can be difficult to model for this type of eruption (Koyaguchi and Woods, 1996); therefore, we assume that there is another source of overestimation of such source parameters in our data not considered in this study.

With this in mind, the mass values obtained for the November 13, 1985, eruption imply that the volume of DRE was between


For the period from 2010 to 2017, minimum and maximum values are reported.

0.07 and 0.11 km<sup>3</sup> when using a density of solid rock of 2,500 kg/m<sup>3</sup> for the NRV (Melson et al., 1990). If we add the volume of the pyroclastic flows (0.009 km<sup>3</sup> , Calvache, 1990), the total volume of the November 13, 1985, eruption was at least 0.12 km<sup>3</sup> . This value is greater than that previously calculated by Naranjo et al. (1986), which was 0.03 km<sup>3</sup> and that calculated by Calvache (1990), which was 0.02 km<sup>3</sup> . With the new data regarding the erupted volume, the November 13, 1985, eruption reaches a VEI of 4 (the lower limit of VEI = 4).

For the September 1, 1989, eruption, if we assume a density of solid rock of 2,500 kg/m<sup>3</sup> (according to the compositional results of Méndez and Valencia, 1991), the volume of dense rock was 1.4 × 10−<sup>3</sup> ± 1.0 × 10−<sup>4</sup> according to the model of Degruyter and Bonadonna (2012). This volume agrees with that obtained by Méndez and Valencia (1991), who used isopach data.

Additionally, it is possible to establish an empirical relationship between RD and RSE with the ash load (L) at the top of the plume. If we neglect those volcanic plumes with H <1 km, that is, the smallest size eruptions with low ash load values, we can construct a fit to obtain a minimum ash load (Lmin) value (mg/m<sup>3</sup> ), knowing the RD (in cm<sup>2</sup> ) or RSE (in Joules) of the volcanic tremor signal associated with the eruptive column, as follows:

Lmin = 68.575mg/cm<sup>3</sup> cm−2×RDcm<sup>2</sup> <sup>+</sup> 394.4mg/cm<sup>3</sup> (17) <sup>L</sup>min <sup>=</sup> <sup>5</sup>×10−4mg/cm<sup>3</sup> cm−2×RSEcm<sup>2</sup> <sup>+</sup> 366.5mg/cm<sup>3</sup> (18)

We choose to fit the minimum value instead of, for instance, an average, since we obtained a wide range of ash load values for similar RD or RSE. In this sense, the minimum value of ash load represents the lower limit of ash load we can obtain for a plume height, although it is possible to obtain higher values for the same plume height, up to the limit of theoretical values by Kaminski et al. (2011). **Figure 11** shows this relationship and the fitted curves.

As it can be observed from **Figure 11**, there is a minimum limit on the ash load depending on the value of RD or RSE; that is, the higher the RD or RSE value, the higher the minimum value of the ash load. Although it is possible to obtain a wide range of ash load values for the same value of RD or RSE, it appears that there is a lower limit on ash load for that value. For instance, for an RSE value of 1 × 10<sup>6</sup> J, the minimum ash load value will always be greater than approximately 850 mg/m<sup>3</sup> . The same relation is valid for RD; for instance, a RD value of 6 cm<sup>2</sup> corresponds to a minimum value of approximately 750 mg/m<sup>3</sup> , and not less than that value. This finding is very interesting for volcano monitoring and forecasting the minimum ash load that an eruption of the NRV will contain based on the seismic signal only.

Moreover, there is a temporal variation of the RSE and RD of volcanic tremors related with the volume of ash for the period from 2015 to 2017. **Figure 12** shows the comparison. From this figure, it is possible to observe three different stages or changes in RD and RSE with the cumulative volume of ash. The first stage (I in **Figure 12**) from March to the end of August 2015 exhibited a regular volume emission of ash, whereas RSE presented increases, and RD was steady, suggesting a semisealed magmatic system interacting with the hydrothermal system, partially blocking the output of solid material to the atmosphere. The second stage (II in **Figure 12**) from the end of August to the beginning of October 2015 exhibited an important increase in ash volume emission associated with a concurrent increase in RD and RSE, interpreted as a less sealed but pressurized magmatic system as a response of a dome emplacement at the crater bottom during September 2015 (SGC, 2015). The third stage (III in **Figure 12**) from March 2016 to the end of October 2016 exhibited an important increase in ash volume emission, whereas RD and RSE remained relatively steady, although RD exhibited a slight increase, implying a more open magmatic system, allowing solid material to be output freely, with a low amount of seismic energy needed to expel it. These stages seem to reasonably explain the current activity of the NRV (July 2018), which is characterized by low ash emissions associated with low energy seismic signal, whereas the dome is growing slowly, indicating that the volcanic system is almost open, allowing the ascent of solid material easily. These findings have some important implications for risk assessment: very strong eruptions, larger than the one of November 13, 1985, at the NRV, would probably require a drastic change in the conduits, such as blocking or pressurization. Such an imminent event should then be manifested in a significant increase in seismicity at the crater, in addition to deformation signals. On the other hand, the currently open condition of the volcanic system bears the possibility that volcanic eruptions of smaller or medium strength can occur without any considerable changes in seismicity or deformation due to the current open condition of the volcanic system. Therefore, it is mandatory to continue monitoring the activity of the NRV with uttermost vigilance and precision.

On the other hand, the SO<sup>2</sup> flux is one of the most intriguing parameters observed at the NRV over time (Williams et al., 1990). With the aim to observe any relationship of source parameters and SO<sup>2</sup> release, we compared the mass flow and mass of eruptive plumes with those of SO<sup>2</sup> for the period 2015–2017, which was a period with continuous SO<sup>2</sup> measurements with DOAS

instruments (**Figure 1**). **Figure 13** shows this comparison. The SO<sup>2</sup> flux sometimes increased when the mass flow increased, but in other cases exhibited a contrary tendency. A decrease in SO<sup>2</sup> flux in December 2015 does not correspond to any eruption with H >2 km. In contrast, during May and June 2016, an increase in SO<sup>2</sup> was not associated with any eruption, likely suggesting passive degassing in the NRV during that time, that is, release of large amounts of SO<sup>2</sup> without an eruption.

Moreover, with the new calculated volume for the November 13, 1985, eruption, it is possible to revisit the question of the discrepancy between the SO<sup>2</sup> excess released to the atmosphere and the volume of the eruption. According to Krueger et al.

vs. eruptive plume parameters for the NRV for the period from 2015 to 2017. Different ranges of RSE are represented by colored squares. In the plume heights panel, all available data were plotted; for the rest of parameters, only eruptions with H >1 km and duration >240 s were plotted.

(1990), the amount of SO<sup>2</sup> emitted by the eruption was 7x10<sup>8</sup> kg; according to Williams et al. (1990), the magma volume needed to explain such SO<sup>2</sup> is approximately 0.92 km<sup>3</sup> , whereas Sigurdsson et al. (1990) estimated 0.3 km<sup>3</sup> . The new volume for this eruption, as mentioned previously, was 0.12 km<sup>3</sup> , approximately seven

times less than expected according Williams et al. (1990) and only 2.5 times less than the value calculated by Sigurdsson et al. (1990). This discrepancy is less than the factor of 7–30 hitherto accepted and calculated by Naranjo et al. (1986) and Calvache (1990). Although the new volume still is less than that needed to explain

all the SO<sup>2</sup> released, it is in the range of the magma bodies size beneath the NRV derived from seismic velocity anomalies from a 3D seismic tomography (Londoño and Sudo, 2002; Londoño and Kumagai, 2018). According to those seismic tomographic results, it is possible to estimate a shallow magma body (2–3 km depth) of approximately 10 km<sup>3</sup> and another deep magma body (6–8 km depth) of approximately 50 km<sup>3</sup> for 1985. In addition, to explain the constant degassing at the NRV, Williams et al. (1990) suggested a minimum magma volume between 4.6 and 9.2 km<sup>3</sup> beneath the volcano, values that are in agreement with those calculated using seismic tomography for the shallow magma body. If we consider all the SO<sup>2</sup> released to the atmosphere by the NRV from 1985 to 2017, which was approximately 1.5 × 10<sup>10</sup> kg, we can envision one or several magmatic reservoirs of

approximately 60 km<sup>3</sup> in total, in agreement with the estimated volume (summing both magmatic bodies) obtained from seismic tomography. With this new estimated volume, the ratio between the SO<sup>2</sup> emitted to the atmosphere and the ejected magmatic material is 5.1 × 10−<sup>4</sup> , which is two orders of magnitude less than that calculated by Williams et al. (1986). This ratio is similar to or smaller than those calculated for other volcanoes, such as St. Helens (USA) and El Chichón (Mexico) (Williams et al., 1986). On the other hand, if we compare the SO<sup>2</sup> released and the erupted magma for the November 13, 1985, eruption to other volcanoes according to the work of Wallace (2001), the new relation for the NRV fits better in the range of andesitic magma (**Figure 14**). Previously, these results corresponded to the basaltic magma range, which disagrees with the observed products in

fitting curve for the minimum ash load value with respect to the RD and RSE.

the field, classified as andesites and dacites (Melson et al., 1990; Sigurdsson et al., 1990).

It has been argued that the November 13, 1985, eruption of the NRV was modest (Giggenbach et al., 1990; Krueger et al., 1990; Williams et al., 1990), but it is possible that it was not as small as it has been assumed. Several facts support this idea. First, the meteorological conditions that day were not favorable to remotely observe the eruptive plume by weather satellites, and there are not available satellite images for the NRV region at the moment of the eruption according to a search for GOES-6, METEOSAT-2 or GMS-3 satellite images using the National Centers for Environmental Information NOAA web browser.

(https://www.ncdc.noaa.gov/has/HAS.FileAppRouter? datasetname=3645&subqueryby=STATION&applname=& outdest=FILE); Second, SO<sup>2</sup> plumes were observed remotely 15 h later (Krueger et al., 1990) by dedicated satellites that can detect particles of fine ash not observed by ground-based radar and eyes. The SO<sup>2</sup> plume was detected more than 1,000 km away from the volcano to the Atlantic Ocean, suggesting a coexisting very large amount of probably fine ash ejected to the atmosphere, but not deposited close to the volcano. Third, three decades after the eruption we obtained information about a volcanic ash layer of that eruption at least one or two mm thick as far as 215 km far from the volcano (geologist Italo Reyes eyewitness and photographic record, SGC photographic database) in the Belencito municipality in Boyacá Department (120 km to NE of Bogotá), suggesting that the isopachs were probably much more extended than previously calculated by Naranjo et al. (1986). Fourth, the volume of pyroclastic current density (PCD) products could not be accurately estimated, since much of that material was incorporated into the lahar, as pointed out by Calvache (1990). In addition to these facts, the reevaluation of the duration of the eruption performed in this work leads us to conclude that the eruption was at least one order of magnitude larger than previously determined. It is possible that much of the ejected material was incorporated into the lahars (coarse material) or deposited far away (fine material) up to the Atlantic Ocean, it being impossible to realistically estimate the true ejected volume of magmatic material. The newly available data help to constrain the real size of this eruption with a higher precision.

The estimated volume in this work for the September 1, 1989, eruption is in agreement with that obtained by Méndez and Valencia (1991). For this eruption, more data were available, including geological (petrographic, stratigraphic), geophysical (seismic) and observational data, supporting the idea that the combination of different datasets leads to more realistic and consistent volcanic plume parameters.

The estimated volume for the small eruptions of May 29 and June 30, 2012 exhibited a discrepancy with that calculated by Martinez et al. (2012) using an isopachs approach. For the May 29 and June 30, 2012 eruptions, the ash volumes calculated by Martinez et al. (2012) were 1.59 × 10<sup>6</sup> m<sup>3</sup> and 5.83 × 10<sup>4</sup> m3 , respectively, whereas in this study, we calculated a volume between 7.6 × 10<sup>4</sup> (with the model of Mastin et al., 2009) and 1.4 × 10<sup>5</sup> m<sup>3</sup> (with the model of Woodhouse et al., 2013) for May 29 and between 2.2 × 10<sup>5</sup> m<sup>3</sup> (with the model of Mastin

et al., 2009) and 3.1 × 10<sup>5</sup> m<sup>3</sup> (with the model of Woodhouse et al., 2013) for the June 30, 2012 eruptions. This difference may be due to different facts. First, the estimated plume height for these eruptions could be affected by a bent-over situation due to wind at the moment of the eruption; therefore, it is possible that the volume estimated can be biased by the uncertainty in the plume height. For the May 29 eruption, the wind speed at 10 km a.s.l. was 18 m/s, whereas for June 30, it was 24 m/s, according to IDEAM meteorological reports. With this source of error for column height in mind, according to the model of Mastin et al. (2009), the volume of the May 29 eruption was half an order of magnitude less than that of 30 June 2012. It is possible that the volume calculated using the isopachs approach, which uses the plume height as the umbrella-cloud height, was underestimated for the June 30 eruption. Based on the available data and reports, the volcanic plume height of the June 30 eruption was higher than that of May 29; according to the VAAC homepage (Volcanic Ash Advisory, 2010; http://www. ssd.noaa.gov/VAAC/ARCH12/RUIZ/2012E291325.html) and a report from an airplane pilot, the plume height of the May 29 eruption (3:10 am local time) was 5.7 km above the top of the volcano. According to photographic images of SGC, the column height of the 30 June eruption (5:47 pm local time) was at least 7 km above the top of the volcano. Second, the seismic signal of the June 30 eruption was longer than that of May 29. Based on these facts, we conclude that the June 30 eruption was larger than the May 29 eruption. The differences in tephra dispersion may be explained by two facts: first, heavy rain was falling all night long on June 30 in a wide region of the Caldas, Risaralda and Quindío Departments (central Colombia), probably removing part of the fine ash fall in several places, leading to underestimate the tephra dispersion for that eruption (Martinez et al., 2012). Second, it is possible that during the May 29 eruption, the conduit was more sealed than June 30, which involved more pressure and gas, ejecting fine ash over larger distances than the June 30 eruption. The difference in the seismic amplitudes of the signals supports this conjecture; although the June 30 eruption long lasted more than May 29, the latter exhibited a higher

instantaneous seismic amplitude (higher RD) than the former (**Appendix B**).

A continuous decrease in the volume of ash plumes of NRV of about two orders of magnitude from the November 13, 1985 eruption to the September 1, 1989 eruption, and of about 5 orders of magnitude from 1989 to 2012, suggests that a volcanic cycle is ending. Moreover, the correlations between RSE and RD of the volcanic tremor and the ash load, as well as a correlation between RSE, RD and the ash volume for the studied period, allow us to divide the activity of NRV in different stages. These findings could be used to better understand the behavior of NRV in the future.

Our study shows that a holistic analysis of datasets from multiple different sources, such as records of gas, seismicity, and local meteorological and atmospheric conditions, in combination with photographic images and scaling laws based on physical plume models, leads to a significantly improved assessment of eruption source parameters, hence being a promising tool for volcanic monitoring and risk assessment, not only for NRV but also to other active volcanoes.

#### CONCLUDING REMARKS

Volcanic plume parameters such as the mass flow rate, erupted mass, ash flow rate, and volume were calculated for the most recent eruptive period of the NRV, 1985–2017, by integrating different datasets and using different integral volcanic plume models. These new data suggest that the eruption was larger than previously determined. With these new data, the November 13, 1985, eruption had a VEI=4, with a volume of 0.12 km<sup>3</sup> . The eruption lasted longer than initially assumed, based on a detailed analysis of seismic signals associated with the eruption. Moreover, it is possible to establish a relation between the RSE of volcanic tremors and the ash load at the top of the eruptive plume, which allows the RSE of volcanic tremors to be used to provide advice about the risk of volcanic ash for aviation, helping in the volcanic risk assessment of this region.

In addition, the ongoing reactivation of the NRV started in October 2010 and corresponds to the same eruptive cycle of 1985. It is likely that the same magmatic body is acting and releasing large amounts of SO<sup>2</sup> to the atmosphere, in addition to volcanic ash, which has been gradually decreasing over time, suggesting that this cycle is ending. On the other hand, it is argued that currently, a new deep magmatic body is intruding to the S of the NRV (15–18 km, Lundgren et al., 2015), which probably started to move up in 2007 (Londoño, 2016) and was recently highlighted at the surface by the emplacement of a small dome at the bottom of the active crater. This phenomenon indicates that the NRV still is a very active volcano, with the possibility of new eruptions in the near future. Knowing of the current behavior of a volcano by means of the dynamics of the volcanic plume and its

#### REFERENCES


relationship with the seismicity is a useful approach to elucidate future behavior of that volcano. In this study that approach was used for NRV and hopefully it can be applied to other volcanoes.

#### AUTHOR CONTRIBUTIONS

JL had the main idea, wrote portions of the manuscript and processed SO<sup>2</sup> and plume parameters. BG processed the seismic data, collected height plumes from digital images and wrote some portions of the manuscript.

#### FUNDING

This work was partially supported by Universidad Católica de Manizales, Servicio Geológico Colombiano, and Universidad de Caldas, Manizales.

#### ACKNOWLEDGMENTS

We want to thanks to C. Bonadonna and W. Degruyter for their help and for providing computer codes for calculation of volcanic plume parameters. We also thank C. Bonadonna for her careful review of and comments on this paper, E. Kaminski for his suggestions and comments about his model and application to small volcanic plumes, and our colleagues at Servicio Geológico Colombiano (SGC) at the Volcanological and Seismological Observatory of Manizales for their support and suggestions. This work was partially supported by the SGC Direction of Geohazard, Universidad de Caldas and Universidad Católica de Manizales. Luca de Siena, E. Kaminski, and another reviewer, in addition to the associate editor, Yosuke Aoki, and the chief editor, Valerio Acocella, made important suggestions and comments that considerably improved the final manuscript.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart. 2018.00162/full#supplementary-material


at Santiaguito volcano, Guatemala, from infrasound waveform inversion and thermal infrared measurements. Geophys. Res. Lett. 43, 6220–6227. doi: 10.1002/2016GL069098


**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 Londono and Galvis. 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.

## Bayesian Network Modeling and Expert Elicitation for Probabilistic Eruption Forecasting: Pilot Study for Whakaari/White Island, New Zealand

Annemarie Christophersen<sup>1</sup> \*, Natalia I. Deligne<sup>1</sup> , Anca M. Hanea<sup>2</sup> , Lauriane Chardot <sup>3</sup> , Nicolas Fournier <sup>4</sup> and Willy P. Aspinall 5,6

*<sup>1</sup> GNS Science, Avalon, Lower Hutt, New Zealand, <sup>2</sup> Centre of Excellence for Biosecurity Risk Analysis, The University of Melbourne, Melbourne, VIC, Australia, <sup>3</sup> Earth Observatory of Singapore, Nanyang Technological Institute, Singapore, Singapore, <sup>4</sup> Wairakei Research Centre, GNS Science, Wairakei, New Zealand, <sup>5</sup> School of Earth Sciences and Cabot Institute, University of Bristol, Bristol, United Kingdom, <sup>6</sup> Aspinall & Associates, Tisbury, United Kingdom*

#### Edited by:

*John Stix, McGill University, Canada*

#### Reviewed by:

*Stephen R. McNutt, University of South Florida, United States Pablo Tierz, British Geological Survey (BGS), United Kingdom*

\*Correspondence:

*Annemarie Christophersen a.christophersen@gns.cri.nz*

#### Specialty section:

*This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science*

Received: *01 March 2018* Accepted: *01 November 2018* Published: *22 November 2018*

#### Citation:

*Christophersen A, Deligne NI, Hanea AM, Chardot L, Fournier N and Aspinall WP (2018) Bayesian Network Modeling and Expert Elicitation for Probabilistic Eruption Forecasting: Pilot Study for Whakaari/White Island, New Zealand. Front. Earth Sci. 6:211. doi: 10.3389/feart.2018.00211* Bayesian Networks (BNs) are probabilistic graphical models that provide a robust and flexible framework for understanding complex systems. Limited case studies have demonstrated the potential of BNs in modeling multiple data streams for eruption forecasting and volcanic hazard assessment. Nevertheless, BNs are not widely employed in volcano observatories. Motivated by their need to determine eruption-related fieldwork risks, we have worked closely with the New Zealand volcano monitoring team to appraise BNs for eruption forecasting with the purpose, at this stage, of assessing the utility of the concept rather than develop a full operational framework. We adapted a previously published BN for a pilot study to forecast volcanic eruption on Whakaari/White Island. Developing the model structure provided a useful framework for the members of the volcano monitoring team to share their knowledge and interpretation of the volcanic system. We aimed to capture the conceptual understanding of the volcanic processes and represent all observables that are regularly monitored. The pilot model has a total of 30 variables, four of them describing the volcanic processes that can lead to three different types of eruptions: phreatic, magmatic explosive and magmatic effusive. The remaining 23 variables are grouped into observations related to seismicity, fluid geochemistry and surface manifestations. To estimate the model parameters, we held a workshop with 11 experts, including two from outside the monitoring team. To reduce the number of conditional probabilities that the experts needed to estimate, each variable is described by only two states. However, experts were concerned about this limitation, in particular for continuous data. Therefore, they were reluctant to define thresholds to distinguish between states. We conclude that volcano monitoring requires BN modeling techniques that can accommodate continuous variables. More work is required to link unobservable (latent) processes with observables and with eruptive patterns, and to model dynamic processes. A provisional application of the pilot model revealed several key insights. Refining the BN modeling techniques will help advance understanding of volcanoes and improve capabilities for forecasting volcanic eruptions. We consider that BNs will become essential for handling ever-burgeoning observations, and for assessing data's evidential meaning for operational eruption forecasting.

Keywords: Bayesian networks, structured expert judgment, volcano monitoring, eruption forecasting, White Island

#### INTRODUCTION

Volcanoes are complex systems capable of producing hazardous phenomena that can kill or injure people and destroy assets, sometimes with little warning. Interpreting a volcano's state and forecasting the likelihood, extent and intensity of its future activity is extremely challenging. Around the world, volcano observatories are responsible for monitoring volcanoes and interpreting data, often with the added responsibility of providing scientific information to authorities to assist with public safety and civil protection decisions. Typically, multiple scientific sub-disciplines (e.g., geology, seismology, geodesy, geochemistry, remote sensing) contribute to such monitoring and interpretation efforts.

Quantitative support tools for eruption forecasting and decision-support are becoming crucially important for volcano observatories and monitoring groups (Selva et al., 2012; Sparks et al., 2012). Such tools can provide a reproducible, transparent, documented framework that reinforces objective operational forecasting procedures and guidance, and can increase the level of public trust in volcanologists' advice (Barclay et al., 2015). Generally, the focus of such tools is forecasting when and how (e.g., hazard footprint/severity, eruption duration) an imminent eruption will occur.

Quantitative tools to support decision-making during a volcanic crisis are gradually being developed. Event trees are often used to outline possible sequences of events (Newhall and Hoblitt, 2002). As a sequence evolves and new information becomes available, Bayesian methods can be used for updating model outputs (Marzocchi et al., 2008; Lindsay et al., 2010). Bayesian networks (BNs) are concerned with modeling the joint probability distribution of all system variables for their evidential worth when assessing a volcano's state. BNs have been advocated to aid decision-making in volcanic crises for over a decade (Aspinall et al., 2003). They have been applied to retrospectively analyze the 1975–1977 volcanic crisis at La Soufrière volcano, Guadeloupe (Hincks et al., 2014), the 1993 explosion at Galeras volcano, Colombia (Aspinall et al., 2003) and in real-time to the 2011–2012 unrest on Santorini, Greece (Aspinall and Woo, 2014). Sheldrake et al. (2017) developed a BN to evaluate evidence for the cessation in eruptive activity of the Soufrière Hills volcano, Montserrat. Cannavò et al. (2017) introduced BNs to real-time monitoring on Mount Etna, Italy. The flexible framework that BNs offer has also been found useful for probabilistic volcanic multi-hazard assessment of tephra fallout, pyroclastic density currents and rain-triggered lahars at Somma-Vesuvius, Italy (Tierz et al., 2017). Despite these successful case studies, BNs are not widely employed in volcano observatories for forecasting eruptions or as decision-support tools to keep local authorities informed of impending volcanic hazards.

Here, we explore the utility of BN modeling for volcanic eruption forecasting in New Zealand. For this purpose, we developed a pilot project for forecasting the probability of an eruption at Whakaari/White Island volcano, New Zealand. To set up the pilot study, we used a protocol developed for risk assessment studies and described in detail in the **Supplementary Material**. Here, we provide an overview of BN modeling within the context of forecasting volcanic eruptions. We give an overview of volcano monitoring in New Zealand and describe the volcano chosen for the pilot study: Whakaari/White Island. The description of the pilot model is followed by some provisional applications. Although the pilot study did not produce an applicable tool for immediate use, we gained numerous and valuable insights for designing future BN models. We discuss our findings and insights and make recommendations for further work.

#### BAYESIAN NETWORK MODELING

Bayesian networks (BNs) provide a graphical probabilistic framework for modeling complex real-world systems (Koller and Friedman, 2009). They have their origin in the artificial intelligence community (Pearl, 1988), where they were developed to model the top-down (semantic) and bottom-up (perceptual) combination of evidence in reading (Pearl and Russel, 2001), replacing ad hoc rule-based schemes.

A BN is a directed acyclic graph, providing a model structure which represents a set of random variables as nodes (e.g., "Eruption" in **Figure 1**) and the relationships between them as arrows (often called arcs or edges, from graph theory). Arcs point from a "parent" node to a "child" node. A marginal distribution is specified for values relating to each node with no parents, and a conditional distribution is required for values associated with each child node. The absence of an arc between two nodes represents independent or conditionally independent random variables (Korb and Nicholson, 2010). However, if two (unconnected) independent parents share the same child, then the parents can become conditionally dependent when information about the child becomes available (in other words the child is observed). Hence, the (conditional) independence or dependence of two variables represented as nodes in a BN is determined by both the model structure, and the observed or unobserved state of the involved variables. For detailed explanations, theoretical considerations and examples we refer the reader to Pearl (1988), Spirtes et al. (1993), and Murphy (2012).

A BN model structure, and quantitative information about the variables, can be retrieved either from data, when available, or from experts, or a combination of both. BNs are applied in many different domains (e.g., Pourret et al., 2008; Weber et al., 2012), including risk assessment decision support (e.g., Aspinall et al., 2003; Fenton and Neil, 2013; Gerstenberger and Christophersen, 2016).

When implemented as a software program, a BN allows easy but rigorous quantification of the strengths of relationships between different variables. Moreover, the implementation of Bayes' rule (Bayes and Price, 1763), within the program, allows probabilities to be updated in the light of new evidence. Thus, a BN can be used to evaluate formally a variety of different "what if " scenarios in the face of substantial scientific uncertainties: it is this capability that makes the BN framework an invaluable decision resource for supporting volcano forecasting.

#### Developing the Model Structure

Developing the model structure involves defining the variables and identifying possible parent-child relationships. **Figure 1** illustrates different simplified modeling options for a volcanic eruption. For examples **Figures 1A,B** have only two nodes, "Eruption" and "Observable" (e.g., seismicity), while the other examples split the observables into "Geochemistry," "Seismicity" and "Deformation." In these examples, there are correlations between some of the nodes, but this does not imply causation—the observable does not cause the eruption, nor does the actual eruption event itself cause the precursory unrest. There are, however, internal processes that lead to eruption, which also cause observable precursory phenomena.

With these elemental examples in **Figures 1A–D**, the arcs can point either direction; in reality, however, there will be internal processes, such as ascending magma or a magmatic perturbation of the hydrothermal system, that involve underlying latent causal links between precursory observables and eruption. For example **Figure 1E** is a simplification of the model by Hincks et al. (2014) developed to retrospectively analyze the 1975–1977 volcanic crisis at La Soufrière volcano, Guadeloupe. Here, the node "Magmatic process" is the parent node of the eruption and of the observable nodes and thus the model represents cause-and-effect relationships. Generally, the structure depends on data availability and the quantification requirements as further discussed in the next section.

#### Quantifying the Model

The information required to quantify a BN model depends on the type of variable represented by each node. Commonly variables have a finite number of states that are mutually exclusive, and discrete values exhaustively describe these possible node states. For such variables, the dependencies are captured in Conditional Probability Tables (CPTs), as illustrated with our simple examples (**Figure 2** and further explained below).

However, when variables are best characterized by continuous values, the most common approach (apart from asserting some form of discretization) is the assumption of a parametric joint distribution, most often the joint normal distribution. To quantify such models one needs conditional means and variances for the nodes, and regression coefficients for the arcs (Pearl, 1988; Shachter and Kenley, 1989). Unfortunately, often the assumption of joint normality is not validated in practice. Other modeling techniques for continuous variables exist (Hanea et al., 2006; Langseth et al., 2009). In volcanology, the type of probability distribution for a given variable is often unknown, due to data scarcity, and several distributions may seem plausible (Tierz et al., 2016a,b). Alternative distributions could be tested within a BN framework.

For a discrete BN, as in **Figure 1A**, we need to assess the marginal probability distribution of the variable "Observable." For the variable "Eruption," we need to know the probability of "Eruption," given the state of "Observable." In the simplest case, each node has two states, "yes" or "no." Even though there are only two possible outcomes for "Eruption," the model is still probabilistic, since it calculates the probability of either "yes" or "no." To quantify the prior table for the "Observable" node (O), it is sufficient to assess one probability value, P(O = yes) or P(O = no), because given the mutually-exclusive-and-exhaustive requirement: P(O = yes) = 1 − P (O = no), and vice versa. For the same reason, to quantify the CPT of the "Eruption" node (E), given the "Observable" node (O), we only need to assess two probability values, because: P(E = yes|O = yes) = 1 – P(E = no|O = yes) and P(E = yes|O = no) = 1 – P(E = no|O = no). By the law of total probability, the probability of "Eruption" can then be calculated, using simple probability rules, to be P(E = y) = P(E = y|O = y)∗P(O = y)+ P(E = y|O = n)∗P(O = n). Eruption could have several states E1, E2, E3, and E4 distinguishing different eruption styles (e.g., Hincks et al., 2014). In that case, the probabilities to be assessed are P(E1|O = y), P(E2|O = y), P(E3|O = y) and P(E4|O = y), where again one probability can be calculated using the other three because the sum of all four must be one, since the states are mutually exclusive and exhaustive. Additionally, the probabilities P(E1|O = n), P(E2|O = n), P(E3|O = n) and P(E4|O = n) must be assessed.

In case **Figure 1B**, we need to assess the probability distribution of the variable "Eruption," i.e., for the two states E = y and E = n, we need the probabilities P(E = y) and P(E = n), which must add up to one. For the node "Observable," we need the conditional probability depending on the states of "Eruption." So, in the above example, the questions to answer are "what is the probability of observing unrest seismicity given an eruption subsequently occurs?" and "what is the probability of observing unrest seismicity given no eruption occurs?"

The required number of probabilities for a node increases with the number of parent nodes and the number of states of both parent and child node. While there is little difference in assessing the probabilities for cases (**Figures 1A,B**), this changes when comparing the probability estimates required

observables are independent and in (D) they are dependent on eruption and independent of each other, given eruption. Panel (E) is an example of a causal BN, where "Magmatic process" is the parent node for "Eruption" and the observables. "Eruption" and the observables are conditionally independent given "Magmatic process." The model is a simplification of the La Soufrière model (Hincks et al., 2014) that we adapted to our pilot study.

for cases (**Figures 1C,D**), which both have three observables "Geochemistry," "Seismicity" and "Deformation." In case **Figure 1C** we need to assess the marginal distribution of each of the observable nodes, similar to case a). Additionally, we need to assess the probability of eruption given all possible combinations of its parents' states. If each of three parent nodes has two states, the number of combinations is 2<sup>3</sup> = 8. Quantification quickly becomes intractable as the number of nodes increases and nodes have more than two states. Some parent state combinations may happen extremely rarely, engendering large uncertainties. In contrast, the complexity of the probability assessment in case (**Figure 1D**) compared to case (**Figure 1B**) has not changed.

The construction and the quantification of BNs are interrelated, and the chosen structure may change depending on the quantification requirements, data availability, and/or the understanding and representation of the problem. Different modeling choices come with advantages and disadvantages. Using joint normal distributions that preserve the properties of continuous value variables precludes representation of marginal distributions with heavy tails, or tail dependence, as part of the dependence structure. Complex dependence structures may be better represented using large CPTs by discretizing continuous variables into a large number of states. However, this comes with the price of a huge quantification burden, which, in the absence of massive data sets, renders poorly quantified conditional distributions. A mentioned above, data may be unavailable to fully or even partially quantify a chosen BN. When data are absent or incomplete, expert elicitation can contribute to BN quantification.

#### Structured Expert Judgment

Structured expert judgment (SEJ) is the process of eliciting expert knowledge as a form of scientific data (Colson and Cooke, 2017). A variety of expert elicitation protocols have been developed over recent decades, and successfully deployed in numerous domains (O'Hagan et al., 2006; Cooke and Goossens, 2008; Aspinall, 2010; Selva et al., 2012; Hanea et al., 2016). Most follow thoroughly documented methodological rules, but they differ in several aspects, including the way interaction between experts is handled, and the way an aggregated opinion is obtained from individual

CPTs for all observables are assumed to be the same in this simple example and the probability of each observable is assumed to be 50% if an Eruption occurs and 10% otherwise. The calculation of the probability for the observables is given for Seismicity in (A). In (B), Seismicity is observed, and the probability of Eruption is updated according to Bayes' rule. Given a new probability for Eruption due to an affirmative observation of Seismicity, probabilities for Geochemistry and Deformation are also updated because they are conditionally dependent on Eruption state probability (the revised Bayes' calculation is given for the Geochemistry node). Panel (C) shows the correlation matrix of the variables from 100,000 simulated cases.

experts. There is no single, best SEJ protocol: each has strengths and weaknesses.

There are two main ways in which experts' judgments are aggregated: behaviorally, involving striving for consensus via discussion and deliberation (e.g., O'Hagan et al., 2006), and mathematically, involving independent individual expert estimates being combined with a given mathematical rule (e.g., Cooke, 1991). Mathematical rules provide a more explicit, auditable and objective approach. A weighted linear combination of opinions is one example of such a rule. Equal weighting is often used, mostly because of its simplicity. While evidence shows that equal weighting frequently performs well relative to more sophisticated aggregation methods for reliably estimating central tendencies (e.g., Clemen and Winkler, 1999), when uncertainty quantification is sought, performance-based weighting provides superior information (Colson and Cooke, 2017).

One accepted differential weighting scheme is the Classical Model for SEJ (Cooke, 1991). Perhaps its most distinguishing feature is the use of calibration variables to derive performancebased weights, providing an empirical basis for validating experts' judgments. Calibration (or "seed") variables are taken from the problem domain for which, ideally, true values become known post-hoc (Aspinall, 2010). However, this is rarely feasible in practice and so calibration questions with known realizations (values) are used instead. Experts are not expected to know these values precisely, but they are expected to be able to capture them within informative ranges, defined by ascribing suitable values to marker quantiles (e.g., 5, 50, and 95th percentiles). The Classical Model has been the SEJ method most frequently applied for volcanic hazard and risk assessment worldwide, for many years (Aspinall, 2006; Martí et al., 2008; Neri et al., 2008; Wadge and Aspinall, 2014; Bebbington et al., 2018). As a consequence, many volcanologists are familiar with the approach and associated procedures.

Nevertheless, it is challenging to find calibration questions that closely match the types of questions needed to elicit conditional probabilities for a BN, especially if there are no analogs from real-life observations. For our pilot study, we did not have the resources to develop appropriate seed questions. We used SEJ with average weighting of experts, and introduced the concept of calibration to the participating experts for possible future applications.

#### Benefits of BN Modeling Within and Beyond the Pilot Study

Behind the intuitive visualization of a complex relational problem, which allows collaborative drafting and quantification of models, the most palpable advantage of a BN is its rigorous probabilistic foundation. BNs offer a flexible platform and can easily combine expert input, incomplete data sets, and other disparate sources of information. For example, different expert panels can contribute sub-models, which can then be combined to represent a complex system. If partial data sets are available, the sub-models' parameters can be fitted from data. When suitable data sets are available, entire sub-models (the structure and parameters) can be learned from data, using machinelearning techniques (Murphy, 2012), potentially extending and complementing expert knowledge.

Once the BN is built and fully quantified, its main use is to update distributions given new data or additional observations. This is referred to as instantiation, inference, or (less commonly) bi-directionality (Gerstenberger et al., 2015). Evidence added to one node will change the probabilities of all dependent nodes, regardless of the direction of the arcs. This is illustrated in **Figure 2**, which uses the elemental example of **Figure 1D**. We assume a monitoring period of 1 month. The probability of "Eruption" in any 1 month is assumed to be 5% as shown in **Figure 2A**. If an Eruption occurred, we assume that any precursory variable was observed in 50% of the prior months. The probability of any of the observations is 10% in a month when no Eruption occurs the following month as shown in the CPTs in **Figure 2A**. If "Seismicity" is observed, Bayes' rule can be applied to update the probability of "Eruption." Given an increased probability of "Eruption" the probabilities of the dependent nodes "Geochemistry" and "Deformation" also increase. It may seem counterintuitive that the observation of "Seismicity" has any impact on the probabilities of seeing changes in other observables. However, the BN has implicit connections between observable nodes via hidden processes relating to "Eruption." The CPTs express these dependencies which, in the real world, are driven by underlying magmatic processes that are not directly modeled in this variant of a BN model. **Figure 2C** shows the correlations between the four variables in the BN model. We used the BN software Netica (Norsys, 1995-2018) to simulate 100,000 realizations of the joint distribution of the four variables. We calculated these correlations to illustrate that dependence through correlations may result from the chosen CPT values despite the lack of causal relation between these variables.

The bi-directionality of BNs is a large advantage over event trees: BNs can be used to analyze the dependencies and can advance the understanding of the system. This characteristic can also be beneficial in operational applications; for example, if one loses seismograph network coverage due to a telecommunications signal failure, the BN allows one to infer what the likely seismicity level would be, given other nonseismic observations. Even more important, the BN allows one to make use of "negative evidence": if gas flux suddenly decreases, is it due to a conduit blockage (dangerous) or to a change in gas exsolution in the reservoir (likely benign)? The two carry quite different hazard implications, and both scenarios need to be accommodated and ascribed relative probabilities. Inference as to which is the actual cause may be weighed by what other observables are indicating.

#### PILOT STUDY OF A DISCRETE BN MODEL TO FORECAST ERUPTIONS ON WHAKAARI/WHITE ISLAND

This section provides some background on volcano monitoring in New Zealand before describing Whakaari/White Island, the volcano chosen for the pilot study. We briefly outline the motivation for the pilot study before describing pilot model structure, the estimated probabilities and the eruption probabilities. We briefly discuss other findings from the workshop and close with a demonstration of typical BN calculations for two individual experts.

#### Volcano Monitoring in New Zealand

In New Zealand, GNS Science, through the GeoNet project, conducts national volcanic monitoring (New Zealand Ministry of Civil Defence Emergency Management, 2015). GeoNet issues notifications of any change in volcanic alert level status through Volcanic Alert Bulletins to the Ministry of Civil Defense and Emergency, other agencies, and the media. Volcanic Alert Bulletins are also published on GeoNet's website (GeoNet, 2018).

GeoNet coordinates the volcano monitoring team, which consists of GNS Science staff based at three sites. The volcano monitoring team meets regularly to review the status of all 12 monitored New Zealand volcanic centers, and to set Volcano Alert Levels (Potter et al., 2014) and the Color Codes of the International Civil Aviation Organization. Team members prepare Volcanic Alert Bulletins when required. The team is responsible for providing scientific advice to emergency management authorities at the national, regional, and local level (New Zealand Ministry of Civil Defence Emergency Management, 2015). In addition to these legislative requirements, the volcano monitoring team regularly estimates the probability of forthcoming eruptions for internal health and safety policy requirements (Jolly et al., 2014; Deligne et al., 2018).

GeoNet data is available free of charge. While GeoNet is committed to transparent data discoverability, the process of operationalizing data discovery for non-continuous data (e.g., monthly gas flights, gas isotope sampling results) was in its early stages at the time of our pilot study, and therefore not readily available for the model development. Thus, the parameters of the BN model were estimated by expert elicitation rather than from data.

#### Whakaari/White Island Volcano

Whakaari/White Island volcano is of one New Zealand's most active volcanoes and it is a major tourist attraction. Located about 50 km off the coast of the North Island (**Figure 3**), it is the northernmost subaerial active volcano of the Taupo Volcanic Zone (Cole and Lewis, 1981). Its emerged area of 3.3 km<sup>2</sup> is the summit of the much larger White Island Massif (Cole and Nairn, 1975). The volcano is andesitic to dacitic in composition and is formed of two overlapping cones with a succession of lava flows, breccias, agglomerates, and unconsolidated beds of ash and tuffs containing lava blocks (Black, 1970). Its main topographical feature (Main Crater) consists of three sub-craters: western, central and eastern (Houghton and Nairn, 1989, 1991), with the current active vent in the eastern sub-crater.

**Figure 4** illustrates a conceptual model of the volcano. The magmatic system is perceived to consist of a deep collective reservoir (4–7 km) that feeds a shallower convective reservoir (1– 2 km) from which small amounts can be injected into an upper conduit to the surface (Clark and Otway, 1989; Houghton and Nairn, 1989; Cole et al., 2000; Kilgour et al., 2016). Werner et al. (2008) proposed that gas and magma are transported from deep to shallow levels within a closed system (magma convection), while an open system characterizes the upper conduit. Steaming ground areas, hot springs, fumaroles, and an acidic crater lake are the surface expressions of the volcano hydrothermal system which has existed for at least 10,000 years (Giggenbach and Glasby, 1977). Several attempts have been made to characterize the volcano's hydrothermal system (e.g., Ingham, 1992; Nishi et al., 1996a). From the location of volcanic earthquakes, Nishi et al. (1996b) suggested that the active hydrothermal system is located below Main Crater but its extent and evolution with time (e.g., presence of a seal) remain poorly understood. The Crater Lake level had several filling/evaporative cycles that correlate with varying discharge of the springs (Christenson et al., 2017).

Volcanic activity ranges from fumarolic and hydrothermal during quiescence, to phreatic, phreatomagmatic, Strombolian and effusive during more active periods (Cole and Nairn, 1975; Houghton and Nairn, 1989; Nishi et al., 1996b; Chardot et al., 2015). A major eruptive magmatic episode occurred between 1976 and 2000 and formed the current active crater (**Figure 3**); this episode comprised several cycles, with intraepisode clusters of activity from 1976 to 1993, 1998, 1999, and 2000. There were at least 250 eruptions during this quartercentury period, ranging from localized steam and mild ash eruptions to eruptions with ejecta outside the Main Crater (G. Jolly, personal communication). These eruptions are not classified by eruption style, but 45 had ballistic and/or surge impacts beyond the crater complex (**Figure 3B**), suggesting an eruption rate of 0.15 large eruptions/month. Assuming a Poisson distribution for the number of eruptions in a month, the probability of one or more eruptions within 1 month can be calculated from the rate R like 1-exp(-R). Thus, the probability of one or more eruptions within 1 month impacting beyond the 1976–2000 crater complex is 14%. However, the 1976–2000 eruptive episode altered the volcanic system and the prior rate is not necessarily representative of activity since.

The most recent eruptive episode began with unrest in August 2011 (Chardot et al., 2015), and continued through the end of 2016. During this episode there were eight eruptions (**Table 1**), half of which had ballistic and/or surge outside the 1976–2000 crater complex and would have posed safety concerns if they had coincided with site visits to the island. Calculating an average rate of eruptions impacting beyond the crater rim for the period from 2001–mid 2018, there were four eruptions within 210 months, which equals 0.02 per month. The Poisson probability for an eruption within 1 month impacting beyond the 1976–2000 crater complex is 2%. There was also one magmatic effusive eruption in form of a lava dome (**Table 1**). There were no eruptions in the 2 years leading up to the expert elicitation workshop central to our pilot study.

Eruptions at Whakaari can be preceded by increased levels of seismicity (Latter et al., 1989), magnetic changes (Hurst et al., 2004) and deformation within Main Crater (Clark and Otway, 1989), although eruptions can occur with no useful short-term precursory activity to indicate that an eruption is imminent. In the case of the examples cited, changes in monitored parameters were observed months before the eruption.

Volcano monitoring at Whakaari has been ongoing since 1967 and part of the GeoNet project since its inception in 2001. As of August 2018, monitoring includes continuous visual observations (three on-island webcams), seismic (two continuous seismic broadband stations), deformation (two GPS stations) and SO<sup>2</sup> emission measurements (two miniDoas sites) (**Figure 3B**), with additional monthly gas flights (CO2, SO2, and H2S emissions) and regular field campaigns (e.g., leveling, fumarole and spring sampling, CO<sup>2</sup> soil gas surveys, magnetic surveys).

Whakaari's seismicity presents a full spectrum of event types, ranging from long-period to volcano-tectonic events and tremor (Sherburn et al., 1998). However, the limited number of local seismometers (and the nearest station off the island being about 50 km away) prevents the accurate and precise location of local seismic events. Therefore, we can determine the depths of larger earthquakes that are recorded off island but have little or no effective depth control on the more frequent small events.

Deformation is mainly assessed using campaign-leveling data, and the sources of the recent changes are interpreted as being due to varying pressurization of the main fumarole field (Peltier et al., 2009; Fournier and Chardot, 2012; Christenson et al., 2017).

#### Motivation for the Pilot Study

The pilot study was motivated by a GNS Science internal presentation on life safety when working on volcanoes. At GNS Science, thresholds of 10−<sup>3</sup> , 10−<sup>4</sup> , or 10−<sup>5</sup> for the hourly probability of a fatality at an active volcano trigger different levels of managerial sign-off for undertaking fieldwork on the volcano (Jolly et al., 2014; Deligne et al., 2018). As input to the life-safety risk calculations, members of the volcano monitoring team regularly estimate the probability of an eruption for New Zealand volcanoes in a state of unrest. Small probabilities are challenging to estimate (Burns et al., 2010), and the team has no shared quantitative tools or models to assist with determining the eruption probabilities. This is in contrast to the GNS seismology

team that has several models to forecast earthquake occurrence on different time scales (Christophersen et al., 2017); the latter are being continuously tested and evaluated in international testing centers (e.g., Gerstenberger and Rhoades, 2010; Rhoades et al., 2016). Recent positive experience with BN modeling for risk assessment in carbon capture and storage (Gerstenberger and Christophersen, 2016) motivated us to explore BNs to address volcanic eruption probabilities.



*Eruptions dates are according to New Zealand standard time; eruptions 1–6 are also described by Chardot et al. (2015), where the dates are reported in Universal Time Coordinated. Eruptions 1, 3, 6, and 7 (marked with* \**) had ballistic and/or surge impacts beyond the 1976–2000 crater complex and thus are of interest for health and safety during field work. The Volcanic Alert Bulletins can be accessed at www.geonet.org.nz/volcano/vab.*

Our goal was not to develop the first operational BN model for eruption forecasting in New Zealand, but to explore the utility of BN modeling for eruption forecasting in principle, and to investigate the challenges and potential benefits of the method. As a consequence, we were experimental in out approach. For example, we did not attempt to constrain the number of variables to a manageable number, quite the opposite; for the first model, we aimed to capture all variables that are regularly monitored. We also did not insist on consistent definition of all nodes, as further explained below.

#### Pilot Model Structure

The aim of the model is to capture the conceptual understanding of the volcanic processes that lead to eruptions and to represent all regularly monitored observables. The pilot BN did not attempt to model dynamic or transient aspects of the volcanic system, such as potential drying of the crater lake, which would engender substantive changes in some variables. The development of the model structure was iterative and involved defining the variables. It took a 2-h meeting for a small team with diverse skill sets to adapt the La Soufrière model to Whakaari and two further hours to draft the initial set of elicitation questions. Ideally there would have been two to three 2-h meetings with a small group of experts from different sub-disciplines to review and fine-tune sub-networks for the overall model. Having a working BN model example (Hincks et al., 2014) helped participants who were new to BN modeling, enabling them to quickly grasp the concepts and understand what we were trying to achieve.

Through feedback with individual experts, we quickly ascertained that individuals from different sub-disciplines had different understandings of the eruption driving processes. As a consequence, several node definitions were purposefully left vague to accommodate different thinking about these processes. A joint discussion with experts from seismology, fluid geochemistry, geodesy, and general geophysics highlighted how the BN framework allowed the different understandings of the volcanic processes to be discussed in an insightful way. Unfortunately, the pilot study had time constraints that did not allow full agreement to be reached on all the nodes.

**Figure 5** presents an overview of all nodes that were elicited. The model structure follows **Figure 1E**, where observables are conditionally independent of eruption given magmatic processes. For ease of handling, we split the model into four areas during the model development process and the expert elicitation. The areas, described in more detail below, are: (1) volcanic processes leading to eruption, (2) observations related to seismicity, (3) observations related to surface manifestations, and (4) fluid-geochemical observations. **Table 2** summarizes the nodes describing the volcanic processes leading to eruption and the eruption nodes.

#### Volcanic Processes Leading to Eruption

The model simplifies the volcanic processes that can lead to eruption into four unobservable nodes (yellow ovals in **Figure 5**). Node 1 is "Gas rich magma ascending" and represents fresh gascharged magma entering the system from depth. Volatiles are driving the ascent of the magma into the upper part of the edifice. Node 2 is "Shallow magma" and represents the presence of a shallow magma reservoir being fed by the deeper reservoir. "Gas rich magma ascending" can fill the shallow magma reservoir and therefore is a parent of "Shallow magma." There is consensus that there has been shallow magma close to the surface at Whakaari for at least several decades (Houghton and Nairn, 1989; Cole et al., 2000). Therefore, we treat this node as a constant (state "yes"). As a consequence changes in its parent or child nodes have no direct influence on the node itself or on other dependent nodes. However, since the presence of "Shallow magma" is a critical component of the conceptual model of Whakaari (**Figure 4**), we chose to keep this node within the model. Node 3 is "Magmatic perturbation of the hydrothermal system." Both "Gas rich magma ascending" and "Shallow magma" can lead to "Magmatic perturbation of the hydrothermal system." Node 4 is "Presence of a hydrothermal system seal," which has no parents. This node describes the partial or full sealing of the hydrothermal system seal that reduces gas emissions and allows gas to accumulate. Gas accumulation pressurizes the conduit and can lead to more explosive phreatic or magmatic explosive eruptions.

The volcanic processes can lead to three types of eruption (orange ovals in **Figure 5**): "Phreatic eruption" (node 5), "Magmatic explosive eruption" (node 6) and "Magmatic effusive eruption (node 7). Phreatomagmatic eruptions are included in

magmatic explosive eruptions. The La Soufrière model only had one eruption node with different states for different eruption types. We struggled to define mutually exclusive and exhaustive states because different eruption types can occur within the 1-month period of interest. We explored the option of estimating the probability of the next eruption within the 1-month period but found it easier to represent different types of eruption by separate nodes. To be consistent with the regular elicitation of eruption probabilities, nodes 5 and 6 are defined as one or more eruptions within the next month "impacting beyond the rim TABLE 2 | The description of nodes 1–4 that capture the hidden processes of the volcano and nodes 5–7 that model three types of eruption.


*These descriptions were given to the experts as workshop notes that are included for all nodes in the* Supplementary Material*. The definitions are purposefully vague to allow for different understandings of the fundamentals of the volcanic system.*

of the 1976–2000 crater complex" (see **Figure 3**). A magmatic effusive eruption (node 7) includes dome development and any lava at the surface. It is likely to happen within the crater, which is at a lower elevation than the area the monitoring team would access. Historically, lava domes at Whakaari are small in comparison to other volcanoes and are therefore not an immediate threat to health and safety during fieldwork. **Figure 5** shows how the eruption nodes are connected to the driving processes.

#### Observations Related to Seismicity

Observations related to seismicity include three types of earthquake occurrence distinguished by their frequency content, and tremor. They reflect the variety of events recorded at Whakaari (Sherburn et al., 1998). The reader is directed to the review by McNutt (2005) for a comprehensive description of each earthquake type. For each of these observables there is an additional node that captures recent occurrence of the respective observables and reflects that the system may have a memory. Thus, recent activity can indicate that fluids or magma have shifted in the system. Here, we only discuss the main process(es) driving each type.

High frequency earthquakes (node 8 for high rate and node 9 for recent high rate) are associated with shear fracture and thus an indication of stress changes. Low frequency and hybrid earthquakes (node 10 for high rate and node 11 for recent high rate) are thought to be associated with fluid processes. Very long period earthquakes are associated with significant fluid movement in the subsurface (node 12 for high rate and node 13 for recent high rate). A further node "Extended duration of earthquake swarm" (node 14) assesses the duration of earthquake activity. We do not distinguish the frequency content of the earthquakes that contribute to the swarm because it may be difficult to measure the frequency content when more than one process is causing the earthquake occurrence. We also consider tremor, which is a persistent seismic signal of varying durations often associated with volcanic eruptions (Konstantinou and Schlindwein, 2003 and references therein). High amplitude of tremor (node 15 for high amplitude and node 16 for recent high amplitude) can reflect the size of an eruption, following McNutt (2005), who showed that higher tremor amplitudes correlate with higher Volcanic Explosivity Index of eruptions. At Whakaari, periods of increasing tremor have been modeled to retrospectively forecast eruptions (Chardot et al., 2015).

All of the nodes representing observations related to seismicity (nodes 8–14) have the same two parents: "Gas rich magma ascending" and "Magmatic perturbation of the hydrothermal system." The tremor nodes (nodes 15 and 16) also depend on "Shallow magma."

#### Observations Related to Surface Manifestations

This part of the model considers the nodes related to deformation and to the crater lake and the temperature of fumaroles. "Anomalous deformation observed by leveling" (node 19) has, as parent, the "Magmatic perturbation of the hydrothermal system" and "Large-scale ground inflation as measured by GPS" (node 22) has as parent the node "Gas rich magma ascending." "High gas emissions through lake (ebullition)" (node 17) and "Lake level change independent of precipitation" (node 18) have as parents the "Magmatic perturbation of the hydrothermal system" and the "Presence of a hydrothermal system seal." "Increase in fumarole temperature" (node 20) and "Increase in crater lake temperature" (node 21) have the parents "Gas rich magma ascending" and "Magmatic perturbation of the hydrothermal system."

One node, "Fresh glass" (node 30), was added during the elicitation workshop and has as parents the "Gas rich magma ascending" and the "Shallow magma" nodes.

#### Observations Related to Changes in Fluid Geochemistry

Elevated gas emissions usually relate to elevated volcanic activity (e.g., Giggenbach and Sheppard, 1989). Observations related to changes in fluid geochemistry include "Elevated gas flux of CO<sup>2</sup> in air" (node 23), "Elevated gas flux of SO<sup>2</sup> in air" (node 24), "Elevated gas flux of CO<sup>2</sup> in fumaroles" (node 25), "Elevated gas flux of SO<sup>2</sup> in fumaroles" (node 26) and "Elevated diffuse (soil) gas emission" (node 27).

"Changes in the composition of fumaroles consistent with the presence of magmatic volatiles" (node 28) and "Changes in the composition of springs and lakes consistent with the presence of magmatic volatiles" (node 29) are also indications of "Gas rich magma ascending." However, the change of the composition would be fed into the fumaroles through the hydrothermal system. Only in case of very high temperature fumaroles (≥800◦C) would the hydrothermal system be circumvented (Giggenbach and Sheppard, 1989). Therefore, the experts decided that only "Magmatic perturbation of the hydrothermal system" is a parent of these two observable nodes (28 and 29).

#### Definitions of States

Clearly defining the nodes and states—so that all experts share the same understanding of the elicitation questions—is an important part of the BN model development. To reduce the elicitation burden we described each variable with two states only: "yes" and "no." Since the probabilities for all states must add to 100%, we only asked the question for the "yes" state and calculated the complementary "no" state probability. During the model development, experts voiced strong reservations about defining variables with hard, definite thresholds. They were concerned that one or more variables might be significantly elevated, such as seismicity prior to the 2014 deadly Mount Ontake eruption (Kato et al., 2015), but still not reach the threshold set to trigger a warning. Near-misses against arbitrary threshold for node states can also lead to negative (Brier) skill score which appear to devalue BN forecast performance, as was found for Soufrière Hills Volcano, Montserrat (Wadge and Aspinall, 2014). Experts in our study wanted to model the observables as continuous probability density function. While the modeling techniques we could access during the pilot model study did not allow for this, we envisaged continuous BN modeling as subsequent model development. As a consequence, we did not define thresholds for the states of the observable nodes because experts had widely varying opinions on what levels are appropriate. Attempting to enforce consensus on issues about which our experts had strong reservations seemed counter-productive to the aim of the study, which was to explore the utility of BN modeling for eruption forecasting in New Zealand.

Instead we asked each expert during the quantification of the model to describe what threshold they had in mind when answering the question. For example, for node 8 "High rate of high frequency earthquakes" we asked, "how do you define "high rate" of high frequency earthquakes?" There is a wide variation in the definitions of thresholds, and for node 8 the answers were: More than 5,10, 20, 30 (two experts), 50, 100 per day, more than two times the background; one magnitude greater than the background rate; a couple of earthquakes per hour for at least half a day, visible with the naked eye on the seismogram. In another example, for node 20, the answers to "how do you define an increase in fumarole temperature" were: + 5 ◦C, +10–15◦C, +10– 20◦C, +20◦C (three experts), +10% of past value, more than 20% change from recent trend, 2–3 times the normal temperature. For node 22, "How do you define elevated gas flux of CO<sup>2</sup> in air" the answers were: More than 250, 500, few hundred, 1,000, 2,000, 2,000, 3,000 tones/day; at least two measurements more than 20% above baseline, increase by over 100% over the background value. Having different thresholds for the states of the observable nodes limits the usability of the pilot study. Although theoretically possible, we did not attempt to group similar answers and use the matching probability estimates to derive a BN with consistent state, because the experts had reservations about using a model with only two states for the observable nodes.

#### The Estimated Probabilities

To quantify the model, the experts estimated 120 probabilities and their 80% confidence intervals, which are summarized in Christophersen (2017). **Figure 6** presents a few questions to show the spread in answers between experts. Broadly, the elicitation questions can be split into four categories depending on the overlap of the experts' uncertainty intervals. **Figure 6a** shows an example of good agreement, where all experts' ranges overlap. **Figure 6b** is an example in which the uncertainty range of one or two experts fell outside the rest of the answers. In **Figure 6c** the experts' answers fell into two groups with some overlap of the uncertainty range, which indicates large uncertainty in the answer. Finally, **Figure 6d** is an example of even larger uncertainty, because the variety of answers covers the entire possible range (from 0 to 100%) and individual ranges tend to be wider. Experts are in good agreement for only 22 out of 120 questions.

For about one third of the questions (43 out of 120), there are one or two experts who answered significantly differently to the others. This may be because they understood the questions differently, or they had different definitions of states in mind. Alternatively, they might have a different understanding of the system or its processes. With more time and resources, it would have been beneficial to explore these outliers as they could help improve model definitions. For about half the questions, there were either two dichotomous groups of responses (17 out of 120), or a very large spread of values (38 out of 120). Thus, for about half the questions there were large uncertainties in the probability estimates.

good agreement of experts' judgments; (B) two apparent outlier credible intervals; (C) two groupings of experts' judgments, and (D) large uncertainties in experts' credible intervals. For each graph the experts' ranges are ordered in terms of decreasing median size and thus the numbering on the y-axis does not refer to the same expert in the different examples.

#### The Eruption Probabilities

The eruption probabilities are a main result from the pilot modeling and are shown in **Figures 7**, **8**. The nodes that are relevant for calculating the eruption probabilities are the four unobservable nodes (yellow ovals in **Figure 5**) and the three eruption nodes (orange ovals in **Figure 5**). These nodes had the same definition of states for all experts and therefore we present the results for each individual expert as well as combined results. Combining experts' judgments becomes problematic only if we wish to condition on any of the observables of the BN (nodes 8–30); this is because then different thresholds, chosen by the experts, would come into play.

**Figures 7**, **8** include the best estimates (circle) for each expert and three composites. Composite 1 was calculated by applying equal weights to the experts' probability estimates for each question. Composites 2 and 3 consider the experts' self-weighting. Experts assessed on a scale from 1 (not very confident) to 10 (very confident) their own expertise in the four different subject matter areas of the BN (i.e., volcanic processes leading to eruption; seismicity; surface manifestations; and fluid geochemistry). For the most confident expert the sum of selfassessments over all subjects was 1.76 times higher than for the least confident expert. For Composite 2, we normalized the experts' self-assessments for each subject separately so that the overall contribution varies between the experts, by

a factor of 1.76 at the extreme. For Composite 3, we first normalized each expert's assessment across all four subjects, so that each expert has an equal contribution overall but with possibly different weightings across subjects depending on their self-assessments. There is negligible difference in the probabilities for the three composites. This may be because we only used a linear scale for weighting. When calibrating the experts in the Classical Model (Cooke, 1991), the weights between experts can differ by orders of magnitude. However, self-weighting is known to correlate poorly with uncertainty judgment performance (Burgman et al., 2011). In our case, the experts suggested the self-weighting themselves as an additional way to express their uncertainty in some areas compared to others.

and for three composite results. Please see text for an explanation of the three composite results.

The graphs also show 80% confidence intervals calculated from the 10th and 90th quantiles, obtained as follows. We fitted beta distributions to each set of lower and upper quantiles for each expert; we then sampled from each beta distribution 1,000 times, re-quantified the BN for each combination and selected the 10 and 90th quantiles of the results.

The composite probabilities per month are around 19% for phreatic (**Figure 7A**), 14% for magmatic explosive (**Figure 7B**), and 5% for magmatic effusive (**Figure 8**). For nodes 5 and 7, the probabilities of individual experts vary significantly, i.e., beyond the estimated 80% confidence interval, while node 6 only has one outlier. The probability range for phreatic eruption is from 0.62% (Expert 9's best estimate) to 60% (Expert 5's best estimate), i.e., two orders of magnitude. For magmatic explosive

eruption the individual experts' probabilities range from 2.5% (Expert 9) to 65% (Expert 7). Magmatic effusive eruption has the least variation in individual results ranging from 0.2% (Expert 9) to 12% (Expert 7), albeit still covering more than an order of magnitude difference. The spread of probabilities reflects large uncertainty between experts' judgments, whereas the wide individual intervals reflect large uncertainty within experts' assessments when asked about these conditional probabilities (**Figure 6**). With more time to clarify some of the questions and the experts' responses, both types of spread would likely decrease.

To compare the model eruption probabilities to the long-term eruption rate of concern to safety when undertaking fieldwork (section Whakaari/White Island Volcano) we sampled the BN to calculate the probability of either one or both types of eruptions (node 5 and nodes 6) occurring in a 1-month period. We used the gRain package (Højsgaard, 2012) to calculate this probabilities (**Figure 7C**) to be around 28%. The results of the individual experts range from 3% (Expert 9) to 81% (Expert 7). Most experts' judgments and the composite results are higher than the Poisson probability of 2%, calculated from the average eruption rate since 2001. Unfortunately, the resources of the pilot study were limited and there was no chance to have thorough discussions with the experts about the findings. Therefore, the results need to be treated with hesitation. Perhaps coincidently, Expert 9 had BN eruption probabilities similar to those implied by observations over the past 17.5 years. We selected the results of this expert and one other to illustrate the power of a BN when evidence in form of observations can be added (section Updating the BN With Evidence).

#### Other Findings From the Workshop

During the workshop we collected feedback from the experts on their impressions of BN modeling and SEJ, as well as general feedback on the workshop itself. Most comments were positive. The main findings were: (1) the majority of experts were interested to learn more about BNs and possibly to be able to run their own models; (2) experts identified further research questions that BNs could help answer such as identifying what monitoring data was most important and testing different conceptual ideas, and (3) there were suggestions to apply BNs to other volcanoes and use it in forecasting routinely, maybe in combination with Bayesian event trees. At the workshop, we did not conduct a calibration exercise that would have allowed us to apply the Classical model, because we did not have the resources to develop calibration questions that were relevant for the conditional probabilities. However, we did introduce to concept of performance weighting and the experts were supportive of the notion of applying this in future elicitations.

#### Updating the BN With Evidence

The fully quantified BN can be used to update distributions given new data or additional observations as already illustrated in **Figure 2**. Setting evidence, as it arrives, is a key step for determining how eruption probability changes when certain observables occur. Analysis of the way different observations influence eruption probability can help to better understand drivers and influences within the volcanic system. Since we had limited opportunity to review the results with the experts the following presentations are for illustrating how a BN works generally, rather than trying to deduce formally volcanic processes integral to Whakaari. For **Table 3**, we select two experts to show the effect of observing either a "high rate of high frequency earthquakes" (node 8) or "elevated gas flux CO<sup>2</sup> in air" (node 23) on the probabilities of "yes" for all other nodes. The selected experts are Expert 9, whose eruption probabilities were the smallest (and similar to the observed Poisson probability since 2001), and Expert 11, whose values are generally midrange, relative to the group. **Table 3** shows the probability of "yes" for each of the 30 nodes in the BN for both these experts. The first column for each expert presents the unconditional probability, i.e., the probability of "yes" in any month for the nodes that are listed in the rows. The second and third columns for each expert have results for two selected observables. Taking the values from each line in the table, one can compare the effect of these observations on all the other nodes. For example, for Expert 9, the probability of a "N5: Phreatic eruption" increases from 0.62 to 1.4% when "N8: High rate of high frequency earthquakes" is observed, and from 9.9 to 25% for Expert 11 under the same condition.

**Table 3** illustrates how adding evidence changes the probability of all nodes that are directly or indirectly connected, i.e., in our case nearly all nodes. For example, observing "High rate of high frequency earthquakes" makes "Elevated gas flux CO<sup>2</sup> in air" more likely (increase from 5.8 to 28% for Expert 9 and 13 to 25% for Expert 11) and observing "Elevated gas flux CO<sup>2</sup> in air" increases the probability for "high rate of high frequency earthquakes" from 1.6 to 7.6% for Expert 9 and from 15 to 29% for Expert 11. The influence of seeing one observable on the likelihood that another being positive can be explained by their indirect linkage via common unobservable volcanic process nodes (yellow nodes in **Figure 5**). In this instance, the CPTs of the observable nodes reflect judgments that these observables are more likely to occur when gas rich magma ascends (node 1), or when there is a magmatic perturbation of the hydrothermal system (node 3). Thus, using the equations outlined in **Figure 2**, the probabilities of nodes 1 and 3 both increase when one of the observables occurs. Higher probabilities for nodes 1 and 3 in turn increase the probabilities of the other observables. This bi-directionality is useful to explore the diagnostic sensitivity of the results (the eruption nodes 5–7) to observing individual nodes, as done in **Table 4**.

An example of independent nodes is nodes 4 and 8. As a consequence of their independence, the probability of node 4 does not change when node 8 is instantiated. In contrast, there is a parent-child relationship between node 4 and node 23. For Expert 11, the probability of node 4 changes when node 23 is instantiated. However, this is not the case for Expert 9, who judged node 4 and node 23 to be independent by providing the same probabilities for the CPT of node 23 regardless of whether node 4 was "yes" or "no."

Looking at the hidden nodes "N1: Gas rich magma ascending" and "N3: Magmatic perturbation of the hydrothermal system," the effect of observing "N8: High rate of high frequency earthquakes" and "N23: Elevated gas flux CO<sup>2</sup> in air" is reversed for both experts: for Expert 9, the probability for node 1 increases from 1 to 38% for observing high frequency earthquakes, compared to only 11% for observing elevated CO<sup>2</sup> in air. This trend is reversed for node 3, where the probability increases from 21 to 46% for observing high frequency earthquakes but to 79% for elevated CO<sup>2</sup> in air. Expert 9's explanation for this is that a high rate of high frequency earthquakes is an indication of rock breaking, which is very likely caused by gas rich magma ascending. Other processes can also cause elevated CO<sup>2</sup> in air and there may also be a delay in CO<sup>2</sup> reaching air when gas rich magma is ascending. Other nodes that have a notable higher probability when conditioned on node 8, rather than on node 23 include "N7: Magmatic explosive eruption" and "N30: Fresh glass." These outcomes are again consistent with fresh magma rising abruptly and breaking rocks along the way. Unfortunately, there was not time to discuss the results with all experts. Doing so, especially as a group process, can help to share, and perhaps resolve, differing understandings of the system.

A final example from **Table 3** might initially appear counterintuitive: the probability of elevated gas flux CO<sup>2</sup> and SO<sup>2</sup> in fumaroles is higher for a high rate of high frequency earthquakes than for elevated CO<sup>2</sup> gas in air. This is the case for both Experts 9 and 11, but more so for Expert 9. Expert 9 estimates that node 1 has a much stronger influence on the gas flux in fumaroles than node 3. As discussed above, the probability of the former is much higher (38 vs. 11%) when a high rate of high frequency earthquakes is observed compared to elevated gas flux CO<sup>2</sup> in air.

**Table 4** shows the probability of observing the different eruption types in the next month for Expert 11, given evidence for the individual observational nodes listed in the rows. For example, the largest increase in the probability of "yes" for "Magmatic explosive eruption" from an individual node comes from node "N28: Changes in the composition of fumaroles consistent with presence of magmatic volatiles" (45% compared to 9.9% for unconditional). This example illustrates how the BN method allows the probabilistic weights of different pieces of evidence to be enumerated, one against another, in a coherent and objective manner.

At the bottom of **Table 4**, we show the probability of the three eruption nodes for all observables happening at the same time, as well as different combinations of nodes. When all observables occur, the probabilities for eruption in the next month are 60% for "N5: Phreatic eruption," 66% for "N6: Magmatic explosive eruption" and 5% for "N7: Magmatic effusive eruption." In the following rows, we combine different nodes to explore which ones are most influential to get closest to the maximum probabilities found when all observables occur. The probabilities of eruption increase most for a combination of nodes from different areas of the BN, namely seismological and changes in fluid geochemistry.

#### DISCUSSION AND RECOMMENDATIONS

The aim of the pilot project was to explore the utility of BN modeling for eruption forecasting rather than to develop an operational model for short-term volcanic hazard assessment. The pilot provided opportunities for productive TABLE 3 | The probability of "yes" for Expert 9 and 11 of observing each node for the unconditional Bayesian network with no evidence added, and evidence added to "yes" for node "N8: High rate of high frequency earthquakes" and node "N23: Elevated gas flux of CO2 in air."


multi-disciplinary and international collaborations, and led to many useful insights that, most likely, would not have emerged without having a structured elicitation framework (detailed in the **Supplementary Material**).

The volcano monitoring team, as key stakeholder for assessing the usefulness of BN modeling for eruption forecasting, was involved in all stages of the pilot project. The team members participated enthusiastically and experts from New Zealand universities readily agreed to participate in the workshop to quantify the model. The strong engagement of volcanologists from within and outside the volcano monitoring team reflects the interest in probabilistic methods for eruption forecasting. The participants saw other possible applications of BNs in their own work, ranging from better understanding the volcano system to deciding what monitoring data were critical and where to put additional monitoring stations.

The development of the BN model structure provided a useful framework for experts from different sub-disciplines to share and synthesize their respective understandings of the volcano system. The graphical representation with simplified causal links proved to be an illuminating "prop" as different experts explained their interpretations of the various elements of the system. Thus, the BN concept enhanced scientific discussion between experts, who already regularly discuss Whakaari.

TABLE 4 | For Expert 11, the probability of each of the three eruption types occurring in the next month (node = "yes") given positive evidence is present for all observational nodes (in the rows), and the same eruption probabilities for different sub-combinations of observational nodes (at the bottom of the table).


*Note that the probabilities now refer to the nodes in the column caption and not in the rows, as in* Table 3*.*

In the following, we highlight issues related to the model complexity, definitions of variables and states, and BN modeling in volcanology.

#### Clarity of Definitions for Nodes and States

Given the time constraints on the pilot study, we did not have the opportunity to discuss all nodes with all experts prior to the elicitation workshop. There was extensive discussion during the workshop although insufficient to gain clarity on all aspects. In particular, the node concerning the "Presence of the hydrothermal system seal" prompted debate among the experts. While written descriptions of each expert's understanding of the node are similar, the spread of probability judgments in questions involving this node indicates large uncertainty. Many experts commented that they were unsure how to deal with this node. This highlights that, while it is important for all experts to have as much clarity and agreement on all nodes in the BN as possible, such challenges are often intrinsic to a complex problem. Their emergence through elicitation and BN building often signal where critical and unresolved issues may exist. Any "hidden" nodes that describe the unobservable processes of the volcanic systems can be expected to be vaguer since experts inevitably have varying understandings of those processes.

For simplicity, the pilot model was configured with only two states for each variable. The experts did not agree that such a simplification was justified due to the continuous nature of data, which characterize most observables, and the wide range of values they can cover. Given the challenge to set thresholds to distinguish between the two states, we did not enforce consensus. Instead the experts provided their individual definition of thresholds when estimating the probabilities (see section Definitions of states). The different definitions of thresholds contributed to the spread in responses (**Figure 6**). Other eruption forecasting models have adopted fuzzy logic to deal with monitoring variables (e.g., Marzocchi et al., 2008; Selva et al., 2012). To distinguish a parameter "anomaly" from its "non-anomaly" state, two separate determinative thresholds are chosen, and a fuzzy function gauges the "degree of anomaly" in between. Within this approach, a probability density function is constructed for the event of interest, e.g., for (magmatic) unrest or eruption. Whilst, for simplicity, our workshop exercise adopted a basic two-state indicator for parameter conditions, for eruption forecasting we recommend working with a fully probabilistic approach that allows for modeling variables measured on a continuous scale with continuous probability density functions, instead of fuzzy logic. New modeling techniques make this feasible (Hanea et al., 2006; Cannavò et al., 2017).

#### Model Complexity

We included in the BN model all observables that are regularly monitored at Whakaari. The subsequent number of nodes in the pilot model created challenges for estimating the required probabilities and for further processing of the data. Within the 2 half-day workshop, our experts succeeded in estimating all required probabilities, as well as provide comments on other aspects of the model and the elicitation process. However, informal feedback indicates that it would have been better to have longer to think about individual variables. This feedback is reflected in the spread of responses to some questions (see **Figure 6**).

Experts commented in the feedback questionnaire that there were too many nodes for observations related to seismicity and to changes in fluid geochemistry. In particular, the nodes related to recent seismicity did not "behave" as expected. **Table 4** indicates that the different seismicity nodes and the different fluid geochemistry nodes have the same effect on the query nodes of eruption. Thus, there could be fewer nodes without loss of information.

For eruption forecasting, the main factor governing the number of BN nodes is the number of variables being monitored and observed. Under crisis conditions, these might comprise four or more data streams; typical examples are: transient or tremor seismicity, such as conduit-related or spatially diffuse VTs and LF/hybrid seismicity; geodetic cGPS deformation for deep pressurization, or EDM distance measurements, tiltmeters, or radar techniques for shallow/surface movements; gas fluxes, from DOAS, COSPEC; thermal and remote sensing images; geophysical measurements such as gravity, resistivity; and hydrological measurements (see e.g., Science Advisory Committee for the assessment of the hazards risks associated with the Soufrière Hills Volcano Montserrat, 2005; Aspinall and Woo, 2014). If, however, the BN needs to reflect time-varying states, the number of nodes overall can easily exceed 12 or more (e.g., Hincks et al., 2014); in this regard, much depends on how close the volcanic system behavior is to stationary. With modern computers, the processing burden for a large BN is not critical and, when data are plentiful, it is feasible to test how much information each node contributes: it generally makes sense to omit nodes (or links) which do not add much information to the outcome or decision. For example, the UNINET BN program (available at: http://www.lighttwist.net/wp/) can automatically eliminate links if an individual conditional correlation is below some threshold. Similarly, tests of a BN for lahar probability on Montserrat, which re-learned its correlations at each time step starting from time zero, stabilized after about 10 lahar events had been observed (T. Hincks, pers. comm.).

The other key factor, which unquestionably will influence the tractability of a complex BN in a crisis application, is whether there is a person available with the time and knowledge to run the BN and communicate the results. In principle, a volcano observatory should have someone who continually monitors hazard and risk levels, just as colleagues monitor physical variables such as seismicity and fluid geochemistry in real time.

#### BN Modeling in Volcano Monitoring

Our pilot project and several recent publications of BN applications in volcano monitoring and volcanic multi-hazard assessment (Aspinall and Woo, 2014; Cannavò et al., 2017; Sheldrake et al., 2017; Tierz et al., 2017) clearly demonstrate the usefulness of BN modeling for different aspects of volcano monitoring, including eruption forecasting. We briefly focus below on three areas that are important for future BN model development for eruption forecasting: continuous variables, dynamic system and expert involvement.

#### Continuous Variables

One of the main challenges in the model development of the pilot study was defining thresholds for the states of each node, because the experts had reservations about representing continuous observable data with just two states in the BN model. The answer to this challenge is working with continuous variables to reflect the nature of most monitoring data. However, there is less software available for this task, and more importantly most modeling techniques for continuous variables assume joint normality of the data. The assumption of joint normality seems valid on Mount Etna, Italy (Cannavò et al., 2017), where the volcanic activity appears much more regular than most volcanoes around the world. However, the upper tails of many volcanological datasets exhibit dependencies upon each other. For example, while observing an extreme value of any one particular variable is unlikely, if one variable, such as seismicity, is very high, it is more likely that another, e.g., gas flux, is also high. Thus, the assumption of joint normal properties cannot properly represent tail dependence (e.g., Joe, 2014). One possible solution is to condition the BN on a period of unrest and only consider the distribution of variables for that time period.

We have started to work on a BN model to answer the question whether a period of unrest will lead to a large eruption for New Zealand volcanoes. We recommend investing time and effort into developing BN modeling techniques for eruption forecasting to expand the availability of tools for future hazard and risk assessments.

#### Dynamic System

Volcanoes are dynamic systems with magma undergoing profound changes in physical properties during ascent prior to an eruption (Sparks, 2003). As a consequence, the relationship between system variables may change considerably over time. Dynamic BNs can model sequences of events with changing dependencies. Basic applications in volcanology include Aspinall et al. (2006) and a simplified time-stepping BN used by Aspinall and Woo (2014), while Hincks et al. (2014) outline the potential for developing models that include temporal relationships between nodes of a volcanological dynamic BN. We recommend further exploring the use of dynamic BNs to get a better understanding of the volcanic processes.

#### Stakeholder and Expert Involvement

Volcano experts assess crisis situations and advise on what might happen next. We do not intend for BNs or any other quantitative tools to replace the role of a volcano monitoring team. However, quantitative decision-support tools are crucial to provide a reproducible, transparent, and documented framework for giving advice during a crisis.

It is important for the experts who use quantitative models to understand model limitations. Therefore, we recommend that the wider volcano monitoring team is involved in the model development and that a BN model owner, who orchestrates the development of and maintains the model when it is complete (or hands it over to another to run for repeated analyses), is identified.

In the future, we recommend engaging with stakeholders beyond the volcano monitoring team during all stages of the model development for wider model buy-in.

Despite efforts to build a library of worldwide volcanic unrest data (Newhall et al., 2017), future BN model developments for volcano hazard and risk assessments are likely to involve expert judgment due to limited data on most volcanoes, even in well-monitored volcanic centers, such as those in New Zealand. We recommend following structured expert elicitation procedures (e.g., Hanea et al., 2016). We also recommend weighting experts according to their ability to quantify topicspecific uncertainties based on seed items with known values (e.g., Colson and Cooke, 2017). In the pilot study we did not use weighting because we did not have the resources to develop calibration questions appropriate for target questions on conditional probabilities. It will be easier to develop appropriate calibration questions for continuous BNs when asking for uncertainties associated with physical data, rather than with probabilities.

We strongly encourage discussion among experts to share their knowledge and understanding of the volcanic system in an environment conducive for open discussion yet. At the same time, the elicitation procedure should allow experts to express their own scientific beliefs without peer-group or institutional pressures.

When developing a model for a volcano monitoring team, we recommend the inclusion of external experts for an outside perspective. This additional logistical effort and requirement of resources will have to be balanced carefully with the purpose of the model.

For the development of the model structure, it would be useful to select an interested available representative of each sub-discipline. The representative experts can then explain their rationale about the chosen variables and their relationships. Unfortunately, there is little existing guidance on qualitative expert elicitation, such as the development of the model structure. Research into procedures for this is warranted.

#### CONCLUSIONS

We have started exploring the potential of BN modeling for eruption forecasting in New Zealand. While our pilot study of a discrete model to forecast eruptions on Whakaari did not yield a tool ready for application, it provided substantial benefits to the science team involved. In particular, the development of the model structure allowed experts from different sub-disciplines to share their respective understanding of the mechanisms and processes leading to eruption. The simplified graphical presentation of the volcanic system highlighted assumptions that were made by individual sub-disciplines but not necessarily widely appreciated before.

We have found that BNs offer a flexible framework to address many questions in volcano monitoring and volcanic hazard and risk assessment. We anticipate that the BN approach will become essential for handling ever-burgeoning observations and amounts of monitoring data, and indispensable for assessing their evidential meaning for operational eruption forecasting. Further research into these techniques, in particular continuous and dynamic modeling, will extend the scope for useful applications.

Our four key recommendations for future work are: (1) building causal BN sub-models that allow experts from different sub-disciplines to express their knowledge and understanding of particular volcanic processes, which can then be combined into an over-arching volcanic system network; (2) applying BN modeling techniques for continuous variables that more naturally reflect volcano-monitoring data; (3) carefully choosing the number of variables to be modeled; and (4) using robust methods, including structured expert judgment.

Moreover, it is vital that time and effort is invested in developing any BN forecast decision support tool well before the next volcanic crisis starts. Once unrest data is coming in, attention should be given to updating the quantitative aspects of the network with the data observed.

#### ETHICS STATEMENT

This study was carried out in accordance with the recommendations of Code of Ethical Conduct, Massey University, New Zealand. The study was evaluated to be of low-risk of harm to the participants. Consequently a Low Risk Notification was submitted to Massey University. All participants gave written informed consent.

#### AUTHOR CONTRIBUTIONS

AC initiated and led the pilot project. AC, ND, and LC were part of the initial team that adapted the structure of La Soufrière volcano, Guadeloupe, to Whakaari, New Zealand and drafted the first questionnaire to elicit the conditional probability tables. NF was the official project owner; he also helped revise the model structure and advised on the experts to include in the elicitation. AC organized and facilitated the expert elicitation workshop, where ND, LC, and NF were three of the 11 participating experts, and ND facilitated an exercise on calibration questions. AC processed the results, compiled the report of preliminary results, and led the writing of the manuscript. WA provided guidance

#### REFERENCES


and support throughout the study. AH reviewed the preliminary results and guided the development of an initial continuous BN for New Zealand, with LC and NF being the key contributors. All authors contributed to writing of the manuscript.

#### FUNDING

The pilot project was supported by GNS Science Strategic Science Research funding and by GNS Science Core Research Programme funding.

#### ACKNOWLEDGMENTS

We thank Gill Jolly for setting up the initial team and getting the pilot project started. Rob Buxton and Sally Potter were part of the initial team and we thank Sally for overseeing the ethics application process. Katy Kelly and Kat Hammond helped with setting up the questionnaires and administrative support for workshop organization and report formatting. Art Jolly, Steve Sherburn, Bruce Christenson, and Tony Hurst revised the initial model structure and Tony helped describing all variables for workshop notes. We thank our 11 experts at the elicitation workshop for their essential and constructive contibutions: LC, ND, NF, Tony Hurst, Art Jolly, Geoff Kilgour, Jan Lindsay, Agnes Mazot, Jon Procter, Brad Scott and Steve Sherburn. Without the support and enthusiasm of the volcano monitoring team to explore BN modeling the study would not have been possible. We thank Brad Scott for providing information on Whakaari for completing the manuscript. We are grateful to David Rhoades and Ken Gledhill, who reviewed the manuscript before submission. We thank an anonymous reviewer, Stephen McNutt, Pablo Tierz and the associate editor John Stix for their valuable comments, which improved the manuscript.

#### SUPPLEMENTARY MATERIAL

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/feart. 2018.00211/full#supplementary-material Supplementary material provides more detail on "Quantifying the pilot model" and includes the workshop notes and the questionnaires.

urgent decision support under uncertainty. J. Appl. Volcanol. 3, 1–12. doi: 10.1186/s13617-014-0012-8


CO2 in a saline aquifer. Int. J. Greenhouse Gas Contr. 51, 317–329. doi: 10.1016/j.ijggc.2016.05.011


**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 Christophersen, Deligne, Hanea, Chardot, Fournier and Aspinall. 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.

## Temporal Changes of Seismic Velocity Caused by Volcanic Activity at Mt. Etna Revealed by the Autocorrelation of Ambient Seismic Noise

#### Raphael S. M. De Plaen1,2 \*, Andrea Cannata<sup>3</sup> , Flavio Cannavo' <sup>4</sup> , Corentin Caudron<sup>5</sup> , Thomas Lecocq<sup>6</sup> and Olivier Francis <sup>1</sup>

<sup>1</sup> Faculté des Sciences, de la Technologie et de la Communication, University of Luxembourg, Luxembourg, Luxembourg, <sup>2</sup> Centro de Sismología y Volcanología de Occidente, Centro Universitario de la Costa Sur, Universidad de Guadalajara, Puerto Vallarat, Mexico, <sup>3</sup> Dipartimento di Scienze Biologiche, Geologiche e Ambientali - Sezione di Scienze della Terra, Università degli Studi di Catania, Catania, Italy, <sup>4</sup> Etnean Observatory, Istituto Nazionale di Geofisica e Vulcanologia, Catania, Italy, <sup>5</sup> Univ. Grenoble Alpes, Univ. Savoie Mont Blanc, CNRS, IRD, IFSTTAR, ISTerre, Grenoble, France, <sup>6</sup> Royal Observatory of Belgium, Brussels, Belgium

#### Edited by:

Zhong Lu, Southern Methodist University, United States

#### Reviewed by:

Matthew Haney, Alaska Volcano Observatory (AVO), United States Valerio Acocella, Università degli Studi Roma Tre, Italy

> \*Correspondence: Raphael S. M. De Plaen raphael.deplaen@gmail.com

#### Specialty section:

This article was submitted to Volcanology, a section of the journal Frontiers in Earth Science

Received: 08 March 2018 Accepted: 21 December 2018 Published: 17 January 2019

#### Citation:

De Plaen RSM, Cannata A, Cannavo' F, Caudron C, Lecocq T and Francis O (2019) Temporal Changes of Seismic Velocity Caused by Volcanic Activity at Mt. Etna Revealed by the Autocorrelation of Ambient Seismic Noise. Front. Earth Sci. 6:251. doi: 10.3389/feart.2018.00251 On active volcanoes, ambient noise-based seismic interferometry can be a very useful monitoring tool as it allows to detect very slight variations in seismic velocity associated with magma transported toward the surface. However, the classical cross-station approach occasionally fails to detect seismic velocity changes related to eruptive activity, even on very active, well-instrumented volcanoes such as Mt. Etna. In this work, we explored an improved ambient noise-based monitoring strategy by performing the autocorrelation of seismic noise recorded at Mt. Etna volcano, by three stations located close to the active summit craters, during April 2013–October 2014. Such an interval was chosen because of the number and variety of eruptions. In place of the classical cross-correlation, we implemented the phase cross-correlation of each component with itself, which does not require normalization of the signals. The detected seismic velocity variations were very consistent for all three stations throughout the study period, mainly ranging between 0.3 and −0.2%, and were time-related to both sequences of paroxysmal eruptions and more effusive activities. In particular, we observed seismic velocity decreases accompanying paroxysmal eruptions, suggesting an intense pressurization within the plumbing system, which created an area of extensional strain with crack openings. In addition, seismic velocity variations over time were analyzed in the light of ground deformation data recorded by GPS stations and volcanic tremor centroid locations and displayed a particularly strong correlation with the former. Finally, we showed that, although the investigated frequency band (1–2 Hz) contained most of the volcanic tremor energy, our results did not indicate a particular contamination of seismic velocity variation measurements by variations of tremor sources. Ultimately, our investigation highlights a better way to implement noise-based seismic monitoring techniques. The near-field sensitivity of the autocorrelation helped improve our understanding of the relationship between variations of seismic velocity, ground deformation and the pressurization dynamics of volcanic plumbing systems which, in turn, allows for better monitoring implementations of seismic interferometry on other volcanoes.

Keywords: volcano monitoring, ambient seimic noise, autocorrelation, Mt. Etna volcano, seismic noise interferometry, volcanic tremor, seismic velocity variation

#### INTRODUCTION

Magma intrusion and pressurization of the plumbing system can change the seismic velocity structure of a volcano (e.g., Brenguier et al., 2008, 2016). Sometimes, such variations can be detected by classical tomography, as during the March 2009 eruption of Redoubt Volcano, where the presence of a mobile phase at shallow depth beneath the volcano was linked to a reduction in S-wave velocity (Kasatkina et al., 2014).

However, the variations in seismic velocity associated with transport of magma toward the surface are often too small to be detected by tomography techniques (e.g., Brenguier et al., 2016). In these cases, seismic interferometric techniques have turned out to be very effective. In particular, the ambient noise interferometry, based on the fact that the Earth is not static but permanently vibrating, has many advantages (e.g., Brenguier et al., 2016): (i) it has a very high precision, much higher than classic tomography techniques (it is able to detect velocity variations much lower than 1%; e.g., Brenguier et al., 2016; Donaldson et al., 2017); (ii) unlike coda wave interferometry applied on repeating earthquakes (e.g., Cannata, 2012; Hotovec-Ellis et al., 2014), ambient noise interferometry allows a continuous monitoring of the medium velocity changes; (iii) it is a non-invasive or destructive monitoring method. For these reasons, measurement of seismic velocity by ambient noise interferometry is a promising tool to monitor volcanoes (e.g., Brenguier et al., 2008; Duputel et al., 2009). For example, Donaldson et al. (2017)showed a remarkable correlation between relative velocity and the radial tilt measured at K¯ilauea summit. This suggested that during inflation phenomena, since the volumetric strain is extensional above the inflation source and compressional at the sides, most of the edifice is dominated by compression closing cracks, producing higher seismic velocities, while the extension area above the source is dominated by opening pores and crack, producing lower seismic velocities.

Cross-correlation functions, used to perform interferometry, are typically reconstructed from the signals acquired by pairs of stations, allowing to continuously and accurately monitor the temporal changes in seismic velocity (Hadziioannou et al., 2009). More recently, implementations of this technique used passive recordings at individual stations to detect changes of seismic velocity in the crust by cross-correlating each component of each individual station with itself, or autocorrelation (e.g., Hobiger et al., 2014; De Plaen et al., 2016; Yukutake et al., 2016).

In this work, we implemented the autocorrelation of ambient seismic noise using the phase cross-correlation to monitor the activity of Mt. Etna volcano during April 2013—October 2014. Such a time period was chosen because of the number and variety of eruptions taking place on the volcano summit area. In order to better understand the source mechanisms of the detected medium changes, the seismic velocity patterns are compared with ground deformation, volcanic tremor, and volcanological data.

#### BACKGROUND

Mt. Etna lies at the convergence between the African plate and the Eurasian plate. Located near the second most populated city of the Sicily, Catania, it is a very active volcano with a series of brief but powerful eruptive episodes (paroxysms) almost every year at least since 2011. It is in an almost constant state of activity, and has documented records of historical volcanism dating back to 1500 BCE. The current activity ranges from effusive to Strombolian, to paroxysmal activities.

Since January 2011, a new cone grew up at the summit area of Mt. Etna, following a series of paroxysms (e.g., Behncke et al., 2014; De Beni et al., 2015). Informally named the New South East Crater (NSEC, **Figure 1**), the cone was built up of material accumulated from lava fountains.

After 10 months of quiescence, the NSEC resumed its episodic activity in February 2013, producing 19 paroxysms until early December (e.g., Cannata et al., 2015; Spampinato et al., 2015). In particular, 13 paroxysms took place from February to April (E26-E38 in **Table 1**; the nomenclature of the episodes derives from De Beni et al., 2015), and six from October to December (E39-E44 in **Table 1** and De Beni et al., 2015). All the paroxysms at NSEC showed similar features. In particular, the following phases of activity were observed (Behncke et al., 2014): (i) minor explosive activity, (ii) more vigorous Strombolian activity often accompanied by lava discharge, (iii) lava fountaining with voluminous ash ejection and lava flows, and (iv) a waning phase with transition to mildly Strombolian activity and the end of the eruption.

In mid-December 2013, the activity transitioned from brief and violent paroxysms, to long-lived and less explosive episodes (De Beni et al., 2015). The initial explosive activity (phase "i" and "ii") did not culminate in sustained lava fountaining and voluminous tephra emission. This type of activity largely continued in 2014. The main eruptions took place in January-April and June 2014 at NSEC (E47-E48 in **Table 1** and De Beni et al., 2015).

Successively, a 5-week-long period of intense Strombolian and effusive activity occurred in July-August 2014 (E49 in **Table 1** and De Beni et al., 2015) from a fracture, located ∼1 km north of the NSEC, on the eastern flank of the North-East Crater (NEC) (e.g., Spina et al., 2017), the highest of Etna's

Crater; VOR, Voragine). The purple dashed line is the approximate extent of the E49 fracture system. The stations are <5 km apart and station ECPN is the closest to

summit craters (which erupted for the first time in 1911; Ponte, 1920). This activity was probably linked to the NSEC plumbing system (e.g., Viccaro et al., 2016), although it did not contribute to the evolution of the NSEC cone. Before the end of this eruption, on 8 August, explosive activity resumed at NSEC, and

culminated on 11–14 August with vigorous Strombolian activity

NSEC, the most active crater for the studied period.

## DATA AND METHOD

(E50 in **Table 1** and De Beni et al., 2015).

We used data from the Etna seismic network recorded during April 2013—October 2014. The seismic stations consist of broadband three-component Trillium seismometers (Nanometrics), with cut-off period of 40 s and sampling rate at 100 Hz. In particular, we used seismic signals from three stations closest (5 km) to the active craters: ECPN, EPDN, and EPLC (**Figure 1**). This period covered a series of eruptive episodes linked primarily to the NSEC (**Table 1**), with a mixture of brief and violently explosive activity in 2013, and intense Strombolian and effusive activity in 2014.

Seismic velocity changes were measured from seismic noise cross-correlation, following a workflow similar to Lecocq et al. (2014) and De Plaen et al. (2016). Seismic records for all components were pre-processed by carefully checking for their timing and gaps. Then, they were pre-filtered between 0.01 and 8.0 Hz, and resampled to 20 Hz.

The cross-correlation step commonly uses the classical crosscorrelation:

$$C\_{cc}\left(t\right) = \sum\_{\mathbf{r}=\mathbf{r}\_0}^{\mathbf{r}\_0+T} u\_1\left(t+\mathbf{r}\right) u\_2\left(\mathbf{r}\right) \tag{1}$$

where u<sup>1</sup> and u<sup>2</sup> are the cross-correlated seismic traces, t is the lag-time, τ the lag-time shift and T is the correlation window length. This step typically requires to normalize the signal in the time and the spectral domains to suppress high amplitude events and isolated noise sources with defined frequencies (Bensen et al., 2007).

Since we correlate each component with itself, normalizing in the spectral domain by setting the amplitude to 1 for all frequencies removes information on the medium (Hobiger et al., 2014). The contamination of the autocorrelation of ambient seismic noise by high amplitude events, such as earthquakes, is a known issue (e.g., De Plaen et al., 2016). To reduce this vulnerability, we implemented the phase cross-correlation (PCC) of Schimmel (1999) and Schimmel et al. (2011) instead of the classical correlation.

The PCC measures the similarity of the instantaneous phases of the analytic traces and is therefore amplitude unbiased. The main advantage of this property is that it removes the need for temporal or spectral normalizations before the cross-correlation.

$$C\_{pcc}(t) = \frac{1}{2T} \sum\_{\mathbf{r}=\mathbf{r}\_0}^{\mathbf{r}\_0+T} \left\{ \left| e^{i\varphi(t+\mathbf{r})} + e^{i\gamma(\mathbf{r})} \right|^\vartheta - \left| e^{i\varphi(t+\mathbf{r})} - e^{i\gamma(\mathbf{r})} \right|^\vartheta \right\} \tag{2}$$

where γ (τ ) and ϕ (τ ) are the instantaneous phases of the traces u<sup>1</sup> and u2, respectively and the power ϑ controls the sensitivity and the signal-to-noise ratio.

The autocorrelations were computed for each individual day following a bandpass filter between 1.0 and 2.0 Hz. Then, the autocorrelation functions were averaged with a 2-day linear


TABLE 1 | Chronology of the main eruptive activities at Mt. Etna volcano in 2013–2014, with the exception of the 28 December 2014 eruption.

Collected from Bonforte and Guglielmino (2015) and De Beni et al. (2015). Our analysis starts after E33.

stacking to improve the signal-to-noise ratio, while keeping a high temporal resolution.

The fluctuations of seismic velocity in the medium were obtained by comparing each daily autocorrelation to a reference autocorrelation corresponding to a stack of the entire study period. Assuming a homogeneous change in the medium, the relative difference in travel time dt is related to the change in the seismic velocity dv, as:

$$-\text{ dt/t = }d\text{v/v}\tag{3}$$

Relative travel time changes were measured in the frequency domain using the moving-window cross-spectral analysis (MWCS) which has the advantage of clearly defining the bandwidth of the coherent signal (Ratdomopurbo and Poupinet, 1995). Time delays for each window between the two autocorrelations are in the unwrapped phase of the crossspectrum. They were measured from the slope of a linear regression of the samples within the pre-defined frequency band, weighted by the cross-coherence between energy densities in the frequency domain (Clarke et al., 2011). A quality control was obtained from errors, estimated using the cross-coherence values and the squared misfit to the modeled slope. MWCS point measurements with a dt error >0.1 s and a coherence under 0.6 were considered of poor quality and rejected. Variations of seismic velocity were measured on both the causal and the acausal part of the autocorrelation for time lags between 5 and 35 s to prevent direct wave contamination. The velocity variations for each component pair were ultimately averaged by station.

#### RESULTS

#### Station-To-Station Comparison

The seismic velocity variations were globally very consistent for all three stations throughout the study period, with the exception of a minor phase delay for ECPN right before and during some eruptive episodes (E40-E43, **Figure 2**). The few gaps were caused by interruptions in the original data or poor quality measurements that were rejected (July and August 2014, for example). The dataset started with an eruption (E34 on 3 April 2013), for which we therefore have no observation of precursors. A decrease in seismic velocity is observed at all three stations at the same time as the October-December 2013 paroxysmal sequence (E39-E44) and prior to the effusive/Strombolian activities of mid-June, July and August 2014. By contrast, the mainly effusive eruption (E47) initiated at a cluster of vents at the eastern base of the NSEC cone in late January 2014 was preceded by a consistent increase of seismic velocity that coincides with a change in style of activity from paroxysmal to mainly effusive.

FIGURE 2 | Seismic velocity variations at Mt. Etna in 2013-2014. Each line corresponds to the average at each individual station (EPLC, ECPN, and EPDN) of the measured relative seismic velocity variation (dv/v, in %) for each component pair (e.g., ZZ, EE, NN). These dv/v are calculated using the autocorrelation in the 1–2 Hz frequency band, and the cross-correlation functions are stacked linearly over 2 days. The median errors on the dv/v measurements are ∼0.03% for each station. The orange vertical lines and zones are labeled eruptive episodes described in Table 1. The episodes were brief in duration and rather explosive (vertical lines) until E45, in the second half of December 2013, when the activity changed to more long-lived and effusive events that characterized 2014. The purple zones in the top subplot illustrate the timespans enlarged in Figures 3, 4.

**Figure 3** offers a closer view at the October-December 2013 paroxysm (E39-E44). This paroxysm sequence was preceded/accompanied by a clear decrease detected at all stations and was characterized as one of the most violent in the series of paroxysms since January 2011 (De Beni et al., 2015). The decrease in seismic velocity of over 0.1% persisted at all three stations until 11 November (E40), then simultaneously increased at stations EPDN and EPLC. The seismic velocity at station ECPN, however, decreased through another 5 eruptions (E40- 44) until mid-December. As aforementioned, in mid-December the activity at the NSEC transitioned from more to less explosive. We observe that in the meantime seismic velocity increased at all three stations through E45 and E46 to recover previous values later in January 2014. No good reliable seismic velocity measurement could be done during E47 except for station EPLC.

The ground deformation time series (**Figure 3B**) were obtained from the measurement of daily baseline ECPN-EPDN, whose variations over time are sensitive to any change in summit deformation. Most oscillations of baseline length are related to inflation and deflation cycles preceding and accompanying the eruptions, respectively. During the October-December 2013 paroxysm (E39-E44), the baseline length globally shortened (∼-0.8 cm) until early December. It then increased until E47 with the exception of a ∼0.3 cm drop after E46. At a shorter time scale, subtle increase in baseline length preceded episodes E39, E40, E42, E45, and E46. In the course of episode E47, anticorrelated oscillations of seismic velocity changes at EPLC and

baseline length ECPN-EPDN were observed (**Figure 4**, dashed vertical lines).

activity on the eastern flank of the NSEC.

The following activities (E48-50, **Figure 4**) were also clearly preceded by a decrease in seismic velocity at every station. During these longer episodes (E49 and E50 for example), many seismic velocity change measurements were rejected due to high level of errors in dt measurement, although a global increase or recovery could, to a certain extent, be identified. Again, the variation of baseline length between GPS stations ECPN and EPDN (**Figure 4B**) was anti-correlated with the seismic velocity change. The ECPN-EPDN baseline length globally increased until E48, then decreased through E49 and E50 to stabilize after mid-August.

After the recovery, the seismic velocity measurements during E49 and E50 did not provide results of a satisfactory quality (with a coherence under 0.6), causing gaps for all three stations. In the meantime, the correlation coefficient displayed two significant drops for each episode (**Figure 4C**). These episodes were followed by a recovery until late October when stations EPDN and EPLC displayed a new decline of seismic velocity (>0.2%) in early November, only to rise again in December (**Figure 2**). Station ECPN appeared to be relatively stable following E50 recovery (**Figure 2**). These last variations could be related to the powerful 28 December 2014 eruptive episode.

FIGURE 4 | Seismic velocity variation at stations EPCN, EPLC, and EPDN (A), variation in baseline length between GPS stations ECPN and EPDN (B) and correlation coefficient between the daily and the reference Cross-Correlation Function for each component pair at station ECPN (C) between April and October 2014. The green line is the exponentially-weighted moving average with a center of mass of 5 (EWMA) of the baseline variation. Following the Strombolian and effusive activity of E47, a consistent decrease of seismic velocity is observed until E48, also characterized by Strombolian activity and lava flows. An increase is then observed until E49. No reliable seismic velocity change measurement could be done during E49 and E50.

#### Variation of Volcanic Tremor Source Location

The seismic energy in the chosen frequency band is dominated by volcanic tremor, which commonly rises concerns of contamination when seismic interferometry is used for volcano monitoring (e.g., Ballmer et al., 2013). The volcanic tremor centroids (**Figures 5**, **6**) were obtained in near real time at Mt. Etna from 1 h long sliding windows that tracked the spatial evolution of the source (e.g., Di Grazia et al., 2006; Cannata et al., 2013).

In order to locate volcanic tremor sources, we assumed propagation of body waves in a homogeneous medium, and applied a grid-search method based on the spatial distribution of seismic-amplitudes. The body waves assumption is related to the fact that the location result is mostly affected by the seismic amplitudes of the stations closest to the tremor source. As for the grid, we consider a 8 × 8 × 6 km<sup>3</sup> volume with a spacing between nodes of 250 m. The investigated frequency band was 0.5–2.5 Hz, routinely analyzed to highlight the main migrations of the tremor centroid (Patanè et al., 2008; Cannata et al., 2013). The average location errors, calculated by the jackknife method (Di Grazia et al., 2006; Cannata et al., 2013),

were equal to 0.3, 0.4, and 0.7 km for longitude, latitude, and altitude, respectively. Since more than one source can be active simultaneously, the location corresponds to the dominating source or a location between the real sources. During eruptive episodes, the amplitude of volcanic tremor commonly increases and a migration of centroids toward the surface near the active vent can be observed. For example, this migration is illustrated in **Figures 5**, **6** by episode E49 during which, unlike the rest of the study period, the volcanic activity and the tremor centroids mainly focused on the eastern flank of NEC. Between eruptive episodes, centroids migrate back at depth (1–2 km above sea level) and under the Voragine Crater (**Figures 5**, **6**). These differ from the observed variations of seismic velocity. For example, the continuous decrease between E47 and E48 does not indicate a relationship with the corresponding location of tremor sources.

Nevertheless, at times before or during an eruptive episode the temporal variations of volcanic tremor centroids location are correlated with variations of seismic velocity, positively or negatively (**Figure 6**). **Figures 7**, **8** show the spatial interpolation for variation of seismic velocity and the location of tremor source centroids over 5 days before E47 and E49, respectively. The 5 days span offers a compromise between long- and short-term variations to keep the ability to emphasize shallow mechanisms that would cause sudden variations before an eruption. Although episodes E47 and E49 developed on different craters—NSEC and NEC, respectively—the tremor source centroids clearly cluster near the eruptive site, in the vicinity of the observed largest decrease of seismic velocity.

#### DISCUSSION

Although the stations used in this study are all within 5 km of the active craters, station ECPN which is the closest to the NSEC sometimes shows a larger peak-to-peak decrease before eruptions (e.g., E39 on **Figure 3A**) except when the

changed into more effusive and longer. The horizontal dashed lines are the locations of volcano summit craters (NEC; North-East Crater; NSEC, New South-East Crater; VOR, Voragine). All three craters are within 500 m in longitude. Spectrogram of the vertical component of station ECPN (D). The white frame is the frequency band used in this study.

activity migrated to the more distant NEC (E49 on **Figure 4A**). This observation indicates a local sensitivity of the method in contrast with the assumption used to average the relative velocity changes measured over all station pairs in a network of a global homogeneous velocity change in the whole volume of the edifice (e.g., Obermann et al., 2013; Rivet et al., 2015). Averaging results with measurements made at distant stations or station pairs could reduce the capability of precursor detection. This is illustrated by Cannata et al. (2017) who did not observe volcanic precursors in variations of seismic velocity obtained from a large number of stations on Mt. Etna. Ultimately, when pairs of receivers are used, a careful selection of the station pairs mostly affected by the source of change is therefore recommended. When autocorrelation is used, measurements prove to be sensitive to changes in the near field, stations close to the eruptive site should therefore be favored.

A criticism against monitoring using ambient seismic noise cross-correlation is the potential sensitivity of the seismic velocity variations to variations in source distributions and concerns of contamination from volcanic tremor (e.g., Ballmer et al., 2013).

However, although volcanic tremor sources are not suited to calculate Green's functions and to perform imaging studies, they can be used to estimate seismic velocity variations over time, provided that consistent coda arrivals can be reconstructed (e.g., Donaldson et al., 2017). On one hand, using distant parts of the coda of the autocorrelation assumes that the measured stable phases are highly scattered and therefore unaffected by variations of source distributions (e.g., Stehly et al., 2008). On the other hand, changes in frequency content of the noise source have a limited effect on our measurements, since we used the phase cross-correlation, which is amplitude unbiased by design, and the MWCS analysis, in which the amplitude and phase spectra are separated before the measurements are made.

Ultimately, the inconsistent correlation between tremor centroids and seismic velocity variations is an evidence of

a discrepancy between the two observations. A relationship between them is mostly observed at the onset of eruptive episodes. Viccaro et al. (2016) showed that changes in time of volcanic tremor centroids are also strongly linked to the variations in ground deformation patterns using another baseline crossing the summit area. This relationship between tremor source location and seismic or geophysical observations is probably the consequence of the pre-eruptive pressurization which specifically influences all these observations.

Some seismic velocity change measurements were rejected due to high errors and low coherence at the MWCS step during eruptive episodes. Observations of the cross-correlation coefficient between the current and the reference autocorrelation illustrate how the waveforms are dramatically changing during eruptions, with potentially new phases appearing and disappearing until the end of the activity. This decorrelation is typically assumed to be caused by the generation of new scatterers in the medium during the eruption (e.g., Larose et al., 2010; Brenguier et al., 2011; Obermann et al., 2013). In such a case, the observed mechanism is not identified as a change of seismic velocity as the medium is assumed to be significantly

damaged. Since variations of tremor sources are sometimes concomitant with the observation of decorrelation, we cannot entirely preclude some impacts of active sources creating new arrivals in the coda of the autocorrelation function. Nevertheless, the decision to use late part of the coda, which correspond to longer travel times, should reduce such a limitation. Besides, a significant part of the variations in tremor sources occurred just before eruptions and, although contamination from active sources would be expected at such times, these variations did not prevent the observation of precursors from seismic velocity variations. Furthermore, the reliability of these variations of seismic velocity was corroborated by the correlation with ground deformation (**Figures 3**, **4**).

at station ECPN. (B) Seismic velocity variation in % at the considered stations.

Our results give an exclusive insight into the dynamics and the development of explosive and effusive eruptions, specifically when looking at the remarkable link between seismic velocity variation and ground deformation. Several existing studies have associated variations of seismic velocity and cycles of inflation and deflation with changes in magma pressurization (e.g., Brenguier et al., 2008; Bennington et al., 2015; Hotovec-Ellis et al., 2015; Donaldson et al., 2017). The compression associated with an increase in magma pressurization is expected to close pores and cracks in the medium, causing an increase of the elastic moduli and the seismic velocity, along with a potential inflation at the surface (O'Connell and Budiansky, 1974). Both decreases and increases of seismic velocity have been attributed to magma pressurization as a volcano inflates. For example, Brenguier et al. (2008) explained precursory seismic velocity drops measured before several eruptions at Piton de la Fournaise volcano with a dilatation and an opening of cracks at the edifice surface. In contrast, complex patterns of decrease and increase of relative velocity induced by magma migration were observed at Merapi Volcano (Ratdomopurbo and Poupinet, 1995; Wegler et al., 2006; Budi-Santoso and Lesage, 2016). Donaldson et al. (2017) reconciled these observations by connecting the change of seismic velocity to the depth of the source of deformation. The inflation of the pressure source creates an area of extensional strain with cracks opening right above the source and a surrounding area of compressional strain with cracks closing. The deeper the source, the larger the area of extension and, as the source migrates toward the surface, this area of extension becomes smaller.

This mechanism is illustrated in our study by the seismic velocity variation measured throughout the sequence of explosive episodes from October to mid December 2013. All three stations display a consistent decrease in seismic velocity until E40 in November. Station ECPN, closest to the erupting crater, displays a persistent decrease until the end of the sequences while the other two stations start increasing. The explosive nature of this sequence is an indication of a proportional intense pressurization of the volcanic system, which fed the explosive activity until mid-December. The discrepancy in relative seismic velocity between ECPN and the other two stations after E40 would indicate a source of inflation that migrated toward the surface and a smaller associated extensional area which only impacts the closest station.

As the pressurization of the system decreased, the activity at the NSEC then evolved to a less explosive style, probably associated with a reduced pressurization of the system, and the seismic velocity was eventually restored to previous values. This recovery could also explain why no station displayed any significant precursor before the long-lasting episode E47 that started on 22 January 2014.

Our investigation emphasizes how the location of source pressurization in the volcanic plumbing system significantly impacts the measured seismic velocity changes. Existing studies already showed contrasts such as at Piton de la Fournaise and K¯ilauea volcanoes where the measured pre-eruptive change of seismic velocity can, respectively decrease or increase as a result of the depth of sources of deformation and the location of the stations used (e.g., Brenguier et al., 2008; Donaldson et al., 2017). Here, we also explored how lateral variation of the source of pressurization combined with changes in types of activity also required a monitoring strategy that accounts for changes of seismic velocity in the near field. The near-field sensitivity of the autocorrelation allowed us to successfully detect previously undetected pre-eruptive signals several days before most eruptions. The reliability of these early signs of eruption is better assessed in the light of ground deformation measurements. This success offers the opportunity to better implement noise-based seismic monitoring techniques on other volcanoes and, in turn, obtain a more comprehensive understanding of their dynamics when combined with other observables.

### CONCLUSIONS

We implemented the autocorrelation of ambient seismic noise using the phase cross-correlation to monitor the activity of Mt. Etna volcano. Using the continuous signal recorded at three individual, three-component, broadband seismic stations located in proximity to active vents, we successfully retrieved changes in seismic velocity several days ahead of most eruptions and unveiled associated pressurization dynamics within the subsurface. By contrast, passive interferometry averaging pairs of stations failed to unambiguously detect precursors.

All the used stations showed a very coherent evolution throughout the investigation period, despite being at different distances to the eruptive site. However, a discrepancy between the three stations appears at times when the source of pressurization is suspected to be shallower, toward the end of the explosive sequence or before an effusive episode. These different signatures which distinguish explosive and less explosive episodes are better understood in the context of the evolving pressurization pattern of the magma system. This property is particularly useful to study the co-eruptive evolution of a volcano to potentially forecast the end of a paroxysm. The seismic velocity variation was specifically useful with explosive sequences and so was the evolution of the cross-correlation coefficient between the current and the reference autocorrelation function during the effusive sequence, respectively.

Our results are correlated with variations in tremor source before eruptions. Although we do not definitively preclude a possible contribution of co-eruptive seismic sources in the decorrelation observed during volcanic eruption, the combination of the phase cross-correlation and the movingwindow cross-spectral analysis should reasonably mitigate such an influence. Our results, also strongly correlated with ground deformation, did not indicate a particular contamination of seismic velocity variation measurements by variations of tremor sources. Ultimately, our results also prove that the autocorrelation of ambient seismic noise can be used to monitor volcanoes with different types of activity, even on sparse networks.

### AUTHOR CONTRIBUTIONS

RD is the main investigator on this research. He selected and implemented the method for the seismic interferometry and wrote most of the manuscript. AC provided the seismic data, contributed to parts of the manuscript and the tremor sources location. FC provided the baseline changes and contributed to parts of the manuscript. CC participated in the interpretation of the results and the elaboration of the manuscript. TL helped shape the methodology, the interpretation, and the text. OF supervised the work and contributed to the manuscript.

#### FUNDING

This work was funded by the University of Luxembourg.

#### REFERENCES


#### ACKNOWLEDGMENTS

We are indebted to the technicians of the INGV, Osservatorio Etneo for enabling the acquisition of seismic data. AC thanks the project ICE-VOLC (PNRA14\_00011) funded by Programma Nazionale Ricerche in Antartide. Original data used in this manuscript are available from these authors: AC for volcanic tremor data (andrea.cannata@unipg.it), FC for GPS data (flavio.cannavo@ingv.it).

monitoring crustal temporal variations using ambient seismic noise. Geophys. J. Int. 186, 867–882. doi: 10.1111/j.1365-246X.2011.05074.x


with ambient seismic noise. J. Geophys. Res. Solid Earth 118, 6285–6294. doi: 10.1002/2013JB010399


the July 2014 eruption of etna volcano (Italy). Bull. Volcanol. 79, 79–51. doi: 10.1007/s00445-017-1132-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 De Plaen, Cannata, Cannavo', Caudron, Lecocq and Francis. 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.