An eruption forecast is a probabilistic assessment of the likelihood and timing of volcanic activity. The forecast may also include information about the expected style of activity (Section 1.6), the duration of an eruption, and the degree to which populations and infrastructure will be affected (Sparks, 2003). A prediction, in contrast, is a deterministic statement about where, when, and how an eruption will occur, and a prediction will either be correct or incorrect. Short-term forecasts primarily use monitoring data (principally seismic, deformation, heat flux, volcanic gas, and fluid measurements) to detect and interpret periods of unrest, whereas long-term forecasts primarily rely on the geologic record of past eruptions. Long-term forecasts assess eruption potential and hazards over the lifespan of a volcano and are independent of short-term forecasts.
Volcano science has demonstrated undeniable advances in using pattern recognition in monitoring data and the geologic record to anticipate eruptions and make statistical forecasts. Case studies in Boxes 3.1 and 3.2 highlight notable instances of the use of quantitative monitoring data to estimate the timing of future eruptions, but Table 3.1 also points to the challenges involved in forecasting, including some eruption patterns that were not anticipated. It is not straightforward to quantify forecast success, and so the table includes a short discussion of each event.
An alternative and potentially superior approach involves forecasting using physics- and chemistry-based models, such as those discussed in Chapters 1 and 2, informed by monitoring data, an approach used in weather forecasting. Developing such models is a tremendous challenge. At present no single physics-based model can explain the full range of volcanic activity or account for the complexities inherent in volcanic systems. Achieving a paradigm shift, from pattern recognition to model-based forecasting, will require improved constraints on plumbing system geometry and nonlinear material response, and improved understanding of the connections between subsurface processes and monitoring data.
Geophysical and geochemical monitoring data are used to detect unrest and enable short-term probabilistic forecasts based on pattern recognition
in monitoring time series. As an example, the Mount St. Helens post–May 18 eruptions in 1980 were successfully forecast based on patterns of precursory seismic tremor and localized deformation that consistently preceded events (Malone et al., 1981; Swanson et al., 1983). Relatively frequent eruptions of Kilauea volcano, Hawaii, have led to clear seismic and deformation precursors prior to eruptions. However, many other volcanoes, particularly those that erupt violently, have had limited or no historical eruption observations and few quantitative measurements. In these cases, observations from similar volcanoes are
TABLE 3.1 Examples of Forecasts and Missed Opportunities
|Mount St. Helens (United States), 1980||Lateral blast and VEI 5 eruption occurred after a period of elevated seismicity, dramatic dome growth, and phreatic explosions (Lipman and Mullineaux, 1981).||Yes and no. Although unrest was observed for months and led to heightened surveillance, the timing, directionality, and scale of the eruption was not anticipated and 57 individuals perished.|
|Kilauea (United States), 1983–?||Longest eruption sequence (currently ongoing) in Hawaii’s history corresponds to the principal locus of eruptive activity~15 km from the summit on the East Rift. Repeated episodes of lava fountaining and effusion were well monitored. The geodetic network at the summit is used to anticipate downrift eruptive activity with hours of warning (Anderson et al., 2015).||Generally yes as the eruption progressed. Observations led to a viable model, which was used to estimate a high likelihood of events during certain time intervals. The long duration of activity (more than 30 years) was not anticipated.|
|Nevado del Ruiz (Colombia), 1985||Modest eruption spawned large lahars, which was a hazard previously recognized in both historical and geologic observations. Phreatic eruptive activity and elevated volcanic gases pointed to unrest beginning 1 year prior to eruption (Pierson et al., 1990).||No. Although elevated activity spurred volcano study and the development of hazard maps, government agencies were unable to forecast the primary hazard and provide guidance in a timely manner. More than 25,000 deaths resulted.|
|Pinatubo (Philippines), 1991; see also Box 3.1||Largest eruption of the last hundred years. Its potential Plinian eruption size was anticipated based on studies of previous eruption deposits. Eruption evolution was relatively “well behaved,” with seismic precursors, phreatic explosions, tiltmeter inflation, increasing sulfur output, then increasingly violent magmatic eruptions (Newhall et al., 1996).||Yes, in the sense that evacuations (out to 40 km) were issued prior to the paroxysmal event and valuable property at Clark Air Force Base was moved in a timely manner. Still, some 800 people were killed, largely due to roof collapse and lahars.|
|Soufriere Hills (Montserrat), 1995–?||Long-term small to moderate ash eruptions beginning in 1995 were later accompanied by lava-dome growth and pyroclastic flows that forced evacuation of the southern half of the island and ultimately destroyed the capital city of Plymouth, causing major social and economic disruption. To date, there have been four phases of eruption, separated by periods of up to 2 years with no residual surface activity.||Yes and no. Cyclic short-term (6–12 hour) precursors, such as increased seismicity and inflation, were successfully used to anticipate the most dangerous times and the most likely timing of dome-collapse pyroclastic density currents. However, the long-term behavior and ultimate duration of the eruption have been difficult to anticipate.|
|Hekla (Iceland), 2000; see also Box 3.2||An eruption producing an 11-km-high ash plume was accurately forecast (Höskuldsson et al., 2007). Timely notifications were made to the National Civil Defense of Iceland within 20 minutes of unrest and 40 minutes prior to the inferred onset of eruption. The Civil Aviation Administration was also notified.||Yes. A well-monitored volcano presented precursory earthquake activity that was well understood and similar to a previous eruptive episode. The data were used to anticipate an eruption that occurred shortly (~1 hour) after initial unrest.|
|Mount Ontake (Japan), 2014||Unanticipated (small) phreatomagmatic eruption occurred with no recognized warning, despite an extensive monitoring network (Kato et al., 2015).||No. Hindsight analysis indicates subtle anomalies that were not recognized at the time of eruption or that required timeintensive laboratory analyses. 57 individuals were killed.|
|Villarrica (Chile), 2015||Open-vent volcano experienced seismic and infrasonic anomalies and elevated lava lake activity for months leading up to a short-lived paroxysmal eruption. The activity represented the first explosive eruption in more than 30 years.||Yes. Alert levels were incrementally raised by Chilean authorities and tourists were kept away from the hazard zone. The severity and duration of the paroxysmal event was not anticipated.|
|Calbuco (Chile), 2015||Closed-vent volcano, quiet since 1972, erupted suddenly and intensely with a VEI 4 eruption.||No. Anomalous seismicity was noticed prior to eruption. In hindsight, only hours of limited seismicity preceded eruption. The monitoring network was limited.|
NOTE: VEI, Volcano Explosivity Index.
used to inform probabilistic assessment (e.g., Ogburn et al., 2016).
Due to the pervasive lack of robust monitoring data and the limitations of models used to forecast eruptions, volcano monitoring agencies typically issue qualitative alerts. The USGS provides four levels of alerts: normal (or background), advisory (signs of unrest), watch (escalating unrest), and warning (dangerous eruption under way).1 Every active or potentially active volcano in the United States is assigned a value on this scale. Such alert levels can be a starting point for civil authorities to activate plans for alert, evacuation, and shutdown of critical infrastructure, on time scales of hours to months (Aspinall et al., 2003; Punongbayan et al., 1996; Voight, 1988).
A growing array of precursory phenomena signal unrest preceding most eruptions. Technological advances and the expansion of monitoring infrastructure at some volcanoes over the past few decades allow quick detection and interpretation of signals of unrest days to months prior to an eruption (e.g., Ewert et al., 2005). Eruptions at monitored volcanoes generally occur with at least a few hours of warning in the form of anomalous seismicity (Figure 3.1), ground deformation, hydrologic changes, and/or heat flux and gas emissions (Tables 1.1 and 1.2; Aiuppa et al., 2007). Integrated studies that combine gas emission data, ground deformation data, seismic signals, and novel petrologic techniques provide compelling evidence of magma movement months to weeks before some eruptive episodes (Kahl et al., 2013; Figure 3.2). Novel analytical techniques applied to continuous seismic signals have made it possible to detect subtle changes in seismic wave speed (Figure 3.3), interpreted as magma pressurization and ascent in the mid- to shallow crust (e.g., Brenguier et al., 2008; Obermann et al., 2013) prior to eruption, and to track dike propagation in the shallow crust without the dif-
ficulty of locating individual earthquakes during intense activity (Taisne et al., 2011). Other signals that magma is entering the shallow crust include phreatic explosions; swarms of shallow earthquakes; low-frequency earthquakes and volcanic tremor (e.g., Chouet and Matoza, 2013); changes in fluid discharge, chemistry, and temperature (e.g., White and McCausland, 2016); short-wavelength deformation; and SO2/CO2 gas ratios that deviate significantly from the baseline recorded during quiescence (e.g., Figure 3.2). Such signals are thought to indicate an increased likelihood of an eruption (e.g., Moran et al., 2011).
Anticipating eruptions within minutes to hours before they occur is often based on rapidly intensifying or changing indicators of unrest, such as the onset of strong volcanic tremor, ground tilt, or hydrologic changes (e.g., Figure 3.1). For example, repeating earthquakes transitioning into “gliding” volcanic tremor consistently preceded explosions of Redoubt volcano, Alaska, in 2009 and have since been linked to frictional
faulting in a highly stressed region of the conduit (Dmitrieva et al., 2013). Similarly, an exponential increase in the number of earthquakes has been interpreted as an exponential increase in the probability of eruption (Endo and Murray, 1991).
Forecasting the magnitude, style, and duration of an eruption remain major challenges for short-term forecasting. Recent efforts to analyze eruption databases for indicators of eruption volume are promising (e.g., Bebbington, 2014), with the caveat that the record of observed eruptions is biased toward small to moderate events at frequently active volcanoes, whereas the prehistoric geologic record is biased toward larger eruptions (Kiyosugi et al., 2015).
In general, detecting the onset of unrest has been far more successful than anticipating the evolution of a volcanic eruption once it has begun. Future short-term eruption forecasts must become more adept at incorporating disparate geophysical and geochemical data gathered during ongoing eruptions to create ensemble forecasts that anticipate possible changes in eruptive activity. As indicated in Table 3.1, key questions during eruptions concern the likely duration of eruptive activity, the nature of pulsatory or intermittent activity, the significance of a hiatus in eruptive activity (is the eruption over or has it only paused?), and changes in the style of activity (e.g., switch from explosive to effusive activity).
Key parameters in dynamic forecast models include the location, composition, and volatile content of the magma as well as mass fluxes of magma and gases. Once an eruption commences the combination of eruption flux and geodetic data can be used to constrain total magma chamber volume, pressure, and volatile content (e.g., Anderson and Segall, 2013; Mastin et al., 2009a). Potentially, erupted rocks and minerals could be analyzed immediately to provide information on the pressure, temperature, volatile content, and composition of the deep magmatic system feeding the eruption. Active and passive source seismic experiments with high-density coverage will continue to improve four-dimensional imaging of the volcano’s plumbing system potentially in near real time (e.g., Kiser et al., 2013; Ulberg et al., 2014). Finally, remotely measured gas compositions combined with thermodynamic modeling, melt inclusion volatile contents, and solubility data could help constrain magma depth and quantity (e.g., Edmonds et al., 2001; Iacovino, 2015).
Many episodes of unrest do not culminate in eruption and better assessments of the proportion of unrest episodes that end with magma intrusion into the crust are needed (Phillipson et al., 2013). On the other hand, some explosive magmatic eruptions, such as the 2015 VEI 4 eruption of Calbuco, Chile, are preceded by surprisingly little seismicity (Romero et al., 2016). Relatively small explosive eruptions may be triggered when
gas pathways are sealed by formation of a magma plug or precipitation of minerals in the hydrothermal system. This sealing process can be manifested by fluctuations in gas emissions, tilt, or long period seismicity (e.g., Cruz and Chouet, 1997; Fischer et al., 1994; Johnson et al., 2014; Nishimura et al., 2012; Rodgers et al., 2015; Stix et al., 1993; Voight et al., 1998).
Short-term forecasts are generally considered successful only when they lead to evacuation of exposed assets or populations from the hazard zone in a timely manner (Winson et al., 2014). Short-term forecasting depends greatly on the quality and quantity of ground- and space-based monitoring infrastructure (Section 1.4), the length and completeness of the baseline monitoring record, and the ability to interpret these data in a timely manner using some combination of experience as well as numerical and empirical models (Clarke et al., 2013; Peltier et al., 2005). In practice, short-term forecasting using empirical or statistical models of time series is hampered by limited sample size (for example, the limited number of times similar activity has been observed previously or the limited number of instruments deployed on a volcano). Short-term forecasts based on physics and chemistry models, whether deterministic or stochastic, are not yet used in practice due to model complexity and recalcitrant model parameters. As a result, short-term forecasts are not routine.
Long-term forecasts are used to estimate the likelihood and magnitude of eruptions over the life cycle of a volcano. These forecasts are relevant for land use planning over time scales of years to decades (Marzocchi and Bebbington, 2012) to more than tens of thousands of years for proposed underground nuclear waste repositories (e.g., Yucca Mountain, Nevada). Developing long-term forecasts requires reconstructing a volcano’s eruptive chronology through field study (e.g., Hildreth et al., 2012) and radiometric dating. Difficulties arise due to a lack of sufficient age determinations and a variety of biases, including bias toward large events preserved in the geologic record (e.g., Kiyosugi et al., 2015), preservation bias influenced by climate, bias toward the best mapped regions of Earth, and bias toward the most recent events that are most prevalent at the surface.
Tephrachronology and deposit mapping are the most important tools for understanding magnitudes and frequencies of past eruptions and for inferring potential future activity, including large-magnitude, infrequent events (Crandell and Mullineaux, 1978; Newhall et al., 1996; Power et al., 2010; Sherrod et al., 2008). For example, annually laminated lake sediments reveal more than 100 small VEI 2 events at the basaltic and currently open-vent volcano Villarrica over the last 600 years (Van Daele et al., 2014), but field mapping indicates that a VEI 5 caldera-forming eruption occurred in the last 10,000 years—a significant and high-impact departure from the historical record. Pinatubo’s 1991 eruption (Box 3.1), which was the largest of the last 100 years (Newhall et al., 1996), was anticipated based on field mapping of voluminous ignimbrite deposits of older eruptions. Thus, analysis of the geologic record (Chapter 2) and models of eruption processes to interpret the geologic record (Section 1.7) are critical to long-term forecasts (see Section 3.2).
A key problem is how to transform observations and models of the long-term behavior of the crust and mantle into long-term forecasts of magma ascent and eruption. For example, how does recognition of an electrically conductive body in the crust or mantle change long-term eruption forecasts for the next year or decade? Images of the crust and mantle developed from seismic tomography (Figure 2.4), magnetotellurics, geochemical models, and other technologies help us delineate the presence of magma in the subsurface, but the images are static and difficult to relate to the comparatively instantaneous process of dike ascent. One solution to this problem lies in modeling. That is, rather than simply recognizing a seismic tomographic anomaly in the mantle, the challenge is to create dynamically consistent models of how that anomaly changes the probability of magma ascent and eruption on a scale relevant to individual volcanoes. The problem would be relatively simple if a correlation could be identified between a single variable, say, seismic velocity perturbation in the subsurface (Figure 3.3), and eruption rate at the surface. Such direct correlations of single parameters have not yet been identified, and it is likely that future models will rely on a range of observations.
Understandably, most people living near volcanoes are less concerned about whether the volcano will erupt than with the consequences of eruption. Underestimating eruption consequences has contributed to the worst volcano disasters, such as at Nevado del Ruiz, Colombia, in 1985 when lahars killed tens of thousands of people (e.g., Voight, 1990).
Most forecasts of eruption hazards depend on numerical models that simulate transport phenomena—such as the development of eruption plumes, pyroclastic flows, tephra fallout, lava flows, and lahars—given that a specific type of eruption has occurred (Section 2.3). These models can be tuned to account for a range of erupted volumes, informed by mapping. Monte Carlo simulations are used to estimate the conditional probability that a flow will inundate a specific area, or that tephra fallout will exceed a given thickness (e.g., Favalli et al., 2009; Iverson et al., 1998; Jenkins et al., 2012; Wadge et al., 1994). These conditional probabilities are also used to set priorities for instrument deployment and to help authorities formulate evacuation plans and other responses to volcanic activity.
Hazard maps are developed from a combination of geologic data and numerical models to display the forecast impacts of volcano eruptions. Maps can be based on specific scenarios or probability models (e.g., Neri et al., 2015; Figure 3.4). Currently, most hazard maps identify zones that have been inundated in the past. Because the geologic record is biased, the community is moving toward model-based hazard maps, using Monte Carlo simulations and models such as those described in Section 2.4. This approach places a high premium on model validation and verification, on how to use the geologic record to formulate model inputs, and an unbiased understanding of the life cycle of volcanoes.
For syn-eruptive forecasts, both the forecasts and the hazard maps are updated during ongoing activity. For example, the maps may update areas likely to be inundated given ongoing lava flow activity (e.g., Cappello et al., 2016). Tremendous potential exists for assimilating remotely sensed data into numerical models (Section 2.3) during eruptions to provide critical updates to hazard forecasts. Stimulated by the 2010 Eyjafjallajökull eruption that disrupted air traffic over the Atlantic and much of Europe, significant progress has been made in using satellite data and dispersion models (Section 2.3) to characterize volcanic ash emissions and mass eruption rates, and to forecast and track plume trajectories after an eruption has begun (e.g., Bursik et al., 2012; Merucci et al., 2016; Pavolonis et al., 2013; Stohl et al., 2011). A more pressing need is to use these observational methods and models to accurately forecast ash concentration in airspace downwind of the volcano in the days following an eruption. Similarly, emerging remote sensing technologies, including near-real-time four-dimensional morphological mapping and new space-borne lidars (e.g., Hughes et al., 2016), will likely improve syn-eruptive forecasts, which are crucial for identifying potential changes in eruptive activity, change in topography during eruptions, and the likely duration of eruptive events.
Probabilistic forecasts of volcanic eruptions are intended to account for uncertainties about when a volcano will erupt, the magnitude of the event, and the risks to people and infrastructure. Both short-term forecasts, prepared when eruption precursors are observed, and long-term forecasts, prepared before there are signs of volcanic unrest, generally follow the same steps (Aspinall et al., 2016):
- Develop a conceptual model of how the volcano and its magmatic system work, using diverse geologic, geochemical, and geophysical data, drawing on patterns of activity at the volcano or analogous volcanoes. In preparing probabilistic forecasts, it is essential to focus on the types of activity that are possible, given how magma is stored and ascends in a particular system. Models address questions such as are vents distributed, or will future eruptions likely occur from a single vent? What are the likely products of volcanism? What are the likely volumes of future eruptions?
- Assess rates of activity. Long-term forecasts use historical observations, radiometric dates, stratigraphy, and mapping to construct a chronology of past volcanic eruptions. A statistical model is then used to transform the chronology into a forecast of future activity. The
- Assess the potential location of activity, particularly the locations of future vents. Long-term forecasts use the distribution of past vents, sometimes augmented by geologic or geophysical data, to create statistical models of probable vent locations. Short-term forecasts use geophysical or geochemical data to forecast potential dike intrusion and vent locations. However, even with high-resolution networks of instruments, the location of vents may remain highly uncertain.
- Assess the potential magnitude of activity. For long-term forecasts, magnitude is estimated from past events. Volumes of past eruptions, for example, can be used to create a probability density function of volume. Short-term forecasts of magnitude are not the norm, although the eruption volume has sometimes been roughly estimated from the magnitudes of geophysical signals in the context of the geologic record of past eruptions (Anderson and Segall, 2013).
- Assess the potential impacts of activity. Numerical models are used to estimate how far volcano products such as lava flows and tephra will extend from eruptive vents, given an eruption of a specific magnitude and style. The output of these models is usually probabilistic—for example, the likely mass loading due to
key sources of uncertainty are incompleteness in the geologic record and changes in eruptive behavior over time. Short-term forecasts are sensitive to changes in unrest and use information such as changes in rate of earthquakes or seismic energy release, deformation, or gas flux.
tephra accumulation at a specific location, given the volume and duration of the eruption, and other model parameters.
These steps can be summarized graphically with an event tree (Marzocchi et al., 2008; Neri et al., 2008; Woo, 2008). Nodes of the tree are defined as events (e.g., the volcano erupts, the magnitude is VEI 2, and a lava flow is produced).2 Different nodes in a given branch are alternative events with their own probabilities, often assigned by expert judgment (Aspinall et al., 2003). Another common approach is a logic tree, which relies on the types of models discussed in this report. In logic trees, the nodes are models and alternative models for recurrence rate or vent location are each represented as a node on the graph (Figure 3.5). The transition probability is the weight assigned to each model. By assigning weights to ensemble models and calculating the probable outcomes, the sensitivity to model assumptions can be assessed directly. Event trees are easier to use and faster to implement than logic trees. Consequently, logic trees have historically been used for long-term forecasts, and event trees have been used for short-term forecasts.
Linking Monitoring and Process: Moving Toward Physics-Based Forecasting Models
Cutting-edge data analysis leading to improved understanding of how signals in monitoring data reflect key volcanic processes is critical for improving forecasting accuracy and moving beyond pattern recognition toward physics- and chemistry-based forecasting models. Particularly important are geophysical and geochemical analytical techniques that image changes in space and time, including the following:
- Documenting ambient noise and shear-wave splitting observations of wave speed changes prior to eruption by conducting experiments at more volcanoes, and correlating changes with changes in deformation and other geophysical measurements such as gravity and electrical resistivity;
- Integrating continuous Global Positioning System (GPS) and frequent interferometric synthetic aperture radar (InSAR) time series to elucidate changes in magma reservoir pressure both between and prior to eruptions;
- Testing models of volcanic source excitation by, for example, correlating seismicity with stress changes inferred from deformation observations and/or changes in gas volume or chemistry;
- Analyzing chemical and physical changes in volcano hydrothermal systems as eruption precursors and acquiring syn-eruptive measurements to evaluate eruption progress (e.g., transitions from phreatic to magmatic eruption); and
- Using continuous high-temporal-resolution and high-spatial-resolution volcanic plume gas composition and flux measurements to test models of changes in magma reservoir permeability, volatile content, redox, and temperature prior to and during eruptions.
An improved understanding of seismic wave generation, including low-frequency earthquakes and tremor, could allow these signals to be incorporated into dynamical models. Changes in stress, documented by volcano tectonic earthquakes and changes in seismic velocities, could be jointly analyzed with geodetic, gas, and gravity measurements to image subsurface magma transport (Box 3.3). Once an eruption commences, the combination of eruption flux and geodetic data can be used to constrain total magma chamber volume, pressure, and volatile content (Anderson and Segall, 2013; Mastin et al., 2009a). Eruption models conditioned on these and other observations (gas emissions, gravity, and seismicity) could be updated to yield probabilistic forecasts of future behavior (e.g., Segall, 2013), analogous to data assimilation methods in meteorology and other fields. It will be a significant challenge to develop and test such models on active volcanoes. Physical–chemical models of ash dispersal, lava flow, and, to lesser degree, pyroclastic density current inundation are more advanced and so offer more near-term promise for this approach.
Expanding Monitoring Efforts: On the Ground and from Space
Tremendous strides have been made in developing techniques to forecast eruptions in the short term.
Eruptions can be forecast using monitoring data on gas emissions, volcanic earthquakes, deformation, and other geophysical signals. Together, these phenomena are sensitive indicators of potential eruptions. Yet, in practice there is a dearth of monitored volcanoes and a paucity of coordinated monitoring studies. Even in the United States, only a subset of volcanoes are well monitored by ground-based instrumentation, and they tend to be volcanoes that erupt relatively frequently, typically producing small-magnitude events, or that are located in high-risk areas. There is a critical need for more comprehensive volcano monitoring using ground-based seismic, geodetic, and gas sensing tools. In particular, high-resolution degassing and hydrologic data are gen-
erally less available than seismic and geodetic data, and instruments such as the miniature differential optical absorption spectrometer, multigas, and high-temporal-resolution ultraviolet and infrared cameras (Table 1.1) need to be incorporated into permanent sensor networks. When unrest begins, the basic infrastructure will need to be rapidly augmented with additional sensors and more diverse and emerging technologies, such as drones and rapid petrologic analyses. Open sharing of all data in near real time, emulating the successes of the seismologic community, will be vital.
Increased spatial and temporal resolution of satellite-borne remote sensing instruments (Table 1.2) is also crucial.Thermal sensors such as ASTER (Advanced
Spaceborne Thermal Emission and Reflection Radiometer) have high spatial resolution but low temporal resolution and so rarely provide timely observations of thermal signals such as small lava flows within craters (e.g., Reath et al., 2016). Similarly, rapid surface deformations cannot be adequately monitored with infrequent InSAR passes. For example, the planned NASA–Indian Space Research Organisation synthetic aperture radar mission provides 12-day repeat passes, which are too coarse for monitoring or documenting the evolution of eruptions. With a larger constellation of satellites, this repeat time could be reduced. It is still unclear if the increases in CO2 emissions that can precede eruptions are detectable using current satellites (e.g., Orbiting Carbon Observatory-2) because of relatively high detection limits and low temporal resolution. Additional satellites, automated detection of anomalies via those satellites (e.g., Wright et al., 2004), as well as open access to existing data streams would significantly improve monitoring.
The paucity of well-observed large eruptions poses a different set of challenges. There is only about a 1 percent chance that a VEI ≥6 event will happen in a given year. Though relatively infrequent, the consequences of these large eruptions are grave (Figure 1.2). Thus, it is critical that the volcano science community prepare to make comprehensive and high-quality observations of the next major eruption, regardless of where on Earth it is located. It is likely that the next major eruption will occur at a completely unmonitored and poorly characterized volcano, because (1) instrumentally monitored volcanoes tend to be those which have erupted in recent history, and (2) long periods of repose may be directly correlated with erupted volume (e.g., Passarelli and Brodsky, 2012). Thus, the initial detection of precursory unrest prior to a major eruption is likely to be made via satellite or local reports of felt seismicity, ground cracking, phreatic explosions, and/or increased gas emissions, all of which may not become apparent until late in the precursory sequence. For example, precursory unrest began only a few months before the paroxysmal eruption of Mount Pinatubo in 1991. A further complication is that a large eruption may not be immediately apparent from initial precursory unrest.
Satellite-borne measurements provide a global picture of where on-land volcanoes are deforming (e.g., Fournier et al., 2010), in some cases documenting the assembly of potentially eruptible magma bodies. However, forecasting the location, timing, and magnitude of major eruptions on the basis of this information remains challenging. One way to balance the tradeoff between long repose between major eruptions and our need to mitigate their dire consequences is to work toward sparse ground-based monitoring of all potentially active volcanoes (such as one or two seismometers), noting that six instruments constitute a well-monitored volcano (Winson et al., 2014) and that monitoring strategies need to be tailored to the type of volcano in question. The utility of sparse ground-based observations can be dramatically increased by scanning for all signs of unrest, including deformation, increased heat flux, and gas emissions using satellite-borne instrumentation, ideally at least daily because of the sometimes short times between the initiation of unrest and the onset of eruption (Figure 2.5). Detection of unrest that appears to herald a major eruption would then need to be followed by rapid deployment of a dense, multiparameter network of telemetered ground-based instrumentation. Such an effort would require significant resources and advance planning, developing algorithms for automated processing and scanning of satellite data, tasking satellite-borne instruments to collect more frequent observations of restless volcanoes, a cache of ground-based instrumentation, a response plan specifying the selection of personnel and procedures for import and installation of instruments, and advance coordination with monitoring agencies worldwide.