Thoughts on the criteria to determine the origin of volcanic unrest as magmatic or non-magmatic

As our ability to detect volcanic unrest improves, we are increasingly confronted with the question of whether the unrest has a magmatic origin (magma on the move) or a non-magmatic origin from a change in the hydrothermal system (fluids that are not magma on the move) or tectonic processes. The cause of unrest has critical implications for the potential eruptive hazard (e.g. used in constructing Bayesian Event Trees), but is frequently the subject of debate, even at well-studied systems. Here, we propose a set of multi-disciplinary observations and numerical models that could be used to evaluate conceptual models about the cause of unrest. These include measurements of gas fluxes and compositions and the isotopic signature of some components (e.g. H2, He, C, SO2, H2O, CH4 and CO2), the spatial and temporal characteristics of ground deformation, thermal output, seismicity, changes in gravity, and whether there is topographic uplift or subsidence spanning hundreds to thousands of years. In several volcanic systems, both magmatic and non-magmatic unrest is occurring at the same time. While none of these observations or models is diagnostic on its own, we illustrate several examples where they have been used together to make a plausible conceptual model of one or more episodes of unrest and whether eruptions did or did not follow the unrest. This article is part of the Theo Murphy meeting issue ‘Magma reservoir architecture and dynamics’.


Introduction
This special issue explores an emerging conceptual model of trans-crustal magmatic systems (TCMS) where magma storage and differentiation occurs at several locations throughout the entire crustal column [1,2]. This conceptual framework is being driven by petrological, geochronological, and geochemical studies of magma storage conditions that show (1) many large eruptions tap multiple melt sources, (2) large melt bodies are probably transient features, and (3) crystals carried by the transporting melt have been stored at a range of pressures and temperatures [2][3][4]. The implications of the TCMS model for eruption dynamics and forecasting are still being explored [1]. In particular, observations such as ground deformation, seismicity, and volcanic degassing are the basis of volcano monitoring, and interpretations of these data need to be reassessed within the newly-developed framework [1].
For example, most eruptions are presaged by at least a few days to weeks of volcanic unrest [5,6] -how does this unrest relate to the TCMS paradigm? However, some volcanic eruptions have little to no measured unrest before eruption [7,8] -is this consistent with the TCMS conceptual model? We define volcanic unrest "as the deviation from the background or baseline behaviour of a volcano towards a behaviour which is a cause for concern in the short-term because it might prelude an eruption" from Phillipson et al., (2013) [6]. Given this definition, since the level of background activity varies for each volcano, the threshold for unrest is determined separately for each volcano [9]. However, while the unrest "might prelude an eruption," unrest may not result in an eruption at least in the short term -these events are sometimes called "failed eruptions" [10]. According to the TCMS paradigm [1], unrest can be both magmatic and/or non-magmatic ( Fig. 1a,b) -caused by destabilization of magma layers, volatile accumulations, or both. Unrest can result in eruption or not depending on the size, location, and rapidity of the destabilization (Fig. 1c), possibly involving multiple layers [1]. To assess the eruptive threat from a given episode of unrest, it is critical to understand its origin. If unrest is caused by magma layer destabilization and movement, Sparks and Cashman (2017) consider it more ominous than volatile accumulation and release from magma that moved during a previous destabilization, or changes in the shallow hydrothermal system [1].
To quantify the possible eruptive threat from unrest episodes, several groups have developed Bayesian Event Trees (BET, Fig. 2) that include an assessment of the likelihood that the unrest is caused by "magma on the move", geothermal or tectonic activity, or some other process [9,[11][12][13]. The eruption hazard from non-magmatic unrest (geothermal, tectonic, volatile accumulation and release, or other) is generally lower than magmatic unrest. But, these event trees include the possibility of hazards from non-magmatic unrest, including eruptions, for example through sector collapse, phreatic explosions, or a tectonically induced fracture [11,12]. Non-magmatic unrest includes "fluids on the move" [13], where the fluids include brines, gas, supercritical fluids or a combination of these as long as partially molten rock (magma) is not migrating [13]. These fluids could have their origin from magma, for example volatiles derived from stagnant [9] cooling/crystallizing magma batches (e.g., Fig. 1a,b) and are called "magmatic fluids" [14], from meteoric/groundwater water mixing in the hydrothermal system, or a combination of the two that are referred to as "hydrothermal fluids." This distinction between magmatic and nonmagmatic unrest is not yet standardized. For example, a comprehensive review of global caldera unrest between 1984-2014 noted that when known, the root cause for all unrest at calderas was magmatic [15]. By the definitions used here, not all of these caldera unrest episodes are from "magma on the move." Using BET allows the simultaneous consideration that unrest could be magmatic or non-magmatic when evaluating the outcomes of the unrest episode.
To implement BET frameworks such as those shown in Fig. 2, the top level question for evaluating hazard from an unrest episode is determining the probability that unrest signifies magmatic or non-magmatic processes. Specific monitoring criteria that can distinguish between the origins of unrest have been proposed at a few volcanic systems (Table 1) [13]. Yet, this is a challenging question to answer, as there has been a long history of debate about magmatic . . vs. non-magmatic sources of unrest even at well studied volcanoes [14,16]. On the other hand, our ability to monitor unrest on a global basis is increasing thanks to a growing number of space and ground-based observations [17,18], and the exposure of global populations to volcanic hazards [19] motivates improved assessment of the origin of volcanic unrest. Our goal in this paper is to review the characteristics of unrest that could be diagnostic of magmatic or nonmagmatic processes and present some case studies where these characteristics have been used to develop conceptual models of the causes of unrest. Certainly, the TCMS paradigm reinforces the idea that both magmatic and non-magmatic unrest could be occurring at the same time [20], but we suggest that the critical question is still whether there is evidence for "magma on the move" irrespective of evidence for additional hydrothermal activity. We propose a check-list of criteria that could be consulted during new episodes of activity to assess the origin of unrest and prioritize new diagnostic observations. Our focus is on silicic systems because these are the sites of the most explosive and dangerous eruptions on Earth, but we include some useful examples from other systems as well.

Methods: Types of data to distinguish cause of unrest
In this section we review monitoring criteria that have been used to assess whether unrest is magmatic or non-magmatic. None of these non-eruptive criteria are unique -the best way to determine if unrest is magmatic is to see if it culminates in an eruption of new magma. This is necessarily retrospective. But unrest due to "magma on the move" doesn't always result in eruption [10]. In practice, no single non-eruptive criterion is used to determine if unrest is magmatic [13]. For some volcanoes, there is a history of relating unrest to eruption, so there may be specific quantitative thresholds that can be used to assess the likelihood that unrest is magmatic (e.g., Popocatépetl and Cotopaxi, Table 1). However, since these criteria are specific to each volcano, we do not attempt to assign quantitative values -instead we discuss the value and qualities of the different types of data. The use and combination of these data-streams to assess the nature of unrest is explored further in the case studies in section 3.

(a) Gravity change
A key difference between magmatic and non-magmatic fluids is that their density differs by a factor of 50% or more -thus measurements of gravity change associated with unrest that determine the density of the moving fluids could be used to diagnose the cause of unrest if the mass change is large enough [21]. Gravity change has been reported at several volcanoes with fluid densities ranging from 142-1115 kg m −3 (likely water and/or gas) [22,23] to 2192-3564 kg m −3 (partially molten rock) [24,25] to a combination of the two between 1000-2500 kg m −3 [26,27]. Ground deformation is often, but not always associated with gravity changedifferences between the source of gravity change and deformation can illuminate the origin of unrest [28]. Gravity measurements are difficult to make and several corrections must be applied in order to infer the density values -for example, changes in aquifer levels impact the surface gravity and are not always well constrained [21,24]

(b) Ground deformation
Ground deformation can be caused by magmatic or non-magmatic unrest [1,36]. But several spatial and temporal characteristics of the deformation, as well as the relation of deformation changes to coincident or delayed changes in other data types (e.g., seismicity, gravity, and degassing) could be more diagnostic for the origin of unrest. Here, we briefly describe a few of these deformation characteristics.
Size of deformation/Depth of source: The size of the deforming region is related to the size and/or depth of the source. A deeper source causes deformation over a larger area of the surface, but the deformation pattern is non-unique -a large shallow source could also match the data. If there is additional data from petrology and/or other geophysical methods, the source depth can be better constrained [37]. Because hydrothermal systems are usually <10 km deep [38], source depths greater than this likely indicate magmatic unrest. But it is possible that magmatic fluids at depths >10 km are the source of unrest without "magma on the move" [39] as part of the TCMS model [1].
Relation to degassing and seismicity: At several volcanoes, ground deformation changes rate and even direction over short time periods (days to years) and these changes are correlated with changes in seismicity and degassing that can be used to diagnose the origin of unrest [14,20,40]. We will discuss how these different datasets can be combined together below in the case studies for individual volcanoes (section 3).
Relation to geomorphic uplift: The repose interval between eruptions is frequently hundreds to thousands of years (or longer), and it may take a similar amount of time to accumulate magma between eruptions. Over these timescales, geomorphic uplift can provide a record of magma injection. These geomorphic observations complement the shorter duration geodetic observations spanning years-decades, but are limited to areas where geomorphic features are created (e.g., shorelines, river and coastal terraces, etc.) and preserved. In fact, uplift of tens to hundreds of meters has been recorded in some volcanic areas over the past centuries to millennia and have been related to magmatic unrest without eruption (e.g., Laguna del Maule, Chile; Ioto (Iwo Jima), Japan; Socorro, New Mexico) [41][42][43]. There is at least one example of 20 m of uplift being related to non-magmatic unrest in the Gulf of Naples [44]. On the other hand, there are also examples of volcanic areas that have little geomorphic uplift (less than a few tens of meters) despite significant or persistent geodetic deformation rates [45][46][47]. The inference is that these geodetic rates are not sustained for centuries or millennia, such that the deformation is caused by transient magmatic unrest or episodic non-magmatic unrest [45][46][47].
Temporal evolution of ground deformation: Several different types of analytic and numerical models can be used to assess if unrest detected by ground deformation is of magmatic or nonmagmatic origin. While these models are non-unique, they are useful for developing testable hypotheses about the nature of unrest. For example, ground deformation signals at several silicic [48-51]) and basaltic [52-56] volcanoes show either exponential or a double exponential pattern in time. These signals are predicted by several models that couple Newtonian magma flow in a conduit from a deep pressurized source (usually in the lower crust) into a shallow pressurized reservoir. Variations of the coupled reservoir-conduit models incorporate two coupled shallow reservoirs [55] and reservoirs surrounded by a viscoelastic media [57]. The exponential and double exponential patterns arise from the deep source pressurization function. A deep source pressurization increase followed by a constant pressure results in a double exponential, while a deep source constant pressure results in a single exponential. The quasi-exponential trends in ground deformation data can thus indicate magma injection but can also be caused by reservoir pressurization in a viscoelastic media. The ambiguity between these interpretations can be . Ground deformation can also be caused by movement of non-magmatic multiphase (liquid, gas, super-critical fluids) and multicomponent (e.g., H 2 O-CO 2 ) hydrothermal fluid flow and poroelastic deformation [61-63]. The diagnostic rates, patterns, total magnitudes, and durations of ground deformation from these non-magmatic models can be similar to magmatic models of ground deformation [61], and so other types of data (gravity, seismicity, gas, etc.) are important in distinguishing the cause of ground deformation. Unfortunately, these models have been rarely used to predict ground deformation time series. Instead of modeling the temporal evolution, most of the ground deformation studies have only used a set of either cumulative or average displacements. Hence, the potential ability to assess whether the exponential or double exponential signals result from magma injection or hydrothermal flow for time scales of 1-10 years is still an area of active research.

(c) Temperature
Some volcanoes and fumarole fields (e.g., Momotombo, Nicaragua [64]; Satsuma-Iwojima, Japan [65,66]; Kudryavy, Kuril Islands [67]) have continuous high temperatures at the surface (700-900 • C) that imply shallow magma, but without evidence of magma migration [9]. In some cases, these high temperatures have been ongoing for decades to centuries, and crystallization and/or convection of a stagnant magma body can explain the degassing and high surface temperatures [9]. Globally, there are dozens of such stagnant, but degassing magma bodies (e.g., Pleistocene restless volcanoes and calderas with fumaroles and sometimes ground deformation and seismicity [68]), that may be considered in a "dormant" state ( Fig.1) [1]. On the other hand, spatially large and high temperature thermal anomalies that produce incandescence at the summit, for example eruptions at Popocatépetl, Mexico (Table 1) and the lava lakes, such as that at Villarrica, Chile [69], can indicate magmatic unrest near the surface. Transient increases in temperature and area of the fumarole field provide more ambiguous evidence of magmatic or non-magmatic sources. Sandri et al., (2017) [13] notes that at some volcanoes, temperature increases of 120-200 • C are hydrothermal (non-magmatic) and increases >300 • C are magmatic. Kern et al., 2017 [70] describes the thermal detection of the increase in area of the fumarole field and the increase of water vapor degassing before the 2016 eruption of Sabancaya. The increase of water vapor can clearly be related to "fluids on the move" but it is unclear if the driving force to originally cause fluid movement was the introduction of new magma (i.e., "magma on the move").

(d) Seismicity
At some volcanoes, particularly long-dormant ones, swarms of volcano-tectonic (VT) earthquakes may occur before other signs of unrest or before there is a clear sign that the unrest is of magmatic origin. For example, unrest at Cerro Chiles, Ecuador-Colombia [71] and Sabancaya, Peru [72] started as earthquake swarms with an unclear magmatic connection, but as time progressed, evidence for magmatic intrusion developed in both locations.
Swarms of VT earthquakes are common at volcanoes -that is many earthquakes of about same size that happening within the same volume over a short duration [73]. However, an informal poll of experts estimates that only about 1/3 to 1/10 of swarms precede eruptions (McNutt, unpublished data). The idea is that over time, numerous swarms occur, but only a fraction of them are followed immediately by eruptions. These values would be initial probabilities of the first branch of BET (a more formal elicitation or similar study is needed to yield further insights). The low rate of precursory swarms (or high rate of false alarms or intrusions) is broadly consistent with the TCMS model. Magma may move between different components of TCMS but swarms are not necessarily pre-eruptive. VT swarms typically have rates about one order of magnitude above background, and swarm durations range from hours to months or more with a mean of . .  [73]. Pre-eruptive swarms have systematically longer durations by about a factor of two [73].
A systematic pattern of seismic event types has been observed and has been described as the Generic Volcanic Earthquake Swarm Model (GVESM [74]). The sequence consists of VT events that reach a peak rate of occurrence, followed by relative quiescence. Next to appear are lowfrequency (LF) events (also called long-period or LP) followed by tremor. LF events and tremor often have similar frequencies (1)(2)(3)(4)(5) and are thought to represent transient (LF) and sustained (tremor) fluid processes in sections of conduits at shallow depths. The fluids may be magma, water (e.g. groundwater), gases, or any combination of these. An uptick in seismicity often precedes eruptions on a time scale of hours; this may be VT or LF events or an increase in tremor amplitude. Strong tremor generally accompanies eruptions with the amplitude (using a metric known as reduced displacement) being roughly proportional to the VEI [75]. Swarms of deeper VT events often follow eruptions as stresses re-adjust in the vicinity of magma chambers. The time frames of swarms that follow the GVESM are highly variable; examples are given in [74] Not every earthquake swarm at volcanoes follows the GVESM sequence. There are cases of some elements missing; for example, the Redoubt 1989 sequence was missing VT events, and instead had LP events, tremor, and the onset of eruption 23 hours after the LP events began [76]. Presumably the deeper ascent of magma from 10 to 1 km was relatively aseismic.
The GVESM is a conceptual model that is consistent with vertical ascent of magma and/or volatiles. Most VT events occur at depths of 3-10 km whereas LF events occur at depths of 1-3 km and tremor even shallower. Observations of increased steaming are common around the time of LF event onset. A complication is that a water/gas front may rise in advance of the magma, and this would also give the observed sequence of events. If the tremor and/or LF events are caused by boiling of hydrothermal waters, the reduced displacement is generally less than 5 cm 2 [77,78]. Eruption tremor (magma) is considerably stronger.
In addition to the increased rate of events that makes up a swarm, the distribution of event sizes may also change. Seismologists refer to this as the frequency-magnitude distribution or bvalue (b is the slope in the relation Log 10 N = a -bM, where N is cumulative number of events, M is magnitude, and a and b are constants). Laboratory studies show that b increases (steeper slope or more small events) with higher pore pressure or thermal gradient, and decreases (shallower slope or more large events) with higher stresses. Thus b-values can shed further light on processes and likelihood of eruption. For example, the 2006 Augustine eruption was preceded by a decrease in b-value months before it began [79].
In terms of BET scenarios, VT swarms alone likely give probabilities of eruption of 1/10 to 1/3 as mentioned above. If the b-values become lower, this suggests increased magmatic pressure. If VT events are followed by LF events, the probability of an eruption (i.e., a magmatic origin) increases, although it is hard to be more quantitative at this stage. An increase in b-value at this point could indicate either increased pore pressure (hence water) or increased thermal gradient (magma or water). The probability increases again with the onset of volcanic tremor. So overall, the presence of different types of events, which forms the basis of indicating adherence to the GVESM, suggests a higher likelihood of eruption than the presence of any one type of seismicity alone. However, the presence of the GVESM sequence does not unambiguously show that magma or water or gases are the cause. Additional characteristics of the seismicity can help resolve the ambiguity. For example, a recent paper by Passarelli et al. (2018) [80] determined that lateral magmatic dike intrusion was the most likely cause of a swarm at Jailolo volcano. This was based on careful assessment of hypocenter migration, focal mechanisms, non-double-couple components, surface fracture orientation, and tectonic considerations.
Deep long-period (DLP) events are also recorded at volcanoes. At some volcanoes their rate of occurrence is approximately steady state, such as Kilauea of Pinatubo, although they were only recognized in hindsight [82]. At other volcanoes, DLP events mostly occurred after eruptions, such as at Mount Spurr in 1992 [83]. All three cases cited here suggest the involvement of magma. However, DLP events beneath Mammoth Mountain, California, have been linked to CO 2 release at the surface that may or may not be related to "magma on the move" [84]. Many DLP events in the Cascades, Japan, and Alaska are not associated with eruptions or unrest.
Another class of earthquakes is known as very long-period (VLP) events. The period is 10 seconds or longer (up to 50 seconds) and both shallow and deeper events have been recorded. Two well-studied examples are at Stromboli, Italy associated with gas slug motion through magma [85], and at Aso, Japan associated with expansion and contraction of a shallow geothermal feature [86]. Again similar seismograms are recorded with and without magma involvement.
Intrusion and pre-eruption swarms share common features with the exception of shallow harmonic tremor near the eruption site when an eruption is the outcome of unrest [87]. This suggests that intrusions represent the same suite of processes as pre-eruptive swarms but the magma stalls before reaching the surface (see also [10]).

(e) Gas flux and composition
Magmas at depth have numerous chemical species dissolved within them that exsolve as magma rises through the crust. The major gas species are usually H 2 O, CO 2 , SO 2 , H 2 S and the halogen halides (e.g., HF, HCl) but other minor components include noble gases (e.g., He), H 2 , CH 4 , CO, COS and metallic species. The composition of the gas mixture will depend on factors including the pressure of release of the gas (e.g., CO 2 tends to be exsolved deeper in the crust than other species and higher pressures favour H 2 S over SO 2 ), the temperature of the system, mixing between magmatic and other volatile sources and interactions with ground water or hydrothermal systems that might strip or scrub out soluble species [88,89].
Where present, gas emissions at restless volcanic systems can be characterized by a range of surface manifestations. In some (usually basaltic) systems like Etna (Italy) and Villarrica (Chile) gas is emitted from a magma-air interface at a visible lava lake or down within the vent/conduit. Other systems are characterized by fumarolic emissions, hotsprings, emissions through lakes or diffuse seeping out as soil gas.
As SO 2 emissions from a magma occur at a lower pressure than CO 2 and are favoured by high temperature and low pressure conditions, an increase in SO 2 flux at a volcano may herald the arrival of new magma into the shallow system. For example, SO 2 flux measurements were an important contributor to successful prediction of the June 1991 eruption of Mount Pinatubo. Measurements in mid-May (500 tonnes/day) indicated that unrest involved intrusion of magma. By late May the flux had increased tenfold, interpreted to imply that (1) magma was rising and/or (2) a hydrothermal system was being boiled and removed, allowing more of the SO 2 , that had previously been scrubbed by it, to reach the surface. Both of these scenarios pointed to an increased hazard from the volcano. The signals were not always straightforward to interpret, however. On 5 June there was a sudden, short-lived drop in SO 2 flux (to 260 tonnes/day), even as seismicity was increasing. This may have been caused by plugging or sealing of the system inhibiting gas escape. On June 7, as a new dome was extruded, emissions increased again and the last pre-paroxysmal measurement (10 June) was >13,000 tonnes/day [90].
Changes in SO 2 /CO 2 ratios have been suggested to be of use to forecast magma movement and hence dangerous unrest resulting in explosive eruptions at both systems with strong hydrothermal processing (e.g., the gas emissions through the lake at Poás volcano, Costa Rica in 2014 [91]) and the more dominantly magmatic emissions from systems like Villarrica, Chile [92], Etna and Stromboli volcanoes, Italy [93,94]. However, these systems go through long periods of unrest with known magmatic involvement and these studies are more aimed at finding precursors to dangerous changes in this background unrest rather than establishing whether or not there is a degree of magmatic involvement. Gas fluxes and ratios were also used to give key insights into Increases in CO 2 /SO 2 and H 2 S/SO 2 over the months prior to the eruption recorded in fumarole gas samples, were used to suggest a deep degassing source associated with an input of fresh magma (most likely of mafic composition). The quantity of SO 2 degassed was used to argue for the presence of an exsolved fluid phase in the pre-eruptive magma body, which could have played a key role in the eruption's explosivity [95] The presence of SO 2 and HCl, clearly magmatic gases, does not necessarily imply the migration of a magma though, as exsolution of both species can occur through convection in the magmatic plumbing system and/or crystallization of a stagnant, cooling magma batch and gas exsolution [9]. For example, Kawah Ijen, Indonesia has mainly been in a state of non-magmatic unrest for years, with the occurrence of only phreatic or geyser-like eruptions [38]. Nevertheless, the volcano hosts the largest reservoir of acidic surface water on Earth, continuously fed by the input of magmatic gases and volatilized metals [9].
At many systems, such as caldera systems, that start to show new unrest after periods of quiescence, the surface manifestations in terms of degassing might, at least initially, be more subtle and build up more slowly than in cases like Pinatubo. In cases where diffuse emissions of gases emitted deeper in the Earth's crust dominate (usually  However, not every volcano has a large hydrothermal system clearly linked to the magmatic system like Laguna del Maule and Lastarria-Lazufre (described in section 3). In Iceland, for example, there is no significant hydrothermal system at Hekla, indicating that any significant partially molten volume is deep (> 14 km) [101]. However, Krafla and Grimsvotn support hydrothermal systems and shallower magma [101]. It might seem obvious that if there is no surface manifestation of a hydrothermal system, then hydrothermal activity could not be a source of unrest, and a magmatic origin for the unrest is more likely. However, further investigation is required since there are "blind" geothermal systems. In these systems, near-surface permeability conditions do not allow a hydrothermal system to develop above a magmatic heat source, and the hot fluids are forced to move 10 or more km horizontally [102].
Water outputs from hydrothermal systems can manifest in numerous ways (e.g., hot springs, mud pools and volcanic lakes/streams) and their chemistry can be intimately linked to fumarolic gas chemistry (see section 2e). For example, volcano lake compositions are strongly influenced by volcano fluid inputs although modulated by meteoric inputs and physical, chemical and biological processes within the lake. These in turn might be dominated by magmatic gases or the products of their interactions with the edifice/crustal rocks during transport. Major variations in lake composition often result from a changing volcanic input composition or magnitude and are thus useful for volcano monitoring [103]. Similarly, thermal spring compositions are modulated by external inputs (e.g., meteoric water), magmatic fluids and interactions with the edifice and crustal country rock. Despite this complex interplay between sources modulated by feedbacks such as fluid pH and temperature changes in thermal spring chemistry and the resulting streams or rivers can be useful indicators of variations within the system [104,105]. In submarine systems, magmatic intrusions have been shown to change the He, CO 2 , H 2 and CH 4 , and the 3 He/heat ratio from hydrothermal fluid samples for months before returning to background levels over longer time periods [106].
A high heat flux at a volcano is ultimately driven by magmatic activity, but is also responsible for hydrothermal activity. The amount of interaction between the hydrothermal system and the magma below can be measured by monitoring the discharge in Cl from streams and constraining the heat flow [107]. Increased heating of the hydrothermal system and increased heat flux in general can indicate new magma in the system or increased transport of existing magmatic fluids, triggering further unrest [108]. Water flux variations have been seen in the water level of crater lakes after the El Chichón, Mexico eruption in 1982 and a rise in the lake at Poás, Costa Rica before eruption [9]. Such water level measurements are not frequently made. Changes in water levels can also be monitored by geophysical methods like self potential, microgravity, resistivity or other surveys [9].
(g) Drilling: Direct sampling and monitoring of the subsurface The monitoring techniques mentioned so far (gas and water sampling, seismicity, temperature, ground deformation, gravity, etc.) are restricted to making measurements at the surface or in shallow boreholes [109,110], such that only inferences can be made about conditions at depth causing unrest. In a few cases, drilling several km into upper crustal magmatic and hydrothermal areas has directly measured parameters that are critical for developing numerical models: the composition, physical properties (e.g., density, permeability, thermal conductivity, etc.), stress state, and temperature at depths closer to where unrest is occurring [111]. Some of these wells were designed for primarily scientific objectives [112][113][114][115] whilst others were for geothermal energy development. In several cases, the drilling revealed features that were unsuspected based on surface data alone, like lower temperatures than expected [116] and very abrupt transitions between solid and partially molten rock [117][118][119]. To provide ground truth to our near-surface monitoring data (e.g., the location of sub-solidus conditions, super-critical fluids, partially molten rock, etc.) and develop numerical models, future drilling and installation of deep observatories have been suggested [120][121][122]. Such drilling would also better constrain the location and properties of high enthalpy geothermal systems that could be a high quality energy source [123].

Case studies: Combining data-streams to understand unrest
In the last section, we reviewed the general data-streams that have been used to estimate the likelihood that unrest is magmatic. In the following section, we briefly discuss nine specific cases where multiple of these criteria have been combined to assess the origin of unrest. This is not intended to be an exhaustive review of such studies but is rather intended to illustrate key points via examples from a range of volcanic systems. Our examples are determined by those systems where suitable studies exist, but are also chosen to span different tectonic settings (rifts, subduction zones, hotspots) as classified by [68] that have been shown to have different characteristic relations between ground deformation and eruption globally [124]. More examples of caldera systems and more detail on restless episodes are available in [15,38]. Here the focus is on how multiple parameters have been used to assess if unrest requires "magma on the move" (even without eruption) and what additional data are required to assess this. Some of these systems (like Campi Flegrei and Yellowstone) have been extensively discussed already in the literature and so we include briefer comments and references. We spend more time on lesser-studied systems where multi-parameter observations of the unrest are in the nascent stages of being synthesized into conceptual models of the magmatic system. increases that killed trees) and changes to the hydrothermal system [32]. The unrest spans more than 20 km laterally from a large Pleistocene silicic caldera to the younger dacitic Mammoth Mountain and surrounding basaltic eruptions. The various manifestations of caldera unrest have been related to magma injection, but recently Hildreth (2017) [32] proposed that the unrest in different parts of the Long Valley system could have different causes. Specifically, the uplift, earthquakes and other activity in the Pleistocene caldera could be caused by degassing of stagnant magma (that he attributes to "second boiling") [32]. He questions whether the gravity change in the caldera that has been attributed to magma intrusion because of the high density [24,25] could not be due to non-magmatic processes (i.e., the gravity change "signal" is being misinterpreted because it is actually "noise"). On the other hand, he attributes the CO 2 flux and earthquakes beneath and near Mammoth Mountain to the intrusion of mafic magma. A lingering question at Long Valley is if the caldera uplift since 1980 (about 0.8 m) is related to gas pressurization, why has there not been subsidence of approximately equal magnitude as the uplift as seen at Yellowstone or at least appreciable subsidence as at Campi Flegrei (Fig 3)? One possibility is that the depressurization may take a longer time period than the available observations. For example, gas uplift of a dome (5 km diameter) of 20 m in the Gulf of Naples has been proposed to have lasted 12,000 years [125] -but the contrast with Yellowstone over the last decades where significant subsidence has occurred is striking (Fig 3). Over the last 333,000 years, uplift of the caldera as recorded in lake sediments and the Hot Creek flow is constrained to be of order 40 m [32]. Either the caldera uplift events like that over the last 40 years don't eventually lead to equal subsidence and are infrequent (i.e., repeat every ∼ 8000 years) or are eventually counterbalanced by subsidence, consistent with non-magmatic unrest driven by cycles of nearly equal pressurization/uplift and depressurization/subsidence.

(b) Campi Flegrei, Italy -subduction zone
Campi Flegrei caldera has experienced several episodes of unrest (e.g., total ground uplift over 3 m, Fig. 3b) during the 20th century without eruption [15]. The caldera last erupted in 1538 and its activity is of great concern to the 360,000 people who live in the caldera and the 3 million residents of neighboring Naples [126]. There have been decades of discussion about whether unrest at Campi Flegrei is magmatic, non-magmatic, or both [16,22,125] and a comprehensive review of the relevant datasets and arguments is beyond the scope of this paper. Many studies have shown the value of combining multiple datasets when inferring the cause of unrest. One argument for a non-magmatic origin of unrest is the delay between increased diffuse degassing following the uplift pulses lasting about 100 days [127]. When combining geophysical and geochemical data time series spanning multiple decades (e.g., Fig. 3b [125] argue that the 1982-1984 unrest is primarily magmatic while the unrest since 2005 is non-magmatic while other papers flip the interpretations for these two episodes [108,128]. Offsets in coastal features show episodic cycles of decimeter uplift and subsidence over decades-centuries related to a combination of magmatic and non-magmatic processes [129], including a permanent uplift of as much as 33 m inferred to be magmatic intrusion in the last 2000 years [130].

(c) Yellowstone, USA -hotspot
The Yellowstone caldera has an observational record of cyclic inflation and deflation (Fig. 3c) spanning almost a century [14]. Similar patterns of uplift and subsidence occur over millennia and have resulted in net geomorphic change of only a few tens of meters [45]. Changes in deformation from inflation to deflation are correlated with seismic swarms (Fig. 3c)  hydrothermal discharges. Ground deformation in different parts of Yellowstone is frequently anti-correlated -when the Norris Geyser Basin changes from uplift to subsidence, the Sour Creek resurgent dome deformation has the opposite sense (Fig. 3c). The deformation sources responsible for the inflation and deflation cycles are located within the large body of partial melt that underlies the caldera, in both the shallower rhyolitic (depth ∼5-15 km) and the deeper basaltic (depth ∼15-20 km) sections of the tomographically imaged TCMS extending from the surface to the lower crust at depths of ∼50 km, and with a volume of ∼56,000 km 3 [26, 131,132]. That changes in deformation in different parts of the caldera occur together and rapidly (within a few days to weeks, for example in early 2014, Fig. 3c) suggests coupling of low viscosity fluids. The deformation overturn and seismic swarms are thought to be produced by the breaching of a sealed layer that mobilizes magmatic fluids in the shallow hydrothermal system [133]. The driving mechanism responsible for the broad caldera uplift is thought to be basaltic magma injection (related to a large CO 2 flux), cooling rhyolitic magma releasing fluids, and a significant contribution from hydrothermal processes. The exact coupling between the hydrothermal and magmatic systems and how much deformation is of hydrothermal origin is currently unknown [14]. Yet, the correspondence between the magnitude of inflation and deflation over years to millennia suggests a role for a recoverable process like pressurization and depressurization by fluids on the move, as opposed to persistent uplift expected from repeated magmatic intrusions   [134]. Uturuncu is a dacitic stratovolcano of Pleistocene age and Lazufre is about 10 km from two Pleistocene-Holocene-age andesitic-dacitic arc volcanoes called Lastarria and Cordón del Azufre. A connection between the Lazufre magmatic system and the active hydrothermal system at Lastarria 10 km away is suggested [135][136][137], but not yet confirmed. Both Uturuncu and Lazufre-Lastarria are likely associated with TCMS spanning hydrothermal systems at the surface to geophysical anomalies (e.g., low electrical resistivity and low seismic velocities) indicating partial melt through the mid-to lower crust [134]. Unrest is occurring at multiple depths at approximately the same time within the system, and it has been proposed that unrest of magmatic origin is the cause of deep unrest [138,139] while non-magmatic activity may be the cause of shallower unrest [135,140,141]. A different interpretation is that all of the unrest is of non-magmatic origin and caused by cycles of vertical up and down movements related to volatile-driven pressurization and depressurization [39,134]. The non-magmatic origin would be consistent with no geomorphic uplift over thousands of years [47]. On the other hand, at Lazufre, there is evidence of ∼ 500 m of uplift over the past 0.4 Ma that could be related to magmatic intrusions over the longer time period [47,135]. To further assess whether there is "magma on the move" at Uturuncu and Lazufre, analysis of seismic data at Lazufre (e.g., shallow and deep LP events) and measurements of gas at both Lastarria and Uturuncu are ongoing.

(e) Laguna del Maule, Chile -subduction zone
Laguna del Maule (LdM) is a large center of silicic volcanism in the southern Andes with significant unrest since at least 2007 (uplift of more than 20 cm/yr, earthquakes, gravity change) but no historic eruptions [142]. The ground uplift is modeled as an inflating sill at 4.5 km (magmatic unrest) [49], while the gravity change is primarily caused at shallower levels by low-density fluids (non-magmatic unrest) [23] that are not related to any known surface geothermal system, but such a system could be hidden beneath the lake or displaced 15 km (Baños Campanario [143]).

(f) Cordón Caulle, Chile -subduction zone
Cordon Caulle (Southern Chile), is a ∼10 km long large rhyolitic fissural system that has had three large rhyodacitic eruptions (VEI 3-5) in [1921][1922]1960 and 2011-2012 with nearly the same chemical composition [144,145]. The volcano also hosts one of the largest hydrothermal systems in the Southern Andes, with several acid fumaroles and even a small geyser [146]. The 9-month 2011-2012 eruption is the only one with instrumental observations at this volcano, and was preceded by several years of InSAR-detected ground deformation from 2007-2009 and early 2011 (Fig. 4a). Seismic observations are only available since mid 2010, and show a clear seismicity increase two months before the eruption [147], in agreement with the preeruptive ground deformation. Despite the seismic and deformation pre-cursors [145], there was no observed increase in temperatures of the fumarole fields even three weeks before the eruption with respect to a measurement made more than one year before ( Fig. 4b-d). After the eruption, a temperature increase is visible at the lava flow, but again, there is no significant change in the fumaroles (Fig. 4e). If there was magma or fluids on the move in the months before the eruption in 2011, it seems that the hydrothermal system did not record the change in terms of temperature at the surface. This does not rule out that the hydrothermal system responded to changes in seismicity and ground deformation over shorter time scales of a few months to a few days, as it did in 2007-2008 (Fig. 4b), but were not recorded due to the poor temporal sampling of the available thermal data in 2011. Or perhaps the unrest in the magmatic system did not impact the temperatures of the hydrothermal system at the surface. This might be expected if the magma transfer from depth to the surface was highly localized and happened rapidly, but this partitioning of the hydrothermal system from the rest of the magmatic system calls into question the utility of monitoring surface temperature change as an indicator of magma or fluids on the move at some volcanoes. Unfortunately, there were no gas measurements in the volcano hydrothermal system in the months preceding the eruption and right after it to detect possible changes in the shallow hydrothermal system.
The volcano underwent almost 1 m of uplift right after the end of the 2011-2012 eruption in three discrete pulses during 2012-2018 (Fig. 4a), ongoing as of May 2018, and with very similar spatial footprints [60]. This post-eruptive inflation is not associated with abnormal seismicity [50] or temperature change (Fig. 4b). The fact that the ground deformation followed an exponential trend during the first three years after the eruption [50], as predicted by the magma injection models (section 2b), strongly suggests that the post-eruptive deformation is of magmatic origin. Finite element simulations show that viscoelastic relaxation is of secondary importance with respect to magma injection [60]. Alternatively, inflation after eruption is possible with no recharge for an incompressible magma [57]. While the Cordón Caulle erupted magmas are highly compressible [145], the compressibility of the post-eruptive magma intrusions is not known. The deformation sources are significantly deeper than the inferred depths of the shallow hydrothermal system [50]. All these lines of evidence suggest that the 2012-2018 post-eruptive inflation is most likely of magmatic origin. However, volatile exsolution from depressurization [36] (so-called "first boiling") without significant magma input is another possibility that cannot be excluded in the absence of other datasets and of simulations that can predict the ground deformation temporal evolution.

(g) Aluto, Main Ethiopian Rift (MER) -rift
A recent global compilation suggested that the predictive relationship between deformation and eruption in rift settings was lower than that derived from the global dataset as a whole [124] Therefore, understanding how to combine multiple datastreams to interpret signals of unrest at volcanic systems within rifts is of key importance. The East African Rift (EAR) is often used as a case study for continental rifting and hosts numerous restless volcanoes [148][149][150].
Here we focus on Aluto volcano in the MER, as an example of an EAR volcanic system that has been the focus of recent inter-disciplinary study to elucidate the processes driving its unrest. Aluto is a restless silicic volcano/caldera in the MER that last erupted about 400 years ago. The style and volume of recent eruptions suggests that silicic eruptions occur at an average rate of 1 per 1000 years, and that future eruptions of Aluto will involve explosive emplacement of localized pumice cones and effusive obsidian coulees of volumes in the range 1-100 x 10 6 m 3 [151]. Aluto has a well-developed hydrothermal system and the complex has been targeted for geothermal development. Diffuse volcanic degassing also takes place at a number of sites across the volcano and it is evident that the pre-existing structures dictate gas and hydrothermal fluid ascent to the surface [152].
Aluto has undergone multiple uplift and subsidence events since 2003 without eruption [149] (Fig. 3d) and understanding the causes of these unrest events has required drawing data together from multiple techniques. Detailed analysis of the InSAR timeseries shows 2 episodes (in 2004 and 2008) of accelerating uplift and edifice-wide inflation, followed initially by rapid subsidence and then slower deflation [153]. The location of the uplift source is roughly centered within the caldera and is constant between 2004 and 2011. Modelling of the uplift finds the best-fit with a spherical point source at a depth of 5.1 km beneath the surface. Deep well observations and magnetotelluric (MT) results place the main geothermal reservoir at a depth of >2 km [154,155]. Geochemical modelling of eruptive products suggests that silicic magmas at Aluto are generated and stored at depths of 4-6 km [156]. This suggests that the 5-km inflation source is most likely located between the upper boundary of the magmatic reservoir and base of the geothermal system or within a volatile rich cap in the magma storage zone. At Aluto there is no evidence in the deformation pattern to suggest contracting sources elsewhere around the complex, so it is inferred that inflation was fed from a depth greater than 5 km. As discussed above, deformation data alone do not allow us to unambiguously differentiate between whether the fluid is gas, aqueous fluid, magma, or a combination of these.   [153] argue that as most peralkaline volcanoes are considered to have a volatile-rich cap at around 5-6 km depth and that this zone is consistent with the modelled inflation source depth that magmatic fluid injection or intrusions into this cap is the most likely source mechanism for uplift at Aluto.
Further information is contained within the deflation signal at Aluto [153]. The roll-over from uplift to subsidence takes place over a timescale of a few months with analytical source models suggesting that the deflation deformation source is at 3.5 km. This short timescale and shallower depth are strongly suggestive of the migration of magmatic or hydrothermal fluids and degassing with fluids being removed from deep within the geothermal reservoir [157,158]. Gas geochemistry (CO 2 -δ 13 C [153]) suggests that the magmatic and hydrothermal reservoirs of Aluto are physically connected. Combining the gas and InSAR data with the other constraints on the system described above led   [153] to favour a coupled magmatic-hydrothermal process as the mechanism for the Aluto unrest. In their model: • The uplift is caused by a fresh magmatic fluid pulse or intrusion into a shallow crustal reservoir at 5 km. • The inflating source region is not well sealed and once pressure builds past a threshold fluids and gas may then leak into the geothermal reservoir and ascend to the surface along fault pathways, resulting in sharp deflation. magmatic intrusion as the main cause of unrest. It is also consistent with seismic data which has been interpreted to show signals from both the hydrothermal system and the deeper partially crystalline, magmatic mush [159][160][161]. It is interesting to note, however, that a recently compiled satellite time series implied no temperature change of the fumaroles along Aluto's major faults associated with the deformation cycles between 2004 and 2016 [162].
Rift maturity within the EAR varies, for example, from immature in the Kenyan Rift to intermediate-mature continental rifting in the MER to incipient seafloor spreading in Afar [163]. As well as understanding the signals and implications of unrest in general, key open questions remain about how the symptoms of unrest and its course and consequences (e.g., progression to eruption) change with such variations in rifting processes along the EAR. Drawing together the present literature with future studies combining multiple and extended time series datastreams from restless volcanoes in the Afar, MER and Kenyan rift segments will be key to developing a systematic understanding of volcanic unrest in this region [150,164,165]

(h) Santorini, Greece -subduction zone
The history of activity at Santorini volcano has been characterized by huge explosive eruptions, sometimes accompanied by caldera formation [168]. Activity between these large-scale events is characterized by intra-caldera edifice construction and lower intensity explosive activity. This is the type of activity that characterizes the current behavior of the system with the last eruption of the intra-caldera Kameni Islands occurring in 1950, 3 decades before the installation of the first instrumental monitoring network [168]. After 60 years of seismic and geodetic quiescence a new period of unrest began in January 2011. Multiple small earthquakes were detected within the caldera, many located along the Kameni line, a fault system thought to have been responsible for the delivery of magma to the surface during previous eruptions [169,170]. Permanent GPS stations and satellite interferometry showed that Santorini was deforming radially and inflating, with parts of northern Nea Kameni rising as much as 15 cm [171]. The rates of seismicity and deformation had returned to baseline levels by September 2012. Modelling of the deformation suggested that over the course of the unrest, two pulses of volume change of 14 -23 million cubic metres occurred at about 4 km depth (Fig. 3e), beneath the northern caldera [59,172]. Measurements of diffuse CO 2 degassing flux through a small part of the Kameni Islands showed insignificant changes ( Fig. 3e) with the onset of unrest [173]. Evidence for the magmatic involvement in the unrest comes from the coincidence of the modelled deformation depth with petrological evidence for the depth of the shallow magma storage zone [174], the H 2 /H 2 O and CO 2 -δ 13 C compositions of the Kameni fumarole gases that indicated an increase in temperature of the system [175] and increases in gas 3 He/ 4 He ratios indicative of new magma arriving in the system from depth [176]. This evidence does not rule out the hydrothermal system playing a role in driving the unrest however. Radon-CO 2 -δ 13 C systematics can even be used to suggest the evolved dacitic nature of the shallow intrusion, again consistent with petrological studies [173,174]. Data from previous Kameni eruptions suggest that this new pulse of magma delivered to the shallow storage zone is a significant fraction of the expected volume of the next eruption should it occur within the next few years [170].

Proposed check list for determining cause of volcanic unrest
Based on the review of physical mechanisms and case studies presented in the last two sections, we propose a check-list of criteria ( Table 2) that could be consulted during new episodes of unrest to assess the origin of unrest and prioritize new observations that could test hypotheses and help to diagnose the underlying origin and hazards associated with the unrest. Whilst we have focused on subaerial systems, many of the criteria in Table 2 [177] and hydrothermal fluid samples [106]. However, as on land, there are ambiguities in determining the cause of unrest -the same event on the East Pacific Rise in 1995 can be interpreted as of tectonic (non-magmatic) [178] and magmatic [179] origin.
It is critical that the data collected in Table 2 are available as a time series spanning multiple decades (e.g., Fig. 3). Not only do these time series help define when unrest begins and ends, but the time series reveal trends that are used to infer the origin of unrest [125]. A few of the key observations needed in Table 2 can be made from space (e.g., Fig. 4), albeit with limits on spatial and temporal resolution. However, many key variables at volcanoes (e.g., seismicity, gravity change, CO 2 and other trace gases) cannot be measured from space (or at least easily with current satellites) and so enhanced ground networks are needed. Unfortunately, very few volcanoes have several of the datasets in Table 2 available as a time series and only one or two have the entire suite. Clearly, there is a need to increase the observational record, and so we encourage further discussions to identify the key volcanoes around the world and prioritize the most critical multiparameter observations from Table 2 to be made. Not all datasets can be collected at all volcanoes -e.g., the fluxes of some gases may not be measurable above background, some volcanoes are open and do not have deformation [124], etc. But where available, as we have shown through the nine examples in section 3, combining geophysics and geochemical data from Table 2 together can be used to develop conceptual models. Yet observations alone will not answer the question of the origin of unrest as magmatic or non-magmatic in many cases -consider the volume of data generated for Campi Flegrei and the debate that still exists about the origin of unrest. Refined interpretations of the origin of unrest may require new datasets (e.g., drilling [115]) and numerical models that can integrate geophysical and geochemical data together [125,180].
The ultimate goal is to use the observations from Table 2 in BET. This is already happening at a few volcanoes that do have multi-decade records of observations (e.g., Fig.1), but newly identified restless volcanoes (e.g., Aluto) will take time to develop such records. An open question for research is the extent to which we can use the observations used in BET from one volcano and apply them to another [181].

Conclusion
A major question in volcano hazard forecasting is understanding the origin of unrest (i.e., magmatic vs. non-magmatic). The importance of this question is growing as we are increasingly able to measure unrest episodes through space and ground based observations. While it is rarely possible to precisely determine the origin of unrest without eruption, schemes like Bayesian Event Trees (BET) are able to incorporate the inherent uncertainty in assessing the cause of unrest into forecasts. We have compiled a set of observations ( Table 2) that might be developed in the future to assign probabilities to the origin of unrest, for example via a BET approach. From several case studies it is clear that combining multiple types of observations is criticalno one type of data uniquely pinpoints the origin of unrest. One goal for compiling our list is to help prioritize future observations at volcanoes where the cause of unrest is unclear. But surface observations will not be sufficient as ambiguities remain in interpretation. Further development of numerical models and drilling into these systems to test these models is likely required to advance understanding [120]. We encourage the community to refine and improve upon this list. Further, several restless volcanoes indicate unrest of both magmatic and non-magmatic origin is occurring simultaneously, in combination or in close proximity. Models of a heterogeneous magmatic systems where unrest in different subregions could have different origins is consistent with the TCMS paradigm [1].
Data Accessibility. All data presented here are publicly accessible through published scientific papers.
Authors' Contributions. The paper was conceived by MEP, TM and SM, and all authors contributed to writing and editing the paper.
Competing Interests. The authors declare that they have no competing interests.       Sandri et al., (2017) [13] to distinguish non-magmatic unrest.
Y/N means that the criteria is whether the the phenomena have been observed (yes/no). Information on the numeric probabilities is not available.