Corresponding author: P. Dalban Canassy, Laboratory of Hydraulics, Hydrology and Glaciology (VAW), ETH Zurich, 8092 Zurich, Switzerland.(email@example.com)
 In order to improve our understanding of the dynamics of potentially unstable steep glacier tongues, we monitored during summer 2010 the micro seismicity of Triftgletscher, Switzerland. Our system, comprising 8 three‒component seismometers coupled with the ice surface, was installed upstream of the glacier's tongue, which is likely to evolve toward an unstable regime. Complementary surface motion and proglacial runoff measurements allowed the icequake activity to be interpreted in terms of glacier dynamics and hydraulics. The strong contrast in seismic wave velocities due to the underlying bedrock was taken into account using a three‒dimensional (3‒D) velocity model, implemented in a nonlinear probabilistic location procedure allowing to accurately define the hypocenter uncertainty. We located 120 icequakes, with a focal depth accuracy that allowed distinguishing between shallow events (87 events) and near‒bedrock icequakes (33 events). The first motions of most of the deep events argue against pure shear sources expected in case of stick‒slip motion, and our suggested source mechanism is a superimposed tensile crack and shear dislocation. The analysis of surface strain showed that near‒surface events represent the opening and transverse extension of surface crevasses in a confluent flow regime.
 There exists a wide variety of seismic signals related to glaciers and their dynamics. A subset of these signals, which we will refer to as “icequakes,” is directly emitted by brittle deformation within the glacier ice or its underlying bed. Precise characterization of icequake sources therefore offers valuable insights into local stresses, faulting, and basal sliding. Icequake hypocentral location is a particularly important source parameter. The less accurately it is known the more difficult it is to interpret icequake records in terms of glaciological processes.
 The most common icequakes are near‒surface events (0–30 m deep), which mark the growth of crevasses [Neave and Savage, 1970; Mikesell et al., 2012]. Roux et al.  used these events to detect changes in the surface strain tensor in response to lake calving and drainage events on Alpine glaciers. A number of previous investigations of surface icequakes have provided only rough constraints on source depth [Dalban Canassy et al., 2012; Walter et al., 2008; Weaver and Malone, 1979; Roux et al., 2008]. However, a reliable determination of the focal depth of near‒surface events could reveal vertical crevasse maximal extension, and thus help estimate vertical variations of the Arrhenius constant of Glen's Flow Law [Paterson, 2002].
Deichmann et al.  were the first to constrain reliable hypocenters for icequakes near the bed of a Swiss Alpine glacier. Such basal events occur less frequently than surface crevassing events and usually after periods of water‒enhanced glacier sliding, when the glacier recouples to its bed [Walter et al., 2008]. More quantitative interpretation of basal icequakes is inhibited by high vertical uncertainties of hypocenters, which do not allow for distinction between events occurring at the ice‒bedrock interface, within the underlying substrate or within the overlying ice layers.
 Icequakes measured near Antarctic ice streams [Smith, 2006; Danesi et al., 2007] have been interpreted as stick‒slip motion, which is driven by the accumulation and release of elastic strain resulting in sudden sliding episodes of the ice mass along the underlying bedrock [Winberry et al., 2013]. Such result was derived based on compelling geodetic measurements [Wiens et al., 2008; Walter et al., 2012] or epicenters and first motion polarity associated with P waves [Zoet et al., 2012b; Anandakrishnan and Bentley, 1993]. However, uncertainties in source depths inhibit an accurate location with respect to the ice stream base, which, in turn, leaves room for speculations on the nature of stick‒slip motion.
 In contrast to shallow seismicity, the source mechanisms of deeper icequakes remain elusive and are currently debated in the glaciological community. Using moment tensor inversions, Walter et al.  show that a subset of basal events beneath Gornergletscher (Switzerland) is related to tensile fracturing associated with opening of cracks. On the other hand, Zoet et al. [2012a] suggest that seismicity recorded near the base of various glaciers is related to shear dislocation associated with stick‒slip motion. Classifying under which glaciological conditions the two proposed mechanisms occur requires, once again, reliable constraints on hypocentral locations and meaningful uncertainties thereof.
 Finally, icequake hypocenters also play a crucial role in the context of breaking‒off events of hanging glaciers [Faillettaz et al., 2008; Faillettaz et al., 2011]. Damage theory predicts crevasse development and shear failure above the frozen bed prior to the breaking‒off [Pralong et al., 2005]. Therefore, accurately located icequakes could serve as indicators for imminent breaking‒off events. So far, seismic early warning systems have relied on single‒station‒derived event statistics involving no spatial resolution of icequake sources [Faillettaz et al., 2008; Faillettaz et al., 2011].
 The above examples illustrate that hypocenter locations are key to correct interpretation of the occurrence of icequakes. Events recorded on four or more stations can in principle be located, especially since glacier ice is often highly homogeneous [e.g., Deichmann et al., 2000]. However, as study sites are often difficult to access and approximate icequake locations are usually unknown prior to instrument deployment, seismometer network geometries are rarely ideal. This is a considerable drawback, because source‒station geometry determines location uncertainties, together with inaccuracy in determining arrival times of seismic phases (henceforth “picking”). Uncertainties may furthermore increase, as the seismic velocity model of the glacier bed is usually poorly known at best.
 For the analysis of icequake source mechanisms in terms of glaciological processes, dynamic and hydraulic measurements represent a useful complement to hypocenter location. The relationship between deep events and basal sliding fluctuations pointed out by Walter et al.  could not have been derived without precise GPS and borehole pressure data. Similarly, Roux et al.  demonstrated the predominant control of glacier surface deformation on recorded shallow seismic activity using detailed geodetic measurements of the surface displacements.
 The present study focuses on seismic data from Triftgletscher, Switzerland, which has been monitored since 2007 due to its steep tongue prone to break‒off [Dalban Canassy et al., 2011; Dalban Canassy et al., 2012]. Using continuous seismic records from an on‒ice campaign network, we determine icequake hypocenters with a nonlinear probabilistic approach [Lomax et al., 2000] and a three‒dimensional P wave velocity model that accounts for the complex topography of bedrock and glacier surface. We assess location accuracy and precision by relocating blasts with known locations and by analyzing traditional error ellipsoids and confidence regions of the probability density function (PDF). Our results show that the PDF for icequakes located inside the network is well‒defined, and corresponding location accuracy and precision are small (in the range between a few meters to a few tens of meters). Icequakes located outside of the network show a complex PDF. For these events traditional error ellipsoids provide a poor approximation of location precision. Nevertheless, we achieve a location accuracy of a few tens of meters for these events. Our locations confirm icequake sources near the glacier surface as well as near the bed. With the help of seismic waveform characteristics, glacier surface motion, geodetic measurements, and proglacial runoff data, we interpret the seismic location results in terms of glacier dynamics.
2 Study Site
 Triftgletscher glacier is located between the Gadmer and Hasli valleys in the Bernese Alps (Switzerland) (Figure 1a). It flows from 3380 to 1651 m a.s.l. over a distance of 5.1 km and covers a surface of 15 km2. The glacier tongue extends from 2350 to 2050 m a.s.l on a 35° steep section and ends in a basin bordered on the northern side by a riegel.
 In absence of englacial temperature measurements, the thermal regime of the glacier in the study area remains unclear. However, some clues can be considered for allowing us to assume with a fair confidence that the ice is temperate at our study site: First, the maximal elevation of the accumulation area of Triftgletscher (3380 m a.s.l), which appears relatively low with respect to other glaciers where a polythermal regime was observed [Suter and Hoelzle, 2002]. Second, we could not observe any refreezing closure in three boreholes drilled to the bedrock vicinity and filled with water over the 3 weeks of observation. In this context, and in the absence of definitive evidences, we assume the ice in the investigated area to be temperate.
 Over the last 15 years, Triftgletscher has retreated substantially and a proglacial lake containing 5×106 m3of water [Grischott et al., 2010] formed in front of the glacier terminus. The retreat resulted in the destabilization of the tongue, increasing its flow speed and causing frequent avalanches on both sides of the terminus [Dalban Canassy et al., 2012]. If the entire ice tongue slides into the lake or if volumes of ice avalanches exceed 106 m3, potential flood waves are a threat to the downstream Gadmertal valley [Dalban Canassy et al., 2011].
3 Field Measurements
 The goal of the field experiments was to characterize the dynamics of the potentially unstable tongue of Triftgletscher. However, rugged topography prevents direct on‒ice monitoring of the glacier tongue [Dalban Canassy et al., 2012]. Consequently, field experiments were focused on the weakly sloped area upstream (south) of the unstable tongue (Figure 1a).
3.1 Seismic Monitoring and Blasts
 Between 14 July and 5 August, a seismic network was installed at altitudes between 2400 and 2350 m a.s.l (Figure 1b). The instrumentation consisted of one Geometrics GEODE seismic recorder connected to 8 three‒component, 8 Hz borehole seismometers (Geospace GS‒11D) installed in ca. 1 m deep boreholes. Each sensor continuously recorded the three components of ground velocity (vertical, north‒south, east‒west) at a sampling frequency of 1000 Hz. The presence of surface crevasses as well as cables between sensors and GEODE limited the network aperture. Six sensors were arranged in an approximate circle. The remaining two sensors were placed north of the circle within a crevasse field near the head of the unstable glacier tongue.
 In order to constrain englacial and subglacial seismic velocities, three explosive charges (Riodin, ≈8 kg) were detonated at the bottom of boreholes, (b1, b2, and b3 with respective depths of 106, 120, and 104 m). In addition, a fourth explosion (b4) (≈50 g) was detonated at the glacier surface (Figure 1b).
3.2 Surface Motion
 In order to measure glacier surface motion, two transverse profiles of six and five stakes respectively were installed into 1 m deep boreholes within the study area (Figure 1b). The positions of the top of each stake were surveyed with differential GPS (Leica System 500) once or twice a day depending on weather conditions. Following Sugiyama and Gudmundsson , we estimate positioning errors of 3 and 5 mm in the horizontal and vertical direction, respectively.
3.3 Glacier Geometry
 In order to approximate the glacier's surface geometry, we use a Digital Elevation Model (DEM) derived from aerial photogrammetry (Swisstopo pictures) (Figure 1b). At a grid spacing of 5 m, the estimated uncertainty is ±0.2 m.
 Bedrock topography was determined from helicopter‒based radar surveys carried out in spring 2008 and winter 2012. The data were acquired with the University of Münster Airborne Ice Radar (UMAIR) [Blindow, 2009]. The system has a central frequency of 30 MHz and is designed as a sling load from a helicopter which makes it suitable for smaller glaciers with steep surrounding terrain. The procedure is described in detail in Ryser et al. . Following the method of Gabbi et al. , 15 profiles (red lines in Figure 1b) were processed to derive the ice thickness in the study area (see example in Figures 2a and 2b). Finally, bedrock elevation at the profile points was deduced by subtracting ice thickness from the glacier surface elevation.
 We subsequently applied the ice‒thickness estimation method (ITEM) [Farinotti et al., 2009] in order to interpolate bedrock topography between radar measurements. This technique takes into account ice deformation via inversion of the Glen's flow law [Glen, 1955]. The associated uncertainty is ±15% of the ice thickness. The results reveal a basin shape in the study site, which is open to the north toward the glacier tongue (Figures 2c and 2d). With steep slopes at its walls, the basin contains a nearly flat section of several hundred square meters in its middle (Figure 2c). Within our study site, the ice thickness ranges from around 40 m on the sides to 140 m in the central part (Figure 2d).
 A water pressure sensor (Keller DCX‒22) was installed in the proglacial stream about 10 m from the tongue to measure water level every 15 min (red point in Figure 1a). Atmospheric pressure influence is removed with the help of additional pressure measurements outside the stream within close proximity. The uncertainty is 4.5×10−4 bar. We assume no significant water transit time between the tongue and the study area, as well as no significant temporal changes of the stream bed topography. Moreover, there are no stream tributaries other than the glacier tongue. This offers the possibility to take changes in stream height as a proxy for runoff in the glacier catchment.
4 Icequakes Detection and Location
4.1 Seismic Data Processing
 As a first step, icequakes were detected in the continuous data using a standard short‒term window over long‒term window average (STA/LTA) trigger algorithm, [e.g., Allen, 1978] with window lengths of 80 and 800 ms, respectively. In this procedure, the root mean square of the two concurrent windows are compared and an event was declared when the resulting ratio exceeded a trigger threshold (set here equal to 3) on at least five stations within a 2 s time period. These conservative settings allowed detection of rather strong events with high signal‒to‒noise ratios that facilitated manual picking of first arrivals. A total of 241 events were detected using these settings. Local tectonic earthquakes as detected by the Swiss Seismological Service (http://www.seismo.ethz.ch/eq/latest/index_EN) were removed leaving a total of 209 events for further analysis. Event durations ranged from 0.1 to 1 s, with a mean of 0.3 s, which are significantly shorter than those typically observed for ice or rock falls (typically >1 s) [Roux et al., 2008; Dalban Canassy et al., 2012]. Consequently, it can be expected that our data set is not contaminated by signals from ice or rock falls. For each event, P wave arrival times were manually picked on vertical components at all stations. This was achieved by looking for clear changes in the signal‒to‒noise ratio and in the frequency content between signal and noise. Each arrival time was associated with an uncertainty of 3 ms, which was derived by looking at waveforms with different signal‒to‒noise ratios and different frequency content of the first arriving phase (Figure 3).
4.2 3‒D Seismic Velocity Model
 In previous studies, different types of seismic velocity models were used to locate icequakes. The use of a homogeneous velocity model with a constant velocity for ice may be justified to locate shallow icequakes [Roux et al., 2010] or deep icequakes if only direct wave arrivals (arrivals traveling directly from the sources to the receiver through the ice) are considered [Deichmann et al., 2000; Walter et al., 2008]. If both direct and critically refracted arrivals (arrivals that are critically refracted at the underlying bedrock and travel along the ice‒bedrock interface) are used to locate icequakes, results with a 1‒D velocity model will be inaccurate and a 3‒D velocity model should be used that takes into account the bedrock topography beneath the glacier [Roux et al., 2008].
 To investigate whether critically refracted arrivals play a significant role in our data set, we analyzed data recorded from a blast (blast 2) close to the bedrock by plotting the vertical components of the seismic signal recorded at each sensor against the hypocentral distance (Figure 4). While arrivals at stations s3 to s8 correspond to the direct P wave, arrivals at stations s1 and s2 can be clearly associated with a critically refracted phase, as indicated by a different slope in the arrival‒time‒distance plot (Figure 4). This suggests that first‒arriving refracted waves can be expected in our data, and that a 3‒D P wave velocity model reflecting the bedrock topography should be used to locate icequakes in our study region. The 3‒D P wave velocity model used in our study was developed in three steps:
 Digital elevation models of bedrock and glacier surface topography were derived from radar measurements (see 3.3) and aerial photographs, respectively.
P wave velocities of air and the granitic bedrock were set to 0.333 kms−1and 5 kms−1[Walter et al., 2010], respectively. P wave velocity of ice (3.68 kms−1) was derived from arrival time data of direct arrivals recorded from blast b4 located at the surface (Figure 1b). Our estimate is close to that found by similar studies on Alpine glaciers [Walter et al., 2009; Deichmann et al., 2000].
 We parameterize our velocity model by cubes (cells) with constant P wave velocity (Figure 5a). Total dimension of our model is 600 m, 550 m, and 300 m in east‒west, north‒south, and vertical directions, respectively. We chose a cell size of 5 m in horizontal and vertical directions, which corresponds to 120, 110, and 60 cells in east‒west, north‒south, and vertical directions, respectively. This value represents a compromise between computation time of travel times, arrival time precision, and resolution of the steep bedrock topography. A P wave velocity is assigned to each corner of each cell depending on whether it is located in the air, ice, or bedrock (Figure 5a). The P wave velocity at the center of each cell is then obtained by linear interpolation between P wave velocities at all eight corners.
 Our approach yields a 3‒D model with constant P wave velocities for air, ice, and bedrock as defined by glacier surface and bedrock topographies (Figure 5b). The validity of the assumption to represent ice and bedrock with constant velocities depends on the ratio between dominant wavelength and size of expected fractures, which should be greater than one [Gischig, 2007; Roux et al., 2008; Roux et al., 2010]. Indeed, the presence of fractures filled with air and/or water may locally cause significant slow‒down in P wave speed [Endres et al., 2009]. On Triftgletscher, we observed dominant frequencies in the range of 10–50 Hz, which corresponds to wavelengths of 75–360 m and 100–500 m in the ice and bedrock, respectively. Crevasses along the central flow line of Triftgletscher showed a maximum width of 15–20 m. Furthermore, crevasses on temperate Alpine glaciers are likely not deeper than 25–30 m [Paterson, 2002, p.189]. These values are significantly smaller than expected wavelengths in the ice, which means that cracks do not significantly interfere in the propagation of the recorded seismic waves. Unfortunately, we do not have any information on density and size of potential cracks into the bedrock. Given the expected wavelengths of 100–500 m however, and the size of our study area, we believe that potential cracks in the bedrock do not play a significant role.
4.3 Nonlinear, Probabilistic Icequake Location
 Preliminary location results provided by a traditional linearized location technique [Lee and Stewart, 1981] roughly indicated a number of icequakes clusters located far outside of the array. For these events, location uncertainties can be highly nonlinear, i.e., nonellipsoidal in shape [Husen et al., 2003], and, consequently, a traditional ellipsoid is not a valid approximation of location uncertainties. We therefore used a nonlinear probabilistic approach as implemented in the software package NonLinLoc (NLLoc) [Lomax et al., 2000] to locate icequakes in our study. NLLoc computes the posterior probability density function as defined by Tarantola and Valette  and Moser et al. . The posterior PDF represents a complete, nonlinear probabilistic solution to the location problem, including information on uncertainty and resolution. It can be irregular in shape and show multiple maxima [e.g., Lomax et al., 2000; Husen et al., 2003].
 In NLLoc, synthetic travel times are computed using a finite‒difference solution of the Eikonal equation [Podvin and Lecomte, 1991], which has been proven to work reliably in velocity models with strong velocity contrasts as ours. The posterior PDF is computed using the Oct‒Tree importance sampling algorithm [Curtis and Lomax, 2001], which provides an efficient and reliable sampling of the solution space [Husen et al., 2003]. We use a picking error of 3 ms (as estimated in section 4.1) to represent our measurement errors; we decided not to consider model errors since the original formulation of Tarantola and Valette  requires them to be Gaussian distributed, which is likely not the case for most studies [Pavlis, 1986].
 We represent location results by means of a maximum likelihood hypocenter location and associated confidence regions describing location uncertainty. The latter were derived for a single event (Figure 6) as well as for multiple events (Figure 7). Confidence regions were obtained by computing the cumulative sum of sorted PDF values, normalized with respect to the maximum PDF value and finding the PDF value that corresponds to a given confidence level [Moser et al., 1992]. Each confidence region is described by a 3‒D volume, bounded by an isosurface derived from the associated PDF value. It corresponds to the volume where the icequake has a given probability to be located, as defined by the confidence level. Shape and extent of this volume is defined by the location uncertainty. For multiple events, we compute a cumulative PDF following the same approach. In this case the confidence region does not outline the volume for a single event but for multiple events, for which individual confidence regions overlap.
4.4 Assessing Location Accuracy and Precision Using Relocation of Blasts
 The inherent coupling between seismic velocities and hypocenter locations demands to assess the reliability of the chosen velocity model [Husen et al., 2003]. We define location accuracy as the difference between true and relocated hypocenter locations, whereas location precision is given by formal location errors as computed by the location program [Husen and Hardebeck, 2010]. Location accuracy is mainly influenced by the quality of the chosen velocity model. Uncertainties in picked arrival times in combination with number and spatial distribution of stations, for which arrival times could be picked, mainly affect location precision. We use data from three blasts (b2, b3, b4) to compute differences between true and relocated hypocenter locations and to assess location accuracy. Relocated hypocenter locations were computed using our 3‒D P wave velocity model. We assess location precision by analyzing shape and size of confidence regions for each blast, as well as the semiaxes of the associated 68% confidence ellipsoids derived from traditional covariance matrix [Lomax et al., 2000]. It is important to note that assessing location accuracy and precision by relocating blasts is only valid for icequakes that occur at similar locations as the blasts. In our study the chosen blasts are located at the surface and inside the network (blast b4), close to the bedrock and inside the network (blast b3), and close to the bedrock but outside the network (blast b2).
 Differences between true and relocated hypocenter locations for blasts b2, b3, and b4, as well as lengths of semiaxes of associated 68% confidence ellipsoids are given in Table 1. As can be inferred from Table 1, location accuracy is in the range of a few meters for blasts b3 and b4, which locate inside the network. As expected, location accuracy is much poorer (in the range of 15–20 m) for blast b2, which locates outside the network. Although location accuracy is worse for blast b2, it is still in the range of estimated uncertainties in bedrock topography (see section 3.3).
Table 1. Location Accuracy (Discrepancy Between the True and Relocated Hypocenter (m)) and Precision (Semiaxes of the 68% Confidence Ellipsoid (m)) for Blasts b2–b4
Epicenter Accuracy (m)
Focal Depth Accuracy (m)
Semiaxes of 68% Confidence Ellipsoid (m)
16, 20, 24
15, 26, 55
33, 63, 118
 Confidence regions for each blast and associated 68% confidence ellipsoids are shown in Figure 6. Blasts b4 (Figures 6a and 6b) and b3 (Figures 6c and 6d) show well‒defined hypocenter locations as indicated by confidence regions that are ellipsoidal in shape and compact, which is underlined by a good agreement with the 68% confidence ellipsoids showing maximal horizontal semiaxes of 26 and 20 m, respectively. Confidence regions are larger in focal depth for blast b3 indicating that focal depth for deep icequakes inside the network is less well‒constrained than for shallow icequakes, which is supported by the associated 68% confidence ellipsoids with a vertical semiaxis twice as large for blast b3 (24 m for b4 and 55 m for b3). Nevertheless, focal depth for b3 is still constrained within the lower half of the glacier consistent with a deep source location.
 Confidence regions for blast b2 show a rather complex topography (Figures 6e and 6f), which is consistent with a hypocenter location outside the network [Husen et al., 2003]. Due to the poor azimuthal coverage, the epicenter location is poorly constrained in southwest to northeast direction, associated with a maximal horizontal semiaxis of the 68% confidence ellipsoid of 118 m. Focal depth for blast b2 is surprisingly well‒constrained with a clear maximum close to the bedrock; the length of the vertical semiaxis of the associated 68% confidence ellipsoid is 33 m. Note that due to the nonellipsoidal shape of the confidence regions, the 68% confidence ellipsoid provides only a poor quantification of the location precision, particularly in focal depth (Figure 6f). For this reason, we decided not to use 68% confidence ellipsoid to quantify location precision. Instead, we will use in the following the 70% confidence region computed for individual and multiple events. The observation that focal depth is relatively well‒constrained for blast b2 is likely caused by the fact that, at stations s1 and s2, refracted phases arrive first (Figure 4). Due to their downward oriented take‒off angles, these arrivals provide important constraints on focal depth. The observation that the confidence regions suggest a higher probability of the hypocenter location to be in the bedrock is clearly an artifact and a consequence of our model parameterization, which does not account for uncertainties in bedrock topography. In our model we place the bedrock interface at a certain depth and model it as first‒order discontinuity. As a consequence, arrivals at more distant stations (s1 and s2) are modeled as critically refracted phases, which due to their downward oriented take‒off angles yields hypocenter locations close to the bedrock interface. In reality, however, the bedrock interface is located deeper since the shot b2 was located clearly in the ice. This discrepancy between modeled depth and true depth of the bedrock interface leads to hypocenter location just below the bedrock interface, which in reality was located above the bedrock interface.
 Our results show that we can achieve a location accuracy of a few meters for shallow and deep icequakes inside the network. For these events hypocenter locations are well‒constrained as indicated by confidence regions that are ellipsoidal in shape and compact. Location accuracy increases to 15–20 m for deep icequakes located outside the network but is still within the range of uncertainties in bedrock topography. Due to arrivals of refracted phases at distant stations, focal depth for these events is relatively well‒constrained with a clear maximum close to the bedrock. Epicenter location for these events is poorly constrained, in particular, in southwest to northeast direction. We, therefore, conclude that our combination of 3‒D velocity model and network geometry allows to distinguish between shallow and deep icequakes based on their hypocenter location and associated confidence regions.
5 Icequakes Spatial Distribution
 Of the 209 initially detected icequakes, 195 have a maximum likelihood location within the study area. These events were analyzed in the following way:
 For each event, the PDF was visually inspected and 120 well‒constrained hypocenters were retained. For these hypocenters, the PDF was compact in shape and showed a clear single maximum.
 From these 120 events, 33 and 87 icequakes were classified as deep and shallow, respectively. We defined icequakes as shallow and deep if the corresponding 70% confidence region was located in the bottom half and in the top half of the glacier, respectively.
 We performed a simple spatial clustering analysis of all the 120 icequakes by computing the 70% confidence level of the cumulative PDF following the above mentioned procedure (see section 4.3). The rational for this analysis is that icequakes with overlapping confidence regions should be associated to the same spatial cluster since their hypocenter locations cannot be separated. Following this logic, it is not meaningful to split two clusters if their corresponding hypocenter locations and uncertainties overlap. The resulting isosurface of the 70% confidence level of the cumulative PDF shows four regions that can be clearly separated in space (Figure 7a). Two of these regions (labeled SC1 and SC2) locate in the upper half of the glacier; the other two regions (labeled DC1 and DC2) locate in the bottom half of the glacier. Following our definition for shallow and deep icequakes (see above), we conclude that icequakes located in the regions SC1 and SC2 are shallow (at a confidence level of 70%) and that icequakes located in regions DC1 and DC2 are deep (at a confidence level of 70%). Furthermore, each region forms a spatial cluster that is clearly separated from the other clusters.
 In a final step, we cross‒correlated waveforms of all the 120 icequakes to identify families of events that show a high degree of similar waveforms (cross‒correlation coefficient >0.8) (see section 5.3). Icequakes within these families are likely to have similar source mechanisms and locate very close in space.
 We emphasize that the volume enclosed by the isosurface at the 70% confidence level of the cumulative PDF (Figure 7a) represents the combination of the location uncertainty of individual icequakes and the actual size of each icequake cluster. The fact that epicenters included in cluster SC2, which locates within the network, are best constrained suggests that location uncertainties dominate. In order to further evaluate this matter, we identified a highly coherent icequake family within cluster SC1 (see family 6 in section 5.3). For this family, all event pairs exhibit a cross‒correlation coefficient of 0.9 or higher. We interpret such high coherency as evidence that the icequake sources occurred essentially at the same location, as even minor variations in path effects would reduce waveform similarities. Moreover, this highly coherent event family includes 22% of all SC1 events, and the associated cumulative PDF volume enclosed by the 70% confidence level isosurface amounts to 56% of the volume of the entire cluster. In other words, even if we confine cluster SC1 to those events, whose locations nearly coincide (based on waveform similarity), the volume enclosed by the 70% confidence level isosurface shrinks by less than 50%. This further suggests that location uncertainties are the main control of the size of the volumes enclosed by the isosurfaces shown in Figure 7.
 Significant difference in the location uncertainty can be noticed from one cluster to another. Among the deep clusters, DC2 is substantially better constrained than DC1 as indicated by the smaller volume of the 70% confidence level of the cumulative PDF. Likewise, clusters SC2 is better constrained than cluster SC1. Moreover, clusters DC1 and SC1 location uncertainties are strongly elongated toward the western boundary, while the associated maximum likelihood locations locate close to the seismometer array. Such orientation of location uncertainties is a consequence of array geometry and icequake epicenter locations at the periphery of the array.
 Figures 7b and 7c show longitudinal and transversal cross sections of the investigated area, with projections of the maximum likelihood points of epicenters located in a 15 m band width. In both profiles, deep and shallow hypocenters appear to be unambiguously disconnected, which supports the above mentioned clustering distribution and validates that both types of icequakes can be distinguished with our location procedure.
5.1 Deep Icequakes
 Deep sources represent 27.5% of the data set, which makes Triftgletscher a particularly suitable study site for near‒bedrock emissivity investigation. Vertical component seismograms of a representative deep icequake (from DC1), examined by means of particle motion analysis, as well as hypocentral locations are presented in Figures 8a, 8b, and 8c. The waveform contains impulsive, high‒frequency P and S arrivals, and a maximum P wave amplitude on z component. It lacks a notable surface wave (Figure 8a). These are the main characteristics of icequakes with focal depths well below the vertical extension of the surface crevasse zone [Deichmann et al., 2000; Walter et al., 2008]. Despite their clear onsets, we did not use any S waves in the location procedure, because blast b4 did not provide any information about shear wave speed.
 The NLLoc location for the deep event (Figure 8b) is poorly constrained along the line connecting the maximum likelihood epicenter and the center of the seismometer circle. This is a result of array geometry and the fact that the icequake epicenter located at the edge or outside the array. The focal depth (Figure 8c) exhibits a maximum likelihood point at the ice‒bedrock interface. However, the PDF shows a higher probability that the event locates just beneath the ice‒bedrock interface than within the basal ice layer. The same was observed for blast 2 (Figure 6f). As the blast clearly occurred in the ice, this suggests that the high location probability below the bedrock is an artifact due to our model parameterization that does not account for uncertainty in bedrock topography (10–20 m). Moreover, our deep icequake seismogram shares qualitative characteristics with seismograms of intermediate (below surface crevasse zone, but well above glacier bed) events [Walter et al., 2009; Deichmann et al., 2000]. We, therefore, conclude that the deep event most likely locates at the ice‒bedrock interface or within the basal ice, while there exists a non‒negligible probability that it may also locate as far as 50–100 m below the glacier bed.
 The epicentral distribution of the deep sources (maximum likelihood points and 70% confidence surface of cumulative PDF) is shown in Figure 9a. Apart from four outlying epicenters, events scatter within the two deep clusters DC1 and DC2. Cluster DC1 consists mainly of events detected at the beginning of the monitoring period, while DC2, which includes less events, was mainly active at the end. The absence of deep event locations with epicenters located inside the network deserves further comment at this point. The well‒constrained locations of deep and shallow blasts (Figure 6) show that location close to the network center is possible. Furthermore, we visually inspected all events detected with our STA/LTA algorithm finding no further signals with deep icequake characteristics. We, therefore, conclude that the absence of deep icequake locations beneath our seismic network does indeed reflect a lack of deep seismicity in this region.
5.2 Shallow Icequakes
 Shallow icequakes typically pose the vast majority of seismic events on Alpine glaciers [Deichmann et al., 2000; Walter et al., 2008; Roux et al., 2010; Dalban Canassy et al., 2012; Mikesell et al., 2012]. Figure 8d shows a typical shallow icequake seismogram. It clearly shows around 5.5 s the high amplitude of the dominant phase associated with surface wave arrival (Rayleigh wave, denoted R in Figure 8), which is the main waveform characteristic of a near‒surface icequake [e.g., Deichmann et al., 2000; Faillettaz et al., 2008]. For the event shown, the associated epicenter lays inside the network and is therefore well‒constrained (Figure 8e). At a 90% confidence level the hypocenter locates within the upper half of the glacier thickness, with a maximum likelihood about 10 m below the surface Figure 8f.
 Within the study area, detected shallow events mainly concentrate inside and west of the network in the two clusters SC1 and SC2 (Figure 9b). SC1 lays outside of the array near crevasses ends. It mostly includes icequakes which occurred during the second half of the investigation period. Cluster SC2 is located inside the network. It shows some elongation oriented from south‒west to north‒east. Most of its events were recorded toward the beginning of the monitoring period. The sources are concentrated near opening crevasses, but not specifically near the extremities, as opposed to the SC1 events.
5.3 Waveform Similarities
 In order to quantify the similarity of our waveforms, we cross‒correlate all pairs within the set of 120 located events. We use the vertical component records of sensor s1 because of its central position with respect to our clusters. The cross‒correlation coefficient ranges between 0 and 1 with the latter indicating a perfect match. Results are presented by means of a dendrogram computed using an average linking method (Figure 10a). The dendrogram shows that on average all shallow events correlate poorly with deep events (cross‒correlation coefficient of 0.35 or less, see red cross in Figure 10). This indicates that such a relatively simple similarity analysis based on one station is an efficient means to automatically separate an icequake data set into shallow and deep event signals.
 Clusters DC1, DC2, and SC1 contain subclusters (“families”) whose members have an average mutual cross‒correlation coefficient of 0.8 or higher (colored patches in Figure 10a). Seventy‒five percent of the deep events in DC1 are organized in four families composed of two to six events. Sources in DC2 do not fall into families according to our definition of a mutual cross‒correlation coefficient exceeding 0.8. The three deep events in family 16 (cluster DC1) show a surprisingly low correlation with all other events, even those belonging to the same cluster. Concerning shallow icequakes, there exist seven families including 68% of shallow cluster SC1, with as much as nine and seven events (families 6 and 7 respectively). Fifty percent of all events belonging to shallow cluster SC2 are part of four families, with the largest one including five icequakes (family 1). Note also that three families of respectively two, two, and three shallow events (brown patches in Figure 10a), neither located in SC1 nor SC2, are found. As they correspond to isolated sources, we did not consider them in the similarity analysis.
 The observation of event families points to different source mechanisms within a cluster. In case of shallow events, this possibility appears well‒supported given the clearly separated crevasses in close proximity of SC1 and SC2. Our threshold choice of 0.8 for the average mutual cross‒correlation coefficient defining a subcluster family is arbitrary. However, it is in line with the message of Figure 10a that event groups of a single cluster show different degrees of similarity. In order to explain subclustering and as a basis for the following discussion, we note that both different fault mechanisms (source effect) and different source locations (path effect) can potentially effect waveform similarity. Consequently, highly coherent icequake families are likely a result of repeated fracturing on a single fault, as it may be illustrated by events of families 7 and 14 (Figures 10b and 10c).
 Finally, we noticed differences in time intervals between consecutive events. For example, the deep events of family 14 occur every 4 to 13 min, whereas the intervals of shallow family 7 range from a few hours to more than 1 day. Explanations for such differences in interevent time intervals are speculative at this point, but may be connected with different stress regimes at the glacier surface and close to its bed.
6.1 Near‒Bedrock Seismicity Versus Runoff
 Figure 11 shows the occurrence times of deep icequakes throughout the study period (blue and magenta stars) superimposed on water level measurements (given in hydrostatic pressure) of the proglacial stream at the glacier tongue. The time series show that deep icequake detection was suppressed during peaks in stream height. Instead, deep icequakes occurred primarily after stream height peaks. This is also true for the most prominent peak around 23 July during which substantial surface uplift was measured (dashed line in Figure 11).
 At this point we assume that changes in water level of the proglacial stream represent changes in glacier runoff. Furthermore, we assume that during runoff peaks, englacial water storage increases to the point where the subglacial drainage system is no longer efficient enough to evacuate the water input [Bartholomaus et al., 2008]. Consequently, the subglacial water pressure increases. Likewise, we deduce that during falling or low runoff, the subglacial water pressure falls.
 In this sense, the stream height and uplift time series suggest that during the monitoring period, the subglacial drainage system was channelized rather than distributed: only during the largest runoff peak (corresponding to the stream height peak around 23 July), substantial glacier uplift occurred. During all other peaks, runoff was evacuated quickly enough to keep subglacial water pressures below ice‒overburden pressure [Schoof, 2010].
 We do acknowledge that the seismic detection time series may be biased toward periods of low stream height: these times correspond to low glacier runoff during which the seismic background noise may be relatively low. Consequently, the trigger algorithm is more sensitive, which explains the increased number of detected events during most stream height minima (Figure 11, top). Nevertheless, our result that deep icequakes occur primarily during falling subglacial water pressures agrees with experiments from Gornergletscher [Walter et al., 2008].
6.2 Deep Icequake Fault Mechanisms
 In order to constrain the source mechanisms of our deep icequakes, we determined the first motion polarities for each event (Figure 12). Note that the high signal‒to‒noise ratio characterizing the deep events onset as well as their short mean duration (0.3 s) prevent them to be obscured by other events and therefore assure a reliable constraint of the first motion. Out of the 29 events, 21 exhibit dilatational (“down”) first motion at all stations. Seven events have compressional “up” first motion at all stations and one event shows a polarity mixture. We interpret the predominant single polarity character (i.e., same polarity at all stations) as evidence for strong isotropic components of the underlying moment tensors [e.g., Julian et al., 1998]. Coseismic volumetric changes of the source region due to tensile faulting provide the most straightforward explanation for such a strong isotropic component. In this view, a tensile crack opening emits purely compressional first arrivals and a tensile crack closing emits purely dilatational first arrivals. This source mechanism has been determined from full waveform moment tensor inversions of a deep icequake cluster beneath Gornergletscher (Switzerland) [Walter et al., 2010]. Like the deep events presented in the present study, the Gornergletscher icequakes occurred primarily during falling or low water pressures [Walter et al., 2008]. However, the mixed polarity of our event 22 (Figure 12) cannot be explained with a pure tensile crack moment tensor. Instead, we suggest that in addition to the tensile crack component, there should exist some shear component which produces mixed polarity first motions [Julian et al., 1998] in case of event 22. However, we can only state that for 28 out of our 29 deep icequakes, the tensile crack component dominates the first motion polarity.
 We emphasize that first motion analyses are subject to ambiguities if station coverage is limited. This is the case for our deep icequakes as they locate outside the seismic network. Under these conditions, the quadrantal distribution of first motion polarity, which is characteristic for pure shear dislocations, may not reveal itself, because all recording stations may locate within the same quadrant of the focal sphere [Aki and Richards, 1980]. Although we consider it unlikely that this be the case for 28 of the 29 analyzed events, we cannot reject the possibility of pure shear sources based on first motion analysis alone. Nevertheless, we can check if the first motions are consistent with pure shear dislocations during seismogenic stick‒slip motion across the glacier bed. Such events may be a prevalent phenomenon beneath Antarctic ice streams, during which the glacier suddenly slips across its bed, analogous to earthquake rupture [Anandakrishnan and Bentley, 1993; Smith, 2006; Danesi et al., 2007; Walter et al., 2012; Winberry et al., 2013].
 In order to model the first motions of a stick‒slip motion, we employ pure basal shear sources, whose fault planes are parallel to the local radar‒derived glacier bed. The shear slip direction (given by “rake” parameter, as described, e.g., in Aki and Richards ) is chosen such that coseismic horizontal motion coincides with the displacements of stakes 10 and 7 for events in clusters DC1 and DC2, respectively. Figure 12 shows first motion polarities of such a hypothetical pure shear source (“beach ball plots”). The results indicate that for most events, the projected “up” first motions (red crosses) are not confined to the black patches (compressional) of the beach ball diagrams. Similarly, the projected “down” first motions (green crosses) are not confined to the white patches (dilatational) of the beach ball diagrams. This implies that the first motion polarities of most events are incompatible with the stick‒slip mechanism, unless the assumed fault geometry disagrees substantially with the surface and/or radar‒derived bed topography. For events 1, 10, and 15, the compressional first motion polarity agrees with the stick‒slip mechanism at five or six out of the eight stations. Varying the assumed rake by ±10° can reduce the number of stations with incorrect polarity to one.
 In summary, first motion analysis alone cannot fully determine icequake source mechanisms for our network geometry. For example, we cannot rule out shear faults, which differ strongly from the stick‒slip geometry. Such unexpected shear geometry may be the result of a likely complicated local stress field at the glacier bed. Nonetheless, the first motions do argue against pure shear sources expected from stick‒slip motion accounting for most basal motion, as previously observed in Antarctica [Zoet et al., 2012a]. Instead, our preferred source mechanism is a superimposed tensile crack and shear dislocation.
6.3 Shallow Seismicity and Surface Deformation
 Waveforms of our near‒surface icequakes show only compressive first motions, and they are dominated by the Rayleigh phase, both typical features for events associated with crevasse formation or lateral extension [e.g., Neave and Savage, 1970; Walter et al., 2009; Roux et al., 2010]. Along Triftgletscher's central flow line, crevasse formation occurs near the southern edge of our study site, close to the center of our seismometer circle (Figure 13, red area). As the ice flows northward toward the steep tongue, the crevasses are slightly rotated counter‒clockwise (Figure 13, blue area). This is a consequence of the confluence with a side‒tributary glacier, which adds inertia to the main Triftgletscher branch at its western side. Further down glacier, the surface crevasses are extended eastward and further opened in the eastern half of the study site (Figure 13, cyan area). Further west, closer to the confluence, the crevasse pattern is destroyed (Figure 13, green area).
 Our surface seismicity concentrates at the southern end of the crevasse field. This suggests that surface seismicity is mostly connected with the initial opening and lateral extension, rather than the advection of surface crevasses. The reason is that once crevasses form, they relax near‒surface stresses as they further open, allowing for higher strain rates even in the absence of further fracturing [Vaughan, 1993; Harper and Humphrey, 1998; Roux et al., 2010].
 Strain analysis of the surface deformation from 17 July to 2 August (derived following Lindsay and Stern ) shows that crevasses seismicity does not correlate with strain magnitude. In fact, seismicity is strongest in cell D corresponding to smallest strain (Figure 13). However, the strain axis orientation in cell D is such that the compression axis is most aligned with the main crevasses orientation [Harper and Humphrey, 1998], which is given by the clockwise rotation induced by the confluent regime (as crevasses exit the blue area). This suggests that strain orientation, rather than strain magnitude controls the deepening and lateral extension of crevasses. Furthermore, as the crevasses leave cells A and B, they are destroyed due to misalignment of the initial pattern of the most visible crevasses and the strain axes. Down glacier of cells C and D, on the other hand, this misalignment is less pronounced and the initial pattern of most crevasses is preserved, which allows the crevasses to be advected.
 We located 120 icequakes on Triftgletscher, with focal depths close to the glacier surface and in the bedrock vicinity. Deep sources represent 27.5% of the data set, which makes this area a particularly suitable study site for near‒bedrock emissivity investigation. The use of a 3‒D model allowed good constraints on focal depth for sources located outside of the array. Source locations were corroborated by results of a waveform analysis, i.e., waveform characteristics and cross correlation, which represents a promising method to obtain basic information about focal depth and cluster reliability.
 Both deep event clusters were located in steep terrain, confirming the observations previously reported on Gornergletscher [Walter et al., 2009]. The first motions of most of these events argue against pure shear sources expected in case of stick‒slip motion, and our preferred source mechanism is a superimposed tensile crack and shear dislocation, supported by results of a basal mechanisms analysis. In this context, our work shows that deep seismicity cannot be assumed to be an indicator of stick‒slip motion.
 Surface icequakes primarily relate to opening and lateral extension of crevasses in areas rather than where strain amplitudes are the biggest. The absence of detected icequakes in the downstream part of the study area, yet highly crevassed, reflects the reduction of the seismogenic character of fractures once they are opened.
 As a next step, a more extended seismic array could allow a more precise location of deep sources emitting on Triftgletscher, and a better apprehension of the different information contained in the associated waveform could be provided by the use of broadband sensors. Additionally, a full waveform inversion may help to elucidate basal faulting by providing information on fault planes of crack opening and crack closing events.
 This work was supported by the EU‒FP7 “ACQWA” Project (www.acqwa.ch) under contract 212250. We are grateful to many members of the VAW for help in fieldwork, to the Swiss Seismological Service of ETH and more particularly Franz Weber who brought invaluable assistance to design the seismic array. The Swiss military provided valuable support to us by transporting equipment via helicopter. We warmly thank Norbert Blindow who carried out the helicopter‒based radar profiles, as well as N. Deichmann for having supplied routines allowing focal sphere computation. The technical assistance provided by Anthony Lomax for the use of NLLoc is greatly appreciated.