Seismic constraints on rock damaging related to a failing mountain peak: the Hochvogel, Allgäu

Large rock slope failures play a pivotal role in long‐term landscape evolution and are a major concern in land use planning and hazard aspects. While the failure phase and the time immediately prior to failure are increasingly well studied, the nature of the preparation phase remains enigmatic. This knowledge gap is due, to a large degree, to difficulties associated with instrumenting high mountain terrain and the local nature of classic monitoring methods, which does not allow integral observation of large rock volumes. Here, we analyse data from a small network of up to seven seismic sensors installed during July–October 2018 (with 43 days of data loss) at the summit of the Hochvogel, a 2592 m high Alpine peak. We develop proxy time series indicative of cyclic and progressive changes of the summit. Modal analysis, horizontal‐to‐vertical spectral ratio data and end‐member modelling analysis reveal diurnal cycles of increasing and decreasing coupling stiffness of a 260,000 m3 large, instable rock volume, due to thermal forcing. Relative seismic wave velocity changes also indicate diurnal accumulation and release of stress within the rock mass. At longer time scales, there is a systematic superimposed pattern of stress increased over multiple days and episodic stress release within a few days, expressed in an increased emission of short seismic pulses indicative of rock cracking. Our data provide essential first order information on the development of large‐scale slope instabilities towards catastrophic failure. © 2020 The Authors. Earth Surface Processes and Landforms published by John Wiley & Sons Ltd.


Introduction
Gravitational mass wasting is the dominant geomorphic process shaping mountain peaks. The release of large rock volumes on steep slopes marks the start of a sediment cascade that evolves to the terrestrial sediment flux into the oceans. Its suddenness, velocity and limited predictability renders gravitational mass wasting hazardous yet hard to constrain in terms of drivers and triggers, and mechanisms and their temporal evolution. Classic approaches to studying mass wasting processes are either ex post investigations of the failure mechanism (Frayssines and Hantz, 2006), volume and environmental conditions before and during the event (D'Amato et al., 2016), long-term monitoring, often with remotely sensed imagery, or, increasingly, point measurements from dedicated sensors at sites with known or expected activity (Collins et al., 2018;Di Maio et al., 2010;Dixon et al.,2018;Lévy et al., 2010;Walter et al., 2020). None of these approaches can provide detailed continuous insight into the activity with high temporal resolution, beyond the "point dimension" and throughout the wider process domain, including the downslope pathway of a failing rock mass.
A mass wasting event is preceded by a phase of preparation, during which a set of intrinsic and extrinsic processes drive a rock mass to a state of instability, making it susceptible to trigger mechanisms such as loading with precipitation, seismic ground shaking, wind force fluctuation, and freeze-thaw transitions (see for example (Stock et al., 2013) for a summary of trigger mechanisms). While the preparation phase is crucial for hazard anticipation, its mechanisms are little known. It is clear, however, that rock slope failures predominantly develop in the steepest rock flanks of mountains under high topographic stresses (Leith et al., 2014). There, differential stress conditions can be above the micro crack initiation threshold and favour crack coalescence along the weakest trajectories in the rock mass, i.e. the critical path (Einstein et al., 1983). Once a rock mass starts to move, deformation accumulates along that critical path, and shear stress concentrates on a decreasing number of remaining rock bridges, initiating subcritical and critical fracture propagation in these preferential damage zones (Kemeny, 2003). In the final phase of rock slope failure, when most rock bridges have been destroyed, the rock mass often exhibits cyclic accelerations and decelerations referred to as "stick-slip failure" (e.g., Schöpa et al., 2018;Yamada et al., 2016). This behaviour is thought to be controlled by macro-roughness along the sliding plane that inhibits sliding and promotes stress concentration along the affected obstacles in the sliding plane and release in episodic deformation (Borri-Brunetto et al., 2003;Saettele et al., 2015). In this phase, changes in hydrostatic pressure and seismic excitation can initiate significant acceleration of a sliding rock mass by enhancing shear stress on remaining obstacles. This is due to transiently enhanced overall shear stress, reduced total friction of the surrounding sliding plane, and the effect of water on rock fracture mechanical strength (Voigtländer et al., 2018). As shear planes are mostly inaccessible, insight into spatial and temporal patterns of stress concentration and fracture propagation is difficult to achieve with classic approaches, especially under natural conditions.
In the last decade, networks of seismic sensors have been established as a complementary tool to overcome some of the limitations of other measurement and monitoring approaches. Seismic sensors can record ground motion caused by processes acting at or close to the Earth's surface. In general, seismic networks can i) detect mass wasting activity, ii) locate and track the process in space, and iii) infer kinetic and anatomic details of a process event. The application fields of environmental seismology (Burtin et al., 2014;Larose et al., 2015) comprise, for example, time-resolved investigation of the evolution of slope instabilities (e.g., Mainsant et al., 2012;Schöpa et al., 2018), detection and quantification of event activity from the catchment to the global scale (e.g., Dammeier et al., 2011;Ekström and Stark, 2013;Fuchs et al., 2018;Lacroix and Helmstetter, 2011), tracking of mass movements in space (e.g., Burtin et al., 2016;Cook et al., 2018;Pérez-Guillén et al., 2019;Walter et al., 2020), inversion of seismic signals for event kinetics (e.g., Allstadt, 2013;Ekström and Stark, 2013), and attribution of events to drivers and triggers (e.g., Burtin et al., 2013;Helmstetter and Garambois, 2010).
In addition to investigating the seismic signals emitted by mass wasting processes, it is possible to use random background signals (ambient noise) to survey the material properties of landforms (Welch and McLamore, 1973). There are several techniques that are sensitive to state changes of the near-surface materials in eroding landscapes. A classic technique is the horizontal to vertical spectral ratio (HVSR) analysis (Nakamura, 1989;Nogoshi and Igarashi, 1971). It has been widely used to estimate the near-surface structure of sites (for a review see Del Gaudio et al., 2014), though there is controversy about the confidence and robustness of quantitative inversions for material properties (see for example discussion in Overduin et al., 2015). Bottelin et al., (2013) used seismic sensors to monitor changes in the vibration frequencies of an unstable rock column in the French Alps as a function of temperature and precipitation (i.e., modal analysis).  applied the same technique to the Matterhorn, identifying cyclic as well as cumulative damage evolution. Seismic noise cross correlation (Sens-Schönfelder and Wegler, 2006;Snieder, 2004) has the potential to detect subtle, relative changes in the seismic wave velocity, termed dv/v throughout this text. The technique can be applied using the individual components of a single seismic station or records from seismic station pairs, given that both stations are within the coherence range of the ambient wave field. The seismic wave velocity is affected by a series of factors, such as rock temperature, ground moisture, accumulated rock damage and mechanical stress (Clements and Denolle, 2018;Larose et al., 2015;Snieder, 2004). HVSR, vibration frequency and dv/v are valuable proxies of material state, because they provide information on the average state of the medium within the seismic network aperture, not just at a point. Moreover, these proxies can be recorded continuously, at arbitrary high resolution and with little infrastructural effort, so that the evolution of the material state can be tracked over time and changing boundary conditions.
Here we study the summit of the Hochvogel, a 2592m high solitary peak in the South German Alps. We exploit the utility of a small, telemetered seismic network on the summit, which was operational during a period in which a large rock volume near the summit was undergoing failure preparation. We investigate seismically evident patterns of summit deformation, how these are controlled by environmental conditions, and how they translate into mechanistic activity.

Study Site and Instrumentation
The Hochvogel is a solitary, 2592m high peak in the Eastern Allgäu Alps, on the border between Germany and Austria. It is formed of folded and thrusted upper Triassic dolomite (Hauptdolomit), which forms competent decimetre to metre thick beds, 20°dipping to the NNW in the summit area. Extensive vertical fracture sets run perpendicular to the general dip, giving rise to an intensely fractured fabric. Formation of large fractures in the summit area was first recorded in the late nineteenth century, and a progressively accelerating rock slide has developed since the 1950s. Catastrophic failure could mobilise up to 260,000 m 3 of rock in at least six subunits, ranging from 8000 to 148000 m 3 (Krautblatter et al., 2019) along a failure plane located on two large, NE-SW trending fracture systems. In 2018, the southeastern, main fracture had already opened by 2-4 m, had a dimension of about 5m width and 30m length, and continued to open at a rate of a few mm per month (Krautblatter et al., 2019). The unstable rock mass has slid downslope by several metres and developed a severe state of fracturing ( Figure SI 1-3). The main fracture can be traced more than 60m downwards, yielding a minimum sliding plane surface area of 2000 m 2 . Crack meter data from a 2020 installation (Leinauer et al., 2020) display potentially precipitation driven cycles with a duration of about five days, that start with sliding (i.e., opening of the main crack) and end with a slight backward rotation of the displaced mass. Sliding on a premature plane with remaining rock bridges is inferred to cause rhythmic major rock fracturing events, determined by deformations stress buildup and release.
To monitor activity at the Hochvogel summit as well as any mass wasting processes on the southeastern mountain slope, we installed a nested network of seismic stations (Figure 1). At the summit we installed a small aperture network (33m average, 73m maximum spacing) consisting of six PE6B one-component 4.5Hz geophones, three on the less fractured, relatively stable German side (SA 11 ÀA 13 ) and three on the failing Austrian side (SA 14 ÀA 16 ). The sensors were connected by 100m long cables to a Nanometrics Centaur data logger that recorded the ground velocity values at 200Hz with a gain of 40. These seismic data were transmitted to the GFZ data server every five minutes. The network was operational from 10 July 2018. However, after a lightning strike of the station on 1 August 2018 no data were recorded for 43 days until 12 September 2018. From this date we operated a set of two PE6B 4.5Hz geophones (SA 21 a 1-component and SA 22 a 3-component sensor) and Digos DataCube 3 ext loggers, installed in an aluminium case with a 2cm thick concrete bed. To record slope activity below the summit, we installed five seismic stations with spacings between 0.9 and 1.6 km. Station HV 1 contains the same setup as SA 22 and was used in this study as additional summit station. Here we only consider data from the small aperture array at the summit, 418 M. DIETZE ET AL.
for the periods 10 July-1 August and 12 September-12 October.

Data Processing
All seismic data were processed with the software R v. 3.6.3 (R Development Core Team, 2020), using the package eseis v. 0.5.0 (Dietze, 2018a(Dietze, , 2018b. The Supporting Information (SI) contains dedicated R scripts of all major processing and analysis steps. The underlying data is available via GFZ Data Services (Dietze and Krautblatter, 2020). All seismic data were deconvolved to correct for the instrument response, yielding ground velocity (m/s) time series. For modal analysis we calculated averaged Welch (1967) spectra for non-event periods (Bottelin et al., 2013) of the summit network sensors. For this we calculated an STA-LTA ratio (Allen, 1982) and flagged periods with a ratio > 1.5 as eventcontaminated. These periods were excluded from the subsequent step of spectra calculation. The spectra were calculated for 5 s windows without overlap, which were then stacked to non-overlapping hourly spectra. Each of these spectra was subsequently normalised. Finally, we executed a 5 sample running average filter in frequency and time. The PE6B geophone sensors have a natural frequency of 4.5 Hz. However, they are also sensitive below their corner frequency, including frequencies below 2Hz (Overduin et al., 2015). Still, the fundamental frequency of the studied rock mass might be below this threshold (e.g. Bottelin et al., 2013;Lévy et al., 2010). Therefore, we restrict our analysis to the lowest stable frequency band visible in our data.
For the HVSR analysis we calculated spectra of station HV 1 and SA 22 for the vertical and horizontal components within 5 s windows with 50 % overlap. We then calculated the ratio of the corresponding horizontal and vertical spectra and stacked these to hourly results. To quantify the spectral modes inherent to the HVSR data, we applied end-member modelling analysis (EMMA) to the data set. EMMA (Dietze and Dietze, 2019) is a statistical technique, which allows unmixing multimodal data, and describing it as a linear combination of loadings (the fundamental underlying spectra distributions) and scores (the relative contribution of each loading to each sample). EMMA was originally developed for decomposition of high resolution grain-size spectra, but it can be extended to other kinds of compositional data (i.e., data that is non-negative and where each sample is described by a constant sum). To be able to apply EMMA, we truncated the HVSR data set to the most variable frequency range and normalised the spectral ratios of each hourly time slice between 0 and 1. Since we were interested in identifying the position and shape of the spectral modes as well as their relative contributions to a temporal sample, but not in the absolute intensity of the HVSR amplitudes, this normalisation step is valid. We followed the generic deterministic EMMA protocol (using the EMMAgeo package v. 0.9.8, Dietze and Dietze, 2019) and tested between two and six end-members for plausibility of the results. These plausibility checks were based on model quality (i.e. time-resolved and frequency-resolved variance explained by the model), the degree of unmixing of the frequency distributions and the interpretative meaningfulness of the unmixed frequency distributions and their evolution with time.
We estimated relative velocity changes from ambient noise cross correlation. First, the seismic traces were pre-processed to reduce the impact of erratic, high amplitude signals. We trimmed the seismic records to 10 min long segments, downsampled to 50 Hz, detrended and filtered in the 4-16 Hz frequency range. We normalised amplitudes by spectral whitening, setting all the amplitudes in the Fourier spectrum to 1 and performed 1-bit normalisation. In the first time period, we performed cross correlations between all vertical component sensor pairs. In the second period, since only two stations were operating, located on different sides of the main fracture, we performed single station cross correlation (Hobiger et al., 2014) on station SA 22 using the three combinations north-vertical, east-vertical, and east-north. We stacked the obtained correlation functions to 30 min non-overlapping averages and stored them in a correlation matrix C(t i ,τ) where t i and τ is the lapse time in the correlation process. To infer relative velocity changes dv/d, we used the stretching technique (Sens-Schönfelder and Wegler, 2006): Each of the 30 min correlation functions were be stretched or compressed depending on their relative velocity change compared to a reference (τ). For the two time periods, we computed the average correlation function (τ) and used it as a reference. We calculated the correlation coefficient R(t i ,ε j ) between stretched versions of this reference with each correlation functions stored in C(t i ,τ) such as where ε j indicates stretch-values tested in the time window [τ 1 τ 2 ]. We used τ 1 ¼ 0:5 s to avoid unwanted source effects and used a window duration of 3 s. For the second time period, we averaged the measurements between the three combinations by stacking their three corresponding matrices R(t i ,ε j ). Finally, we extract those dv(t i )/v values for which ε j yielded the maximum correlation coefficient R(t i ,ε j ), and only kept values where the correlation coefficient was at least 0.5. In addition to the continuous analysis, we isolated discrete seismic events in the data of the summit network. We picked events using a classic STA-LTA ratio routine (Allen, 1982), applied to the 10-40Hz filtered signal envelopes. The routine was applied with a low detection threshold (on-ratio = 3, off-ratio = 1, STA-window = 0.5 s, LTA-window = 120 s). This yielded many false positives but also guaranteed detection of small events. All potential events were tested against automatic rejection criteria. We required an event to have a minimum duration of 0.2 s as well as a maximum duration of 10 s. We assumed that events shorter than 0.2 s are associated with random sources such as rain drop impacts , and events longer than 10 s are unlikely to be linked to discrete rock fracturing activity at the summit (e.g., Senfaute et al., 2009). A further rejection criterion imposed that an event had to be detected by at least two stations with a time difference of not more than 0.1 s. This rule enforces that signals must travel across the entire network in a time that corresponds to an apparent seismic wave velocity of at least 730 m/s, a threshold well below typical values for limestone of different origin and degree of fracturing (Assefa et al., 2003;Helmstetter and Garambois, 2010). Signals with longer detection time differences are usually a result of the coincidence of unrelated signals or waves travelling through air (about 330 m/s velocity) before coupling to the ground. All remaining events were then checked manually for plausibility: telltales of genuine geomorphic signals include spindle shaped seismograms, and consistent amplitude and frequency decrease as well as onset shifts with distance between source and sensor (for details see Hammer et al., 2015;Hibert et al., 2011;Provost et al., 2018;Vouillamoz et al.,2018).
Meteorological data for the summit were approximated by scaling hourly values of a station near Oberstdorf, 12 km to the west (DWD, 2019). Temperature scaling by -0.6°C per 100m altitude increase results in a downward shift of 12°C of summit temperatures. For precipitation the German Weather Bureau offers hourly gridded data (DWD, 2019). However, for the time of interest there are data gaps. Therefore, we calculated a transfer function between gridded values at the peak and the Oberstdorf data for adjacent periods with data and use this relationship to infer precipitation data also for the peak during relevant data gaps.

Meteorological conditions
Meteorological dynamics define an important set of boundary conditions for the activity of the summit. The scaled air temperature records ( Figure 2) are dominated by the diurnal cycle, with amplitudes of about 18°C in the first instrumentation period (10 July-01 August) and about 22°C in the second period (10 September-12 October). Superimposed on these fluctuations was a rising trend from 10 to 23°C (interrupted by stable temperatures between 16-24 July) in the first period and a somewhat decreasing trend from 14-10°C during the second period. Precipitation events occurred throughout the record time (Figure 2 e), but lasted only a few hours to a day, with usually less than 6 mm/h peak intensities. Precipitation events did not coincide with changes in any of the seismic data sets.
Continuous seismic observations at the summit Continuous seismic observations included the results of the modal, HVSR and dv/v analyses, which will be used to infer material property changes of the rock mass. Here, we offer a condensed, synoptic report of the multitude of observed metrics. The SI contain results from all sensors. In general, sensors from the instable mass on the southeast side of the Hochvogel summit (SA 14 , SA 15 , SA 16 , SA 22 ) did not record a systematic pattern of vibration frequency evolution on the vertical component. In contrast, stations on the stable mass, opposite (SA 11 SA 12 ), did reveal consistent vertical component vibration frequency bands. These were best expressed at station SA 11 , located closest to the summit and the main fracture (see Figure SI 5 and 2 a). We used the two available three-component stations HV 1 (first period) and SA 22 (second period) for HVSR analysis. Both stations were located at the same spot, on the instable mass adjacent to the main crack. The dv/v analysis yielded between-station correlation coefficients > 0.5 for the station pair SA 12 and SA 13 , farthest away from the severely fractured mountain top, while all other pairs did not show meaningful signal coherence. Likewise, the between-components analysis for SA 22 during the second period did not yield correlation coefficients above 0.5 and results were thus discarded.
The two monitoring periods (10 July to 01 August and 12 September to 12 October, separated by 43 days without data) showed contrasting results of the analysed monitoring metrics. The lowest continuously visible vibration frequency band of the vertical component of stations installed near the top of the stable northern rock mass (SA 21 and SA 22 , Figure 2 a) ranged between 24 and 28 Hz. There were also higher frequencies present in the data set (see Figure SI 5) but they were either multiples of the lower frequency band from Figure 2 a or frequencies corresponding to higher order bending modes (Bottelin et al., 2013). In the first period, the 24-28 Hz band exhibited prominent multi-day cycles, consisting of four to seven days of continuous rise of the vibration frequency by two to three Hz followed by a sudden drop, almost back to the initial values. Four such cycles are visible in Figure 2 a, with drops around 10, 15, 21, 29 July. In the second period, the 24-28Hz vibration frequency band was dominated by diurnal cycles without clear multi-day effects. On the instable southern rock mass, no persistent vibration frequency pattern was found in the vertical component (see Figure SI 5). 420 M. DIETZE ET AL.
The horizontal seismic sensor component (Figure 2 b) exhibited a persistent vibration frequency band between 7 and 10 Hz, though weakly expressed, with diurnal cycles throughout both periods. Multi-day patterns were suppressed (first period) or absent (second period). A data gap during the second period was due to event contaminated time slices, which were removed, thus affecting the moving window temporal averaging process.
Diurnal cycles were also visible in the HVSR results (Figure 2 c), most clearly during the first survey period (10 July to 01 August). The higher resolution of this record in the time dimension allowed a more detailed analysis of the data. This revealed that the HVSR was around 5-7Hz during night time and around 9-10Hz during daytime. This diurnal pattern was distorted during episodes when the vertical vibration frequency drops occurred (Figure 2 a).

End-member modelling
To better characterise and quantify the underlying frequency patterns and their contribution to the data through time, we performed end-member modelling analysis on the 2-15Hz band of the normalised HVSR data (Figure 3 a). Since the number of end-members is not defined a priori, we tested solutions with 2-6 end-members. Models with two (62.2 % of total variance explained for first period, 55.9 % for second period), three (75.6 %, 70.0 %), five (77.9 %, 70.8 %) and six (80.0 %, 75.3 %) end-members showed mostly lower performance than the model with four end-members (80.6 %, 71.9 %). The tested explained variance is a measure of the ability of the model to describe the input data set. Usually, models with too few or too many end-members show poor explanatory power. More importantly, models with other than four end-members did not yield meaningful unmixing results. Specifically, these models exhibited i) multimodal end-member loadings with modes at frequencies not visible in the input data, ii) broad frequency distributions, or largely overlapping and thus not unmixed end-members (see SI for all tested models). The two four-end-member models for the survey periods 10 July-01 August and 12 September-12 October allowed describing the data set in the time domain (i.e., the hourly time windows of HVSR data) with R 2 ¼ 0:89 (first period) and R 2 ¼ 0:79 (second period). Thus, the evolution of the frequency end-members with time could be modelled in a way that 89 and 79 % of the data's variance can be explained, respectively. The frequency domain (i.e., the model representation of individual HVSR distributions, cf. Figure 3 c) is described with R 2 ¼ 0:73 (first period) and R 2 ¼ 0:65 (second period). This lower R 2 value in the second period was caused by a general lack of frequencies higher than 10Hz in the input data set ( fig. 3 c). The shapes of the end-member loadings (Figure 3 d) agree with the visual inspection of the raw HVSR data (Figure 2 c). During the first survey period, two recurring frequency modes at 5.4 and 9.0Hz are represented by EM 1 and EM 3 , while the transitional stages between these two are formed by EM 2 (modal frequency at 7.0 Hz). EM 4 covers a broader frequency band, ranging from 10-15 Hz. During the second survey period, EM 1 to EM 3 had consistently shifted their modal frequencies, to 6.4 Hz, 7.4Hz and 9.8 Hz, respectively. The EM 4 from the first period was no longer visible, but EM 3 showed a broad shoulder towards higher frequencies. During this second period, EM 4 was in the frequency range 2-5Hz with a modal frequency of 3.2 Hz. This new low frequency variability in the data is also visible in the normalised (Figure 3 a) as well as the raw (Figure 2 c) HVSR data set. In terms of end-member scores (Figure 3 b), EM 1 provided high contributions to the HVSR data mostly during the early hours of the day, and EM 3 during the late hours. EM 2 takes an intermediate role, between EM 1 and EM 3 . In the first survey period, EM 4 mainly contributed to HVSR data during episodes of enhanced seismic event rates (see below). Noise cross correlation analysis yielded meaningful results only for the station pair SA 12 ÀÀA 13 (Figure 2 d). Even there, the correlation among the stations did not reach > 0.8 before 16 July. From that date on, the 1-day running averaged dv/v data shows a diurnal pattern superimposed on a longer-term trend. That longer term trend is better visible in the 3-day running average curve: Coincident with the periods of vertical vibraion frequency (Figure 2 a)

Discrete seismic events
The summit network recorded a series of discrete signals of diverse properties with respect to duration, evolution and spectral content. The STA/LTA event detection algorithm and application of the subsequent rejection criteria yielded a total of 2490 and 845 potential near-surface events for the first and second survey period, respectively. One class of detected events had the hallmarks of rock cracking (e.g., Senfaute et al., 2009): one to two seconds long, erratic, single pulses suddenly emerging from an otherwise silent record, with frequencies predominantly between 30 and 80 Hz. These events showed a distinct evolution during the two survey periods. Between 10 July and 01 August, the average event rate was 100 cracks per day, with episodes of enhanced activity on 10-12, 15-16, 21-22, 24-26 and 27-28 July (Figure 2 f). In the second period, 12 September to 12 October, the average crack rate was lower, 18 per day. Episodes with enhanced crack activity occurred also during this period (e.g., 18, 20, 24-28, 03-07 October), but significantly less pronounced compared to the earlier interval. We note that the episodes of enhanced activity did not systematically coincide with precipitation events or weekend days when mountaineering traffic was likely more frequent. Individual events had an average duration of 0:58 þ0:44 À0:02 s (median and quartile range) during the first and 1:57 þ0:95 À0:48 s during the second period, both following a log-normal distribution (Figure 5 a). The signals were clustered at the diurnal scale, occurring preferentially at day time (Figure 5 b). That trend is even more pronounced during the second period.
Among the cases rejected because of their duration were events as shown in Figure 4. These signals exhibited short, reg-FIGURE 4. Seismic signals due to different sources, as recorded by the Hochvogel summit network (station SA 22 ). The plots show spectrograms (scaled between -160 and -100 10log 10 (m 2 /s 2 )/Hz, except for c), which is scaled in the range -180 to -120) of the signals and also seismograms as white line overlays. a) Sequence of steps generated by a person walking towards and past the station along the summit. b) Signature of a small mass wasting event at the summit. c) Seismic signals during a rain event (note order of magnitude lower amplitudes). d) Seismic signature of a stress release event as described in Figure  ularly spaced pulses (about 1.3 pulses per second) of systematically increasing and and decreasing amplitude (Figure 4 a), or signals without distinct pulses but rather a spindle shaped energy envelope, increasing in amplitude for 4 s and decreasing for 7 seconds (Figure 4 b). Likewise, during times of precipitation (Figure 4 c), the sensors recorded up to 0.2 s long single pulses with mainly 60-80Hz frequency content and decreasing recurrence intervals but higher amplitudes as the rain intensity increased (not shown).

Discussion
Individual signals of summit activity According to our seismic data, cyclic activity in the Hochvogel summit occurs on diurnal and multi-day scales (Figure 2). The diurnal cyclicity is expressed in systematic 1-2Hz changes of the vertical and horizontal component's vibration frequency, in the HVSR bimodal frequency distribution, and in dv/v fluctuations of about 0.2 %. These patterns are clearest in the more intact, northwest facing rock mass, and less so in the heavily fragmented rock mass facing towards the southeast. Multi-day cyclicity was prominent during the first survey period (10 July-01 August), expressed in several seismic observables. The vertical component vibration frequency increased gradually over four to six day periods, alternating with two days of decreasing trends. Similarly, a pattern of 4-6 days stability with 2 day long interruptions of the bimodal frequency distribution pattern was found in the HVSR data. During these interruptions, dv/v dropped by about 0.2 %, and the rate of short duration seismic events in the summit area increased by a factor of 0.5 to 7. The significance of these coincidences is hard to estimate. However, tentatively we infer that these temporally connected signals are being driven by the same underlying mechanisms. To probe the nature of these drivers, we consider the individual seismic proxies (Figure 2) in more detail.

Modal analysis
The frequency with which an object vibrates is at first order defined by its geometry and material properties. While the geometry of the Hochvogel summit can be considered relatively constant, at least during the time period covered in this study, the material properties are subject to changes. These can include thermal changes or variations in water content due to the local weather, and also internal mechanical changes due to rock mass fracturing, or the failure of rock bridges (Lévy et al., 2010;Moore et al., 2010). Similar to the range of values found in the Hochvogel summit, Lévy et al. (2010) reported a resonance frequency drop of almost 1 Hz, paired with an increased rate of discrete seismic emissions in a limestone cliff in the Western Alps, just prior to a 21,000 m 3 large failure event. Bottelin et al. (2013) measured and modelled the resonance frequency of a 1,000 m 3 limestone slab in the Western Alps, finding systematic, thermally driven changes in the vibration frequency. At the diurnal scale, increasing temperatures can cause an increasing vibration frequency, with a time lag of a few hours due to heat dissipation into the rock. Bottelin et al., (2013) attribute this to stiffening of the interface between the column and the rock mass due to thermal expansion, likely involving temporary closure of rock fractures. In contrast, at the annual scale, rising temperatures result in a lower vibration frequency (with a time lag of three months) due to a decreasing Young's modulus. Following this explanation of empirical and numerical modelling results, we interpret the diurnal cyclicity of our vibration frequency data as representing the daily opening and closing of fractures in the rock mass, resulting in increased stiffness during warming and decreased stiffness during cooling.
In contrast, we did not observe a direct or lagged relation between vibration frequency and air temperature on the multi-day scale. This may be due, in part, to the fact that the long-term temperature range is smaller than the diurnal range. Note also that our periods of observation were significantly shorter than the time lag of three months reported by Bottelin et al. (2013). Lévy et al. (2010) interpreted vibration frequency drops as a global parameter characterizing the mass and its coupling to the stable rock mass. In the absence of any other likely cause we propose that slow increases in frequency were caused by cumulative increases of stress within the rock mass, which were then released episodically during a few days when frequency drops back to the values before the phase of stress increase. Similar drops of the vibration frequency have also been reported to correlate with increased rock mass mobility values (e.g., Burjánek et al., 2018) and changes in the elastic moduli have been reported under stress in laboratory experiments (Barton, 2007;Draebing and Krautblatter, 2012).
The different vibration frequency windows of the vertical component (24-28 Hz, Figure 2 a) and the horizontal component (8-10 Hz, Figure 2 b) may represent different modes of movement, possibly stamping in the vertical direction and either bending or rotating in the horizontal component. These mode dependent frequency differences were also described and modelled by Bottelin et al. (2013). Whether there might be even lower frequencies, beyond the sensitivity limit of the geophones remains unresolved. In view of the complex structural situation, a yet unclear size and shape of the developing failure plane, and under-constrained geometric boundary conditions, numerical modelling of the rock mass properties appears not meaningful in the Hochvogel case and we desist from the deduction of more detailed insights about the quantitative physical meaning of the vibration frequency values.
Horizontal-to-vertical spectral ratio HVSR data are thought to represent site resonance, solicited predominantly by S-waves and to a lesser degree by Rayleigh waves (Del Gaudio et al., 2014). Peaks in the ratio reflect the frequency at which the power of the horizontal component of a seismic signal significantly exceeds that of the vertical component, thus indicating a site specific resonance frequency. The method has been used in steep rock slope landscapes to monitor changes in system and material states, signaled by changes in the resonance frequency (e.g. Burjánek et al., 2012;. These studies predominantly focused on seasonal to multi-year time spans to explore reversible and irreversible effects contributing to failure preparation. Like  we have found multiple discrete vibration frequency bands in the HVSR data (see SI). Here, we focus on the lowermost visible one, around 8-10Hz (Figure 2 c), which was also flagged in the vibration frequency analysis (Figure 2 b), indicating the spectral amplification effect due to the contribution of the horizontal component.
Our measurement intervals were not long enough to compare our data against reported findings on multi-seasonal time scales. However, we can explore the diurnal and the multi-day cyclicity patterns in the data. Our end-member modelling analysis has helped to identify and quantify the shape of the underlying frequency distributions and their contributions to the resulting normalised HVSR time series (Figure 3). We interpret EM 1 , which predominantly contributes to the data during the night time (see downwards pointing arrows in Figure 3 b) as indicative of a system with a lower resonance. This is supported by the evolution of the vibration frequency of the horizontal component (Figure 2 b)  and spectral resolution. The lower resonance frequency may be caused, as described above, by thermal shrinkage which results in opening of small cracks and an overall decrease in the stiffness of the summit rocks. Vice versa, EM 3 is the high resonance mode of the system, predominantly active during day time (green line and upwards pointing arrows in Figure 3 b). EM 2 can be regarded as the intermediate state between the mutually exclusive EM 1 and EM 3 . We interpret EM 4 as noise or residual end-member, accounting for the variability within the data set that is not covered by the three other end-members (for details see Dietze and Dietze, 2019). During the second survey period, later in the year, EM 4 had a different frequency range, around 2-5 Hz, also visible in the raw HVSR data (Figure 2 c). This frequency band had no special weight during the first survey period. Note that the models for the two instrumentation periods were not calculated with any shared parameters. Thus, the resulting sets of end-members are independent of each other. This is obvious for the noise end-members EM 4 . In contrast, the similar though throughout higher frequency end-members 1-3 point at a persistent driving force. Overall, the model quality ( R 2 ¼ 0:89 for the first and R 2 ¼ 0:79 for the second survey period) is comparable with results from other studies (cf. Dietze and Dietze, 2019), implying that 89 % and 79 % of the system's variance can be explained by just four end-members. Between the two monitoring intervals, end-members EM 1 to EM 3 moved to higher frequencies. This is likely due to the seasonally increasing temperature effect: overall higher temperatures within the rock result in thermal expansion and a stiffening of the rock mass. Likewise, the night time EM 1 contributed systematically less to the data set in the second survey period (decrease from 42 % to 19 %, see Figure 3 d), while the day time EM 3 gained importance (from 19 % to 27 %). We attribute this to the decreasing time for the system to swing and remain in the low frequency mode during night time later in the year, due to the overall thermal input and heat legacy, and despite the gradually shorter day lengths. In the first survey period EM 4 contributions were slightly stronger during the episodes of signal disturbance (yellow bars in Figure 2). However, since we clipped the HVSR data at 15Hz to isolate the part of the frequency spectrum that is dominated by daily variations, interpretation of this trend may be not meaningful. During the second monitoring interval, this effect was not found. In summary, we consider the HVSR data (Figure 2 c) and the estimated end-member loadings (Figure 3 d) and scores (Figure 3 b) to be a high resolution proxy for reversible material stiffness, as constrained by thermodynamics, predominantly at the diurnal scale, but also with an underlying seasonal trend.
Seismic wave velocity evolution Relative changes of the seismic wave velocity are the summed effect of many contributing factors, and as for the vibration frequency, these can be exogenic or endogenic in nature. Changes in seismic wave velocity can derive from variations of water saturation or freezing in pore space and fractures, ambient stress, for example exerted by gravity-driven deformation or ice pressure, and reflect changes in fracture extension and density (Barton, 2007;Draebing and Krautblatter, 2012;Krautblatter and Draebing, 2013). Temperature-driven changes are evident at the diurnal scale, with relatively low dv/v values at the start and end of the day and relative maxima in the middle of the day (Figure 2 d). In contrast, the longer-term trends in temperature are not reflected in the dv/v data set. Transitions from positive to negative temperatures, which would cause freezing of water and thus a drastic increase of the seismic wave velocity (James et al., 2019), did not affect the dv/v time series. Elsewhere, the effect of rainwater infiltration is a lowering of the relative seismic wave velocity, both for deep groundwater systems (Clements and Denolle, 2018) and shallow soil water content in the vadose zone . However, rain (Figure 2 e) had no easily discernable effects on the dv/v data from the Hochvogel summit network. We remind that we have no direct rain measurements from the Hochvogel summit but used the rain gauge data from a station 12 km to the west. Under the rapidly changing and filigree weather conditions in the Alps, the representativeness of this rain data set may be limited. However, in the dv/v data we see very few erratic changes that could be attributed to unrecorded rain events.
More importantly, we found cyclic patterns of slowly rising and suddenly dropping dv/v values, with an amplitude of about 0.2 % (Figure 2 d, bold black line), coincident with the periods of disturbance of the other proxy data. In the absence of meteorological forcing, we attribute these cycles to reversible, rock-internal material state changes. Increased stress has been reported to yield higher dv/v values (Sens-Schönfelder and Larose, 2010), and significant reductions in dv/v values have been recorded immediately prior to the initiation of mass wasting processes (Mainsant et al., 2012). Following these findings, we interpret meteorologically unrelated and reversible dv/v cycles as proxy for stress accumulation and dissipation inside a rock mass.
Emission of discrete seismic signals The different types of short seismic signal pulses recorded by the Hochvogel summit network (Figure 4) have distinct properties, which make them straightforward to attribute to likely sources. Repeated, regularly spaced pulses with gradually increasing and decreasing amplitude (Figure 4 a) are typical for people approaching and passing a seismic station (e.g., Vouillamoz et al., 2018). Several such sequences were recorded during installation and maintenance visits of the stations, and others registered only during day time when hikers frequent the summit. Spindle shaped continuous signals with rapidly increasing and less rapidly decreasing amplitudes lasting several seconds (Figure 4 b) are indicative of small mass wasting events down the summit flanks. Similar signals have been described in detail by numerous other studies (e.g. Burtin et al., 2016;Helmstetter and Garambois, 2010;Hibert et al., 2011;Lacroix and Helmstetter, 2011) and their frequency spectrum is similar to artificial rock drop experiments . High-frequency signals unique to independently registered rain events (Figure 4 c) are evident at all sensors but with random, incoherent drop impact signal timings at stations of the network. The pauses between the 0.05-0.2 s long pulses decrease after the onset of a rain event and become less evenly spaced. The frequency content of the signals is also higher (60-80 Hz) than for other recorded signals. We interpret such low amplitude signals as rain drop impacts at or very close to the seismic sensors. Finally, sudden pulses of 30-80Hz signals without preceding or subsequent activity (Figure 4 d) are most likely the result of sudden releases of accumulated stress by failing rock bridges or extension of cracks in the rock mass. Such signals have been reported under natural (Vouillamoz et al., 2018; and laboratory (Murton et al.,2016) conditions. We use this type of seismic signal as proxy for stress release, e.g. due to rock cracking.
Stress release signals had durations of 0:58 þ0:44 À0:02 s (median and quartile range) during the first, and 1:57 þ0:95 À0:48 s during the second period (Figure 5 a). This increased duration of pulses may be due to the greater average travel distance of a signal to a sensor: only 2 stations instead of up to six during the first 425 SEISMIC CONSTRAINTS ON ROCK DAMAGING RELATED TO A FAILING MOUNTAIN PEAK: THE HOCHVOGEL, ALLGÄU period, which results in extension due to scattering as waves travel through the rock medium, especially given that during the second period the two stations were separated by the main crack through the summit. More interestingly, stress release events appear to be linked to daylight conditions or, more specifically, potential insolation duration. During July-August, topography corrected sunrise was between 05:30 and 05:45 local time and sunset between 21:05 and 21:20, yielding a duration of potential insolation of between 15:20 and 15:45 hours. In contrast, during September-October sunrise was between 06:45 and 07:30 and sunset between 18:45 and 19:45, separated by 11:15 to 13:00 hours of potential sunlight (see horizontal bars in Figure 5 b). Reflecting these insolation times, stress release events occurred preferentially between 05:00 and 22:00 local time during the first recording period, with a multimodal activity distribution. During the second survey period such events clustered between 7:00 and 21:00, with a pronounced maximum between 09:00 and 12:00. These patterns are in contrast with results from other alpine locations, such as the Matterhorn (Coviello et al., 2015), Hörnligrat (Hasler et al., 2012) and Jungfraujoch (Girard et al., 2013). These latter sites show a higher crack signal density during cold periods, when air temperature drops below 0°C, hence during the colder hours of the day, when freezing drives cryogenic processes in water/ice-filled clefts. During the instrumented period, the Hochvogel did not experience negative temperatures, and it remains to be seen whether or not cryogenic processes change the activity pattern during the winter season.
At the multi-day scale, stress release events were not randomly distributed. Instead, we observed discrete bursts of a 150-600 % activity increase above a relatively stable base line of 100 cracks per day (median) for the first recording period, and 18 cracks per day during the second period. These clusters coincided with the episodes of disturbance in the continuous proxy records (Figure 2, yellow bars). During these episodes, the vibration frequency of stations dropped, the bimodal frequency pattern of HVSR data was interrupted, and the local seismic velocity decreased.
An exception to the collocation of continuous seismic proxy adjustment and bursts of cracking is the prominent spike in stress release activity on 25 July. This could signal that there is a hierarchy of stress build up and release cycles within the Hochvogel summit, which has not been covered in its entirety by our short surveys. Alternatively, stress release in the summit is a process with cyclic and stochastic elements.
During the first survey period there were up to seven sensors in operation, with some brief, incidental intermissions of data transmission. In contrast, during the second survey period only two stations were operating, continuously and without any data gap. The average number of stations that detected the same seismic event was 3.5 during the first and 2 during the second survey period. Hence, there may be a network bias inherent to the time series. However, even if the number of true events were underestimated during the second period, it is unlikely that the general pattern of a significantly reduced event rate and overall number during that period is an artifact. Therefore, we conclude that there was a stark contrast in stress release activity between the first and the second survey period, potentially indicating a change in the sliding regime with reducing obstacles, i.e., rock bridges. A robust interpretation of this difference would only be possible in the context of a more extensive data set.

Synthesis
With the interpretation of individual proxy data in hand, we can now investigate underlying system-wide dynamics. On the diurnal time scale, most likely due to thermal forcing, the Hochvogel summit experiences a cyclic increase and decrease of stress, but also strengthening and weakening of the coupling between the strongly fragmented southeastern and the less severely fragmented northwestern parts. Stress evolution is reflected throughout both survey periods and by multiple proxy data: i) the dv/v data, which records the effects of increased stress as higher dv/v values, and ii) discrete stress release signals predominantly during daylight conditions. The coupling evolution is evidenced by i) the vertical vibration frequency of the northwestern summit part, ii) the horizontal vibration frequency of the southeastern instable summit part, and iii) the HVSR end-member time series. The predominant effect of thermal forcing may be interpreted based on i) the clear diurnal air temperature signal in line with the patterns of all seismic proxy data, ii) the agreement of day lengths and stress release signals at the seasonal scale, and finally iii) the absence of evidence for any other plausible forcing mechanism.
The short sub-diurnal reaction time of the system implies that both stress release and coupling adjustment must be implemented at or very close to the surface, and are thus unlikely to have a significant effect on the wholesale slope instability, especially at many metres depth. Heat as driver for this system is in agreement with the effects that we have measured seismically. That heat can be provided either by direct insolation or as sensible heat of the air circulating around the summit and into the cracks. It would be dissipated by conduction, either directly within the rock or through percolating water, which is able communicate thermal effects deeper into the rock, and thereby attenuate the expressions of heat input close to the surface. Regardless of the mechanism, heat diffusion into the rock would result in time lags of several hours (e.g., 37 h for 50cm and 60 h for 1 m, Mulas et al., 2020) and rapid dampening (e.g., Collins et al., 2018) of the thermal input signal. Indeed, the end-member evolution with day time (especially for EM 1 and EM 3 , Figure 3 b) shows that the low frequency EM 1 usually decreases gradually with the course of the day, whereas the high frequency EM 3 correspondingly increases throughout day time. The coupling adjustment shows no signs of progressive evolution. Continuously less cracks capable of closing during day time and thus a systematic decrease in the vibration frequency is not evident from our data during the study interval. We rather see the opposite effect: a systematically rising frequency of the end-member's modal frequencies. However, at least one full year of monitoring would be required to investigate whether the vibration frequency returns to the values after one annual cycle through all seasons, or whether there is an offset indicative of progressive weakening (cf. . Stress release signals in the form of short seismic pulses do also argue for an irreversible component in the consequences of the diurnal forcing regime. The thermally driven physical weathering very close to the surface contributes to flaking (Collins et al., 2018) and leads to extended water pathways, which in turn increases infiltration capacity and circulation dynamics, basic foundations for subsequent weathering processes. On top of diurnal cycles, the trends of dv/v values continuously increasing for four to seven days before dropping for a few days, argue for a systematic conditioning of the Hochvogel summit that results in increased accumulated stress that is episodically released and thereby emits an increased amount of short seismic pulses beyond the typical daily rate. The dv/v proxy, especially as it is based on the 4-16Hz frequency window, representing wave lengths of tens of metres, integrates mechanical effects significantly beyond just the top few millimetres of the rock mass (Larose et al., 2015;Snieder, 2004). Thus, together with the non-diurnal cyclicity pattern, we see evidence for a process that is disconnected from the set of meteorological drivers and rather points at an internal mechanism that is driven by endogenic dynamics of the instable rock mass. We have measured the dv/v time series not between sensors on the instable southeastern rock mass. However, through frictional coupling and still intact rock bridges, the that instable mass also results in increased loading of the adjacent northwestern part of the summit, causing stress in that rock mass, which we have probed by our sensors.
Stick-slip behaviour has been found in several large mass wasting events prior to catastrophic failure (e.g., Poli, 2017;Schöpa et al., 2018;Yamada et al., 2016). Whereas this precursor signal appears to bear an exploitable potential for hazard early warning strategies during the final part of the preparation stage of an event, little is known about the earlier stages of preparation. This is because stick slip motion is overall hard to detect when the intervals between slip events are long in relation to the monitoring period. Seismically, slip events manifest as periods of increased seismic activity, for small events predominantly through the emission of short pulses (< 2 s) of 30-80Hz signals in the proximity of seismic sensors (Vouillamoz et al., 2018;, and up to 30 s, low-frequency (2-9 Hz) signals for larger events at up to 30 km distance to the sensor (Poli, 2017). In the Hochvogel summit, we also evidence 1-2 day long periods of increased seismic activity exhibiting these pulse-like properties, that result in a drop of the dv/v stress proxy, separated by several days of just background activity during which we see the buildup of stress. We suggest that this multi-day cyclic pattern may be explained as an early stage of stick-slip evolution of the instable rock mass at the summit, stressing that further, independent data on the surface expression of such a process are required to robustly confirm this interpretation. Under the assumption of its validity, we propose that during a stick phase gravity pulls the rock mass down slope, causing elastic deformation (background seismic pulse emissions), which causes buildup of stress (increase of dv/v) until eventually the rock mass slips and thereby emits more seismic signals and the stress within the rock mass is reduced. We anticipate that as failure preparation of the Hochvogel rock mass progresses, this frequency of slip events would increase, and immediately prior to failure, these patterns can turn into a continuous tremor-like signal with a frequency of several per second (Schöpa et al., 2018).
Throughout the analysed proxy data we identified superimposed cycles of mechanisms that drive the evolution of the slope instability. In addition, the period of increased stress release pulses on 25-26 August (Figure 2 f) points at a further mechanism, which may be of stochastic or higher hierarchy nature. The differences in all proxy time series between the first and second survey period may suggest also a seasonal component contributing to the preparation phase of the failure. However, in order to reveal these potentially additional components, and to fully constrain the progressive evolution towards the final stage of the mass movement, we would need a longer, complete data set from a constant and sufficiently dense seismic network on the summitobviously a task of future research at this site.
We have presented a multi proxy based anatomy of a major future slope failure during a key stage of its preparation phase, which is inherently difficult to survey. We propose that it is possible to generalise the Hochvogel observations to also describe the fate of other high mountain peaks prone to failure. Most seismic proxies can be measured and interpreted in a similar way as for the case presented here. However, the outcomes of this study have also shown that it will be key to design a system that is reliable and capable of operating under the hostile environmental conditions encountered at this solitary peak. Such a system should ideally also probe the surface expressions of the evolution the rock mass undergoes, for example by dedicated direct measurements of crack opening, reference point position tracking and surface as well as below-surface temperature and moisture logs. That way, the seismically sensed effects can be linked more directly to other first order physical quantities and thus allow for a distributed, continuous and effective survey of a hitherto enigmatic yet essential phase in the evolution of landscapes dominated by steep slopes.