Interactive comment on “ Validity , precision and limitations of seismic rockfall monitoring ” by Michael Dietze

This manuscript presents a new approach for detecting and locating rockfalls using seismic signals, applied to a case study in the Swiss Alps. I find the manuscript well written, well organized and the results interesting. Validity and precision of the method have been carefully discussed, while I found the discussion about possible limitations a bit dry. I suggest to improve this part, especially given the fact that several manual adjustments and optimizations are needed in post-processing. Below are some minor comments about the main text and the supplementary materials.


Study area
The Lauterbrunnen Valley in the central Swiss Alps is a deglaciated U-shaped valley.It is flanked by up to 1000 m high, Mesozoic limestone cliffs with sometimes almost vertical walls (88.5 • ) and several hanging valleys that host more than 70 waterfalls.Talus slopes at the base of the cliff, reaching around 150 m above the valley floor, argue for substantial and sustained rockfall.The steepest wall section separates the town of Mürren above the cliff from the town of Lauterbrunnen in the valley (fig.1).Our study focused on this wall, which has minimal snow and vegetation cover throughout the year.The wider area contains further rockfall-prone locations that can deliver rockfall signals, such as the steep slopes of the Chänelegg and the ridge south of the Ägertenbach (fig. 1 a).The steep topography of the Lauterbrunnen Valley :::: with : a :::: few ::::: small ::::: ledges :::: (fig.:: 3 :: b) implies a significant free fall phase of detached rocks, followed by rockmass impacts on the cliff face or the talus slopes below, perhaps grading into moderate translocation processes on the less than 250 m long depositional areas.Rockfall activity in the Lauterbrunnen Valley has been monitored by repeated TLS since 2012 (Strunden et al., 2014), yielding 122 detected rockfalls (523.72 m 3 in total) over an 18 month investigation period.These events appear to be evenly distributed throughout valley walls (15.13 events per year and km 2 ) with most frequent events being smaller than 1 m 3 .

Equipment and deployment
High resolution point clouds ( :::: with : a limit of detection ::: (i.e., ::: the ::::::: smallest ::::::::: resolvable ::::: length ::::::: fraction :: at ::: the :::: cliff ::::::: surface) :: of about 11 cm ) were generated by TLS, using an Optech ILRIS-LR terrestrial light detection and ranging (lidar) scanner with : a scan frequency of 10 kHz and a reflectivity of 80 % at 3 km distance.Scans were recorded during two field campaigns on 22 September 2014 and 28 October 2014.The TLS data collection and processing approach used in this study is identical to that of previous work conducted in the same study area (for details see Strunden et al., 2014).To ensure sufficient overlap and to avoid topographic shading effects, the study area was scanned from five different positions (see fig. sensor was installed in a small hand dug pit at 30-40 cm depth.Seismic activity was recorded for 89 days, between 1 August and 28 October 2014.In this study only the period bracketed by the two TLS surveys is used (22 September-28 October).
For further analyses a digital elevation model (DEM) of the wider study area with 5 m grid size (swissALTI3D) was used, transformed to the UTM coordinate system and resampled to 10 m grid size.

Lidar data processing
Point clouds were processed with the "Joint Research Center 3-D Reconstructor 2" software, adjusted manually and merged using control points and a best fit algorithm to minimize differences in the overlapping data.Rockfall detachment locations and volumes were calculated from the two data sets using the inspection tool and the cut and fill algorithm.Photographs recorded during scanning were used to confirm that the detected volume changes were not caused by processes other than rockfall (e.g., vegetation growth).Measurement uncertainty was estimated based on scan differences from stable control regions (for details see Strunden et al., 2014).Detachment area coordinates were obtained by georeferencing the rasterised point cloud data on referenced topographic maps and orthoimages.Given the typical rockfall volumes < 1 m 3 :: Strunden et al., 2014), location uncertainty should mainly result from the georeferencing process and is quantified by the root mean square error (RMSE).All location coordinates were rounded to the full meter and transformed to the UTM coordinate system.

Seismic data processing: Event detection
A single seismic station records 200 samples per second and spatial :::::::: geometric :::::: signal component, resulting in more than 311 million measured values per day.Hence, potential rockfall events must be identified (picked) automatically from the stream of data.However, for rockfall events with volumes usually below 1 m 3 (Strunden et al., 2014) it is challenging to find reasonable parameter settings for any picking algorithm.Therefore, the seismic time series of all operating stations were manually screened during a control period, 22 September-1 October, to find reference events for parameter definitions.
When the onset of an event is recorded it will not affect the LTA value but have a significant effect on the STA value, thus increasing the ratio.When the seismic signal returns to background, the STA values approach the LTA value again, which lowers the ratio towards one.The STA-LTA-ratio picker thus has four relevant parameters: the lengths of the STA window and LTA window, a threshold value to define the start of an event and another threshold value defining the end of an event.
The detected potential events should agree with these observations.All successfully evaluated events were used for subsequent analyses.
All picked events were clipped with a buffer of 3 s before and after the event and then migrated.Locations with a likelihood quantile below ::::::::::::: cross-correlation ::::: value ::: R 2 ::::: below ::: the : 0.95 ::::::: quantile were removed and the remaining values were normalised between 0 and 1. Events located along the margin pixels of the distance map of the study area were rejected.Only events inside the area of interest (cf.fig. 1) were used for validation.The threshold quantile value of 0.95 to clip location areas is arbitrary though in the range of values from the literature (cf.Burtin et al., 2014) :::::::::::::::: (Burtin et al., 2014).To investigate the effect of varying this value on the number of rockfall locations inside the resulting uncertainty polygon was tested by changing the value from 0.9-1.0 and recording the number of TLS-based detachment locations and corresponding downslope trajectories, which remained inside the uncertainty polygons.
Location differences ∆P max were calculated as the minimum planform Euclidean distance between the highest value of the seismic location estimate (P max ) and the downslope trajectory line of the corresponding TLS-based detachment pixel.
The direction of the trajectory line was defined by the average cliff face azimuth (99 ± 44°).This approach was chosen because seismic signals can only be emitted at the detachment zone or rockfall impact sites below it, and since the cliff face is nearly 90°:steep there is a high likelihood that the rock mass will follow the line of steepest descend without much deviation.
Uncertainties arising from deviations of the rock mass from this line cannot be accounted for.

Lidar-detected rockfalls
Between 22 September and 28 October, ten rockfall events were detected by TLS.The events were spread over the entire monitored part of the cliff, but the southern section, near stations "Sweaty Herbs" and "Confident Pulse", hosted 50 % of all events.The smallest detected rockfall (event 5 in table 1) had a volume of 0.053 ± 0.004 m 3 while the largest rockfall (event 10 in table 1) had a volume of 2.338 ± 0.085 m 3 .The average volume of rockfalls in this period was 0.482 m 3 .A summary of all rockfall events including location coordinates based on TLS and seismic data is shown in table 1.With only one exception (event 6), all rockfalls detached from the lower part of the cliff, some almost at the base (cf.fig. 3 b, table 1).The georeferenced RMSE in the event locations was between 4.8 and 17.5 m.The range in RMSE values calculated depends on the number of identified ground control points (between 8 and 17 per scene) as well as the size and perspective of the referenced image.
Manual screening of seismic records during the control period (22 September and 1 October) yielded evidence of two rockfalls, events 7 and 10 of the final data set (table 2).One of these rockfalls (event 7, fig. 4 b) generated two short, distinct bursts of seismic energy, less than 2 s apart, followed by a rise of the seismic signal about 7.5 s later (cf.:: see : fig. 2 for details).
The first burst contains frequencies between 30 and 60 Hz, while the second peak mainly has frequencies below 20 Hz.The subsequent strengthening signal is again dominated by frequencies between 30 and 80 Hz.The entire sequence was recorded by all operating stations, though with different amplitudes, from about ± 0.38 µms −1 ::::: µms −1 : at station "Sweaty Herbs" to ± 4.9 µms −1 ::::: µms −1 : at station "Funny Rain".The maximum time offset between event onsets at the stations was 0.51 s.The STA/LTA :::::::::::: STA-LTA-ratio : values reached up to 7 for the first two peaks and decreased below 2 before grading to the next rise.
Based on the above characteristics of event 7 and similar properties for event 10 from the control period, the parameters for event picking of the entire data set were defined.The STA/LTA : , ::: i.e., ::: the ::::::::::::: STA-LTA-ratio threshold to define the start of an event was set to 5, the threshold for defining the end of an event to 3. Note that this approach does not yield a correct start and (semitransparent line graph) as well as the picker algorithm characteristics (STA-LTA-ratioas well as , : "on"-and "off"-thresholds).end time.However, the location approach is not based on exact onset times but is used with the addition of a 3 s wide buffer before and after an event.The minimum SNR of an event at the picking station "Gate of China" was set to 6.
Thus, to eliminate false events picked by the STA/LTA ::::::::::::: STA-LTA-ratio approach the minimum duration of an event was set to 0.5 s to remove rain-related picks and the maximum duration was set to 20 s to remove earthquakes.Additionally, the ::: The minimum average SNR value among all stations was set to 6.The STA/LTA-picking :::::::::::: STA-LTA-ratio : approach yielded a total of 2175 potential events.After application of the automated rejection criteria the number decreased to 511.These 511 events had to be manually screened and included 455 spurious or unknown events, 19 short earthquakes and 37 potential rockfall signals.The most common spurious event type was associated with train traffic.This type of signal could not be eliminated by any automatic routine and had to be removed manually.The remaining earthquakes had an average ::::::::::::: STA-LTA-based : duration of 11.9 +4.6 −4.0 s (median and quartiles) and were also removed manually.The 37 detected potential rockfall events had ::::::::::::: STA-LTA-based durations of 4.7 +2.8 −2.0 s.Several of the potential rockfall events had very weak seismic signals, with average SNRs below 8 (n = 8 :::: eight ::::: cases) but the majority generated average SNRs of 11.2 +2.8 −2.6 .Although the SNR of an event is somewhat related to its duration, this relationship is rather weak (r = 0.37).Likewise, the correlation coefficient between SNR and the log of the maximum signal envelope amplitude (246 +138 −108 counts) is not higher than 0.44.interpretation ::: (not :::::::: including :::::::: subsequent ::::: talus :::: slope ::::::: activity).: SNR denotes range of signal-to-noise ratios among all recording stations.
ID Time (UTC) duration (s) SNR f def ault (Hz) f opt (Hz) A (nms −1 ) range, there is no such clear result for the former approach.The solid black lines in fig.6 show two velocity ranges with high P max values, between 1000 and 1800 ms −1 and between 2200 and 3000 ms −1 .Due to the recent deglaciation and persistent rockfall activity, the limestone cliffs of Lauterbrunnen appear rather compact and only marginally weathered.Thus, there is no reason to assume much lower values than those of 2000-3300 ms −1 for S-waves in limestone from empiric tests (Bourbie et al., 1987).Accordingly, the first local maximum at lower velocities did not yield any consistent rockfall locations along the 10 cliff, even when the other criteria clearly pointed at a rockfall.The average P max ::: R 2 values for the higher velocity range from a broad plateau of equally likely velocities including 2700 ms −1 .Thus, based on information from both approaches, the average seismic wave velocity for running the location routine was set to 2700 ms −1 .Without the existence of independent locations of rockfall detachment zones, seismic velocity can only be constrained with low uncertainty by active seismics.

Location of rockfalls
Applying the location routine to the 37 potential rockfall events placed nine of them in the area of interest covered by our TLS surveys and the seismic network (cf.fig.1).Eight further events were located along the west-facing valley side.Most of these had poor location constraints due to low SNR or inappropriate fits of the overall time delays of the signal envelopes.The other events could either only be located along the margins of the distance maps as the closest approximation for more distant sources, or were located west of the Lauterbrunnen Valley, higher in the catchment.One event, which showed all characteristics of a very proximal rockfall and subsequent rock avalanche but exhibited an extraordinarily wide frequency range (cf.event 10 in fig.8) could successfully be located within the area of interest by manually setting the location frequency window to 5-35 Hz.
Thus, after extensive processing and manual verification, all ten TLS-detected rockfalls could be independently located by the seismic approach.SNRs of all ten events were above 5 and up to 59, depending on the magnitude of the event and the distance of the source to a seismic station.With the exception of the manually adjusted settings for event 10, the default settings resulted in an average difference between TLS (i.e., line of steepest descend from detachment zone) and seismic location of

+59
−29 m.The maximum difference was 761 m (event 1, cf table 1) because a significant part of the location estimate polygon for this event, including the location of P max , was placed on the other valley side, separated from the cliff face by the entire valley floor.However, all TLS-based events were located within the default 95 % threshold uncertainty areas ::::::::: uncertainty :::: areas ::::::: defined :: by ::: the :::: 0.95 ::::::: quantile, most of which were elongated by several 100 m in the north-south direction in plan view (cf.table 1).
Increasing the quantile thresholds to define the uncertainty polygons for each location estimate reduces their area, which eventually leads to a drop of the number of matches with TLS-based event location (fig.7).Up to a threshold value of 0.9726 :::: 0.973, all ten rockfalls are included in the uncertainty areas.

Rockfall location
All ten TLS-based rockfall events were confirmed with an average location error along the rockfall trajectory of 33 m when the frequency window of the location algorithm was adjusted manually.Without this optimisation, which is only possible when reference data are available, the location deviation was 81 m on average.This is comparable with errors of about 80 m from a rock avalanche study on Montserrat, Lesser Antilles, using a similar location approach with a network of 11 stations (Levy et al., 2015).However, that study :: had :: a ::::: larger ::::::: network :::::::: aperture ::: and : focused on event volumes of 10 3 -10 6 m 3 .Instead, rock mass volumes in the Lauterbrunnen Valley were generally well below 1 m 3 and our study had only four operating seismic stations, organised in topology and station spacing comparable to those from other studies (Lacroix and Helmstetter, 2011;?;Burtin et al., 2016) ::::: (Hibe The TLS-based detachment locations and their rockfall trajectories are within the areas defined by the 0.95 quantile threshold (fig.8).Only when independent constraints on the location of the seismically recorded events are available, is it possible to investigate the validity and effectiveness of this arbitrary threshold.In this study area, the threshold can be increased up to 0.9726 :::: 0.973 : to still provide a valid uncertainty estimate for possible rockfall locations/trajectories.Effectively, this means that the area of each uncertainty polygon can be decreased by 45 %.
An important issue is that for some rockfalls the best location estimate (P max ) is above the actual rockfall detachment zone.
This may be related to the extreme topography of the Lauterbrunnen Valley.The studied rockwall is up to 800 m :::: high, yet it is represented by as little as four plan view pixels in the 10 m DEM and distance maps (cf.ranges of z seis in table 1).Arguably, the lateral offset of rockfall location P max from the line of steepest descend is more important from a hazards point of view.
Assigning the locations of the ten seismically detected rockfalls to those detected by TLS is unambiguous in most cases.
However, rockfalls with comparable volumes from similar detachment heights can be hard to distinguish.For example, events 3 and 4 are located 44 m apart, at 1108 and 1064 m asl., and released 0.201 and 0.175 m 3 of rock, respectively.Accordingly, their seismic waveforms and PSDs (fig.8) look very similar and there remains ambiguity about the seismic identification as stated in table 1.This has consequences for the temporal information associated with the seismic data.But in this case, both events occurred on 20 October, one at 3 pm, the other at 7 pm.Ambiguity also arises for events 1 and 2. However, there the rockfall volumes allow for a better matching with the seismic results.Event 1 entrained 0.201 m 3 whereas event 2 displaced :::::::: mobilised : only 0.063 m 3 from a near identical position and fall height.Accordingly, the emitted seismic energy of event 1 should be significantly higher than event 2, which is reflected in the corresponding PSD, where event 1 shows a much longer and more powerful signal.Hence, if the geometric properties of the released rock masses are sufficiently distinct, it is possible to disentangle nearby events from the detailed seismic information.
The largest event (2.338 m 3 ) did not yield the highest signal intensities or longest duration, and vice versa for the smaller events.The combination of released volume, detachment height above cliff base, the number, distance and strength of intermediate impacts, the degree of fragmentation during the fall phase and the fate to the rock mass on the talus slope (direct deposition, subsequent downhill translocation, entrainment of impacted talus) result in a polymorphic seismic signal, which complicates direct links of seismic parameters with geometric or kinetic properties of the detected rockfalls at this spatial The apparent :::::: seismic detection limit for rockfall volumes in the Lauterbrunnen Valley is well below 1 m 3 .This is remarkable given that the stations are mostly more than one km apart and that most of the rockfalls used for validation originated at the lower cliff parts, resulting in limited kinetic energy upon impact.Location feasibility is however not only determined by the rockfall volume and drop height.The distance between impact location and location of the seismic stations, the inelastic attenuation properties of the rock and the energy dissipation due to rock fragmentation (e.g., ?) ::::::::::::::::::::: (e.g., Hibert et al., 2011) also determine the potential to successfully locate the rockfall.The possibility to analyse rockfalls as small as 0.053 m 3 , impacting at distances of 170-1950 m from the seismic stations, makes seismic monitoring a method that is able to reveal events well below the resolution of most other post-event survey techniques with the exception of TLS surveys.
Unlike other rockfall survey techniques, seismic methods allow for monitoring of rockfalls with high temporal resolution, down to fractions of a second.During the first half of the monitored month only two rock masses were released, while the other half of the month saw the majority of events.Beyond this, the high temporal resolution allows connecting the events to ambient conditions and trigger mechanisms, and to study process interactions (cf.Burtin et al., 2014) :::::::::::::::::::::::::::::::::::::: (e.g., Helmstetter and Garambois, 2010; Burtin

Rockfall anatomies
Seismic monitoring allows detailed insight into the dynamics of rockfalls.The example event (fig.2) consists of three distinct phases and lasts in total for almost a minute.Phase 1 (less than one s duration) is the first notable seismic activity after minutes of calm at all stations.It reflects the seismic signal associated with initiation of the rockfall event.The high fre-quency content most likely corresponds to the ::: may ::::: either :::::::::: correspond :: to ::: the ::::::: rebound :: of ::: the :::: cliff :::: after :::::::::: detachment :: of ::: the ::::: mass :::::::::::::::::::::: (e.g., Hibert et al., 2011) or ::: the : opening and propagation of fractures rather than impacts of a moving rock mass.This ::: The :::: latter : interpretation is supported by seismic records from the Illgraben, Rhone Valley, Switzerland, that show an exponentially increasing density of signals, which indicate cracking or fracture propagation (Zeckra et al., 2015) starting days before a 10 4 m 3 large rock avalanche took place (Burtin et al., 2016).The spectral properties of these signals (short, less than 1 s pulses at 20-50 Hz), recorded by a seismic station about 150 m away from the initiation zone of the rockslide are very similar to the first phase of the rockfall from the Lauterbrunnen Valley (fig.2).
Phase 2 (one s duration) begins 1.7 s after this fracture propagation phase and may reflect :: the : impact of the released rock mass on the cliff face.The predominantly low frequency content implies that the mass is still intact upon the first collision.Low frequencies can only be generated by large rock masses that convey a high momentum rather than a series of smaller particles hitting a surface simultaneously (Burtin et al., 2016).The strong impact likely caused fragmentation of the rock, because there is no low frequency content in any of the later signals from this event.The rock fragments experienced a free fall phase (calm period in all signal waveforms) of approximately 7.5 s, corresponding to a drop height of 271 m.With a detachment elevation between 1048-1123 m asl.this places the impact somewhere in the central part of the talus slope that reaches from 910 m asl.
at the cliff base to 820 m asl. on the valley floor.
Phase 3 (about 40 s duration) represents the continuous impact of the fragmented rock mass on the talus slope for tens of seconds.This activity very likely graded into a phase of downslope translocation of debris and entrainment of further talus, because the .:::: The : PSD of phase 3 shows a gradual shift from higher to lower frequencies with time.This is likely due to the longer lasting downslope transport of larger particles due to their higher momentum.

Conclusions
The detachment locations of ten rockfall events, totalling a volume of 4.789 ± 0.100 m 3 , were detected by TLS over 37 days.
Using broadband seismometers, these events were independently detected and located with an average deviation of 81 +59 −29 m.
Further seismic rockfall signals were detected and located outside this instrumented cliff area.The seismic signatures allow i) insight into the dynamics of single events, ii) quantification of the exact event onset time and duration, and iii) calculating minimum fall heights.Volume estimates based on the emitted seismic energy or peak ground acceleration were not possible for the small rockfall events identified in this study.This was mainly due to the influence of intrinsic factors, such as the proportion of energy consumed for fragmentation during the event or contribution of mobilised debris to the seismic signals upon impact on the talus slope.However, the ::: The : method allows detecting rockfalls with volumes as small as 0.053 m 3 .
It is possible to monitor rockfalls sensu strictu ::::: stricto : with a significant free fall phase and a pronounced short impact phase.
This extends the previous field of applications of environmental seismology to more extreme settings.Furthermore, seismic monitoring is not restricted to the instrumented cliff face but allows detection of rockfall signals from other areas such as the other valley wall and locations higher up in the catchment, though with sometimes only poor constraints on the location of these events.
There is significant potential to optimise the parameters for event location but there is no straightforward way to do this without independent auxiliary information.Hence, a realistic location error range along the trajectory of released rocks is 52-140 m (interquartile range).The height and location of the detachment zone can only be provided by seismic methods if the detachment process can be recorded and the subsequent impacts of the released rock mass can be located with sufficient confidence to allow back-calculation of the falling time.Rockfall release zones that are separated below the level of seismic location confidence can be deciphered from each other if the released volumes are different from each other and generate sufficiently distinct seismic characteristics.Hence, combining seismic and TLS methods can provide a very detailed complementary picture of rockfall activity.

Data and code availability
The seismic data used in this study is available in the supplementary materials, along with a detailed documentation about how to use it to reproduce the results of this study.The digital elevation model data set cannot be made freely available, ::: but :::: may ::: be ::::::: replaced :: by ::::::::: equivalent :::: data :: to :::::::: reproduce ::: the :::::: results.TLS point cloud data are available upon request.
Author contributions.Michael Dietze contributed to seismic fieldwork and data analysis.Solmaz Mohadjer contributed to TLS fieldwork and data processing.Jens M. Turowski, Todd A. Ehlers and Niels Hovius contributed to equipment provision, project planning and data analysis.All authors contributed to manuscript preparation.

Figure 1 .
Figure 1.The study area Lauterbrunnen Valley.(a) Schematic map with location of seismic stations, TLS positions and anthropogenic noise sources (settlements, technical infrastructure).(b) Photograph of the instrumented east-facing cliff face of the Lauterbrunen Valley with the Breithorn and Tschingelhorn in the background.Seismic stations (yellow stars) are separated by 1200 m on average.

Figure 3 .
Figure 3. Rockfall detachment zones determined from TLS mapping.(a) Overview (aligned point cloud data) of the :::: about ::: 2.7 :: km ::::: long, instrumented east-facing wall ::::: stretch : of the Lauterbrunnen Valley with rockfall detachment zones (red dots) and seismic stations (yellow stars, station names :: and ::::::: distances : see fig. 1).(b) Close-up of the southern rock wall section with the detachment zones of events 3 --: 5 at elevations less than 100 m above the talus slope.(c) Boxes show rockfall detachment patterns on the rock wall.Released rock volumes and uncertainties are given below each box.Event numbers are the same as in (a) and tables 1 and 2).

Figure 7 .
Figure 7. Number of rockfall trajectories inside location estimate polygons as function of minimum location estimate quantile.

Table 2 .
Rockfall events detected by seismic monitoring.IDs correspond to those in table 1.