the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 05 Jun 2026
Method for near real-time detection of snow avalanches using Distributed Acoustic Sensing
Abstract. We present a novel method for near real-time snow avalanche detection using Distributed Acoustic Sensing (DAS). A ∼10 km long telecommunication cable permanently installed along the avalanche-prone Flüelapass road (Swiss Alps) was continuously monitored over a full winter. Avalanches, including events that did not physically reach the cable, were clearly recorded and confirmed with photographic evidence. To discriminate avalanches from anthropogenic signals, we introduce a dual-frequency short-term over long-term average attribute that produces coherent high-value spatio-temporal signatures for avalanches, while vehicles predominantly generate negative values with pronounced move-out. The workflow consists of (1) a quasi-instantaneous threshold-based trigger to detect onset time and location, followed by (2) a rapid waterfall image analysis to estimate event extent and invalidate traffic-induced alerts. The first step issues alerts with millisecond-scale latency and meter-scale spatial resolution. The second step introduces additional latency, as it requires the event to sufficiently develop in order to assess its spatio-temporal morphology and confirm or discard the initial trigger. Our system issued alerts only 4.5 ‰ of the time when the road pass was open (i.e. 2.5 hours over 23 days), demonstrating the robustness against traffic, and 0.36 ‰ of the time when the pass was closed (i.e. 55 minutes over 108 days). Among those, a total of 73 potential avalanches were identified, most of them occurring during three independently documented avalanche episodes. These findings demonstrate that DAS represents a viable and cost-effective solution for operational real-time avalanche monitoring, with potential applicability to broader natural hazard detection.
- Preprint
(19283 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-2373', Antoine Turquet, 16 Jul 2026
-
AC1: 'Reply on RC1', Pascal Edme, 12 Oct 2026
We thank the reviewer for his careful assessment of our manuscript and their constructive comments and suggestions. Our detailed point-by-point responses are provided below.
1) The proposed detector seems to work well with avalanches that exhibit weak or delayed low-frequency STA/LTA responses relative to their full-band responses. However, large avalanches or the avalanches that are very close to the cable may generate higher/faster low-frequency STA/LTA responses and thus lower HVR amplitudes. Such events could therefore be missed or potentially misclassified as moving anthropogenic sources. It would be useful for the authors to discuss this potential limitation and clarify whether events with strong simultaneous low- and high-frequency responses were observed in the dataset.
We thank the reviewer for this insightful comment. We agree that events producing nearly simultaneous full-band and low-frequency STA/LTA responses may result in very weak or even absent HVRs. This is, in essence, the principle underlying our discrimination between avalanches and anthropogenic noise, particularly vehicle traffic. We acknowledge for instance that a very short avalanche occurring close to the cable may therefore go undetected. This represents a potential limitation of the proposed attribute, although we did not observe such cases in our dataset. Larger avalanches, however, are unlikely to be affected by this limitation because of their greater spatial and temporal extent. As the moving seismic source approaches or moves away from the cable, the evolving signal is likely to generate some temporal separation between the full-band and low-frequency STA/LTA responses, thereby producing detectable HVRs. Moreover, even when the two responses are nearly simultaneous, an HVR may still develop if the full-band response is sufficiently stronger than the low-frequency response.
Finally, it is important to emphasize that STA/LTA responses reflect relative changes in signal energy rather than absolute amplitudes. Consequently, the proposed method is relatively insensitive to signal strength and can detect even weak events, provided they emerge sufficiently above the background noise. In practice, our main challenge appears to be an excess rather than a lack of sensitivity, resulting in a substantial number of false-positive triggers. These were systematically investigated through extensive visualization and analysis of the DAS recordings. This detailed examination also gives us confidence that no clearly identifiable avalanche signatures were missed by the detection algorithm. We will further clarify this point in our response to RC1’s comment 3 andRC2’s first comment.
2) The (1/\sqrt{2}) term in Equation 2 is not sufficiently explained. In its current form, assuming both STA/LTA quantities fluctuate around one during stationary background noise, the resulting attribute does not appear to fluctuate around zero.
We thank the reviewer for pointing out that the (1/\sqrt{2}) term in Eq. (2) was insufficiently explained. This term was introduced empirically, based on the observed behavior of the two STA/LTA attributes under ambient noise conditions. While both STA/LTA ratios fluctuate around unity, their combination in Eq. (2), (\sigma-|\sigma_{\mathrm{LF}}-1|), was observed to fluctuate around approximately (1/\sqrt{2}). Subtracting this constant therefore brings the resulting attribute close to zero under background noise conditions.
We acknowledge that this offset is not a universal theoretical value. It depends on the statistical properties of the STA/LTA fluctuations and may vary with the DAS interrogator, acquisition settings, and noise characteristics, particularly the frequency-dependent noise levels. More generally, the term (1/\sqrt{2}) could be replaced by an adjustable constant (C), calibrated from the background noise statistics for a given acquisition configuration. Importantly, this constant simply shifts the attribute and does not affect the underlying discrimination mechanism, provided the detection threshold is adjusted accordingly. We will clarify this empirical choice and its potential generalization in the revised manuscript.
3) A discussion of false negatives would be valuable for understanding the limitations and operational applicability of the method. Were there independently observed avalanches that the system failed to detect?
We thank again the reviewer for this important comment. As discussed in our response to Comment 1, our detection strategy deliberately favors sensitivity, at the expense of a relatively large number of false-positive triggers. These were subsequently examined and classified through detailed visual analysis. In addition, the complete continuous DAS dataset was visualized, rather than only the triggered windows, providing further confidence that no obvious avalanche signatures were missed by the detection algorithm.
Independent observations are unfortunately limited. Webcam observations, available only from 25 March onwards, provide confirmation of a few events near the top of the pass, while Figure 12 compares our detections with avalanche activity independently reported in the SLFPro database. The good temporal consistency, particularly during the main avalanche episodes, further supports our results. However, SLFPro reports are neither exhaustive nor sufficiently localized to establish event-by-event correspondence with our detections. Consequently, the limited independent observations do not allow us to reliably assess the number of missed avalanches or estimate event-level recall.
4) It would be a nice add-on if the authors could discuss the seismo-acoustic pathways of avalanche signals, signal phases and limitations of DAS more explicitly.
We thank the reviewer for this interesting suggestion and agree that a more detailed investigation of the seismo-acoustic wavefield generated by avalanches would be scientifically valuable. However, we deliberately chose to focus this study on the practical challenge of avalanche detection, rather than on the detailed characterization of avalanche-generated signals. Our primary objective is to present a simple, computationally efficient detection strategy that can discriminate avalanche activity from anthropogenic noise.
A detailed investigation of the underlying physical mechanisms would require a substantially different analysis and falls beyond the scope of the present study. We believe that maintaining this focus is important for the clarity and conciseness of the manuscript, while emphasizing its methodological and operational contributions. A more detailed characterization of these physical mechanisms, particularly their influence on avalanche signatures and DAS detection capabilities, is the subject of ongoing research, including within our own group, with the aim of presenting the findings in future publications.
Citation: https://doi.org/10.5194/egusphere-2026-2373-AC1
-
AC1: 'Reply on RC1', Pascal Edme, 12 Oct 2026
-
RC2: 'Comment on egusphere-2026-2373', Anonymous Referee #2, 31 Aug 2026
Major comments:
247-252: This appears to be the main summary for the method's performance. There does not seem to be a specific section for avalanche detection statistics, even if using only confirmed avalanche events (line 289). I would expect the authors to report how many confirmed avalanches were detected and missed, together with event-level precision and recall where the available observations allow. If the independent catalog is insufficient for these calculations, this limitation and the resulting inability to estimate false negatives should be stated.
251: A quantitative comparison with conventional STA/LTA is also needed. Saying that it would 'likely' trigger quasi-continuously is insufficient.
Lines 323-328: The wording appears to signal that the 73 potential avalanches identified occupy the entire 4.5 per mille (open road) and 0.36 per mille (closed road) of detections. There should be more clarity, not only in this section, but also before in the manuscrupt regarding what these numbers actually represent. For example, explicitly mentioning that the possible avalanches were manually picked from the trigger report.
Line 335-336: Applicability to other hazards is not established. Lahars and debris flows can produce coherent activity along channels, potentially resembling the traffic signature that this method rejects, so transfer would require hazard-specific and site-specific calibration and validation.
Minor comments:
Line 21: “several hundred million Swiss francs”
Line 30: Review also https://doi.org/10.1029/2020JF005741 and https://doi.org/10.1002/2014GL061254 for applications in other geographic contexts.
Line 44: See https://doi.org/10.1016/j.jvolgeores.2018.09.007 and https://arc.lib.montana.edu/snow-science/objects/ISSW2024_O8.2.pdf
Lines 45, 50: remove the space in “Pérez- Guillén”
Line 46: “Detections at distances of up to 15 km from an avalanche”
Line 71 and Fig. 9 caption: replace “noises” with “noise” or “noise sources”
Line 80: “tap tests on site”
Lines 90, 93: “As an additional assessment tool” and “Finally, observations…”
Fig. 1 caption: “Map data”
Fig. “Signals at distances <1 km are mostly generated by cars”
Fig. 5 caption: “in the early afternoon of 15 April”
Lines 128–135: standardize the notation as (f)–(k)
Lines 139–140: use “in real time” and remove the comma in “interpretability, and reproducibility”
Line 153: “channel by channel”
Line 166: “snow avalanches” (no hyphen)
Line 198: the apostrophe after filter, at the end right after the colon:
Line 205: “small avalanches”
Fig. 8 caption: “each individual car”
Fig. 9 caption: “an NVR patch”
Line 222: “referred to as”
Fig. 11 caption: “ “Burst events” instead of "Bursts" (singular) and “right hand side panel” should be “right-hand panel”.
Line 246: “closely spaced”.
Line 262: “vehicle-related” (with hyphen)
Line 264: “truck-generated” (with hyphen)
Line 280: “on the night of”
Line 288: insert a comma after the parenthesis ending “raw strain-rate data)”
Line 313: “system’s sensitivity”.
Lines 324–325: weird wording: revise as “while near-field signals generated by cars very rarely triggered the detector”.
Line 334: “decision-making”
Line 339: “MATLAB”
121: Maybe this is discussed later, but could this inform any potential for avalanches whose path do not intercept the FO?
128: This paragraph is formatted as an image caption. Ideally, it can be rewritten in prose without explicit panel mentions. Instead, something like "this avalanche does not include clear modes in its time-space domain as shown in Fig. 5c". See: https://publications.copernicus.org/for_authors/manuscript_preparation.html
119-123 and Fig. 3: Explain how the exact 19 March event time was independently constrained when the deposit was photographed on 25 March (was this aided by the DAS results?). Fig. 3 should include the data of the actual photo collection.
Citation: https://doi.org/10.5194/egusphere-2026-2373-RC2 -
AC2: 'Reply on RC2', Pascal Edme, 12 Oct 2026
We thank the reviewer for the careful assessment of our manuscript and their constructive comments and suggestions. Our detailed point-by-point responses are provided below,
247-252: This appears to be the main summary for the method's performance. There does not seem to be a specific section for avalanche detection statistics, even if using only confirmed avalanche events (line 289). I would expect the authors to report how many confirmed avalanches were detected and missed, together with event-level precision and recall where the available observations allow. If the independent catalog is insufficient for these calculations, this limitation and the resulting inability to estimate false negatives should be stated.
We thank the reviewer for this suggestion and agree that event-level precision and recall would be valuable indicators of detection performance. However, as also discussed in our response to Reviewer 1 (Comment 3), it is not possible due to the lack of a comprehensive independent avalanche catalogue. Although few events were confirmed by webcam observations, and Figure 12 provides a qualitative comparison with avalanche activity reported in the SLFPro database, these observations are insufficient to establish a precise number of avalanches or calculate reliable event-level precision and recall. We will clarify this limitation in Section 4 and distinguish more explicitly between the statistics of our detection workflow (2,643 triggers, 73 potential avalanche events, including 20 with unambiguous signatures) and independently validated detection performance. We will also explicitly state that the absence of a comprehensive reference catalogue prevents a reliable assessment of false negatives. Finally, we would like to emphasize that our analysis was not limited to the triggered windows: the complete continuous DAS dataset was also visualized, providing additional qualitative confidence in the sensitivity of the proposed method.
251: A quantitative comparison with conventional STA/LTA is also needed. Saying that it would 'likely' trigger quasi-continuously is insufficient.
We agree with the reviewer that our statement regarding quasi-continuous triggering with conventional STA/LTA is not supported by a quantitative comparison.
Our statement was motivated by the high level of anthropogenic activity recorded along the road, when traffic-related signals are present almost continuously somewhere along the cable (as illustrated, for example, in Figure 6). A meaningful quantitative comparison with conventional STA/LTA would require defining a separate detection configuration, including the frequency band, STA/LTA threshold and potentially additional processing choices. These parameters could strongly influence the resulting triggering statistics, making such a comparison less straightforward than it might initially appear. As the objective of this study is to introduce a detection strategy, rather than to benchmark different detection algorithms, we prefer not to extend the analysis in this direction. We will therefore remove the speculative statement regarding quasi-continuous STA/LTA triggering and retain the quantitative triggering statistics of our proposed method.
Lines 323-328: The wording appears to signal that the 73 potential avalanches identified occupy the entire 4.5 per mille (open road) and 0.36 per mille (closed road) of detections. There should be more clarity, not only in this section, but also before in the manuscrupt regarding what these numbers actually represent. For example, explicitly mentioning that the possible avalanches were manually picked from the trigger report.
We thank the reviewer for pointing out this possible ambiguity. The reported fractions of 4.5‰ (open road) and 0.36‰ (closed road) do not represent the duration of the 73 identified potential avalanches. Rather, they correspond to the cumulative fraction of monitoring time during which at least one channel anywhere along the cable exceeds the HVR triggering threshold, irrespective of the source responsible for the trigger. These fractions therefore include all triggering activity, including false positives associated with traffic and other sources.
The main point we wish to emphasize is the substantial reduction in the amount of data requiring further examination. All 2,643 initial triggers were subsequently examined and manually classified, leading to the identification of 73 potential avalanche events. We will clarify this distinction in the revised manuscript, particularly in the Abstract, Section 4 and Conclusions, to avoid suggesting that these fractions correspond exclusively to avalanche activity.
Line 335-336: Applicability to other hazards is not established. Lahars and debris flows can produce coherent activity along channels, potentially resembling the traffic signature that this method rejects, so transfer would require hazard-specific and site-specific calibration and validation.
We thank the reviewer for this important observation and agree that the applicability of our method to other natural hazards has not been demonstrated. While the general detection principle may be relevant to phenomena such as lahars or debris flows, their signal characteristics may indeed differ substantially from those of snow avalanches, potentially requiring different discrimination criteria. We will therefore moderate the corresponding statement in the manuscript to avoid implying that such applicability has already been established.
Minor comments
We thank the reviewer for these careful editorial suggestions and for the additional relevant references. We will implement the proposed corrections throughout the manuscript.
Line 121 —Maybe this is discussed later, but could this inform any potential for avalanches whose path do not intercept the FO?
We thank the reviewer for this observation. Indeed, an avalanche does not necessarily need to intersect the fiber-optic cable to be detected, as the associated seismic signals can propagate through the ground over some distance. However, quantifying the detection range is particularly challenging, as it depends on numerous factors, including avalanche size, distance from the cable, local propagation conditions, and fiber coupling. This question is currently being investigated through ongoing research at other sites in Switzerland. Given this complexity, a quantitative assessment would require a dedicated study and falls beyond the scope of the present manuscript, which focuses on the development and evaluation of the detection methodology.
128: This paragraph is formatted as an image caption. Ideally, it can be rewritten in prose without explicit panel mentions. Instead, something like "this avalanche does not include clear modes in its time-space domain as shown in Fig. 5c".
We agree with the reviewer that this paragraph reads too much like an extended figure caption. We will revise it to focus on the main observations and their implications for avalanche discrimination, using Figures 5 and 6 as supporting illustrations rather than describing their individual panels sequentially.
Lines 119–123 and Figure 3
We thank the reviewer for pointing out that the chronology of this event was insufficiently explained. The avalanche deposit was discovered and photographed during a field visit on 25 March, whose original purpose was to install a webcam to independently monitor wet-snow avalanche activity. This discovery prompted us to revisit the DAS recordings, in which we retrospectively identified the corresponding avalanche signal on 19 March. Thus, the event was first identified through field observations, independently of the DAS analysis, while its timing was subsequently established from the DAS recordings. We will clarify this sequence in the manuscript and explicitly indicate the photograph acquisition date (25 March) in the Figure 3 caption.
Citation: https://doi.org/10.5194/egusphere-2026-2373-AC2
-
AC2: 'Reply on RC2', Pascal Edme, 12 Oct 2026
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 283 | 196 | 31 | 510 | 23 | 23 |
- HTML: 283
- PDF: 196
- XML: 31
- Total: 510
- BibTeX: 23
- EndNote: 23
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Dear Editors and Authors,
I have read the manuscript, 'Method for near real-time detection of snow avalanches using Distributed Acoustic Sensing.' The study is very interesting and presents valuable findings, and the manuscript itself is clear and well-organized. The work represents a strong contribution to DAS-based avalanche monitoring, particularly because it addresses the important operational challenge of distinguishing avalanche signals from anthropogenic noise.
I have only a few comments and suggestions that may help clarify the methodological scope and communicate the results more effectively.
General comments
1) The proposed detector seems to work well with avalanches that exhibit weak or delayed low-frequency STA/LTA responses relative to their full-band responses. However, large avalanches or the avalanches that are very close to the cable may generate higher/faster low-frequency STA/LTA responses and thus lower HVR amplitudes. Such events could therefore be missed or potentially misclassified as moving anthropogenic sources. It would be useful for the authors to discuss this potential limitation and clarify whether events with strong simultaneous low- and high-frequency responses were observed in the dataset.
2) The (1/\sqrt{2}) term in Equation 2 is not sufficiently explained. In its current form, assuming both STA/LTA quantities fluctuate around one during stationary background noise, the resulting attribute does not appear to fluctuate around zero.
3) A discussion of false negatives would be valuable for understanding the limitations and operational applicability of the method. Were there independently observed avalanches that the system failed to detect?
4) It would be a nice add-on if the authors could discuss the seismo-acoustic pathways of avalanche signals, signal phases and limitations of DAS more explicitly.