Autoencoder-based feature extraction for the automatic detection  of snow avalanches in seismic data

Simeon, Andri; Pérez-Guillén, Cristina; Volpi, Michele; Seupel, Christine; van Herwijnen, Alec

doi:10.5194/gmd-18-8751-2025

Articles | Volume 18, issue 22

https://doi.org/10.5194/gmd-18-8751-2025

© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-18-8751-2025

© Author(s) 2025. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 18, issue 22

Development and technical paper

|

20 Nov 2025

Development and technical paper |

| 20 Nov 2025

Autoencoder-based feature extraction for the automatic detection of snow avalanches in seismic data

Andri Simeon, Cristina Pérez-Guillén, Michele Volpi, Christine Seupel, and Alec van Herwijnen

Download

Final revised paper (published on 20 Nov 2025)
Preprint (discussion started on 27 May 2024)

Interactive discussion

Status: closed

CEC1:
'Comment on gmd-2024-76', Juan Antonio Añel, 15 Jun 2024

Dear authors,

Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".

https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
The problem is that you have not published the code and data necessary to replicate your manuscript. Our policy clearly states that all the code and data used in a manuscript must be published at the submission time in one of the acceptable repositories listed in our policy, and that the Code and Data Availability section must contain the details (links and DOIs) for such repositories. Instead this section in your manuscript reads "The code and data to develop the final models used in this study will be made available on GitLab and EnviDat"
You have provided internally a internet address containing part of these assets (not all of them, according to my understanding). This is not enough. First, the WSL server is not a repository that complies with the standards required for scientific publication; second, all the information must be available to every potential reader in Discussions to facilitate the peer-review and comments by the community, and sharing it privately with the editors fails to comply with the Discussions peer-review process.
Therefore, please, publish your code in one of the appropriate repositories, and reply to this comment with the relevant information (link and DOI) as soon as possible, as we can not accept manuscripts in Discussions that do not comply with our policy. Therefore, the current situation with your manuscript is irregular.
In this way, if you do not fix this problem, we will have to reject your manuscript for publication in our journal.
Also, you must include in a potentially reviewed version of your manuscript the modified 'Code and Data Availability' section, containing the links and DOI of the repository containing code and data.
Finally, in the current git that you have provided for the assets, there is no license listed. If you do not include a license, the code is not "free software/open-source"; it continues to be your property and nobody can use it, despite you make it public. Therefore, when uploading the code and data to the new repository, you should add a license. You could want to choose a free software/open-source (FLOSS) license. We recommend the GPLv3. You only need to include the file 'https://www.gnu.org/licenses/gpl-3.0.txt' as LICENSE.txt with your code. Also, you can choose other options that acceptable repositories provide: GPLv2, Apache License, MIT License, etc.
Juan A. Añel

Geosci. Model Dev. Executive Editor

Citation: https://doi.org/10.5194/gmd-2024-76-CEC1
- AC1: 'Reply on CEC1', Andri Simeon, 21 Jun 2024
  
  Dear Juan A. Añel
  
  We apologise for the misunderstanding and any inconvenience it might have caused.
  In the meantime, we have added the code and data used in the manuscript to Zenodo. As suggested, we also included a license text.
  DOI: 10.5281/zenodo.12162570
  Link: https://zenodo.org/records/12162570
  The GitLab repository is found at: https://gitlabext.wsl.ch/simeonan/code-egu-paper
  
  Sincerely,
  Andri Simeon
  
  Citation: https://doi.org/10.5194/gmd-2024-76-AC1
RC1:
'Comment on gmd-2024-76', Anonymous Referee #1, 15 Jul 2024

The authors have applied deep learning autoencoder models for the automatic and unsupervised extraction of features from seismic records. These extracted features were then used in classifiers to identify snow avalanches. This study presents a novel and relevant approach to enhance machine learning predictions, which could be useful not only for identifying snow avalanches but also for detecting other types of natural events. The overall methodology is well-defined, and the manuscript is well-written and easy to follow. I recommend the publication of this manuscript after the following issues are addressed

Major:
1) Given that the models tend to miss the onset of an avalanche, the authors should have included a scenario where only verified avalanches were used, excluding non-verified ones during the training of the autoencoders. While I am not suggesting that this must be incorporated in the revised version, as this conclusion emerged only after the study was completed, it is still worth mentioning.
2) How were the machine learning algorithms implemented, including details such as programming languages and libraries used?

Minor:
Line 218: “The best model from the cross-validation procedure (Table F2) was composed of convolutions with kernel size 20 (or 0.1 s) and stride 10. “ Was the MSE (Mean Squared Error) the primary metric used for classification?
Line 228: “As an activation function, we use the leaky rectified linear unit (leaky ReLU; (Xu et al., 2015))”. Was the activation function unchanged during the hyperparameter optimization process?
Line 318: Please replace “This for” by “For this”

Citation: https://doi.org/10.5194/gmd-2024-76-RC1
- AC2: 'Reply on RC1', Andri Simeon, 28 Aug 2024
  
  We would like to express our gratitude to the reviewer for taking the time to evaluate and provide valuable feedback on our manuscript. The responses to the individual comments are found in the attached PDF.
  
  Citation: https://doi.org/10.5194/gmd-2024-76-AC2
RC2:
'Comment on gmd-2024-76', Anonymous Referee #2, 27 Jul 2024

The paper is dealing with snow avalanche detection using autoencoded seismic data. It uses one study side in Davos and events were picked and the performance tested. The article is well written and of interest for publication. However, the following should be adressed.
Discuss the effects of various avalanche types in relation to models and autoencoders. Explain how the findings can be applied to other study sites and what the specific considerations are in this context. That could be also highlighted in the comparison with the former studies.
The conclusions and the further use should be more clear.
Please, avoid repetitions throughout the article.

Citation: https://doi.org/10.5194/gmd-2024-76-RC2
- AC3: 'Reply on RC2', Andri Simeon, 28 Aug 2024
  
  We are thankful to the reviewer for the valuable suggestions on how to improve our manuscript. We appreciate you to evaluate our work and we will revise the paper following the suggestions. Please find the responses to the individual comments in the attached PDF.
  
  Citation: https://doi.org/10.5194/gmd-2024-76-AC3

Peer review completion

AR – Author's response | RR – Referee report | ED – Editor decision | EF – Editorial file upload

AR by Andri Simeon on behalf of the Authors (11 Sep 2024) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (05 Oct 2024) by Le Yu

RR by Anonymous Referee #3 (24 Oct 2024)

Suggestions for revision or reasons for rejection

This study develops deep-learning models to identify seismic signals of avalanches.
I am a seismologist working on non-earthquake signals (avalanches, landslides, glaciers...).
But I have no expertise on deep learning methods.
The introduction and the seismological methods are well described. But I had trouble to understand the deep-learning methods.
I think the authors did not make much effort to make their work understandable and interesting to readers that are not familiar with deep learning methods.
I did not know many terms (f1 score, recall ...), but I learned from other sources. But readers of GMD are probably more familiar with deep learning methods than me.
I am thus not able to criticize the part on deep leaning methods.
The results of the model seem correct, but not as good as previous similar studies (eg, Provost et al), and maybe not good enough for real time warning.

Main comments

1) Objectives
It is not clear to me what are the long-term objective of this work, beyond showing that machine learning can detect seismic signals of avalanches, more or less as well as experts?
If the final goal is to monitor avalanches and investigate their flow properties, experts could do that better and faster than machines, given the relatively small rate of avalanches.
If the goal is real-time warning, do you think that there would be enough warning time (from detection until propagation to roads, villages ...) to allow mitigation actions?

2) Time window
Seismic signal is divided in windows of 10 s, much smaller than the typical duration of avalanche seismic signals.
I understand that a short window is preferable for early waring (but is this your goal and is it feasible?).
But for classification, I think a longer time window (about 60 s) would improve the results by allowing to better detect the characteristic spindle space of avalanche signals and the consecutive P, S and coda waves of earthquakes

3) Network based attribute
Why did you choose not to use network attributes, although you have a network of 5 sensors and network attributes were found to improve the results in the study of Provost et al. (2017)?
For instance, computing cross-correlation between sensors provides the peak correlation and time delays, which could be useful to distinguish avalanches from earthquakes or noise sources.
The correlation between sensors is much larger for distant sources (earthquakes, quarry blasts) than for nearby signals (avalanches and noise).
The time delay depends on apparent velocity, which is larger for deep and distant sources (earthquakes) than shallow and nearby source (avalanches, acoustic waves).

4) Infrasound
Each seismic sensors was colocated with an infrasound sensor. Why didn't you use infrasound data?
I guess that the amplitude of the infrasound signal (relative to the seismic signal) should be much higher for avalanches than for earthquakes?

Minor comments;
Figure 8: what are the main features shown in x axes? ("ES[3]" "DISTQ3Q1" "F21"...?).

Figures 2 and 12: Could you also show the spectrogram of the signal? Avalanches are often easier to identify by looking at the spectrograms than seismograms.

l349 "two-thirds of the avalanches were not verified". You mean, by cameras or radar? But they were verified by experts?

l424-425 What do you mean by " Thus, the models could, in turn, be considered to annotate large datasets, which in turn can be used to detect fine precursor signals."?

l462 "Also, these event types typically generate signals with a higher signal-to-noise ratio than avalanches".
I don't agree. All gravitational and tectonic processes generate a majority of small events (eg, Gutenberg Richter law for earthquakes). The minimum size is always the detection threshold.

l469 "In contrast to our approach, they only considered avalanches verified on camera images or manually 470 picked events."
I don't understand. I thought you also manually picked the avalanche seismic signals and verified part of them with radar and cameras?

Hide

RR by Tate Meehan (08 Jan 2025)

RR by Kevin Hank (13 Jan 2025)

RR by Anonymous Referee #6 (01 Feb 2025)

Suggestions for revision or reasons for rejection

The study at hand investigates the possibility of detecting avalanches in seismic sensor signals with the help of representation learning. It demonstrates that an auto-encoder can learn useful features from seismic sensor signals in an unsupervised fashion. The usefulness of those features becomes apparent when training RF classifiers on top of the autoencoder to address binary avalanche detection tasks. Those validation tasks have a high scientific significance and are relevant for operational services in the avalanche warning community.

Scientific significance: The study could contribute significantly to the current field of avalanche detection from seismic sensors. Previous work is limited to random forest approaches and manually engineered features. The authors' approach to learning a representation of the avalanche features and training a classifier on top is a well-respected method in machine learning. While the benefits of this method have not been very well presented in the paper, I acknowledge and recognize the high scientific significance of this work, as well as its novelty.

Scientific quality: The scientific approach is sound and convincing. However, the usefulness of the novel methodological approach is not evident, given that related work outperforms this approach. Potential advantages of the novel method are claimed, but they are neither backed by literature nor experiments. Moreover, the experimental setup must be revised to address relevant concerns regarding the data processing (concerning data split and data normalization).

Scientific reproducibility: The ML part of the work is fully reproducible. The data and the code are accessible and executable. The experimental setup to collect the data is made available and could be reproduced for different site locations. It is unclear how to reproduce the processed data from the raw data.

Presentation quality: The presentation quality could be improved significantly. All requests on that end are "minor," i.e., they involve mainly restructuring bits and pieces, including literature to back up claims, building a more substantial narrative throughout the paper, and editing the paper to respect "context-content-conclusion." Even though my vote is low in that category, I am convinced that the authors can address all those concerns.

Detailed feedback can be found in the referee report.

I am listing here the major revisions requested that I consider the most relevant (emboldened in the feedback document):

Presentation:
- Research questions and contributions should be described more explicitly.
- Improve the narrative/storyline of the paper.
- Structure the paragraphs with "context – content – conclusion" format.
- Use names/identifiers (for models) and technical terms consistently throughout the paper.
- Use booktabs to format the tables.
- Improve figures by adding more descriptive subcaptions. Each figure should bring one main point across. Make the content in the figures more accessible (see comments).
- Discussion: Back up some of the claims on the ML side with literature.

Methods:
- Data split: How is data leakage via correlated samples prevented? Are you employing a stratified split? Are you separating sensors/locations?
- Normalization of the data: This is a major revision that is necessary. See the referee report for reasoning.
- Balancing of data: Address this in the paper (explain what you do to address the imbalance).
- Learning curves: Include them in the appendix so the reader can see whether the models are learning well and whether they underfit/overfit.
- More runs: Run the models at least 3 times with different random seeds. If you have already done that, indicate this in the paper and add error bars where applicable. This is necessary to compare the different models. The results are so close that the differences might not be significant if re-run.

Code and Data:
- Make sure to store the data on zenodo separately from the code.
- Minor revisions on the code; see referee report.

I recommend the paper for publication after major revisions. I consider the work done here an essential contribution to the field - a step towards automating avalanche detection and warning. Using representation learning in avalanche detection is a major novelty, and the work presented here has the potential (after revisions) to be an outstanding example of using representation learning for a relevant task.

Referee Report: PDF

Hide

ED: Reconsider after major revisions (06 Feb 2025) by Le Yu

AR by Andri Simeon on behalf of the Authors (19 Mar 2025) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (27 May 2025) by Le Yu

RR by Anonymous Referee #3 (04 Jun 2025)

RR by Anonymous Referee #6 (11 Jun 2025)

RR by Tate Meehan (02 Jul 2025)

ED: Publish subject to minor revisions (review by editor) (19 Aug 2025) by Le Yu

AR by Andri Simeon on behalf of the Authors (25 Aug 2025) Author's response Author's tracked changes Manuscript

ED: Publish as is (09 Oct 2025) by Le Yu

AR by Andri Simeon on behalf of the Authors (14 Oct 2025)

Short summary

Avalanche detection systems are crucial for forecasting, but distinguishing avalanches from other seismic sources remains a challenge. We propose novel autoencoder models to automatically extract features and compare them with engineered seismic features. These features are then used to classify avalanches and noise events. The autoencoder feature classifiers exhibit the highest sensitivity in detecting avalanches, while the engineered seismic classifier performs better overall.