the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 28 Apr 2026
Self-Supervised Contrastive Learning in the Context of Volcano-Seismic Datasets
Abstract. Volcano-seismic datasets are expensive to label due to the requirement for expertise to understand the signals and the time-intensive nature of extracting and labeling different events that are occurring. This work evaluates whether self supervised methods can enable volcanologists to gain knowledge about the content of volcanic datasets without the use of labels, or reduce the amount of labels required. The aim of this work is to compare several common techniques and illustrate their usefulness for the volcanic community, where labeled data is an even more precious commodity than the wider seismic community. Experiments have been performed on three real-world datasets containing isolated volcano-seismic datasets from Llaima volcano, Colima volcano, and Mount Etna. Time-Series Representation Learning via Temporal and Contextual Contrasting (TS-TCC) shows particularly high performance in this task for finding structures in an self-supervised fashion. This indicates the untapped potential of self-supervised training to aid in different data analysis tasks within the volcano-seismology community.
- Preprint
(2292 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-1201', Anonymous Referee #1, 10 Jul 2026
-
RC2: 'Comment on egusphere-2026-1201', Anonymous Referee #2, 17 Aug 2026
This is a useful paper on an important problem. The comparison of SimCLR and TS-TCC on the Llaima, Colima and Etna datasets is relevant. The Llaima result is encouraging, and the Etna example is probably the most useful part of the paper because it is closer to a real monitoring catalogue.
I do not think the paper is ready in its present form. The main issue is that the study is presented as unsupervised, but labels appear to be used at several stages when the pipeline is chosen. The evaluation is also too limited, the baseline is weak, and the data and code statement needs to be replaced. These points need substantial revision.
Main comments
- The current pipeline is not fully unsupervised
This is the main issue for me. The encoder is selected using supervised accuracy on Llaima. The manuscript also says that contrastive parameters are selected partly from supervised classification, UMAP is hand-tuned to obtain a discriminative representation, and the number of clusters is fixed from the known number of classes. It is also not clear whether the best clustering algorithm was chosen after looking at the test labels.
Please show clearly where labels are used: architecture choice, SSL tuning, UMAP settings, cluster number, clustering method and final evaluation. Ideally, rerun the main experiment with those choices fixed without labels. Otherwise, describe the work as label-informed or semi-supervised rather than fully unsupervised.
- The evaluation needs more information
Please add Adjusted Rand Index and Adjusted Mutual Information. Colima is quite imbalanced, so V-measure alone is not enough. A class-by-cluster contingency table would also show whether the improvement is shared across event types or is mainly coming from the large classes.
Tables 2 and 3 should include all combinations of representation and clustering method, not only the best one. Repeat the runs with several random seeds and report the mean and spread. Give the UMAP seed, initialisation, metric and software version. It would also help to cluster in the encoder space, or after PCA, so the reader can see how much of the result depends on UMAP.
For Etna, add at least internal cluster scores and a stability check. The visual result is interesting, but the current interpretation is entirely qualitative.
- The augmentation choice needs a test
Window warping changes the local time scale and therefore changes frequency content. Frequency is also used to separate and interpret the event groups in this paper. I am not convinced that the representation should be invariant to that change without further evidence.
A small ablation would be enough: jitter only, jitter plus amplitude scaling, and jitter plus window warping, using the same training budget and more than one seed. Please also explain why each transformation is a plausible nuisance change for these signals and not a possible change of event class.
- The baseline is too weak
UMAP applied to the prepared waveform is a useful lower bound, but it is not a strong baseline for learned feature extraction. Please add at least one conventional feature set, such as spectral or cepstral descriptors, and one learned non-contrastive baseline such as a convolutional autoencoder. Scattering features would also be relevant.
If the paper keeps a broad claim about TS-TCC, a more recent time-series SSL method would be useful as well. If that is outside the scope, narrow the claim to a comparison of the two selected methods.
- The novelty claim should be toned down
The paper should not say that it introduces contrastive learning to volcano-seismic analysis. The contribution is a comparative application of SimCLR and TS-TCC to three volcano-seismic datasets, including an automatically picked Etna catalogue. That is still a useful contribution.
The related-work section also needs to be expanded. Please discuss the close seismic SSL work, including Yang et al., SeisCLIP and SeisLM, and the more recent time-series methods TS2Vec, CA-TCC and TF-C. Unsupervised seismic clustering and cross-volcano transfer work, including the studies by Ardid and colleagues, should also be placed in context.
- The work must be reproducible
The current data and code statements amount to asking readers to contact the corresponding author. That is not enough for a benchmark paper. Please release the code, configuration files, splits, seeds and package versions in a public versioned repository.
The Llaima dataset DOI should be given directly. For Colima and Etna, release the processed event sets if possible. If the waveforms cannot be redistributed, provide the station and channel information, exact time windows, pick lists, filtering steps and scripts needed to recreate the datasets.
- The Etna section needs a few basic details
Please report the number of detections before filtering, the number removed by each filter, and the final number of events in each split. Explain how the spectra and envelopes in Figure 11 were calculated. Cluster 2 contains only 99 samples, so it needs some discussion: is it a stable rare family, an outlier group, or an artefact of the chosen settings?
The paper also uses the word noise for both waveform noise and mistakes from automatic picking. Please keep those two things separate.
Minor comments
- The raw citation keys "smith1987iris, suarez2008international" are still in the introduction and need to be fixed.
- The manuscript needs a careful language check. Examples include "volcano-seimsic", "eachother", "disciminating", "reproducable", "expectd", "window wawrping" and "generealising".
- Use one form consistently: SimCLR, K-means, HDBSCAN and self-supervised when used as an adjective.
- Spell out the LAV station name and provide network, station, location and channel codes for all datasets.
- State the value of beta used in the V-measure equation.
- Give the train, validation and test sizes for Llaima and Colima, and all final split sizes for Etna.
- Event-wise maximum-amplitude normalisation removes absolute amplitude information. Please discuss what that means for monitoring applications.
- Check the UMAP figures for print and colour-vision accessibility. Marker shapes would help.
- The conclusion should be more measured. The Colima score and the qualitative Etna result are promising, but they do not yet show generally high performance across noisy volcano-seismic datasets.
- The author-contribution statement is too broad. Please state who prepared the data, implemented the methods, designed the experiments and interpreted the results.
Citation: https://doi.org/10.5194/egusphere-2026-1201-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 323 | 243 | 25 | 591 | 23 | 25 |
- HTML: 323
- PDF: 243
- XML: 25
- Total: 591
- BibTeX: 23
- EndNote: 25
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript presents an interesting and potentially high-impact study on the application of Self-Supervised Contrastive Learning to volcano-seismic datasets. The work addresses a relevant issue—the cost and time-intensive nature of labeling volcanic datasets—and I believe it is valid and merits publication in Natural Hazards and Earth System Sciences (NHESS).
However, before the manuscript can be accepted, it requires some revisions to improve its overall completeness, the fluency of the writing, and the factual contextualization of the work. Please address the following main points:
1. Introduction, Problem Framing, and the "Lack of Labels" Claim
The introduction currently feels a bit sparse, especially regarding the description of the seismic signals and the presentation of the methods. Furthermore, the authors state that there is a lack of available labels for training supervised machine learning models on volcano-seismic tasks. Is this absolute statement entirely accurate? While creating specialized ML datasets is indeed time-consuming, volcanic observatories worldwide have maintained extensive, labeled seismic catalogs for decades. The authors should nuance this statement, perhaps clarifying the distinction between the availability of raw observatory catalogs and curated ML-ready datasets. Additionally, it lacks clarity regarding the "practical monitoring objectives" this specific method aims to achieve. The authors should better clarify how their approach stands out from the state of the art and how it practically aids monitoring compared to existing methods.
2. Dataset Justification and Ground Truth (Mt. Etna)
Regarding the Mount Etna dataset, the authors claim that there is no ground truth public catalog available and that the dataset was analyzed in an exploratory manner for this reason. This assertion is highly questionable. Mount Etna is one of the most heavily instrumented and closely monitored volcanoes in the world, managed by INGV, which undoubtedly maintains detailed and robust operational catalogs. Is it truly possible that no certain labels exist for Etna? If the authors mean that a specific, pre-formatted public machine-learning dataset was not readily accessible for their study period, they must state this explicitly. Claiming a general "lack of ground truth" for Etna undermines the credibility of the premise and overlooks the extensive monitoring efforts present at this volcano.
3. Fluency and Completeness of the Methods
The methodology section requires more thorough explanations and a more fluid narrative. Currently, the presentation of the techniques (such as SimCLR and TS-TCC) feels unbalanced. I recommend expanding this section to provide a broader, more balanced overview, properly justifying the choice of all the adopted strategies (rather than focusing heavily on TS-TCC at the expense of SimCLR) and ensuring the text flows more naturally for the reader.
4. Discussion and Conclusions
The Discussion and Conclusion sections need to be better argued and characterized. Currently, they appear slightly weak compared to the amount of work that was conducted. It is crucial that the conclusions refer more specifically and explicitly to the obtained results and the figures presented in the text, explaining in detail how the metrics and visible clusters support the final claims regarding the utility of SSL for the volcanic observatory community.
5. Improvement of Figures (Waveforms)
The images presented in the manuscript need to be revised and replaced. In particular, the figures displaying the waveforms (e.g., the time and frequency representations of the clusters) require significant graphical improvement to make them clearer and more readable. These figures are fundamental for visually supporting the effectiveness of the cluster separation proposed in the text.
I am confident that by addressing these points, the manuscript will reach the level of clarity and maturity required for publication in NHESS.
I have provided a few suggested references below. However, I highly recommend expanding the reference list further, particularly in the introduction, to better frame the study:
1. Abed, W., Zali, Z., Sciotto, M., et al. (2026). Hidden patterns in volcanic seismicity: deep learning insights from Mt. Etna’s 2020–2021 activity. *Scientific Reports*, 16, 6155. [https://doi.org/10.1038/s41598-026-36677-x](https://doi.org/10.1038/s41598-026-36677-x)
2. D’Auria, L., Koulakov, I., Prudencio, J., et al. (2022). Rapid magma ascent beneath La Palma revealed by seismic tomography. *Scientific Reports*, 12, 17654. [https://doi.org/10.1038/s41598-022-21818-9](https://doi.org/10.1038/s41598-022-21818-9)
3. Gammaldi, S., Cabrera-Pérez, I., Koulakov, I., D’Auria, L., Barberi, G., García-Hernández, R., et al. (2025). Seismic tomography of a newborn volcano. *Geophysical Research Letters*, 52, e2025GL114932. [https://doi.org/10.1029/2025GL114932](https://doi.org/10.1029/2025GL114932)
4. Gammaldi, S., Donne, D. D., Cantiello, P., et al. (2025). A near real-time framework for monitoring very-long-period signals at volcanoes. *Scientific Reports*, 15, 41626. [https://doi.org/10.1038/s41598-025-25636-7](https://doi.org/10.1038/s41598-025-25636-7)
5. Grimaldi, A., Amoroso, O., Scarpetta, S., et al. (2026). Single-station analysis of Campi Flegrei (Italy) seismic signals using multiscale entropy and unsupervised learning. *Scientific Reports*, 16, 7669. [https://doi.org/10.1038/s41598-026-38257-5](https://doi.org/10.1038/s41598-026-38257-5)
6. Tan, X., et al. (2025). A clearer view of the current phase of unrest at Campi Flegrei caldera. *Science*, 390, 70-75. [https://doi.org/10.1126/science.adw9038](https://www.google.com/search?q=https://doi.org/10.1126/science.adw9038)