General comments

This study examines future avalanche activity in the Swiss Alps with a modelling approach that combines downscaled climate projections, a detailed snow stratigraphy model, and machine learning models. The models predict the number of days with dry and wet snow avalanches under different climate scenarios at 7 high-elevation locations, finding an overall decrease in dry snow avalanches and an increase in wet snow avalanches. These trends are interpreted by showing how the snowpack structure will change at different elevations.

The overall quality of the manuscript is high, with a clear presentation of the methods and results and interesting discussion and conclusions. The study design combines several state-of-the-art methods to investigate a significant topic that fits well within the scope of The Cryosphere. I recommend publication after addressing the following comments.

Specific comments

• Climate warming versus climate change: From what I understand “climate warming” primarily refers to the temperature component of “climate change”, as opposed to the entire climate system (precipitation, feedback loops, etc.). Were the authors deliberate in how they used each term throughout the manuscript (e.g., “climate warming” in the title suggests the study focuses on the response to warming temperatures)?
• Explanation of climate models and data: The climate data used for the study should be explained in more detail. The SNOWPACK model is thoroughly explained, but there is comparatively little information about the EURO-CORDEX model and CH2018 datasets, especially for readers unfamiliar with these products. Please briefly introduce and define GCMs and RCMs and explain the differences between the three RCP scenarios. Listing the specific GCM/RCM models in Table B1 has minimal meaning without explaining what they are and how they differ. Also, specifically for this study, it would be valuable to comment on how well the CH2018 datasets resolve fluctuating weather systems over a season and whether they are appropriate for predicting realistic snowpack stratigraphies. Do they produce smoothed average values, or do they capture realistic storms interspersed with high-pressure weather? For example, warming-related intensification of heavy snowfall is mentioned in line 38, would this be reflected in the data? The importance of resolving extreme weather events is addressed in the discussion section, but commenting on these aspects earlier would help readers interpret the results.
• Quantile mapping method: Please explain the basic principles behind quantile mapping to provide an overview of its function and application. Was the goal bias correction, downscaling, elevation adjustment, etc., and why was it chosen over alternative statistical methods?
• Figures: The overall figures quality is high, but some figures are confusing because they mix multiple data types and axes into a single graphic. It would be worth examining whether any subfigures should be split or omitted. Examples where data/axes do not fit with the rest of the figure include the left column in Fig. 3, the right column in Fig. 4, the mix of subfigures in Fig. 5, and the validation column in Fig. C2. Also, several axes label relative differences as fractions, but the manuscript text discusses them in terms of percentages. Perhaps it would be easier to interpret if the axes were labelled with percentage values (i.e., -10% instead of -0.1)?
• Impact of validation findings: 3b suggests the model chain has a systematic bias towards underpredicting wet avalanche days at low elevations and overpredicting at high elevations. Are there any known reasons for this and could it be corrected? Furthermore, how does this impact the results? Fig. 3 and 4 show distinct trends in wet avalanche activity by station elevation, but are these results exaggerated due to this bias?
• Spatial/frequency distribution: The discussion acknowledges that the model only addressed one component of avalanche risk: the likelihood of triggering. While the importance of destructive size is discussed, the spatial/frequency distribution of avalanches is also a commonly recognized component. The European Avalanche Warning Services defines avalanche danger based on snowpack stability, frequency distribution, and avalanche size. Similarly, the conceptual model of avalanche hazard defines the "likelihood of avalanches" as a function of "sensitivity to triggers" and "spatial distribution." The spatial component of hazard should be acknowledged and discussed. For example, the results show some elevation trends that may relate to frequency distribution, but in general the machine learning models primarily predicted likelihood of triggering. Could the methods be adapted to further examine this, either with spatially distributed snowpack models (e.g., by aspect), or perhaps training machine learning models to predict distribution?
• References: The manuscript is well cited, but perhaps the reference list is a little long at 5 full pages. Also, I think Verfaillie et al. (2018) is very relevant to this study, and perhaps the recently published review by Eckert et al. (2024).

Technical comments

• Line 110: Appendix B is mentioned before Appendix A. Consider reordering.
• Line 178: How was three-day sum of new snow height calculated and was it done the same way for SNIOWPACK output and AWS measurements?
• Line 181-183: It seems odd to introduce a figure in an appendix before a figure in the main body. I would consider reordering (assuming Fig. 2 is more relevant than Fig. C1).
• Line 198: “and with AWS measurements” would better explain relative differences than “or with AWS measurements”.
• 3: Do you have any idea why the ensemble of climate models resulted in more spread for wet avalanche days than the dry avalanche days? Was the wet avalanche machine learning model more sensitive to a specific input?
• Line 265: An aside comment… Could you split the dry snow avalanche days by weak layer type to count the number of days with avalanches on persistent versus non-persistent grain types? What other information about the avalanches could be derived from snowpack models (avalanche problems, aspect-trends, etc.)?
• Line 297: “and despite” instead of “as despite”.

References

Eckert, N., Corona, C., Giacona, F. et al. Climate change impacts on snow avalanche activity and related risks. Nat. Rev. Earth Environ., 5, 369, https://doi.org/10.1038/s43017-024-00540-2, 2024.

Verfaillie, D., Lafaysse, M., Déqué, M., Eckert, N., Lejeune, Y., and Morin, S.: Multi-component ensembles of future meteorological and natural snow conditions for 1500 m altitude in the Chartreuse mountain range, Northern French Alps, The Cryosphere, 12, 1249–1271, https://doi.org/10.5194/tc-12-1249-2018, 2018.

General comments:

The manuscript assesses future projected changes in the amount of avalanche triggering at 7 high-elevation stations in the Swiss Alps, distinguishing between dry- and wet-snow avalanches. Output from an ensemble of 8 climate models were statistically downscaled and used to drive a detailed snow cover model to describe the snow stratigraphy, from which avalanche risk was evaluated based on a machine learning model. Results show an overall decrease in avalanche activity, with increases in wet-snow avalanche activity at high elevations. Snow stratigraphy simulations propose a mechanistic explanation for these changes that snow depth and snowfall declines alone cannot explain.

The overall quality of the study is good -- the authors communicate novel methods clearly for the most part; the language and graphics describing and interpreting results seem to flow well, with a thoughtful discussion; and the study's question addresses an impactful problem within the journal scope. I have few very minor comments/suggestions, following those of Referee #1, and ultimately recommend the manuscript to be published after addressing them.

Specific comments:

[Line 54 and 76] What constitutes as a “low-elevation” range, and what uncertainties does the “spatial statistical transfer” from lower- to higher-elevation locations introduce that the quantile mapping approach may not address?

[Line 92 on quantile mapping] I was similarly unsure of the purpose, inner-working, and justification of the multivariate quantile mapping approach. It would be helpful to provide a brief walkthrough and justification of the technique. How exactly are values transformed from CORDEX output to the daily values we use to force SNOWPACK? Does it preserve storm characteristics well enough to allow SNOWPACK to appropriately resolve avalanche-relevant stratigraphy?

[Line 296 on differences across studies] Could the authors comment what they deem the most important factor(s) is/are to reconcile reported differences in warming-induced avalanche trends?

Technical comments:

[Abstract] Consider providing corresponding numbers of avalanche days in the abstract to give context to the reported percentages.

[Line 187 / Figure C1] Please clarify to avoid misinterpreting what values are being compared here for validation. Does this mean, for instance that the 22 values (winters 2001 through 2022) were averaged for 1 Dec, 2 Dec, etc. in the model, and then compared these averages to the observed averages’ corresponding day of year? Or, was the entire cold-season averaged, and each season was compared?

[Figures 2, 5, and 6] Consider adding elevation values beneath station IDs to help readers interpret results without needing to refer to Fig 1 (or 3). Consider adding a horizontal line at 0 in Fig 2 to emphasize the model’s target value.

[Line 304] “Enhanced temperatures hinder the formation of weak layers by directly affecting the temperature gradient across the snowpack.” Could the authors briefly describe the mechanism for the change in the snowpack’s temperature gradient that hinders weak layer formation? Is the gradient itself stronger or more uniform, and why?