General comments

This paper investigates the potential to forecast natural dry-snow avalanches with snowpack simulations. Avalanche observation data are used to train and validate models that predict the probability of natural avalanche occurrence, as well as the probability of different avalanche sizes. Comparing to benchmark models that only considered the amount of new snow, they show improved model performance by adding snowpack stratigraphy and stability information, especially in regions prone to persistent weak layers. As a final step, model predictions were validated with regional avalanche danger ratings, which illustrated the potential for this model to support avalanche forecasting.

The study is interesting, thoughtfully designed, and well written (especially considering the complexity of the subject and data sets). The research is relevant to the avalanche research and forecasting community and fits well within the scope of NHESS. I recommend publication after addressing the following comments.

Specific comments

My main comment is that it could be clearer how aspect dependent information was applied in the model training and validation. Overall, the data and methods are presented very clearly, but understanding how aspect information was applied required a fair bit of extra effort which I think could easily be improved by providing more details. Some examples:

• Consistent terminology. It would help to explicitly describe values as “aspect-specific” throughout the manuscript when referring to AvD/nAvDs. “aspect-specific” and “slope-specific” are used interchangeably (e.g., lines 255, 336).
• 3.1.1 could provide a clearer description of how avalanche days were counted by aspect in the training data. My interpretation is that for each day of the study period Y is calculated 4 x (number of stations) times and then each value of Y is matched to the explanatory variables for the corresponding virtual slope simulation to build the training dataset. Are the flat field simulations discarded?
• Line 150 states the height of new snow was independent of aspect, which I assume means these variables were taken from flat field simulations. But if this version of SNOWPACK models snow redistribution, then shouldn’t your profiles have different amounts of HN on different aspects and lead to a higher likelihood of natural avalanches on lee aspects?
• Can you comment on the impact of not considering aspect in the AvD definition used in the validation part of the study (i.e., line 253)? The model predicts aspect-specific probabilities, but these are evaluated on whether avalanches were observed anywhere in the region. The relevance of this assumption should be justified.
• Given SNOWPACK’s virtual slopes have limited verification studies, it could be interesting to briefly comment on whether this study finds they add value relative to flat field simulations only. Are there any insights or recommendations about virtual slopes to share with future researchers?

Technical comments

• Introduction: I appreciate how the introduction clearly addressed the limitations of using avalanche observations for model verification. The objective and structure of the paper is also very clear.
• Line 107: Can you state how many avalanches were included in the AV3 dataset (as done for AV1 and AV2)?
• Line 120: It may be better to mention wind transport was enabled here rather than line 152. Also, can you describe whether this SNOWPACK setup determines snowfall with precipitation gauges or snow height measurements (since new snow height is an important variable in this study)?
• 1: This figure (and the entire Data section) is organized in a way that makes it very easy to understand the different datasets used in various parts of the study. Thanks.
• 2: Can you specify a temporal period for number of avalanche days N(AvD). For example, over a specific study period, seasonal average, or something else?
• Lines 145-146: Some technical notes on choosing variable names and ICSSG standards. Typically, HN3D would suggest a 3-day observation interval, which is different from the sum of observations made at 1-day intervals (due to settlement). Am I correct that in this case, you are summing HN1D values rather than using SNOWPACK to directly get a height of 3-day snowfall? Also, would it be more accurate to call the precipitation particle variable a “thickness” rather than a “depth”? I would interpret depth as the distance from the deepest PP/DF layer to the surface, but summing thicknesses could be smaller if there are other grain types above (e.g., RG/MF). If that is the case, then ICSSG symbol for thickness is D rather than z. Similarly, the standard symbol for grain type is F. I don’t think changing these variables is essential, just something to consider.
• Line 148: Can you briefly describe the main inputs of sn38 and how these differ from the inputs to Punstable?
• 1: It could be clearer here how many avalanche days were computed (i.e., one per station per day?) and how the aspect and elevation information was used.
• Line 181: Variables st and asp are defined but not used in manuscript.
• 209: Variable thr is not defined and appears to only be used in the Appendix.
• Line 208: The subsets are not just based on splits of the AV1 data, but also the snowpack data (i.e., critical grain type). Perhaps it’s better to say “we split the training data…”.
• Line 220: Why was sn38 only used in the binary model and not the continuous model? It would help to list the x variables in the beginning of this subsection (since the variables used in the binary model aren’t described either).
• Line 228 and 243: The idea behind the BS+ metrics could be a bit clearer. Throughout the manuscript they are described as “minority class”, “positive observed outcomes”, “positive events”, “when condition is fulfilled”. I recommend a sentence to explain why these subsets are relevant in the methods and then choosing consistent terms that are more descriptive (e.g., only days with observed avalanches) to use throughout the manuscript.
• Line 252: Why were aspect and elevation neglected when determining AvD here?
• Line 284: According to Table A2 the median is 12 cm not 13 cm.
• Figure 4 is not cited anywhere in the manuscript.
• Fig. 5: The 2020 season seems to standout as anomalous in these figures, was there something unique about that season, such as the prevalence or absence of persistent weak layers?
• Line 316: Should this be dataset AV2 instead of AV3?
• 7: The HN3D model is presented before the Pcrit model in Fig 5, while it’s the other order in Figs 7 and 8. Consistent ordering would be ideal.
• Lines 371-380: This paragraph is a little confusing because not all the values discussed are shown in Fig. 9b, which appears to be due to whether the Pdeep < 0.77 cases are included or not. Which case is more relevant to present here? The text could be clearer about which case is being discussed and which case is shown in the figure.
• Line 395: I am particularly interested in how the extreme cases of widespread versus no activity impact the results at danger level 3. Since the models are fit to these extreme cases, they do not capture the “natural avalanches possible in certain areas” conditions experienced at danger level 3. However, it seems the models produce a desirable result with a wide range of P(AvD) observed at considerable danger in Fig. 8, which speaks to the range of conditions and uncertainty about natural avalanches experienced at considerable danger.
• Line 446: While not published in a peer-reviewed journal, Bellaire & Jamieson (2013) estimated avalanche size from simulated profiles in their 2013 ISSW paper (Fig. 2 is similar to your Fig. 6a).
• 10: This caption could provide higher-level description of what is shown, as it is difficult to understand without also reading the text. Also, what exactly do the contour lines show? Steps in density distributions?
• 11: Why are the pink bars for Rutschblock (T2) only shown for low danger and not others?
• Conclusion: Given the stated objective of the paper was to “investigate whether the instability model developed by Mayer et al. (2022) applied to one dimensional SNOWPACK simulations can be used to predict natural dry-snow avalanches”, I think this question can be more directly answered somewhere in the conclusions to summarize what was learned and how others could apply the model.

References

• Bellaire, Sascha, and Bruce Jamieson. "On estimating avalanche danger from simulated snow profiles." In Proceedings of the International Snow Science Workshop, Grenoble–Chamonix Mont-Blanc, pp. 7-11. 2013. https://arc.lib.montana.edu/snow-science/item.php?id=1740

The authors use simulated snowprofiles from SNOWPACK around AWS as both avalanche day predictor and avalanche size predictor for natural dry avalanches. They validate these indicators using a long-term dataset of re-analyzed danger levels and a short-term avalanche observation dataset. They further show how their predictors improve on simple avalanche day indicators by adding information about stratigraphy and stability.

The study is very well written with good and clear figures and tables. Additional information is found in the appendix. I find the study interesting for a large part of the avalanche community both scientific and practical.

My main critique is that used datasets and methods were developed in other studies. It is very time-consuming to read all these associated studies to gain a better insight into the methods. This is especially true for the binary avalanche day classification.

The discussion does a good job at listing a handful of limitations of this study, however, falls short in explaining what impact these limitations have. I am pointing towards some of these limitations in my comments below.

20: please reconsider this sentence. You write that erroneous forecasts may cause costs as false alarms may lead to economic loss. Isn’t that the same thing?

23: I think accurate forecasting of natural avalanches in space will never be possible. You write about forecasting the location of avalanches also in the first sentence in the abstract which is a bit misleading.

Figure 1: great figure that is very helpful in understanding which datasets serve which purposes.

97: it seems like you are referring to Table 3 before you refer to Table 2.

130: what is the rational behind selecting the deepest WL as Pcrit? If I understand your method correctly, the DH layer at around 50 cm in Figure 3 should be Pcrit? Thank you for clarifying this!

148: I am not familiar with sn38. Could you explain very briefly where and how it is used?

150: What is the rational behind assuming that new snow height is not aspect dependent? You describe in the sentence below that there is a snow redistribution routine in SNOWPACK.

Section 3.1.1: I had to obviously read the Hendrick et al. (accepted) paper to better grasp your definition of avalanche days. I do not think that it is particularly favorable to the readability and comprehensibility of this paper, that one must read up on the methods in another paper. I, however, do think that the algorithm is clever. There are a couple of things that I was wondering about, that do not necessarily have to be answered in a revised manuscript:

• How do the different gap check requirements compare to the size of the forecasting regions (I know that you have dynamic regions) as well as to the typical size one of your observers can cover to do avalanche observations?
• We often say that avalanches are rare events and given that you have had a median of two avalanches per aspect and elevation in an area of 250 km2 around an AWS, I wonder if this is true?

209: I might have missed it, but what is “thr”? threshold?

Section 3.1.2: this section was very hard for me to comprehend and only after reading the results from 4.2 onwards, it became more clear what you were doing.  

Section 3.3: Is it a problem that there is a mismatch between the number of observed avalanches per aspect and elevation within 250 km2 of an AWS in the training dataset (N=2) and the number of observed avalanches regardless of aspect and elevation within 1000 and 5000 km2 which is a minimum of 1? In my mind you are getting more avalanche days due to a less stringent threshold in your validation dataset than in your training dataset.

255: Do I understand you correctly that you are using SNOWPACK simulations forced by Weissfluhjoch data for the entire validation dataset?

260: What is the rational behind removing simulated snow depth < 30 cm?

263: Do you mean that avalanche days were in general associated with new snow, both 24 and 72 hours?

284: 12 or 13 cm?

297: Interesting observation about persistent weak layers needing less new snow for natural triggering. I thought that there was not much difference between the strength of persistent and non-persistent weak layers during and immediately after snow fall. However, non-persistent forms sinter quicker than persistent ones (Alec’s lab study from 2013). And I believe that Ben Reuter and others showed in 2018 that non-persistent forms were initially as weak as persistent forms, however, gaining in strength quicker.

309: What is the physical explanation for taking the mean of both models?

317: A median failure depth of 30 cm for size 1 avalanches is surprisingly high in my mind. Where were you surprised about that result? I must confess that I am positively surprised that simulated weak layer depth is such a good predictor of avalanche size. I thought that discerning avalanche size is much more complex.

Figure 7: median values are not always readable.

Figure 8: the numbers indicating respective portions above and below threshold are not always readable.

395: With the target variable including either a lot or no avalanche activity, what would you expect your results to be if medium avalanche activity days were accounted for? How does the time stamp of 12:00 LT for your model simulations influence the results? Do you foresee some problems with regards to when observers record avalanche activity during the day? I am also convinced that “medium avalanche activity” might characterize many avalanche days with considerable avalanche danger (ref Fig 8).

469: …or get rid of the avalanche danger levels altogether (just my personal opinion and somewhat confirmed by Figure 10)

500: pretty interesting!

Summary

The authors introduce and investigate the predictive skills of two new models: an avalanche day predictor and an avalanche size estimator. Both are based on snow cover model results and observed avalanche activity and especially designed and valid for natural dry-snow avalanche events. The models are trained and tested on individual data sets. The training data consists of two large data sets of avalanche observations and snowpack simulations using the 1-D physics based model SNOWPACK. The validation data sets include avalanche observations, avalanche danger level assessments and snow cover simulation results.

The model for the avalanche day predictor focuses on a RandomForest (RF) model based on derivates of SNOWPACK variables presented by the authors team very recently (Mayer et al., 2022) and is trained using a 3-years data set of avalanche observations data covering the entire Swiss Alps. The estimator model is trained with observed data only, but includes a very large data set of avalanche observations covering 30 years for the Swiss Alps.

To test the performance of both models, the authors test their novel approaches against a very common benchmark model (Height of the three-day sum of new snow; hn3d) and validate their findings against a fully independent data set of avalanche observations. Large parts of the interpretation and discussion is done by comparing the results to a 21-years data set of regional avalanche danger levels assessment.  

Results show good predictive performance results for characterising days with natural dry-snow avalanche activity; especially when natural dry-snow avalanche activity was driven by shallow snowpacks consisting predominantly by persistent weak layers. Compared to the simple benchmark model, performance is very similar and increases slightly when the new approach is combined with the benchmark model. The results for the avalanche size estimator when tested again the independent data set are also encouraging. Finally, both approaches are compared to the regional avalanche danger levels.

Evaluation

The approaches are not fully novel but connect skilfully recent advances with a large data set that represents the currently available golden standard within the avalanche research community.  Research objectives are very clear and concise; methods are well designed, but also how the authors addressed the objectives are mostly well described and easy to grasp.

Language is concise and the manuscript is well written. Major findings are very relevant to the avalanche research and forecasting community. The content fits very nicely into NHESS. There are a few parts of the discussion and interpretation of results that need revision. Having addressed the below stated comments, I recommend publication.

General comments

I have the following general comments:

• In the Abstract(Line 1) and Introduction (Lines 19-21) you give the reader the feeling that you would like to tackle both, very local (path scale) and regional avalanche forecasting. When reading the full manuscript, it becomes obvious that you address regional avalanche forecasting (e.g., Section 3.1.1 or the fact that you compare and discuss results to the regional avalanche danger level). Please be more specific in that case and drop the connection to the local avalanche forecasting.
• The authors compare their avalanche day predictor model to the conventional natural stability index on a 38° steep slope (sn38) which was one of the few indices developed within SNOWPACK to better assess natural avalanche activity (Lehning et al., 2004). The new model outperforms the sn38 (Figure 4), but the authors do not really discuss why this is the case. They state that “The poor performance of sn38 is in line with other studies (Schweizer et al., 2006; Jamieson et al., 2007).” To my knowledge the first study tackles sk38 only, thus skier-triggered scenarios and not spontaneous avalanche activity. The second study compares the natural stability index (sn) based on measurements to natural avalanche activity in the surroundings of the study plot – which is a significant different approach to the one presented (modelled vs. measured). In fact – to my knowledge the only qualitative investigation on the performance of sn38 is given in Lehning et al. (2004). There the authors show reasonable results.
  
  So, it remains difficult to set the presented low performance skills into context. Therefore, it would be very interesting and valuable to tackle in more detail the question, why sn38 has such a low performance compared to the avalanche day predictor model. Both, the conventional and the novel approach, are heavily parameterised by snow density and almost rely on the same concepts: the most important variable for the instability model by Mayer et al. (2022), the viscous deformation rate, shares the identical input parameter as the natural stability index, namely natural shear strength – which in turn is parametrised using snow density. 4 out of the 5 most important features building the RF model rely on snow density. It would be very beneficial for the community if the authors could e.g. use the Weissfluhjoch data set to shed some light into this topic. I know that this represents large efforts, but I believe it would give even more impact to presented results.
• Interconnected to the comment above: How and why is the instability model suited to predict natural avalanche activity, even though it is heavily trained on data that mostly represents skier-triggered avalanche activity?
• The discussion regarding the comparison to the regional avalanche danger levels is nice but needs in a few points a much broader approach: The statement that danger level 3-Considerable needs sub-levels could also be reversed in the fact that the Swiss forecaster need to train themselves more in order to transfer the overlapping parts into the neighbouring classes instead of increasing the level of discretisation. Can you comment on that please. Figures 10-11 are very important but touched very shortly. I would appreciate more details here.

Specific and technical comments

• 1 (Lines 48-49): Why don’t you address all danger levels here? In fact, at danger level 3-Considerable the definition mentions: In certain situations, some large, and in isolated cases very large natural avalanches are possible.
• 1.1. Line 96: Counter for Table Numbering is not in sequential order. You mention Table 3 before you mention Table 2 in the text.
• 2 It would be very helpful to introduce a new habit when using SNOWPACK simulations, namely placing the INI-Files of the model runs into the Appendix.
• 1.1 Definition of avalanche days and non-avalanche days (Lines 197-198): Does that mean that your training data set has no AvD due to a size 4 avalanche?
• 2. Avalanche size estimator: Could you please specify in a little more detail, why you have chosen exactly this approach?
• 4.2 Line 376: Figure 9e does not exist.

Literature

Jamieson, J. B., Zeidler, A., and Brown, C.: Explanation and limitations of study plot stability indices for forecasting dry snow slab avalanches in surrounding terrain, Cold Reg. Sci. Technol., 50, 23–34, 2007.

Lehning, M., Fierz, C., Brown, R. L., and Jamieson, J. B.: Modeling instability for the snow cover model SNOWPACK, Ann. Glaciol., 38, 331–338, 2004.

Mayer, S., van Herwijnen, A., Techel, F., and Schweizer, J.: A random forest model to assess snow instability from simulated snow stratigraphy, The Cryosphere, 16, 4593–4615, https://doi.org/10.5194/tc-16-4593-2022, 2022.

Schweizer, J., Bellaire, S., Fierz, C., Lehning, M., and Pielmeier, C.: Evaluating and improving the stability predictions of the snow cover model SNOWPACK, Cold Reg. Sci. Technol., 46, 52–59, 2006.