the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 30 Jul 2026
Forecasting different types of avalanches based on snowpack and snowfall/snowmelt conditions on the Southeastern Tibetan Plateau
Abstract. Reliable avalanche forecasting is essential for protecting mountain communities and transportation corridors, but estimating avalanche occurrence from monitored meteorological and snowpack conditions remains difficult. Operational assessments often rely primarily on meteorological thresholds, although avalanche responses depend on both type-specific triggering process and the snowpack conditions. Using meteorological, pre-event snowpack, and avalanche observations collected during the 2024 and 2025 snow seasons, we analysed 37 recorded avalanche events, together with corresponding non-avalanche periods. Separate logistic-regression models were developed for dry- and wet-snow avalanches using the intensity and duration of the preceding snowfall or snowmelt process and background snow depth. Their performance was compared with otherwise identical models excluding snow depth. Under leave-one-out cross-validation, including background snow depth increased the area under the receiver operating characteristic curve from 0.72 to 0.83 for dry-snow avalanches and from 0.85 to 0.94 for wet-snow avalanches. For wet-snow avalanches, the true-positive rate increased from 0.79 to 0.92, while the false-positive rate decreased from 0.21 to 0.12. These results demonstrate that pre-event snowpack conditions provides predictive information beyond meteorological forcing alone and improves avalanche forecasting, particularly for wet-snow avalanches.
- Preprint
(2100 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-4540', Jürg Schweizer, 24 Aug 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-4540/egusphere-2026-4540-RC1-supplement.pdfCitation: https://doi.org/
10.5194/egusphere-2026-4540-RC1 -
RC2: 'Comment on egusphere-2026-4540', Anonymous Referee #2, 02 Sep 2026
General comments
This manuscript presents meteorological, snow-depth, and avalanche observations from the Galongla Valley, a region where systematic avalanche data remain relatively scarce. The distinction between dry- and wet-snow avalanches is appropriate, and the study could be useful for local avalanche risk management.
The results suggest that snow depth prior to a snowfall or snowmelt episode contains useful information. However, some aspects of the sampling design and validation currently limit the interpretation of the reported probabilities and forecasting performance.
The manuscript would also benefit from a clearer statement of its original contribution to avalanche forecasting. Moreover, the observations cover only two snow seasons. This relatively short period may limit the generalisability of the conclusions, as the results may partly reflect snowpack characteristics specific to those seasons. This issue is particularly relevant because snow depth is the only snowpack variable considered in the analysis.
Major comments
1. Sample construction and interpretation of probabilities
Section 2.4 is central to the study but remains difficult to reconstruct. Please clarify how avalanches were assigned to monitoring stations, how positive and negative process blocks were identified, whether several samples could belong to the same meteorological episode, and whether the same process could be selected more than once. A schematic time series showing the positive block, excluded periods, buffer, and retained negative blocks would be very helpful.
The datasets contain three selected negative samples per positive sample. Avalanche prevalence is therefore determined by the sampling design and does not represent prevalence among all snowfall or snowmelt processes.
The authors should define the intended prediction unit and explain the consequences of this sampling design. If possible, the model should be calibrated using all eligible processes or corrected for the sampling fractions. Otherwise, the outputs should be described as relative classification scores rather than operational occurrence probabilities.
2. Interpretation of the forcing-duration coefficients
The negative coefficients for SFD and SMD carry considerable interpretive weight. Section 4.1 discusses them at length in terms of settlement, sintering, viscous creep, preferential flow, and restabilisation, while l. 340 presents the role of forcing duration as a relatively unexplored aspect of avalanche initiation. I have three concerns about this interpretation.
First, intensity and duration are mathematically related in the present framework: SFI and SMI are defined as cumulative amount divided by the corresponding duration (l. 117–118 and l. 134–136). Consequently, interpreting the fitted coefficients as fully separate effects of forcing rate and duration is not straightforward. The `ln(1 + x)` transformation further means that there is no simple exact equivalence between the current formulation and a model expressed in terms of cumulative amount and rate. Nevertheless, the interpretation of the duration coefficient may depend on the chosen parameterisation. It would therefore be helpful if the authors could examine this relationship and compare the present model with an alternative formulation using cumulative amount and rate as predictors. This would clarify whether the negative duration effect is robust or partly dependent on how the predictors are defined.
Second, the definitions at l. 114–115 and l. 132–133 do not appear to impose a maximum process duration, whereas the non-event durations reported in the deposited dataset span a comparatively narrow range. Could the authors clarify whether a fixed window or maximum duration was applied when extracting candidate non-event blocks? If event and non-event durations were determined using different observation rules, part of the fitted duration effect could reflect the sampling procedure rather than snowpack behaviour.
Third, the statement at l. 340 that the role of forcing duration has rarely been examined would benefit from qualification. The relationship between loading rate and strength development in buried weak layers is discussed, for example, by Schweizer et al. (2003a) and within the stability-index approach of Conway and Wilbour (1999). I suggest placing the present results more explicitly within this literature and clarifying how the treatment of duration in this study differs from previous approaches.
Finally, the negative association between forcing duration and avalanche occurrence may be specific to the maritime snow climate and snowpack conditions represented in this study. It may not necessarily apply to snowpacks containing persistent weak layers, where cumulative loading over longer periods may play a different role. Marienthal et al. (2015), already cited in the manuscript, report an association between deep-slab avalanche days and precipitation accumulated over the preceding seven days. This provides a useful contrast and could help define the limits within which the present results may be interpreted. Because the response variable is binary and contains no information on avalanche size, the conclusions should remain limited to occurrence probability and should not be extended to event magnitude.
3. Validation and uncertainty
Leave-one-observation-out cross-validation treats each row as independent. However, samples may be grouped by station, season, avalanche path, or meteorological episode. Closely related observations could therefore occur in both the training and test sets, potentially leading to optimistic performance estimates.
I recognise that the limited number of avalanches and the availability of only two snow seasons make fully independent validation difficult. This limitation should be stated clearly, and the results should be described as internal discrimination rather than generalisable forecasting performance. If feasible, an exploratory comparison in which one winter is used for training and the other for testing could provide some indication of temporal transferability, although the outcome would need to be interpreted cautiously because two seasons are insufficient for a robust validation.
The dry-snow model contains only 13 positive events and three predictors. Confidence intervals should therefore be reported for the AUC and threshold-dependent metrics, as well as for differences between the models. Confusion-matrix counts would also help readers interpret the reported rates.
4. Interpretation and operational scope of snow depth
I support replacing “background snow depth” with “pre-event snow depth” or “prior snow depth.” Snow depth is an observable state variable, but it does not represent snowpack conditions in the broader sense because it contains no information on stratigraphy, weak layers, liquid-water content, or stability.
Its association with recorded avalanches may also reflect avalanche size, terrain roughness, the ability of an avalanche to reach the monitored corridor, season, elevation, or detection conditions. These alternative interpretations should be discussed, and the mechanical interpretation in Sect. 4.1 should be moderated.
The study is retrospective and internally validated. The manuscript states that predictors for positive samples are calculated up to the avalanche occurrence time. Please clarify how the model would be applied prospectively to an ongoing snowfall or snowmelt process, including how the observation window would be defined before the outcome is known. Until prospective validation is available, terms such as “candidate forecasting model” or “retrospective classification model” may be more appropriate than claims of demonstrated operational forecasting performance.
Minor comments
1. Please provide concise summary statistics and distributions of the predictors for events and non-events, together with their units and correlations.
2. Some events appear to have zero melt intensity. Please clarify how a zero melt intensity can occur in samples defined as snowmelt events.
3. Please provide additional information in the caption of Fig. 6 and explain the meaning of the red curve.
4. Statements that snow depth has the “largest coefficient magnitude” are not directly meaningful when predictors have different units and scales. Standardised effects or odds ratios for meaningful increments would be more informative.
5. In the abstract, “pre-event snowpack conditions provides” should read “pre-event snowpack conditions provide.” More generally, “pre-event snow depth” should replace “pre-event snowpack conditions” wherever SD is the only snow-related predictor.
Recommendation
The study is based on a valuable observational effort in an underrepresented mountain region. I would encourage the authors to revise the manuscript, with particular attention to the construction of positive and negative samples, the interpretation of probabilities under the sampling design, the interpretation of the forcing-duration coefficients, the limitations associated with the small dataset and short observation period, and the operational interpretation of snow depth. Addressing these points would substantially strengthen the scientific interpretation and potential practical relevance of the work.
References
Conway, H. and Wilbour, C.: Evolution of snow slope stability during storms, Cold Reg. Sci. Technol., 30, 67–77, https://doi.org/10.1016/S0165-232X(99)00009-9, 1999.
Citation: https://doi.org/10.5194/egusphere-2026-4540-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 160 | 120 | 15 | 295 | 20 | 19 |
- HTML: 160
- PDF: 120
- XML: 15
- Total: 295
- BibTeX: 20
- EndNote: 19
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
