the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
A three-stage model pipeline predicting regional avalanche danger in Switzerland (RAvaFcast v1.0.0): a decision-support tool for operational avalanche forecasting
Abstract. Despite the increasing use of physical snow-cover simulations in regional avalanche forecasting, avalanche forecasting is still an expert-based decision-making process. However, recently, it has become possible to obtain fully automated avalanche danger level predictions with satisfying accuracy by combining physically-based snow-cover models with machine learning approaches. These predictions are made at the location of automated weather stations close to avalanche starting zones. To bridge the gap between these local predictions and fully data- and model-driven regional avalanche danger maps, we developed and evaluated a three-stage model pipeline (RAvaFcast v1.0.0), involving the steps classification, interpolation, and aggregation. More specifically, we evaluated the impact of various terrain features on the performance of a Gaussian process-based model for interpolation of local predictions to unobserved locations on a dense grid. Aggregating these predictions using an elevation-based strategy, we estimated the regional danger level and the corresponding elevation range for predefined warning regions, resulting in a forecast similar to the human-made avalanche forecast in Switzerland. The best-performing model matched the human-made forecasts with a mean day accuracy of approximately 66 % for the entire forecast domain, and 70 % specifically for the Alps. However, the performance depended strongly on the classifier's accuracy (i.e., a mean day accuracy of 68 %) and the density of local predictions available for the interpolation task. Despite these limitations, we believe that the proposed three-stage model pipeline has the potential to improve the interpretability of machine-made danger level predictions and has, thus, the potential to assist avalanche forecasters during forecast preparation, for instance, by being integrated in the forecast process in the form of an independent virtual forecaster.
- Preprint
(4782 KB) - Metadata XML
-
Supplement
(35632 KB) - BibTeX
- EndNote
Status: open (extended)
-
RC1: 'Comment on egusphere-2023-2948', Simon Horton, 02 Mar 2024
reply
General comments
This manuscript presents a model chain for producing regional-scale avalanche danger predictions in Switzerland. The key contribution extends a point-scale danger model (Pérez-Guillén et al., 2022) to a regional scale by interpolating across a continuous grid and aggregating within predefined regions. The interpolation and aggregation methods aim to capture relevant processes that influence avalanche danger while aligning with approaches used by human forecasters so that the model chain can be applied as a decision-support tool. The presentation of this model chain is an interesting topic that fits well within the scope of GMD.
The manuscript does an excellent job of communicating a complex topic with clarity and logical progression. It establishes clear objectives, employs sound methodological choices, and draws fair and relevant conclusions applicable to operational avalanche forecasting. I think a few details could be further clarified (explained below), but otherwise recommend the publication of this manuscript.
Specific comments
- Representativeness of the stations. Providing additional information about the stations and snowpack simulations would help readers understand how effectively the training data represents the variability of avalanche conditions within a region. While Pérez-Guillén et al. (2022) likely address some of these details, including more information would offer valuable insights. For example, the number of stations in the dataset, the nature of the simulations (flat field and/or virtual slopes), and whether wind transport was simulated. Without such details, it remains unclear how well the stations capture the full range of expected conditions within each region and how this might impact the resulting predictions. Can we expect this method to predict the most unstable slopes in a region and if not does this create a bias? How well can the interpolation routine capture snowpack conditions not represented in the input data?
- Terrain features. The selection of terrain features for the interpolation routine should be explained in more detail. It is not entirely clear which features are derived from the DSM, nor is the meaning of directional derivatives, difference of Gaussians, and Gaussian pyramids (lines 137 to 141). Some plain-language explanations of what these derived variables are and how they potentially relate to avalanche danger would help. Also, some clarification is needed regarding the interpretation of slope angle, curvature, and aspect at the coarse scale of 32 km², and why these are expected to be relevant. Further explaining the terrain variables would provide readers with important context for interpreting the results.
- The methods section (Sect. 4) has extensive use of mathematical symbols, some of which may be excessive and cause confusion rather than clarity. This is simply a personal preference, but I think it would be clearer to use more plain language and then use symbols strategically where it helps communicate mathematical relationships. Also please check all symbols are unique and defined (e.g., alpha is used differently in line 251 vs alpha in line 281, Ne in line 263 is not defined).
- I agree with the approach to evaluating performance with mean/median accuracy, however, am curious if there were any directional biases in the model in terms of over or under-predicting danger (e.g., for specific regions or danger levels). While this doesn’t need to be fully presented, it would be interesting to comment if this was investigated.
Technical comments
- Line 66: “built” not “build”.
- 1: The IMIS and ZERO-DL station networks are not defined/described anywhere in the manuscript.
- Line 109-11: Data extraction times are unclear. Public forecasts are valid until 17 LT, snow cover data is extracted at 12 LT, but then why is resampled meteorological data centered around 18 LT? Wouldn’t it make sense for all data to be extracted at a single time?
- Line 121: Perhaps state the total dataset size (e.g., number of station-day-danger points).
- Line 130-140: It is not clear how extracting terrain features at a scale of 1 to 32 km2 is capturing the smaller scale topographic properties you say influence avalanches at scales to tens to hundreds of metres. Did you derive slope angle, profile curvature, and aspect from the 25 m DSM and then upscale to coarser grids? Perhaps more details would clarify how terrain characteristics are being captured in the model.
- Line 140: A brief plain language description of the Gaussian pyramid technique would help.
- Fig 2. In the interpolation section, the “etc.” in terrain features is confusing as the methods only list location, elevation, slope angle, curvature, and aspect. Does “etc.” mean to capture the directional derivatives, DOG, and Gaussian pyramids?
- Sect 4.3: It is not clear that three distinct methods were tested (mean, top-alpha, bands). When reading it can be interpreted that top-alpha and band averaging are done in conjunction, rather than two distinct methods.
- Line 366-359: Perhaps I misunderstood the method, but I don’t see how the elevation bands overlap. I would have assumed when you increase the bandwidth you decrease the number of bands accordingly to avoid overlap. What is the motivation for allowing overlap?
- Line 356: Is the “mean method” defined or labelled anywhere? I think the meaning of this method is intuitive but slightly confusing if it is not explicitly defined/labelled anywhere.
Citation: https://doi.org/10.5194/egusphere-2023-2948-RC1
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
239 | 67 | 12 | 318 | 18 | 8 | 9 |
- HTML: 239
- PDF: 67
- XML: 12
- Total: 318
- Supplement: 18
- BibTeX: 8
- EndNote: 9
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1