From the Swiss Alps to the Pyrenees: Evaluating the transferability of machine learning models for avalanche forecasting

Abstract

Avalanche forecasting is crucial for ensuring safety and mobility in snow-covered mountainous regions worldwide. While traditional avalanche forecasts relied solely on human expertise, machine learning models have been developed in recent years and are already operational in Switzerland. Such models can rapidly assimilate large amounts of complex data, providing objective and high-resolution predictions of the avalanche danger level and avalanche activity, assisting forecasters with a “second opinion”. These Random Forest models were originally trained on a historical database containing 20 years of Swiss meteorological measurements, snow cover simulations, danger level assessments, and avalanche observations. The performance of the models was comparable to the experience-based avalanche forecasts in the Swiss Alps, although some variations were observed among different regions. For the future operational extension of the models, there is the challenge of addressing the spatial variation of predictive performance across snow climates, and hence verifying the capability to transfer such models to other mountain regions outside of Switzerland. In this study, we test for the first time the avalanche danger and wet-snow avalanche activity models using a database recorded during three winter seasons in Val d’Aran, a small county on the Atlantic watershed of the Spanish Central Pyrenees. The Aran Avalanche Center issues a daily regional avalanche forecast to inform the public about the avalanche hazard. To prepare this bulletin, forecasters analyze extensive data sources, including weather forecasts and meteorological and avalanche activity observations. Additionally, during the last winter seasons, the SNOWPACK model has provided snow cover simulations for specific orientations (flat field and four mountain slope aspects) driven by meteorological data measured from two automatic weather stations. These data are used as input to run the danger level and wet-snow avalanche activity models, using the same model versions as those operationally used in Switzerland. Even though Val d’Aran is a relatively small mountain region, the high-quality data collected there, including daily avalanche danger level forecasts, snow cover simulations, snowpack stability field assessments, and avalanche activity observations, make it a perfect study site to test and validate the predictions of the Swiss machine learning models in a region with different snow climate conditions. Overall, the danger level models achieved agreement rates of approximately 60%–70% with the public forecast danger levels for dry-snow conditions, whereas the wet-avalanche activity model showed a moderate correlation with the forecast for wet-snow conditions. Our findings demonstrate that machine learning models trained on a dataset from Switzerland can be transferred to another climatic region with only a moderate loss in performance. Therefore, after real-time testing, these models could potentially be integrated into avalanche forecasting, as they are in Switzerland, to support avalanche forecasters in their assessments.