Assessing the performance and explainability of an avalanche danger forecast model

Abstract. During winter, public avalanche forecasts provide crucial information for professional decision-makers as well as recreational backcountry users. While avalanche forecasting has traditionally relied exclusively on human expertise, avalanche warning services increasingly integrate data-driven models to support the forecasting process. This study assesses a random-forest classifier trained with weather data and physical snow-cover simulations as input for predicting dry-snow avalanche danger levels during the initial live testing in the winter season of 2020–2021 in Switzerland. The model achieved ∼ 70 % agreement with published danger levels, performing equally well in nowcast and forecast mode. Using model-predicted probabilities, continuous expected danger values were computed, showing a high correlation with the sub-levels as published in the Swiss forecast. The model effectively captured temporal dynamics and variations across different slope aspects and elevations but showed lower performance during periods with persistent weak layers in the snowpack. SHapley Additive exPlanations (SHAP) were employed to make the model's decision process more transparent, reducing its “black-box” nature. Beyond increasing the explainability of model predictions, the model encapsulates 20 years of forecasters' experience in aligning weather and snowpack conditions with danger levels. Therefore, the presented approach and visualization could also be employed as a training tool for new forecasters, highlighting relevant parameters and thresholds. In summary, machine-learning models like the danger-level model, often considered black-box models, can provide reliable, high-resolution, and comparably transparent “second opinions” that complement human forecasters' danger assessments.