International Snow Science Workshop Proceedings - Montana State University Library

Item: AVALANCHE FORECASTING IN THE PEJO SKI AREA: A DATA-DRIVEN APPROACH USING A SUPPORT VECTOR MACHINE

• • Title: AVALANCHE FORECASTING IN THE PEJO SKI AREA: A DATA-DRIVEN APPROACH USING A SUPPORT VECTOR MACHINE

    Proceedings: Proceedings, International Snow Science Workshop, Whistler, BC, Canada, 2026

    Authors:

    • Christian Brida  [ DICAM - Department of Civil, Environmental, and Mechanical Engineering, University of Trento, Trento (TN), Italy ]  [ Pejo Funivie, Pejo (TN), Italy ]
    • Carlo Bee  [ DICAM - Department of Civil, Environmental, and Mechanical Engineering, University of Trento, Trento (TN), Italy ]  [ ARPAV - Arabba Avalanche Center, Livinallongo del Col di Lana (BL), Italy ]
    • Giorgio Rosatti  [ DICAM - Department of Civil, Environmental, and Mechanical Engineering, University of Trento, Trento (TN), Italy ]

    Date: 2026-09-28

    Abstract: Snow avalanches pose a major hazard in alpine ski areas, where reliable local forecasts are essential for operational decisions such as ski slope opening, avalanche control, and staff deployment. However, this kind of forecasting still relies heavily on expert judgment, making the integration of heterogeneous nivo-meteorological observations challenging and potentially subjective. This study presents an interpretable machine learning framework for supporting local avalanche forecasting in the Pejo ski area (Italian Alps), using a 45-year dataset of weather, snow, and avalanche observations. We developed a Support Vector Machine (SVM) classifier to discriminate between days with and without avalanches. The model was trained using historical nivo-meteorological predictors and avalanche records and then independently validated. The optimised model achieved good predictive performance (F1 = 0.83; ROC–AUC = 0.88) in distinguishing avalanche from non-avalanche days. Model interpretability was addressed using SHAP (SHapley Additive exPlanations), which identified cumulative snowfall, snow depth, temperature trends, and seasonal snowpack evolution as the dominant drivers of avalanche occurrence, consistent with established avalanche processes. To support operational forecasting, the model was integrated into a prototype Telegram-based decision-support system that automatically retrieves real-time observations, generates daily avalanche forecasts, provides SHAP-based explanations for individual predictions, and includes an experimental scenario-analysis module to evaluate alternative meteorological conditions. Operational deployment showed good agreement with field observations. The proposed approach combines predictive performance, physical interpretability, and operational integration, providing a transparent decision-support tool that complements expert judgment in local avalanche forecasting.

    Object ID: ISSW2026_P1.3.pdf

    DOI: https://doi.org/10.15788/1790098764

    Language of Article: English

    Presenter(s): Christian Brida

    Keywords: Artificial Intelligence, Avalanche Forecasting, Decision Making, Quantitative Forecasting, Ski Resort Management, Local Forecasting

    Page Number(s): 63 - 70

  • Get Full Article

Back to Home page