Machine learning for automated avalanche terrain exposure scale (ATES) classification

Abstract. Avalanche risk management is essential for backcountry safety. The Avalanche Terrain Exposure Scale (ATES) classifies mountain terrain based on its potential exposure to avalanche hazards and offers assistance to backcountry users in their terrain assessment. Initially, ATES maps were generated manually, a costly and time-consuming process. Automated ATES model chains (AutoATES) have been developed to address these limitations, but existing approaches require careful parametrisation when applied to novel areas.

This study applies machine learning methods, specifically Random Forests, for automated ATES classification by replacing expert-driven AutoATES classification trees with a data-driven approach. Using a labelled training dataset from the Pirin Mountains, Bulgaria, we trained and evaluated three Random Forest models to assess their potential in classifying avalanche terrain. We analysed the influence of various input features, including slope, potential release areas, and percent canopy cover, on classification performance. Our results indicate that Random Forests offer a robust and scalable method for ATES mapping and that incorporating additional input features can improve classification performance. The accuracies for our Random Forest models on a held-out test set were 79.31 %, 82.32 %, and 80.42 %, demonstrating their potential for automated avalanche terrain classification and supporting safer backcountry decision-making.