Item: MACHINE LEARNING FOR AUTOMATED AVALANCHE TERRAIN EXPOSURE SCALE CLASSIFICATION – A COMPARISON OF TWO APPROACHES
-
-
Title: MACHINE LEARNING FOR AUTOMATED AVALANCHE TERRAIN EXPOSURE SCALE CLASSIFICATION – A COMPARISON OF TWO APPROACHES
Proceedings: Proceedings, International Snow Science Workshop, Whistler, BC, Canada, 2026
Authors:
- Kalin Markov [ National Institute of Geophysics, Geodesy and Geography, Bulgarian Academy of Science, Sofia, Bulgaria ] [ Bulgarian Avalanche Association, Sofia, Bulgaria ]
- Yves Bühler [ WSL Institute for Snow and Avalanche Research SLF, Davos Dorf, Switzerland ] [ Climate Change, Extremes and Natural Hazards in Alpine Regions Research Centre CERC, Davos, Switzerland ]
- Andreas Huber [ Austrian Research Centre for Forests (BFW), Department of Natural Hazards, Innsbruck, Tyrol, Austria ]
- Momchil Panayotov [ University of Forestry, Sofia, Bulgaria ] [ Bulgarian Extreme and FreeSkiing Association - BEFSA, Sofia, Bulgaria ] [ Bulgarian Avalanche Association, Sofia, Bulgaria ]
Date: 2026-09-28
Abstract: Backcountry travel in mountainous regions in Bulgaria has increased significantly in recent decades, heightening the need for effective avalanche risk management. The Avalanche Terrain Exposure Scale (ATES), first introduced in 2004 in Canada, classifies mountain terrain based on its potential exposure to avalanche hazards. Whereas early applications relied on expert-based manual mapping, subsequent studies introduced semi-automated and fully automated workflows. The most recent approaches introduce machine learning methods for automated terrain classification according to ATES criteria. In this study, we compared two machine learning approaches (ATES 1 and ATES 2) for ATES classification for the territory of Pirin National Park, Bulgaria, one of the most popular regions in the country for backcountry skiing. Both approaches used the same training data, a Random Forest model with the same architecture and hyperparameters, and shared input features of slope and percent canopy cover. They differed in the remaining input features used for training procedures and they also used different approaches for determining the potential avalanche release areas (PRA). In addition, in model ATES 1, the runout modelling was performed using AvaFrame/com4FlowPy, while in ATES 2 - from RAMMS::Avalanche runout modelling. The resulting ATES models were evaluated on a set of 36 separate validation polygons distributed throughout the territory of Pirin National Park. The validation polygons were labelled by three independent experts and located outside the valley where training occurred. Both models performed similarly well with respect to calculated performance metrics. In model ATES 1, the pixel-based accuracy was 87%, with 94% of the polygons classified correctly (when taking the majority classification for each polygon). For model ATES 2, both metrics were very similar, with the pixel-based accuracy being 86% and the polygon-based – 92%. The most classification errors occurred in the challenging terrain class for both models. The results demonstrate the ability of machine learning models to generalize well across entire mountain ranges and to adapt robustly to varying input feature sets. At the same time, they indicate that mapping the challenging ATES terrain class remains difficult for predictive models and is therefore associated with the most errors.
Object ID: ISSW2026_P1.33.pdf
DOI: https://doi.org/10.15788/1790098776
Language of Article: English
Presenter(s): Kalin Markov
Keywords: avalanches, ATES, machine learning, com4FlowPy, RAMMS, Bulgaria
Page Number(s): 1039 - 1048
-