Large-Scale Detection of Alpine Glacier Crevasses Using Remote Sensing and Deep Learning

Kurzfassung

Knowledge of crevasse locations is essential for improving our understanding of glacier dynamics. In addition, accurate information on crevasses is crucial for mountaineering safety and enables reliable route planning. However, large-scale monitoring remains difficult because crevasses vary greatly in appearance and often show low contrast on snow-covered surfaces, limiting the effectiveness of traditional detection methods. Here, we present an automated crevasse-detection approach based on a multi-task neural network trained on high-resolution orthophotos of Austrian alpine glaciers. The model was developed using 20 cm aerial imagery of glaciers in the Ötztal and Stubai Alps. Through systematic training and validation, the network achieved 86% detection accuracy across independent test sites, demonstrating robust performance in diverse glaciological settings. The multi-task architecture enables simultaneous feature extraction and classification, efficiently handling the complex spectral and textural characteristics of crevassed ice surfaces. Following successful validation, we applied the method to all glaciated areas in Austria, producing a comprehensive, high-resolution dataset of crevasse locations. We analysed the spatial distribution of crevasses on Austrian glaciers in different mountain regions, including Hohe Tauern, Dachstein, Ötztal, Stubai, Silvretta and Zillertal. Analysis of this dataset reveals spatial patterns of crevasse distribution and quantitative metrics on variations in crevasse density across slope, elevation, velocity, curvature and aspect. The crevasse location dataset provides glacier modelers with detailed boundary conditions for ice dynamics simulations and helps mountaineers plan safe routes. This dataset has already been incorporated into recently published hiking maps by the Austrian Alpine Club and demonstrates how machine learning and open data initiatives can bridge glaciological research and practical applications.