Automated crevasse mapping for Alpine glaciers: A multitask deep neural network approach

Abstract

Glacier crevasses are fractures in ice that form as a result of tension. Information on the location of crevasses is important for mountaineers and field researchers to plan a safe traverse over a glacier. Today, Alpine glaciers change faster than cartography can keep up with up-to-date manually created maps on crevasse zones. For the first time, this study presents an approach for automated crevasse mapping from high-resolution airborne remote sensing imagery based on a multitask deep neural network. The model was trained and evaluated over seven training and six test areas located in the Oetztal and Stubai Alps. By simultaneously preforming edge detection and segmentation tasks, the multitask model was able to robustly detect glacier crevasses of different shapes within different illumination conditions with a balanced accuracy of 86.2 %. Our approach is applicable to large-scale applications as demonstrated by creating high-resolution crevasse maps for the entire Oetztal and Stubai Alps for the years 2019/2020. Spatial and temporal transferability was proven by creating high-quality crevasse maps for all glaciers surrounding Großglockner, Piz Palü, and Ortler. The here presented datasets can be integrated into hiking maps and digital cartography tools to provide mountaineers and field researcher with up-to-date crevasse information but also inform modelers on the distribution of stress within a glacier.