Deep learning outperforms existing algorithms in glacier surface velocity estimation with high-resolution datathe example of Austerdalsbreen, Norway

Harald Zandler^1*Jakob Abermann¹Benjamin Aubrey Robson²Alexander Maschler³Thomas Scheiber³Jonathan Lee Carrivick⁴Jacob Clement Yde³

• ¹Department of Geography and Regional Science, University of Graz, Graz, Austria
• ²Department of Earth Science, University of Bergen, Bergen, Norway
• ³Department of Civil Engineering and Environmental Sciences, Western Norway University of Applied Sciences, Sogndal, Norway
• ⁴School of Geography and water@leeds, University of Leeds, Leeds, United Kingdom

Remote sensing is a key tool to derive glacier surface velocities but existing mapping methods, such as cross-correlation techniques, can fail where surface properties change temporally or where large velocity variations occur spatially. High-resolution datasets, such as UAV imagery, offer a promising solution to tackle these issues and to study small-scale glacier dynamics, but new workflows are required to handle such data. Therefore, we tested the potential of new deep learning-based imagematching algorithms for deriving glacier surface velocities across the ablation area of a glacier with strong spatial variability in surface velocities (<5 m/yr to >100 m/yr) and substantial changes in surface properties between image acquisitions. For a thorough comparison of state-of-the-art methods and sensors, we applied three different techniques (cross-correlation using geoCosiCorr3D, feature tracking with ORB using SeaIceDrift and the new deep learning-based method using ICEpy4D) and three different platforms (Sentinel-2, PlanetScope, UAVs) to estimate glacier surface velocities. Results showed lowest errors for velocities derived with the deep learning-based approach applied to UAV imagery (RMSE=2.17 m/yr, R²=0.99), followed by cross-correlation using Sentinel-2 images (RMSE=21.0 m/yr, R²=0.59) and the deep learning-based approach with PlanetScope data (RMSE=21.28 m/yr, R²=0.36). Cross-correlation with geoCosiCorr3D resulted in comparably high errors with the UAV dataset (RMSE=36.22 m/yr, R²=0.24), whereas ORB-based feature tacking showed lowest performance with all sensors. Spatial patterns of computed velocities indicate that applying existing cross-correlation methods for areas with regular displacements or low glacier velocities yields suitable results on UAV data, but innovative deep learning-based approaches are required for resolving rapid changes in velocities or in surface properties. This novel method benefits from improved keypoint detection and matching through training using neural networks and data characterized by challenging geometries, outlier minimization and more robust descriptors by applying cross-attention layers. We conclude that continued development of deep learning-based feature tracking approaches for glacier velocity computations may substantially improve UAV-based velocity derivations applied to challenging situations. This method is able to deliver reliable displacement data in situations where traditional methods fail, which implies a new level of detail in understanding and interpreting glacier dynamics.