Automating glacier facies classification: pan-European dataset and deep learning baseline

Konstantin Maslov, Thomas Schellenberger, Prashant Pandit, Claudio Persello, Alfred Stein

Glacier facies play a critical role in understanding the mass balance of glaciers, offering insights into accumulation and melting processes. Large-scale mapping of glacier facies from satellite data is therefore essential for monitoring glacier response to climate change and informing climate policies. In this study, we present the largest glacier facies dataset ever compiled for Europe, comprising 31 glaciers, 92 Landsat and Sentinel-2 scenes, 137592 expert point labels and eight classes—five glacier facies (ice, snow, debris, firn and refrozen-like) and three miscellaneous classes (shadow, water and cloud)—encompassing a wide variety of surface conditions. A confident learning method pruned 16% of ambiguous expert labels overall. A compact and straightforward convolutional neural network reached a macro-average F1 score of 82% on the complete cleaned data or 74% on the full, unpruned data, and 82.3±10.5% glacier-wise. This performance remains consistent across different regions and sensors. When the facies products were regressed against World Glacier Monitoring Service records, they showed moderate, yet significant correlation with the surface mass balance measurements globally (r = 0.65, RMSE = 0.60 m w.e., where 1 m w.e. = 1000 kg m^-2 denotes metre of water equivalent) and competitive correspondence for glacier-specific calibration (r = 0.79, RMSE = 0.28 m w.e.). Overall, the dataset and baseline show that large-scale glacier facies classification can be achieved with high consistency. By providing both the dataset and baseline classification models, we aim to support the broader community in developing more advanced methods for glacier facies mapping to enhance our understanding of ongoing glacial changes.