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

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 dataset of visual proxies of glacier facies ever compiled for Europe, comprising a sample of 31 glaciers, 92 Landsat and Sentinel-2 scenes, 137 592 expert point labels and eight classes—five facies-related classes ( 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 F 1 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 classification products were regressed against World Glacier Monitoring Service records, they showed a moderate, yet significant correlation with the surface mass balance measurements. Overall, the dataset and baseline show that large-scale classification of glacier facies proxies can be achieved with high consistency. By providing both the dataset ( https://doi.org/10.5281/zenodo.18469893 ) and baseline classification models ( https://github.com/konstantin-a-maslov/glacier_facies_classification ), we aim to support the broader community in developing more advanced methods for glacier facies mapping to enhance our understanding of ongoing glacial changes.