Machine learning for mapping glacier surface facies in Svalbard

Glaciers are dynamic and highly sensitive indicators of climate change, necessitating frequent and precise monitoring. As Earth observation technology evolves with advanced sensors and mapping methods, the need for accurate and efficient approaches to monitor glacier changes becomes increasingly important. Glacier Surface Facies (GSF), formed through snow accumulation and ablation, serve as valuable indicators of glacial health. Mapping GSF provides insights into a glacier\'s annual adaptations. However, satellite-based GSF mapping presents significant challenges in terms of data preprocessing and algorithm selection for accurate feature extraction. This study presents an experiment using very high-resolution (VHR) WorldView-3 satellite data to map GSF on the Midtre Lovénbreen glacier in Svalbard. We applied three machine learning (ML) algorithms—Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Machine (SVM)—to explore the impact of different image preprocessing techniques, including atmospheric corrections, pan sharpening methods, and spectral band combinations. Our results demonstrate that RF outperformed both ANN and SVM, achieving an overall accuracy of 85.02 %. However, nuanced variations were found for specific processing conditions and can be explored for specific applications. This study represents the first clear delineation of ML algorithm performance for GSF mapping under varying preprocessing conditions. The data and findings from this experiment will inform future ML-based studies aimed at understanding glaciological adaptations in a rapidly changing cryosphere, with potential applications in long-term spatiotemporal monitoring of glacier health.