Abstract
Snow accumulation and ablation play a pivotal role in the mass balance of mountain glaciers. Accurate mapping of snow extends on glaciers enhance our understanding of climate change impacts (Larocca et al.) and serve as valuable input for glacier surface mass balance models (Rabatel et al.). For example, the snowline altitude at the end of the summer season serves as a proxy for the Equilibrium Line Altitude and strongly correlated with the annual mass balance (Rabatel et al.). The largest study mapped 3489 snowlines of 269 glaciers of the ~275.000 glaciers from Landsat data between year 1984 and 2022 and found an increase in snow line elevation of approximately 150 meters. This significant rise indicates a reduction in glacier accumulation zones, suggesting negative mass balance trends driven by rising temperatures and shifting precipitation patterns. Regional variability in snowline changes underscores the complex interplay of climatic and local factors, emphasizing the importance of snowline monitoring for glacier health and climate change studies (Larocca et al.). However, manual delineation is time consuming and band- and index threshold-based approaches often fails in areas of steep and complex mountainous terrain and in homogenous snow conditions. The spectral similarity to firn and ice and varying illumination conditions in steep terrain are posing significant challenges for accurate large-scale and long-term snow monitoring Advanced machine learning techniques offer promising solutions to address these limitations by automating snowline detection and improving accuracy under diverse conditions (Prieur et al.). Designed to overcome these limitations by improving accuracy and scalability, this study presents first steps towards region-scale snow cover extend mapping using machine learning. We manually digitized 312 snowlines on 41 glaciers in Scandinavia (13), Svalbard (9), and the European Alps (19) from Sentinel 2 data in the period 2015 to 2023, encompassing a wide range of seasonal snow conditions., Using this benchmark, we trained several machine learning models, including pixel-based classifiers such as Support Vector Machine, Random Forest, and XGBoost (Chen et al.), and U-Net (Ronneberger et al.), a Fully convolutional neural network , and compared them against threshold-based approaches as baselines. The results from Scandinavia demonstrate the superiority of machine learning methods. While threshold-based approaches, such as the Normalized Difference Snow Index (NDSI > 0.4) and Near-Infrared (NIR > 0.11), achieved an Intersection over Union (IoU) score of 0.7147, U-Net significantly outperformed with an IoU of 0.9456. Random Forest was the next best-performing method with an IoU of 0.8957, followed by XGBoost (0.8899) and SVM (0.8887). Adding elevation models and slope data to the classifiers resulted in only marginal performance improvements. This significant improvement highlights the potential of U-Net to accurately capture fine-scale snowline features, especially in heterogeneous and complex mountainous environments with spectrally similar classes such as firn and ice and paves the way for accurate, low-cost, automated and large-scale snow mapping on glaciers.