Glacial Lakes Segmentation Using Multispectral Remote Sensing Data and Deep Learning Models

Abstract

The identification and delineation of glacial lakes through segmentation is crucial for tracking glacial changes and evaluating potential dangers from sudden flood events (GLOFs). These floods can severely impact populated areas and man-made structures downstream. Recent advances in high-quality satellite imagery have sparked increased attention toward using advanced machine learning methods, particularly deep learning, to enable precise and automated glacial lake detection. In this study, we explore the effectiveness of deep learning-based pointwise semantic segmentation for glacial lake mapping using multisource remote sensing imagery, including both optical and synthetic aperture radar (SAR) data. We experiment with a novel stack combination that fuses optical and SAR features to evaluate its effectiveness for glacial lake segmentation. To this end, we trained and evaluated four segmentation models—U-Net, U-Net++, Attention U-Net, and DeepLabV3+—on the enriched dataset. The results demonstrate noticeable performance improvements across all models, affirming the quality and complementary nature of the fused input stack. The study further investigates the spatiotemporal generalization capabilities of these models, examining how well they perform when transferred to unseen glacierized regions with differing topographic and climatic characteristics. A key finding is that integrating multispectral and multitemporal inputs, particularly through combined optical and SAR data, substantially enhances segmentation accuracy. Moreover, the choice of model architecture plays a critical role in achieving effective generalization across diverse terrains. This research contributes to the development of robust and scalable tools for longterm glacial lake monitoring, providing valuable insights for climate change impact assessment and disaster risk mitigation.