Multi-Sensor Fusion and Deep Learning Approaches for Semantic Segmentation of Glacial Lakes: A Comparative Study for Coastal Hydrology Applications

Monitoring glacial lakes is critical for assessing climate change impacts and mitigating glacial lake outburst flood risks. This study evaluates three deep learning architectures, U-Net, Simple Convolutional Neural Network (CNN), and Atrous Spatial Pyramid Pooling SegNet (ASPP SegNet), for binary semantic segmentation of glacial lakes using multi-sensor optical satellite imagery (Sentinel-2). Incorporating data augmentation and custom evaluation metrics (IoU, F1-score, validation loss), the results show that Simple CNN achieves the highest IoU (0.9155) and F1-score (0.9557). At the same time, ASPP SegNet demonstrates superior generalization with the lowest validation loss (0.03337). U-Net also delivers a reliable performance, albeit slightly lower. Visual and quantitative assessments highlight the advantage of multi-scale, context-aware architectures in delineating fragmented lake boundaries. This comparative study provides practical guidance for deep learning model selection in remote sensing-based glacial and coastal hydrology monitoring. Future work will explore temporal modeling, multi-class segmentation, and the integration of optical, radar, and elevation data for improved resilience.