Hazard assessment of glacial lakes in Himalaya through erosional features detection using deep learning

Glacier retreat under climate change has significantly increased the number and size of glacial lakes in the Eastern Himalayas over the past three decades, increasing the risk of glacial lake outburst floods (GLOFs). The formation and expansion of glacial lakes are interlinked with the evolution of erosional features under glacier retreat. Therefore, this research proposes a framework for detecting key erosional features, namely cirques, valleys, flow channels, and retreating glaciers leading to glacial lake evolution, followed by hazard assessment of existing lakes and identification of potential future lakes using a heuristic approach. This study employs an EfficientNet-B0 convolutional neural network (CNN)-based model to detect the erosional features and glacial lakes in the Eastern Himalayas through Google Earth imagery from 2015 to 2022. The trained CNN model achieved high accuracy, with intersection over union (IoU) scores exceeding 0.88 for all features. The trained CNN model was tested on 12,968 images, and results were validated using performance metrics, highlighting the robustness of the model. The CNN model detected 2,647 glacial lakes in the study area. The hazard assessment identified 235 extremely dangerous, 357 very highly dangerous, 175 highly dangerous, 1269 moderately dangerous, 99 low dangerous, and 512 non-dangerous lakes. The results also identified 174 locations where glacial lakes are likely to form in future. The proposed method can be used to update the database of dangerous glacial lakes and locations of future lakes. This will assist in early warning and proactive disaster management, contributing to the safety of vulnerable mountainous communities and future GLOF studies.