Integrating satellite remote sensing and deep learning for mapping glacier boundary and its changes in the Himalaya

Abstract

Himalaya glaciers are changing rapidly, and therefore, need to be mapped with greater frequency. However, traditional mapping methods are challenging, especially over debris-covered areas and accumulation zones. These areas can easily be confused with bare rocks and perennial snow patches, respectively. In this study, we integrated glacier-dynamic variables, surface velocity and elevation change, along with multispectral data from Sentinel-2 and Landsat-8, InSAR coherence and backscatter from Sentinel-1, DEM derivatives, and multi-year minimum NDSI (perennial snowpatches) into five improved DL models (UNet++, M-LandsNet, GlacierNet-2, GlaViTU, and SAUNet) for delineating glaciers in the five basins of the Himalaya. We identified the most relevant layers that retain up to 80% of the information, reducing redundancy. This modified UNet++ achieved the highest mIoU of 0.92–0.98, surpassing the 18-layer input (mIoU = 0.84–0.94) in discriminating snow patches from small glaciers. The modified UNet++ model was applied to generate outlines of more than 500 glaciers in three epochs: 2010, 2019, and 2024. Lake-terminating glaciers retreated 2–3 times faster than land-terminating glaciers, with the highest frontal retreat rate observed in the Chhombo Chhu watershed (∼18–22 m yr−1), while debris-covered glaciers such as Gangotri showed lower rates (<10 m yr−1). This study highlights the importance of glacier dynamics as the input layer for detecting different types of glacier surfaces (e.g., debris, snow), discriminating between similar adjacent off-glacier surfaces, and delineating multi-temporal mountain glacier boundaries. This study shows the potential of integrating satellite data with deep learning methods to efficiently map mountain glaciers, which may be useful in investigating their response to the changing climate.