Glacial lake mapping using remote sensing Geo-Foundation Model

Abstract

Glacial lakes are vital indicators of climate change, offering insights into glacier dynamics, mass balance, and sea-level rise. However, accurate mapping remains challenging due to the detection of small lakes, shadow interference, and complex terrain conditions. This study introduces the U-ViT model, a novel deep learning framework leveraging the IBM-NASA Prithvi Geo-Foundation Model (GFM) to address these issues. U-ViT employs a U-shaped encoder–decoder architecture featuring enhanced multi-channel data fusion and global-local feature extraction. It integrates an Enhanced Squeeze-Excitation block for flexible fine-tuning across various input dimensions and combines Inverted Bottleneck Blocks to improve local feature representation. The model was trained on two datasets: a Sentinel-1&2 fusion dataset from North Pakistan (NPK) and a Gaofen-3 SAR dataset from West Greenland (WGL). Experimental results highlight the U-ViT model’s effectiveness, achieving an F1 score of 0.894 on the NPK dataset, significantly outperforming traditional CNN-based models with scores below 0.8. It excelled in detecting small lakes, segmenting boundaries precisely, and handling cloud-shadowed features compared to public datasets. Notably, the U-ViT demonstrated robust performance with a 50% reduction in training data, underscoring its potential for efficient learning in data-scarce tasks. However, its performance on the WGL dataset did not surpass that of DeepLabV3+, revealing limitations stemming from differences between pre-training and input data modalities. The code supporting this study is available online. This research sets the stage for advancing large-scale glacial lake mapping through the application of GFMs.