Satellite-based supraglacial lakes (SGLs) mapping is crucial for assessing their impacts on ice motion and glacier stability. The Segment Anything Model (SAM) proposed as the first foundation model in image processing, provides a new opportunity for automated SGLs extraction. However, its potential has not been thoroughly explored, and the reliance on manual prompts heavily hinders large-scale mapping of SGLs. Therefore, we introduced an adapting SAM for automated SGLs extraction, and comprehensively assessed its potential across Greenland Ice Sheet (GrIS). The adapting SAM was first fine-tuned on small datasets derived from Sentinel-2 imagery, and then was evaluated across five major basins of GrIS using Sentinel-2 and Landsat-8 imagery. Results demonstrated strong generalization capabilities across varying training size, different regions, and multi-source satellite images, achieving an average F1-score of 87.77%, which represented an average improvement of 19.18% over UNet and DeeplabV3 + . Notably, even when fine-tuned with 20 samples, the adapting SAM maintained a high average F1-score of 85.73%. Furthermore, the adapting SAM was utilized to large-scale mapping of SGLs in Isunnguata-Russell glacier, and was superior to an existing product derived based on water index, identifying 5.14 times more lakes. Overall, this study offers valuable insights for SGLs mapping based on foundation model.
Chai et al. (2025) studied this question.