CryoSentinel: A Multimodal Foundation-Model Segmenter for Glacial Lakes in High Mountain Asia from Sentinel-1 SAR, Sentinel-2 Optical, and Copernicus DEM Imagery

Abzal Abdrash 

Glacial-lake outburst floods (GLOFs) are the dominant climate-driven hazard in High Mountain Asia, and reliable lake-extent segmentation is the prerequisite for every downstream early-warning workflow. We present CryoSentinel, a multimodal foundation-model semantic segmenter built on the IBM/ESA TerraMind 1.0 Large encoder (1.1 B parameters) with a UperNet decoder, fine-tuned on 5,614 co-registered chips spanning the Tien Shan, Zhetysu Alatau, and Ile Alatau ranges of Central Asia. The model jointly ingests Sentinel-1 GRD SAR (VV, VH), Sentinel-2 L2A optical (12 bands), and Copernicus DEM 30 m elevation at 10 m / pixel and is trained under a 17-km-buffered spatial-block-split protocol that rules out lake-level leakage. On the validation split it attains an IoU of 0.9557 at threshold 0.5 with flip-only test-time augmentation, a +4.27 percentage-point improvement over the most directly comparable prior result (Adhikari and Regmi, 2025; arXiv:2512.24117; Sentinel-1-only, IoU 0.9130). On a held-out spatial test split it reports IoU 0.8918; an independent MNDWI cross-check reveals that 7 of the 665 test chips are Kumar-polygon mislabels rather than model failures, and dropping these raises the label-corrected test IoU to 0.9082. We document the eleven configuration corrections that lifted our finetune from 0.825 to 0.9557 IoU, the spatial-block-split protocol, and the label-noise audit procedure in full so the result is reproducible end-to-end. Code (Apache 2.0), pretrained weights, and the 42,237-chip dataset (ODC-By 1.0) are publicly released and assigned permanent DOIs.