Free satellite ML model predicts Himalayan glacial lake bursts at 0.89 ROC

A new arXiv paper by Matthew Kahn, Milan Arjel, Nirmala Adhikari, Mingmar Sherpa, and James Pope tests whether free satellite data alone can predict which Himalayan glacial lakes are at risk of bursting and when triggers arrive. The study combines two satellite signals: radar interferometry that measures slow sagging of moraine dams, and satellite weather data that tracks weeks of heavy rainfall or heat stress on primed lakes. Using 589 dated glacial lake outbursts from the HMAGLOFDB database plus thousands of catalogued landslides, the authors matched each event against similar unfailed sites, then validated under spatial cross-validation that withholds entire map tiles to prevent models from memorizing training locations.

Antecedent weather proved effective at timing hazards, hitting ROC 0.73 for large moraine- and ice-dammed bursts, 0.83 for rainfall-triggered landslides, and 0.82 for small glacier pond floods. Terrain-based susceptibility scoring was weaker once compared honestly against nearby sites—0.76 for bursts, 0.71 for landslides, and just 0.54 (chance) for small floods—though the burst signal strengthened to 0.89 within Nepal alone. Notably, five deep-learning models failed to decisively beat a simple gradient-boosted tree baseline, which was itself reproducible by a three-rule decision tree on ruggedness and monsoon rainfall. The paper closes with a ranked Nepal watchlist as a prioritisation aid, not a prediction, and clarifies where free satellite data hits its limits.