IceWatch: a multimodal deep-learning framework for monitoring glacial lake outburst floods, evaluated at Shisper Glacier, Karakoram

Abstract. Lakes dammed by surging glaciers can fill and drain within one season, often under cloud. We present IceWatch, which combines a convolutional classifier of Sentinel-2 scenes (RiskFlow), forecasts of glacier velocity (TerraFlow) and land-surface temperature (TempFlow), and a probabilistic fusion stage (FusionFlow). We test it at Shisper Glacier, Karakoram, where an ice-dammed lake drained four times between 2019 and 2022, with all data taken over the lake and its dam, and scenes labelled from a published chronology. Trained on earlier years and tested on 2022 to 2025, RiskFlow detected both post-outburst scenes of 2022 with 2 false alarms in 102 other scenes. With each year held out in turn, it detected 4 of 7 post-outburst scenes with no false alarm in 151 others, whereas a water-index rule detected all 7 but raised 26 false alarms. A model trained on 103 other ice-dammed lakes did not transfer. TerraFlow was 30 % more accurate than persistence, and TempFlow beat persistence and climatology two days ahead. Three of the four outbursts followed cooler-than-normal weeks, and neither an agreement rule nor the fusion improved on RiskFlow. IceWatch can confirm drainage from the next clear scene; anticipating outbursts will require observations of lake filling.