Machine learning-based probabilistic prediction of glacial lake formation using erosional and topographic features

This study develops a probabilistic framework to predict the probability of glacial lake formation (PGLF) in the Eastern Himalaya by integrating key erosional and topographic features. Using Google Earth imagery and digital elevation models within a 3 × 3 neighbourhood grid structure, we evaluated three predictive models: Logistic Regression (LR), Artificial Neural Network (ANN), and Bayesian Neural Network (BNN). BNN outperformed LR and ANN with an AUC of 0.878, while also estimating both aleatoric and epistemic uncertainties (10⁻³ to 10⁻⁴), enhancing prediction confidence. Neighbouring lakes, cirques, gentle slopes, and retreating glaciers emerged as the most influential predictors, demonstrating the importance of geomorphology, which is often omitted from prior models. The proposed approach offers a transferable framework for identifying high-risk glacial lake formation sites, supporting regional hazard mitigation, early warning systems, and sustainable water resource management in the Himalaya and other glaciated regions. Future improvements should integrate moraine development chronologies, automate data preparation, and incorporate field validation to further refine predictive accuracy and inform global mountain hazard management efforts.

Glacial lake formation in high mountain regions, particularly the Himalayas, is accelerating due to climate-driven glacier retreat, increasing the risk of glacial lake outburst floods (GLOFs) that threaten downstream populations and infrastructure. While climate governs meltwater availability, the formation and evolution of glacial lakes are primarily controlled by geomorphological features such as cirques, valleys, flow channels, retreating glaciers, and neighbouring lakes. However, most predictive models overlook these controls, limiting hazard forecasting capabilities.