Machine Learning-Based probabilistic prediction of glacial lake formation using erosional and topographic features

Abstract

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. 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.

Keywords: Early warning; Eastern Himalaya; Erosional features; GLOF; Google Earth images.

Fig. 1

Location and topographic overview of the study area in the Eastern Himalaya, covering parts of Arunachal Pradesh (India) and adjacent regions of Tibet. The top panels illustrate the regional context, the specific study boundary, and the grid overlay used for spatial analysis. The bottom panel shows an SRTM 30 m (DEM) of the study area, with elevation ranging from 340 to 7,252 m. The figure is prepared using Google Earth Pro version 7.3.6.10201 (https://earth.google.com) and ArcMap 10.8 (https://www.esri.com/en-us/arcgis/products/arcgis-desktop/resources). The region exhibits complex mountainous terrain, extensive valley networks, and high-altitude zones where glacier retreat and glacial lake formation processes are active.

Fig. 2

Flowchart depicting the study approach.

Fig. 3

Representative examples of mapped glacial erosional features in the Eastern Himalaya based on high-resolution Google Earth imagery, generated using Google Earth Pro version 7.3.6.10201 (https://earth.google.com). Yellow lines delineate the boundaries of (top) glacial lakes, (middle) cirques, and (bottom) glacial valleys. These features govern the spatial occurrence, expansion, and hydrological connectivity of glacial lakes, influencing downstream flood hazards and landscape evolution. The systematic mapping of these features forms a critical input for the probabilistic prediction of glacial lake formation.

Fig. 4

Mapped glacial erosional features in the Eastern Himalaya are shown using high-resolution Google Earth images, generated using Google Earth Pro version 7.3.6.10201 (https://earth.google.com). Visual examples of retreating glaciers (top two rows) and associated flow channels (bottom two rows) in a high mountain region. Yellow lines delineate glacier margins and hydrological flow paths. Retreating glacier images show ice thinning, exposed rock surfaces, and the formation of proglacial lakes, indicative of significant ice loss. Flow channel images illustrate the evolution of drainage networks in response to glacial recession. These visual observations provide contextual support for identifying terrain susceptible to glacial lake formation and hydrological transformation.

Fig. 5

ANN architecture.

Algorithm 1 Training a ANN and probability estimation for a feature class

Algorithm 2 Training BNN and uncertainty quantification with variational inference

Fig. 6

Model accuracy based on ROC curve for the test dataset: (a) LR, (b) ANN, and (c) BNN.

Fig. 7

LR model-based probability map of glacial lake formation generated in ArcMap 10.8 software (URL: https://www.esri.com/en-us/arcgis/products/arcgis-desktop/resources). The LR-derived map shows colour-coded probability, where red indicates high and dark green low probabilities of lake formation.

Fig. 8

ANN model-based probability map of glacial lake formation generated in ArcMap 10.8 software (URL: https://www.esri.com/en-us/arcgis/products/arcgis-desktop/resources). It displays colour-coded probabilities, where red indicates high and dark green low chances of lake formation.

Fig. 9

Spatial distribution of the BNN model based on PGLF and its associated variance, generated in ArcMap 10.8 software (URL: https://www.esri.com/en-us/arcgis/products/arcgis-desktop/resources). The top panel shows PGLF values categorised from very low (0.0–0.2, green) to very high (0.8–1.0, red), indicating areas with varying likelihood of glacial lake development. The bottom-left panel depicts variance in PGLF predictions at a scale of 10⁻⁴, while the bottom-right panel shows variance at 10⁻³. Higher variance areas (purple to light brown) indicate greater uncertainty in model predictions and identify zones requiring further investigation.