Snow avalanche susceptibility, hazard, and exposure assessment in the Western Himalaya using machine learning and numerical modelling

Abstract

Snow avalanches pose a significant threat to infrastructure and communities. Avalanches widely affect the Western Himalayan basins every year. This study evaluates avalanche susceptibility and hazard in the Chandra-Bhaga and Upper Beas basins of the Western Himalaya using machine learning and numerical modelling. A variety of machine learning algorithms - including Random Forest, Support Vector Machine, Logistic Regression, and Artificial Neural Network - were tested and compared using a comprehensive set of avalanche predictive factors, to assess avalanche susceptibility at the basin scale. The random forest model achieved 88.73% accuracy and an area under the curve (AUC-ROC) of 0.95. 1,484 potential avalanches were simulated for hazard and exposure analysis. Findings reveal that ~8% of the region is highly susceptible to avalanches, particularly in Lahaul and Spiti. With a snow release-depth of 0.5 m originating from the high and very-high avalanche susceptible slopes, ~161 buildings and 7 lakes are exposed to potential avalanches. In a worst-case scenario with a 3-meter avalanche release-depth, the exposure significantly increases to ~557 buildings and 9 lakes. The findings of the study are crucial for site specific detailed avalanche forecasting and can serve as a base for identifying avalanche hotspots.