Assessment of the Avalanche Susceptibility Using Multiple Machine Learning Algorithms in Western Himalayan Watersheds

Snow avalanches are masses of snow that descends rapidly down a slope and has the potential to cause fatalities and damage infrastructure including roads, dams, and buildings. These events are a common natural hazard in the glaciated and snow covered areas of the Indian Himalayan Region (IHR). The rising air temperatures due to global warming have led to early wet snow formation, contributing to an increased frequency of avalanches in recent years across the IHR. Avalanche susceptibility is crucial for avalanche forecasting and infrastructure planning. In the current study, snow avalanche susceptibility is modelled using multiple machine learning techniques in the Chandra-Bagha and Upper Beas basins covering parts of the states of Himachal Pradesh and Jammu and Kashmir, Western Himalaya. The study evaluates 24 predictive variables, including topographic, hydrological, cryospheric, geological, climatic, and anthropogenic layers using various machine learning algorithms. The random forest technique produced promising results with an accuracy of 88%. The results are presented as avalanche probabilities, which are then reclassified into five classes for susceptibility mapping. Further, the predictive variables are ranked based on their influence on the accuracy of the machine learning algorithm. Valley depth, snow cover duration and distance to lineaments are identified as the the most important variables for predicting snow avalanches in the region. 

Keywords: Snow Avalanche, Machine Learning, Himalaya, Random Forest