A progressive snow avalanche Hazard assessment framework coupling interpretable machine learning with dynamic simulation

Snow avalanches severely damage infrastructure in high-mountain regions. Susceptibility maps show where avalanches may occur but do not quantify hazard intensity, which limits risk-informed decisions. We propose a decision-oriented framework that combines interpretable machine learning with dynamic simulation, applied to the Eastern Himalayan Syntaxis. A remote-sensing inventory of 1233 avalanche events was compiled and validated by field surveys. A Bayesian-optimized XGBoost model produced susceptibility maps. SHAP analysis identified critical thresholds of key drivers. These thresholds were used to automatically map potential release areas under general and extreme scenarios. The release areas were then input into the Rapid Mass Movement Simulation (RAMMS) to generate hazard intensity maps and assess road exposure along the Galongla section of the Zhamo road. The model achieved high predictive performance (AUC = 0.932). High-susceptibility zones concentrated in the Duoxiongla and Galongla corridors. Four dominant factors were NDVI (15.6%), aspect (11.0%), slope (9.1%), and roughness (8.9%). Under general and extreme scenarios, avalanche-affected areas covered 54.6% and 66.2% of the study region, and 16.7% and 25.8% of road segments faced high risk (impact pressure > 500 kPa). The framework provides practical support for hazard zoning, infrastructure prioritization, and scenario-based risk evaluation in data-scarce high-mountain environments.