Avalanche Susceptibility Mapping with Explainable Machine Learning: A Case Study of the Kanas Scenic Transportation Corridor in the Altay Mountains, China

Avalanche susceptibility mapping is vital for disaster prevention and infrastructure safety in cold mountain regions under climate change. Traditional machine learning (ML) approaches have demonstrated strong predictive capacity, yet their limited interpretability and difficulty in identifying threshold effects hinder their broader application in geohazard risk management. To overcome these limitations, this study develops an explainable ML framework that integrates remote sensing data, topographic and climatic variables, and SHapley Additive exPlanations for the Kanas Scenic Area transportation corridor in the Chinese Altay Mountains. The framework evaluates five classifiers: Random Forest, XGBoost, LightGBM, Soft Voting, and Stacking, and using sixteen conditioning factors that capture topography, climate, vegetation, and anthropogenic influences. Results show that LightGBM achieved the best performance, with an AUC of 0.9428, accuracy of 0.8681, F1-score of 0.8750, and Cohen’s kappa of 0.7366. To ensure transparency for risk decisions, SHAP analyses identify Terrain Ruggedness Index, wind speed, slope, aspect and NDVI as dominant drivers. The dependence plots reveal actionable thresholds and interactions, including a TRI plateau near 5–7, a slope peak between 30° and 40°, a wind effect that saturates above about 2.5 m s−1, and a near-river high-risk belt within 0–2 km. The five-class map aligns with independent field observations, with more than three quarters of events falling in moderate to very high zones. By integrating explainable ML with remote sensing, this study advances avalanche risk assessment in cold region transportation corridors and strengthens the robustness of regional susceptibility mapping.