Avalanche Susceptibility Assessment in the Kanas-Hemu Scenic Area of Xinjiang Using Coupled Frequency Ratio and Machine Learning Models

In the context of global climate change, avalanche disasters frequently occur in the Kanas-Hemu scenic area of Xinjiang, China, posing continuous threats to regional disaster prevention and mitigation, transportation safety, and tourism. To improve the accuracy of avalanche susceptibility assessment, this study aims to construct and compare coupled models that integrate frequency ratio (FR) and machine learning (ML) methods, systematically evaluate avalanche susceptibility along key transportation routes in the study area, and identify the key influencing factors and their spatial distribution characteristics. A total of 11 factors from three categories, including topography, meteorology, and snow cover, were selected. Six susceptibility assessment models were constructed by combining FR with various ML algorithms (SVM, MLP, XGBoost). The SHAP method was employed to interpret the contribution of each factor. The results indicate that the coupled models (FR-SVM, FR-MLP, FR-XGBoost) outperformed their corresponding single ML models. Among them, the FR-XGBoost model achieved the best overall performance, with an AUC of 0.897. Slope gradient and NDVI were identified as the most important influencing factors across all models. Besides, spatial distribution analysis reveals that high and very high susceptibility zones are primarily distributed in a strip-like pattern along gullies and major transportation routes with significant topographic relief in the northwestern and southwestern parts of the study area. This study demonstrates the superiority and applicability of coupled FR-ML models in avalanche susceptibility assessment. The findings can provide a scientific basis for local avalanche risk prevention and control, transportation safety assurance, and the development of a tourism early warning system.