Integrating machine-learning-based landslide susceptibility and vulnerability for slope-unit risk assessment

Abstract

Landslide susceptibility identifies terrain associated with landslide occurrence, whereas risk prioritization also requires information on the vulnerability of threatened elements. This letter presents a slope-unit framework linking machine-learning susceptibility prediction, SHAP-based interpretation, and vulnerability-based relative risk classification. Four models were evaluated using a regional test dataset, and the selected model was applied to 393 slope units. XGBoost achieved the highest area under the receiver operating characteristic curve (AUC = 0.952), followed by random forest (0.904), logistic regression (0.662), and support vector machine (0.600). Paired DeLong tests indicated significant differences between XGBoost and logistic regression (LR) and support vector machine (SVM), but not random forest (RF). SHAP analysis identified slope, relief, and elevation as the leading predictors, accounting for 62.83% of normalized grouped importance. High and very high relative risk occupied 21.34% of the mapped slope-unit area. Of 69 very-high-susceptibility units, only 13 were classified as very high risk after vulnerability integration. The framework supports the identification of priority slope units for targeted field investigation, monitoring, and mitigation planning.