In this study, a novel interpretable model based on SHAP and XGBoost is proposed for the interpretation of landslide susceptibility evaluation at global and local levels. First, 10 condition factors and 5 rainfall factors were collected, and r.slopeunits software was used to delineate slope units as evaluation units. The sample was divided into two subsets by 7:3 for model training and model testing. Then, XGBoost and 3 machine learning methods (RF, LR, and SVM) were compared for landslide susceptibility evaluation. Finally, factor importance ranking, factor dependence analysis, and single sample interpretation were implemented separately using SHAP. An integrated framework incorporating both global and local explanations was proposed based on SHAP for interpreting landslide susceptibility assessment results and the interactions between influencing factors. In addition, the one-factor dependence plot of SHAP revealed the nonlinear response of landslides to the influence factor, indicating the early warning threshold of the influence factor.