Abstract: The complex terrains in Southwest China often lead to frequent landslides． Past studies mostly focus on single geomorphic types, overlooking systematic comparison of model applicability and heterogeneity of dominant factors across different geomorphic scenarios．Traditional “black-box” models often struggle to reveal effects of geomorphology on landslides． In this study we establish a “Machine Learning-Geomorphic Zoning-Interpretability” framework for Chengkou county (mountainous), Liangping district (hilly), and Zhongjiang county (low hill-plain)． Random forest (RF), XGBoost, and LightGBM algorithms were employed to assess landslide susceptibilities． SHAP method was used to quantitatively evaluate contributing factors under different geomorphic settings． XGBoost model demonstrated strong robustness, achieving best performance in mountainous areas (AUC = 0.923), but with slightly lower performance in hilly and plain areas． Geomorphic differences significantly modulated dominant factors． Landslides were found primarily driven by elevation and human activities (POI kernel density, distance to roads) in Chengkou． Lithology and roughness index dominated in Liangping． But rainfall and elevation were key factors in Zhongjiang． In this study we could verify the advantages of SHAP in breaking the “black box” nature of past models． We also elucidated differentiated driving mechanisms of landslide development from a geomorphic perspective． We therefore provide some solid scientific basis for refined disaster prevention in regional landslides．