The loess hilly area is one of the areas with a high incidence of geological disasters, and it is urgent to use appropriate evaluation factors and training models to conduct research on the susceptibility assessment of geological disasters. Kangdian Town, Gongyi City, the township hardest hit during the "7·20" extremely heavy rainstorm in Zhengzhou, was taken as the study area. Based on satellite remote sensing interpretation, field survey, UAV aerial photography and relevant data collection, an evaluation system covering 13 influencing factors of three main control factors, namely loess interface, human engineering activities and hydrodynamic effects, was constructed. CatBoost model, XGBoost model and LightGBM model were used to carry out the evaluation study of geological disaster vulnerability. Based on the machine learning model with the best performance, SHAP(shapley additive explanations) algorithm was used to complete the global interpretation of characteristics and dependency analysis. The results show that the CatBoost model has higher accuracy than other models (XGBoost and LightGBM), and performs the best in AUC(area under curve) value, accuracy, precision, recall, F₁ score, and field validation. The proportion of areas with extremely high, high, medium, low, and extremely low susceptibility is 3.19%, 1.40%, 2.04%, 5.93%, and 87.44%, respectively. The extremely high and high susceptibility areas are mainly distributed on both sides of gullies with strong human activities, and slope cutting and building are important causes of geological disasters. The aim of this study is to optimize the modeling approach, investigate the uncertainty and interpretability of the modeling process, explain and analyze the decision-making mechanism of machine learning susceptibility, and provide scientific basis for geological disaster prevention and control in the loess hilly area of western Henan.