Abstract: Objectives: In order to improve the efficiency of landslide extraction and explore the spatial-temporal distribution characteristics of regional landslides and single landslide, a rapid landslide extraction model is designed to provide a scientific basis for landslide disaster prevention and management. Methods: Based on multi-temporal domestic high-resolution remote sensing satellite images, ALOS 12.5mDEM, and historical landslide hidden danger points, this study selected 8 landslide-prone townships located in the northwest of Fengjie County to form the study area. Feature optimization based on SHAP interpretation framework and Bayesian hyperparameter automatic optimization based on Optuna framework are introduced into XGBoost algorithm to construct and optimize the landslide extraction model. The study realized rapid extraction of landslide spatial information and quantitative analysis of landslide spatial-temporal distribution in 2013, 2015, 2018 and 2020. Results: In the comparison of models accuracy, Accuracy, Precision, Kappa coefficient and AUC value of landslide extraction model constructed by optimized XGBoost basic algorithm are 96.26%, 90.91%, 0.8602 and 0.9705, respectively. It is higher than GBDT, LightGBM and Adaboost. Conclusions: From 2013 to 2020, the overall development degree of landslides in the study area is relatively high, and the spatial distribution of landslides is uneven among villages and towns, and the landslides is distributed on both sides of the river valley and the river, showing the characteristics of regional concentration. From 2013 to 2020, Miaowan landslide has high development intensity, strong activity, repeated resurrection phenomenon and induced new landslides. The slope of the landslide is about 25° ~ 45°, and it’s topographic characteristics change little. It’s significant changes mainly focus on color, texture, geometry and vegetation coverage.