Abstract: In the field of modeling the susceptibility of county-level earthquake-induced landslides, comparative analysis research on ensemble learning models is still insufficient. This study takes Luding County as an example, where detailed information on 502 landslide events was collected based on historical records, remote sensing data, and field surveys to construct a sample database. Twelve key influencing factors were selected, including digital elevation model (DEM), slope, aspect, profile curvature, plan curvature, topographic position index, lithology, and the distances to roads, faults, rivers, and seismic sources, as well as land use. After conducting collinearity tests on these factors, the Bayesian optimization method was used to optimize ensemble learning models such as AdaBoost, GBDT, XGBoost, LightGBM, and CatBoost, thereby constructing a landslide susceptibility assessment system. The accuracy and predictive precision of each model were evaluated through ROC curves and frequency ratio comparison. The results show that after Bayesian optimization, the performance of all models was significantly improved, with the BO-GBDT model's AUC value increasing by 7.4%. Frequency ratio analysis further indicated that the BO-XGBoost model had the strongest identification capability at high susceptibility levels, with a frequency ratio of 13.73, demonstrating its superior risk area identification capability. Although the BO-GBDT model had a slight advantage in AUC value (0.974), its frequency ratio at high susceptibility levels was only 10.14. Therefore, the selected BO-XGBoost model reveals that high-susceptibility areas are mainly concentrated along the Dadu River and its nearby towns and transportation routes, while the landslide susceptibility in the eastern mountainous area is relatively low, providing important guidance for future research and disaster prevention and mitigation efforts.