Strong earthquake-induced landslides have the characteristics of a large number, wide distribution, and large scale, which seriously threaten the safety of people's lives and property. Landslide susceptibility mapping (LSM) can quickly predict the spatial prone-area distribution, which is of great significance for reducing risky disasters post-earthquakes. However, in the study of co-seismic landslide LSM, how to select the landslide negative samples and integrated machine learning model to improve the evaluation accuracy still needs further investigation. In this research, the landslides induced by the Wenchuan Earthquake were selected as a case study in mountainous areas. Firstly, 10 landslide influencing factors such as topography, geological environment, and seismic parameters were selected to analyze the landslide spatial distribution; secondly, collinearity analysis was used to test data redundancy. Non-negative sample points of sampling strategies were randomly selected at the extremely low susceptible regions by frequency ratio method (FR). Finally, Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and their optimal models were used to map the co-seismic landslides susceptibility, conduct a comparative study of the models and carry out the accuracy assessment. The results show that: ①The landslide spatial distribution is controlled by multi-level factors; ②the accuracy of the models is FR-RF (AUC=0. 943) > FR-GBDT (AUC=0. 926) > RF (AUC=0. 901) > GBDT (AUC=0. 856). Selecting the landslide-negative samples at low-prone areas could significantly improve the LSM accuracy, and the research results can provide a reference for selecting landslide-negative samples and constructing the evaluation models, as well as for providing theoretical support for disaster prevention and mitigation of post-earthquake.