Search

Objective Strong earthquake-induced landslides are characterized by large number, wide distribution and large scale, and seriously threaten people's lives and property. Landslide susceptibility mapping (LSM) can quickly predict the spatial distribution of prone areas, which is highly important for reducing the risk of post-earthquake disasters. However, in the studies of coseismic landslide LSMs, how to select negative landslide samples and integrate machine learning models to improve the evaluation accuracy still needs further investigation. Methods In this study, the landslides induced by the Wenchuan earthquake in mountainous areas are selected as a case study. First, 10 landslide influencing factors, such as topography, geological environment, and seismic parameters, are selected to analyse the spatial distribution of landslides. Then, collinearity analysis is used to test data redundancy, nonnegative sample points from the sampling strategies are randomly selected in the extremely low susceptibility regions by the frequency ratio (FR) method. Finally, gradient boosting decision tree (GBDT), random forest (RF), and their optimal models are used to predict coseismic landslide susceptibility, conduct a comparative study of the models and carry out an accuracy assessment. Results The results show that ① the spatial distribution of landslides is controlled by multiple factors, and ② the accuracy of the models is FR-RF(AUC=0.943)>FR-GBDT(AUC=0.926)>RF(AUC=0.901)>GBDT(AUC=0.856). ③ Selecting negative landslide samples in low susceptibility areas could significantly improve the accuracy of LSMs. Conclusion The research results can provide a reference for selecting negative landslide samples and constructing evaluation models, as well as for providing theoretical support for post-earthquake disaster prevention and mitigation.