Considering the effect of non-landslide sample selection on landslide susceptibility assessment

Crafting landslide susceptibility mapping is pivotal for the effective management of landslide risks. However, the influence of non-landslide sample selection on the modeling performance of landslide susceptibility assessment models remains a crucial challenge to overcome. This article employs Huize County as the research area and identifies 12 factors that exert influence. In this study, we utilized Extreme Gradient Boosting and Random Forest algorithms, and four methods (Whole-area random selection method, Buffer method, Frequency Ratio method, and Analysis Hierarchy Process) were employed to select non-landslide samples for constructing the landslide susceptibility assessment model. The findings revealed that the model performance derived from different non-landslide sample selections exhibited significant variations, and the models obtained using the buffer zone method, frequency ratio method, and AHP method performed better than the full-area random selection model. Among the evaluated models, the AHP method for non-landslide sample selection demonstrated the most optimal performance, with an AUC of 92.17% for XGBoost-AHP and 91.64% for RF-AHP. Based on the SHapley Additive explanation (SHAP), the main variables impacting landslide danger in the research area were elevation, NDVI, and peak seismic acceleration. This study provides theoretical support for the selection of non-landslide samples in landslide susceptibility assessments and interpretable AI research.