An optimized non-landslide sampling method for Landslide susceptibility evaluation using machine learning models

Machine learning models are extensively utilized in landslide susceptibility (LS) mapping. However, the conventional selection of non-landslide samples often contains numerous flaws, potentially leading to unscientific and inaccurate LS assessments. In this study, an optimized non-landslide sampling method (ONLSM) was creatively proposed for evaluating the LS of the Wanzhou section in the Three Gorges Reservoir, China. Initially, ground deformation rates were measured using Interferometric Synthetic Aperture Radar (InSAR). Concurrently, a bias-standardized information value (BSIV) model was employed to assess the LS, based on critical landslide-causing factors (geology, topography, hydrology and environment). Then, the LS were categorized into five susceptibility levels: very low, low, moderate, high and very high. Subsequently, non-landslide samples were selected from points with ground deformation rates ranged between + 5 mm/yr and − 5 mm/yr in very low susceptible level areas. Finally, LS maps were generated based on the proposed ONLSM in conjunction with support vector machine (SVM), random forest (RF), and gradient boosting decision tree (GBDT) models. The ONLSM-based maps exhibited superior accuracy compared to those produced by traditional non-landslide sampling methods (TNLSM) using the same machine learning models. The area under the receiver operating characteristic curve (AUC) values for ONLSM reached 0.974 (ONLSM + SVM), 0.977 (ONLSM + RF), and 0.986 (ONLSM + GBDT). It also indicated that the GBDT model based on ONLSM was more suitable for LS evaluation.