LAUSR

Abstract The quality of “non-landslide” negative samples may result in unreasonable prediction results for machine learning (ML) models. The aim of this study is to improve the performance of ML models by perfecting the quality of “non-landslide” samples in landslide susceptibility modelling so as to produce more reliable susceptibility maps. The middle and lower reaches of Jinsha River basin (MLRJB) were chosen as the study area, and the elevation, slope aspect, curvature, lithology, distance to faults, slope of slope, slope of aspect, precipitation, land use, and NDVI were considered as predisposing factors for landslide susceptibility mapping. Firstly, three “non-landslide” samples are randomly selected from the low-slope area, landslide-free area and very low susceptibility area based on fractal theory (FT) model generation, and then three sample scenarios are constructed with 4445 landslide positive samples. Next, the performance of cross-application of three sample scenarios in the support vector machines (SVM) and naive Bayes (NB) models are compared and evaluated based on the statistical indicators such as accuracy, recall, precision, receiver operating characteristic (ROC) curve, and area under the ROC curve (AUC). The evaluation results show that the “non-landslide” negative samples generated on the basis of FT model are more reasonable and that the hybrid method supported by FT and ML models exhibits the highest prediction efficiency, around 94% overall accuracy produced by scenario-FT, followed by scenario-SS (87%) and scenario-RS (65%). Finally, with the validation of landslide and unstable slopes data, the landslide susceptibility map produced by the hybrid method composed of FT model and the SVM model is the ultimate output product for landslide prevention.