Research on machine learning forecasting and early warning model for rainfall-induced landslides in Yunnan province

Abstract

Abstract Landslides are highly destructive geological disasters that pose a serious threat to the safety of people’s lives and property. In this study, historical records of landslides in Yunnan Province, along with eight underlying factors of landslide (elevation, slope, aspect, lithology, land cover type, normalized difference vegetation index (NDVI), soil type, and average annual precipitation (AAP)), as well as historical rainfall and current rainfall data were utilized. Firstly, we analyzed the sensitivity of each underlying factor in the study area using the frequency ratio (FR) method and obtained a landslide susceptibility map (LSM). Then, we constructed a regional rainfall-induced landslides (RIL) probability forecasting model based on machine learning (ML) algorithms and divided warning levels. In order to construct a better RIL prediction model and explore the effects of different ML algorithms and input values of the underlying factor on the model, we compared five ML classification algorithms: extreme gradient boosting (XGBoost), k-nearest neighbor (KNN), support vector machine (SVM), logistic regression (LR), and random forest (RF) algorithms and three representatives of the input values of the underlying factors. The results show that among the obtained forecasting models, the LSM-based RF model performs the best, with an accuracy (ACC) of 0.906, an area under the curve (AUC) of 0.954, a probability of detection (POD) of 0.96 in the test set, and a prediction accuracy of 0.8 in the validation set. Therefore, we recommend using RF-LSM model as the RIL forecasting model for Yunnan Province and dividing warning levels.