Comparative study on landslide susceptibility of different evaluation units based on LightGBM-SHAP

Abstract

Abstract Landslide is a common natural disaster, which has a serious impact on human life, property safety and socioeconomic development. Landslide susceptibility zoning can predict the spatial distribution of landslide occurrence probability. Based on grid units, slope units and terrain units, this study explore the influence of different evaluation units on regional landslide susceptibility zoning. Taking Yunyang County as a case study, 15 influencing factors such as elevation, slope and curvature were selected to establish a geospatial database, and the light gradient boosting machine (LGBM) algorithm was used to const-ruct the landslide susceptibility model (LSM). The results show that the accuracy of LSM constructed by different evaluation units is diffe-rent. Among them, the LGBM model based on grid units has the highest accuracy, with an accuracy of 0.7589, F1-Score of 0.7453, and the area under curve (AUC) values in training data set and verification data set were 0.8998and 0.8099, respectively. In addition, SHaply Additive ExPlanation (SHAP) is used to explain the model. The global interpretation shows that elevation, distance from river and distance from road have great influence on landslide in the study area. Local interpretation found that elevation, distance from the river and distance from the road have a greater impact on Jiuxianping landslide. This study can provide scientific reference for LSM construction and disaster prevention.