An interpretable model for the susceptibility of rainfall-induced shallow landslides based on SHAP and XGBoost

Abstract

The machine-learning “black box” models, which lack interpretability, have limited application in landslide susceptibility mapping. To interpret the black-box models, some interpretable machine learning algorithms have been proposed recently. Among them is SHaply Additive ExPlanation (SHAP), which has attracted much attention because of its ease of operation and comprehensiveness. In this study, a novel interpretable model based on SHAP and XGBoost is proposed to interpret landslides susceptibility evaluation at global and local levels. The established evaluation model provided 0.75 accuracy and 0.83 AUC value for the test sets. The global interpretation shows that the peak rainfall intensity and elevation are the dominant factors that influence the occurrence of landslides in the study area. The combination of local interpretation and field investigations can provide a comprehensive framework for evaluating designated landslides, and it can also be used as a reference for preventing and managing the hazards of landslides.