An interpretable hybrid model for predicting step-like landslide displacement: a case study in the Three Gorges Reservoir

Predicting landslide displacement is essential for ensuring reliable landslide warning and forecasting. Artificial intelligence models have been used to predict the displacement of step-like landslides. However, the quantitative assessment of the influence of rainfall and reservoir water level on landslide deformation is still challenging, and the interpretability of models for step-like landslides could be further enhanced. This study proposes an interpretable hybrid model by combining Light Gradient Boosting Machine (LightGBM) with Partial Dependence Plot (PDP) and Shapley Additive exPlanations (SHAP). LightGBM-SHAP-PDP is used to predict displacement and quantitatively interpret the influential factors on displacement. LightGBM is used to predict monthly displacement of Baishuihe landslide, and SHAP and PDP jointly explain the relationship between input parameters and landslide displacement from both global and local perspectives. LightGBM-SHAP-PDP are compared with particle swarm optimization- support vector machine (PSO–SVM), extreme learning machine (ELM), and random forest (RF). The results indicate that LightGBM-SHAP-PDP outperforms other models. When the reservoir water level declines to 145 m or cumulative displacement over the past 2 months exceeds 50 mm, the displacement increases significantly. LightGBM-SHAP-PDP has excellent prediction performance and interprets quantitatively the relationship between input and displacement, which provides reference for landslide warning threshold.