Insights into landslide susceptibility in different karst erosion landforms based on interpretable machine learning

The aim of the present study was to assess differences in the conditioning factors and the performance of landslide susceptibility mapping (LSM), employing the SHapley Additive exPlanations (SHAP) model to gain profound insights into the intrinsic decision-making mechanism of LSM in diverse landforms. Two typical karst erosion landforms were selected as the research areas. Based on 15 conditioning factors, LSMs for the two areas were developed using the Bayesian optimization random forest (RF) and eXtreme Gradient Boosting (XGBoost). The SHAP model was used to explore the landslide formation mechanisms from both global and local perspectives. The results show that the area under the curve (AUC) values of the XGBoost models were 0.791 and 0.761, and the AUC values of the RF models were 0.844 and 0.817, in the two different landform areas, respectively. The RF model's accuracy was higher than that of the XGBoost model in both regions. In the low-elevation hills area, the primary three conditioning factors were identified as slope, topographic relief and distance from the river. Conversely, in the microrelief and mesorelief low mountain area, the predominant conditioning factors were elevation, distance from the river and distance from the road. Both karst landform areas exhibited a high sensitivity to the distance from the river, indicating its significant interaction with other factors contributing to landslide occurrences. Notably, the RF model demonstrated superior performance compared to the XGBoost model, rendering it a more suitable choice for conducting landslide susceptibility mapping research in karst erosion landform areas. In the present study, a comprehensive explanatory framework based on the RF-SHAP model was proposed, which enables both global and local interpretation of landslides in various karst landscapes. Such an approach explores the intrinsic decision-making mechanism of the model, enhancing the transparency and realism of landslide susceptibility prediction results. This work has three stages. In Stage A, compile historical landslide data, create geographic spatial databases, and partition training and test sets in a 7:3 ratio; in Stage B, construct landslide susceptibility mapping) (LSM) models using XGBoost and Random Forest algorithms and assess their performance with metrics like area under the curve (AUC) to determine the optimal model for predicting landslides; and in Stage C, incorporate SHapley Additive exPlanations (SHAP) model to rank factors and provide global interpretation for the entire region's landslides and local interpretation for a single landslide case.image