A Comparative Study of Model Interpretability Considering the Decision Differentiation of Landslide Susceptibility Models

The “black-box” nature of machine learning (ML) and deep learning (DL) models has raised concerns about the trustworthiness of landslide susceptibility mapping (LSM) results among users. Existing studies have applied many techniques to interpret LSM models, but they predominantly focused on individual model interpretations, lacked comparisons of interpretation results across different models, and failed to fully explore the potential of explainable artificial intelligence (AI) techniques in LSM. This study develops an innovative model interpretation framework based on the Shapley additive explanation (SHAP) method and different ML and DL models, to analyze the decision mechanisms differences and discuss the geospatial heterogeneity of landslide conditioning factors (LCFs). A geospatial database is constructed, including historical landslides, 16 common LCFs, and three earthquake-related LCFs for two study areas: Zigui and Jiuzhaigou. The data are then divided into training and testing sets in a 7:3 ratio for four models: random forest (RF), extreme gradient boosting decision tree (XGBoost), residual network, and densely connected convolutional networks (DenseNets). Finally, global and local interpretations are provided using the SHAP method. The analysis indicates that: 1) XGBoost consistently outperforms the other models in both study areas, achieving Kappa coefficient (Kappa), overall accuracy (OA), and area under the receiver operating characteristic curve (AUC) values of 0.9416, 0.9738, and 0.9757 for Zigui, and 0.8525, 0.9337, and 0.9312 for Jiuzhaigou and 2) the global interpretation shows that the same LCFs play different roles in the XGBoost and DenseNet, reflecting different decision mechanisms among LSM models. Moreover, local interpretations demonstrate that the same LCFs contribute differently in the two areas, highlighting the geospatial heterogeneity in LSM.