Mapping unit selection in landslide susceptibility prediction: A comprehensive explainable framework integrating multi-method interpretability

Current research on revealing the similarities and differences of machine learning models in different mapping units for landslide susceptibility frequently employs only accuracy evaluation methods, yet seldom explores interpretability in greater depth. Moreover, single-type interpretability methods are limited in terms of applicability, and their explanatory capacity is not comprehensive. In response to these challenges, a comprehensive interpretability framework has been proposed to reveal the intrinsic mechanisms of machine learning models for landslide susceptibility prediction (LSP) across various mapping units. Pingyang County is selected as the case study, we first select 16 landslide environmental factors and 225 landslide samples. Then, the slope unit-random forest (RF) model and grid unit-RF model are performed for LSP. Finally, a comprehensive interpretability method is proposed to analyze the explainability of different machine learning models for LSP. This method encompasses mean squared error (MSE) feature importance analysis, single-feature analysis based on partial dependence plot (PDP) and individual conditional expectation (ICE) methods, interaction analysis based on area under the receiver operating characteristic curve (AUC), and full-feature Shapley additive explanations (SHAP) value analysis for individual samples in an exploratory manner. Results indicate that: (1) Both models achieve AUC accuracies above 0.9, indicating highly accurate predictions. The comprehensive interpretability method reveals that the LSP results of the slope unit-RF model are more consistent with geological data; (2) Environmental factor importance is similar in both models; however, the terrain moisture index and normalized difference vegetation index have a greater impact in the slope unit-RF model. This study reveals that the slope unit-RF model exhibits larger SHAP values for individual samples through a comprehensive single-sample full-feature SHAP value analysis. By incorporating real-world case studies, the analysis of explainability results indicates that the model’s predictions are more reliable for terrain assessments within the slope unit framework.