Ensemble Landslide Susceptibility Modeling Based on Spatial-Matrix Coupling and Uncertainty Analysis

Landslides are common and highly destructive geological hazards, and accurately identifying landslide-prone areas is of great significance for disaster prevention and mitigation. To address the limitations of traditional landslide susceptibility models—such as insufficient generalization capability, strong spatial heterogeneity, and high predictive uncertainty—this study proposes an integrated landslide susceptibility modeling approach that incorporates spatial matrices and uncertainty analysis. The proposed method ensembles four base models, including Logistic Regression, Random Forest, Maximum Entropy, and a Graph Neural Network. Node-level uncertainty is quantified using prediction variance. Three types of adjacency matrices—geographical, environmental, and prediction-based—are constructed and adaptively fused via an attention mechanism. Within a two-layer graph convolutional network framework, multi-source information is jointly propagated and probability estimates are calibrated. A case study in Linxiang City, Hunan Province, China demonstrates that the proposed model achieves an AUC of 0.937 and a landslide identification rate of 94.96%, significantly improving the accuracy and reliability of landslide recognition.