Uncertainty-aware ensemble learning and dynamic threshold optimization for landslide susceptibility mapping

Abstract

Landslides represent a prevalent and devastating geological hazard. Identifying areas susceptible to landslides is vital for disaster prevention and reduction. However, traditional models suffer from limited predictive accuracy, strong regularity in breakpoint selection for susceptibility zoning, and inconsistent predictions across different models, resulting in uncertainty in susceptibility assessment. To address these issues, this study proposes an innovative intelligent landslide susceptibility mapping approach that integrates ensemble learning, multi-model uncertainty analysis, and dynamic optimization. Focusing on Linxiang City, Hunan Province, China, this research synthesizes historical landslide inventories and field-identified unstable slopes as positive samples. Three base models were constructed: logistic regression (LR), random forest (RF), and graph neural network (GNN). Ensemble learning using the stacking method was applied to combine these models. The ensemble further incorporates prediction uncertainty estimation and multi-dimensional k-nearest neighbor (KNN) adjacency matrix. Utilizing an attention mechanism, the model dynamically integrates geographic features, environmental factors, and prediction outputs. The final output is a prediction model that synthesizes spatial structure information and prediction uncertainties. For susceptibility mapping, this study proposes a dynamic optimization approach combining Natural Breaks, Frequency Ratio, and Equal Interval methods, determining optimal threshold combinations through relative density distribution of landslide occurrences to enhance susceptibility classification rationality. Model performance was evaluated and compared using area under roc curve (AUC), where a larger AUC signifies higher predictive accuracy. The results show that the ensemble model outperformed all others with an AUC of 0.95, compared to the base models' AUCs of 0.82 (LR), 0.84 (RF), and 0.87 (GNN). This demonstrates that the ensemble learning methods that incorporate uncertainty achieve higher accuracy in risk identification than conventional models. The dynamic classification method also shows a better performance over conventional approaches in high-susceptibility classification precision and landslide density differentiation.