Explainable AI for Landslide Susceptibility Modeling in Bavaria: Integrating Neural Networks and SHAP Analysis

Abstract

Landslides are a significant threat to infrastructure, livelihoods, and lives, with their occurrence and magnitude expected to rise due to climate change. As demographic and migration dynamics lead to more people living in vulnerable areas, understanding and predicting landslide susceptibility becomes increasingly important to avoid economic, ecologic, and human losses, especially in mountainous regions. Hence, this thesis presents a state-of-the-art landslide susceptibility model, using slope units for enhanced terrain segmentation and a dense neural network for accurate susceptibility prediction. The findings are illustrated in a comprehensive susceptibility map and are further interpreted using SHAP, a game-theoretic approach that provides both global and local explainability of the influence of various features within the model . The model achieves a ROC-AUC score of 0.953 and a Precision-Recall AUC of 0.844, indicating its high predictive performance. SHAP analysis reveals that geological features, soil types, and land cover are the most influential features in predicting landslide susceptibility, while stream power index and flow accumulation play a minor role. The susceptibility is notably high in slopes ranging from approximately 700 m to 1500 m in altitude, as well as slopes with a higher variability in slope angles. When benchmarked against an existing, yet methodologically vague and incomplete, susceptibility map, the results and assessed metrics suggest that the proposed model performs well, particularly in predicting deep landslides compared to shallow slope failures. However, the application of the model to an updated landslide inventory reveals challenges. The predicted susceptibility values for areas with new events are relatively low, indicating a potential bias in the inventory containing data structures that the model has not yet learned from. Overall, the susceptibility map shows robust performance based on the available data but might encounter generalization issues when considering a broader inventory, especially since the current dataset appears to have a spatial selection bias.