Interpretable landslide susceptibility evaluation using a BP neural network model optimized by multiple algorithms

Current landslide susceptibility evaluation methods predominantly rely on statistical techniques and machine learning models, both of which are prone to errors related to sample quality and parameter selection. The complexity of model training and the uncertainty of prediction results limit the broader application and development of machine learning models in this field. To address these issues, this study evaluates landslide susceptibility in Zigui County, Yichang City, using a backpropagation (BP) neural network optimized by three algorithms: Bayesian optimization, sparrow search algorithm (SSA), and gorilla troops optimization (GTO). Common susceptibility evaluation factors, such as elevation, NDVI, and stratigraphic lithology are used as inputs. Three hybrid models— Bayesian-BP, SSA-BP, and GTO-BP are constructed and trained using the respective algorithms. The models are comprehensively evaluated using statistical standards such as AUC value, F1 score, and accuracy, with K-fold cross-validation for robustness. Additionally, SHAP (shapley additive explanations) is used to enhance interpretability of the models. The results show that the accuracy, precision, F1 score, and other indicators of the three optimized models are significantly higher than those of the standalone BP model, confirming the effectiveness of the optimization strategies. Among the optimized models, the GTO-BP model exhibits superior overall performance and is better suited for landslide susceptibility evaluation in Zigui County. In addition, the SHAP analysis provides interpretable results, offering valuable technical support for future landslide prevention and mitigation efforts in the region.