Enhancing landslide susceptibility mapping analysis through neighborhood feature aggregation and interpretable machine learning: Implications for disaster prevention in valley-type urban planning.

Geological disasters, as major natural hazards threatening regional sustainable development, have made risk assessment and planning governance core issues in global spatial safety management. This study, set against the backdrop of Shiquan County, a typical mountainous river-valley-type city in China, conducts landslide susceptibility prediction. We evaluated the performance of six machine learning models (LR, SVC, BP neural network, RFC, CatBoost, LGBM) in landslide susceptibility mapping. We proposed an optimization method based on "Neighborhood Feature Aggregation (NFA)". Among them, the CatBoost model performed the best. The study found that the proportion of geological disaster susceptibility zoning in Shiquan County exhibits a gradient distribution and shows a linear distribution along river valleys spatially. Based on the research findings, urban disaster prevention strategies are proposed from the perspective of social risk, providing references for other mountainous valley-type cities with characteristics similar to those of Shiquan County.