From pixels to policy: Multi-scale flood susceptibility mapping using interpretable machine learning for urban resilience

Urban flooding poses escalating risks globally as climate change intensifies extreme precipitation events and urbanization accelerates impervious surface expansion. This study develops a comprehensive, interpretable framework for high-resolution flood susceptibility assessment in Jeju City, South Korea—a volcanic island city where rapid urbanization is transforming porous basaltic terrain into flood-vulnerable urban landscapes. We integrated multi-resolution gridded geospatial data (10 m, 30 m, 100 m) with four tree-based machine learning algorithms (Decision Tree, Random Forest, XGBoost, CatBoost) and SHapley Additive exPlanations (SHAP) to create an interpretable flood susceptibility system aligned with administrative boundaries for evidence-based urban planning. Random Forest at 10 m resolution emerged as the optimal configuration (AUC = 0.919), balancing predictive accuracy with superior recall performance and generalization stability essential for policy applications. SHAP-based interpretation revealed elevation, proximity to water bodies, and slope as dominant flood drivers, with anthropogenic infrastructure factors showing secondary influence. District-level analysis uncovered pronounced spatial heterogeneity in flood mechanisms: reservoir-dominated patterns in Ildo districts, river-influenced dynamics in coastal Samdo areas, and drainage-constrained vulnerabilities in Yeongdam-dong. This heterogeneity underscores the limitations of uniform city-scale models and validates the necessity of localized risk assessment. The framework's alignment with administrative boundaries enables evidence-based policy formulation, supporting targeted interventions from green infrastructure deployment to drainage retrofitting. By demonstrating how interpretable machine learning can bridge high-resolution geospatial analysis with jurisdictional planning frameworks, this research advances both methodological rigor and practical applicability for climate-resilient urban governance. The methodology's modular design facilitates transferability to other topographically complex coastal and island cities worldwide.