Geospatially interpretable long-term flood risk assessment using integrated weighting and LightGBM–SHAP

Long-term flood risk assessment is vital for identifying persistent hazard-prone areas and supporting risk-informed spatial planning in rapidly urbanising regions. This study developed an interpretable geospatial framework integrating AHP, LightGBM–SHAP, and reliability-weighted game-theoretic weight fusion for the Beijing–Tianjin–Hebei (BTH) region using annual flood-related dataset for 2001–2025. Negative sample selection was refined using HAND-based low-likelihood screening and buffer-based exclusion around flood-labelled pixels. The LightGBM component showed strong internal predictive performance, with AUC, TSS, F1-score, and Kappa values of 0.92, 0.69, 0.84, and 0.69, respectively. Independent validation of the final flood risk index for 2022–2025 further confirmed its ability to identify flood-prone areas, with annual AUC values ranging from 0.63 to 0.83. High-risk areas were mainly concentrated in the central-eastern and southeastern plains, especially along the Beijing–Langfang–Tianjin–Cangzhou corridor. Flood risk fluctuated markedly and showed a weak increasing trend, with stable and increasing-risk areas accounting for 87.8% and 11.7%. Topographic factors dominated risk patterns, followed by hydrometeorological, urbanisation, and ecological factors. The framework supports long-term flood risk mapping, driver attribution, and spatial prioritisation of risk reduction measures.