Urban flood risk projection under climate change: an approach based on explainable and optimized ensemble learning

Under climate change, urban floods occur more frequently and respond nonlinearly, complicating forward looking risk projection. Taking Beijing as a case, we assemble 12 indicators spanning hazard, exposure, and environment, and pair multiple RCP emission pathways with SSP socioeconomic scenarios. We develop an optimized Stacking ensemble with four heterogeneous base learners and a Bayesian-optimized LightGBM meta-learner, achieving ~92% test accuracy and more stable performance than single-model baselines. SHAP values and partial dependence plots quantify nonlinear effects and thresholds of key drivers (e.g. building density, population, GDP, slope, and extreme rainfall), enabling decision support under consistent RCP–SSP settings. Projections show a persistent center-to-southeast high-risk gradient, with stronger growth under RCP4.5 and RCP8.5. The lowest-risk class decreases from ~60.3% (2010s) to ~50% (2030s–2060s), indicating expansion of moderate-to-high risk zones. Scenario-based maps and interpretable responses provide targeted guidance for flood-control planning and resilience-oriented infrastructure investment.