Ensemble Machine Learning-Based Feature Selection for Flood Susceptibility Mapping Under Climate and Land Use Change Scenarios

Effective flood susceptibility mapping (FSM) is critical for risk-informed environmental planning and climate adaptatioAn strategies. However, the complexity of flood-influencing factors limits the efficacy of the FSM algorithms. This study presents an in-depth comparison of feature selection techniques using ensemble machine learning algorithms to identfiy key factors influencing flooded areas in South Korea. The analysis incorporated historical rainfall data (1980–2023), simulated Land Use and Land Cover (LULC) scenarios, and climate projections based on CMIP5 and CMIP6 (RCP4.5, RCP8.5, SSP245, SSP585). The results showed that Variance Inflation Factor (VIF) performed best in feature selection by reducing redundancy while retaining essential hydrological and topographical predictors. Model performance was evaluated using multiple metrics, with Gradient Boosting (GB) achieving the highest accuracy (ROC-AUC: 0.93), followed by Random Forest (RF) (ROC-AUC: 0.875) and Extra Trees (ET) (ROC-AUC: 0.85). FSM outputs revealed that GB classified over 12% of the region as high flood risk, particularly in densely urbanized and low-lying areas, whereas RF and ET identified broader moderate-risk zones. Future projections suggest increased flood exposure due to intensified monsoon rainfall and urban expansion. While GB performed best under extreme climate conditions, RF provided reliable medium-impact predictions. This study introduces a novel approach by integrating heterogeneous data (multi-scenario climate and land-use projections) into ensemble learning to reduce prediction bias, enhacing the precision and and robustness of FSMs. These findings are crucial for adaptive flood risk management, spatial planning, and informed decision-making amid dynamic environmental changes.