Surrogate-based optimization of land-use-informed LID strategies for enhanced urban flood resilience

Urban flooding, driven by rapid urbanization and climate change, poses a critical challenge to resilient urban development. Although low-impact development (LID) practices are effective for mitigating flood hazards, identifying optimal LID combinations across heterogeneous land-use settings requires computationally efficient decision-support approaches. This study presents an optimization framework that couples a personal computer storm water management model (PCSWMM)-driven machine learning surrogate model with a genetic algorithm (GA) to identify optimal land-use-specific combinations of LID facilities. PCSWMM rainfall-runoff simulations were used to train the surrogate model, enabling rapid and accurate predictions of runoff responses under diverse LID combination scenarios. The surrogate was subsequently coupled with a GA to search for optimal combinations that efficiently minimize runoff volume. The framework was applied to a highly urbanized catchment in Gwangju, South Korea. The analysis considered five LID types across five impervious land-use categories and rainfall conditions spanning return periods of 5–500 years, storm durations of 1–12 h, and LID coverage of 3–9% of the impervious area. Optimal LID combinations were strongly influenced by land-use type and storm duration and were comparatively less sensitive to return period; detention facilities performed best for short-duration storms, whereas infiltration facilities excelled during longer events. Even during an extreme storm event, the optimized combinations reduced the inundation area by up to 20%. By integrating geospatially explicit land-use information with advanced computational modeling and optimization, the proposed framework provides a scalable decision-support tool for sustainable urban flood management.