High-resolution urban flood risk mapping using indicator cokriging integrated with physical modelling and observational data

Abstract

This study presents a geostatistical data fusion framework using indicator cokriging (ICK) to enhance urban flood risk assessment by integrating observational data, such as citizen-generated reports or limited sensor networks, with high-resolution simulation outputs. Indicator cokriging assimilates binary flood indicators and critical auxiliary variables, specifically simulated inundation depth and distance-to-flood raster, which captures proximity to inundation boundaries, to generate probabilistic flood maps. The framework's performance was evaluated through synthetic experiments and a field-derived data application in the Oncheon-cheon catchment, South Korea. Synthetic tests showed that ICK achieved classification accuracy improvements of 16% to 27% over traditional indicator kriging under low to moderate observation densities, and maintained stable performance even under substantial rainfall input uncertainty (a 50% decrease to a 100% increase relative to baseline conditions). In the real-world scenario, the model attained a perfect hit rate (1.00), although with moderate overall accuracy (0.64) due to localised false positives near documented flood boundaries, reflecting cautious spatial extrapolation from limited data. The proposed framework demonstrates scalability and adaptability, offering practical and robust flood mapping capabilities that are essential for resilient urban water management and show strong potential for broader applicability across diverse urban contexts.