INTEGRATING ARTIFICIAL INTELLIGENCE AND GEOGRAPHIC INFORMATION SYSTEMS FOR REAL-TIME URBAN FLOOD PREDICTION

Abstract

Urban flooding is among the most damaging and rapidly intensifying climate hazards, yet conventional hydrological models often lack the speed and spatial granularity required for real-time warning. This study proposes and evaluates an integrated framework combining machine learning with geographic information systems (GIS) to predict urban flood inundation in near real time. The framework fuses static geospatial layers—elevation, land use, drainage density, soil permeability—with dynamic inputs including rainfall intensity and antecedent moisture, using a machine-learning model to predict flood depth and extent at fine spatial resolution. We assembled a dataset from a mid-sized flood-prone city, integrating remote-sensing imagery, digital elevation models, and historical flood records. A gradient-boosting and neural-network ensemble was trained to classify flood-risk grid cells and estimate inundation depth, with GIS providing spatial feature engineering and visualization. Model performance was benchmarked against a traditional hydrological baseline. The integrated model achieved 0.89 accuracy and an AUC of 0.93 in identifying flooded cells, substantially outperforming the baseline, while reducing prediction latency to seconds per update. Feature-importance analysis identified rainfall intensity, elevation, and drainage density as dominant predictors. Spatial validation confirmed that predicted inundation zones aligned closely with observed flood footprints. The framework demonstrates that AI–GIS integration can deliver accurate, high-resolution, low-latency flood prediction suitable for early-warning systems. Contributions include a reproducible integration architecture, empirical performance benchmarks, and practical guidance for deployment in urban resilience planning. Limitations regarding data availability and transferability across cities are discussed.