Enhancing urban flood risk assessment: A PCA-integrated deep learning surrogate for hazard and damage prediction

This study proposes an integrated deep-learning framework for rapid urban flood hazard mapping and damage prediction from compact rainfall descriptors. A simulation-based reference database linked synthetic rainfall scenarios with grid-based flood-depth maps generated by a validated rainfall-runoff and inundation model and category-specific damage indicators derived from damage assessment. For hazard mapping, a principal component analysis-integrated deep neural network predicted retained component scores and reconstructed full inundation maps through inverse transformation. For damage prediction, two pathways were compared: a hybrid pathway using surrogate-generated inundation maps and a latent-assisted deep neural network using principal component analysis-derived inundation features. Across real rainfall events, refinement reduced the mean root mean square error from 0.11 to approximately 0.09 m and increased the mean Nash-Sutcliffe efficiency from 0.48 to approximately 0.57, although further improvement remains necessary under complex observed conditions. The latent-assisted damage pathway achieved a mean normalized root mean square error of 0.02 and a mean coefficient of determination of 0.86 across ten damage indicators, outperforming the hybrid pathway. Overall, the framework reproduced dominant inundation patterns and provided reliable scenario-level damage estimates, offering a computationally efficient basis for rapid flood assessment and forecast-linked decision support.