Comparative Evaluation of Deep Learning–Based Super–Resolution Models for Urban Flood Mapping

Abstract

ABSTRACT Urban flood forecasting benefits from high‐resolution inundation maps, but fine‐grid hydrodynamic simulations are computationally costly. We compared three CNN–based super–resolution (SR) models, ResUNet, EDSR, and RCAN, for downscaling physics–based simulations in downtown Portland, Oregon, using paired flood maps at 1 m (HR) and both 4 and 8 m (LR). Performance was assessed using image level metrics (PSNR, SSIM) and flood specific indicators: CSI for flood extent, RMSE for water depth accuracy, and a depth–based severity classification. At 4× upscaling, all SR models outperformed the LR baseline; RCAN performed best (PSNR +57%, SSIM +31%, RMSE −73%, CSI +53%), followed by EDSR (PSNR +50%, SSIM +30%, RMSE −64%, CSI +45%) and ResUNet (RMSE −55%, CSI +40%). Analysis of class–wise recall showed RCAN leading for non–flood (98.06%, +6.59 pp) and severe flood (96.48%, +16.90 pp), while EDSR led for mild flood class (97.95%, +6.49 pp). Errors were most pronounced along wet–dry boundaries and in complex urban geometries, where RCAN and EDSR reduced error magnitude more effectively than ResUNet. Models with larger numbers of parameters required longer training times. Furthermore, the computational cost further increased with more training epochs and especially at 4× upscaling relative to 8×, reflecting differences in model complexity and scaling configuration. Taken together, these findings support SR as a practical complement to physics–based modeling for real time forecasting and planning, while also providing guidance for selecting architectures under varying computational budgets.