FLO-SR: Deep learning-based urban flood super-resolution model
Hyeonjin Choi, Hyuna Woo, Minyoung Kim, Hyungon Ryu, Jun‐Hak Lee, Seungsoo Lee, Seong Jin Noh
Abstract
Urban flooding, intensified by both climate change and urbanization, requires high-fidelity and computationally efficient modeling frameworks for effective risk assessment and mitigation. This study presents FLO-SR, a deep learning-based super-resolution (SR) model, to enhance the spatial resolution of urban flood simulations while significantly reducing computational costs. FLO-SR leverages a convolutional neural network (CNN) to convert low-resolution (LR) flood maps into high-resolution (HR) outputs. The model was validated using two distinct flood events: Hurricane Harvey in Houston, Texas (synthetic scenario using bicubic interpolation) and an urban flood event in Portland, Oregon (physics-based simulation scenario). FLO-SR was evaluated in terms of image similarity, flood depth, and inundation extent. FLO-SR achieved accuracy improvements in both cases at scale factors of 2, 4, and 8×, with average RMSE reductions of 56.2, 32.4, and 10.7 % in Houston and 24.5, 33.8, and 44.1 % in Portland. However, performance at the 8× scale was limited due to challenges in reconstructing fine-scale flood features and spatial discontinuities in LR inputs. To address this, future improvements should incorporate hydrodynamic constraints and enhance model generalization. Despite these limitations, FLO-SR combined with physics-based modeling achieved up to 63 % and 45.7 % runtime reductions when reconstructing 2 m from 4 m and 4 m from 8 m simulations, respectively, highlighting its potential for real-time urban flood forecasting.