A terrain-guided and physics-constrained dual-stream network for real-time high-resolution urban flood mapping.

Abstract

High-resolution (HR) flood inundation mapping is vital for emergency response, yet meter-scale hydrodynamic simulations remain computationally prohibitive over large domains. While deep learning-based super-resolution (SR) offers an efficient alternative, conventional architectures suffer from micro-topographic feature dilution, drainage channel disconnection, and unphysical high-ground ponding at large scaling factors. To overcome these challenges, this study develops a DEM-guided Dual-Stream Super-Resolution framework (DDS-SR) integrated with physics-informed soft regularizers. DDS-SR decouples DEM from low-resolution (LR) inundation depths via independent deep encoding streams and dynamically fuses terrain boundaries through a Cross-Modal Attention (CMA) mechanism. Furthermore, domain-specific physical constraints-slope consistency, water-surface continuity, and depth smoothness-are embedded to penalize hydraulic and topological violations. Systematic evaluations across flood scenarios in Houston demonstrate that DDS-SR markedly outperforms single-channel (FLO-SR) and early-concatenation (UNet-SR) baselines. The results indicate that the proposed approach achieves higher reconstruction fidelity and more stable physical behavior. Under an extreme 8× upscaling setting (8 m to 1 m), DDS-SR achieves an SSIM > 0.90, a median cross-sectional NSE of 0.92, and a low false alarm rate of 0.069, effectively eliminating staircase-like artifacts and over-smoothing. Coupling coarse-resolution hydrodynamic simulation with DDS-SR inference generates a full-domain 1 m inundation map in an estimated 1.98 minutes, achieving an approximately 501-fold speedup over native 1 m hydrodynamic modeling. Cross-domain validation (Houston to Southern Brooklyn) further confirms its spatial transferability. DDS-SR provides an operational, physically consistent paradigm for real-time urban flood management.