CA-DSUNet: A Novel Physics-Prior-Guided Dual-Stream Deep Coupling Network for Flood Inundated Area Forecasting in Emergent Support

Abstract

Rapid forecasting of spatiotemporal flood evolution is critical for emergency response. However, traditional hydrodynamic models face computational bottlenecks, while purely data-driven models lack physical constraints, frequently generating topologically implausible overspill in complex terrains. Here we propose CA-DSUNet, a physics-prior-guided dual-stream neural network for inundation forecasting. The architecture uses a lightweight, CPU-based cellular automata (CA) module to compute gravity-driven expansion envelopes from local topography, establishing an antecedent physical boundary. A dual-stream encoder then fuses a five-day historical flood sequence with static topographic constraints to perform nonlinear residual mapping strictly within this prior envelope, balancing hydrodynamic momentum capture with resistance to static-background overfitting. To supervise training, we couple signed distance fields with Otsu adaptive thresholding and cumulative precipitation to reconstruct continuous flood trajectories from sparse satellite observations. Evaluated on Poyang Lake and transferred to the East Dongting Lake basin, CA-DSUNet overcomes the identity mapping trap common to time-series models and suppresses disordered overflow. The model achieves an overall intersection over union (IoU) of 0.948 on Poyang Lake and maintains an IoU of 0.990 under cross-basin transfer, while significantly reducing false negatives at highly transient flood frontiers. By bounding deep feature extraction with fundamental physical constraints, this framework provides an accurate, generalizable, and computationally efficient solution for real-time flood forecasting.