SwinFlood: A hybrid CNN-Swin Transformer model for rapid spatiotemporal flood simulation

Deep learning-based flood prediction methods have demonstrated significant potential for rapid simulation and early warning of flood disasters. Existing flood surrogate models typically require developing diverse deep-learning architectures based on multiple flood-driving factors, making it challenging to apply these models to different flood scenarios within a consistent network architecture. The temporal resolution of predicted flood maps is also inherently constrained by input flood-driving factors. This paper conceptualizes flood modeling as the translation from coarse-grid to fine-grid flood maps and proposes a novel spatiotemporal flood simulation method termed SwinFlood. The flood-driving factors are unified into two-dimensional coarse-grid hydrodynamic features and fused with fine-grid static terrain features. Utilizing the Swin Transformer for deep feature extraction, the model ultimately outputs fine-grid flood maps. A multi-level model evaluation approach is implemented to systematically assess the performance of the SwinFlood model at global, local, and pixel levels. The proposed model is applied to a complex urban–rural catchment in the upper reaches of the Shenzhen River. Compared to physics-based models, the results demonstrate that the SwinFlood model effectively captures the spatiotemporal variations of water depth, velocity, and river discharge, achieving a speed-up ratio exceeding 1900. The SwinFlood model outperforms traditional purely CNN-based models with comparable parameters. This study provides an efficient and accurate deep-learning method for real-time flood simulation and management.