Deep learning surrogate models for spatiotemporal prediction of coastal flooding inundations in Tianjin, China

The coastal region of Tianjin, China. Due to the sea-level rise under climate change and the increase in extreme weather events, this study aims to simulate the spatiotemporal variation process of coastal flooding using deep learning surrogate models. Four models—U-Net, CNN-LSTM, ConvLSTM, and CNN-Transformer—were developed and evaluated. The training dataset was generated using the LISFLOOD-FP hydrodynamic model under extreme sea-level scenarios. These models were applied to predict the spatiotemporal dynamics of inundation extent and water depth in Tianjin. Model performances were compared based on prediction accuracy and efficiency. All models achieved high accuracy using only DEM, land cover, coastline, and sea level time series as inputs. The U-Net model showed the best performance (MAE = 0.0125 m, RMSE = 0.0486 m, R² = 0.9935), with 98.52 % classification accuracy for flood severity. This study highlights that deep learning models offer high computational efficiency, strong predictive accuracy, and a streamlined modeling process, enabling rapid large-scale spatiotemporal simulations once trained. Furthermore, the model exhibits strong generalization capability across different extreme sea-level scenarios. This study not only provides a novel data-driven approach for simulating coastal flooding dynamics but also offers valuable insights into the flood risk response in low-lying coastal cities under climate-induced sea-level rise, supporting early warning systems and adaptive flood management strategies. • Coastal flood inundation processes are simulated using a hydrodynamic model. • Four deep learning models are developed to serve as surrogates for the hydrodynamic model. • U-Net shows the best spatiotemporal prediction performance for inundation depth. • Deep learning models significantly improve simulation efficiency.