Chao Tong

Frequent flooding induced by climate change has been posing growing threats to human lives and property worldwide. Most of 2D and 3D physics-based hydrodynamic (PBH) models are generally too time-consuming to support decision-makers to react in time during an urgent flood event. Therefore, highly efficient and accurate flood inundation prediction models are urgently needed to improve emergency response capabilities. In this paper, a novel model based on a 1D deep residual neural network with Squeeze-and-Excitation (ResNet-SE) attention is proposed for fast forecasting dynamic evolution of 2D fluvial flood inundation. Different from widely-used 2D image-oriented DL models, our model directly consumes original 1D nodal hydraulic output vectors from TELEMAC-2D, avoiding interpolation errors introduced by mesh rasterization. The results of the well-trained ResNet-SE model compare with those from the TELEMAC-2D and a convolutional neural network (CNN) model. As indicated by several quantitative assessment matrices, the present model outperforms CNN model and conforms closely to the TELEMAC-2D in high precision. Furthermore, we demonstrate that ResNet34-SE (34 layers) stands out distinctly among ResNet-SEs of 18 and 50 layers in flood inundation prediction. The average $MAE$ and $RMSE$ metrics of ResNet34-SE under three unseen flood events reach 0.04623 m and 0.08052 m in water depth prediction. As for spatial flood extent assessment, this model also delivers impressive performance, with the average $Precision$, $Recall$, $F1$, and $Accuracy$ scores reaching 96.92%, 99.69%, 98.28%, and 98.55%, respectively. Furthermore, current deep-learning (DL) model achieve a speedup of 300 to 400 times over traditional PBH models, enabling the generation of 24-h flood evolution maps within seconds. The proposed ResNet-SE model demonstrates good simulating accuracy of fluvial flood dynamics for proactive emergency warning systems and flood risk mitigation planning within the Lushui River basin.