Author(s): Yaoxing Liao; Zhaoli Wang; Xiaohong Chen; Chengguang Lai

Linked Author(s):

Keywords: Convolutional neural network; Deep learning; Fast simulation; Urban pluvial flood

Abstract: Rapid prediction of urban floods is crucial for disaster prevention and mitigation. However, physics-based models often demand significant computational time, making them less suitable for time-sensitive scenarios. This study explores a deep learning (DL) approach employing convolutional neural network (CNN) combined with physics-based model for fast urban flood prediction. The results indicate that: 1) The inundation water depths predicted by the CNN model closely match those predicted by the physics-based model, with average PCC, MAE, and RMSE metrics during a test rainstorm reaching 0.983,0. 020 m, and 0.086 m, respectively. 2) The CNN model accurately reproduces the trend of water depth in each grid cell over time. 3) The predictive performance of the CNN model surpasses that of extreme gradient boosting (XGBoost) model, followed by multi-objective random forest (MORF) model and K-nearest neighbor (KNN) model. 4) The computation speed of the CNN model is extremely fast, and is 600 times faster than that of the physics-based model. The CNN model serves as a powerful surrogate model for rapid simulation of urban pluvial floods, offering a reference for the utilization of DL in early warning and mitigation of urban flood disasters.

DOI: https://doi.org/10.3850/iahr-hic2483430201-488

Year: 2024