A novel deep learning method for urban flood prediction

The acceleration of urbanization and the occupation or destruction of vast cultivated land and gardens in the process of urban construction diminish the natural capacity of urban areas to absorb and regulate rainwater. Consequently, urban flooding risk significantly escalates. Accurate prediction of flood maximum depth and flood arrival time is critical for minimizing casualties and property damage. Conventional flood prediction methods based on hydrology and hydrodynamics are renowned for their precision but are often hindered by computational complexity and low efficiency. Moreover, due to the complexity of generating datasets, the available data for flood prediction is limited. Furthermore, cumulative errors may arise during the flood routing prediction process, leading to a gradual decline in prediction accuracy over time. To solve the above problems, this paper proposes to use Generative Adversarial Networks to augment the original dataset. Additionally, a novel approach is introduced that combines GAN with Squeeze-and-Excitation (SE) to enhance the perception ability of important information and provide ample sample data for precise predicting the urban flood routing processes. In response to the problems of slow calculation speed and high demand for computational resources in traditional hydrodynamic prediction of flood trends, this paper proposes an Unet-LSTM model combined with a Long Short Term Memory network and an Unet-Attention model based on an attention mechanism to address the cumulative error problem in the frame-by-frame prediction process of urban flood routing trends. Comparing the above two models with the Unet model, the Unet-Attention model focuses on more critical information in the data, while the Unet-LSTM model learns temporal features between data. The experiment shows that the method proposed in this paper can improve the accuracy of prediction, predict the flood routing process of longer time series, and reduce the cumulative error problem in the prediction process.