Data-Driven Flood Emulation: Speeding Up Urban Flood Predictions By Deep Convolutional Neural Networks

Computational complexity has been the bottleneck for applying physically based simulations in large urban areas with high spatial resolution for efficient and systematic flooding analyses and risk assessment. To overcome the issue of long computational time and accelerate the prediction process, this paper proposes that the prediction of maximum water depth can be considered an image-to-image translation problem in which water depth rasters are generated using the information learned from data instead of by conducting simulations. The proposed data-driven urban pluvial flood approach is based on a deep convolutional neural network trained using flood simulation data obtained from three catchments and 18 hyetographs. Multiple tests to assess the accuracy and validity of the proposed approach were conducted with both design and real hyetographs. The results show that flood prediction based on neural networks use only 0.5% of the time compared with that of physically based models, with promising accuracy and generalizability. The proposed neural network can also potentially be applied to different but relevant problems, including flood analysis for flood-safe urban layout planning.