Deep Learning-Based Rapid Flood Inundation Modeling for Flat Floodplains With Complex Flow Paths

Abstract

Flood inundation emulation models based on deep neural networks have been developed to overcome the computational burden of two-dimensional (2D) hydrodynamic models. Challenges remain for flat and complex floodplains where many anabranches form during flood events. In this study, we propose a new approach to simulate the temporal and spatial variation of flood inundation for a floodplain with complex flow paths. A U-Net-based spatial reduction and reconstruction method (USRR) is used to find representative locations on the floodplain with complex flow paths. The water depths at these locations are simulated using one-dimensional convolutional neural network (1D-CNN) models, which are well-suited to handling multivariate timeseries inputs. The flood surface is then reconstructed using the USRR method and the simulated flood depths at the representative locations. The combined 1D-CNN and USRR method is compared with a previously developed approach based on the long short-term memory recurrent neural network (LSTM) models and a 2D linear interpolation-based SRR method. Compared to the LSTM model, the 1D-CNN model is not only more accurate, but also takes less time to develop. Although both surface reconstruction methods take <1 s to produce an inundation map for a specific point in time, the USRR method is more accurate than the SRR method, leading to an increase of 5.6% in the proportion of correctly detected inundation area. The combination of 1D-CNN and USRR can detect over 95% of the inundated area simulated using a 2D hydrodynamic model but is 98 times faster.