Spatiotemporal flood depth and velocity dynamics using a convolutional neural network within a sequential Deep-Learning framework
Mohamed M. Fathi, Zihan Liu, Anjali M. Fernandes, Michael T. Hren, Dennis O. Terry, C. Nataraj, Virginia Smith
Abstract
Computational hydrodynamic models support river science and management. However, current physics-based models face computational challenges; they require extensive processing time for large-scale two-dimensional flood simulations. Despite the success of Deep Learning (DL) applications in generating inundation maps, accurate prediction of unsteady flood hydrodynamic maps remains challenging. This paper compares traditional approaches to a novel DL approach, which integrates convolutional neural networks with long short-term memory, to deliver precise, rapid, and continuous simulation of the spatiotemporal dynamics of river floods. This is the first DL framework able to generate essential hydrodynamic variables: water depth, velocity magnitude, and flow direction maps. Water depth and velocity magnitude predictions across the testing dataset are robust, with average RMSE of 0.14 m and 0.02 m/s, respectively. The DL predictions are 415 times faster compared to traditional computational approaches, representing a paradigm shift in hydrodynamics modeling that advances long-term flood simulations and resilient river management. • This paper compares a novel hydrodynamic DL framework to a physics-based model. • CNNs are integrated with LSTMs to capture spatiotemporal dynamics of flood events. • The framework enables rapid and reliable generation of hydrodynamic parameters. • The approach generates water depth, velocity magnitude, and flow direction maps.