An integrated deep learning framework enables rapid spatiotemporal morphodynamic predictions toward long-term simulations

Abstract

Physics-based morphodynamic modeling is essential for advancing river management science and understanding Earth's geomorphological evolution processes. However, their computational demands and long processing times hinder long-term applications. This paper introduces and tests a robust Deep Learning (DL) framework that opens the door to overcoming these challenges through integrating convolutional neural networks (CNNs) with long short-term memory (LSTM) architectures, trained using outputs from the physics-based HEC-RAS model. This framework facilitates rapid and continuous spatiotemporal predictions of hydrodynamic parameters and morphodynamic responses of flood events. Hydrodynamic predictions showed strong performance across the testing dataset, with mean RMSEs of 0.15 m and 0.04 m s−1 for water depth and flow velocity, respectively. Bed change predictions also demonstrated promising results, with normalized RMSE of 27 % and R2 of 0.93. This novel approach generates predictions 4700 times faster than traditional physics-based computational models, representing a paradigm shift in long-term river evolution simulations and opening new opportunities for fluvial morphodynamic modeling.