Predicting river bathymetry for data sparse regions with a GANs model

River bathymetry data are needed for studying and simulating fluvial systems, but these data are not easy to collect. Previous studies have attempted to reconstruct river bathymetry with inverse hydraulic simulations. However, the results were unstable, and the information related to flowrates and velocities, which are necessary for inverse hydraulic simulations, are even harder to get than bathymetry data. This study attempts to address the scarcity of bathymetry data by applying a deep learning approach to predict river bathymetry in data sparse regions. Generative adversarial neuron networks (GANs) is a deep learning model developed to generate realistic data. One of the common applications of GANs is to augment insufficient datasets with limited training data. This study takes the advantage of GANs to generate realistic bathymetry data for data sparse regions. The model is designed to use simple river geometry features, such as channel width, thalweg position, and the lowest elevation, to generate a 2D cross-section profile. The model is validated on reaches at different sites including the Mississippi River in Minnesota and the Brazos River in Texas. The root mean square error (RMSE) in elevation at each generated cross-section is used to evaluate the performance of the GANs model in terms of the river geometry. The perimeter and cross-section area which are critical hydraulic features are also considered. The GANs model shows better performance in validations than interpolations methods. Specifically, the cross-section perimeters and areas are within 5% error. The RMSE normalized by the vertical range of the dataset is found to be less than 20%. In comparison, the interpolation methods produce 10% error for hydraulic features and up to 30% error for geometric features.