Bed Topography Inference from Velocity Field Using Deep Learning

Measuring bathymetry has always been a major scientific and technological challenge. In this work, we used a deep learning technique for inferring bathymetry from the depth-averaged velocity field. The training of the neural network is based on 5742 laboratory data using a gravel-bed flume and reconstructed velocity fields, namely the topographies were obtained from real-world experiments, and the velocity fields were estimated using a statistical model. To examine the predictive power of the proposed neural network model for bathymetry inference, we applied the model to flume experiments, numerical simulations, and field data. The results showed the model properly estimates topography, leading to a model for riverine bathymetry estimation with a 31.3% maximum relative error for the case study (confluence of the Kaskaskia River with the Copper Slough in east-central Illinois state, USA).