CP-2021-13
Deep learning technique for fast inference of large-scale riverine bathymetry
Ghorbanidehno, Hojat, Jonghyun Lee, Matthew Farthing, Tyler Hesser, Eric F. Darve, and Peter K. Kitanidis
Advances in Water Resources 147, 103715, https://doi.org/10.1016/j.advwatres.2020.103715 (2021)
Riverine bathymetry is of crucial importance for shipping operations and flood management. However, obtaining direct measurements of depth is not always easy. Conversely, with recent advances in sensor technology, indirect measurements can be obtained and used to estimate high-resolution river bed topography. Physics-based inverse modeling techniques have been used to estimate bathymetry using indirect measurements like flow velocity at the surface. However, these methods are computationally expensive for large-scale problems. Recently, deep learning has opened a new door toward knowledge representation and complex pattern identification in many fields; however, these techniques have not been used for high-dimensional riverine bathymetry problems since they require a large amount of data in the training phase to have a good estimation performance that can be generalized for new river profiles. Also, unless one reduces the dimension of the problem, these methods can have a computationally expensive similar to that of physics-based techniques. Here, we develop a new deep learning framework for riverine problems that can be trained using only a few river profiles and in a computationally efficient way that allows finding solutions on personal computers. The proposed method exploits the spatially local connection between the observations and river bed profile and combines a fully connected Deep Neural Network (DNN) with Principal Component Analysis (PCA) to image river bed topography using depth-averaged flow velocity observations. The new method is presented and applied to three riverine bathymetry identification problems. Results show that the proposed method achieves satisfactory performance in bathymetry estimation, providing a powerful data-driven technique for riverine bathymetry in terms of prediction quality, robustness, and computational cost that requires only a relatively small number of training samples.