Bathymetric Prediction and Uncertainty Quantification Using a Bayesian Deep Neural Network Based on Gravity Data

As seabed topography is closely related to the ocean gravity field, utilizing gravity data for seabed topography inversion has become the mainstream method. Although conventional deep neural network (DNN) methods have great potential in bathymetric prediction, they can neither evaluate the uncertainty of the prediction process nor the impact of uncertainty on prediction results, which limits their practical application value. To address this problem, a Bayesian DNN (BDNN) method is proposed for bathymetric prediction and uncertainty quantification. This method introduces Monte Carlo (MC) dropout variational inference into the architecture of a conventional DNN. Thus, the model achieves uncertainty quantification of prediction results with only a small amount of network structure changes. In addition, the captured uncertainty is fed back into the network training process to constrain the model parameters and calibrate the bathymetric prediction results. The experimental results show that the proposed BDNN model provides more reliable and accurate bathymetric prediction results than the conventional DNN and seabed topography inversion models. Moreover, the uncertainty results quantified by the model have a significant spatial correlation with the seabed topography, providing high confidence in the prediction results and reducing the risk in the interpretation of seabed topography, thus proving the potential of BDNN for accurate bathymetric prediction from gravity data.