High-Resolution and Accurate Spatial-Temporal Prediction of Oceanographic Fields via Sparse Observations from Marine Vehicle Network using Deep Learning and Data Assimilation

High-resolution and accurate prediction of the states of ocean temperature, salinity, and flow fields with the observation of a network of marine vehicles are significant for a variety of civilian and military applications. Relative to the high-dimensional ocean states, the observations by the marine vehicles are sparse and noisy. Thus, how to implement high-resolution and accurate prediction of the oceanographic fields spatially and temporally based on the sparse and noisy data streams provided by a marine vehicle network is an important issue. This paper investigates this issue and presents a data-driven prediction framework, which uses a deep neural network model to predict the mean-field of the ocean states and the Gaussian process to predict the residual. Based on the model, the Kriged Kalman filter method is used to update the network parameters and to predict the ocean states with sensing data of marine vehicles. The deep learning model is based on the ConvLSTM and considers the joint prediction of temperature, salinity, and flow fields. To account for the high-dimension of the network parameters, the filtering is implemented with the ensemble Kalman filter. A simulation demonstration is given to show the learning of the deep neural network parameters with the data assimilation method.