Oceanic 3-D Thermohaline Field Reconstruction With Multidimensional Features Using SABNN

Abstract

Aiming at the problems of missing data and outliers in ocean observations and incomplete characterization of thermohaline related features, a 3-D thermohaline reconstruction model of the ocean based on multisource data are proposed. Multisource data from remote sensing and Current and Pressure recording Inverse Echo Sounders were used to analyze the projection relationship between 12-D features, such as sea surface temperature, bidirectional propagation time, and seafloor current velocity, and the distribution of ocean temperature and salinity at different depths (10–1000 m). A Bayesian optimization algorithmic framework is used to evaluate and gradually remove uncertainty from currently known data during the iterative process by extracting network parameters from the approximate probability distribution. More informed decision making improves the stability of the iterative process and reconstruction. In addition, a self-attention mechanism is introduced to dynamically focus on the dependencies between features of different dimensions by calculating the correlation matrix between features at arbitrary locations, enabling the model to more comprehensively characterize the thermohaline distribution and its changes. A Self-attentive Bayesian neural network (SABNN) model is established through empirical regression. The reconstructed model is validated using observational data from the Gulf of Mexico, and the experimental results show that the SABNN model has a significant improvement in temperature and salinity reconstruction accuracy compared with other network models or methods, with the RMSE and $R^{2}$ improved by more than 29.68%, 21.14% and 31.01%, 37.33%, respectively.