Predicting subsurface thermohaline structure from remote sensing data based on long short-term memory neural networks

Abstract

Abstract Satellite remote sensing can detect and predict subsurface temperature and salinity structure within the ocean over large scales. In the era of big ocean data, making full use of multisource satellite observations to accurately detect and predict global subsurface thermohaline structure and advance our understanding of the ocean interior processes is extremely challenging. This study proposed a new deep learning-based method—bi-directional long short-term memory (Bi-LSTM) neural networks—for predicting global ocean subsurface temperature and salinity anomalies in combination with surface remote sensing observations (sea-surface temperature anomaly, sea-surface height anomaly, sea-surface salinity anomaly, and the northward and eastward components of sea-surface wind anomaly), longitude and latitude information (LON and LAT), and subsurface Argo gridded data. Because of the temporal dependency and periodicity of ocean dynamic parameters, the Bi-LSTM is good at time-series feature learning by considering the significant temporal feature of the ocean variability and can well improve the robustness and generalization ability of the prediction model. For December 2015 as an example, our average prediction results in an overall determination coefficient (R2) of 0.691/0.392 and a normalized root mean square error of 0.039\xa0°C/0.051 PSU for subsurface temperature anomaly (STA)/subsurface salinity anomaly (SSA) prediction. This study sets up different cases based on different sea-surface feature combinations to predict the subsurface thermohaline structure and analyze the role of longitude and latitude information on Bi-LSTM prediction. The results show that in the prediction of STA, the contribution of LON\xa0+\xa0LAT to the model gradually increases with depth, whereas in the prediction of SSA, LON\xa0+\xa0LAT maintains a relatively significant contribution to the model at different depths. Meanwhile, in the STA and SSA prediction, the LAT plays a more significant role than LON. We also applied the model to bi-directional prediction for different months of 2010 and 2015 to prove the applicability and robustness of the model. This study suggests that Bi-LSTM is more advantageous in time-series modeling for predicting subsurface and deep ocean temperature and salinity structures, fully takes into account the timing dependence of global ocean data, and outperforms the classic random forest approach in predicting subsurface thermohaline structure from remote sensing data.