Data driven three-dimensional temperature and salinity anomaly reconstruction of the northwest Pacific Ocean
Journal Articles
Overview
Research
Identity
Additional Document Info
View All
Overview
abstract
By virtue of the rapid development of ocean observation technologies, tens of petabytes of data archives have been recorded, among which, the largest portion are those derived from the orbital satellites, embodying the character of ocean surface. Nevertheless, the insufficiency of information below the subsurface restricts the utilization of these data and the understanding of ocean dynamics. To circumvent these difficulties, we present the spatially three-dimensional reconstruction of ocean hydrographic profiles at depth based on the satellites and in-situ measurement data. In this manuscript, long short-term memory network (LSTM) and Gaussian process regression (GPR) methods are invoked to predict the temperature and salinity profiles in the northwest Pacific Ocean, and to improve computational and storage efficiency, the proper orthogonal decomposition (POD) method is subtly incorporated into these two models. LSTM and GPR show satisfactory results, with the root mean square error (RMSE) of temperature is less than 1.45, and the RMSE of salinity is less than 0.19. The incorporation of the POD method substantially accelerates efficiency, particularly in the LSTM model, which improves 7.5-fold without significant accuracy loss. The sensitivity of different sea surface parameters on the reconstructed profiles reveals that sea surface height anomaly and latitude significantly influence the reconstruction of temperature anomaly (TA) and salinity anomaly (SA) profiles. Besides, sea surface salinity and sea surface temperature anomalies can improve the model's estimation ability for the upper TAs and SAs, respectively. The contribution of monthly climatology to temperature and salinity profile estimation is also explored in this paper. It is shown that adding monthly mean climatology to the input of the model can achieve more accurate estimates.
