An inversion method of subsurface thermohaline field based on deep learning and remote sensing data

Abstract

ABSTRACT Seawater temperature and salinity are basic marine environmental parameters, which can be used to calculate other marine environmental parameters. However, most of the on-site observation data have the problems of uneven spatial distribution and time discontinuity, and it is difficult for remote sensing observation methods to obtain subsurface information. In this study, we proposed a deep learning model with combining remote sensing temperature and salinity as well as in-situ measured data by Argo profiles, and the nonlinear relationship was revealed. An effective and direct inversion method was realized for underwater three-dimensional thermohaline structure based on remote sensing temperature and salinity at accurate points on global scale. The SST data were obtained from the FY3C-VIRR daily sea surface temperature product, the SSS data were acquired from the SMAP Level 3 8-day running averages sea surface salinity product, and the Argo scatter data were got from the ‘Global Ocean Argo Scatter Data Set’. Based on the temporal and spatial location information of the data, this paper matched the remote sensing data from 2016 to 2019 with the Argo data, 64751 valid pairing points were obtained. The deep learning model was constructed as a multilayer perceptron model with 5 hidden layers. The RMSE of temperature had a maximum value of 2.106°C in 130 m depth and a minimum value of 0.367°C in 1000 m with an average of 1.174°C for validation dataset. And the RMSE of salinity had a maximum value of 0.356 psu in 0 m and a minimum value of 0.045 psu in 1000 m with an average of 0.202 psu. Compared with other methods based on fixed mesh products, this study realized a direct inversion method at any location in the global ocean and improved the inversion accuracy, which provides a reliable data support for refined marine monitoring.