Retrieving Global Subsurface Salinity from Satellite Observations Using Deep Learning

The satellite observations have been providing ocean data with wide spatial coverage for couple decades, substantially advancing the ocean and atmosphere science. Such ocean data measurements have been extended from surface to subsurface and deeper ocean by the deep ocean remote sensing (DORS) methods both dynamically and statistically, removing the constraint of sea surface on satellite observes. In this study, we successfully retrieved the global subsurface salinity anomaly (SSA) based on the sea surface information through deep learning methods. The surface information observed from satellites included sea surface temperature, sea surface salinity, sea level anomaly, and sea surface wind, and the Argo data was also employed as target data, being used for validation and testing. Monthly mean SSA from 10m to 2000m depth during 2010-2017 has been estimated, and the results showed that this inverse model could precisely retrieve SSA both spatially and temporally, with most of the coefficient determination larger than 0.9. Moreover, the trained deep learning model could accurately estimate SSA with the new inputs of surface information, thus, SSA before Argo era, during 1987-2005, had been estimated based on the surface satellite observations. Overall, the proposed method improved the global SSA estimation and can help identify and describe the oceanic salinity structure. Moreover, it would be a useful and practical technique for studying salinity variability at subsurface and deeper ocean and its role in changes of ocean dynamics and even climate change.