A deep learning approach to estimate ocean salinity with data sampled with expendable bathythermographs

Abstract

Expendable bathythermographs (XBTs) have made possible the build-up of an impressive dataset of temperature in the upper one thousand meters of the ocean. In the absence of direct measurements of salinity, the salinity profile associated with the temperature may be estimated by regression from temperature data or by objective analysis and data assimilation techniques. With the advent of the Argo program, the number of in situ salinity sampling has increased considerably. However, it is still far from ideal and XBTs continue to provide invaluable contribution to oceanography. Here, considering the large amount of data available, and motivated by the increasing use of machine learning to solve complex problems, a deep learning model based on a feed-forward neural network was used to estimate salinity directly from the temperature measurements. The model was independently trained and evaluated with two different datasets: (1) the sampled temperature and the associated salinity in XBT datasets, estimated with traditional methodologies, and (2) data sampled by conductivity-temperature-depth (CTD) profilers. The fitted deep learning model was then used to estimate the salinity from independent XBT datasets. The results show that the model’s accuracy increases substantially when longitude, latitude, depth, and month of the samplings are considered. Compared with four traditional methods, the deep learning performed much better, particularly near the ocean’s surface. In general, the results are highly accurate. However, when the CTD-trained model is used to predict salinity with XBT temperature input, the estimates are less accurate when compared with the salinity in the original XBT data. This seems to be caused by poorer estimates of salinity obtained by other methods in the original XBT datasets.