Item

Abstract

XBT has been proven to be an invaluable tool in the measurement of temperature in the upper layers of the ocean. The effortlessness and low cost of XBT-based surveys has made possible the build-up of an impressive dataset used for inferring dynamic properties such as dynamic height, geostrophic velocity, and sound speed. However, for the proper estimation of such properties, the corresponding salinity fields are also required. Due to the inherent difficulty of in-situ measurements, the salinity can be estimated from XBT data based on the close relationship that exists between temperature and salinity in most of the ocean’s waters. The salinity field can be also estimated by objective analysis based in coarse in-situ observations or by data assimilation techniques, combining numerical observations and models. In recent years, with the increasing power of computers and the availability of immense amounts of data, the use of Neural Networks has been increasingly used to solve complex problems. An effort is underway to develop a Machine Learning (ML) model for the estimation of salinity along temperature profiles measured by XBTs. The preliminary results of a model using a stack of fully connected neural networks are promising. Using data from the AX18 and AX08 lines in the South Atlantic, it is found that when the input variables include the XBT location (longitude and latitude) and the depths, the accuracy is considerably increased as compared with when only temperature is used. In the sequence, the model will also incorporate remotely sensed surface properties.