Abstract: Linear and non-linear empirical models for salinity (S) are estimated from the Argo temperature (T) and salinity (delayed) data. This study focuses on the reconstruction of salinity in the upper 1200m of the eastern North Atlantic Ocean, a region characterized by the presence of many different water masses. While previous studies have found it necessary to split this region by boxes to fit different polynomial models in each box, a unique model valid for the entire region is fitted here. Argo profiles are randomly distributed on two sets: one for fitting the models and one for testing them. Non-linear regressions are built using neural networks with a single hidden layer and the fitting data set is further divided into two subsets: one for adjusting the coefficients (training data) and one for early stopping of the fitting (validation data). Our results indicate that linear regressions perform better than the climatologic T–S relationship, but that non-linear regressions perform better than the linear ones. Non-linear training using a three-data subsets strategy successfully prevents overfitting even when networks with 90 neurons in the hidden layer are being trained. While the presence of local minima may complicate the generalization of non-linear models to new data, network committees (created by training the same network from different random initial weights) are shown to better reproduce the test data. Several predictors are tested, and the results show that geographical, or surface, information does provide significant information. These results highlight the potential applications of future satellite missions measuring sea-surface salinity to reconstruct, when combined with temperature profiles, vertical salinity profiles. [Copyright &y& Elsevier]