Sea Surface Salinity is an important parameter to learn oceans ’ impact on the global climate. SMOS (Soil Moisture and Ocean Salinity) developed by the European Space Agency is one of the satellites dedicated to measuring ocean salinity. However, due to influences like Radio Frequency Interference (RFI), the accuracy of SMOS salinity products hardly achieve the expected results. In order to improve the accuracy of the sea surface salinity products, a deep neural network based sea surface salinity retrieval algorithm is proposed in this paper. The experiment takes the central Pacific Ocean (150°E—180°, 5°—30°N) as the research area, using the Argo measured salinity data as the reference. First, SMOS L1C and L2 level products are matched in time and space with the Argo salinity data. Then, according to the ocean remote sensing and radiative transfer theory, seven parameters that affect ocean salinity are selected, including Brightness Temperature (TB), Sea Surface Temperature (SST), Rainfall Rate (RR), Significant Wave Height (SWH), Zonal Wind Speed (ZWS), Meridional Wind Speed (MWS) and Evaporation (Eva). The DNN (Deep Neural Network) is developed and optimized with the K-fold cross-validation method. And the sea surface salinity data is retrieved using Argo salinity as the ground-truth. The experiments show that the mean absolute error of the sea surface salinity data obtained by the algorithm proposed in this paper is 0.159, and the root mean square error is 0.195, which are better than the accuracy of SMOS salinity products.