Sea salinity is one of the indicators of the global water cycle and affects the surface and deep circulation of the ocean. While passive microwave satellite sensors have been used to monitor sea surface salinity (SSS), the uncertainties from radio frequency interference (RFI) and low sea surface temperature often result in large errors, especially in river-dominated coastal seas. This study investigated the improvement of the Soil Moisture Active Passive (SMAP) SSS over five river-dominated oceans over the globe using three machine learning approaches (i.e., random forest (RF), support vector regression (SVR), and artificial neural network (ANN)). Four SMAP products and four ancillary data used in the SMAP SSS retrieval algorithm were used as input variables to the machine learning models. The results showed that all models improved the SMAP SSS product by up to 28% reduced in the root mean square error (RMSE) for validation, and RF yielded better performance than SVR and ANN. The calibration and validation RMSEs by RF were 0.203 and 0.556 practical salinity unit (psu), while those of SMAP SSS were 0.774 psu. The improved SSS well captured the spatiotemporal patterns of SSS for not only low but also high salinity water for all five regions. The proposed approach can be used to operationally improve the global SMAP SSS product including other coastal areas and the near Polar regions in the future.