Objectives SSS (sea surface salinity) is an important physical and chemical indicator describing ocean state, and is also one of the most critical variables in global water cycle. In order to achieve large-scale and continuous observation of the ocean salinity, SMOS (soil moisture and ocean salinity) mission is launched by European Space Agency to access brightness temperature for SSS retrieval. However, researches have pointed out that the accuracy of SMOS ocean salinity products cannot fully achieve scientific expect. Thus, a novel and effective method is desired for improving SMOS salinity.

Methods This paper selects study area in the Southeast Pacific, where the region is relatively less affected by radio frequency interference. The downloaded array for real-time geostrophic oceanography SSS data, SMOS level2 ocean salinity products and the Auxiliary data in the whole year of 2018 are matched both temporally and spatially. Based on the remote sensing mechanism and radiation transfer theory, five variables, including the first Stokes parameter (T_H+T_V), SST, UN10, VN10 and Ω are selected as sensitive factors. The ocean salinity retrieval model based on random forest is first established, and GridSearchCV(GS) algorithm is applied to optimize superparameters of the initial model, which is able to improve SSS accuracy further.

Results The experiment results show that: (1) For objective measurement, the mean absolute error (MAE) and root mean square error (RMSE) are adopted to compare the accuracy of the produced SSS from two established models and the SMOS products, in terms of Argo in-situ SSS as reference. According to experiments, MAEs of basic RF(random forest) model and GS‐RF model are both 0.08, which is evidently lower than SMOS SSS product􀆳s 0.66. Similarly, RMSEs of two models are also remarkable lower than SMOS􀆳s 0.93, while GS‐RF model is 0.14, indicating its slight advantage over RF model, which is 0.15. (2) All SSS retrieved from RF model and GS‐RF model distribute concentratedly around 36 psu, while SMOS ocean salinity product occasionally presents abnormal SSS that beyond 40 psu or below 30 psu, which is inappropriate. (3) In the testing set, error of SMOS SSS can reach -12.9 psu and 6 psu, while errors of two established models are no more than ±1.3 psu, noticeably lower than SMOS product. Meanwhile, GS‐RF model also present slight advantage than basic RF model on error measurement.

Conclusions Through experiments and analyses, we found that it is feasible to establish sea surface salinity retrieval methods based on random forest model, by deliberately selecting SSS sensitive factors, and advancedly optimizing super-parameters of model, especially with GridSearchCV algorithm. This paper has achieved reliable and satisfactory results, which have greatly improved the accuracy of ocean salinity data compared with SMOS products, thus benefit for relevant marine studies.