Inversion of sea surface salinity in the western Pacific Ocean based on machine learning

To improve the accuracy of the SMOS satellite and explore the applicability of machine learning models at varying distances from the coastline, this study employs three machine learning models—Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN)—using seven sea surface characteristic parameters as retrieval factors to investigate sea surface salinity (SSS) inversion for the year 2022 in a selected area of the Northwest Pacific (0–30°N, 120–150°E). The inversion results show that: (1) The RF and SVM models performed well, followed by KNN. The root mean square errors (RMSE) of the three models compared with Argo in-situ measurements were 0.2741 psu, 0.2829 psu, and 0.3357 psu, respectively. All models outperformed the SMOS Level 2 salinity product, significantly improving the satellite's accuracy. (2) The error ranges for all three models were relatively concentrated; however, their performance declined in October. In regions with lower salinity values, the output salinity values were also less accurate. (3) All three machine learning models demonstrated high applicability even within 200 km of the coastline, an area typically challenging for salinity retrieval.