Correction of Satellite Sea Surface Salinity Products Using Ensemble Learning Method

Although salinity satellites can provide high-resolution global sea surface salinity (SSS) data, the satellite data still display large errors close to the coast. In this paper, a nonlinear empirical method based on random forest is proposed to correct two Soil Moisture and Ocean Salinity (SMOS) L3 products in the tropical Indian Ocean, including SMOS BEC and SMOS CATDS data. The agreement between in-situ data and the corrected SMOS data is better than that between in-situ data and the original satellite data. The root-mean-square deviation (RMSD) of the satellite SSS data decreased from 0.366 to 0.275 and from 0.367 to 0.255 for SMOS BEC and SMOS CATDS, respectively. The effect of the correction model was better in the Arabian Sea than in the Bay of Bengal. The RMSD of corrected BEC (CATDS) SSS was reduced from 0.44 (0.48) to 0.276 (0.269), and the correlation coefficient was increased to 0.915 from 0.741(0.801) in the Arabian Sea while the correlation coefficient improved less than 0.02 in the Bay of Bengal. The cross-validation results highlight the robustness and effectiveness of the correction model. Additionally, the effects of different features on the correction model are discussed to demonstrate the vital role of geographical information in the correction of satellite SSS data. The proposed method outperformed other machine-learning methods with respect to the RMSD and correlation coefficient.