Most of the current studies on Sentinel-1/2 synoptic inversion of vegetation soil salinity were based on Sentinel-2 spectral information and Sentinel-1 backscattering coefficients, without considering the two aspects that Sentinel-2 spectral information was susceptible to soil brightness and Sentinel-1 backscattering coefficients were susceptible to soil roughness and moisture. Therefore, in order to further improve the accuracy of Sentinel-1/2 synoptic inversion of vegetation soil salinity, the Sentinel-1 backscatter coefficients were corrected with a water cloud model to eliminate the influence of vegetation. Then, the corrected backscatter coefficients and Sentinel-2 texture features screened by VIP, OOB and PCA were used to construct soil salinity inversion models based on RF, ELM and Cubist. The results showed that the correlation between the radar backscatter coefficient and the soil salinity was improved to some extent after the removal of vegetation effects by the water cloud model. For the coupled models of different variable selection methods and different machine learning methods, OOB had the best performance in soil salinity inversion when being coupled with RF, ELM and Cubist, with R2 above 0.750 for both modeling and validation sets. And OOB-Cubist model had the highest accuracy and R2v/R2c was 0.955, which had good robustness. It provided some ideas for further applications of machine learning in collaboration with physical models and optical satellites in collaboration with radar satellites in soil salinity inversion.