AimsSoil salinization has been an important environmental problem globally, particularly in oasis areas in arid zones. The advantages of using multi-source data, combining radar and optical remote sensing data, and applying machine learning-based algorithms to these data could be beneficial for addressing the soil salinization problem.MethodsThis study combines the environmental covariates extracted from the Gaofen-3 (GF-3) radar data, Landsat-8 multispectral data, and digital elevation model (DEM) data to explore the advantages of radar remote sensing in detecting soil salinity. The soil salinity distribution degree in the Keriya Oasis is mapped using a machine-learning-based method, and the advantages of different sensor images in predicting soil salinity are evaluated. Three soil salinity inversion models are constructed using measured electrical conductivity (EC) data, the random forest (RF), gradient boosting tree (GDBT), and extreme gradient boosting (XGBoost) models.ResultsThe best accuracy corresponding to an R2 of 0.87, and a root mean square error (RMSE) of 6.02 is achieved by the RF model on the GF-3 + Landsat-8 data. Therefore, the use of multi-source data is a more effective method for mapping soil salinity in the study area. The mapping results of the optimal model demonstrate that natural factors significantly influence the distribution of soil salinity.ConclusionThe radar polarization decomposition characteristics are incorporated into the inversion of soil salinity modeling as an environmental covariate, providing an innovative and efficient method for soil salinity estimation in arid areas.