Owing to the all day and all weather advantages of radar remote sensing and the strong penetrability of microwaves, information from radar may be supplementary to that of optical sensors, and thus facilitating the research of soil salinization using both radar and optical images. However, at present, few quantitative studies on soil salinization have been carried out by using polarimetric synthetic aperture radar (PolSAR) image and polarization characteristic parameters. Moreover, different variables extracted from optical and radar images as well as DEM data have been adopted to retrieve soil salinity by previous scholars. As to their retrieval efficiencies and comparative advantages, there are still some uncertainties and confusions which should be explored comprehensively to locate those variables with strong universality. Taking Juyanze, which is located at southeastern Ejina Banner in Inner Mongolia, as the study area, six types of variables including band reflectance, vegetation index, salinity index, polarimetric SAR parameter, land surface temperature and topographic factor were extracted based on Sentinel-2, Radarsat-2, Landsat-8 and SRTM DEM data. Variable optimization strategy was adopted to screen the optimal variable of each variable type and their combinations, and then multiple random forest (RF) and support vector machine (SVM) soil salinity prediction models were established and evaluated. The optimal model was used to predict soil salinity in Juyanze area, which was expected to provide practical reference for soil salinity monitoring in arid area. The results showed that variables such as short-wave infrared band (B11), canopy response salinity index (CRSI), extended ratio vegetation index (ERVI), salinity index Ⅱ rededge3 (S2re3), single scattering (FOdd), land surface temperature (LST) and total catchment area (CA) had high universality for soil salinity monitoring. For single variable models, the salt prediction accuracies were ranked in descending order as topographic factor, polarimetric SAR parameter, land surface temperature, salinity index, vegetation index and band reflectance. Multi-variable combination can effectively improve the model accuracy and stability. With the addition of environmental variables, R2 of the optimal model was increased by 0.117 and the corresponding RMSE was decreased by 2.556 percentage points when all six types of variables were involved in the model. RF model was more suitable for soil salt inversion in arid areas than SVM, and the RF model based on the optimal total variable group had the highest accuracy. The inversion results showed that the soil was mild salinized in northeast part and areas around Swan Lake, while in southwest paleolake basin, severe soil salinization was generally occurred.