Accurate soil salinity monitoring is crucial for sustainable soil use and management. While most existing studies rely on optical remote sensing for salinity estimation, the potential of polarimetric synthetic aperture radar (PolSAR) data, particularly its polarimetric decomposition characteristics, remains underexplored. This study focuses on the Yutian Oasis in southern Xinjiang, China, to investigate the potential of PolSAR data for estimating soil salinity in arid regions through the integration of multi-source remote sensing data (including RADARSAT-2 C-band SAR, Sentinel-2, and topographic data). From the multi-source dataset, 121 features were extracted, and correlation analysis identified 52 variables significantly correlated (P < 0.05) with soil electrical conductivity (EC). These variables were then further screened using three feature selection algorithms: Recursive Feature Elimination (RFE), Boruta, and Variable Importance in Projection (VIP), to mitigate high-dimensionality and collinearity. Subsequently, three machine learning models-Multi-Layer Perceptron (MLP), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)-were employed to construct soil salinity inversion models. The results revealed that the Boruta-MLP model outperformed other strategies in both the calibration and validation phases, demonstrating strong generalization capabilities. For validation, the Boruta-MLP model achieved an R-2 of 0.819, with RMSE and MAE values of 5.767 and 3.800, respectively. Variable sensitivity analysis indicated that key SAR features-including the backscatter cross-polarization ratio (sigma(0)_VV/sigma(0)_VH), radar vegetation index (RVI_sigma(0)), and volume scattering index (VSI_sigma(0))-along with SAR polarimetric decomposition components (Alpha, Entropy, MF4CF_theta_FP) and texture features (Contrast_sigma(0)_VH, Dissimilarity_sigma(0)_VH, and Homogeneity_sigma(0)_VH), play crucial roles in soil salinity estimation. This research underscores the critical role of SAR data and advanced feature selection in soil salinity estimation, offering a robust framework for arid region salinity mapping through multi-source data integration and machine learning optimization.