Abstract: Soil moisture has great significance in the researches of hydrology, meteorology and agriculture yield estimation. The quad-polarimetric SAR images can provide a lot of polarimetric features, the significance of the features in surface parameter retrieval have attracted attentions in previous researches with no final conclusions because of the complexity of terrain scattering. In this paper, random forest regression (RFR) is used for both soil moisture retrieval and the importance evaluation of polarimetric features of Radarsat-2 images in winter wheat fields. According to the score of importance, feature selection and combination are done for modelling. We evaluate the retrieval accuracy of models with different feature combinations. The results show that models of important features selected by RFR have RMSE(root mean square error) less than 6% which are better results compared to traditional models; when compared with support vector regression and artifical neural networks, the RFR also shows best retrieval accuracies, which proves that RFR is suitable for soil moisture retrieval and feature selection. The high retrieval accuracies of LBC-CPD(linear backscatter coefficients-Cloude-Pottier decomposition) and LBC-CPR(linear backscatter coefficients-circular polarimetric ratio) indicates these features can improve the retrieval accuracy of soil moisture.