Water content affects rice quality, which also has an important impact on rice storage, transportation, acquisition and processing. The annual loss of production caused by grain deterioration was up to 10 million tons, and the economic loss was up to 20 billion. Therefore, detecting rice moisture content accurately is beneficial to improve rice quality and reduce yield loss. A new method based on dielectric properties was proposed to detect the moisture content of rice in this study. Firstly, the dielectric properties (relative dielectric constant and dielectric loss factor) of 120 copies of rice of Japonica No.3 with different moisture contents were measured with impedance analyzer and self-made coaxial cylindrical capacitor at 201 discrete frequencies over the frequency range of 1 kHz-1 MHz, and the moisture contents of rice were measured by dry weight method. Secondly, sample set partitioning based on joint x-y distances (SPXY) was used to subset partitioning. Uninformative variables elimination (UVE) and successive projection algorithm (SPA) were applied to extract the characteristic variables of dielectric parameters ((the relative dielectric constant, dielectric loss factor and relative dielectric constant combined with dielectric loss factor). And the effect of SPA was compared with that of UVE to determine the optimal method for characteristic variable selection simultaneously. Finally, the support vector regression (SVR) machine and multiple linear regression (MLR) were adopted to establish the relationship models with two kinds of characteristic variables, single variables and full variables for predicting rice moisture content. And the performances of all the models were evaluated by the determination coefficient and root mean square error for calibration set and prediction set. The least square method was used for linear regression of predicted moisture content and measured moisture content at different temperatures, and the temperature compensation was carried out for the prediction results. The performances of the best model to predict different varieties of rice moisture content were explored to determine the applicability of the model. The research results showed that the relative dielectric constant decreased with the increase of the measurement frequency between 1kHz and 1MHz. When the frequency was greater than 300 kHz, the dielectric loss factor decreased with the increase of frequency and increased with the increase of water content. The measurement frequency and moisture content had an obvious effect on the dielectric properties of rice. Based on SPXY, 72 samples were partitioned to a calibration set and 48 samples to a prediction set. SPA was more effective than UVE in selecting useful information from the whole spectra of dielectric constant and dielectric loss factor. The model established by using the combination of relative dielectric constant and dielectric loss factor at multiple frequencies had better performance in predicting moisture content, which compared with the single dielectric parameter at a single frequency. Compared with MLR, SVR had better performance in predicting moisture content. The results showed that the support vector machine regression model based on the combination of relative dielectric constant and dielectric loss factor and SPA gave the highest correlation coefficient of predication set (0.980) and the lowest root mean square error of predication set (0.403%). When the best model was used to predict the water content of different varieties of rice, the prediction results were more accurate. Compared with the measured water content by the drying method, the prediction error was concentrated within ±0.5%. The study provided a reference for improving the accuracy of the grain moisture detection device.