Abstract: Abstract: Rice variety identification is important in seed industry to assure rice seed purity and quality. The objective of this study was to assess the feasibility of a rapid and nondestructive determination of varieties of rice seeds using multispectral imaging system. A total of 250 seeds (five varieties with 50 seeds each) were provided by Institute of Rice Research, Anhui Academy of Agricultural Sciences, Hefei, China. The seeds were divided into two groups such as calibration set (40 seeds of each variety) and validation set (10 seeds of each variety). The multispectral imaging analysis was performed using the VideometerLab equipment (Videometer A/S, H?rsholm, Denmark) which acquired the multispectral images at 19 different wavelengths from the visual to the lower wavelengths of the NIR region (in the range of 405-970 nm). Image segmentation was performed using the VideometerLab software version 2.12.23. Background of the image was removed by a Canonical Discriminant Analysis (CDA) and rice seed images were segmented using a simple threshold. Having been segmented into a region of interest (ROI), the images were then used to get reflectance spectral of ROI. Image features data, including area (mm2), width/length, roundness, hunter L*, A*, B* values of rice seeds were extracted from the image analysis. Principal component analysis (PCA) with reflectance spectral and image features data was performed to examine the qualitative difference of these five rice varieties using the first two score vectors. Least squares support vector machine (LS-SVM) was used to obtain the discrimination model of varieties of rice seeds, particle swarm optimization (PSO) was designed to search the optimal values of SVM parameters to improve the search efficiency. The method presented in this paper with reflectance spectral and images feature data (Model 1) was compared with other different 3 models. For Model 2, the method used was the same as Model 1 with the spectral features data only, Model 3 was traditional LS-SVM from which optimal values of SVM parameters were obtained by cross validation with spectral and image features data, and Model 4 was principal component analysis-back propagation neural network (PCA-BPNN) with first four score vectors. The results showed that the discrimination accuracy of rice seeds of each variety from the Model 1 that combined with spectral and image features was up to 90% in validation set and 100% in calibration set, respectively, higher than other methods used in the experiments. It can be concluded that the spectral data or the image features data were both the key factors for the discrimination of the varieties of rice seeds, and differences among different varieties of rice seeds did exist and groups were apparent. The model with LS-SVM based on PSO can get better parameters to improve the classification ability than traditional LS-SVM based on cross validation. The model proposed in this paper using the combined spectral and image features data had the best prediction abilities, with the mean accuracy of the total five varieties seeds up to 94% in validation set. Multispectral imaging combined with the proposed algorithm has been proved to be a very powerful and attractive tool for classifying varieties of rice seeds because it is nondestructive, simple, rapid, and no pretreatments required.