The method of identifying maize seed purity by analysising seeds spectral images of a small amount of near infrared bands was developed to satisfy the needs of rapid inspection and automatic sorting of maize hybrid seeds. The spectral images of hybrid and female parent of 5 maize varieties at 4 short wave near infrared bands in transmission mode and 4 medium wave near infrared bands in reflection mode were collected. Black-white calibration, median filtering, otsu method were applied to remove the noise and extract the seeds from background. Texture features were extracted by histogram statistics(HS) and gray level co-occurrence Matrix(GLCM). Splicing the feature data at each wavelength, principal component analysis(PCA) and orthogonal linear discriminant analysis(OLDA) were applied to reduce dimensions and obtain the best separability of subspace. The transmission and reflection spectral image purity identification model was built by support vector machine (SVM). The average correct identification rate of 5 maize varieties was above 85% both in transmission and reflection models. This research show that it is feasible to use spectral images of a small amount of near infrared bands to identify the purity of maize hybrid seeds.