In order to establish the better maize variety identification models based on single kernel samples, the near-infrared spectral measurement methods were studied by comparing the direction of the maize seed’s embryo, diffuse reflectance and transmission mode, devices of holding the sample according to their impacts on the performance of variety identification models. Partial least squares-discriminant analysis (PLS-DA) was used to compress the pretreated spectral data into 9 variables, and then the identification models were built based on biomimetic pattern recognition (BPR). The results show that with the maize grain’s embryo facing the light source the models can be made perform better than with embryo backing toward the light source, diffuse reflectance mode is better than transmission mode, and small sample pool performs better than the small aperture. The measurement method of acquiring the diffuse reflectance near infrared spectra of maize by small pool with the seed embryo facing the light source can make models have the best performance. The average correct identification rate of the models is 94.6%, and the average correct rejection rates for the varieties not belonging to the models reached 96.5%.