Maize is one of the important food source, which is widely planted in China. The selection of excellent maize varieties is the key to agricultural production and breeding. However, there are wide varieties of maize on the market at present. In this paper, the extreme learning machine (ELM) model of maize varieties identification based on hyperspectral image technology was proposed to solve the problem of maize varieties identification. In this study, eight varieties of maize seeds were regarded as research objects, and 480 experiment samples were divided into training sets and test sets in a 2∶1 ratio, with 320 and 160 samples respectively. The images of maize seeds in the 935.61~1 720.23 nm were obtained by a hyperspectral acquisition system. Regions of interest (ROI) of 10×10 pixels in germ were selected after correction, and the average spectrum in the region was extracted as the original spectral data. Due to the large noise at both ends and less effective information of the original spectrum, in order to enhance the signal-to-noise ratio, spectral bands of maize seeds in the range of 949~1 700 nm were selected as effective bands for analysis. Due to the strong interference of irrelevant information during data collection, the spectral bands information after denoising was processed by Savitzky-Golay smoothing. The smoothing point was set to 3. Maximum normalization (MN) was used to pretreat based on SG smoothing. After pretreatment, feature wavelength variables were extracted by competitive adaptive reweighted sampling (CARS), successive projection algorithm (SPA) and CARS+SPA, CARS-SPA. The wavelength reflectance was used as the input matrix X, and the sample varieties 1, 2, 3, 4, 5, 6, 7, 8 were used as the output matrix Y. (SG+MN)-ELM, (SG+MN)-CARS-ELM, (SG+MN)-SPA-ELM, (SG+MN)-(CARS+SPA)-ELM, (SG+MN)-(CARS-SPA)-ELM were established. The experiment results showed that (SG+MN)-(CARS-SPA)-ELM model had the best identification performance compared with others, and the average identification accuracy of training sets and test sets was 98.13%, indicating that CARS-SPA secondary screening feature wavelength variables were more sensitive, which could represent all wavelengths information. The ELM model had better qualitative identification performance. It could realize the identification of maize varieties. This study provides a new idea and method for rapid and accurate identification of maize and other crop seeds.