As one of the most widely planted crops in China, the yield of corn is of great significance to Chinas food security. Since different varieties have different characteristics, scientific seed selection according to planting conditions can significantly improve the yield and reduce the cost of production. Still, the appearance of different corn seeds is extremely similar, which leads to a certain degree of difficulty in scientific seed selection. In this study, based on near-infrared spectroscopy combined with the Kernel Extreme Learning Machine (KELM) to construct a discrimination model for the classification of corn varieties, the use of sweet glutinous yellow corn, sweet princess, Chang sweet, golden superman, sweet No.5 five kinds of maize seeds, each kind of 6 grains as a sample, a total of 126 samples as the object of the study, the near-infrared spectroscopy data collected by the standard normal variate transformation (SNV) treatment. Competitive Adaptive Re-weighted Sampling (CARS) was used to downscale the dataset. The samples were randomly divided into training and test sets according to the ratio of 5∶1 to explore the effect of the Northern Goshawk Optimization Algorithm (NGO) on the performance of the KELM model. The two important parametric regularization parameters C and Gaussian kernel function γ of the KELM model was optimized using the NGO algorithm, particle swarm algorithm (PSO), and gray wolf algorithm (GWO), respectively, and the C and γ corresponding to the highest accuracy rate of the 50-50 cross-validation recognition were selected as the modeling parameters to build the KELM classification model. The KELM models performance is established after each algorithms optimisation is compared. It is found that the performance of the KELM model established after optimization by the NGO algorithm is higher than that of the KELM model optimized by the other two algorithms, and the recognition accuracy of the test set can reach 100%. The CARS-NGO-KELM, CARS-PSO-KELM and CARS-GWO-KELM models are built based on CARS dimensionality reduction, and the results show that the NGO algorithm still performs better in the face of dimensionality reduction data, and its test set accuracy and F1 value both reach 100%. To verify the effect of sample size on the model, the KELM model was retrained using a total of 90 samples after synchronizing the sample size of each species. The results showed that after synchronizing the number of samples of each species, the performance of each model was improved on both the training and test sets. In this study, a variety of optimization algorithms are introduced to construct a machine learning model based on near-infrared spectroscopy and the recognition accuracy is increased to 100%, which realizes a fast, non-destructive, and accurate variety identification of maize seeds. The results of the study provide a new method for the rapid identification of maize varieties and also have certain significance as a guide for the supervisory authorities.