Hyperspectral imaging technology is used to sort varieties of seeds. However, the overall performance of prediction models decreases when they are used to test the same variety of seeds from different years or seasons. Prediction accuracy is susceptible to the influence of time and thus depends on the training set used to build the model. In this study, a model updating procedure of hyperspectral imaging data for classification of maize seeds using a clustering algorithm was proposed to maintain the accuracy and robustness of the model. A total of 2000 seeds of four typical maize varieties grown in China in three different years were used for classification based on a least-squares support vector machine classifier. After determining and applying the model parameters, the updated model achieved an overall accuracy rate of 98.3%, which is higher than the 84.6% accuracy obtained using the non-updated model. The accuracy rate of the updated model was 94.8% when testing with the Kennard-Stone algorithm, which is commonly used for selecting datasets. The proposed model updating method can successfully update seed data for cross-year model building and thus improve the overall accuracy for predicting of maize seeds harvested in different seasons.