In order to solve the problem, the performance of the wheat seed purity detection model decreased due to sample imbalance and band information redundancy in the process of hyperspectral imaging. A seed purity hyperspectral detection model was proposed by combining the synthetic minority oversampling technique (SMOTE) with uninformative variables elimination (UVE) and support vector machine (SVM). In this model, the SMOTE was used to expand the minority class (impurity) samples of the wheat seeds to improve the sample imbalance. At the same time, the UVE was used to select the high-dimensional hyperspectral features, and the SVM model was constructed to further reduce the risk of model overfitting caused by feature redundancy. Results showed that: The average accuracy, precision, and negative sample detection rate of the five types of wheat seeds are 95.98%, 94.94%, and 89.32%, respectively, which are 3.89%, 7.18%, and 12.42% higher than the traditional methods, respectively. The proposed method has a good application prospect in the detection of wheat seed purity based on hyperspectral imaging technology.