China is a major cotton-growing country with numerous cotton varieties, each exhibiting significant differences in yield and fiber quality. However, the current management of cottonseed varieties is disorganized, resulting in severe homogenization and the presence of counterfeit and mislabeled varieties. The detection of cottonseed variety information has become a critical issue for the Chinese cotton industry. In this study, we collected near-infrared (NIR) spectral data from six cottonseed varieties and constructed a GAN for cottonseed NIR data (GAN-CNIRD) model to generate additional cottonseed NIR data. The Euclidean distance method was used to label the generated NIR data according to the characteristics of the true NIR data. We then applied Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), and Normalization algorithms to preprocess the combined dataset of generated and real cottonseed NIR data. Feature wavelengths were extracted using Bootstrap Soft Shrinkage (BOSS) and Competitive Adaptive Reweighted Sampling (CARS) algorithms. Subsequently, we developed Linear Discriminant Analysis (LDA), Random subspace method (RSM), and convolutional neural network (CNN) models to classify the cottonseed varieties. The results showed that for the LDA model, the use of feature wavelengths extracted after Normalization-BOSS processing achieved the best performance with an accuracy of 97.00%. For the RSM model, the use of feature wavelengths extracted after SNV-CARS processing achieved the best performance with an accuracy of 98.00%. For the CNN model, the use of feature wavelengths extracted after MSC-CARS processing achieved the best performance with an accuracy of 100.00%. Data augmentation using GAN-CNIRD-generated cottonseed data improved the accuracy of the three optimal models by 6%, 5%, and 6%, respectively. This study provides a crucial reference for the rapid detection of cottonseed variety information and has significant implications for the standardized management of cottonseed varieties.