Hyperspectral imaging combined with generative adversarial network (GAN)-based data augmentation to identify haploid maize kernels

Abstract

The rapid and efficient selection of qualified haploid maize kernels plays an important role in maize breeding. However, considering the rarity of haploid maize kernels, usually only a limited data set can be obtained, which easily leads to poor generalization ability and instability of established classifiers. In this work, hyperspectral imaging (HSI) combined with generative adversarial network (GAN)-based data augmentation method was used to identify haploid maize kernels. Two variants of GAN (deep convolutional GAN (DCGAN) and conditional GAN (CGAN)) were used to expand the spectra of haploids and diploids of two maize varieties respectively. After training of 10,000 epochs, fake spectra which were very similar to real spectra were generated, and the similarity between them was analyzed. To verify the reliability of this method, various classifiers were used, and the classification accuracy was compared while keeping the test set unchanged. The results showed that both DCGAN and CGAN could improve the accuracy of each classifier by more than 10 % on average, and CGAN improved the accuracy of each classifier higher than DCGAN. In addition, it also indicated that HSI combined with GAN-based data augmentation method had a good application prospect in the identification haploid maize kernels.