Data augmentation using improved cDCGAN for plant vigor rating

The supervised deep learning models rely on large labeled training samples, which is a common challenge affecting current plant phenotyping studies. One practical approach to alleviate the insufficient training samples is data augmentation. In this study, we investigated the data augmentation approach using improved cDCGAN (conditional deep convolutional generative adversarial network) for vigor rating of orchid seedlings, a significant but labor-intensive task in modern commercial greenhouse. Various modifications on the architecture of cDCGAN network were explored for generating high-quality fine-grained RGB plant images with designated class labels. ResNet deep learning classifier was employed for performance evaluation throughout the whole analysis. On the small training sets, which obtained obviously worse ResNet classification results than bigger sets, cDCGAN was employed to generate additional plant images. The synthesized images provided a significant boost in classification performance, up to a 0.23 increase in the testing F1 score after data augmentation, achieving comparable results with that obtained with larger training sets without data augmentation. Different size of real and augmented training sets for optimal classification was systematically evaluated. The advantage of the improved cDCGAN architecture with added bypass connections was also demonstrated. The proposed data augmentation approach might be extended to deal with the common challenge of insufficient data size in other plant science tasks.