A Few-Shot Wheat Seed Spectral Classification Method Based on Improved WGAN-GP Data Augmentation and Transfer Learning with a 1D CNN

Highlights Optimized WGAN-GP with physical constraints for high-fidelity spectral data generation. Used 1D CNN to classify augmented datasets and evaluated enhancement strategy impact. Transferred learning from rice seed spectra to improve wheat seed classification. Developed an online platform to verify real-time performance and stability. ABSTRACT. Visible and near-infrared spectroscopy (Vis/NIR) technology enables rapid and non-destructive identification of seed varieties. However, in few-shot scenarios, the scarcity of labeled data limits the deployment of deep learning models. This study proposes a method based on an improved Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and transfer learning to address the few-shot wheat seed spectral data classification problem. First, high-fidelity spectral data was generated using an improved WGAN-GP model for data augmentation, and a mixing strategy with varying ratios of generated data to real data was designed. Then, based on the augmented dataset, a 1D-CNN model was trained to validate the effectiveness of the generative model and the impact of different augmentation strategies on the classification performance. Experimental results show that the optimal data augmentation strategy improved the classification accuracy by 12.83%, achieving 84.62% compared to the baseline without data augmentation. Subsequently, a pretraining process was conducted on a 1D-CNN model using a rice seed spectral dataset. Through transfer learning, the model was fine-tuned using the augmented wheat seed dataset, ultimately achieving a classification accuracy of 94.87% on the wheat seed test set. Additionally, the developed online detection platform verified the potential for engineering application of this method. Experimental results demonstrate that the combined approach of data augmentation and transfer learning significantly improves classification accuracy, validating its effectiveness and feasibility in few-shot seed spectral data classification. Keywords: Deep learning, Few-shot, Transfer learning, Vis/NIR, Wheat seeds, WGAN-GP.