A hybrid Conv1D-GRU model with spectral augmentation for non-destructive rice seed vigor detection

Seed quality is closely associated with rice yield and grain quality, and seed vigor is a key indicator for seed quality evaluation. High-vigor seeds usually show stronger resistance to environmental and biotic stresses, thereby improving germination and seedling establishment. Thus, rapid and accurate detection of rice seed vigor is essential for breeding, storage management, and crop production. In this study, a non-destructive rice seed vigor detection method based on near-infrared (NIR) spectroscopy, spectral augmentation, and Conv1D-GRU classification was developed. Rice seed samples with different vigor levels were prepared through artificial aging, and seed-level NIR spectra were acquired using a NIR spectrometer. Spectral preprocessing was applied to reduce noise, enhance relevant spectral features, and correct scattering effects. Sparse representation and dictionary learning were used to augment the training spectra and improve sample diversity. In the Conv1D-GRU classifier, the Conv1D layers extracted local spectral features from adjacent wavelength regions, while the GRU layer captured wavelength-order contextual information across the spectral sequence. The key hyperparameters of the classifier were optimized using an integrated population search algorithm. Experimental results showed that the proposed method achieved test accuracies of 0.9844, 0.9740, and 0.9818 for conventional japonica rice, indica-japonica hybrid rice, and japonica glutinous rice, respectively. Compared with PLS-DA, SVM, XGBoost, 1D-CNN, and GRU models, the Conv1D-GRU classifier showed better overall performance under the current experimental conditions. These results indicate that the proposed NIR spectroscopic method provides a promising non-destructive approach for rice seed vigor detection and has potential for seed quality evaluation and agricultural production management.