Hyperspectral retrieval of nitrogen content in garlic plants via cVAE-GAN-based data augmentation and ResNet architecture

Garlic’s long growth cycle and field sampling difficulties limit data availability for hyperspectral inversion of nitrogen content. To overcome this, we propose a hybrid modeling framework combining a conditional variational autoencoder-generative adversarial network (cVAE-GAN) with a convolutional block attention module-ResNet (CBAM-ResNet). Hyperspectral data were collected by unmanned aerial vehicles during key growth stages and preprocessed with Savitzky–Golay smoothing. The cVAE-GAN, optimized with a composite loss and multicriteria evaluation, generated synthetic hyperspectral-nitrogen datasets to expand limited samples. Based on this, the CBAM-ResNet model enhanced spectral feature extraction and focused on informative bands for nitrogen retrieval. Experiments demonstrated that augmenting with 270 synthetic samples achieved the best results, with R2=0.9867, RMSE = 0.0554 mg/g, and MAE = 0.0451 mg/g. Compared with models trained only on real data, R2 increased by 0.1051, whereas RMSE and MAE decreased by 0.0977 and 0.0877 mg/g, respectively. These results confirm that the proposed cVAE-GAN-CBAM-ResNet framework effectively addresses sample scarcity and significantly improves the accuracy and generalization of nitrogen monitoring in garlic, offering a reliable, efficient, and nondestructive solution for precision agriculture.