Collaboration of hyperspectral data and generative adversarial networks for improved nitrogen nutrition diagnosis and nitrogen requirement estimation in winter wheat

Nitrogen nutrient diagnosis and nitrogen requirement (NR) estimation are key components for accurate and precise crop fertilizer management. Owing to the limitations of field data collection, the number of measured samples is usually small and unbalanced, resulting in errors in model estimation accuracy. Challenges remain in accurately obtaining nitrogen nutrient diagnostics and estimating nitrogen fertilizer requirements. In this study, hyperspectral canopy data and measured data of winter wheat were acquired. The generative adversarial networks (GAN) was used to generate the winter wheat canopy hyperspectral dataset, and the original dataset, the GAN balanced dataset and the GAN hybrid dataset were constructed. The nitrogen concentration and biomass were estimated by combining partial least squares regression (PLSR), Gaussian process regression (GPR) and one-dimensional convolutional neural network (1D-CNN) models. Based on the estimation results, the nitrogen nutrient index (NNI) was calculated via the critical nitrogen dilution curve, and the NR estimation model was established with integrated consideration of days after sowing, nitrogen recovery efficiency, and the NNI. The results show that the GAN can meet the extension needs of small sample datasets, and the quality of the generated data is reliable enough at epoch=2000 and performs best when the amount of generated data reaches two times the original amount of data. Among the three models, GPR had the highest accuracy in estimating nitrogen concentration, whereas the 1D-CNN performed best in estimating biomass. Compared with the original dataset (R² = 0.88 for nitrogen concentration and R² = 0.82 for biomass), the R² values for nitrogen concentration and biomass estimation were 0.94 and 0.91 on the GAN balanced dataset and 0.97 and 0.92 on the GAN hybrid dataset. Compared with those of the original dataset, the R² values for estimating nitrogen concentration and biomass improved by 10.2 % and 12.1 %, respectively. R²=0.90 and RMSE=0.11 for the estimation of the winter wheat NNI based on nitrogen concentration and biomass were further obtained, with R²=0.80 and RMSE=22.86 for the estimation of NR. This study demonstrated the potential of the GAN application in hyperspectral data generation, which provides strong support for the precise management of nitrogen fertilization in winter wheat.