Improved inversion of soil total nitrogen content using conditional generative adversarial networks for hyperspectral data augmentation

It is often required for the inversion accuracy of the soil TN content under small sample size using visible and near-infrared (Vis-NIR) spectroscopy. In this study, a data augmentation framework was proposed to invert the soil TN content using generative adversarial networks (GANs). Specifically, a conditional generative adversarial network (CGAN) architecture was employed to improve the quality of the generated spectral data. The generation process was also guided using auxiliary information. Three types of the adversarial generative networks were evaluated: a standard GAN, a label-conditional generative adversarial network (LCGAN) with the soil TN content values as the conditional labels, and a variable importance in projection (VIP)-CGAN with the extremum scores as the conditional vectors in the feature wavelength sets. The feature wavelength was refined to identify the appropriate extreme points on the VIP score curve. Feature wavelengths were then extended with the less affected by the noise and external environmental interference. The more effective performance was validated this approach, compared with the higher VIP scores alone. The Vis-NIR spectra were collected in situ from the agricultural soils in the field. The experimental dataset was constructed with the TN content from the field and laboratory. Qualitative and quantitative assessments were performed on the fidelity of the synthetic samples that generated by the standard GAN, LCGAN, and various configurations of the VIP-CGAN. The results show that the best performance was observed in the VIP-CGAN variant with 9 extended feature wavelength bands (referred to as VIP-CGAN(T9)). The generated samples were achieved in the maximum mean discrepancy (MMD) and Fréchet inception distance (FID) scores as low as 0.003 and 0.005, respectively. There was the high statistical consistency between the generated data and the original data distribution. The reliable synthetic spectra were generated to fully learn the relationship between constraints and features. An enhanced dataset was constructed to combine the real samples with the synthetic samples after VIP-CGAN. The data augmentation was evaluated after enhancement. A systematic test was carried out to evaluate the predictive performance of three regression models—partial least squares regression (PLSR), support vector regression (SVR), and a one-dimensional convolutional neural network (1D-CNN). The optimal performance was achieved using synthetic samples generated by VIP-CGAN (T9) at a proportion of 300% (three times the size of the original training set). The PLSR model was attained a coefficient of determination (R²) of 0.86 with a root mean square error (RMSE) of 0.028 g/kg; The SVR model was achieved an R²of 0.84 and an RMSE of 0.009 g/kg; And the 1D-CNN model was performed best, with an R² of 0.88 and an RMSE of 0.026 g/kg. There was the significant improvement over the baseline models that trained only on the original dataset. The generator was conditioned on the VIP-selected wavelengths. The spectral regions were associated with the chemical information that related to the nitrogen compounds and organic matter, thus forming a physically meaningful constraints. The data augmentation was provided a more robust training environment for the regression models, in order to effectively mitigate the overfitting. In conclusion, an effective framework was obtained to enhance the hyperspectral inversion accuracy of the soil TN content under small sample size. The finding can provide a sound solution to the small samples in the soil Vis-NIR spectroscopy. The future applications of the soil properties can be extended to explore the more advanced generative architectures.