Combining deep convolutional generative adversarial networks with visible-near infrared hyperspectral reflectance to improve prediction accuracy of anthocyanin content in rice seeds

Abstract

Anthocyanin is a crucial reference indicator for evaluating the quality of rice varieties, making it significant to rapidly establish a non-destructive detection method for anthocyanin in rice grains. This study constructs a 1D-DCGAN (One-dimensional deep convolutional generative adversarial network) strategy optimized for one dimensional spectral data and a 1D-CNN (One-dimensional convolutional neural network) model, achieving high-quality generated sample effects and more accurate anthocyanin predictions within a limited dataset. The SG (Savitzky-Golay)-1D-CNN significantly outperforms LSR (Least squares regression), SVM (Support vector machine) and BPNN (Backpropagation neural network) in the test set, with R² (Determination coefficient), RMSE (Root mean square error) and RPD (Residual predictive deviation) values of 0.83, 10.99, and 2.45, respectively. Furthermore, using DCGAN-generated samples to train the SG-1D-CNN by adding a certain number of generated samples can enhance the model's performance in the test set. When the number of added samples is 60 (75% of the original training set sample size), the SG-DCGAN-1D-CNN (Savitzky-Golay deep convolutional generative adversarial network one dimensional convolutional neural network) exhibits the best performance, with R², RMSE, and RPD reaching 0.87, 9.40, and 2.88, respectively. The DCGAN-1D-CNN (Deep convolutional generative adversarial network one dimensional convolutional neural network) method based on this strategy is expected to provide new insights into precise prediction for multi-variety rice seeds.