Codebook-constrained Spectral Synthesis: A Vector Quantized-Augmented Generative Algorithm Framework for Predicting Corn Starch Content

Starch, the predominant component of maize, is integral to numerous industrial processes, including ethanol production and adhesive manufacturing. Accurate and efficient quantification of starch content in maize is essential for optimizing industrial productivity and reducing operational costs of maize production. This study aimed to develop a robust predictive model using hyperspectral imaging combined with advanced generative algorithms to enable the rapid, non-destructive determination of maize starch content. A total of 80 maize samples were collected from various locations in Ningxia, China, to establish a quantitative relationship between hyperspectral responses and starch content. The dataset was augmented using a Vector Quantized Variational Autoencoder (VQ-VAE), which discretizes continuous spectral features into a structured codebook space, significantly enhancing data diversity and model generalization capabilities. The experimental results demonstrated that the dataset augmented with VQ-VAE significantly improved the accuracy and stability of the regression models. Compared to conventional autoencoders (AE) and Generative Adversarial Networks (GAN), VQ-VAE-generated data preserved the critical characteristics of the original dataset while expanding its distribution range, thus enhancing the robustness of the model. A Convolutional Neural Network (CNN) model trained on this enriched dataset achieved an R2 value of 0.7510 and a Root Mean Square Error (RMSE) of 0.4413 on the test set, underscoring its superior predictive accuracy and generalization potential. This study provides a novel methodological framework for the rapid, non-destructive, industrial-scale determination of maize starch content.