Non-destructive prediction of the moisture content of individual wheat kernels combining hyperspectral imaging and...

Moisture content is a crucial factor that significantly impacts grain quality and safety. In this study, hyperspectral imaging (HSI) and Wasserstein generative adversarial networks (WGAN) were employed to predict the moisture content of individual wheat kernels. To address the varying data sizes in the visible and near-infrared (Vis-NIR) and short-wave infrared (SWIR), WGAN was designed with two separate network architectures for augmenting spectral and moisture content data. The generated data was validated as plausible by comparing it with real data and applying the t-SNE dimensionality reduction algorithm. The optimal regression models, obtained by selecting the optimal pre-processing method and the most effective wavelengths in the Vis-NIR and SWIR, respectively, were the 1st-derivative-SPA-CNN (R_P² = 0.9371, RMSEP = 0.5717, RPD = 3.8955, RER = 15.2713) in the Vis-NIR and the 1st-derivative-ReliefF-CNN (R_P² = 0.8095, RMSEP = 1.0009, RPD = 2.3201, RER = 8.6116) in the SWIR. Model inversion was applied to single pixels to visualize the spatial distribution of moisture content within individual wheat kernels. Overall, this study proposes a novel approach to enhance the accuracy of predicting moisture content, thereby improving grain processing quality and safety monitoring.