Weijun Xie

Maize kernel hardness is a critical indicator for processing quality and storage stability. To overcome the drawbacks of conventional destructive and time-consuming methods, this study proposes a non-destructive approach using hyperspectral imaging (374.98–1038.79 nm). A novel spectrum-image fusion strategy was implemented to develop separate, lightweight, and high-precision prediction models for high- and low-hardness kernels. Various preprocessing, feature selection, and modeling algorithms were systematically compared. For high-hardness kernels, an SVR model with SNV preprocessing and SPA+CARS selection achieved outstanding performance (test set R²=0.9518, RPD=4.5533) using only 30 key wavelengths, reducing data dimensionality by over 75%. For low-hardness kernels, a robust 1D-CNN model was constructed by fusing 15 spectral features with 10 image features (test set R²=0.926, RPD=3.6772). The study successfully enables rapid and accurate hardness detection, providing a practical tool for industrial quality control and grading.