Detection of Frost-Damaged Maize Kernels for Food Quality Screening Using Hyperspectral Imaging and 2D-COS-Assisted Deep Learning.

Maize is a crucial cereal crop susceptible to subzero temperature stress, which causes physicochemical damage to kernels and compromises their processing quality and food value. Traditional analytical methods are destructive, labor-intensive, and unable to meet the requirements of rapid and high-throughput quality screening in the food supply chain. This study established a nondestructive detection method combining hyperspectral imaging (900-1700 nm) and deep learning for the identification of frost-damaged maize kernels. Multiple scattering correction (MSC) was verified as the optimal spectral preprocessing approach for improving spectral stability and classification performance. For feature extraction, Two-Dimensional Correlation Spectroscopy (2D-COS) effectively enhanced frost-induced spectral variations by resolving overlapping and weak spectral responses, outperforming CARS and VCPA-IRIV and extracting 432 characteristic bands associated with frost damage. The constructed 2D-COS-CNN model achieved a prediction accuracy of 0.9700 and a recall of 0.9647, with a calibration set accuracy of 0.9967, demonstrating excellent classification capability and reliable generalization performance. The integration of 2D-COS-based feature enhancement and CNN-based nonlinear discrimination provides an effective strategy for rapid and nondestructive identification of frost-damaged maize kernels, offering a reliable analytical tool for grain quality evaluation, damage-level sorting, and processing suitability assessment in the food supply chain.