Chemometric evaluation and machine learning classification for near-infrared hyperspectral analysis of heat-damaged soybeans

Imported soybeans frequently undergo thermal damage during prolonged transit, necessitating robust detection methods. This study applied chemometric analyses to characterise interrelationships among key quality parameters—including colour coordinates ( L *, a *, b *), protein content, water-soluble protein, acid value, nitrogen solubility index (NSI), fat content, and moisture content—in thermally damaged soybeans. PCA, hierarchical clustering, and Pearson correlation identified moisture content, water-soluble protein, and a * as core indicators: grey relational analysis and entropy weighting assigned metric weights to establish a thermal damage assessment framework. We integrated near-infrared hyperspectral imaging (NIR-HSI) with these core indicators and machine learning to develop a classification model, evaluating the impact of sample grinding on performance. After MSC-CARS preprocessing, the CNN model achieved 99.13% training and 88% testing accuracy, providing an efficient non-destructive methodology for industrial quality assessment. • Moisture, water-soluble protein, and a* value were core thermal damage indicators. • Grey relational and entropy weighting established a comprehensive evaluation model. • CNN applied to NIR-HSI data yielded 88% classification accuracy. • Fine grinding enhanced CNN performance by improving spectral discriminability. • MSC-CARS streamlined feature extraction for superior classification.