Hyperspectral imaging combined with genetic algorithm-optimized machine learning for the rapid and nondestructive quantification of ginsenosides and polysaccharides in Panax ginseng

Ginseng, a high - value food and medicinal ingredient, is influenced by origin. Prior studies used spectroscopy for component analysis, but model optimization, band application, and field use remain challenging. This study presents a novel method using hyperspectral imaging to create a dual - function non - destructive model. It can measure ginsenoside (RS) and polysaccharide (RP) contents simultaneously and identify the geographic origin.The integrated methodology includes two parts. First, it uses component - specific algorithms for quantitative prediction. ELM performs well in RS prediction in VNIR band (R² = 0.86), and GA - optimized back - propagation neural network is excellent in RP prediction (R² = 0.85). Second, it classifies the origin. It can use second - order derivative preprocessing with particle swarm optimization (accuracy = 0.91 in VNIR) or standalone processing (accuracy = 0.91 in short - wave infrared). The developed four - dimensional optimization framework ("component - band - model - feature") shows VNIR's detection advantages for both components and improves feature selection stability by 30% through GA iterative voting. Therefore, these findings establish an efficient, nondestructive ginseng quality - control tool enabling food producers to verify composition and origin during raw material screening and production standardization. • Ginseng composition is strongly affected by its geographic origin • A nondestructive model for determining ginseng composition is developed • An extreme machine learning model excels in predicting ginsenoside content • Second-order derivative preprocessing is employed for ginseng origin identification