Hybrid near-infrared and chemical-based machine learning enhances the reliability of chili powder origin classification

Accurate origin classification of chili powder is essential for consumer trust and regulatory compliance. In this study, combined near-infrared (NIR) spectroscopy and chemical composition analysis were integrated with machine learning to classify domestic (n = 54, Korea) and imported (n = 66, China and Vietnam) chili powder samples. Baseline analysis revealed systematic differences: domestic powders showed higher protein, calcium, and moisture contents, whereas imported samples contained more organic acids, sugars, and capsaicinoids. Using 16 NIR bands selected by the least absolute shrinkage and selection operator (LASSO), support vector machine (SVM) models achieved high accuracy, with Savitzky–Golay first derivative plus standard normal variate preprocessing yielding the best performance. The hybrid models enhanced reliability. NIR alone achieved high origin-classification accuracy in this dataset using as few as four selected bands; however, NIR combined with organic acid variables (NIR + org) consistently achieved 100% accuracy and showed improved probability reliability. Shapley additive explanation analysis showed that while O–H and C–H overtone bands drove the NIR spectral band-only models, the hybrid models emphasized organic acids and proximate components, providing clear chemical interpretability. The findings demonstrated that integrating NIR with targeted chemical variables enables robust, reliable, and interpretable origin classification, offering rapid screening and regulatory assurance.