Discrimination between domestic and imported red chili powders using laser-induced breakdown spectroscopy: variable selection and comparative evaluation of machine learning algorithms

Abstract

Ensuring the authenticity of domestically produced red chili powder is crucial for consumer trust in Korea. Here, laser-induced breakdown spectroscopy (LIBS) was combined with systematic variable selection and multiple machine-learning algorithms to classify 79 commercial samples (43 domestic and 36 imported). From the LIBS spectra, nine representative elemental emission lines (Mg, Li, Rb, Ca, K, Na, H, C, and O) were chosen via interclass-distance analysis, and all 511 non-empty variable subsets were modeled using linear discriminant analysis (LDA), k-nearest neighbors (k-NN), Naive Bayes (NB), support vector machine (SVM), and random forest (RF). The highest leave-one-out cross-validation (LOOCV) accuracy was 97.5% and it was achieved with only two or three variables (Mg + Li, Mg + Li + Rb, or Mg + Li + Ca), showing that a compact feature set suffices for reliable discrimination. Bootstrapped random-split validation (100 independent 70 : 30 partitions) further confirmed robustness: mean accuracies were consistent with LOOCV within one standard deviation, and variable-selection frequencies were strongly concentrated on Mg, Li, and Rb (with Ca occasionally replacing Rb). Across models, Mg, Li, Rb, and Ca consistently exhibited the strongest discriminative power, reflecting agro-environmental and geological contrasts between production regions. Overall, LIBS with principled variable selection provides a rapid, cost-effective, and reproducible framework for differentiating domestic and imported red chili powders.