Comparative analysis of spectral variable selection methods for NIR-based multi-component detection of Xanthoceras sorbifolium Bunge seed kernels

The moisture, oil, protein, and nervonic acid contents are crucial criteria for assessing the nutritional value and processing quality of Xanthoceras sorbifolia Bunge seed kernels (XSKs). Near-infrared spectroscopy (NIRS) offers a viable alternative to conventional chemical methods for rapid and non-destructive sample analysis. This study employed a highly adaptive method known as competitive adaptive reweighted sampling (CARS), which facilitates variable selection for different components and streamlines the feature-variable screening process. Using this method, various algorithms were applied to establish NIRS detection models for the four components of XSKs, and accuracy evaluations were conducted using the prediction set determination coefficient (${\text{R}}_{\text{P}}^{\text{2}}$), root-mean-square error (RMSEP), and other parameters. The results indicated that CARS effectively extracted key information on various components in NIRS. Based on the feature variables selected by CARS, the partial least squares regression (PLSR) model exhibited optimal performance regarding moisture (${\text{R}}_{\text{P}}^{\text{2}}$ = 0.978, RMSEP = 0.117 %) and nervonic acid content (${\text{R}}_{\text{P}}^{\text{2}}$ = 0.988, RMSEP = 0.072 %). Meanwhile, the particle swarm optimization-enhanced support vector regression (PSO-SVR) model and back-propagation neural networks (BPNN) achieved superior performance for protein (${\text{R}}_{\text{P}}^{\text{2}}$ = 0.989, RMSEP = 0.243 %) and oil content (${\text{R}}_{\text{P}}^{\text{2}}$ = 0.978, RMSEP = 0.467 %), respectively. These models enable rapid and non-destructive detection of multi-component content, thereby facilitating the expeditious identification and selection of XSKs with high nutritional value and superior quality.