A hybrid LD-PCA approach for dimensionality reduction and classification in near-infrared spectroscopy: A case study...

Near-infrared spectroscopy (NIRS) data are typically characterized by high dimensionality and complexity, making dimensionality reduction essential for reliable classification. Conventional principal component analysis (PCA) preserve variance but ignores class information, while linear discriminant analysis (LDA) enhances separability yet suffers from singularity in the intra-class scatter matrix. To address these limitations, this study proposes a hybrid feature extraction method, Linear Discriminant Principal Component Analysis (LD-PCA). LD-PCA integrates the inter-class separability of LDA with the variance preserving capability of PCA, while incorporating identity matrix regularization to enhance stability. The proposed method was evaluated on a near-infrared dataset of 59 exported green tea samples comprising three different quality grades. Three classifiers-Naïve Bayes, partial least squares discriminant analysis, and random forest-were employed, and the performance of LD-PCA was benchmarked against PCA, LDA, Elastic Net and Kernel PCA for the task of classifying green tea by quality grade. Across 100 Monte Carlo experiments, LD-PCA consistently achieved average accuracy, precision, recall and F1-score above 95%. Furthermore, the projection weights of LD-PCA effectively highlighted wavelength regions associated with key chemical components, confirming its spectrochemical interpretability. These results confirm that LD-PCA offers a robust, interpretable, and balanced dimensionality reduction strategy for high-dimensional NIRS analysis.