Interpretable wavelength selection for complex matrices: A multi-stage feature engineering framework for baijiu characterisation using near infrared spectroscopy

Severe collinearity, broad-band overlap, and high-dimensional redundancy remain major challenges for near infrared (NIR) analysis of complex liquid matrices such as Chinese Baijiu. To address feature redundancy, selection variability, and limited model interpretability, this study proposes a multi-stage feature-selection framework termed AW-IVCPA-APSO. The framework was evaluated using 482 base Baijiu samples from 35 production batches under a batch-exclusive 5 × 4 nested cross-validation strategy. The AW-IVCPA-APSO approach reduced the original 1226 spectral variables to 59.0 ± 7.4 features, with a Jaccard similarity of 43.9 ± 6.4% across the batch-disjoint outer folds. Among the evaluated classifiers, XGBoost achieved the highest mean Macro-F1 of 94.05 ± 2.20% and Accuracy of 93.98 ± 2.26%, demonstrating a favorable balance among dimensionality reduction, classification performance, and feature-selection reproducibility. SHAP-based interpretation further showed that the discriminative information was mainly concentrated around 5800–6100 cm -1 , together with several complementary spectral regions, and exhibited consistent importance across different tree-based models and production batches. The dominant region is primarily associated with C–H first-overtone absorptions and may reflect the combined spectral contributions of ethanol and other C–H-containing organic constituents in the complex Baijiu matrix. Overall, AW-IVCPA-APSO provides a compact and interpretable chemometric approach for rapid and non-destructive classification of base Baijiu and demonstrates its potential for NIR-based analysis of complex food matrices.