To address the challenge of discriminating Baijiu base liquor grades using near infrared (NIR) spectroscopy caused by highly similar chemical compositions and severe spectral overlap, this study proposes a robust NIR spectroscopy analytical strategy based on dual-domain feature decoupling. Unlike traditional approaches that treat background variations as noise, the proposed method simultaneously decodes chemical absorption features and spectral morphological features from the same signal. Specifically, first-order Savitzky–Golay (SG) derivatives are employed to enhance narrow-band absorption peaks associated with functional groups such as C–H and O–H, while continuous wavelet transform (CWT) is utilized to extract global morphological variations related to scattering effects and baseline drift. To improve the robustness of high-dimensional feature selection, a stability-guided screening method based on dynamic reference features is introduced to effectively retain consistently informative variables. Experimental results demonstrate that this method significantly reduces feature dimensionality while boosting classification accuracy to 97.56% in cross-validation and 97.78% on an independent test set. Further analysis indicates that spectral morphological features provide complementary discriminatory evidence to chemical absorption information, thereby enhancing model robustness under complex matrix conditions. This study provides a practical reference strategy for the grading of Baijiu base liquor.
Zhang et al. (Sat,) studied this question.