Discrimination of Chinese Baijiu Base Liquor Grades Using Near-Infrared Spectroscopy Based on SHAP and Multi-scale Pyramid Feature Fusion

Abstract

To address the challenges of high-dimensional redundancy in near-infrared spectroscopy (NIRS) for grading Chinese Baijiu base liquor and to overcome the limitations of traditional feature selection methods—which often lose local spectral shape information due to an excessive focus on discrete wavelengths—this study proposes a dual-feature optimization framework. The methodology initiates by circumventing conventional distance metrics through a graph-based splitting strategy that leverages spectral clustering to characterize the manifold structure of samples within the feature space. Following this, a "point-plane" fusion mechanism is introduced: SHAP (Shapley Additive Explanations) values derived from an eXtreme Gradient Boosting (XGBoost) model are utilized to quantify the global contribution of each wavelength, thereby facilitating the selection of key discrete wavelengths (points). Simultaneously, multi-scale pyramid features are incorporated to extract local texture and shape information (representing the 'planes') from the spectra via hierarchical statistical pooling, resulting in a more refined representation of NIRS data. Ultimately, the XGBoost model is implemented for classifying base liquor grades. On the graph-based hold-out partition, the final SHAP–Pyramid–XGBoost model achieved an Accuracy of 0.9688 and a Macro-F1 score of 0.9696. Repeated nested stratified five-fold cross-validation with five repeats yielded a more conservative Accuracy of 0.9083 ± 0.0237 and a Macro-F1 score of 0.9063 ± 0.0241. Evaluation on three public spectroscopic benchmark datasets further indicated that the contribution of pyramid fusion was dependent on the dataset and wavelength-selection method. These results support promising internal discrimination performance on the present Baijiu dataset, while independent multi-site validation remains necessary.