SPIRAL: A probabilistic deep learning framework for Chinese liquor (Baijiu) classification via near-infrared hyperspectral imaging

Chinese liquor (Baijiu) is a complex alcoholic beverage whose chemical composition determines its alcohol content, aroma type, and brand. Accurate identification of these properties is critical for quality control and food safety. Conventional analytical methods, such as gas chromatography–mass spectrometry, are often time-consuming and destructive. In this study, we introduce near-infrared hyperspectral imaging (NIR-HSI) combined with the spectral-probabilistic-inference-and-recognition-for-alcoholic-liquor (SPIRAL) framework for rapid, non-destructive analysis of Chinese liquors. SPIRAL employs adaptive variance learning to model spectral uncertainties arising from the complex chemical matrix and measurement variability. To validate the proposed framework, we developed the Chinese liquor hyperspectral imaging dataset (CLHID), comprising 1000 NIR-HSI images of 49 liquors, covering 9 alcohol contents, 5 aroma types, and over 30 brands. SPIRAL achieved excellent classification performance on CLHID, with F1-scores of 99.72% for alcohol content, 92.82% for aroma type, and 89.71% for brand identification, highlighting its strong potential for industrial food quality assurance applications.