Abstract
Black tea is a widely consumed beverage whose high economic value has led some producers to illegally add artificial colorants such as Sunset Yellow, Tartrazine, and Ponceau 4R, posing health risks. Although near-infrared (NIR) spectroscopy offers a rapid, non-destructive detection method, its use in trace-level colorant detection is limited due to low adulterant concentrations and interference from natural tea pigments. Hence, we developed a rapid, non-destructive method for detecting trace adulteration (from 0.1 to 0.5 g·kg(-1)) in black tea with artificial colorants using a handheld near-infrared spectrometer. To enhance sensitivity to low-level adulteration, we proposed a novel Spectral Multi-scale Attention Fusion Network (SMAFNet), designed to dynamically integrate multiscale features. SMAFNet consists of spectral preprocessing, multi-scale feature extraction, and cross-scale attention fusion modules. Comparative experiments with traditional machine-learning models demonstrated that SMAFNet achieved superior performance even at low adulteration levels. Sample sets (each including 36 samples) adulterated with Sunset Yellow, Tartrazine, and Ponceau 4R, SMAFNet achieved accuracies of 97.22-100%, F1-scores of 0.9879-1.00, and 100% recall. These findings confirm the feasibility and robustness of combining NIR with SMAFNet for the rapid and discriminative detection of trace colorants in black tea, offering a practical framework for on-site food safety monitoring and quality control.