Eunice Oluwabunmi Owoola ,  Yingzhe Lu ,  Qinglu Yang  ,  Mochen Liu ,  Jing Wang  ,  Yinfa Yan ,  Yao Lu

Abstract 

Toxic Illicium adulterants are visually similar to Illicium verum (genuine star anise), making rapid, non-destructive authentication difficult in routine food safety inspection. This study developed a fruit-level hyperspectral discrimination framework based on a deep derivative-enhanced spectral attention network (DSAN) to distinguish Illicium verum from toxic Illicium adulterants (Illicium lanceolatum and Illicium henryi). Intact Illicium samples were imaged using visible/near-infrared (Vis/NIR) hyperspectral imaging, and fruit-level representative spectra were extracted from segmented plate-level hyperspectral cubes for subsequent analysis. A lightweight wavelength-wise spectral attention mechanism was applied to a derivative-enhanced input formed by concatenating raw spectra with their first- and second-order numerical derivatives, enabling the network to emphasize discriminative spectral features. The evaluation of model performance was conducted using plate-grouped leave-one-plate-out validation with 10 repeated runs. The results were then compared with four classical chemometric models and five deep learning baselines. To prevent the leakage of information, feature filtering and z-score normalization were performed using training-fold statistics. DSAN achieved an accuracy of 99.13 ± 0.41%, a balanced accuracy of 99.03 ± 0.43%, a recall of 98.54 ± 0.73% for toxic adulterants, and an F1-score of 98.90 ± 0.52%. In comparison, the baselines achieved 80.25-97.08% accuracy, 79.93-96.53% balanced accuracy, 74.69-96.88% recall, and 75.96-96.26% F1-score. These results demonstrate the effectiveness of derivative-enhanced spectral attention learning for fruit-level discrimination of genuine Illicium verum from toxic Illicium adulterants.

Paper Linkage:https://doi.org/10.1016/j.foodcont.2026.112560