Non-destructive geographical origin authentication of licorice using hyperspectral imaging and a component-aware Swin...

Accurate authentication of geographical origin is critical for ensuring the quality, safety, and traceability of medicinal and functional food materials such as Glycyrrhiza uralensis Fisch. (licorice). However, reliable origin discrimination remains challenging because licorice samples from different producing regions often exhibit subtle compositional differences but highly overlapping spectral signatures. In this study, a hyperspectral imaging (HSI)-based deep learning framework, termed Component-Aware Swin Transformer (CAST), was proposed for non-destructive origin authentication of licorice. A sliding-window-based local principal component analysis (PCA) guided by the Landgrebe criterion was first employed to reduce spectral redundancy under high-dimensional small-sample conditions. A PCA-aware multi-scale feature extractor incorporating a Component-Aware Squeeze-and-Excitation mechanism was designed to adaptively recalibrate heterogeneous principal components. In addition, a 3D Swin Transformer block was introduced to capture long-range spectral-spatial dependencies, while a joint optimization strategy combining Center Loss and Label Smoothing improved feature compactness and class separability. Experiments conducted on 1045 licorice samples from three major producing regions showed that the proposed CAST model achieved an overall accuracy of 95.22%, outperforming the strongest baseline HyperSFormer by 2.54 percentage points. These results demonstrate the potential of integrating hyperspectral imaging with deep learning for reliable food origin authentication and intelligent quality control.