Lightweight identification of Angelica sinensis via deep learning-guided spectral feature reduction

Angelica sinensis (Danggui) is a popular botanical dietary supplement and functional food ingredient widely consumed in Asian cuisine. However, it is frequently adulterated with morphologically similar species, such as Levisticum officinale, Angelica pubescens, and Angelica acutiloba, which compromises food safety and consumer interests. To address this issue, a rapid and non-destructive identification framework was developed using near-infrared hyperspectral imaging (HSI) combined with deep learning. A Gramian Angular Field-Convolutional Neural Network (GAF-CNN) was constructed to transform 1D spectral data into 2D images, enabling the capture of complex non-linear spectral-spatial features regarding chemical composition. The GAF-CNN model achieved an overall classification accuracy of 98.52% on the independent test set. To improve model interpretability and reduce computational redundancy, gradient-based saliency analysis was employed to visualize the decision-making process, pinpointing 30 characteristic wavelengths primarily associated with O-H and C-H stretching vibrations. Remarkably, when these selected bands were applied to a simplified Linear Discriminant Analysis (LDA) classifier, the model achieved 100% accuracy, precision, and recall. These results demonstrate that the proposed deep learning-guided feature reduction strategy can successfully extract the most discriminative chemical markers, providing a theoretical foundation for developing low-cost, portable multispectral detection devices for quality control in the herbal food market.