A dual-branch CNN-encoder model for rapid geographic traceability of Astragalus using near-infrared spectroscopy

The geographical origin of medicinal herbs significantly affects their chemical composition and pharmacological efficacy. Therefore, developing rapid and non-destructive origin identification techniques is essential for quality control in traditional Chinese medicine (TCM). Astragalus membranaceus, a representative herb with both medicinal and nutritional value, is highly sensitive to environmental conditions across different producing areas. This study proposes a novel classification approach that integrates near-infrared (NIR) spectroscopy with a dual-branch deep neural network architecture, termed Dual-TCNet, for rapid origin traceability of Astragalus. The model combines convolutional neural networks (CNN) and a modified encoder to extract local and global spectral features, respectively, and employs a multi-head attention mechanism to fuse multi-scale representations. Experiments were conducted on 588 Astragalus samples collected from five geographical regions. The proposed Dual-TCNet achieved an accuracy of 98.31%, a precision of 98.39%, and a recall of 98.32%, all outperforming baseline models including SVM, RF, and 1D-CNN. These results demonstrate that the proposed method offers high classification performance and practical potential for efficient and non-destructive traceability of TCM origins.