Non-destructive geographical traceability of American ginseng using near-infrared spectroscopy combined with a novel deep learning model

Abstract

American ginseng is a renowned medicinal herb that falls under the category of medicine food homology. The pharmacological benefits of American ginseng vary based on its origin, and accurately tracing its origin in a non-destructive and quick manner remains a challenge. This study presents an approach that utilizes Near-infrared (NIR) spectroscopy and a novel deep learning model called AGOTNet to accurately identify the origin of American ginseng. This approach offers the benefit of being rapid and non-destructive. The AGOTNet utilizes three external self-attention modules of different sizes to create its backbone for extracting multi-level features (local and global features) and multi-varieties features (data and dataset-level features). The classification head network, consisting of fully connected layers, employs these features effectively to determine the origin of American ginseng. AGOTNet and its four competitors are trained and tested using a dataset containing 2240 samples from five different origins. The experimental results demonstrated that the proposed method outperformed the other four methods, achieving overall accuracy, precision, recall, F1 score, MCC, and AUC values of 98.95 %, 98.97 %, 98.96 %, 98.95 %, 98.65 %, and 99.60 % respectively for the testing samples. The contents of six ginsenoside components in samples were determined simultaneously using HPLC. The study applied partial least-squares-discriminant analysis and principal components analysis to discover the specific ginsenoside components that are impacted by the origin of the American ginseng samples and to classify them accordingly. In conclusion, it is possible to employ NIR spectroscopy combined with deep learning models to rapidly and non-destructive identify the source of American ginseng.