A new data processing strategy combined with a convolutional neural network for rapid and accurate prediction of geographical classifications of natural products

Geographic traceability has attracted attention in the field of food and natural products. It is related to security and quality and is inseparable from the vital interests of producers and consumers. A conventional method for geographic traceability is to combine chemical fingerprints with chemometrics. However, insufficient information is provided from a single detection system for high accuracy traceability. This study uses 'Two-Elements-Multi-Information-Fingerprints' (TEMIF) to obtain more fingerprint information, implemented on Cnidium monnieri (L.) Cuss from six different provinces in eastern China. Due to differing polarities, the chemical compositions of the sample can be divided into weak polar parts and strong polar parts. Different chromatographic systems were used to detect different polar compositions. All chromatograms obtained were deconstructed and reconstructed into TEMIF. The frequently used classification models have poor performance in such complex data. Therefore, a convolutional neural network (CNN) was used to process TEMIF. As a result, the more information carried on the fingerprint, the higher is the model's classification accuracy. The accuracy of the CNN model was much higher than that of the traditional machine learning model. Under the same data, support vector machine has the highest accuracy of 62.67% among the three traditional machine learning models, while CNN model can reach 99.80%. On the premise of enriching chemical information, our strategy greatly improved accuracy and made the geographic traceability model more rapid. The online automatic geographic traceability could be realized using a combination of the model and chromatographic workstation.