Identification of geographical origin and adulteration of Northeast China soybeans by mid-infrared spectroscopy and spectra augmentation

Mathematical models based on infrared spectroscopy and machine learning have been successfully used to trace the origin of soybeans. However, as previous research reported, it is necessary to employ spectra data that undergo multiple pre-processing operations in order to achieve optimal accuracy during model training. And these established models are only capable of predicting samples with identical spectra pre-processing. Specifically, baseline correction, which necessitates individual processing of each spectrum, requiring substantial investments of time and human resources with a large dataset. In this study, the spectra augmentation technique was proposed based on the theory of data augmentation, in order to simplify or even eliminate the pre-processing steps for the prediction dataset. The technique utilized a combination of the standard spectra pre-processed data and the “boost data” to train models, specifically, a total of 180 spectra, including 90 pre-processed standard spectra and 90 “boost” spectra. The “boost” data refers to data without the standard spectra pre-processing. On the prediction dataset without the standard spectra pre-processing, the model with the spectra augmentation technique had an accuracy of 0.91 for the recognition of Northeast China soybeans, while the accuracy of the model with the training method frequently reported in previous studies only reached 0.71, demonstrating that the model trained with the proposed technique possessed higher robustness and generalization capabilities. The spectra augmentation technique can maintain high accuracy while simplifying spectra pre-processing steps on prediction data, therefore, providing a more efficient and expedited method for practical food traceability and authentication.

Graphical abstract