Online detection of naturally DON contaminated wheat grains from China using Vis-NIR spectroscopy and computer vision

Deoxynivalenol (DON) contamination of wheat grains is a serious problem in China, and it is necessary to remove contaminated wheat before it enters the consumer market. In this study, visible-near infrared (Vis-NIR) spectroscopy and computer vision techniques were combined to simulate online discrimination between normal and DON-contaminated wheat grains. Naturally growing wheat samples were collected from several of the main wheat-producing areas in China, the reference DON contents were measured by using liquid chromatography serial triple quadrupole mass spectrometer (LC-MS), and then wheat samples were divided into two categories according to the national standard of 1 mg kg−1. The characteristic spectral variables, colour and texture features were extracted and integrated for chemometric analysis. Principal component analysis based on fusion features indicated better clustering than with just spectral features. Subsequently, linear discriminant analysis modelling based on spectra and texture features achieved the best discrimination with an accuracy of 95.06% and 91.36% for calibration and validation sets respectively, which was 5% higher than with just spectral features, and the false positive rates (FPR) were the lowest: 3.41% and 10.42% for calibration and validation sets respectively. The internal scanning results of whole wheat flour indicated that the higher the content of DON, the looser the binding of starch granules, which could cause the textural change of wheat grains. The research showed that Vis-NIR spectroscopy combined with computer vision has the potential to be used in the non-destructive and online detection of DON-contaminated wheat grains; further study on the interference from complex environments is still need for actual online detection.