Prediction of Deoxynivalenol contamination in wheat kernels and flour based on visible near-infrared spectroscopy, feature selection and machine learning modelling.
Journal:
Spectrochimica acta. Part A, Molecular and biomolecular spectroscopy
PMID:
39805244
Abstract
Contamination of wheat by the mycotoxin Deoxynivalenol (DON), produced by Fusarium fungi, poses significant challenges to the quality of crop yield and food safety. Visible and near-infrared (vis-NIR) spectroscopy has emerged as a promising, non-destructive, and efficient tool for detecting mycotoxins in cereal crops and foods. This study aims to utilize vis-NIR spectroscopy, coupled with a feature selection technique and machine learning modelling, to predict and classify DON contamination in wheat kernels and flour. A total of ninety-five samples, collected from commercial wheat fields in Lithuania and Belgium, were scanned using a vis-NIR (400-1650 nm) spectrophotometer. The DON content was subsequently determined by a liquid chromatography-mass spectrometry (LC-MS). The data were preprocessed and analyzed using random forest classifier (RFC), and regressor (RFR), extra trees classifier (ETC), and regressor (ETR), AdaBoost classifier (ABC), and regressor (ABR) for classification and regression tasks, respectively. To enhance model accuracy, recursive feature elimination (RFE) algorithm to select the most informative wavebands was applied, and the random over-sampler (ROS) was adopted to mitigate the imbalance of data in DON classes. Results showed that the feature selection approach improved the prediction and classification accuracy of the models. Notably, the performance of the algorithms was better for the flour samples compared to the kernels. The most effective DON prediction model was achieved with the ETR-RFE modelling approach for the flour samples, demonstrating high accuracy [determination coefficient (R) = 0.94 and root mean square error of prediction (RMSEP) = 3.42 mg.kg]. On the other hand, the ETC applied to the full spectrum data along with ROS, achieved the highest classification accuracy of 89.5 %. These results demonstrate the potential of using vis-NIR with RF-RFE modelling approach, for rapid analysis of DON levels in wheat kernel and flour.