Data fusion of visible near-infrared and mid-infrared spectroscopy combined with feature selection and machine learning for rapid discrimination of fusarium head blight infection in wheat kernel and flour

Abstract

Fusarium head blight (FHB) is a significant crop fungal disease that downgrades the yield quality and affects food safety. There is a necessity for the development of fast and cost-effective detection approaches of FHB to meet the needs of the food industry, as the traditional methods are slow, costly, difficult, and expose chemicals to the environment. Visible near-infrared (vis-NIR) and mid-infrared (MIR) spectroscopy have been used as promising tools for the detection of FHB contamination and mycotoxins in cereal crops and foods. This study explores the potential of data fusion approaches of the vis-NIR (400–1650 nm) and MIR (4000 – 650 cm⁻¹) spectra for FHB detection in wheat kernels and flour of eight varieties of wheat. Spectra concatenation and feature selection methods were utilized as input data for two different machine learning models, namely, random forest (RF), and decision tree (DT). For the selection of the most informative wavebands from both sensors, genetic algorithm (GA), recursive feature elimination (RFE), and principal component analysis (PCA), were employed. Results showed that spectral concatenation data fusion has resulted in very high test accuracy for FHB detection in both kernel and flour, with all models reaching 100% classification accuracy, except the RF-kernel model, which achieved 96.6%. Among the three feature selection algorithms, GA was the best method for the selection of the most informative bands related to FHB, resulting in a correct classification accuracy of 100 %, for both RF and DT modelling tools. For the RFE feature selection method, a lower classification accuracy of 96.6 % was obtained with both RF and DT models in kernels. However, PCA resulted in the lowest accuracies, dropping down by 10.3 % to 17.3 %, compared to that of GA and RFE, respectively. Overall, the proposed data fusion methods allow the non-destructive, rapid, and accurate detection of FHB infection in wheat flour and kernels. This is particularly useful for the flour as it is not possible to visually estimate the infected from healthy samples.