Fusion of Visible−Near-Infrared and Mid-Infrared Spectra for Deoxynivalenol Detection: A Laboratory-Based Study

Deoxynivalenol (DON), a mycotoxin produced by Fusarium species, poses a major threat to wheat safety and quality worldwide. Conventional detection methods provide excellent accuracy and precision but are costly, labor-intensive, and unsuitable for large-scale or on-site screening. Spectroscopic techniques, particularly visible−near-infrared (vis−NIR) and mid-infrared (MIR) spectroscopy, provide rapid and nondestructive analysis with minimal sample preparation. This study investigates whether data fusion of vis−NIR and MIR spectra can enhance DON detection in wheat kernels and flour under laboratory conditions. A total of 143 wheat samples were analyzed using vis−NIR (400−1650 nm) and MIR (4000−650 cm−1) spectroscopy. DON concentrations were determined by high-performance liquid chromatography−mass spectrometry and categorized according to EU regulatory thresholds. Five machine learning models—partial least-squares discriminant analysis (PLS-DA), support vector machine (SVM), decision tree, light gradient boosting machine, and one-dimensional convolutional neural networks—were trained on single-sensor, concatenated, and feature-selected fused spectra. Results showed that vis−NIR provided stronger classification performance than MIR alone, with flour samples consistently outperforming kernels. Spectral concatenation improved performance, with PLS-DA achieving the highest accuracies of 96.5% for flour and 92.7% for kernels. Recursive feature elimination revealed model-dependent effects: tree-based and convolutional models benefited from reduced feature sets, while PLS-DA and SVM performed optimally with complete spectral information. These findings demonstrate that fusion of vis−NIR and MIR spectra, coupled with appropriate model selection, substantially improves nondestructive DON classification in wheat, offering a pathway toward rapid, cost-effective monitoring systems for enhanced grain and flour quality assessment.