Detection of Fungus-Infected Corn Kernels Using Near-Infrared Reflectance Spectroscopy and Color Imaging

Contamination of grain products by fungus can lead to economic losses and is deleterious to human and livestock health. Detection and quantification of fungus-infected corn kernels would be advantageous for producers and breeders in evaluating quality and in selecting hybrids with resistance to infection. This study evaluated the performance of single-kernel near-infrared reflectance spectroscopy (NIRS) and color imaging to discriminate corn kernels infected by eight fungus species at different levels of infection. Discrimination was done according to the level of infection and the mold species. NIR spectra (904 to 1685 nm) and color images were used to develop linear and nonlinear prediction models using linear discriminant analysis (LDA) and multi-layer perceptron (MLP) neural networks. NIRS was able to accurately detect 98% of the uninfected control kernels, compared to about 89% for the color imaging. Results for detecting all levels of infection using NIR were 89% and 79% for the uninfected control and infected kernels, respectively; color imaging was able to discriminate 75% of both the control and infected kernels. In general, there was better discrimination for control kernels than for infected kernels, and certain mold species had better classification accuracy than others when using NIR. The vision system was not able to classify mold species well. The use of principal component analysis on image data did not improve the classification results, while LDA performed almost as well as MLP models. LDA and mean centering NIR spectra gave better classification models. Compared to the results of NIR spectrometry, the classification accuracy of the color imaging system was less attractive, although the instrument has a lower cost and a higher throughput.