A novel hyperspectral-based approach for identification of maize kernels infected with diverse Aspergillus flavus fungi

Authors: TAO, FEIFEI - Mississippi State University; YAO, HAIBO - Mississippi State University; HRUSKA, ZUZANA - Mississippi State University; KINCAID, RUSSELL - Mississippi State University; Rajasekaran, Kanniah - Rajah; Bhatnagar, Deepak

Submitted to: Biosystems Engineering
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/28/2020
Publication Date: 11/20/2020
Publication URL: https://handle.nal.usda.gov/10113/7182562
Citation: Tao, F., Yao, H., Hruska, Z., Kincaid, R., Rajasekaran, K., Bhatnagar, D. 2020. A novel hyperspectral-based approach for identification of maize kernels infected with diverse Aspergillus flavus fungi. Biosystems Engineering. 200:415-430. https://doi.org/10.1016/j.biosystemseng.2020.10.017.
DOI: https://doi.org/10.1016/j.biosystemseng.2020.10.017

Interpretive Summary: Aflatoxins are produced by the aflatoxigenic fungi, A. flavus and A. parasiticus, and crop products contaminated with the dangerous toxins pose a serious health threat to both humans and livestock. It is essential to remove the contaminated products from food and feed supply. Several methods including microbiological or microscopic assays are laborious and time-consuming. Opitcal-based methods using visible (Vis), near infrared (NIR) and hyperspectral imaging, offer rapid, and real-time evaluation leading to elimination of contaminated products. In this work, NIR hyperspectral imaging was applied to discriminate contaminated corn kernels with toxin-producing strains of A. flavus from healthy or contaminated with non-toxin producing strains of A. flavus. Using this method, the contaminated kernels were detected with an accuracy of 95% and it is considered to be more robust and simplified for practical implementation.

Technical Abstract: Near infrared (NIR) hyperspectral imaging over the spectral range of 900-2500 nm was investigated for its potential to identify corn kernels infected with aflatoxigenic fungus. A total of 900 kernels were used with 3 treatments, namely, 300 kernels inoculated with the AF13 (aflatoxigenic) fungus, 300 kernels inoculated with the AF36 (non-aflatoxigenic) fungus, and 300 kernels inoculated with sterile distilled water as control. One hundred kernels from each treatment of 300 kernels were incubated for 3, 5 and 8 days, separately to obtain diverse kernels with different infection times. All kernels were imaged over endosperm and germ sides, separately and the mean absorbance spectra were extracted for individual kernels from each side. Absorbance differences between the endosperm and germ sides were observed over the 1690-1824 and 2254-2398 nm regions. The partial least-squares discriminant analysis (PLS-DA) method was employed to establish discriminant models for 3-class (control, non-aflatoxigenic and aflatoxigenic) and 2-class (aflatoxigenic positive and negative). The models were evaluated 100 times, with calibration and prediction data randomly generated from the full data set. Based on the full 164 spectral points from the same kernel side(s), the overall best mean prediction accuracies achieved 96.31% for the 3-class classification and 97.81% for the 2-class classification. The 3-class and 2-class models using spectra from different kernel sides had the overall best mean prediction accuracies of 91.54% and 95.14%. For process optimization, the random frog (RF) algorithm was employed to select the most informative wavelengths. When using the first 30 variables determined by RF, the simplified RF-PLSDA models yielded a mean overall prediction accuracy of 87.73%. While studying the first 100 variables determined from each of the 100 random runs, it was found that the same 67 variables were selected in every run. These 67 variables were further applied to establish upgraded RF-PLSDA models. These upgraded models achieved an overall mean prediction accuracy of 94.94%. Although the obtained accuracy of 94.94% was slightly lower than the accuracy of 95.16% from the first 100 variables, this model was considered to be more robust and simplified for practical implementation. Furthermore, visualization of the predicted kernel class was proposed and demonstrated in this study, for both 3-class and 2-class classifications.