Skip to main content
ARS Home » Southeast Area » New Orleans, Louisiana » Southern Regional Research Center » Food and Feed Safety Research » Research » Publications at this Location » Publication #317290
Title: Visible near-infrared (VNIR) reflectance hyperspectral imagery for identifying aflatoxin-contaminated corn kernels

Author
item ZHU, FENGLE - Mississippi State University
item YAO, HAIBO - Mississippi State University
item HRUSKA, ZUZANA - Mississippi State University
item KINCAID, RUSSELL - Mississippi State University
item Brown, Robert
item Bhatnagar, Deepak
item Cleveland, Thomas

Submitted to: ASABE Annual International Meeting
Publication Type: Proceedings
Publication Acceptance Date: 5/14/2015
Publication Date: 8/15/2015
Citation: Zhu, F., Yao, H., Hruska, Z., Kincaid, R., Brown, R.L., Bhatnagar, D., Cleveland, T.E. 2015. Visible near-infrared (VNIR) reflectance hyperspectral imagery for identifying aflatoxin-contaminated corn kernels. Proceedings of 2015 ASABE Annual International Meeting. Paper No.:152189995. https://doi.org/10.13031/aim.20152189995.
DOI: https://doi.org/10.13031/aim.20152189995

Interpretive Summary:

Technical Abstract: Aflatoxin contamination can be found in various agricultural commodities. Because of their toxic and carcinogenic properties, both pre-harvest and post-harvest strategies have been developed to minimize the exposure to aflatoxins in food and feed. In this study, visible near-infrared (VNIR) reflectance hyperspectral imaging was employed as a post-harvest strategy to identify aflatoxin-contaminated corn kernels rapidly and non-invasively. Corn ears in the field were inoculated with toxigenic Aspergillus flavus spores at early dough stage to artificially create aflatoxin contamination. Kernels were removed from the inoculated corn ears and scanned using the hyperspectral imagery on the germ side. Subsequently, chemical analyses were performed to determine aflatoxin concentration for each kernel. For all kernel samples, the mean spectra were extracted and compressed using principal component analysis. Both the threshold values of 20 ppb and 100 ppb were applied separately to designate each sample as contaminated or healthy. The algorithm of decision trees was then employed to classify contaminated and healthy samples based on the extracted PCs. The best overall accuracy was 90.00% from the test sets for both 20 ppb based and 100 ppb based classification. Since the gradient of the slope of the reflectance curve ranging from 700 to 800 nm increased as the aflatoxin contamination level of samples increased, the band ratio image of 800 nm to 700 nm was calculated on one hyperspectral image to identify contaminated kernels. The overall identification rate was 80%. Though this accuracy was lower than that using mean spectra analysis, the two band ratio analysis was simpler with less computation time and a lower cost.