Prediction of milled maize fumonisin contamination by multispectral image analysis

Abstract Mycotoxin contamination is a major concern to the maize industry worldwide. Despite the several strategies that have been exploited in an attempt to reduce the severity of this problem, during conducive years, severely contaminated lots are still introduced in the maize processing chain affecting the general quality and safety of the product. As chemical analysis is laborious, time consuming and equipment dependent, more convenient methods are needed for the early identification of contaminated lots. Here a novel approach based on image analysis that provides fast response with minimal equipment and effort is presented. Maize samples were grounded and imaged under 10 different LED lights with emission centered at wavelengths ranging from 720 to 940 nm. The digital images were converted into matrices of data to compute comparative indexes. A three layers feed-forward neural network was trained to predict mycotoxin content from the calculated indexes. The results showed a significant correlation between predictions from image analysis and the concentration of the mycotoxin fumonisin as determined by chemical analysis. The technique developed produces reliable contamination estimates within few minutes and can be readily used to assist lot selection in various steps of the maize processing chain.