LAUSR

The aim of this study is to determine the value of near-infrared spectroscopy (NIRS) as a diagnostic tool for aflatoxin contamination, specifically to rapidly predict levels of aflatoxin, either quantitatively or qualitatively, in ground maize. Maize was collected from inoculated field trials conducted across four sites in Kenya. Inoculated and uninoculated maize ears were harvested, milled, and prepared for NIRS scanning and wet chemistry-based aflatoxin quantification. Several statistical and machine learning techniques were compared. Absorbance at a single bandwidth explained 34 % of the variation in levels of aflatoxin using a regression model while a partial least-squares (PLS) method showed that NIR measurements could explain 42 % of the variation in aflatoxin levels. To compare various methods for their ability to classify samples with high (>100 ppb) levels of aflatoxin, various additional procedures were applied. The k-nearest neighbour classification method yielded sensitivity and specificity values of 0.75 and 0.52 respectively, compared with the support vector machine method with estimates of 0.81 and 0.68, whereas PLS could achieve values of 0.82 and 0.72 respectively. The corresponding false positive and false negative values are still unacceptable for NIRS to be used with confidence, as ~18 % of contaminated ground maize samples would be accepted and 28 % of good maize would be discarded or declared contaminated or downgraded. However, such calibrations could be useful in breeding programs without access to wet chemistry analysis, seeking to rank entries semiquantitatively.