Detection of aflatoxin contamination in single kernel almonds using multispectral imaging system

Abstract

The developed multispectral imaging system was tested in the current study for its detection accuracy of aflatoxin B1 (AFB1) in single kernel almonds. Based on our previous study, characteristic wavelengths for aflatoxin B1 (AFB1) were selected from the hyperspectral imaging data and a prototype of almond sorter was designed and developed using multispectral imaging system. Five levels of AFB1 concentrations, viz. 0.25, 0.50, 0.75, 1.00 μg/kernel were used for testing of the sorter. Artificial neural network (ANN) was used to predict the AFB1 contamination levels and four discriminant models, viz. support vector machine (SVM), logistic regression (LR), linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) were analyzed to classify the contaminated kernels from the normal ones. Results indicated that ANN was very efficient in prediction of AFB1 threshold contamination levels of 0.250 µg/g with accuracy of 99.300%, 98.100%, 95.700%, 86.200% and least error rate of 0.042, 0.077, 0.109 µg/g, 0.187 µg/g respectively for training, cross-validation, prediction and testing using external test set data. The two-class LR model accurately discriminated the contaminated from the noncontaminated almonds with a threshold of 0.25 µg/g and the correct classification rate of 90.800% and 95% for cross-validation and external test datasets, respectively. The multispectral imaging system demonstrated its potential application to rapid online detection of AFB1 contamination levels in single almond kernels and discrimination of contaminated kernels at the same time.