Non-destructive hyperspectral imaging in both olive fruit and powder for aflatoxin detection and estimation by machine learning

Aflatoxins are toxic and carcinogenic mycotoxins posing a significant threat to human health and food safety. A non-destructive hyperspectral imaging (HSI) system to automatically detect aflatoxin contamination in olive fruit and powder by machine learning was proposed. Imaging was conducted in the 418 to 1072 nm wavelength range. For whole fruit analysis, Linear Discriminant Analysis (LDA) achieved 100% accuracy in binary classifying healthy and contaminated samples. The Support Vector Machine method also reached 98.75% accuracy for the same purpose. For powdered olive samples, PLSR based on full spectral data, yielded coefficient of determination (R 2 ) values of 0.9986 and 0.9858, for calibration and validation disjoint data sets, respectively. Furthermore, combining Decision Tree with a Learning Automata algorithm extracted the 15 optimal most discriminant (effective) wavelength (EW) values, enabling data dimension reduction without a significant loss of discrimination power. Using 15 effective wavelengths, LDA model had a maximum accuracy of 100%. PLSR model developed using the selected effective wavelengths also had robust performance, with R 2 of 0.89, validation set. Findings confirm the high potential of hyperspectral imaging for non-destructive and accurate detection of fungal toxin contamination in plant food, suggesting its potential as a rapid and reliable method in food industry. • Automatic HSI estimation of aflatoxin contaminants in olive fruit by machine learning • Optimal prototype with discriminant wavelengths for both detection and estimation • High coefficient of determination R 2 =0.985 in olive powder contamination regression • HSI shown as alternative to costly destructive conventional contaminant detection • Contaminant estimation has significant implications for public health and food industry