Classification of aflatoxin B1 naturally contaminated peanut using visible and near-infrared hyperspectral imaging by integrating spectral and texture features

The aim of this study is to carry out a non-destructive, hyperspectral imaging-based method to discriminate between normal and naturally aflatoxin B-1 (AFB(1)) contaminated peanuts. Two varieties of peanut were imaged under a hyperspectral imaging system in the spectral range from 400 to 1000 nm. The reference AFB(1) levels were measured by enzyme linked immunosorbent assay (ELISA) method, then the peanuts were divided into two categories according to the threshold of 20 ppb. The spectral, color and texture features were extracted and integrated to discriminate AFB(1) contaminated from normal peanuts. Different pretreatment methods were conducted on the full spectra, the linear discriminant analysis (LDA) results indicated that firstly Savitzky-Golay smoothing (SGS) then standard normal transformation (SNV) could achieve the best discrimination with an accuracy of 90% and 92% for calibration and validation sets respectively. The LDA results of feature integration proved that the optimized spectral features combined with the texture features realized the best classification, with an accuracy of 91% and 94% for calibration and validation sets respectively. Finally, the performance of partial least squares discrimination analysis (PLS-DA) and support vector machine (SVM) was compared with the LDA, the SVM with RBF kernel revealed the best results with an accuracy of 93% and 94% for calibration and validation sets respectively. This study presented the potential of hyperspectral imaging in direct AFB(1) contamination classification of peanut, and demonstrated that the combination of texture and spectra features could improve the modelling results.