Quantitative analysis of aflatoxin B1 in moldy peanuts based on near-infrared spectra with two-dimensional convolutional neural network

• Published in: Infrared physics & technology 2023-06, Vol.131, p.104672, Article 104672
• Main Authors: Jiang, Hui, Deng, Jihong, Zhu, Chengyun
• Format: Article
• Language: English
• Subjects: Aflatoxins B1; Convolutional neural network; Gramian angular summation field image coding; Near-infrared spectroscopy; Peanut
• ISSN: 1350-4495; 1879-0275
• DOI: 10.1016/j.infrared.2023.104672

•This study proposes a new determination method of AFB1 content in peanuts.•Gramian angular summation field was used to encode the NIR spectral features.•Two-dimensional convolutional neural network was designed to determine the AFB1 in peanuts.•The model-based deep learning has a promising application prospect in spectral analysis. A two-dimensional convolutional neural network (2D-CNN) model based on Gramian angular summation field (GASF) image coding was proposed to detect aflatoxin B1 (AFB1) in moldy peanuts with high precision. Fourier transform near-infrared spectrometer was employed to acquire near-infrared spectra of moldy peanuts. The GASF was used to encode the NIR spectral features after principal component analysis. A convolutional neural network (CNN) framework was designed and the GASF-2D-CNN model was developed to realize the quantitative detection of AFB1 in peanuts. The results obtained showed that compared with the traditional chemometrics models and one-dimensional CNN (1D-CNN) model, the performance of GASF-2D-CNN model was significantly better than the previous two types of detection models. The root mean square error of prediction (RMSEP), coefficient of determination (RP2), relative percent deviation (RPD), and ratio of performance to interquartile range (RPIQ) of the best GASF-2D-CNN model were 2.0 μg·kg−1, 0.99, 8.3, and 9.3, respectively. The study reveals that GASF image coding technique can effectively mine the potential feature information in NIR spectra, and this technique will have a promising application prospect in the multivariable calibration of spectroscopy.