Quantitative aflatoxin B1 detection and mining key wavelengths based on deep learning and hyperspectral imaging in subpixel level

Abstract

Aflatoxin is a serious threat to agricultural safety and human health. This study used a one-dimensional convolutional network to detect peanut aflatoxin B1 content and mine the aflatoxin key wavelengths based on spectral information. First, predicting AFB1 content performance was compared between single-branch and dual-branch convolutional networks. Pearson correlation coefficient of the dual-branch network was 0.77 for toxin content prediction, which was better than the single-branch network model. Then, the key wavelengths of deep learning mining were evaluated using support vector machines under different training folds. The experimental results showed that the key wavelength was optimal with 8 training folds. Finally, deep learning and traditional methods were performed to mine the key wavelengths. The experimental results indicated that the key wavelengths mined by deep learning were better than the traditional method, and the accuracy of deep learning was 91.30% in the test set. The most important wavelength of deep learning was 445 nm. The study indicated that the model accurately predicted toxin content and selected key wavelengths. The model provides an effective method for aflatoxin rapid detection.