Non-Destructive Discrimination of Sunflower Seeds with Different Internal Mildew Grades by Fusion of Near-Infrared Diffuse Reflectance and Transmittance Spectra Combined with 1D-CNN

Abstract

Internally mildewed sunflower seeds, which cannot be recognized and discarded based on their appearance, pose a serious risk to human health. Thus, there is a need for a rapid non-destructive mildew grade discrimination method. Currently, few reports are available regarding this process. In this study, a method based on the combination of the near-infrared diffuse reflectance and near-infrared diffuse transmission (NIRr-NIRt) fusion spectra and a one-dimension convolutional neural network (1D-CNN) is proposed. The NIRr-NIRt fusion spectra can provide more complementary and comprehensive information, and therefore better discrimination accuracy, than a single spectrum. The first derivative (FD) preprocessing method could further improve the discrimination effect. By comparison against three conventional machine learning algorithms (artificial neural network (ANN), support vector machine (SVM), and K-nearest neighbor (KNN)), the 1D-CNN model based on the fusion spectra was found to perform the best. The mean prediction accuracy was 2.01%, 5.97%, and 10.55% higher than that of the ANN, SVM, and KNN models, respectively. These results indicate that the CNN model was able to precisely classify the mildew grades with a prediction accuracy of 97.60% and 94.04% for the training and test set, respectively. Thus, this study provides a non-destructive and rapid method for classifying the mildew grade of sunflower seeds with the potential to be applied in the quality control of sunflower seeds.