Identification of Geographical Origin of Semen Ziziphi Spinosae Based on Hyperspectral Imaging Combined with Convolutional Neural Networks

The accurate identification of geographical origin is one of the main challenges in quality assessment of semen ziziphi spinosae. This paper proposed resnet50-based and shuffle-based convolutional neural network (CNN) models (CNN-resnet and CNN-shuffle for hyperspectral image input, as well as CNN-resnet1D and CNN-shuffle1D for spectra input) to discriminate semen ziziphi spinosae from five geographical origins. Specifically, minimum noise fraction combined with mask was adopted to remove background pixels. Savitzky-Golay smoothing, followed by standard normal variate, was applied to denoise in pixel-wise. Mahalanobis distances were calculated based on the average spectra to identify and exclude outliers. Partial least squares discriminant analysis and support vector machine were also adopted to compare with the four CNN networks, particular in model accuracy, computing time, and parameters. Overall, CNN-shuffle was the best, with accuracy (0.902) comparable to CNN-resnet and in less time (0.348 s) and with fewer parameters (5607). Feature wavelengths were selected based on the learned weight coefficients in self-defined layer of CNN-resnet1D and CNN-shuffle1D models. They were related to lipid and protein, of which contents were statistically analyzed. Both lipids and proteins were key parameters, and lipids played a greater role in the identification. Spinosin as a quality marker was measured via standard HPLC method, which was comparably discussed with the hyperspectral imaging (HSI). The potential application of HSI with CNN in quality control of semen ziziphi spinosae within the modern system of mass production was higher compared with HPLC.