Identification of the Origin of Ziziphi Spinosae Semen Based on Near-infrared Spectroscopy and One-dimensional Convolutional Neural Network

In order to facilitate the rapid identification of the origin of Ziziphi Spinosae Semen and to conduct a quantitative analysis of adulteration, raw spectra of samples originating form five regions—Hebei, Henan, Shandong, Shanxi, and Shaanxi were collected using near-infrared spectroscopy (NIRS). Additionally, the moisture, fat, and protein contents of these samples were measured. Five preprocessing methods were employed to eliminate noise from the data. Qualitative and quantitative models were developed utilizing partial least squares (PLS) regression and a one-dimensional convolutional neural network (1DCNN) with a custom selection layer. Furthermore, competitive adaptive reweighted sampling (CARS) and the successive projection algorithm (SPA) were applied to identify the most relevant wavelengths for model optimization. The results showed that there was no significant difference between samples from Hebei and Shandong. However, notable differences were observed when compared with other origins. The 1DCNN preprocessed with Savitzky-Golay derivative 1st (SGD1) performed best in the classification model with 91.11% accuracy in the test set. For Xing Ziziphi Spinosae Semen samples exhibiting varying levels of adulteration, the 1DCNN model after SG derivative preprocessing outperforms the partial least squares regression (PLSR) model by a coefficient of determination greater than 0.86 and a residual prediction deviation (RPD) greater than 2.50. In conclusion, this study indicates that combining near-infrared spectroscopy with a one-dimensional convolutional neural network can realize high-precision classification and quantitative assessment of adulteration in Ziziphi Spinosae Semen sourced from different geographical locations.