Extraction and classification of origin characteristic peaks from rice Raman spectra by principal component analysis

Abstract

Rice origin identification can provide brand protection for rice with geographical indications. Therefore, it is necessary to design a convenient and fast detection method to meet the demands of the market. Based on Raman spectroscopy, this study involved the extraction and classification of characteristic spectral peaks of rice grain samples of the same cultivar but different places of origin. Raman spectra were acquired from 80 samples of rice with four different places of origin (Longjing 31 cultivar), and spectral information was extracted for analysis. First, the rice spectra were pretreated using a baseline correction and range normalization. Second, from the 400−1600cm−1 and 2800−3200cm-1 spectral region starting from the first four principal components of the regression coefficients extracted from principal component analysis (PCA), further screening produced eight spectral peaks characteristic of the place of origin: 476 cm−1, 867 cm−1,940 cm−1, 1121 cm−1, 1342 cm−1, 1384 cm−1, 1462 cm−1, and 2914 cm−1. These were also assigned to functional groups, revealing subtle differences in the nutritional content dependent on place of origin. Third, eight characteristic values extracted by PCA were used to establish a four-layer 8-9-6-4 (input-hidden-hidden-output) back propagation (BP) neural network structure as a rice-origin identification model. Finally, the model was used to train 80 rice samples in place-of-origin classification, and the average prediction accuracy of the cyclic test for the training samples reached 97.5 % after five epochs; for the other four epochs the accuracy ranged from 98.75%–96.25%. These results show that the model is feasible as a tool for the identification of rice types of the same variety and that it can effectively identify rice from different areas.