Multi-classification identification of PLS in rice spectra with different pre-treatments and K/S optimisation

Abstract

Rice from neighbouring areas and of similar varieties can have a negative impact on classification and identification systems. Hence, it is important to validate methods of identifying similar varieties of rice from neighbouring areas to ensure the traceability of food quality. In this study, Raman spectroscopy and chemometric methods were used to identify and classify 14 similar varieties of rice from neighbouring areas in Northeast China. The Raman spectra of 630 single-grain rice samples were collected, and the original spectra were pre-processed using 9 methods, including Savitzky–Golay (SG) smoothing, normalisation (NL), and polynomial curve fitting (PCF). The original full spectrum (FS), characteristic band spectrum, and the characteristic wavelengths were subsequently extracted via the successive projection algorithm (SPA) in the range of 400–1700 cm−1 as the partial least squares (PLS) modelling input data. A contrast analysis of the impact of changes in the number of rice species on the accuracy of different modelling, training, and validation sets was performed. Thereafter, the models were optimised using the Kennard–Stone (K/S) sample set partition method. Results showed that different pre-processing methods have varied effects on the modelling. In the FS and SPA feature extraction methods, the fourth polynomial fitting of the PCF had the best recognition effect after pre-processing, and the SG-NL-PCF modelling effect in the band spectrum was optimal. At the same time, with an increase in rice varieties, the recognition rate of the PLS classification model showed a downward trend. At 14 species, FS-PCF-PLS and characteristic bands SG-NL-PCF-PLS and SG-NL-PCF-SPA-PLS were predicted. The correct rate of set recognition was 81.90 %, 92.86 %, and 86.67 %, respectively. After the sample set was divided with K/S, the correct rate of recognition of the three prediction sets was 96.67 %, 97.14 % and 90.95 %, respectively. The Raman spectroscopy technology, combined with the PLS method, can be used to quickly and non-destructively identify multiple types of rice from neighbouring areas. The pre-processing and K/S sample set partition methods can significantly improve the model prediction accuracy.