Constructing an origin discrimination model of japonica rice in Heilongjiang Province based on confocal microscopy Raman spectroscopy technology

Abstract

An origin discrimination model for rice from five production regions in Heilongjiang Province was constructed based on the combination of confocal microscopy Raman spectroscopy and chemometrics. A total of 150 field rice samples were collected from the Fangzheng, Chahayang, Jiansanjiang, Xiangshui, and Wuchang production areas. The optimal sample processing conditions, instrument parameter settings, and spectrum acquisition techniques were identified by investigating the influencing factor. The Raman spectra of milled rice within the range of 100-3200 cm^-1 were selected as the raw data, and the optimal preprocessing method combination consisting of normalization, Savitzky-Golay smoothing, and multivariate scatter correction was identified. Subsequently, the competitive adaptive reweighted sampling and discrete binary particle swarm optimization algorithms were employed to optimize the feature wavelength selection, resulting in the screening of 226 and 1899 feature wavelength variables, respectively. Using the full Raman spectrum data and feature wavelength data as inputs, partial least squares discriminant analysis, support vector machine and extreme learning machine origin discrimination models were constructed. The results indicated that the BPSO-GA-SVM model exhibited the best predictive ability, achieving a testing set accuracy of 86.67%.

Keywords: Confocal microscopy Raman spectroscopy; Feature spectral selection; Japonica rice; Origin discrimination model.

Fig. 1

Raw Raman spectrum of different sample processing statuses.

Fig. 2

Raw Raman spectra of different excitation wavelengths.

Fig. 3

Raw Raman spectra of different laser intensity.

Fig. 4

Comparison diagram before and after removing cosmic ray.

Fig. 5

Raman spectra of different collection methods.

Fig. 6

Raman spectroscopic characteristic peaks of rice.

Fig. 7

The original spectral and spectra preprocessed sequentially by Nor, SG, and MSC.

Fig. 8

The running process of CARS filtering variables.

Fig. 9

The running process of BPSO filtering variables.

Fig. 10

Confusion matrix of the BPSO-GA-SVM model in the prediction process.

Fig. 11

Distribution map of japonica sampling sites in Heilongjiang Province. Note: The distribution map was generated by ArcGIS Pro software (https://www.esri.com/en-us/arcgis/products/arcgis-pro.) and was based on the latitude and longitude data measured at the sampling sites.