基于高光谱成像的茶叶中EGCG分布可视化

针对目前关于表没食子儿茶素没食子酸酯(epigallocatechin gallate,EGCG)在茶叶中的分布缺乏可视化表达问题,该文采用高光谱成像技术以实现EGCG在茶叶中的分布可视化。通过高光谱成像仪采集茶叶的光谱信息,按照标准方法 HPLC(high performance liquid chromatography)法测量茶叶的EGCG浓度。运用化学计量学方法建立光谱与EGCG浓度之间的回归模型。为寻求相对较优的模型效果,对光谱进行不同的预处理,以确定最优的预处理方法;采用4种建模方法建立回归模型,以确定最优的建模方法;对光谱进行特征波段选择,以降低数据冗余提高模型的稳定性和运算速度。最后,将高光谱图像中像素点对应的光谱变量导入模型,从而生成EGCG浓度分布图。结果表明:可见-近红外光谱与EGCG浓度之间具有很强的相关性,其回归模型的决定系数达到0.905,利用高光谱成像技术对茶叶中EGCG分布进行可视化可行。通过对不同品种、叶位的茶叶中EGCG浓度分布进行可视化,能够为高EGCG浓度茶树品种的培育、EGCG代谢规律的分析以及茶树采摘部位的识别提供有效手段。

EGCG distribution visualization in tea leaves based on hyperspectral imaging technology

Abstract: EGCG (epigallocatechin gallate) is an important functional material in tea, and it is regarded as an indispensable index for evaluating the quality of tea as it''s of great benefit to health. With the difference of tea varieties and physiological parts of tea plant, the distribution of EGCG is different. Visualization of EGCG distribution contributes to analyze the distribution and metabolism of EGCG directly. However, no research on the visualization of EGCG distribution in tea leaves has been reported till now. This study took advantage of hyperspectral imaging technology and chemometrics method to realize visualization of EGCG distribution in fresh tea leaves. On the basis of visualization, distribution characteristics of EGCG between different tea varieties and different leaf positions were studied. The operation procedure of visualization was mainly divided into 5 steps: 1) Acquisition of physical and chemical information. To obtain the physical information, 486 fresh leaves from the 1st to the 6th leaf positions at the tender shoots of tea plants with 3 varieties were gathered first, hyperspectral images of these fresh leaves were collected by a hyperspectral imager, and then average spectral information used to build models was extracted from the hyperspectral images. To acquire the chemical information, the fresh leaves were freeze-dried, ground into powder, sieved and heated by water-bath to obtain the EGCG solution, and the EGCG concentration was determined through HPLC (high performance liquid chromatography) at last. 2) Samples division and spectral preprocessing. In order to divide the samples reasonably, an interval-extraction method was adopted to ensure the distribution uniformity of chemical values. All the samples were divided into calibration set and prediction set in a ratio of 2:1. Due to the limited performance of hyperspectral imager, obvious noise region of the spectra was eliminated first in order to avoid the interference to subsequent analysis. For 2 common issues during spectral acquisition, i.e. random noise and baseline drift, the SG (Savitzky-Golay) smoothing and baseline correction were performed. Through comparing different preprocessing methods, it was found that the unprocessed spectra showed the best performance. 3) Model establishment and analysis based on full efficient spectra. To determine the best modeling method, PCR (principal component regression), PLSR (partial least squares regression), RBFNN (radial basis function neural network) and LS-SVR (least squares support vector regression) models between full efficient spectra and EGCG concentration values were established respectively. The results showed that the nonlinear models had better performance, and by comparing the evaluation parameters of different models, LS-SVR was chosen as the best modeling method. 4) Model establishment and analysis based on feature bands. The full efficient spectra contain 478 variables, which carry rich information, and cause a collinear problem between variables at the same time. To reduce the data redundancy and the complexity of the model based on full efficient spectra, SPA (successive projection algorithm) was employed to select feature bands, and the LS-SVR model based on feature bands showed better performance compared with the LS-SVR model based on full efficient spectra, with the Rp2 (determination coefficient of prediction set) and RPD (residual prediction deviation) that is the ratio of standard deviation of measured values to root mean square error of prediction set reaching 0.905 and 3.248 respectively. 5) Generation of EGCG distribution map. Inputting the feature bands of each pixel selected by SPA in the testing hyperspectral images into the SPA-LS-SVR model, the EGCG concentration of each pixel could be calculated, so the distribution maps of EGCG in fresh tea leaves were generated finally. This study proved that EGCG distribution visualization in fresh tea leaves can be realized by hyperspectral imaging technology and chemometrics method. Through the analysis of EGCG distribution between different tea varieties and different leaf positions, the distribution showed significant differences. This study provides an effective method for cultivation of tea plant variety with high EGCG concentration, analysis on the metabolism rule of EGCG and recognition of tea shoots.