| 摘要: |
| 儿茶素和咖啡碱是茶叶品质的重要评价指标，为了探索深度卷积神经网络(CNN)结合可见近红外光谱(Vis/NIR)用于茶叶儿茶素和咖啡碱无损快速检测的可行性，本研究通过高效液相色谱来测定茶叶中的儿茶素和咖啡碱含量，并与样本的光谱信息建立对应关系；采用回归分析和CNN建模构建了光谱与茶叶内含物的定量关系；采用竞争自适应重加权采样(CARS)和连续投影算法(SPA)选择特征波长，用于开发基于这些特征波长的简单模型。结果表明:4种儿茶素和咖啡碱含量从第1叶位到第6叶位呈现出逐渐降低的趋势；提取特征波长不仅减少了光谱变量数，还获得了比全谱更优或接近的模型性能；CNN在回归分析和特征提取中均表现出良好的性能，预测儿茶素和咖啡碱最优模型的决定系数(R2)和残余预测偏差(RPD)分别达到了0.93和3.28以上。因此，卷积神经网络结合可见近红外光谱可以对儿茶素和咖啡碱的含量进行快速无损检测。 |
| 关键词: 卷积神经网络 可见近红外光谱 新鲜茶叶 儿茶素 咖啡碱 |
| DOI：10.11841/j.issn.1007-4333.2021.11.11 |
| 投稿时间：2021-01-29 |
| 基金项目:国家自然科学基金项目(31771676)；国家重点研发计划项目(2018YFD0700501)；浙江省科技计划项目(2017C02027-01)；高校基本科研业务费专项资金项目(2015QNA6005) |
|
| Spectral rapid detection of phytochemicals in tea (Camellia sinensis)based on convolutional neural network |
|
LI Xiaoli,ZHANG Dongyi,DONG Yulun,JIN Juanjuan,HE Yong
|
| (College of Biosystem Engineering and Food Science/Key Laboratory of Spectroscopy Sensing of Ministry ofAgriculture and Rural Affairs, Zhejiang University, Hangzhou 310058, China) |
| Abstract: |
| Catechins and caffeine are the key indicators of tea quality. In order to explore the feasibility of visible/near infrared spectroscopy(Vis/NIR)combined with convolutional neural network(CNN)for nondestructive and rapid detection of catechins and caffeine of fresh tea, this study used high performance liquid chromatography to determine the content of catechins and caffeine in tea, and established a corresponding relationship with the spectral information of the sample. Regression analysis and CNN modeling were used to construct the quantitative relationship between the spectrum and contents of tea. In addition, the competitive adaptive reweighted sampling(CARS)and successive projections algorithm(SPA)were used to select the characteristic wavelengths for the development of simple models. The result showed that: The contents of four catechins and caffeine displayed a gradual decrease trend from the first leaf position to the sixth leaf position. The extracting characteristic wavelengths not only reduced the number of spectral variables, but also obtained models performance that were better or close to that of the full spectrum. CNN performed well in the regression analysis and feature extraction, the coefficient of determination(R2)and residual prediction deviation(RPD)of optimal models for predicting catechins and caffeine reached 0. 93 and 3. 28, respectively. Therefore, the convolutional neural network combined with visible/near infrared spectroscopy could quickly and non-destructively detect contents of catechins and caffeine. |
| Key words: convolutional neural network visible/near infrared spectroscopy fresh tea catechin caffeine |
