基于漫反射光谱的初制绿茶含水率无损检测方法

Nondestructive measurement of moisture content of green tea in primary processing based on diffuse reflectance spectroscopy

茶叶含水率是影响茶叶加工品质的一项重要指标。为了实现茶叶加工中含水率的快速检测,该文提出了一种应用漫反射光谱技术的绿茶初制过程中含水率无损检测方法。采用波长范围在325～1075nm的可见-短波近红外光谱仪,对炒青绿茶在8个加工工序中随机抽取的568个茶叶样本进行漫反射光谱扫描,光谱扫描后立即测量样本的含水率。对于得到的光谱数据,采用小波变换降低其信息维度并提取小波系数,比较小波低频系数对于光谱特征信息的提取能力,结果显示,小波低频系数能够有效提取原始光谱数据中的特征信息。采用3种回归算法:偏最小二乘回归、神经网络和支持向量机分别建立含水率的测量模型。比较发现支持向量机回归模型的结果最优,建模相关系数为0.9985,预测相关系数为0.9875。研究结果表明,漫反射光谱可以用于绿茶含水率的无损、快速检测,小波变换是一种有效的光谱特征提取算法,而且支持向量机回归算法具有高精度和强泛化能力,可广泛用于回归分析。

Moisture content of tea is an important index affecting processing quality of tea.To realize fast measurement of moisture content of tea in processing,this paper put forward a nondestructive way to measure moisture content of green tea in primary processing based on diffuse reflectance spectroscopy technique.A visible-near infrared（Vis/NIR） spectroradiometer was adopted for scanning diffuse reflectance spectra of 568 samples in the range of 325-1 075 nm wavelengths.These samples were from eight procedures in primary processing,moisture contents of samples were immediately measured after spectral scanning.For obtaining high-dimension spectral data,wavelet transform（WT）was used to reduce of dimensionality and extraction of wavelet coefficients.The capability of low-frequency wavelet coefficients was evaluated for extracting spectral characteristic,and the result indicated that it was effective to mine characteristic information from spectra by wavelet transform.Three regression algorithms including partial least square（PLS）,artificial neural network and least square support vector machine（LS-SVM）were used to develop models for determination of moisture content respectively.It could be found that LS-SVM model obtained the optimal result with rc=0.9985 and rv=0.9875.These results indicated that it is feasible to measure moisture content of green tea nondestructively and fast based on diffuse reflectance spectroscopy,WT is an effective method for extraction of characteristic from spectra,and LS-SVM algorithm can be broadly used for regression analysis with high precision and strong generalization.