小波分析的茶鲜叶全氮含量高光谱监测

茶是世界上最受欢迎的饮料之一，而氮素（N）是影响茶叶品质的主要成分之一，因此快速准确地估算N素含量至关重要。由于测定N含量的化学方法繁琐耗时，利用高光谱对茶鲜叶中N含量进行预测，利用连续小波转换（CWT）提取的小波系数，探究CWT不同分解层数对于N素含量的估测能力，并讨论了不同波长选择算法所建模型的预测效果。首先，采集广东省英德市茶园的151个茶鲜叶样品高光谱数据，将获得的原始光谱通过卷积平滑（SG）、去趋势（Detrending）、一阶导数（1st）、多元散射校正（MSC）和标准正态变量变换（SNV）五种预处理方法进行预处理并作为参考。其次，采用连续小波对原始光谱进行初步处理生成多尺度小波系数，并进行相关性分析, 分别利用连续投影算法（SPA）、竞争性自适应加权采样法（CARS）和变量组合集群分析（VCPA）方法进一步优化CWT变换后光谱数据的变量空间，最后，以特征变量为输入使用PLSR建立了N素定量监测模型，并对比不同尺度不同方法估算N素的效果。结果表明，连续小波分析方法可有效提升茶鲜叶光谱对N素含量的估测能力，明显优于常规光谱处理方法。经连续小波分解后，对茶鲜叶N素的预测能力随分解尺度的增加整体呈逐步降低的趋势，其中在1~6尺度连续小波变换后的光谱与茶鲜叶N素存在良好的相关性，表明小尺度的连续小波分解可有效应用于茶鲜叶N含量的监测。基于CWT（1）-VCPA方法建立的模型精度最高，且变量数相比于全波段减少了99.34%，其建模与预测R2达到0.95和0.90，相比于传统光谱处理方法，精度提升了11% ，证明CWT-VCPA可以有效降低光谱维度并大幅提升模型精度。实现了茶叶N素含量的高效量化预测，为评估茶叶的其他成分提供了可靠技术参考。

Estimation of Total Nitrogen Content in Fresh Tea Leaves Based on Wavelet Analysis

Tea is one of the most popular beverages globally, which is greatly affected by the content of nitrogen (N) in quality. Due to the complicated and time-consuming method for determining N content in fresh tea leaves by traditional chemical analysis, this paper proposes a means for N content prediction by hyperspectral technique. The wavelet coefficients extracted by continuous wavelet transform (CWT) technology are used to estimate N content by different decomposition layers of CWT. Moreover, the predictive effects of models built by different wavelength selection algorithms are also discussed. Several 151 hyperspectral data of tea samples were collected from tea gardens in the Yingde City of Guangdong Province. The original spectra data are preprocessed by smoothing (SG), detrending (Detrending), first derivative (1st), multiple scattering correction (MSC), and standard normal variable transformation (SNV) while comparing with CWT. Then, continuous wavelet multi-scale analysis is applied to process the original spectrum for generating wavelet coefficients, and Pearson correlation analysis was also performed. Next, three kinds of methods, including successive projections algorithm (SPA), competitive adaptive weighted sampling (CARS) and variable combination population analysis (VCPA), are adopted to optimize the variable space of the spectral data after CWT transformation. At last, quantitative models of N content prediction are established and compared by PLSR with characteristic variables selected by the three above mentioned methods as input. The overall results show that the continuous wavelet analysis algorithm can improve the model’s efficiency for estimating the N content of the fresh tea leaves by hyperspectral data. Furthermore, it has better performance than other conventional spectral preprocessing methods significantly. With continuous wavelet decomposition, the precision of the model for N content prediction gradually decreases with the increase of the decomposition scale.There is a good correlation between the spectrum after the continuous wavelet transforms on the scale of 1~6 and the N in fresh tea leaves,which shows that the small-scale continuous wavelet algorithm can be well applied to monitor N content in fresh tea leaves. The model established by CWT (1scale)-VCPA method has the best performance, andthe number of variables is reduced by 99.34% compared to the full band. The R2 of the calibration model and prediction model respectively, are 0.95 and 0.90. Compared with the traditional spectral processing method, the accuracy is improved by 11%. It is proved that the combination of CWT-VCPA can obviously reduce the spectral dimension and improve the accuracy of the model. This research achieves an efficient way for N content prediction of tea, which provides a technical basis and reliable reference for other components evaluation of tea.