基于卷积神经网络和近红外光谱的太平猴魁茶产地鉴别分析

太平猴魁茶因其特有的“喉韵”深受广大消费者喜爱，不同产地太平猴魁茶市场价格相差较大，如何实现产地精准鉴别是目前促进绿茶产业发展的关键因素。依赖于人工经验的感官评审方法主观性强、稳定性差，无法应用于实际生产检测过程。作为目前主要的检测分析方法，化学分析方法周期长、检测成本高，而且目前没有用于茶叶产地鉴别的统一标准。近红外光谱（NIR）作为一种无损检测分析方法，具有快速、非破坏性、无污染等特点，但是不同产地太平猴魁茶主要内含成分及其含量基本相同，不同产地样本光谱特征峰分布相似，导致常规分析方法无法有效选择特征变量。卷积神经网络（CNN）作为经典深度学习网络模型之一，具有强特征提取和模型表达能力。采用太平猴魁茶产地光谱特征分析，利用一维卷积神经网络模型（1-D CNN）提取太平猴魁茶NIR特征，提出一种基于1-D CNN和NIR的太平猴魁茶产地鉴别分析方法。试验以6个不同产地共120个样本为研究对象，分析10 000~4 000 cm-1范围内的光谱信息；将样本随机划分为训练集（84，占70%）和测试集（36，占30%），分别讨论不同间隔采样、网络结构、卷积核大小及激活函数对产地鉴别结果的影响，并引入Dropout方法对比分析模型过拟合现象；最终建立一个具有9层结构的1-D CNN模型。蒙特卡罗试验结果表明，相比于基于原始光谱数据（40.57%，7.06）和PCA方法（31.93%，6.96）的太平猴魁茶产地预测模型准确率和标准差，基于1-D CNN的太平猴魁茶产地鉴别模型预测精度和稳定性更高，其测试集预测准确率平均值和标准差分别为97.73%和3.47。因此，1-D CNN可有效提取太平猴魁茶不同产地NIR特征，提高太平猴魁茶产地鉴别精度，为太平猴魁茶精准产地鉴别及溯源分析提供参考。

Geographical Origin Discrimination of Taiping Houkui Tea Using Convolutional Neural Network and Near-Infrared Spectroscopy

Taiping Houkui Tea, one of China’s precious tea series, occupies an important position in public consumption and the tea market. The price of Taiping Houkui tea varies greatly from different geographical origins, and accurate geographical origin discrimination is currently an important factor in promoting the green tea industry. Sensory evaluation methods that rely on the manual experience are highly subjective and poor instability and cannot be applied to the actual analysis process. As the main detection and analysis method at present, the chemical analysis method is time-consuming and laborious. More importantly, there is currently no uniform standard for the geographical origin discrimination of tea. Near-infrared spectroscopy (NIR), as non-destructive testing and analysis method, has the characteristics of fast, non-destructive, and non-polluting. However, the types and contents of main components of Taiping Houkui tea from different origins are similar, which results in the same spectral peak distributions of various samples, and conventional analysis methods are limited for selecting feature variables. As one of the typical deep learning network models, convolutional neural network (CNN) has strong feature extraction and model expression capabilities. Based on the analysis of the spectral characteristics of Taiping Houkui tea from different geographical origins, the 1-dimension CNN (1-D CNN) is used to extract the NIR features, and a discriminant method combing NIR with 1-D CNN is explored to identify the geographical origin of Taiping Houkui tea in this work. In this paper, 120 samples were collected from 6 different geographical origins. The NIR were sampled from 10 000~4 000 cm-1 and preprocessed by standard normal variate (SNV). The sample is randomly divided into a training set (84, 70%) and test set (36, 30%), and the effects of CNN structure, convolution kernel size, activation function and other parameters on the analysis results were discussed separately. As a result, a 1-D CNN model with 9-layer was constructed for the geographical origin discrimination of Taiping Houkui tea. The principal component analysis (PCA) was compared, and the Monte-Carlo method was used to evaluate the stability and robustness of the proposed method. Compared with the prediction accuracy and standard deviation of the models based on original spectral data (40.57%, 7.06) and the PCA method (31.93%, 6.96), the prediction accuracy and stability of the 1-D CNN-based geographical origin discrimination model are higher, and the average prediction accuracy and standard deviation of the testing set are 97.73% and 3.47, respectively. The comparison results demonstrate the proposed 1-D CNN model can effectively extract NIR features and has the ability to identify the geographical origins of Taiping Houkui tea, which provides an effective method for the identification and traceability analysis of the geographical origin and production of valuable tea products such as Taiping Houkui tea.