Geographical Origin Identification of Lycium Barbarum Using Near-Infrared Hyperspectral Imaging
WANG Lei1, 2, QIN Hong1,2*, LI Jing3, ZHANG Xiao-bo3, YU Li-na1, 2, LI Wei-jun1, 2, HUANG Lu-qi4*
1. Institute of Semiconductors, Chinese Academy of Sciences, Beijing 100083, China
2. Center of Materials Science and Optoelectronics Engineering & School of Microelectronics, University of Chinese Academy of Sciences, Beijing 100049, China
3. State Key Laboratory Breeding Base of Dao-di Herbs, National Resource Center for Chinese Materia Medica, China Academy of Chinese Medical Sciences, Beijing 100700, China
4. State Key Laboratory Breeding Base of Dao-di Herbs, China Academy of Chinese Medical Sciences, Beijing 100700, China
摘要： 宁夏产地的宁夏枸杞属于《中华人民共和国药典》收录的道地药材，药用价值较高、消费者青睐度更高，然而优质宁夏枸杞的种植面积较小、产量较低、枸杞子市场以乱充好、以其他产地冲抵道地产区产品的现象频发。因此，建立快速有效的宁夏枸杞产地鉴别模型对监督市场具有重要的意义。日常的市场交易枸杞子的鉴定一般凭借经验，但是该方法误差较大，可信度较低。传统的理化实验鉴别周期较长，非专业人员无法操作。近些年一些学者研究发现不同产地的宁夏枸杞成分含量具有差异性，然而枸杞子样本较小、形状不规则、成分分布不均匀，近红外光谱鉴别通常需要碾碎成粉末然后采集光谱信息，无法做到无损批量地采集枸杞子数据来鉴别枸杞子产地。近红外高光谱图像结合了近红外光谱和图像，包含丰富的空间信息和光谱信息，可以实现无损批量地采集非均匀样本光谱信息。利用近红外高光谱图像对甘肃、青海、新疆、宁夏和内蒙5个产地的宁夏枸杞进行产地鉴别。使用近红外高光谱图像系统采集了1 650个样本数据之后，通过阈值分割、图像去噪等方法提取出感兴趣区域(region-of-interest, ROI)。对比了ZCA白化(zero-phase component analysis whitening)预处理方法和常用的标准化(normalization)预处理方法，实验结果表明ZCA白化预处理是一种有效的高光谱数据预处理方法，可以去除特征之间的相关性，提升产地鉴别模型的准确率。对预处理后的数据采用偏最小二乘降维(partial least squares based dimension reduction, PLSDR)降低模型复杂度，结果表明经过ZCA白化预处理后的数据可以由288维特征降低到4个主成分，使得去除相关性后的特征可以被更少的隐藏特征来表示，这样可以极大程度上降低模型复杂性。最后，将降维后的特征输入到不同的分类器中进行训练，包括支持向量机(support vector machine, SVM)、Fisher线性判别分析(linear discriminant analysis, LDA)和Softmax分类。其中，基于ZCA白化、PLSDR和Softmax分类的模型表现最好，在测试集上的平均准确率为94.06%，可以有效的鉴别宁夏枸杞产地。
关键词：近红外；高光谱；宁夏枸杞；产地鉴别；ZCA白化；Softmax
Abstract：Lycium barbarum produced in Ningxia belongs to the genuine regional drugs contained in the Pharmacopoeia of the People's Republic of China. Due to the small planting area, low yield, high medicinal value and high consumer preference, the market is filled with chaos, and the phenomenon of passing others origins off as Ningxia happens occasionally. Therefore, it is of considerable significance to establish a rapid and effective geographical origin identification model of Lycium barbarum to supervise the market. In the process of market transactions, discriminating origin of Lycium barbarum is often based on experience, which has much error and low credibility. The traditional physical and chemical experiment has a long identification cycle and can't be operated by non-professionals. In recent years, some scholars have found that the content of Lycium barbarum in different producing areas is different. However, because of the small sample size, irregular shape and uneven distribution of components, the near-infrared spectroscopy identification technique often needed to smash Lycium barbarum to collect spectral information. Near-infrared hyperspectral image technology combined with near-infrared spectroscopy and image technology, which contains rich spatial information and spectral information, can achieve non-destructive acquisition of spectral information. In this research, near-infrared hyperspectral image technology was used to discriminate the geographical origin of Lycium barbarum samples, which were gathered from Gansu, Qinghai, Xinjiang, Ningxia and Inner Mongolia in China. After collecting the hyperspectral information of 1 650 samples by hyperspectral image system, the region of interest (ROI) was effectively extracted by threshold image segmentation and denoising. During the pretreatment process, the comparison between zero-phase component analysis (ZCA) whitening results and normalization results indicated that ZCA whitening was an effective spectral preprocessing method to remove correlation between features and improve the accuracy of the model. Partial least squares based dimension reduction (PLSDR) was used to reduce the complexity of the model for the preprocessed data. The experimental results indicated that the data after ZCA whitening pretreatment could be reduced from 288-dimensional features to 4 principal components, which made the correlation-removed features can be represented by fewer hidden features. Finally, the dimensionality-reduced features were fed to different classifiers to train model, including support vector machine (SVM), linear discriminant analysis (LDA) and softmax regression. Among those models, the average recognition accuracy based on ZCA whitening, PLSDR and softmax regression was 94.06% on the test set. The results demonstrated that the proposed method could effectively discriminate the origin of Lycium barbarum samples.
[1] Gong G, Dang T, Deng Y, et al. International Journal of Biological Macromolecules, 2018, 109: 611.
[2] Forino M, Tartaglione L, Carmela Dell’Aversano, et al. Food Chemistry, 2016, 194：1254.
[3] Yang R F, Zhao C, Chen X, et al. Journal of Functional Foods, 2015, 17：903.
[4] Lu W, Jiang Q, Shi H, et al. Journal of Agricultural and Food Chemistry, 2014, 62(37)：9073.
[5] Li Q, Yu X, Xu L, et al. Food Chemistry, 2017, 221：1113.
[6] Yin W, Zhang C, Zhu H, et al. PLoS One, 2017, 12(7)：e0180534.
[7] WANG Huan, JIN Zhe-xiong, WEN Mei-jia, et al(王 欢, 金哲雄, 温美佳，等). Research and Practice on Chinese Medicines(现代中药研究与实践), 2014, 28(6): 21.
[8] QU Yun-qing, ZHANG Tong-gang, LIU Dun-hua(曲云卿, 张同刚, 刘敦华). Food & Machinery(食品与机械), 2015, 31(2): 76.
[9] Tang J L, Wang D, Zhang Z G, et al. Computers and Electronics in Agriculture, 2017, 135: 63.
[10] Chen H, Tan C, Lin Z, et al. Spectrochimica Acta Part A, 2018, 189: 183.
[11] Dang T, Sethu V, Ambikairajah E. Compensation Techniques for Speaker Variability in Continuous Emotion Prediction. IEEE Transactions on Affective Computing, 2018.
[12] Lee J, Chang K, Jun C H, et al. Chemometrics and Intelligent Laboratory Systems, 2015, 147: 139.
[13] Naseer N, Noori F M, Qureshi N K, et al. Fronties in Human Neuroscience, 2016, 10: 237.
[14] Dai X, Duan Y, Hu J, et al. Infrared Physics & Technology, 2018, 97: 25.
