基于高光谱技术的不同地理尺度贡菊产地鉴别研究  

Geographical origin authentication of Gongju at different spatial scales based on hyperspectral technology

作　　者：郭雪 白瑞斌 王慧[1] 李卫文[2] 董玲[2] 孙嘉慧 张小波[1] 杨健[1] GUO Xue;BAI Rui-bin;WANG Hui;LI Wei-wen;DONG Ling;SUN Jia-hui;ZHANG Xiao-bo;YANG Jian(National Key Laboratory for Quality Ensurance and Sustainable Use of Dao-di Herbs,National Resource Center for Chinese Materia Medica,China Academy of Chinese Medical Sciences,Beijing 100700,China;Key Laboratory of Horticultural Crop Germplasm innovation and Utilization(Co-construction by Ministry and Province),Institute of Horticulture,Anhui Academy of Agricultural Sciences,Hefei 230001,China)

机构地区：[1]中国中医科学院中药资源中心道地药材品质保障与资源持续利用全国重点实验室,北京100700 [2]安徽省农业科学院园艺研究所农业农村部园艺作物种质创制与利用重点实验室(部省共建),安徽合肥230001

出　　处：《中国中药杂志》2024年第22期6073-6081,共9页China Journal of Chinese Materia Medica

基　　金：中国中医科学院科技创新工程项目(CI2023E002);中药全产业链质量技术服务平台项目(2022-230-221);国家中医药管理局高水平中医药重点学科建设项目(ZYYZDXK-2023244);财政部和农业农村部国家现代农业产业技术体系项目(CARS-21)。

摘　　要：贡菊为《中国药典》收录的五大药用菊花品种之一,近年来其栽培产地发生了明显变化,导致药材品质良莠不齐。该研究以安徽、云南、重庆等地产出的贡菊为研究对象,通过花冠朝上(A)、花基部朝上(B)等不同模式分别在可见-近红外(VNIR)和短波红外(SWIR)波段采集高光谱数据。采用多元散射校正(MSC)、Savitzky-Golay平滑(SG)、一阶导数(D1)、二阶导数(D2)和标准正态变换(SNV)等5种方法对高光谱数据进行预处理后,结合偏最小二乘法判别分析(PLS-DA)、随机森林(RF)和支持向量机(SVM)3种算法,分别从省域和安徽省内市县域2个地理尺度建立贡菊产地鉴别模型,以预测结果的准确率作为评价指标筛选最佳模型,同时使用混淆矩阵评估模型分类性能。结果表明,2种地理尺度的贡菊产地鉴别模型均以花基部朝上(B)采集模式结合D2预处理和RF分类方法为最优,全波段(VNIR+SWIR)建模效果略优于单一波段,省域和市县域预测集准确率分别达到99.69%、99.40%。进一步采用竞争性自适应重加权算法(CARS)、连续投影算法(SPA)和空间迭代收缩法(VISSA)筛选特征波长建模,采用CARS筛选的特征波长数量较少,且2种地理尺度模型优化后预测集准确率亦可达到99.56%、98.65%,基本实现与全波段模型相似的识别效果,但去除冗余变量大大降低了模型的复杂性。该研究建立的高光谱技术结合化学计量学方法可实现不同地理尺度贡菊的产地鉴别,为贡菊产地快速检测体系的构建及专属小型化仪器装备系统的开发提供了理论依据和技术参考。Gongju(Chrysanthemum morifolium)is one of the five major medicinal Chrysanthemum varieties included in the Chinese Pharmacopoeia.In recent years,its cultivation areas have changed significantly,resulting in mixed quality of the medicinal herbs.In this study,Gongju cultivated in Anhui,Yunnan,Chongqing,and other places were selected as research objects.Hyperspectral data were collected in the visible-near-infrared(VNIR)and short-wave infrared(SWIR)bands using different modes,such as corolla facing up(A)and flower base facing up(B).After pre-processing the hyperspectral data using five methods,including multiplicative scatter correction(MSC),Savitzky-Golay smoothing(SG),first derivative(D1),second derivative(D2),and standard normal variate(SNV),partial least squares discriminant analysis(PLSDA),random forest(RF),and support vector machine(SVM)were used to establish origin identification models of Gongju at the two geographical scales of the province and the city-county in Anhui province.The accuracy of the prediction results was used as an evaluation index to select the optimal models,and the classification performance of the models was evaluated by confusion matrix.The results showed that the flower base facing up(B)collection model combined with second derivative pretreatment and RF method was the best model for both geographical scale identification models.The modeling effect of the full-band(VNIR+SWIR)was slightly better than that of the single band,with the accuracy of the prediction set in the province and city-county regions reaching 99.69%and 99.40%,respectively.The competitive adaptive reweighted sampling algorithm(CARS),successive projections algorithm(SPA),and variable iterative space shrinkage approach(VISSA)were further used to screen the feature wavelength modeling.The number of feature wavelengths screened by CARS was fewer,and the prediction set accuracy of the two geographical scales models after optimization could reach 99.56%and 98.65%,which was basically comparable to the full-band model.However,

关 键 词：高光谱技术 贡菊 产地鉴别 化学计量学 地理尺度 

分 类 号：R282.5[医药卫生—中药学]