摘要
食药植物玛咖富含多种营养成分,极具药用价值。采用近红外漫反射光谱,对采自秘鲁及云南共139份玛咖样品进行产地鉴别。采用多元信号校正结合二阶导数和Norris平滑预处理光谱,利用光谱标准偏差初步选择光谱波段(7 500~4 061cm^(-1)),结合主成分-马氏距离(principal component analysis-mahalanobis distance,PCA-MD)筛选出适宜的主成分数为5。基于所筛选的光谱波段及主成分数,利用"模群迭代奇异样本诊断"方法剔除2个异常样品后,分别采用竞争自适应重加权法(competitive adaptive reweighted sampling,CARS)、蒙特卡洛-无信息变量消除法(monte carlo-uninformative variable elimination,MC-UVE)、遗传算法(genetic algorithm,GA)和子窗口重排(subwindow permutation analysis,SPA)四种方法筛选光谱变量信息,利用模型集群分析(model population analysis,MPA)思想对所筛选的光谱变量信息进行评价。结果显示,RMSECV(SPA)>RMSECV(CARS)>RMSECV(MC-UVE)>RMSECV(GA),分别为2.14,2.05,2.02,1.98,光谱变量数分别为250,240,250和70。采用偏最小二乘判别分析法(partial least squares discriminant analysis,PLS-DA)对四种方法筛选的光谱变量建立判别模型,随机选择97份样品作为建模集,其余40份样品作为验证集。通过R2,RMSEC和RMSEP分析可知,R2:GA>MC-UVE>CARS>SPA,RMSEC和RMSEP:GA<MC-UVE<CARS<SPA,且GA,MC-UVE,CARS和SPA四种方法筛选的光谱信息所建立的产地判别模型预测正确率分别为95.0%,92.5%,90.0%和85.0%。四种方法筛选的光谱信息所建立的产地判别模型均具有较好的预测性能,其中GA法所筛选的光谱信息建立的判别模型更准确。该方法的建立旨在为中药材鉴别和品质评价奠定基础。
Medicinal and edible plant Maca is rich in various nutrients and owns great medicinal value.Based on near infrared diffuse reflectance spectra,139 Maca samples collected from Peru and Yunnan were used to identify their geographical origins.Multiplication signal correction（MSC）coupled with second derivative（SD）and Norris derivative filter（ND）was employed in spectral pretreatment.Spectrum range（7 500~4 061cm^（-1））was chosen by spectrum standard deviation.Combined with principal component analysis-mahalanobis distance（PCA-MD）,the appropriate number of principal components was selected as 5.Based on the spectrum range and the number of principal components selected,two abnormal samples were eliminated by modular group iterative singular sample diagnosis method.Then,four methods were used to filter spectral variable information,competitive adaptive reweighted sampling（CARS）,monte carlo-uninformative variable elimination（MC-UVE）,genetic algorithm（GA）and subwindow permutation analysis（SPA）.The spectral variable information filtered was evaluated by model population analysis（MPA）.The results showed that RMSECV（SPA）〉RMSECV（CARS）〉RMSECV（MC-UVE）〉RMSECV（GA）,were 2.14,2.05,2.02,and 1.98,and the spectral variables were 250,240,250 and 70,respectively.According to the spectral variable filtered,partial least squares discriminant analysis（PLS-DA）was used to build the model,with random selection of 97 samples as training set,and the other 40 samples as validation set.The results showed that,R2：GA〉MC-UVE〉CARS〉SPA,RMSEC and RMSEP：GA〈MC-UVE〈CARS〈SPA.For the spectral information selected by the four methods,GA,MC-UVE,CARS and SPA,the model prediction accuracy were 95.0%,92.5%,90.0% and 85.0%,respectively.Compared with the four methods,we could know that the origin discriminant models built based on spectra information filtered by the four methods possess good estimated performance.Among them,the model built based on the spectra information filtered
出处
《光谱学与光谱分析》
SCIE
EI
CAS
CSCD
北大核心
2016年第2期394-400,共7页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金项目(31460538
81260608)
云南省自然科学基金项目(2013FD066
2013FZ150)资助
关键词
玛咖
近红外光谱
鉴别
光谱信息筛选
模型集群分析
Lepidium meyenii Walp.
NIR spectroscopy
Identification
Spectral information screening
Model population anal ysis
