高光谱成像技术鉴别红景天的品种  

Identification of Rhodiola varieties by Hyperspectral imaging technology

作　　者：李涛[1] 钟玉琴 曲明亮 LI Tao;ZHONG Yuqin;QU Mingliang(West China School of Pharmacy,Sichuan University,Chengdu,Sichuan,610041 P.R.China)

机构地区：[1]四川大学华西药学院,四川成都610041

出　　处：《华西药学杂志》2021年第5期526-530,共5页West China Journal of Pharmaceutical Sciences

基　　金：四川省科学技术厅应用基础研究计划项目(编号:2020YJ0275);四川省中医药管理局全国第四次中药资源普查科技专项项目(编号:2019PC003)。

摘　　要：目的采用高光谱成像技术结合化学计量学方法建立红景天的分类判别模型。方法首先获取935~1720 nm大花红景天、长鞭红景天和狭叶红景天的高光谱图像信息,然后分别提取红景天感兴趣区域的反射光谱值,在采用标准正态变量变换法(SNV)和多元散射校正法(MSC)对原始的高光谱数据进行预处理后,分别建立偏最小二乘判别分析模型(PLS-DA),通过对比两种预处理方法对PLS-DA模型判别结果的影响,得到SNV为较优的预处理方法。采用载荷系数法(x-LW)和竞争自适应重加权算法(CARS)提取经SNV预处理后数据的特征波长,并分别建立基于特征波长的线性判别分析(LDA)、PLS-DA和概率神经网络(PNN)分类模型。结果与x-LW特征提取方法比较,CARS算法提取的特征波长建立的LDA、PLS-DA和PNN模型有更好的分类性能;PNN的分类性能优于PLS-DA;SNV-CARS-LDA为区分不同品种红景天的最优判别模型,对训练集和测试集的识别率均为100%。结论高光谱成像技术与化学计量学相结合的方法具有很大的在线检测潜力,可以快速、无损地鉴别红景天品种。OBJECTIVE To investigate a classification method of Rhodiola variety using hyperspectral imaging technology combined with chemometrics.METHODS Firstly, the hyperspectral image information of R.crenulata,R.fastigiata,and R.kirilowii was taken in the visible/near infrared(VIS-NIR) range(935-1720 nm).Then, the reflection spectral value of the interested area of Rhodiola was extracted respectively.After the two pretreatment methods of standard normal variable transformation(SNV) and multiple scattering correction(MSC) were used to preprocess the original hyperspectral data, the partial least squares discriminant analysis(PLS-DA) model was established.Comparing the effects of the two pretreatment methods on the discriminant results of PLS-DA,it was concluded that SNV was the better pretreatment method for discriminating model.The characteristic wavelengths of the preprocessed SNV data was extracted using the x-loading weights(x-LWs) and the competitive adaptive reweight sampling method(CARS).The classification model of linear discriminant analysis(LDA),PLS-DA and probabilistic neural network(PNN) based on the characteristic wavelengths was established respectively.RESULTS The results showed that the LDA,PLS-DA and PNN models based on the characteristic wavelengths extracted by the CARS algorithm had better classification performance than the x-LW method.The classification performance of nonlinear model PNN was better than that of linear model PLS-DA,and the SNV-CARS-PLS-DA was the best discriminating model for Rhodiola with the accuracy of 100% for both training sets and testing sets.CONCLUSION The integration of hyperspectral imaging and chemometrics has a great potential for on-line detection and was feasible to rapidly and non-invasively discriminate Rhodiola variety.

关 键 词：大花红景天 长鞭红景天 狭叶红景天 高光谱成像技术 标准正态变换法 多元散射校正法 载荷系数法 竞争自适应重加权算法 线性判别分析 偏最小二乘判别分析 概率神经网络 

分 类 号：R917[医药卫生—药物分析学]