高光谱成像技术结合特征波长优化对苍术颗粒剂生产厂家的可视化判别研究

Research on the Visualization Differentiation of Atractylodes Lancea Granule Manufactures Based on Hyperspectral Imaging Technology Combined With the Selection of Characteristic Wavelengths

为了给苍术颗粒剂基于高光谱成像的可视化区分提供理论指导,选用竞争性自适应重加权采样法(CARS)和相关性分析(CA)进行两次特征波长选择,提出了利用近红外高光谱成像技术对苍术颗粒剂产品溯源的新方法。874~1 734 nm波段范围内采集150个来自三个生产厂家的苍术颗粒剂高光谱图像,提取感兴趣区域(ROI)的光谱反射率值作为鉴别模型的输入变量,采用邻近算法(KNN)、误差反向传输神经网络(BPNN)、偏最小二乘法判别分析(PLS-DA)、最小二乘支持向量机(LS-SVM)建立四种算法(分类器)的判别模型。通过对模型效果的评价标准(预测集总体判别率以及kappa系数)来判别三个不同厂家苍术颗粒剂的区分效果。除KNN模型外,预测集的判别率都是100%, kappa系数均为1。为了加快运算速度,研究通过CARS、随机蛙跳算法(RF)、连续投影算法(SPA)和序列前向选择(SFS)算法初步选择特征波长;采用CARS, RF, SFS和SPA结合CA算法取得了4组最优波长。分别得到4个(975, 1 220, 1 419, 1 476 nm)、 2个(1 005, 1 442 nm)、 4个(924, 1 005, 1 419, 1 584 nm)和3个(948, 1 146, 1 412 nm)最优波长,并分别建立了KNN, BPNN, PLS-DA和LS-SVM判别模型。在筛选三种最优算法的情况下,能够以较少的特征波长个数获得的最好建模效果为:CARS-CA-LS-SVM模型中预测集总体判别率是100%, kappa系数为1。将CARS-CA筛选出波长变量的每个像素点光谱数据输入到LS-SVM模型中,将判别结果用不同颜色直观显示。该研究为快速无损进行苍术颗粒剂产品溯源提供了思路,为今后开发相关机构的快速监管提供了技术支持。

In order to provide theoretical guidance for the visualization differentiation of Atractylodes Lancea granules based on hyperspectral imaging,competitive adaptive reweighted sampling(CARS)and correlation analysis(CA)was used to select two characteristic wavelengths.A new method for traceability of Atractylodes Lancea granules using near-infrared hyperspectral imaging technology was proposed.Hyperspectral image of 150 Atractylodes Lancea granules from three manufacturers in the range of 874~1734 nm,extracting the spectral reflectance value of the region of interest(ROI)as the input variables for the identification model,and using the proximity algorithm(k-nearest neighbor,KNN),back-propagation neural networks(BPNN),partial least squares-discrimination analysis(PLS-DA)and least square support vector machine(LS-SVM)to establish discriminant models of four algorithms(classifiers).The discrimination effect of three different manufacturers of Atractylodes Lancea granules was discriminated by the evaluation criteria of the model effect(predictive set overall discriminant rate and kappa coefficient).Except for the KNN model,the discriminant rate of the prediction set was 100%,and the kappa coefficient was 1.In order to speed up the operation,this study selected the characteristic wavelengths by CARS,random frog(RF),successive projections algorithm(SPA)and sequential forward selection(SFS)algorithm,and used CARS,RF,SFS,and SPA combined with the CA algorithm to achieve four sets of optimal wavelengths.Four(975,1220,1419,1476 nm),two(1005,1442 nm),four(924,1005,1419,1584 nm)and three(948,1146,1412 nm)optimal wavelengths were obtained respectively,and KNN,BPNN,PLS-DA,and LS-SVM discriminant models were established.Therefore,in the case of screening three optimal algorithms,the best modeling effect that can be obtained with fewer feature wavelengths was:the overall discriminant rate of the prediction set in the CARS-CA-LS-SVM model was 100%,the kappa coefficient was 1.Finally,the spectral data of each pixel of the wavelength variables selected by CARS-CA were input into the LS-SVM model,and the discrimination results were visually displayed in different colors.This study provides a method for the rapid and lossless traceability of Atractylodes Lancea granules product,and provides technical support for the rapid supervision of related organizations in the future.