为建立一种快速无损的黄芪药材产地和产出模式识别方法，利用高光谱成像系统(光谱范围:400～1000、900～1700 nm，检测时间15 s)，对80份不同产地、不同产出模式的黄芪药材进行检验。采集的高光谱数据集高度相关且数量较大，需要建立稳定可靠的数据降维和分类模型，首先使用归一化、高斯平滑和构建掩模等方法对原始光谱进行预处理，再分别采用主成分分析(PCA)、偏最小二乘法判别分析(PLS-DA)和竞争自适应重加权采样(CARS)构建高光谱降维数据集，对降维后的光谱图像和光谱曲线分别使用支持向量机(SVM)、多层神经网络模型(FFNN)和卷积神经网络(CNN)进行训练。结果显示，在对黄芪高光谱图像数据集进行PLS-DA之前，应用CARS作为变量选择方法，该降维算法在CNN测试集中的准确率、精确度和召回率均达到100％，F1值(F1-score)和ROC曲线下的面积值(area under the curve of ROC，AUC)均达到1。该方法方便快捷，样品用量少且无损样品，为黄芪药材产地和产出模式的快速识别提供技术支持。
Fast identification of origins and cultivation patterns of Astragali Radix by dimension reduction algorithms of hyperspectral data
This study aims to establish a rapid and non-destructive method for recognizing the origins and cultivation patterns of Astragali Radix.A hyperspectral imaging system (spectral ranges:400-1000 nm,900-1700 nm;detection time:15 s) was used to examine the samples of Astragali Radix with different origins and cultivation patterns.The collected hyperspectral datasets were highly correlated and numerous,which required the establishment of stable and reliable dimension reduction and classification models.Firstly,the original spectra were preprocessed by normalization,Gaussian smoothing,and masking.Then,principal component analysis (PCA),partial least squares-discriminant analysis (PLS-DA),and competitive adaptive reweighted sampling (CARS) were performed to reduce the dimension of the hyperspectral data.Finally,support vector machine (SVM),feedforward neural network (FFNN),and convolutional neural network (CNN) were used for data training of the spectral images and spectral curves with dimension reduction.The results showed that applying CARS as a variable selection method before PLS-DA on the hyperspectral data of Astragali Radix achieved the accuracy,precision,and recall of 100％ on the CNN test dataset.The F1-score and area under the curve of ROC (AUC) reached 1.This method is convenient,quick,sample-saving,and non-destructive,providing technical support for rapid identification of the origins and cultivation patterns of Astragali Radix.
Astragali Radixhyperspectrumpartial least squares-discriminant analysiscompetitive adaptive reweighted samplingconvolutional neural network
NETL
NSTL
万方数据
