三七是一种传统的中药材, 具有较高的药用价值。 目前市场上中药售假的现象屡见不鲜, 许多不法商贩将三七支根或剪口粉末假冒主根粉末销售, 严重损害了消费者的利益。 利用高光谱技术结合多元分析方法实现三七不同部位粉末的快速无损鉴别。 通过高光谱成像系统分别采集了三七剪口、 须根和主根粉末在400~1 000 nm范围内的高光谱图像, 共300个样本。 采用Savitzky-Golay（SG）平滑结合标准变量变换（SNV）的方法对高光谱数据进行去噪和消除因散射引起的光谱差异。 为了移除光谱变量中的重迭和冗余信息, 利用竞争自适应重加权采样（CARS）算法和本文提出的一种考虑了变量间交互作用的二进制竞争自适应重加权采样（BCARS）算法进行特征波长选择。 最后分别建立了基于全光谱、 CARS和BCARS特征波长的支持向量机（SVM）和极端梯度提升（XGBoost）分类模型。 结果表明, BCARS-XGBoost模型的分类效果最优, 训练集和测试集的分类准确率分别为100%和99.33%。 与CARS相比, BCARS所选择的特征波长数量较少, 有助于多光谱系统和便携式仪器的开发。 利用高光谱技术结合BCARS-XGBoost模型鉴别三七不同部位粉末是可行的。

Panax notoginseng is a traditional Chinese medical herb with high medicinal value. Nowadays, adulteration is common in the Chinese medicine market, and many unscrupulous traders sell rootlet or rhizome powder as the main root powder, which seriously damages the interests of consumers. Therefore, this study aims to rapidly and non-destructively identify Panax notoginseng powder of different parts by applying a hyperspectral imaging techniques combined with multivariate analysis methods. The hyperspectral images of Panax notoginseng rhizome, fibrous root and main root powder were collected by the hyperspectral imaging system in the spectral range of 400~1 000 nm （a total of 300 samples）. Savitzky-Golay（SG）smoothing combined with Standard Normalized Variate （SNV） was applied to eliminate the noise in spectral data and reduce the spectral difference caused by scattering. In order to remove the overlapping and redundant information in spectral variables, a Binary Competitive Adaptive Reweighted Sampling （BCARS） algorithm that considers the interaction effect among variables proposed in this paper was used to select the feature wavelengths. At the same time, the Competitive Adaptive Reweighted Sampling （CARS） algorithm was also used. Based on the full spectrum, CARS and BCARS feature wavelengths, Support Vector Machine （SVM） and eXtreme Gradient Boosting （XGBoost） classification models were established, respectively. The results showed that the BCARS-XGBoost model had the best performance, with classification accuracies of 100% and 99.33% for the training and prediction sets, respectively. In addition, fewer feature wavelengths were selected by BCARS, which is conducive to developing a multi-spectral system and portable detector. Therefore, it is feasible to identify Panax notoginseng powder of different parts by applying a hyperspectral imaging technique combined with the BCARS-XGBoost model.