金银花和山银花均为忍冬科植物, 二者外形颇为相似, 但化学组成和含量、 功效以及价格存在着显著差异。 一些不法商家为获取超额利润, 以价廉的山银花冒充金银花进行售卖, 消费者仅凭肉眼较难准确鉴别。 现阶段尚缺少金银花与山银花的无损鉴别研究。 近红外(NIR)光谱技术可以实现复杂样品的快速无损分析, 通过结合偏最小二乘判别分析(PLS-DA)等模式识别方法可以实现不同来源样品的鉴别分析。 然而, 光谱变量过多易导致PLS-DA方法出现过拟合的问题。 本研究利用光栅型便携式NIR光谱仪采集了3个产地的643份金银花与本地200份山银花的光谱。 此外, 一个月后收集3个产地金银花与本地山银花样品各50份作为独立验证集。 提出了一种新型模式识别方法——随机检验(RT)-PLS-DA, 并与主成分分析(PCA)、 PLS-DA以及现有的变量筛选-PLS-DA方法如竞争性自适应重加权采样法(CARS)-PLS-DA和蒙特卡罗-无信息变量消除法(MC-UVE)-PLS-DA进行了比较, 利用光谱预处理进一步提高模型的准确性。 结果表明: NIR原始光谱中存在严重的谱峰重叠、 基线漂移及背景干扰; 即使结合优化预处理方法, PCA模型仍无法实现金银花与山银花的准确鉴别; 一阶导数(1st)或连续小波变换(CWT)预处理结合PLS-DA模型可以获得较为准确的鉴别结果, 验证集和独立验证集的鉴别率分别为100%和98%; 3种变量筛选方法-PLS-DA方法中, CARS方法选择变量数最少, RT方法在选择特征变量的同时还可获得满意的鉴别率。 1st-RT-PLS-DA模型最佳, 验证集和独立验证集的鉴别率分别为100%和99.50%。 以上结果表明, 便携式NIR光谱仪结合变量筛选-PLS-DA策略可实现金银花与山银花的准确鉴别, 为中药材掺伪快速鉴定提供了新的思路。

Both Lonicerae Japonicae Flos and Flos Lonicerae are plants of the Caprifoliaceae family. They are rather similar in appearance. However, there are differences in chemical composition, content, efficacy, and price. To obtain excessive profits, unscrupulous merchants sell the cheaper Flos Lonicerae as Lonicerae Japonicae Flos. It is difficult for consumers to distinguish them with the naked eye. Currently, there is no study on the non-destructive identification of Lonicerae Japonicae Flos and Flos Lonicerae. Rapid and non-destructive analysis of complex samples can be achieved using near-infrared (NIR) spectroscopy. The identification of samples from different sources can be achieved by combining pattern recognition methods, such as partial least squares discriminant analysis (PLS-DA). However, an excessive number of spectral variables may easily lead to the problem of overfitting in the PLS-DA method. In this study, 643 spectra of Lonicerae Japonicae Flos from three production areas and 200 spectra of Flos Lonicerae from the local area were collected using a grating portable NIR spectrometer. Besides, 50 samples of Lonicerae Japonicae Flos from each production area and local Flos Lonicerae were collected one month later as the external validation set. A new pattern recognition method, named randomization test (RT)-PLS-DA, was proposed. This method was compared with principal component analysis (PCA), PLS-DA, and existing variable selection-PLS-DA methods, such as competitive adaptive reweighted sampling (CARS)-PLS-DA and Monte Carlo-uninformative variable elimination (MC-UVE)-PLS-DA. The accuracies of the models were further improved with the spectral pretreatments. The results showed that there were severe interferences, including peak overlapping, baseline drift, and background, in the original spectra. Even with optimized pretreatment methods, the accurate identification of Lonicerae Japonicae Flos and Flos Lonicerae cannot be achieved using the PCA method. Accurate identification results could be obtained using PLS-DA with either first derivative (1st) or continuous wavelet transform (CWT) pretreatment, while the identification rates for the validation and external validation sets were 100% and 98%, respectively. Among the three variable selection-PLS-DA methods, the CARS method selected the fewest variables. The selection of feature variables and achieving satisfactory identification rates can be done simultaneously with the RT method. The 1st-RT-PLS-DA model was the best, and the identification rates for the validation and external validation sets were 100% and 99.50%, respectively. The above results indicate that the accurate identification of Lonicerae Japonicae Flos and Flos Lonicerae can be achieved using a portable NIR spectrometer and a variable selection-PLS-DA method, providing a new approach for the rapid detection of adulteration in traditional Chinese medicinal materials.