目的 为实现止嗽青果丸生产过程的质量控制，将遥感领域的像元解混方法引入中药制剂的均匀度评价，以可视化分析止嗽青果丸粉末混合均匀度和蜂蜜混合均匀度。方法 采用SG平滑（Savitzky-Golay，SG）、一阶导数（first derivative，1D）、二阶导数（second derivative，2D）、标准正态变量变换（standard normal variate，SNV）、多元散射校正（multiplicative scatter correction，MSC）等多种预处理方法优化光谱数据，消除背景噪声等干扰。通过利用纯物质光谱，结合混合调谐匹配滤波解混（mixture tuned matched filtering，MTMF）算法，综合目标端元的丰度得分与混合可行性得分，对混合样本像素点进行物质丰度估计与二值化，从而实现药材粉末的空间分布可视化，进一步构建了残差网络（residual network，ResNet）模型对蜂蜜含量进行预测与像素级可视化。结果 ResNet模型与偏最小二乘回归（partial least squares regression，PLSR）、多元线性回归（multiple linear regression，MLR）、岭回归（ridge regression，RR）、最小绝对收缩和选择算子回归（least absolute shrinkage and selection operator，Lasso）、随机森林（random forest，RF）、ElasticNet等传统定量模型相比，预测集R²从0.837 6显著提升至0.931 1。结论 像元解混结合深度学习模型直观展示了药材原料与辅料在制剂中的空间分布，为中药制剂的精细化生产与质量控制提供了新思路。

Objective To achieve quality control during the production of Zhisou Qingguo Wan (ZQW, 止嗽青果丸), the pixel unmixing method from remote sensing was introduced into the uniformity evaluation of traditional Chinese medicine preparations, so as to visually analyze the mixing uniformity of medicinal powder and honey in ZQW. Methods Multiple preprocessing methods, including Savitzky-Golay (SG) smoothing, first derivative (1D), second derivative (2D), standard normal variate (SNV) and multiplicative scatter correction (MSC) were adopted to optimize spectral data and eliminate interferences such as background noise and spectral scattering. Based on pure substance spectra combined with the mixture tuned matched filtering (MTMF) algorithm, the abundance scores and mixture feasibility scores of target endmembers were integrated. Material abundance estimation and binarization at the pixel level of mixed samples were then performed, enabling the visualization of the spatial distribution of medicinal powders. Meanwhile, a residual network (ResNet) model was constructed for the prediction and pixel-level visualization of honey content. Results Compared with traditional quantitative models such as partial least squares regression (PLSR), multiple linear regression (MLR), ridge regression (RR), least absolute shrinkage and selection operator (Lasso), random forest (RF) and ElasticNet, the R² of the prediction set was significantly improved from 0.837 6 to 0.931 1 with the proposed model. Conclusion The method combining pixel unmixing and deep learning intuitively presents the spatial distribution of medicinal raw materials and excipients in the preparation. A novel insight for the refined production and quality control of traditional Chinese medicine preparations is provided.

R283.6

国家自然科学基金项目（82474048）