Rapid identification of Chrysanthemi Flos based on multi-source data fusion and machine learning algorithms

Chrysanthemi Flos (CF), a versatile herb utilized for both medicinal and culinary applications, exhibits a range of visually similar varieties. The majority of extant investigations identify herbal materials merely based on isolated morphological traits or individual chemical markers; systematic research elucidating the correlation between macroscopic phenotypic features and intrinsic bioactive constituents from the holistic Traditional Chinese Medicinal perspective of "differentiating morphological traits and evaluating medicinal quality" remains scarce to date. This study proposes a multi-source data fusion analysis strategy that integrates computer vision with chemical composition analysis. A total of 101 batches of CF samples encompassing seven cultivars, namely Boju (BJ), Chuju (CJ), Gongju (GJ), Hangbaiju (HBJ), Huaiju (HJ), Beijingju (BJJ) and Damaya (DMY), were collected in this study. Python was employed to extract Red-Green-Blue (RGB) features as well as Gray-Level Co-occurrence Matrix (GLCM)-based texture features from the front surface, back surface and powder of CF samples to characterize morphological differences in CF traits. Meanwhile, ultra-high performance liquid chromatography (UPLC) was adopted to quantitatively determine the contents of seven flavonoid and phenylpropanoid components in CF samples of different cultivars, so as to decipher inherent variations in their chemical constitutions. Correlation analysis revealed that RGB color features exhibited significantly positive correlations with chlorogenic acid and apigenin-7-O-glucoside. Meanwhile, GLCM texture features were positively correlated with 3,5-O-dicaffeoylquinic acid. Collectively, these results corroborate an inherent linkage between the macroscopic visual phenotypes of CF materials and their endogenous phytochemical profiles. Four algorithms, namely Random Subspace Method (RSM), Wide Neural Network (WNN), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN), were subsequently employed to establish identification models for CF cultivars. The results demonstrated that all four models exhibited excellent discriminative ability, among which the RSM algorithm achieved the optimal performance, with an accuracy of 94.4% in the training set and 96.7% in the test set. These results construct an innovative methodological framework for fast and precise identification of CF cultivars, and deliver solid data basis as well as practical technical schemes supporting quality appraisal of Traditional Chinese Medicine Materials.