• Developed a multimodal vision framework using root and leaf–flower images • Proposed a YOLOv11-PST model with improved fine-grained phenotypic recognition • Implemented scale–seed ensembling and probability-level fusion for stability • Achieved 98.56% accuracy and 97.77% Macro-F1 across 12 P. lactiflora varieties • Integrated SHAP and Grad-CAM++ to reveal biologically meaningful visual cues The quality of Radix Paeoniae Alba (Baishao), a widely used traditional Chinese medicinal material, is closely linked to its geo-authentic origins, resulting in distinct regional varieties. Conventional assessment methods typically depend on chemical or molecular profiling, which, despite accuracy, are destructive, time-consuming, and unsuitable for large-scale screening. In this study, a computer vision-based framework was established using images of the original plant ( Paeonia lactiflora Pall.) cultivated under uniform conditions. Images of root and leaf-flower organs were captured to extract structural, textural, and chromatic features relevant to varietal differentiation. Machine learning models based on handcrafted features and deep learning models utilizing an enhanced YOLOv11 architecture were systematically compared. A variety-level multimodal fusion strategy integrating root and leaf-flower data through probability-space weighting further improved robustness and balance. Results indicated that root images yielded higher accuracy, whereas leaf-flower images provided more balanced classification across varieties. The multimodal fusion model consistently outperformed single-modality models, achieving a macro-F1 score of 98.56±0.57%. This represented an improvement of approximately 5 percentage points over the best unimodal deep learning models and substantially surpassed traditional machine learning baselines. Interpretability analysis employing SHAP and Grad-CAM++ revealed that the models effectively captured biologically relevant traits such as root branching patterns, contour structures, leaf venation, and petal margins. These findings suggest that image-based phenotypic information can reliably identify P. lactiflora varieties, offering a scalable and interpretable approach to modernizing quality control in traditional Chinese medicine.
Wang et al. (Sun,) studied this question.