Explainable AI -driven deep learning model for the root image-based classification of the medicinal plants Saposhnikovia divaricata, Glehnia littoralis, and Peucedanum japonicum

Understanding how ecological functional traits are encoded and recognized by data-driven models remains a critical challenge in ecological informatics. In particular, root anatomical traits such as xylem coloration and phloem texture reflect plant adaptation to environmental conditions, yet their quantitative interpretation from image data is still limited. In this study, we propose a quantitative explainable artificial intelligence (XAI)-based framework to investigate whether deep learning models capture ecologically relevant root phenotypic traits in root cross-sectional images of three morphologically similar but ecologically distinct medicinal plants ( Saposhnikovia divaricata , Glehnia littoralis , and Peucedanum japonicum ). Rather than focusing on model performance alone, we design a structured analytical pipeline that links model attention to trait-relevant anatomical structures and evaluates their consistency with ecologically interpretable trait signals. Using region-based masking and activation analysis, we quantify how models attend to xylem and phloem tissues and examine how controlled manipulation of color and texture features influences trait-specific attention responses. The results show that model attention is systematically associated with anatomical regions linked to plant functional traits, and that architectures designed for fine-grained recognition exhibit stronger alignment with ecologically meaningful features. These findings demonstrate that interpretable deep learning can be extended beyond prediction to serve as a framework for extracting and validating ecological signals embedded in plant functional traits. The proposed approach provides a pathway for integrating image-based AI with ecological trait analysis, enabling quantitative linkage between image-derived phenotypic traits and environmental adaptation in ecological informatics. • Built an RGB dataset of three ecologically distinct herbal species for trait analysis. • Benchmarked CNN, ViT, and TransFG for trait-based classification, with TransFG best. • XAI maps showed focus on functional traits like xylem color and phloem texture. • Activation Intensity Ratio quantified saliency under color- and texture-enhanced inputs. • Proposed an interpretable AI workflow linking trait monitoring and expert evaluation.