CornCare: A Knowledge-Graph-Enhanced Multimodal Diagnostic Reporting System for Corn Diseases

Accurate and actionable crop disease diagnosis requires not only visual recognition of disease symptoms but also the ability to generate grounded reports that integrate symptom interpretation with agronomic knowledge. Existing image-based plant disease diagnosis methods mainly focus on disease classification and often lack fine-grained symptom description, evidence retrieval, and decision-oriented report generation. To address these limitations, we propose CornCare, a multimodal framework for corn disease diagnosis and diagnostic report generation that combines visual recognition, phenotype captioning, document retrieval, and knowledge-graph-based recommendation support. Given a field corn image, CornCare first localizes disease-relevant leaf regions to reduce background interference. The localized leaf image is then used for disease classification and phenotype caption generation, producing both a disease category and a fine-grained symptom description. These outputs jointly support hierarchical knowledge retrieval, where the disease category narrows the search to relevant expert documents and the phenotype caption retrieves symptom-consistent evidence. The retrieved evidence is further combined with a structured agricultural knowledge graph to generate diagnostic reports with symptom interpretation, likely causes, and management suggestions. Experiments show that CornCare achieves competitive performance in disease identification and phenotype description generation while improving the groundedness, completeness, and practical usefulness of generated diagnostic reports. These results suggest that combining multimodal perception with symptom-grounded knowledge retrieval provides a promising path toward more practical and explainable crop disease diagnosis.