Benchmarking Local Vision-Language Models and Knowledge-Graph-Assisted Differential Diagnosis of Plant Diseases

Image-based plant disease diagnosis is complicated by overlapping symptoms and variable imagining conditions. We evaluated six locally deployable vision-language models (VLMs) as generators of searchable symptom descriptions and integrated the selected model with knowledge-graph-assisted differential diagnosis. The dataset comprised 26,814 images representing 22 diseases, with 21,441 reference images and 5,373 internal-evaluation images. Symptom descriptions were embedded using Qwen3-Embedding-8B and evaluated by Top-1 retrieval, soft voting, and hard voting among same-host reference cases. Hard voting improved internal accuracy for all six VLMs. Qwen3.5-35B-A3B achieved 68.94% accuracy and 66.33% macro F1 at K = 19, compared with 40.87% accuracy and 46.25% macro F1 for a direct-classification baseline using the image, host, and allowed disease labels without retrieval. The selected VLM was integrated with a Neo4j knowledge graph in AI Plant Doctor version 2. In an idealized question-answering simulation using a balanced subset of 440 images and complete, ground-truth-consistent responses, the GraphRAG-only and hybrid retrieval-assisted workflows achieved accuracies of 74.55% and 97.27%, respectively. These values are not estimates of field diagnostic accuracy. This modular and explainable framework provides a promising foundation for plant disease decision support and is available through AI Plant Doctor version 2 at https://pdoc.scnu.ac.kr/pdoc2.