Résumé
Pests and diseases threaten global crop yields, yet the absence of standardized plant-damage datasets limits progress toward general, robust diagnostic tools. Existing resources differ widely in label conventions and scope, hindering interoperability and model generalization. We introduce a fully automated method for harmonizing plant-damage labels across heterogeneous datasets by mapping them to the European and Mediterranean Plant Protection Organization (EPPO) taxonomy. The approach uses large-language-model (LLM) embeddings to capture semantic similarity among label terms, including synonyms, multilingual variants, and vernacular names.Across multiple mapping strategies, embedding-based similarity using OpenAI’s text-embedding-3-large provided the best performance, reaching an F1 score of 0.836 at optimal thresholds and outperforming string-based Levenshtein matching and other LLM baselines. Applying this method, we unified five expert-curated datasets, including the newly released ePhytia collection, yielding 79,808 images mapped to 1895 EPPO-aligned classes.To assess the value of this harmonization, we finetuned a generalist pretrained Vision Transformer for large-scale plant-damage identification. Models trained on LLM-aligned labels consistently surpassed those trained with edit-distance mappings. On independent EPPO test images, our best model achieved 19.4% top-1 accuracy across 1091 classes and 33.1% on the 100 most common classes, demonstrating feasibility at unprecedented label scale. In-dataset evaluation reached 55.8% top-1 accuracy.By grounding label harmonization in an international standard, this work delivers the first large-scale, taxonomy-compliant dataset for in-field plant-damage recognition and establishes a foundation for interoperable diagnostic tools, farmer-facing mobile systems, and plant-health monitoring. We release both the harmonized dataset and the new ePhytia images to support future research.