In modern precision agriculture, computer vision-based early detection and precise control of plant diseases have emerged as critical technologies for safeguarding crop productivity. However, the collection of high-quality field data, essential for tr...
In modern precision agriculture, computer vision-based early detection and precise control of plant diseases have emerged as critical technologies for safeguarding crop productivity. However, the collection of high-quality field data, essential for training deep learning models, is constrained by significant costs and the need for expert labor. While utilizing laboratory data with accurate labels has been explored as an alternative, the severe visual domain gap between controlled laboratory environments and complex outdoor field conditions remains a primary factor hindering the practical deployment of models. Furthermore, pathogen migration and host range expansion driven by climate change are frequently leading to the emergence of novel crop-disease combinations. Conventional diagnostic models, trained on fixed categories, fail to adapt flexibly to these ecological shifts, resulting in structural inefficiency as they require data recollection and retraining for every new outbreak.
To address these limitations in data acquisition and unpredictable ecological changes, this study proposes a strategy to maximize data efficiency through methodological innovation rather than extensive physical data augmentation. First, to enhance field adaptability, we propose a framework integrating Background Re-composition and Unsupervised Domain Adaptation (UDA) that synthesizes laboratory-extracted leaf regions with unlabeled real-field backgrounds and aligns feature distributions via domain adaptation mechanisms, thereby establishing a diagnostic model robust to environmental variations without additional field annotations. Furthermore, to actively address rapidly changing disease spread patterns, a Semantic Feature Disentanglement framework is introduced. Moving beyond conventional classification methods where crop and disease features are entangled, this method explicitly partitions feature vectors and applies contrastive learning and a Local Feature Attention mechanism to independently isolate disease-specific visual features from host attributes. Moreover, by aligning these features with semantic embeddings generated by Large Language Models (LLMs), the model demonstrated improved adaptability to unseen crop-disease combinations compared to baseline models, thereby validating the feasibility of compositional generalization. Ultimately, this study establishes a methodological foundation for a data-efficient plant disease diagnosis system capable of effectively addressing the environmental constraints and ecological variations of real-world agriculture with limited data resources, leveraging the dual approaches of domain adaptation and feature disentanglement.