Bridging the Lab-to-Field gap in plant disease diagnosis through unsupervised domain adaptation enhanced by background recomposition

Early detection of plant diseases caused by pests, pathogens such as viruses is essential for preventing their spread and minimizing crop damage, particularly in indoor farming and orchards. Machine learning-based computer vision approaches have emerged as promising tools for automated disease diagnosis, but their implementation faces critical challenges due to the scarcity of labeled data from real agricultural environments. Consequently, models are primarily trained using laboratory-collected images with monotonous backgrounds, resulting in significantly reduced performance when deployed in real cultivation environments. To address these challenges, this study presents a two-step adaptation method to bridge the gap between laboratory and real-field environments in plant disease diagnosis. Our approach first applies field-adaptive background recomposition for image augmentation, followed by unsupervised domain adaptation, enabling effective disease diagnosis in real agricultural environments. We quantitatively and qualitatively validated the proposed method on disease identification tasks for tomato, chili pepper, grape, and apple leaves, achieving robust performance in field environments without requiring labeled real-field data. • We developed a field-adaptive background recomposition method that augments laboratory images using real-field backgrounds to reduce the domain gap in plant disease classification. • Various unsupervised domain adaptation approaches were employed to further align feature distributions between laboratory and field domains without requiring labeled real-field data. • Finally, we demonstrated the effectiveness of our two-step adaptation method through comprehensive evaluation on various crops, showing significant improvements in disease classification performance under real-field conditions.