Latent diffusion meets semi-supervised learning: Enhancing plant disease detection with minimal labels

Accurate assessment of plant disease severity is essential for sustainable agriculture, but is often limited by the scarcity of annotated data. Generative models offer a promising solution by producing realistic synthetic samples that can enrich training datasets and improve model generalization. However, most existing approaches rely on large, fully labeled datasets, limiting their applicability when expert annotations are costly or difficult to obtain. This study introduces a semi-supervised latent diffusion model for the conditional generation of wheat leaf images with controllable yellow rust ( Puccinia striiformis ) disease severity. The model conditions generation on leaf segmentation masks and disease severity scores, leveraging a small labeled subset and a large set of unlabeled images. This enables the creation of realistic, severity-controllable synthetic images that effectively augment limited training datasets. Experimental results demonstrate substantial improvements in downstream regression performance, particularly under extreme annotation scarcity. Notably, when applying a per-range reduction strategy with only ten labeled images (one per disease severity bin), the baseline regression model trained exclusively on real labeled images achieved an R 2 of 0.44, whereas augmentation with semi-supervised synthetic images increased R 2 to 0.79—a gain of ＋0.35. Improvements were observed across all annotation levels, with consistent reductions in MAE and RMSE, highlighting the effectiveness of combining semi-supervised generative training with balanced sampling. These findings confirm that leveraging unlabeled data in generative diffusion models produces realistic, controllable synthetic images that enrich supervised learning. The proposed approach provides a scalable, cost-effective pathway to reduce dependence on large labeled datasets while maintaining precise control over disease severity.