Hybrid GAN–Diffusion and Lesion-Aware Generative Models for Crop Disease Classification in Precision Agriculture

Crop diseases are an ongoing threat to global food
production, and advancements made in automated diagnosis are
limited due to the public image datasets being sparse and
disorganized. Applications of Generative Adversarial Networks (GANs) to training set augmentation mostly neglect important lesion characteristics or produce images that focus on idealized laboratory conditions, and therefore, miss capturing conditions in the field. To address this gap, we created a hybrid framework that combines conditional GANs, lesion-aware synthesis,diffusion-based refinement, and Cycle GAN-driven domain
adaptation. This framework was validated on Plant Village and
the New Plant Disease dataset for tomato late blight and grape
Esca. We evaluated image quality using SSIM, FID, and LPIPS
metrics. Classification performance was evaluated using CNN
and lightweight models, Mobile Net and Efficient Net. The
pipeline showed a 7.2% increase in recall and 6.6% increase in
F1-score compared to standard augmentation, specifically
targeting the under sampled disease class, where augmentation
impact was most significant. All figures are reproducible given
the five independent iterations per dataset. These were executed
on an NVIDIA RTX 3090 (24 GB) GPU, with PyTorch 2.1. The
additional computational burden that diffusion refinement
entails and the dependence on the quality of segmentation in
lesion-aware synthesis do not overshadow the framework’s
enhancement in reliability and robustness, nor the fully
operational pathway toward edge-ready solutions in precision
agriculture