Abstract
Early identification of grapevine diseases is critical for reducing yield losses and ensuring sustainable viticulture. CNNs trained on benchmark datasets such as PlantVillage often achieve near-perfect accuracy, yet this performance fails to translate to real-world field conditions where lighting, backgrounds, and lesion appearance vary widely. To address challenges of data scarcity and imbalance, this study introduces VitiForge, a novel procedural synthetic imagery pipeline for generating realistic synthetic grape leaf textures representing healthy, Black Rot, Esca, and Leaf Blight conditions. VitiForge is systematically evaluated against GAN-based augmentation through a data ablation study on PlantVillage and FieldVitis, a curated field dataset, using MobileNetV2, InceptionV3, and ResNet50V2 classifiers. Results show that VitiForge significantly improves performance in low-data regimes, enabling model training even without real samples, whereas GAN augmentation proves more effective once sufficient real data is available. On field imagery, VitiForge often matched or surpassed GAN-based methods, particularly when paired with MobileNetV2. These findings highlight the complementary roles of procedural and GAN-based synthetic data: VitiForge offers flexibility and scalability under cross-domain and data-scarce conditions, while GANs enhance realism and variability when ample data exists. Together, they support the development of robust and generalizable models for automated grape disease detection in precision agriculture.