Recent advances in generative artificial intelligence (GenAI) have enabled the creation of high-resolution synthetic images, offering an alternative to traditional data collection for training computer vision models in agriculture. In crop disease diagnosis, synthetic images can supplement datasets when real image acquisition is limited, potentially reducing resource-intensive field collection. Therefore, this study evaluated how different ratios of real-field to Gen-AI-based synthetic watermelon (Citrullus lanatus) disease images (including an additional unknown class) affect EfficientNetV2-L classification performance and feature-space separability. The training dataset was divided into five treatments: H0 (real images only), H1 (synthetic images only), H2 (equal real-to-synthetic ratio), H3 (one real image to ten synthetic images, 1:10), and H4 (H3 plus random images to enhance variability). Models were trained using a custom EfficientNetV2-L architecture with fine-tuning and transfer learning approaches. Treatments H2, H3, and H4 demonstrated strong and consistent performance across all classes, with H2 achieving overall accuracy of 0.80, followed by H3 (0.98) and H4 (0.98). H3 achieved near-perfect precision and recall (0.95–0.99) across all classes, resulting in F1-scores of 0.98. H4 also maintained high precision and recall scores (0.94–1.00), including accurate detection of the additional unknown class (F1 = 0.98). Overall weighted F1-scores increased substantially from 0.72 (H0) to 0.81 (H2) and reached 0.98 in H3-H4, indicating the benefit of hybrid synthetic-real data fusion. These findings show that real-synthetic data fusion enhances model performance and generalization, while synthetic images alone were not effective under the tested conditions.
Rai et al. (Wed,) studied this question.