<![CDATA[Rice is a staple food crop for more than half of the global population, with Asia contributing approximately 90% of global production. Leaf diseases, particularly Blast and Bacterial Blight, are major factors contributing to reduced rice productivity in Indonesia. Conventional detection methods based on visual observation are often subjective and time-consuming, highlighting the need for more reliable automated detection systems. This study aims to implement StyleGAN2-ADA to generate synthetic rice leaf disease images and evaluate its impact on the performance of Convolutional Neural Network-based classification. This research employed a quantitative experimental approach by comparing two scenarios: Baseline without GAN augmentation and Proposed with StyleGAN2-ADA synthetic image augmentation. Two CNN architectures, EfficientNetB0 and ResNet50, were evaluated using accuracy, precision, recall, F1-Score, and confusion matrix metrics. The quality of synthetic images was assessed using the Fréchet Inception Distance. The results demonstrated that StyleGAN2-ADA augmentation improved the overall F1-Score of EfficientNetB0 from 97.62% to 98.20%, with the largest improvement observed in the Blight class, increasing by 3.11%. For ResNet50, the overall F1-Score increased from 97.60% to 98.20%, although the Blast class showed no performance improvement after augmentation. GAN augmentation provided the most consistent benefits for the minority Blight class, while its impact on the Blast class varied across metrics. In EfficientNetB0, improvements in precision and F1-Score were accompanied by a decrease in recall. These findings indicate that model evaluation should consider class-specific performance and the trade-off between precision and recall rather than relying solely on aggregate metrics]]>
