<![CDATA[Rice leaf diseases pose a significant threat to agricultural productivity, yet accurate automated detection remains challenging due to low image quality in field conditions. This study proposes the integration of ESRGAN-based super-resolution with YOLOv12 for rice leaf disease detection using a dataset of 6,204 annotated field images spanning five classes: bacterial leaf blight, brown spot, healthy leaves, hispa, and leaf blast. To prevent data leakage, all images were partitioned into training, validation, and testing subsets prior to augmentation; under the S3 scenario (75:10:15), the training set was expanded from 4,653 to 13,959 images through augmentation. ESRGAN RRDB 4× enhancement was applied exclusively to all test-set images, enabling a clean before-and-after comparison without contaminating training data. The primary finding is that ESRGAN produces a modest but consistent improvement in detection performance: mAP@0.5 increased from 0.949 to 0.955, and mAP@0.5:0.95 increased from 0.910 to 0.925. Per-class analysis shows the largest gains in visually challenging classes, particularly Leaf Blast (+0.073) and Hispa (+0.060). Additionally, four dataset partitioning scenarios (S1–S4) were trained and evaluated under identical settings; as a preliminary observation, the S3 configuration offered a balanced trade-off between training data availability and test-set reliability, though definitive conclusions require further validation through repeated experimental runs.]]>
