<![CDATA[Rice leaf diseases significantly reduce crop productivity, making accurate and reliable automatic detection essential for precision agriculture. Although recent YOLO-based approaches have achieved high detection accuracy under controlled conditions, the robustness of YOLOv11n against diverse lighting conditions in rice leaf disease detection remains insufficiently investigated, particularly under real-world field environments. This study evaluates the robustness of YOLOv11n using synthetic photometric transformations (S0-S11) and real field lighting conditions (L1-L4). Model performance was assessed using Precision, Recall, F1-Score, and mAP@0.5. Under baseline conditions, YOLOv11n achieved a Precision of 0.955, Recall of 0.936, F1-Score of 0.945, and mAP@0.5 of 0.949. The model remained highly robust to mild overexposure, mild underexposure, and saturation variations, while severe underexposure, extreme overexposure, and exposure scaling caused moderate to substantial degradation. Partial shadow produced the most severe performance decline, reducing mAP@0.5 by 90.20%. Furthermore, all real-world lighting scenarios exhibited performance drops exceeding 87%, revealing a pronounced domain gap between synthetic simulations and field conditions. These findings demonstrate that global photometric transformations alone are insufficient to represent complex real-world illumination, providing practical evidence for developing more robust lighting-aware object detection models for agricultural applications and emphasizing the importance of realistic robustness evaluation]]>
Copyrights © 2026
