Deep learning-based detection of Grapevine fanleaf virus in Türkiye: a comparative study of ROBOFLOW 3.0, YOLO-NAS, YOLOV11, and YOLOV12

Grapevine fanleaf virus (GFLV) poses a significant phytopathological threat to global viticulture, demanding the development of rapid and accurate detection methodologies to mitigate economic losses. In this study, four deep learning-based object detection architectures were implemented and comparatively evaluated, ROBOFLOW 3.0 (Fast), YOLO-NAS (Accurate and Medium), YOLOv11 (Fast), and the latest YOLOv12 (Fast), for the automated identification of GFLV symptoms on grapevine leaves. A custom dataset comprising high-resolution images captured under diverse vineyard conditions was curated and meticulously annotated. YOLO models were trained on 100-350 epochs. Model performance was assessed using metrics such as Precision, Recall, Mean average precision at Intersection over Union (IoU)=0.5 (mAP@50), Mean average precision across IoU thresholds (mAP@50–95), and F1 score. Among the evaluated models, YOLOv11 Fast achieved the highest mAP@50 value (93.4%) and an F1 score of 88.3%. In comparison, ROBOFLOW 3.0 Fast reached a mAP@50 of 89.4% and an F1 score of 86%, YOLO-NAS Accurate achieved mAP 92.4% and F1 85.2%, and YOLO-NAS Medium obtained 91.9% and 85.3%, respectively. The latest version YOLOv12 Fast yielded mAP@50 91.3% and F1 86.5%. These results highlight the potential of next-generation YOLO architectures for high-throughput, real-time GFLV detection, offering a scalable and efficient solution for precision viticulture and integrated disease management.