Enhanced Leaf Disease Segmentation Using a Novel Yolov8-Based Framework for Precision Agriculture

Accurate segmentation of leaf diseases is critical for early detection and treatment in precision agriculture. Traditional segmentation techniques often suffer from poor generalization, noise sensitivity, and reduced accuracy when dealing with complex backgrounds or overlapping disease regions. Existing deep learning-based approaches, while powerful, face limitations in balancing detection speed and segmentation precision. YOLOv8, though robust for object detection, requires adaptation for fine-grained segmentation of irregularly shaped leaf disease spots. This work introduces a novel YOLOv8-based segmentation framework optimized for leaf disease identification. The proposed method integrates an improved feature pyramid network with multi-scale attention mechanisms to capture disease patterns across varying sizes and textures. Data augmentation strategies, including random cropping, color jittering, and background normalization, are employed to improve robustness. Post-processing using contour refinement ensures accurate boundary detection of diseased regions. Experimental evaluation on a benchmark plant disease dataset shown a mIoU improvement of 6.4%, Dice coefficient increase of 5.8%, and detection speed of 38 FPS, compared to baseline YOLOv8 models. The proposed framework achieved both real-time efficiency and high segmentation accuracy, making it suiTable.for field-level deployment in smart agriculture.