A Geometry-Aware Road-Constrained Framework for Weed Quantification and Operational Workload Assessment in Vineyard Roads

To address the difficulty of road-constrained weed extraction and operational assessment in orchard road regions under weed encroachment, background interference, and complex illumination, this study developed a vision-based framework integrating road segmentation, in-road weed extraction, spatial quantification, and workload evaluation. A joint image enhancement strategy combining LAB-based luminance correction, HSV-based color gain adjustment, ExG enhancement, and morphological refinement was first applied to improve the separability of green vegetation targets. An improved YOLOv11 with an SE attention mechanism was then used for robust orchard road segmentation. On this basis, road-region constraints and a dual-threshold HSV–ExG strategy were combined to extract in-road weeds and calculate global weed coverage. Furthermore, a geometry-adaptive grid based on actual road boundaries was constructed to quantify grid-cell coverage, aggregation, spatial heterogeneity, and workload index. Results showed that the proposed enhancement method increased the mean and standard deviation of ExG by 21.30% and 19.22%, respectively. The improved YOLOv11 achieved 91.28% precision, 87.52% recall, 93.37% AP50, 68.31% mAP@0.5:0.95, and 89.36% F1-score. Across five sample groups, global weed coverage ranged from 0.6123 to 0.6471, and the workload index ranged from 0.6403 to 0.6859. Overall, the proposed method establishes an integrated image-based analytical pipeline that may support future variable-rate weeding and decision-making after further operational validation.