YOLO-D: A high-precision identification model for sugar beet and weeds at different growth stages

Weed detection is essential for efficient weed management, but agricultural environments and the similarity between crops and weeds make detection challenging. Current deep learning methods often face issues like limited scenario variations, insufficient image samples, and low detection accuracy. Furthermore, weeds and crops vary in shape and color at different growth stages, complicating detection further. To address these issues, we propose a high-precision identification method based on You Only Look Once (YOLO-D) model, for sugar beet (Beta vulgaris) and weeds at different growth stages. The model uses the FasterNet-S backbone to improve feature expression capability and receptor field coverage while maintaining high detection speed. It also introduces the C2F module, incorporating more residual connections to enhance gradient flow within the network structure. An efficient decoupling head is incorporated, reducing computational costs and achieving lower inference latency. The Lincoln beet dataset was used for training and evaluation. Compared to existing studies, the proposed YOLO-D model achieves an overall mean average precision (mAP) improvement of 3.3%, reaching 75.8%. The mAP for sugar beet increases by 1.6%, reaching 87.3%, and the mAP for weeds increases by 5.1%, reaching 64.4%. On the public sesame (Sesamum indicum) dataset, “crop and weed detection data with bounding boxes,” it achieves a high mAP of 88.9%, with an overall mAP increase of 1%.