YOLO-CottWed: A Lightweight Network for Fine-Grained Weed Detection in Cotton Fields

Efficient and robust weed detection is essential for sustainable cotton farming. Existing deep learning methods often face a trade-off between accuracy and computational cost, limiting their practical deployment. To address this, we propose YOLO-CottWed, a lightweight fine-grained weed detection model tailored for cotton fields. The backbone incorporates a novel Depthwise Separable Parallel Convolution (DSPC) module, which not only reduces computational complexity but also enhances model robustness under noisy conditions by simultaneously extracting spatial and channel-wise features. In the neck, a Lightweight Neck module applies channel constraint and projection convolutions to improve multi-scale feature fusion efficiency. The model employs the Extended IoU (EIoU) loss to enhance bounding box localization. Experiments on the CottonWeedDet12 dataset demonstrate that YOLO-CottWed achieves an [email protected] of 91.9%, outperforming YOLOv8n (91.3%) while maintaining a comparable FPS of 233. Moreover, YOLO-CottWed reduces parameter count from 3.0 M to 1.81 M and FLOPs from 8.2 G to 6.2 G. The YOLO-CottWed design can also be transferred to YOLOv10, producing a lighter and faster model with competitive performance. These results highlight the efficiency, robustness, and transferability of YOLO-CottWed, making it suitable for real-time, resource-constrained agricultural applications and precise weed management.