YOLO-CCG: An enhanced model for weed detection in complex cotton fields

Accurate weed detection is essential for site-specific herbicide application in cotton production. However, variations in weed appearance, complex field backgrounds, uneven target distributions, and leaf occlusion can result in missed detections and false positives. To address these challenges, this study proposes YOLO-CCG, an improved YOLO11-based model for weed detection in cotton fields. First, the Coordinate Attention (CA) mechanism is introduced into the backbone to embed spatial coordinate information into channel attention, thereby strengthening target localization and the representation of discriminative weed features. Second, nearest-neighbor upsampling in the neck is replaced with the Content-Aware ReAssembly of Features (CARAFE) module, which performs content-aware feature reassembly to preserve fine-grained spatial information during feature fusion. Finally, GSConv and the VoVGSCSP lightweight module are introduced into the bottom-up path of the neck to replace the corresponding standard convolution and C3k2 blocks, reducing computational complexity without increasing the parameter count. Experimental results show that, compared with YOLO11s, YOLO-CCG achieves relative improvements of 0.4%, 2.2%, 0.8%, and 1.7% in precision, recall, mAP50, and mAP50-95, respectively. These results indicate that YOLO-CCG provides an accurate and computationally efficient solution for weed detection in complex cotton-field environments.