Improved YOLO for Lesion Segmentation of Cotton Verticillium Wilt in the Field

To address the low efficiency and high subjectivity of manual field surveys of cotton Verticillium wilt, this study proposes an instance segmentation model for cotton leaf Verticillium wilt lesions adapted to complex field backgrounds, enabling precise segmentation of Verticillium wilt lesions. Based on in-situ images of infected cotton leaves, YOLOv11 is adopted as the baseline, and a multi-scale feature fusion attention (MFFA) mechanism is proposed. By integrating channel and spatial attention, the model captures extensive contextual information and enhances feature representation, and its performance is compared with mainstream attention mechanisms. A Swin Transformer feature extraction module and the Slide Loss function are also introduced. The results of ablation experiments show that compared with the baseline model, the improved modules enhance the overall performance of the model under the evaluation protocol of this study, and the combination of modules yields higher performance than individual modules. The model integrating all improved modules achieves a Precision of 91.2%, Recall of 83.2%, F1-Score of 87.0%, mAP50 of 87.6%, and mAP50–95 of 67.5%, enabling effective pixel-level identification of small-target lesions under complex field backgrounds. Comparison with representative segmentation models under the identical protocol further shows that the proposed model achieves the best overall accuracy. The improved segmentation model can improve the segmentation accuracy of small-target lesions and enhance overall model performance, and may offer a technical basis for future applications such as efficient field surveys of Verticillium wilt, resistance phenotyping, and precision disease management.