High-throughput Verticillium wilt detection in cotton: A comparative study of faster R-CNN and YOLOv11

Verticillium wilt (VW), a soil-borne fungal disease of cotton, can lead to significant yield loss and has become a growing problem for global cotton production. From a global perspective, the local biotype has a high pathogenicity. Traditional phenotyping and screening methods for resistance to VW are slow, costly, and prone to human error. However, advancements in object detection models can enable automated, high-throughput screening of resistant varieties, therefore, improving speed, reducing costs, and eliminating operator bias. This study develops and evaluates the effectiveness and generalisation of two widely adopted object detection models: the two-stage Faster R-CNN and the single-stage YOLOv11 for VW in cotton stems across various backbone architectures. Digital cameras were used to collect cotton stem images from several fields. The results showed that the Faster R-CNN with the ResNet-101 model achieved a mean average precision (mAP at intersection over union (IOU) of 0.5) between 5 % and 55 % higher for the most complex YOLOv11-x and simpler YOLOv11-n, respectively, on the test dataset. Further evaluation with an independent dataset confirmed that the Faster R-CNN with ResNet-101 was the most robust and generalisable model, achieving a mAP of 85.68 %, outperforming YOLOv11 models by at least 12 % and up to 82 %. However, this enhanced mAP of the Faster R-CNN model incurred a computational cost approximately 8 % higher than that of YOLOv11-x. Nevertheless, in the context of VW detection for cotton breeding, the value of a higher mAP substantially outweighs the value of a lower computational load.