Abstract
Accurate detection of cotton seedlings is essential for monitoring crop establishment. However, manual annotation is time-consuming, and the quality of conventional automatic annotation methods is affected by illumination and background conditions. This study proposes a joint strategy combining green leaf index (GLI) automatic annotation with deep learning detection. Three annotation methods were considered: GLI, excess green minus excess red (ExG–ExR), and hue–saturation–value (HSV). Their nine combinations with three detectors (YOLOv10n, RT-DETR-L, and Mamba-YOLO-B) were evaluated using spatial data partitioning and independent cross-region testing. Under identical post-processing, GLI achieved the highest annotation quality, with an F1 score of 96.18% and a mean intersection over union (IoU) of 89.78% for matched boxes. For initial box drawing only, bounding boxes for 200 targets were generated in 0.399 s, compared with approximately 600 s for manual drawing. Detection accuracy was assessed using mean average precision (mAP) at IoU thresholds from 0.50 to 0.95 (mAP@0.5:0.95). Across three random seeds, Mamba-YOLO + GLI achieved the highest mean source-region mAP@0.5:0.95, at 0.8020. In cross-region testing, Mamba-YOLO + GLI achieved the best detection and counting performance, with an mAP@0.5:0.95 of 0.3446. However, its detection accuracy remained substantially lower than in source-region testing. These results demonstrate that GLI automatic annotation can provide effective training labels for cotton seedling detection. Among the evaluated combinations, Mamba-YOLO + GLI achieved the highest mean source-region mAP@0.5:0.95 and the highest cross-region mAP@0.5:0.95.