Objective Cotton seedling deficiency can significantly reduce crop yield, and early assessment is crucial for effective field management. Traditional field survey methods cannot meet the needs of large-scale assessment. To address this issue, this study proposed a new method for cotton seedling deficiency assessment based on UAV imagery and the WCD-YOLO model.

Methods This model integrates an improved GhostNet as a lightweight backbone network, a BiFPN neck network combined with a P2 detection head to enhance small target detection, and replaces EIoU with the loss function to address dataset imbalance. The method includes regional division of UAV images of cotton fields, calculation of regional seedling deficiency rates, and establishment of a correlation function between seedling spacing and deficiency rate, ultimately enabling regionalized deficiency assessment in cotton fields.

Results Compared with the baseline YOLOv5m, WCD-YOLO reduced the model size to 18.2 MB, which is only 45.27% of the original model, and improved the mean average precision (mAP) by 2.32 percentage points, reaching 86.52%. Compared with mainstream object detection models such as TOOD, CenterNet, YOLOv3, YOLOv7, YOLOv8, and YOLOv10, the mAP of WCD-YOLO was 11.11, 7.97, 3.27, 2.25, 0.93, and 1.01 percentage points higher, respectively, and showed significant advantages in terms of parameter count and computational cost. On the three test images, the regional mean absolute errors (MAE) of this assessment method were 1.72, 2.86, and 1.78, respectively, with a coefficient of determination (R²) of 84.82%.

Conclusion This study provides a reliable solution for early replanting decisions and intelligent cotton field management, contributing to the smooth implementation of large-scale cotton production.