Abstract
Unmanned Aerial Vehicle (UAV) imagery is increasingly used to quantify plant traits such as stand count, providing critical information on crop emergence for replanting and field-management decisions. However, the generalization ability of existing methods remains limited across crop types and field conditions, particularly when only small target-domain datasets are available. This study developed an open-source and generalized plant stand count pipeline based on transfer learning. UAV imagery was collected from cotton, corn, and soybean fields under different years, UAV platforms, cameras, flight conditions, and field environments. Four object detection frameworks, including YOLOv8x, YOLOv9e, YOLOv10x, and RT-DETR-X, were systematically tuned on a three-year cotton dataset in Phase A and subsequently fine-tuned on small soybean and corn datasets in Phase B. Under the main target-domain model-comparison protocol, 30 original training images per crop were expanded to 90 stored images through offline geometric augmentation. The best-performing models achieved mAP@50 values of 0.894 for soybean and 0.948 for corn. A separate label-budget analysis using the original target-domain images without additional offline augmentation showed that cotton-domain pretraining was most beneficial when only one labeled target-crop image was available, increasing mAP@50 from 0.189 to 0.860 for soybean and from 0.135 to 0.527 for corn. The pipeline further converted plant detections into georeferenced spatial point layers, enabling field-scale stand-count mapping from UAV orthomosaics. These results indicate that the proposed transfer-learning pipeline can reduce crop-specific labeling requirements while preserving field-scale stand-count capability. The open-source implementation is available at https://github.com/MUPAAL/Germination_Evaluation.