Accurate monitoring of soybean seedlings in the field is a core component for implementing scientific management during the seedling stage and unlocking yield potential. Traditional manual survey methods are inefficient and highly subjective, making them inadequate for real-time assessment at the field scale. This study addresses challenges such as the small size of individual seedlings, dense inter-plant overlap, blurred boundaries, and complex interferences from soil residue and varying illumination by proposing a high-precision method for soybean seedling instance segmentation and georeferenced localization based on low-altitude (12 m) Unmanned Aerial Vehicle (UAV) imagery. By implementing targeted improvements in the YOLOv11n-seg model, we developed the YOLOv11-seg-SSC model, which integrates the SCSA (Shared Cross-Semantic Space and Progressive Channel Self-Attention) mechanism, the Context-Guided (CG) Block, and a lightweight Slim-Neck structure based on GSConv and VoV-GSCSP. While significantly reducing computational complexity (approximately 9.5 GFLOPs and 2.96 M parameters), the model improved the mean average precision for segmentation (mAP@0.5 Mask) from the baseline of 80.6% to 83.3%, maintained a stable detection mAP@0.5 (Box) at 95.9%, and achieved an overall segmentation precision of 85.1% and recall of 80.3%. This approach not only meets the requirements for near-real-time field processing but also outputs seedling spatial distribution results with true geographic coordinates through georeferenced mapping, thereby providing directly applicable data support for seedling count statistics, missing seedling diagnosis, population spatial pattern analysis, and variable-rate management. This study establishes a complete technical pipeline from precise UAV image segmentation to spatially informed seedling status decision support, offering a theoretical foundation for efficient and accurate monitoring of soybean seedlings in the context of smart agriculture.
Yue et al. (Sat,) studied this question.