PodCountNet: UAV-Based Soybean Pod Detection and Counting for High-Throughput Sink Trait Phenotyping

Photosynthetic source strength is a major determinant of soybean yield formation through its regulation of reproductive sink development. However, high-throughput quantification of sink traits under contrasting photosynthetic source conditions remains a major bottleneck for field-based source–sink physiology research. In this study, three soybean cultivars were subjected to graded artificial defoliation (0%, 25%, and 50%) at the R2 flowering stage to establish different levels of defoliation-induced source limitation. Biological observations showed that mature pod number consistently declined with increasing defoliation intensity across all cultivars, supporting the critical role of flowering-stage source limitation in determining sink formation. To enable efficient quantification of these sink responses, Unmanned Aerial Vehicle (UAV)videos acquired at maturity were used to construct a field soybean pod dataset, and an integrated detection–tracking framework, PodCountNet, was developed by combining an improved SPDM-YOLO detector with the OC-SORT tracking algorithm. The proposed SPDM-YOLO achieved a precision of 88.2% and an mAP@0.5 of 94.5% under diverse canopy conditions. Although single-view counting accuracy varied with canopy architecture and illumination, dual-view fusion effectively reduced systematic errors and, achieving MAE = 66.3 pods and RMSE = 88.0 pods at the individual plot level (n = 27), substantially outperforming single-view front-lit (MAE = 114.2) and back-lit counting (MAE = 172.9). The proposed framework enables accurate, non-destructive, and high-throughput quantification of soybean sink responses under contrasting photosynthetic source conditions, providing an effective phenotyping tool for investigating source–sink coordination and supporting photosynthesis-oriented soybean breeding.