Construction and validation of a YOLO-Soy Model for field soybean seed counting on a self-propelled phenotyping platform

To address the bottlenecks of low efficiency, poor consistency, and inadequate compatibility with high‑throughput phenotyping pipelines inherent in manual field‑based seed counting during soybean breeding, this study developed and validated an enhanced automatic soybean seed detection and counting model, YOLO‑Soy, tailored for complex field environments. Built on a YOLO11n backbone, the model integrates a Zoom multi‑scale feature fusion module, a C2PSA self‑attention enhancement module, a ScalSeq hierarchical feature sequence aggregation module, and a soybean-specific detection head. These additions systematically enhanced the saliency of tiny-seed features under dense occlusion and complex backgrounds and strengthened the capacity for foreground-background separation. Experiments were conducted using two‑year field imagery (2024–2025) and a year-stratified leave-one-year-out cross-validation strategy for training and validation. Ablation study revealed that the four improved modules are functionally complementary, forming a comprehensive pipeline of interference mitigation, scale adaptation, precise feature fusion, and detection output transformation. A single module exhibited limited effect when acting independently, whereas multi-module synergy produced substantial gains. Test-set results demonstrated that the seed counts predicted by the model were highly consistent with manual ground truth, achieving a coefficient of determination ( R ²) of 0.934, a mean relative error of 2.446%, a mean average precision (mAP@0.5) of 0.737, and an inference speed of 58.78 FPS. These metrics satisfy the requirements for real-time field detection. The findings indicated that YOLO-Soy can accelerate the seed‑counting step in variety selection processes, greatly reducing manual workload and subjective errors.