Authors

Abstract

The number of main stem nodes in soybean is an important phenotypic trait for evaluating plant growth, varietal characteristics, and breeding potential. However, automatically detecting soybean main stem nodes under field conditions remains challenging because of plant occlusion, variations in illumination, and complex backgrounds, while manual counting is inefficient for large-scale phenotypic data acquisition. Therefore, this study proposes BGE-YOLOv11, a lightweight model for detecting soybean main stem nodes in complex field environments. Using an image dataset of soybean main stem nodes collected under field conditions, the model was developed by integrating a bidirectional feature pyramid network (BiFPN), Global–Local Self-Attention (GLSA), and EfficientHead into the YOLOv11n baseline to enhance multiscale feature representation and detection performance. The proposed model was evaluated in terms of precision, recall, mean average precision at an intersection over union threshold of 0.5 (mAP50), and model complexity. BGE-YOLOv11 achieved a precision of 94.9%, recall of 92.9%, and mAP50 of 97.6%, representing improvements of 3.9, 4.0, and 3.0 percentage points, respectively, over the YOLOv11n baseline. Moreover, the model contained only 1.87 million parameters, required 5.6 GFLOPs, had a memory footprint of 8.0 MB, and achieved an inference time of 3.0 ms. These results demonstrate that BGE-YOLOv11 improves detection performance while retaining a lightweight architecture. Nevertheless, its performance may be affected by severe occlusion and highly variable field conditions. Future research will focus on expanding the dataset and further improving the model’s generalization ability. This study provides an effective method for the automated counting of soybean main stem nodes and the analysis of phenotypic traits.

Abstract in Chinese

大豆豆荚着生于主茎节点，其数量是影响单株结荚量及大豆考种的重要指标。为实现田间成熟期大豆主茎节点的实时测量，提出一种面向田间成熟期大豆主茎节点数量的快速轻量化检测模型BGE-YOLOv11。针对邻株遮挡与背景干扰问题，通过引入BiFPN、GLSA模块和EfficientHead检测头提升模型的检测能力。BGE-YOLOv11 模型的检测精确率为 94.9%，召回率为 92.9%，mAP50 达 97.6%，参数量为1.87M。与 YOLOv11n基础模型相比，分别提升3.9%,4.0%,3.0%。参数量下降 0.71M，浮点运算数缩减为5.6 G，内存占用量仅为 8.0 MB，推理时间为3.0ms。在降低模型计算量的同时满足检测需求。BGE-YOLOv11 模型实现了田间背景下大豆主茎节点的精准检测及轻量化部署需求，为节点间距、主茎长度等表型参数的精确计算提供数据基础，为田间植株主茎节数的高效统计提供技术支撑，进而可应用于大豆优良品种的筛选与培育。