Authors

Abstract

Accurate soybean field phenotyping is increasingly important for breeding. However, traditional measurement methods are labor-intensive and subjective, while UAV-based approaches are challenged by complex backgrounds and densely distributed small targets. This study first develops UAV-ZSAR to transform oblique UAV images into horizontal-view images and reconstruct plant geometry. A lightweight point-based model, Soy-MOPNet, is then proposed for fast and parallel detection of soybean seeds and stem nodes. The model incorporates the proposed SDConv, optimized hierarchical dilated convolution (HDC) principles, and PBOS to enhance adaptive feature fusion, receptive field design, and multi-branch training stability, respectively. Based on the detected keypoints, six phenotypic traits are extracted in parallel, providing comprehensive support for field phenotyping, breeding selection, and precision agricultural management.

Abstract in Chinese

精准的田间大豆表型分析对育种研究日益重要，但传统表型测量方法费时费力且主观性强，而基于无人机的高通量表型采集又面临田间背景复杂和小目标密集等挑战。本文首先构建了 UAV-ZSAR 方法，将倾斜无人机图像转换为水平视角图像，以恢复植株几何形态。随后，构建了轻量化点式模型 Soy-MOPNet，用于大豆豆粒和茎节点的快速并行检测。该模型引入了所提出的 SDConv、优化的 HDC 原则和 PBOS，分别用于提升自适应特征融合能力、感受野设计效果和多分支训练稳定性。基于检测到的关键点，进一步并行提取了 6 个表型性状，从而为田间表型分析、育种筛选和精准农艺管理提供更系统、更全面的支持。