An automated unmanned aerial vehicle framework for non-destructive soybean leaf morphology extraction using a structure-adaptive deep learning model

Abstract

Soybean leaf morphology is an important breeding trait that requires large-scale phenotyping in commercial breeding programs. Conventional leaf phenotyping still relies on manual destructive measurements, which are labor-intensive and inefficient. Low-altitude unmanned aerial vehicles (UAVs) have emerged as a high-throughput phenotyping platform, but the retrieval of leaf morphology in densely occluded canopies remains challenging due to complex canopy backgrounds, illumination heterogeneity, and leaf overlap under field conditions. To address this issue, this study developed an integrated UAV framework that couples a YOLOv10-based leaf detection module, a novel structure-adaptive segmentation network (DynamicU), and a regression-based trait prediction model for retrieving a set of leaf morphological parameters across 273 soybean genotypes under field conditions. The YOLOv10 detector reliably localized individual leaves under complex canopy conditions, achieving a mean average precision (mAP@50) of 0.84. Subsequently, the DynamicU network, whose architecture was automatically optimized via Emperor Penguin Optimization, achieved a segmentation accuracy of 97.2% and a mean Intersection over Union of 93.8%, substantially outperforming conventional models. Using random forest regression, the framework retrieved relative leaf shape traits, including length-to-width ratio and dissection index, with markedly higher accuracy (R 2 =0.97), compared to absolute morphological traits, including leaf length, width, perimeter, and area (R 2 : 0.76–0.84). Notably, relative leaf shape traits showed positive associations with oil yield per plant and protein yield per plant, supporting their potential as complementary indicators for screening soybean germplasm with differential industrial product output. This end-to-end framework establishes a reliable bridge between UAV remote sensing and leaf-level morphological quantification, advancing high-throughput phenotyping capabilities to support precision breeding in soybean.

Plant phenotyping relevance

UAV画像、葉検出・セグメンテーション・回帰モデルを統合し、圃場でダイズ葉形態を自動定量するフレームワークの開発と性能評価が研究の中心であるため。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.