Deep learning-based high-throughput phenotyping for tiller quantification in interspecific bentgrass hybrids using YOLOv8.

Abstract

Introduction: Tiller production is a critical determinant of turfgrass canopy density and plant performance, yet manual tiller counting is too labor-intensive for large breeding programs. Methods: To address this limitation, we evaluated 770 plants from an interspecific bentgrass hybrid population and developed three automated approaches for tiller quantification: a classical edge-based segmentation pipeline and two deep-learning models, Faster R-CNN and YOLOv8. Using a large annotated image dataset, we assessed each method's accuracy, robustness under occlusion, and computational efficiency. Results: Although two-stage detectors are often expected to provide superior precision for complex plant structures, the one-stage YOLOv8 model achieved the highest accuracy (R² = 0.97) and processed images substantially faster than Faster R-CNN, while both the edge-based method and Faster R-CNN showed reduced performance in dense canopies. Discussion: These findings demonstrate that recall-oriented one-stage detection can outperform more complex two-stage models for phenotyping tasks involving fine, highly occluded structures. The resulting workflow provides a reliable, high-throughput solution for generating biologically meaningful tiller counts and offers a transferable framework for integrating image-derived phenotypes into genetic analyses and breeding pipelines across grass species.

Plant phenotyping relevance

イネ科植物の分げつ数を画像から自動抽出する複数手法を開発・比較検証しており、植物表現型取得法が研究の中心である。

Code and data availability

The paper describes a 770-image annotated tiller dataset, manual counts, and YOLOv8/Faster R-CNN models, but no block contains a data availability statement, deposit, or authors' public URL for the dataset, images, annotations, trained models, or analysis code. All URLs cited are generic third-party libraries (PyTorch,

No evidence-backed public reproduction asset is currently recorded.