Rapid Classification and Deep Learning-Based Development Estimation of the Seeds of Helianthus annuus .

Abstract

Manually counting sunflower seeds on capitula is labor-intensive, requiring approximately one person-hour per head, and can be inconsistent for densely packed heads. Existing phenotyping approaches often depend on laboratory-based equipment, limiting their accessibility. In this study, we developed a benchtop image-based pipeline for rapid, non-destructive estimation of developed and aborted seeds on intact dried sunflower heads. A dataset of 1093 sunflower capitula was imaged under fixed indoor lighting, and individual seeds were annotated as developed or aborted. A YOLOv8m one-stage object detector was trained and evaluated using a counting-focused protocol, in which a single confidence threshold was selected on the validation set and then applied unchanged to an independent test set of 109 images. The baseline model was compared with recent YOLO variants and different augmentation strategies. On the test set, the model achieved a mean absolute count error of 61.3 seeds per image, a mean relative error of 12.0%, and an mAP50 of 0.18 at the locked confidence threshold of 0.15. Only 13.8% of test images had relative errors below 2%. Larger YOLO models and augmentation variants did not improve performance. These findings show that the proposed system provides approximate, non-destructive seed-count estimation under controlled imaging conditions, while highlighting the need for improved localization in dense regions and domain adaptation for fresh heads or field conditions. The annotated dataset and trained model weights are made available to support reproducible research.

Plant phenotyping relevance

ヒマワリ頭花の発達・不稔種子数という植物形質を、画像取得とYOLOによる推定パイプラインで定量化する手法を開発・評価しており、方法が研究の中心である。

Code and data availability

The authors state the source code is available on GitHub and the CVAT-annotated dataset is available via a public share link; the GitHub repository URL is explicitly provided and matches an allowed URL. The dataset link itself is not given, so only the code/checkpoint repository qualifies as an actionable public asset.