Transfer learning models for wheat ear detection on multi-source dataset.

Abstract

Monitoring wheat growth, as one of the most important food grain sources for human nutrition, and forecasting yields are done through different phenological phases. Reliable estimates on yields play a crucial role in securing sufficient food supplies for the world's growing population. Currently, farmers estimate a wheat yield during the later stages of growth and are often biased in this process. Plant breeding scientists use a more accurate approach that collects data on the number of wheat ears manually counted at various locations throughout the field. A sufficiently precise count of wheat ears is one of the most important parameters for reliable early-stage prediction of wheat yield. To support the development of an affordable and trustworthy automated wheat ear detection approach, this work introduces a novel high-quality RGB smartphone image dataset, BioS-Wheat, comprising 5,696 annotated images across four wheat varieties. Additionally, it evaluates six deep learning models for wheat ear detection. Among the F-RCNN-based models, RetinaNet, YOLOv8, and a Vision Transformer-based detector, RT-DETR, achieved the highest mean Average Precision (mAP@50) of 91%, with significantly higher computational complexity. BioS-Wheat complements Global Wheat Head Detection datasets, introducing a meaningful shift in data complexity with high sowing density and minimal row spacing, emphasizing the impact of agronomic diversity on model performance by an increased object occlusion and dense spatial arrangements. Enriched and agronomically diverse datasets support model robustness at different varieties, growth stages, and locations. This work offers a good baseline for establishing the procedure for image crowdsourcing, further dataset expansions, and model improvements.

Plant phenotyping relevance

小麦穂の画像検出による個体群形質推定を対象とし、注釈付きデータセットの構築と複数モデルの評価が中心であるため、表現型計測手法として収載する。

Code and data availability

The paper's BioS-Wheat dataset and analysis code are not yet publicly accessible: the data availability statement says the code is available from the corresponding author upon request and the dataset will be deposited on Zenodo only after publication. No authors' public URL or identifier is provided for either asset.

No evidence-backed public reproduction asset is currently recorded.