Precision Agriculture General Agricultural and Biological Sciences
Barley head detection using UAV imagery and YOLOv10
Abstract
Barley head detection is a crucial task for agricultural applications such as yield estimation and crop monitoring. Unlike wheat, automated barley head detection has not been extensively studied due to challenges posed by its complex head structures and the lack of annotated datasets. In this paper, we leverage YOLOv10, a state-of-the-art object detection framework, to detect barley heads from high-resolution images captured using UAVs. Our training dataset combines UAV-captured barley images, manually annotated images of related cereal crops (Winter Rye, Winter and Spring Wheat, Winter Triticale), and the Global Wheat Head Dataset (GWHD). On a held-out test set of 30 images, the proposed approach achieves a mean Average Precision of 0.903 at IoU threshold 0.5 (mAP50) and a head-counting Mean Absolute Error (MAE) of 2.13. An ablation study confirms the individual contribution of each data source, and we compare YOLOv10x against YOLOv5, YOLOv8, and Faster R-CNN baselines. This work contributes to advancing automated crop monitoring systems in precision agriculture. Purpose: Barley head detection is a crucial task for agricultural applications such as yield estimation and crop monitoring. Unlike wheat, automated barley head detection has not been extensively studied due to challenges posed by its complex head structures and the lack of annotated datasets. Methods: This study employs YOLOv10, a state-of-the-art object detection framework, to detect barley heads from high-resolution images captured using unmanned aerial vehicles (UAVs). The training dataset combines UAV-captured barley images, manually annotated images of related cereal crops (winter rye, winter and spring wheat, and winter triticale), and the Global Wheat Head Dataset (GWHD). An ablation study was conducted to evaluate the contribution of each data source, and YOLOv10x was compared against YOLOv5, YOLOv8, and Faster R-CNN baselines. Results: On a held-out test set of 30 images, the proposed approach achieved a mean Average Precision of 0.903 at an intersection over union threshold of 0.5 (mAP50) and a head-counting Mean Absolute Error (MAE) of 2.13. The ablation study confirmed the individual contributions of the different training data sources. Conclusion: The findings demonstrate the effectiveness of YOLOv10 for accurate barley head detection and counting using UAV imagery. Integrating annotated images from related cereal crops improves the training strategy and supports the development of automated crop monitoring systems for precision agriculture.