Source-linked AI summary
Global Wheat Head Detection (GWHD) dataset: a large and diverse dataset of high resolution RGB labelled images to develop and benchmark wheat head detection methods
E. David, S. Madec, P. Sadeghi-Tehran, H. Aasen, B. Zheng, S. Liu, N. Kirchgessner, G. Ishikawa, K. Nagasawa, M. A. Badhon, C. Pozniak, B. de Solan, A. Hund, S. C. Chapman, F. Baret, I. Stavness, W. Guo
TL;DR
Wheat head detection matters for estimating density and head characteristics, but existing methods are generally trained and validated on limited datasets challenged by variation in genotypes and observation conditions. The paper develops the harmonized GWHD dataset from international contributions, standardizing image scales, patches, and labels. It reports 4,698 patches and 188,445 labelled heads, and proposes guidelines for extending the dataset and benchmarking detection methods.
Problem
Existing wheat head detection methods are generally calibrated and validated on limited datasets, while genotype, developmental, orientation, blur, and overlap variation complicate reliable computer vision.
Method
The paper harmonizes wheat head datasets from multiple institutions by standardizing image resolution and patches, applying consistent bounding-box labelling, and reviewing annotations collaboratively.
Results
4,698 squared patches and 188,445 labelled heads comprise the GWHD dataset, assembled from 2,219 original high-resolution RGB images across 11 sub-datasets.
Takeaways & Limitations
GWHD provides a large, diverse benchmark for wheat head detection and offers acquisition, metadata, and labelling guidelines for future dataset expansion.
Takeaways & Limitations
A 10 cm canopy-height estimation error can induce an 8% sampled-area error, directly transferring to head-density measurement.
Abstract
from arXiv · showhide
Detection of wheat heads is an important task allowing to estimate pertinent traits including head population density and head characteristics such as sanitary state, size, maturity stage and the presence of awns. Several studies developed methods for wheat head detection from high-resolution RGB imagery. They are based on computer vision and machine learning and are generally calibrated and validated on limited datasets. However, variability in observational conditions, genotypic differences, development stages, head orientation represents a challenge in computer vision. Further, possible blurring due to motion or wind and overlap between heads for dense populations make this task even more complex. Through a joint international collaborative effort, we have built a large, diverse and well-labelled dataset, the Global Wheat Head detection (GWHD) dataset. It contains 4,700 high-resolution RGB images and 190,000 labelled wheat heads collected from several countries around the world at different growth stages with a wide range of genotypes. Guidelines for image acquisition, associating minimum metadata to respect FAIR principles and consistent head labelling methods are proposed when developing new head detection datasets. The GWHD is publicly available at http://www.global-wheat.com/ and aimed at developing and benchmarking methods for wheat head detection.
1. Introduction
Wheat head density is an important yield-related trait, but extracting it from imagery remains difficult to scale reliably across diverse field conditions. The GWHD dataset addresses limited, institution-specific training data by harmonizing contributions from nine institutions across seven countries and three continents for benchmarking.
- Wheat head density, the number of heads per unit ground area, is a major yield component still manually evaluated in breeding trials.
- Deep learning methods for wheat head quantification have been demonstrated on high-resolution RGB imagery, but prior evaluations used individual datasets.
- Limited variation in genotypes, geographic areas, and observational conditions makes extrapolation of trained models to new situations difficult.
- Large, diverse shared datasets are needed to compare methods because existing datasets are independently acquired, sparsely labelled, and governed by differing protocols.
- The GWHD dataset harmonizes datasets from nine institutions across seven countries and three continents to support wheat head detection benchmarking.
2. Dataset composition
GWHD combines sub-datasets collected across diverse locations, genotypes, growing practices, cameras, and acquisition configurations. Harmonization rescaled images, divided them into standardized patches, and applied collaborative labelling and review procedures.
- Experiments: GWHD combines sub-datasets collected from nine institutions at ten locations between 2016 and 2019, covering genotypes from Europe, North America, Australia, and Asia.
- Image acquisition: Acquisition conditions span heights of 1.8–3 m, focal lengths of 10–50 mm, and Ground Sampling Distances of 0.10–0.62 mm.
- Data harmonization: Images acquired too early were removed when heads were unclear, while acquisition generally preceded senescence because bending stems increase head overlap.
- Data harmonization: Images were rescaled with bilinear interpolation to make head-level resolution more similar across sub-datasets.
- Data harmonization: Images were split into 1024 x 1024 patches containing roughly 20 to 60 heads, with few heads crossing patch edges.
- Labelling: Bounding-box labels were produced through a shared web platform, with initial labelling and repeated independent review involving approximately 20 contributors.
3. Description of the dataset
The GWHD dataset combines thousands of harmonized RGB image patches and labelled wheat heads from diverse sub-datasets, acquisition conditions, and object appearances. It is positioned as a large open dataset for field plant-phenotyping object detection.
- Dataset scale: 4,698 squared patches contain 188,445 labelled wheat heads from 2,219 original RGB images across 11 sub-datasets.Images average 40 heads, with about 100 head-free images and a maximum of 120 heads per image.
- Dataset scale: The dataset combines images from multiple sub-datasets and acquisition sites, with bounding-box distributions varying in head density and object scale.Bounding-box typical dimension has a median of 77 pixels, while density differences produce multiple distribution peaks.
- Diversity: Images span variable illumination, backgrounds, viewing directions, head orientations, awns, colours, blurriness, and bounding-box shapes.A few heads were cut off when bounding boxes crossed image edges.
- Comparison with other datasets: GWHD is described as the largest freely available open labelled object-detection dataset for field plant phenotyping.The comparison notes that MinneApple has comparable diversity but fewer images and less location diversity, whereas general-purpose datasets cover more object types.
4. Target use case: Wheat Head Detection challenge
The GWHD challenge is designed to benchmark wheat head detection under unseen genotypes, environments, and observational conditions. It uses a geographically separated training setup and object-detection evaluation based on bounding-box overlap.
- Challenge purpose: The challenge benchmarks wheat head detection methods using GWHD training and testing data in an open machine-learning competition.The announced competition was scheduled from May to August 2020.
- Challenge design: Training uses 3,422 images from Europe and North America, representing 73% of the dataset, while testing targets unseen genotypes, environments, and observational conditions.The split is designed to evaluate generalization beyond the training acquisition settings.
- Evaluation metrics: Localization is evaluated with mAP@0.5, where a predicted box is a true positive when its IoU with the closest labelled box is at least 0.5.Among multiple qualifying predictions for one labelled box, the most confident prediction is counted as the true positive.
- Baseline: The baseline uses Faster-RCNN with ResNet34 or ResNet50 backbones and 512x512-pixel inputs.ResNet34 is included because it is less prone to overfitting and faster to train, while input size is constrained by memory.
5. Discussion
The discussion emphasizes acquisition, metadata, and geographic expansion requirements for making GWHD useful for reliable wheat head detection and density estimation. It also identifies unresolved sampling-area and image-resolution issues and proposes future dataset extensions.
- Image acquisition recommendations: Near-nadir imaging, limited overlap, and suitable camera height are recommended to improve visibility and reduce cropped heads.A 2.5 m camera-to-canopy distance is proposed to increase sampled area and reduce edge cropping.
- Image acquisition recommendations: 10 cm of canopy-height estimation error induces an 8% sampled-area error that directly transfers to head-density measurement.Fisheye distortion and non-parallel acquisition geometry further complicate sampled-area estimation.
- Image acquisition recommendations: 21% sampled-area variation separates the lowest and highest heads because reference height remains unresolved.The discussion calls for further investigation of the reference height used to compute sampled area.
- Image acquisition recommendations: Sub-millimetre resolution is suggested for efficient head detection, but the optimal resolution and UAV applicability remain uncertain.The GWHD dataset spans 0.28–0.55 mm GSD, while previous work recommended 0.3 mm GSD.
- Metadata and FAIR principles: FAIR-oriented minimum metadata are recommended because missing metadata hindered precise harmonization, interpretation, and meta-analysis.The metadata specification is presented in Table 4, and camera-to-canopy distance is identified as essential for harmonization and density calculation.
- Need for GWHD expansion: GWHD diversity is geographically uneven: over 68% of images come from Europe and 43% from France, with Africa, Latin America, and the Middle East absent.Future contributions should follow the proposed acquisition and metadata guidelines to preserve consistency and FAIR principles.
- Need for GWHD expansion: Future expansion is planned for wheat-head classification and segmentation, including traits such as head size and flowering state.This extension would require updating the current labels.
6. Conclusion
The paper assembles GWHD to improve wheat head detection and localization through a diverse international dataset, standardized acquisition and labelling guidance, and an open research competition. The dataset is intended to support more accurate and robust methods across wheat-growing regions.
- 6. Conclusion: GWHD is an extensive, diverse wheat head image dataset assembled to develop and benchmark detection methods.It is presented as a collaborative international effort to improve detection and localization accuracy and reliability.
- 6. Conclusion: The contribution includes image-acquisition guidelines, minimum FAIR metadata, and wheat-head labelling guidelines and tools.These resources are intended to support future addition of sub-datasets with broader genotypic and environmental diversity.
- 6. Conclusion: An open research competition accompanies GWHD to identify more accurate and robust wheat-head detection methods across global growing regions.Competition solutions will be made open-source and shared with the plant phenotyping community.