Global Wheat Head Detection 2021: An Improved Dataset for Benchmarking Wheat Head Detection Methods.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• The Global Wheat Head Detection (GWHD) dataset was established in 2020, comprising 193,634 labeled wheat heads from 4700 RGB images across 7 countries.
• The GWHD_2020 dataset garnered significant attention from computer vision and agricultural science communities, highlighted by a Kaggle competition.
• Key areas for improvement identified were dataset size, head diversity, and label accuracy.

Purpose of the Study:

• To enhance the Global Wheat Head Detection dataset by increasing its size, diversity, and label reliability.
• To provide a more robust dataset for training and evaluating wheat head detection models.
• To support advancements in automated crop monitoring and yield prediction.

Main Methods:

• Re-examination and relabeling of the original GWHD_2020 dataset.
• Augmentation of the dataset with 1722 new images from 5 additional countries.
• Inclusion of 81,553 additional wheat head instances.

Main Results:

• The updated Global Wheat Head Detection dataset (2021) is larger and more comprehensive than the 2020 version.
• Increased diversity in wheat head appearance and acquisition conditions.
• Improved label reliability through re-examination and relabeling processes.

Conclusions:

• The enhanced GWHD dataset offers a more valuable resource for research in agricultural computer vision.
• The improvements facilitate more accurate and reliable wheat head detection models.
• This dataset will aid in developing advanced tools for precision agriculture and crop management.