Multi-format open-source weed image dataset for real-time weed identification in precision agriculture

Area of Science:

• Agricultural Science
• Computer Vision
• Robotics

Background:

• Weed management is crucial for crop yield, often relying on robotic systems for targeted herbicide application.
• Effective robotic weed control necessitates sophisticated computer vision algorithms for accurate in-field identification.
• Existing public weed datasets are limited, primarily collected from ground-based perspectives.

Purpose of the Study:

• To introduce a novel, comprehensive agricultural weed dataset.
• To support the development of advanced computer vision algorithms for both ground-based and aerial weeding robots.
• To facilitate advancements in precision agriculture and sustainable weed management practices.

Main Methods:

• Collected a dataset of 3,975 images featuring five common North Dakota weed species.
• Acquired imagery using both handheld cameras and unmanned aerial systems (UAS).
• Annotated images in multiple formats and applied augmentation techniques to simulate real-world conditions.

Main Results:

• Developed a unique dataset capturing weed imagery from both ground and aerial viewpoints.
• The dataset includes five key weed species: kochia, common ragweed, horseweed, redroot pigweed, and waterhemp.
• Images are meticulously annotated and augmented for robust computer vision model training.

Conclusions:

• The open-source dataset is vital for advancing computer vision in precision weeding technologies.
• Enables real-time in-field weed identification for targeted herbicide spot spraying.
• Contributes to more efficient and sustainable agricultural practices through improved robotic weed control.