Weed25: A deep learning dataset for weed identification.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Effective weed suppression is crucial for maximizing crop yields.
• Accurate weed species identification is essential for targeted weeding strategies, reducing crop damage and herbicide use.
• A significant limitation in applying deep learning for weed management has been the scarcity of comprehensive field datasets.

Purpose of the Study:

• To introduce the Weed25 dataset, a novel resource for training deep learning models for in-field weed identification.
• To evaluate the performance of common deep learning object detection models using the Weed25 dataset.

Main Methods:

• The Weed25 dataset was compiled, containing 14,035 images across 25 distinct weed species, including monocots and dicots at various growth stages.
• Deep learning models, specifically YOLOv3, YOLOv5, and Faster R-CNN, were trained using the Weed25 dataset for weed identification.
• Model performance was evaluated based on detection accuracy under consistent training parameters.

Main Results:

• The YOLOv3, YOLOv5, and Faster R-CNN models achieved average detection accuracies of 91.8%, 92.4%, and 92.15%, respectively.
• These results demonstrate the effectiveness of the Weed25 dataset in training robust weed identification models.
• The dataset encompasses diverse weed species and growth stages, crucial for real-world agricultural applications.

Conclusions:

• The Weed25 dataset serves as a valuable resource for advancing automated weed management systems.
• The high accuracy achieved by tested deep learning models highlights the potential for real-time, in-field weed identification.
• Further development of precision agriculture technologies can leverage this dataset for improved crop yield and reduced environmental impact.