TeaWeeding-Action: a vision-based dataset for weeding behavior recognition in tea plantations.

Area of Science:

• Computer Vision
• Agricultural Technology
• Robotics

Background:

• Weed infestations pose significant challenges to tea plantations, impacting crop yields and global food security.
• Intelligent weeding behavior recognition systems are crucial for developing automated agricultural solutions.
• Existing datasets may lack the specificity and diversity required for robust tea plantation weeding analysis.

Purpose of the Study:

• To introduce a novel, publicly available computer vision dataset for analyzing weeding behaviors in tea plantations.
• To facilitate the advancement of intelligent weeding behavior recognition systems.
• To support the development of precision agriculture technologies.

Main Methods:

• Collected 108 HD video sequences and 6,473 annotated images from real tea plantation environments.
• Employed a hybrid data acquisition approach combining field recordings and web-crawled resources.
• Utilized a multi-view acquisition strategy (frontal, lateral, top-down) for 3D understanding and provided annotations in COCO and YOLO formats.

Main Results:

• The dataset encompasses six categories of weeding behaviors, including manual, tool-assisted, and machine-based methods.
• Benchmark evaluations using YOLOv8, SSD, and Faster R-CNN demonstrated the dataset's effectiveness.
• Faster R-CNN achieved a mean Average Precision (mAP) of 82.24% on the dataset.

Conclusions:

• The proposed dataset provides a valuable foundation for developing intelligent weeding robots and precision agriculture monitoring systems.
• This resource will accelerate computer vision applications in complex agricultural settings.
• The dataset contributes to addressing weed management challenges and enhancing agricultural efficiency.