Semi-supervised Learning for Weed and Crop Segmentation Using UAV Imagery.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Weed control is crucial for crop yield and food production.
• Accurate crop and weed mapping is essential for automated weed management systems.
• Existing methods often require large amounts of labeled data, limiting their practical application.

Purpose of the Study:

• To propose SemiWeedNet, a novel weed and crop segmentation method.
• To accurately identify weeds of varying sizes in complex environments.
• To reduce the dependency on large labeled datasets through semi-supervised learning.

Main Methods:

• Developed a unified semi-supervised architecture based on semantic segmentation.
• Integrated a multiscale enhancement module with selective kernel attention for feature highlighting.
• Employed online hard example mining (OHEM) to refine training on difficult-to-distinguish pixels.
• Utilized consistency regularization to leverage unlabeled data and enhance robustness.

Main Results:

• SemiWeedNet demonstrated superior performance compared to state-of-the-art methods on a public dataset.
• The method effectively segments weeds in complex environments with varying sizes.
• Individual components of SemiWeedNet showed promising potential for segmentation improvements.

Conclusions:

• SemiWeedNet offers an effective solution for accurate weed and crop segmentation.
• The semi-supervised approach significantly reduces the need for labeled data.
• The proposed method contributes to the advancement of automatic weed management systems.