Unsupervised semantic label generation in agricultural fields.

Area of Science:

• Robotics and Computer Vision
• Agricultural Technology
• Artificial Intelligence

Background:

• Current agricultural robots rely on deep learning for weed and crop recognition, requiring extensive manual data labeling.
• Manual data labeling is time-consuming, costly, and demands specialized domain expertise, limiting the scalability of these systems.
• There is a need for automated solutions to streamline the training data creation process for agricultural AI.

Purpose of the Study:

• To develop an automated labeling pipeline for crop-weed semantic image segmentation in agricultural fields.
• To enable the training of deep learning models with minimal or no manual image labeling.
• To improve the efficiency and accuracy of perception systems in farm robots.

Main Methods:

• Utilizes RGB images from aerial or ground robots and exploits field row structures for spatially consistent labeling.
• Incorporates an "unknown" class for challenging vegetation to reduce labeling errors and improve consistency.
• Employs evidential deep learning to leverage prediction uncertainty estimates for refining semantic segmentation, particularly for underrepresented classes like weeds.

Main Results:

• The automated pipeline significantly outperforms general-purpose and domain-specific labeling methods in crop-weed segmentation.
• Training models with generated labels improves performance on unseen fields, crop species, growth stages, and lighting conditions.
• Achieved an 88.6% IoU for crops and 22.7% for weeds in sugarbeet fields, surpassing existing methods.

Conclusions:

• The proposed automated labeling pipeline effectively reduces the reliance on manual data annotation for training agricultural AI models.
• The system demonstrates robust performance across diverse field conditions, enhancing the adaptability of farm robots.
• This approach offers a scalable solution for developing advanced perception systems in precision agriculture, improving environmental sustainability.