Weakly Supervised Crop Area Segmentation for an Autonomous Combine Harvester

Area of Science:

• Agricultural engineering
• Computer vision
• Machine learning

Background:

• Deep learning for machine vision offers effective object detection but is hindered by limited labeled agricultural datasets.
• Automatic visual perception is crucial for agricultural navigation systems.

Purpose of the Study:

• To propose a weakly supervised crop area segmentation (WSCAS) method for efficient identification of uncut crop areas.
• To enable automatic path guidance in agriculture with reduced annotation effort.

Main Methods:

• Developed a weakly supervised learning approach using area-specific images for implicit localization.
• Trained a classification model that segments target areas based on learned localization.
• Utilized post-image processing (Canny edge detector, Hough transformation) for edge detection.

Main Results:

• Achieved an intersection over union (IoU) of approximately 0.94 for crop area localization.
• Demonstrated the lowest inference time compared to previous deep learning segmentation methods.
• Successfully detected uncut crop edges for practical application in navigation.

Conclusions:

• The WSCAS method provides real-time crop area inference and localization comparable to existing semantic segmentation techniques.
• This approach is suitable for automatic path guidance systems in agricultural machinery, like combine harvesters.
• Weakly supervised learning offers an efficient solution for agricultural machine vision tasks with limited data.