Performance evaluation of semi-supervised learning frameworks for multi-class weed detection.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Precision weed management (PWM) leverages machine vision and deep learning (DL) for sustainable agriculture.
• Current DL weed detection relies on supervised learning, requiring extensive manual data labeling.
• Label-efficient methods, particularly semi-supervised learning, are emerging as a solution to reduce data annotation efforts.

Purpose of the Study:

• To evaluate a semi-supervised learning framework for multi-class weed detection.
• To assess the effectiveness of the student-teacher framework with improved pseudo-labeling and ensemble students.
• To demonstrate high-performance weed detection with minimal labeled data.

Main Methods:

• Implemented a generalized student-teacher framework for semi-supervised learning.
• Utilized an improved pseudo-label generation module for unlabeled data.
• Employed an ensemble student network to enhance model generalization.
• Tested the framework on FCOS and Faster-RCNN object detection architectures.

Main Results:

• Achieved approximately 76% detection accuracy on CottonWeedDet3 with 10% labeled data.
• Reached approximately 96% detection accuracy on CottonWeedDet12 with 10% labeled data.
• Demonstrated performance comparable to supervised methods using significantly less labeled data.

Conclusions:

• Semi-supervised learning offers a viable and efficient approach for weed detection in precision agriculture.
• The proposed framework effectively reduces the dependency on large labeled datasets.
• This research provides a valuable resource for advancing semi-supervised learning in agricultural applications.