Semi-supervised learning methods for weed detection in turf.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Accurate weed detection is crucial for precision herbicide application.
• Traditional methods rely on extensive manual data labeling for deep learning models.
• This study introduces a semi-supervised learning (SSL) approach for weed detection in turf.

Purpose of the Study:

• To develop and evaluate a novel SSL approach for weed detection in turfgrass.
• To compare the performance of SSL methods against a fully supervised learning (FSL) method (ResNet50).
• To assess the efficiency and accuracy of SSL in terms of labeled data requirements.

Main Methods:

• Implemented and evaluated three SSL models: Π-model, Mean Teacher, and FixMatch.
• Compared SSL performance against ResNet50 using weed and turfgrass image datasets.
• Assessed model performance using classification accuracy and F1 scores with varying amounts of labeled data.

Main Results:

• SSL methods, particularly FixMatch, demonstrated increased classification accuracy compared to ResNet50, even with limited labeled data (100 images/class).
• FixMatch achieved the highest accuracy (≥0.9530) and F1 scores (≥0.951) using only 50 labeled images per class.
• SSL models outperformed FSL in accuracy and efficiency, requiring fewer labeled training images.

Conclusions:

• SSL deep neural networks offer a highly accurate and efficient alternative to FSL for weed detection.
• The proposed SSL approach reduces the time and labor associated with data labeling for precision agriculture.
• This advancement supports more effective and sustainable weed management strategies.