Enhancing Agricultural Image Segmentation with an Agricultural Segment Anything Model Adapter

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• The Segment Anything Model (SAM) offers versatile zero-shot image segmentation.
• SAM exhibits limitations in specialized agricultural applications like crop disease and pest identification.

Purpose of the Study:

• To develop an Agricultural SAM adapter (ASA) to improve SAM's performance in agricultural image segmentation.
• To enable effective zero-shot segmentation for agricultural tasks using domain-specific expertise.

Main Methods:

• Incorporated agricultural domain knowledge into SAM using an adapter technique.
• Leveraged distinctive agricultural image characteristics and user prompts for segmentation.
• Conducted comprehensive experiments to evaluate ASA against the default SAM.

Main Results:

• The ASA demonstrated significant improvements across all 12 agricultural segmentation tasks evaluated.
• Achieved a notable 41.48% average Dice score improvement on coffee leaf disease segmentation tasks.
• Showcased enhanced zero-shot segmentation capabilities in the agricultural domain.

Conclusions:

• The ASA effectively enhances SAM for agricultural image segmentation.
• The proposed adapter technique offers a novel and effective approach for zero-sample agricultural image analysis.
• ASA provides a valuable tool for improving crop monitoring and disease management.