EMSAM: enhanced multi-scale segment anything model for leaf disease segmentation.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Accurate leaf disease segmentation is vital for crop health management but challenged by blurred boundaries and complex features.
• Existing vision foundation models like Segment Anything Model (SAM) show limitations in plant disease image segmentation.
• There is a need for advanced models to achieve fine-grained segmentation of plant leaf diseases.

Purpose of the Study:

• To propose an advanced model, Enhanced Multi-Scale SAM (EMSAM), for fine-grained segmentation of leaf disease images.
• To enhance the capability of handling blurred boundaries and complex feature distributions in plant disease images.
• To improve both segmentation and classification accuracy for plant leaf diseases.

Main Methods:

• Developed Enhanced Multi-Scale SAM (EMSAM) incorporating Local Feature Extraction Module (LFEM) and Global Feature Extraction Module (GFEM).
• Utilized multiple convolutional layers in LFEM for detailed lesion characteristics and fine-tuned ViT blocks with Multi-Scale Adaptive Adapter (MAA) in GFEM for global information.
• Implemented a Feature Fusion Module (FFM) with attention mechanisms and a joint loss function for segmentation and classification.

Main Results:

• EMSAM achieved superior performance on the PlantVillage dataset, outperforming state-of-the-art models.
• EMSAM surpassed the second-best model by 2.45% in Dice Coefficient and 6.91% in IoU score.
• The model demonstrated high Dice Coefficients (0.8354 for moderate, 0.8178 for severe diseases) and 87.86% classification accuracy.

Conclusions:

• EMSAM effectively addresses challenges in plant disease segmentation, particularly blurred boundaries and complex features.
• The proposed model shows significant improvements in both segmentation and classification tasks for leaf diseases.
• EMSAM offers a superior solution for automated plant disease detection and management.