ReLeaf-SAM: reliability-guided detail compensation for SAM-based plant disease segmentation

Area of Science:

• Precision Agriculture
• Computer Vision
• Plant Pathology

Background:

• Automated plant disease analysis requires accurate lesion segmentation for effective management in precision agriculture.
• The Segment Anything Model (SAM) shows promise but struggles with subtle details, occlusions, and complex backgrounds in field conditions.
• SAM's reliance on global cues limits its sensitivity to fine-grained textures crucial for early disease detection.

Purpose of the Study:

• To enhance SAM's capability for precise plant disease lesion segmentation in natural field environments.
• To improve the detection of small and early-stage lesions often missed by standard SAM.
• To develop a robust segmentation method addressing challenges like uneven illumination and occlusions.

Main Methods:

• Integrated a ResNet-50 branch to extract fine-grained local texture features, complementing SAM's global context.
• Developed a reliability-guided variational fusion framework to intelligently combine heterogeneous features based on confidence.
• Employed uncertainty encoding and Kullback-Leibler divergence for robust feature alignment and fusion.

Main Results:

• Achieved high Dice Similarity Coefficient (DSC) scores: 81.05% on PlantSeg, 91.12% on PlantDoc-Seg, and 88.27% on ATLDSD.
• Demonstrated superior performance over state-of-the-art methods in segmenting plant disease lesions.
• Successfully addressed SAM's limitations in capturing subtle disease textures and local details.

Conclusions:

• The proposed enhanced SAM effectively segments plant disease lesions, particularly small and early-stage ones, in complex field settings.
• The reliability-guided variational fusion significantly improves the integration of local and global features for accurate diagnosis.
• This approach offers a reliable tool for automated plant disease analysis, advancing precision agriculture applications.