Precise leaf damage detection across diverse species and environments via a large-scale vision model

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate crop leaf damage detection is crucial for plant health monitoring and yield prediction.
• Traditional deep learning models struggle with generalization across different crop species and field conditions.

Purpose of the Study:

• To propose a novel deep learning paradigm for robust crop leaf lesion segmentation.
• To adapt foundation models for efficient and accurate plant disease detection in agriculture.

Main Methods:

• Integration of the DinoV3 foundation model with the Unet framework.
• Incorporation of a Spatial Prior Module (SPM) and a Projection Module for domain adaptation.
• Experimental validation on coffee and black gram leaf datasets, and the larger AMG_HS dataset.

Main Results:

• The proposed model achieved superior performance (e.g., 78.31% IoU on coffee leaves) compared to standard networks like Unet.
• Demonstrated significant reduction in inference time (93.6%) for high-parameter foundation models.
• Confirmed robustness and scalability on larger datasets and cross-domain reliability.

Conclusions:

• Foundation model adaptation offers a scalable and high-performance solution for intelligent crop protection.
• Coupling customized encoders with foundation models is a superior strategy for cross-domain agricultural tasks.
• The approach enables computationally efficient, real-time plant health monitoring.