A knowledge-driven unified framework for plant disease classification and severity grading via domain adaptation

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate plant disease classification and severity grading are crucial for precision agriculture.
• Current models struggle with novel disease categories and limited plant phenotype representation.

Purpose of the Study:

• To develop a knowledge-driven unified framework for robust plant disease classification and severity grading.
• To enable flexible label space expansion and improve generalization to unseen disease categories.

Main Methods:

• A Common Knowledge Learner consolidates plant and disease features using multi-level contrastive learning.
• Cross-Domain Adaptation with Low-Rank Adaptation (LoRA) bridges in-domain and out-of-domain feature spaces.
• A contrastive feature library facilitates similarity-based reasoning and flexible label expansion.

Main Results:

• Achieved 94.9% disease classification and 90.6% severity grading accuracy in-domain on the Leaf-CG dataset.
• Demonstrated 82.1% true positive rate (TPR) in open-set, out-of-domain conditions.
• Outperformed existing baselines, showcasing strong generalization capabilities.

Conclusions:

• The proposed framework offers a significant advancement in plant disease management through improved classification and grading.
• The model's ability to handle novel diseases and adapt across domains holds great potential for real-world agricultural applications.