Ambiguity-aware semi-supervised learning for leaf disease classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Semi-Supervised Learning (SSL) enhances neural network training using labeled and unlabeled data.
• SSL models generate pseudo-labels for unlabeled data, but errors can reduce accuracy.
• Leaf disease classification benefits from accurate models, but labeled data is often scarce.

Purpose of the Study:

• To develop an Ambiguity-Aware Semi-Supervised Learning method for precise leaf disease classification.
• To enhance pseudo-label quality by implementing a per-disease ambiguity rejection algorithm.
• To reduce the dependency on large, fully labeled datasets in plant pathology.

Main Methods:

• Proposed an Ambiguity-Aware Semi-Supervised Learning approach for leaf disease classification.
• Developed a per-disease ambiguity rejection algorithm to refine pseudo-labels.
• Evaluated the method on coffee and banana leaf disease datasets under various data scenarios.

Main Results:

• The ambiguity rejection algorithm significantly improved the precision of pseudo-labels.
• The semi-supervised method achieved high accuracy comparable to fully supervised models using only 50% labeled data.
• Achieved classification precisions of 99.46% for coffee and 100.0% for banana leaf diseases.

Conclusions:

• The proposed method effectively reduces the need for extensive labeled data in leaf disease classification.
• Ambiguity rejection is crucial for enhancing the performance of semi-supervised learning in this domain.
• The approach offers a viable solution for accurate plant disease identification with limited data.