Data-centric annotation analysis for plant disease detection: Strategy, consistency, and performance.

Area of Science:

• Precision Agriculture
• Computer Vision
• Plant Pathology

Background:

• Object detection models are key for plant disease detection.
• Current research focuses on network architecture and loss functions.
• Data annotation quality and cost are critical but under-investigated factors.

Purpose of the Study:

• To investigate the relationship between data annotation strategies, quality, and model performance.
• To propose a systematic strategy for data annotation in plant disease detection.
• To analyze the impact of annotation inconsistencies on model accuracy.

Main Methods:

• Proposed four annotation strategies: local, semi-global, global, and symptom-adaptive.
• Conducted an interpretability study using class activation maps.
• Defined and investigated five types of annotation inconsistencies and their impact.

Main Results:

• Different annotation strategies yield distinct and remarkable differences in model performance.
• Annotation inconsistencies significantly impact model performance.
• Label inconsistency issues during data augmentation were discussed.

Conclusions:

• A data-centric approach emphasizing annotation strategies is crucial for optimizing plant disease detection models.
• Understanding annotation impact can lead to higher performance and reduced costs.
• Further research should focus on annotation consistency and strategy selection.