Optimized Lightweight U-Net and YOLACT framework for multi-disease severity detection in pome fruit leaves

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Global food demand and plant diseases necessitate automated crop health monitoring.
• Pome fruits (apples, pears) are susceptible to diseases impacting yield and quality.
• Current disease detection methods rely on manual inspection and struggle with complex imagery and coexisting diseases.

Purpose of the Study:

• To develop a novel deep learning framework for multi-disease severity detection and classification in pome fruit leaves.
• To address limitations of existing methods in handling complex imagery and multiple coexisting diseases.
• To introduce a new severity scale for quantifying multiple infections on a single leaf.

Main Methods:

• A dual-model deep learning framework utilizing a fine-tuned MobileNetV2 backbone for feature extraction.
• Integration of Lite-U-Net for semantic segmentation and enhanced Lite-YOLACT for instance segmentation.
• Development of a new multi-disease severity scale and application of improved Grad-CAM for interpretability.

Main Results:

• The proposed framework achieved 95% accuracy in disease severity estimation.
• Successfully identified and graded multiple coexisting infections on single pome fruit leaves.
• The system provides interpretable visual heatmaps for expert validation.

Conclusions:

• The research offers an efficient, interpretable, and scalable deep learning solution for pome fruit crop health monitoring.
• Represents a significant advancement in precision agriculture for disease management.
• Publicly available code and models facilitate further research and application.