Lightweight grading method for potato late blight severity based on enhanced YOLOv8-Unet3Plus network

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Traditional potato late blight severity grading relies on subjective, time-consuming manual measurements.
• Existing methods lack the accuracy and cost-effectiveness required for widespread deployment.
• Objective and efficient grading is crucial for disease management and crop yield optimization.

Purpose of the Study:

• To develop a lightweight, AI-driven model for objective and accurate potato late blight severity grading.
• To enhance the YOLOv8 and UNet3Plus networks for improved performance in disease detection and segmentation.
• To reduce the computational cost and parameter count for practical, cost-effective application.

Main Methods:

• An enhanced YOLOv8 network integrated with Spatial and Channel Reconstruction Convolution, Bi-directional Feature Pyramid Network, and Powerful-IoU loss.
• An optimized UNet3Plus network utilizing Ghost convolution and Multi-Scale Local Response Attention.
• Training and validation on real-world potato late blight datasets for performance evaluation.

Main Results:

• Achieved 95.73% precision for leaf localization.
• Obtained a mean Intersection over Union (IoU) of 82.65% for infected region segmentation.
• Demonstrated reduced parameters and computational cost compared to baseline models.

Conclusions:

• The proposed AI4Science-based YOLOv8-UNet3Plus model offers an effective solution for potato late blight severity grading.
• The model provides objective, accurate, and computationally efficient disease assessment.
• This approach facilitates improved disease management strategies in agriculture.