YOLO-Citrus: a lightweight and efficient model for citrus leaf disease detection in complex agricultural environments

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Citrus leaf disease detection is vital for global production quality and yield.
• Existing methods struggle with occlusion, small lesions, complex backgrounds, high computational costs, and large model sizes.
• These limitations hinder accurate and efficient object detection in agricultural settings.

Purpose of the Study:

• To develop a novel, lightweight, and efficient model for citrus leaf disease detection.
• To enhance detection precision while minimizing computational expenses and model size for practical agricultural use.
• To address limitations of existing methods in handling occlusion, small lesions, and complex backgrounds.

Main Methods:

• Introduction of the YOLO-Citrus model, a YOLOv11-based lightweight architecture.
• Incorporation of three key innovations: C3K2-STA module with Star-Triplet Attention, ADown module for downsampling, and Wise-Inner-MPDIoU loss function.
• Optimization focused on enhancing bottleneck performance, efficient downsampling, and improved bounding box regression.

Main Results:

• YOLO-Citrus achieved 96.6% mAP@0.5 (1.4% improvement over baseline) and 81.6% mAP@0.5:0.95 (1.3% improvement).
• Model size was reduced by 25.0% (from 19.2 MB to 14.4 MB).
• Computational cost decreased by 20.2% (from 21.3 GFlops to 17.0 GFlops), demonstrating significant efficiency gains.

Conclusions:

• YOLO-Citrus offers a highly accurate and efficient solution for citrus leaf disease detection.
• The model's reduced size and computational requirements make it suitable for real-world orchard deployment.
• Validated effectiveness in orchard conditions for early disease detection, precision treatment, and yield protection.