A lightweight co-optimization model for field sunflower disease identification.

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning

Background:

• Accurate crop disease identification is vital for crop quality and yield.
• Existing deep learning models are often too large and lack robustness for field deployment.
• There is a need for lightweight, robust models for on-site crop disease detection.

Purpose of the Study:

• To develop a lightweight and robust deep learning model for crop disease identification.
• To enable practical field application of crop disease identification models on resource-constrained devices.

Main Methods:

• Proposed a lightweight YOLO-CGA model based on YOLOv8n-cls.
• Incorporated a CBAM_ADown module with attention and asymmetric downsampling.
• Replaced C2f with C3Ghost module using ghost convolution for reduced parameters.
• Developed an AFC_SPPF module for multi-scale feature fusion.

Main Results:

• Achieved high accuracy: 98.48% (BARI-Sunflower), 98.32% (Cotton Disease), 91.11% (FGVC8).
• Maintained a lightweight design with only 0.92M parameters.
• Successfully deployed the model on a Raspberry Pi for field application.

Conclusions:

• The YOLO-CGA model effectively balances accuracy, robustness, and lightweight performance.
• Enables real-time, on-site crop disease identification in challenging field conditions.
• Addresses the gap between laboratory models and practical field applications for agriculture.