Exploring the Close-Range Detection of UAV-Based Images on Pine Wilt Disease by an Improved Deep Learning Method

Area of Science:

• Forestry Science
• Computer Science
• Artificial Intelligence

Background:

• Pine wilt disease (PWD) poses a significant threat to forest ecosystems.
• Current methods for detecting PWD-infected trees lack real-time efficiency and accuracy, especially in mixed forests.
• Existing object detection models struggle to balance lightweight design with high performance.

Purpose of the Study:

• To develop a real-time and efficient object detection model for identifying PWD-infected trees.
• To improve upon existing YOLOv5s (You Only Look Once version 5s) algorithms for enhanced PWD detection.
• To provide a reliable tool for forestry management in monitoring and controlling PWD spread.

Main Methods:

• An improved YOLOv5s algorithm, named PWD-YOLO, was developed.
• A lightweight backbone using RepVGG Blocks and GSConv networks was implemented to increase inference speed and reduce complexity.
• A C2fCA module and Bidirectional Feature Pyramid Network were incorporated to enhance feature extraction and multiscale information propagation.

Main Results:

• PWD-YOLO achieved a model size of 2.7 MB, computational complexity of 3.5 GFLOPs, and parameter volume of 1.09 MB.
• The model demonstrated a high inference speed of 98.0 frames/s.
• On a self-built dataset, PWD-YOLO achieved Precision of 92.5%, Recall of 95.3%, and F1-score of 93.9%.

Conclusions:

• The PWD-YOLO model offers superior performance compared to existing object detection models for PWD detection.
• The developed model provides an effective and efficient solution for real-time monitoring and management of pine wilt disease.
• This technology offers crucial technical support for forestry departments in combating PWD.