Detection of Pine Wilt Nematode from Drone Images Using UAV

Area of Science:

• Forest Pathology and Entomology
• Remote Sensing and Geospatial Analysis
• Computer Vision and Deep Learning

Background:

• Pine wilt nematode disease poses a significant threat to forest ecosystems, necessitating rapid and effective monitoring strategies.
• Traditional monitoring methods are often labor-intensive and time-consuming, limiting the scope and speed of disease control efforts.
• Drone-based remote sensing offers a promising approach for timely detection and management of forest diseases like pine wilt nematode.

Purpose of the Study:

• To develop and evaluate an optimized deep learning algorithm for the automatic identification of pine wilt nematode-diseased trees using UAV remote sensing imagery.
• To improve the detection speed and efficiency of existing algorithms, such as YOLOv4, by incorporating lightweight network structures and attention mechanisms.
• To assess the performance of the proposed algorithm against established object detection models in terms of accuracy, speed, and model size.

Main Methods:

• The study employed the YOLOv4 object detection algorithm as a baseline and optimized its backbone feature extraction network using the lightweight MobileNetv2.
• Further enhancements included integrating the CBAM attention module and the Inceptionv2 structure to reduce model parameters and boost identification performance.
• Comparative analysis was conducted against Faster R-CNN, YOLOv4, SSD, and YOLOv5 algorithms using metrics such as average precision, training time, parameter size, and test time.

Main Results:

• The improved MobileNetv2-YOLOv4 algorithm achieved an average precision of 86.85% and an F1 score of 95.60%, outperforming Faster R-CNN, YOLOv4, and SSD.
• The optimized model demonstrated a significantly reduced parameter size (39.23 MB) and a fast single-image test time of 15 ms, comparable to YOLOv5.
• The algorithm exhibited a balanced performance across key indicators, making it more suitable for practical applications, especially on embedded devices for rapid detection.

Conclusions:

• The enhanced MobileNetv2-YOLOv4 algorithm provides an efficient and accurate solution for detecting pine wilt nematode diseased trees from drone imagery.
• The optimized model's reduced complexity and improved performance make it a practical tool for real-time monitoring and management of forest diseases.
• This approach facilitates timely intervention, contributing to more effective control strategies against the rapid spread of pine wilt nematode disease.