EP-YOLO: An Enhanced Lightweight Model for Micro-Pest Detection in Agricultural Light-Trap Environments.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Automated pest monitoring is vital for agricultural early warning systems and food security.
• Existing models struggle with tiny pests and complex background interference, leading to detection errors.
• Challenges include small pest scale, occlusions, and reflections in light-trap sensors.

Purpose of the Study:

• To propose EP-YOLO, an enhanced lightweight detection architecture based on YOLOv8n.
• To improve the accuracy of automated pest detection, especially for micro-targets.
• To overcome physical limitations in detecting extremely small pests in agricultural settings.

Main Methods:

• Developed EP-YOLO, an enhanced lightweight detection architecture.
• Introduced Spatial-to-Depth Convolution (SPD) module to retain spatial pixels of micro-targets.
• Integrated Efficient Multi-Scale Attention (EMA) module to isolate pest features and reduce background noise.
• Evaluated the model on the Pest24 dataset comprising 24 tiny pest categories.

Main Results:

• EP-YOLO achieved mAP@50 of 70.5% and mAP@50:95 of 47.3%, outperforming the baseline.
• Demonstrated significant improvements in detecting specific tiny pests like Rice planthopper (8.4%) and Plutella xylostella (3.1%).
• The model successfully addressed challenges posed by small scale and background interference.

Conclusions:

• EP-YOLO effectively overcomes physical limitations in detecting tiny pests.
• The proposed architecture provides a robust and deployable solution for real-time agricultural monitoring.
• Enhanced pest detection capabilities contribute to improved agricultural early warning systems and food security.