AgriPest-YOLO: A rapid light-trap agricultural pest detection method based on deep learning

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning for pest control

Background:

• Manual pest identification is labor-intensive and time-consuming.
• Automated pest monitoring using light-trap images faces challenges due to scale variation, complex backgrounds, and dense pest distribution.
• Existing vision technologies struggle with rapid and accurate detection in these conditions.

Purpose of the Study:

• To develop a lightweight pest detection model (AgriPest-YOLO) that balances efficiency, accuracy, and model size.
• To improve the accuracy and speed of automated pest detection in agricultural light-trap images.
• To address the challenges of scale variation, complex backgrounds, and dense pest distribution.

Main Methods:

• Proposed a coordination and local attention (CLA) mechanism to enhance pest feature extraction and reduce noise interference.
• Introduced a grouping spatial pyramid pooling fast (GSPPF) module to enrich multi-scale pest feature representation.
• Implemented soft-Non-Maximum Suppression (NMS) in the prediction layer to optimize detection of overlapping pests.

Main Results:

• AgriPest-YOLO achieved 71.3% mean Average Precision (mAP) on a large-scale dataset with 24 pest classes and 25,000 images.
• The model outperformed classical and lightweight pest detection models in accuracy and speed.
• Demonstrated a superior balance between model size, detection speed, and accuracy.

Conclusions:

• AgriPest-YOLO offers an accurate and efficient solution for real-time, multi-class pest detection from light-trap images.
• The model is a key component for advancing pest forecasting and intelligent pest monitoring technologies.
• The proposed CLA mechanism and GSPPF module effectively address challenges in light-trap image pest detection.