A CNN-Transformer Hybrid Framework for Multi-Label Predator-Prey Detection in Agricultural Fields

Area of Science:

• Agricultural Entomology
• Computer Vision
• Machine Learning

Background:

• Accurate identification of predator-pest relationships is crucial for effective biological control in agriculture.
• Current image-based methods face challenges in recognizing insect co-occurrence under complex field conditions, limiting their ecological applicability.

Purpose of the Study:

• To develop a robust deep learning framework for accurate multi-label recognition of predator-pest combinations.
• To improve the ecological applicability and accuracy of image-based insect identification in agricultural settings.

Main Methods:

• Proposed a hybrid deep learning framework integrating Convolutional Neural Networks (CNNs) and Transformer architectures.
• Introduced a novel co-occurrence attention mechanism to capture semantic relationships between insect categories.
• Employed a pairwise label matching loss function to enhance ecological pairing accuracy.

Main Results:

• Achieved an F1-score of 86.5% and mAP50 of 85.1% on a dataset of 5,037 images.
• Demonstrated strong generalization to unseen predator-pest pairs with an average F1-score of 79.6%.
• Outperformed established baselines such as ResNet-50, YOLOv8, and Vision Transformer.

Conclusions:

• The developed model offers a robust and interpretable approach for multi-object ecological detection.
• The framework shows practical potential for deployment in smart farming, UAV-based monitoring, and precision pest management systems.