Tree-Guided Transformer for Sensor-Based Ecological Image Feature Extraction and Multitarget Recognition in...

Area of Science:

• Computer Vision
• Ecological Modeling
• Artificial Intelligence

Background:

• Farmland ecosystems exhibit complex pest-predator interactions, challenging current image recognition and ecological modeling.
• Sensor-driven computer vision tasks require robust methods for multitarget recognition in agricultural settings.

Purpose of the Study:

• To develop an advanced AI framework for accurate pest-predator recognition and ecological analysis in farmland.
• To enhance image-based multitarget recognition using ecological knowledge and hierarchical classification.

Main Methods:

• A tree-guided Transformer framework with a knowledge-augmented co-attention mechanism was developed.
• A hierarchical ecological taxonomy and an ecological knowledge graph were integrated for semantic reasoning.
• A multimodal dataset of 60 pest and predator categories was created for evaluation.

Main Results:

• The framework achieved high precision (90.4%), recall (86.7%), and F1-score (88.5%) in image classification.
• Detection tasks showed 91.6% precision and 86.3% mAP@50, with 80.5% co-occurrence accuracy.
• Hierarchical reasoning and knowledge-enhanced tasks yielded F1-scores of 88.5% and 89.7%, respectively.

Conclusions:

• The proposed framework effectively extracts structured, semantically aligned image features under real-world sensor conditions.
• This interpretable and generalizable approach advances intelligent agricultural monitoring and ecological understanding.