A Multi-Scale Vision-Sensor Collaborative Framework for Small-Target Insect Pest Management.

Area of Science:

• Agricultural Science
• Computer Vision
• Environmental Monitoring

Background:

• Small-target pests present significant challenges in agriculture due to their size, background complexity, and environmental dependence.
• Current pest management often relies on experience, hindering a transition to data-driven, precise control strategies.

Purpose of the Study:

• To develop a multi-scale vision-sensor collaborative recognition method for accurate and stable identification of small-target pests.
• To integrate pest ecological mechanisms with visual data for improved pest management in diverse agricultural settings.

Main Methods:

• A multi-scale visual feature module was designed to enhance fine-grained details of small targets in deep networks.
• Environmental sensor data (temperature, humidity, illumination) were incorporated as priors to modulate visual features.
• A vision-sensor collaborative discrimination module was implemented for multimodal fusion and pest prediction.

Main Results:

• The proposed method achieved high accuracy (93.1%), precision (92.0%), recall (91.2%), and F1-score (91.6%) on a multimodal dataset.
• Performance significantly surpassed traditional machine learning, single-scale deep learning, and vision-only baselines.
• Balanced performance was observed across various small-target pests, including aphids, thrips, and whiteflies.

Conclusions:

• The multi-scale vision-sensor method effectively addresses challenges in small-target pest recognition under complex agricultural conditions.
• Integrating environmental priors and multi-scale visual modeling is crucial for enhancing recognition accuracy and stability.
• This approach facilitates data-driven pest management, promoting precise and green agricultural practices.