Pest detection in dynamic environments: an adaptive continual test-time domain adaptation strategy

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Pest management is crucial for agricultural production and food security, with early detection enabling timely intervention.
• Traditional object detection models struggle with real-world, non-i.i.d. data, leading to reduced accuracy in diverse environments.
• Existing methods often fail to adapt to the variability encountered in different agricultural settings.

Purpose of the Study:

• To develop an accurate and robust pest detection method for diverse agricultural environments.
• To address the challenge of domain shift in pest detection datasets.
• To improve the performance of object detection models in unseen environments through test-time adaptation.

Main Methods:

• Proposed the CrossDomain-PestDetect (CDPD) method, built upon the YOLOv9 object detection model.
• Incorporated a test-time adaptation (TTA) framework featuring Dynamic Data Augmentation (DynamicDA) and a Dynamic Adaptive Gate (DAG).
• Integrated object detection with image segmentation in a Multi-Task Dynamic Adaptation Model (MT-DAM) with adaptive feature fusion and self-supervised learning for TTA.

Main Results:

• Without TTA, CDPD showed a 7.6% mAP50 increase in the original environment and a 16.1% increase in the target environment over the baseline.
• With TTA, the mAP50 score reached 73.8% in an unseen target environment, significantly outperforming the baseline.
• The method demonstrated improved detection capabilities through adaptive feature fusion and self-supervised adaptation during testing.

Conclusions:

• The CDPD method effectively enhances pest detection accuracy in diverse and unseen environments.
• Test-time adaptation is a viable strategy to improve the robustness of object detection models in agriculture.
• The proposed approach offers a significant advancement for automated pest management and crop protection systems.