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

Abstract

Pest management is essential for agricultural production and food security, as pests can cause significant crop losses and economic impact. Early pest detection is key to timely intervention. While object detection models perform well on various datasets, they assume i.i.d. data, which is often not the case in diverse real-world environments, leading to decreased accuracy. To solve the problem, we propose the CrossDomain-PestDetect (CDPD) method, which is based on the YOLOv9 model and incorporates a test-time adaptation (TTA) framework. CDPD includes Dynamic Data Augmentation (DynamicDA), a Dynamic Adaptive Gate (DAG), and a Multi-Task Dynamic Adaptation Model (MT-DAM). Our DynamicDA enhances images for each batch by combining strong and weak augmentations. The MT-DAM integrates an object detection model with an image segmentation model, exchanging information through feature fusion at the feature extraction layer. During testing, test-time adaptation updates both models, continuing feature fusion during forward propagation. DAG adaptively controls the degree of feature fusion to improve detection capabilities. Self-supervised learning enables the model to adapt during testing to changing environments. Experiments show that without test-time adaptation, our method achieved a 7.6% increase in mAP50 over the baseline in the original environment and a 16.1% increase in the target environment. Finally, with test-time adaptation, the mAP50 score in the unseen target environment reaches 73.8%, which is a significant improvement over the baseline.