Continual Test-Time Adaptive Object Detection in Remote Sensing Images via Spectral Routing and Historical Reconstruction

Remote sensing object detection (RSOD) remains challenging due to significant domain shifts caused by variations in sensors, geographic regions, illumination, weather, and object scales, which severely degrade model performance when deployed on unseen target domains. While test-time adaptation (TTA) enables model updating using unlabeled test data, existing methods typically assume relatively stable target distributions and struggle under the dynamic and sequential conditions of remote sensing scenarios. Continual test-time adaptation for object detection (CTAOD) further extends TTA to streaming test data, but current approaches often rely on global updates with fixed criteria, resulting in noisy feature propagation, error accumulation, and catastrophic forgetting. In this paper, we propose CTAOD-RS, a novel framework for continual test-time adaptive object detection in remote sensing imagery. The proposed framework leverages Spectral Routing Adaptation to decompose features into frequency and spatial components, suppressing noise and clutter, and employs History-Informed Reconstruction to dynamically determine when model adaptation is necessary, avoiding unnecessary updates and mitigating catastrophic forgetting. Extensive experiments on benchmark RSOD datasets, including the newly introduced DIOR-C corrupted benchmark designed to simulate realistic remote sensing domain shifts, demonstrate superior detection accuracy and robustness compared to existing TTA and CTAOD methods, while maintaining high efficiency. The code will be released at https://github.com/rosou1/CTAOD-RS.