Key points are not available for this paper at this time.
Current state-of-the-art change detection methods face a critical dilemma: achieving robustness against non-stationary pseudo-changes (e.g., seasonal shifts and vegetation phenology) while preserving fine-grained boundary details often requires computationally expensive models. Existing lightweight approaches usually sacrifice one aspect for the other, leading to either excessive false alarms or blurred object boundaries. To address this challenge, we propose a novel coarse-to-fine refinement network, termed CFR-Net. CFR-Net adopts a dual-time Mamba-based encoder to suppress non-stationary variations and capture robust spatiotemporal dependencies, followed by a cascaded dual-decoder architecture that progressively refines change boundaries from coarse localization to fine delineation. Extensive experiments on three benchmarks (LEVIR-CD, GZ-CD, and SYSU-CD) demonstrate that CFR-Net consistently outperforms state-of-the-art methods. Specifically, on the LEVIR-CD dataset, our model achieves an F1-score of 97.46% and requires only 6.54 G FLOPs, striking an optimal balance between accuracy and efficiency. These results confirm the effectiveness of CFR-Net for robust and efficient high-resolution remote sensing change detection. Our code is available at https://github.com/wyyyz-web/CFR-Net.git
Wang et al. (Thu,) studied this question.