Frequency-aware dual-domain network for remote sensing optical image change detection

Change detection (CD) in remote sensing aims to identify variations in the same geographical area over multiple periods, providing critical data support and a decision-making basis for diverse scientific fields and real-world applications. Traditional deep learning methods for CD typically extract features from bi-temporal images and fuse them to produce change maps. However, simple feature concatenation or differencing often fails to capture complex change patterns, particularly in cases involving subtle or semantic-level changes. To address these limitations, we propose the frequency-aware dual-domain network (FADDNet), an end-to-end space–frequency joint framework for optical CD. FADDNet incorporates frequency information to mitigate the detail loss caused by downsampling. The dual-domain encoder explicitly models spectral differences to recover high-frequency features, whereas the semantic align fuse module employs spatial attention for progressive fusion, effectively transferring informative spatial features and suppressing semantic inconsistencies. By integrating complementary information from multiple domains, frequencies, and scales, FADDNet forms a coherent system that enhances CD accuracy. Experiments on the change detection dataset and learning vision and remote sensing change detection demonstrate that FADDNet consistently outperforms state-of-the-art baselines, achieving superior performance in both precision and recall.