Abstract: The significant performance degradation of change detection models, caused by distribution shifts due to different regions and sensors in satellite remote sensing images, was investigated. This study aims to enhance the model's adaptability to unlabeled target domains in unsupervised scenarios.A novel cross-domain change detection network, Swin-CDCD, which integrates Swin Transformer and adversarial feature alignment, was proposed. First, a siamese Swin Transformer backbone was constructed as the feature extraction network, leveraging its hierarchical window attention mechanism for powerful long-range dependency modeling to significantly improve the representation of change features in bi-temporal remote sensing images. Furthermore, a self-attention-based discriminator was designed to meticulously capture and align inter-domain feature distribution discrepancies through global dependency modeling. Lastly, a dual-space feature alignment (DFA) strategy was introduced, which innovatively conducted joint adversarial learning on both the holistic temporal features and the difference features, effectively enhancing the model's sensitivity to genuine changes and reducing false alarms and missed detections caused by domain shift. Extensive bidirectional cross-domain experiments on the LEVIR-CD and WHU-CD datasets demonstrated that the proposed method significantly outperformed various existing unsupervised domain adaptation baselines. Specifically, it achieved F1 scores of 82.71% and 79.04%, and Intersection over Union (IoU) values of 70.52% and 65.35% on the L2W (LEVIR-CD→WHU-CD) and W2L (WHU-CD→LEVIR-CD) tasks, respectively, confirming its superior detection accuracy and generalization capability in complex cross-domain scenarios.

Key words: remote sensing, change detection, unsupervised domain adaptation, Transformer, adversarial training