SwinGAN-CD: An End-to-End Cross-Domain Change Detection Framework for Global Feature Modeling and Multi-Dimensional Feature Alignment

In remote sensing change detection tasks, significant cross-domain data distribution discrepancies severely weaken model adaptability, especially in scenarios where the target domain lacks labels. Existing methods struggle to effectively address domain shifts under such unsupervised conditions. To tackle this, this paper proposes SwinGAN-CD, an unsupervised cross-domain change detection method that uses the Swin Transformer as the feature extraction backbone combined with adversarial networks. By introducing the long-range dependency modeling capability of the Swin Transformer as the encoder and constructing a joint discriminator to align multi-dimensional feature spaces between the source and target domains, the method effectively enhances the model’s cross-domain recognition ability for change regions. Furthermore, a Self-Attention based Discriminator (SA-D) is proposed to enable more thorough and efficient feature alignment during adversarial training. Additionally, a Dual-space Feature Alignment (DFA) strategy tailored for change detection tasks is introduced to specifically align global features and change features across pre- and post-event phases, thereby improving detection accuracy. Cross-domain experiments conducted on the WHU-CD and LEVIR-CD datasets demonstrate that the proposed method achieves higher detection accuracy and stronger robustness under unsupervised conditions compared to existing approaches, attaining F1 scores of 82.71% and 79.04%, respectively, both of which are state-of-the-art. This study provides an effective unsupervised solution for cross-domain change detection in remote sensing.