Dual-Consistency Input-Space Domain Adaptation for Remote Sensing Image Semantic Segmentation via SSC-CycleGAN
Longbao Wang, Yiding Ma, Meng Ding, Yinqi Luan, Shun Luo, Xiaoyang Meng, Yueyang Mao, Hongmin Gao · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
In remote sensing image semantic segmentation, unsupervised domain adaptation (UDA) addresses two key challenges: scarce labeled data and poor cross-domain generalization. By narrowing domain shift between labeled source and unlabeled target domains, UDA boosts model performance on target domains. However, existing UDA struggles with significant scale and style discrepancies in cross-domain remote sensing images, limiting real-world efficacy. To address this, we propose a two-stage input-space adaptation method based on a dual-consistency generative network. In the input-space alignment stage, the network integrates two dedicated modules: a structure-preserving module in the generator that retains source details and mitigates scale differences, ensuring structural consistency; and a style enhancement module in the discriminator that explicitly quantifies style disparities and strengthens style representation, thereby completing the holistic alignment. These modules enforce dual-consistency constraints, enabling the generation of images that are both structurally sound and stylistically congruent with the target domain. In the segmentation adaptation stage, a contextual consistency training strategy further improves target-domain context utilization. Experiments on Potsdam and Vaihingen datasets validate the method’s effectiveness, especially in scenarios with pronounced scale and style variations.