Domain adaptation for remote sensing image semantic segmentation with prototype-driven domain disentangle alignment

Existing remote sensing domain adaptation (DA) methods for semantic segmentation typically achieve domain feature alignment by learning domain-invariant features. However, these methods may inadvertently retain domain-specific features within the extracted representations, resulting in insufficient alignment of domain semantic feature information. To address this issue, we propose a prototype-driven domain disentangled alignment (PDDA) framework to mitigate the impact of domain-specific information from extracted features on the misalignment of semantic feature information. Specifically, we design a domain feature disentangle with prototypes (DFDP) method to automatically disentangle domain-invariant and domain-specific representations by explicitly modeling domain-invariant and domain-specific prototypes, respectively. Then, we construct a domain-specific guided feature alignment (DSFA) strategy, which leverages domain-specific representations as complementary guidance to improve the semantic alignment of domain-invariant features. In addition, we propose a domain-invariant induced self-learning (DISL) mechanism, which exploits pseudo-label learning guided by domain-invariant features to further strengthen cross-domain semantic feature alignment. Extensive experiments on both ISPRS (Potsdam and Vaihingen) and LoveDA (Rural and Urban) datasets robustly attest to the preeminence of the proposed PDDA over other advanced DA methods. The code will be available at https://github.com/XZhang878/PDDA.