Class-Domain Incremental Segmentation for Remote Sensing Images

Significant progress has been made in class-incremental (learning new classes without forgetting old ones) and domain-incremental (adapting to data from different distributions) semantic segmentation for remote sensing images. However, in real-world deployment, class-space changes and distribution shifts may co-occur between old and new data. Existing incremental learning methods typically address only one type of shift, struggling to handle joint class and domain incremental learning. To achieve class-domain incremental segmentation, we propose CDISeg, a novel framework that enables cross-domain knowledge accumulation through feature synthesis. CDISeg employs a temporary style encoder while repurposing the segmentation model’s backbone as the content encoder. By enforcing orthogonality between their outputs, the model disentangles image content from style, thereby preserving domainspecific style features throughout incremental learning steps. The framework synthesizes old-domain features by projecting old-domain styles onto new-domain content, which supports the preservation of old knowledge, extends new classes to past domains, and facilitates the learning of old classes over new-domain images. Additionally, we introduce class-aware style randomization to enhance feature disentanglement and improve synthesis quality. Extensive experiments on ISPRS and Open-EarthMap datasets demonstrate the remarkable superiority of CDISeg in enabling models to progressively acquire new classes from new domains while recognizing all learned classes across all encountered domains.