Bridge: Leveraging Vision Foundation Models for Efficient Cross-Domain Remote Sensing Segmentation

Adapting large-scale Vision Foundation Models (VFMs), pretrained on natural images (NI), to remote sensing (RS) downstream tasks such as Domain Adaptation Semantic Segmentation (DASS) though Parameter-Efficient Fine-Tuning (PEFT) is a promising yet challenging direction. However, due to the large domain gap between NI and remote sensing images (RSI), as well as the inherent intra-domain variations in RS DASS task, existing PEFT methods struggle to achieve satisfactory adaptation. To address these challenges, we propose Bridge, a framework with small trainable parameters specifically designed for RS DASS. Bridge introduces a set of learnable Domain- Invariant Tokens that capture transferable knowledge shared across NI and RSI domains. To complement this, a Spatial Prior Module (SPM) is proposed to extracts multi-scale spatial priors that provide auxiliary guidance for Domain-Invariant Token Adaptation (DITA), which produces Domain-Specific Tokens. Finally, Domain-Aware Injection (DAI) injects the Domain-Specific Tokens into intermediate VFM features, aligning them with RS distributions and facilitating effective DASS task adaptation. With significantly fewer trainable parameters, Bridge efficiently transfers NI-pretrained VFMs to RS DASS task. Extensive experiments demonstrate that Bridge substantially outperforms state-of-the-art PEFT approaches. Remarkably, compared with the current SOTA, Bridge achieves a significant improvement in mIoU. The code will be released at Bridge.