Answers are grounded in this paper and cite their sources.
Referring Remote Sensing Image Segmentation (RRSIS) aims to predict the segmentation masks of corresponding ground objects based on text descriptions. Achieving accurate cross-modal alignment between text descriptions and visual content is crucial for this task. However, most existing methods rely heavily on feature-level cross-attention mechanisms based on the Transformer architecture. This approach is often inefficient and prone to information redundancy, which in turn interferes with the alignment process; moreover, due to the lack of domain prior for semantic understanding of ground objects, it remains challenging to achieve precise cross-modal alignment. To address the above issues, this paper proposes a cross-modal hierarchical alignment network (CMHA-Net) for RRSIS, which is guided by the selective state space model (SSM) and domain prior. Specifically, CMHA-Net adopts MambaVision to replace the Transformer architecture as the backbone for visual feature extraction and uses BERT to encode text. The proposed semantic selective alignment module (SSAM) leverages the domain prior provided by the pre-trained remote sensing vision foundation model (RS-VFM) to guide the selection of cross-modal features and the SSM for multi-directional weighted visual scanning, dynamically screening the visual regions with the highest semantic relevance, thereby achieving efficient cross-modal alignment. Furthermore, the prior-guided fine-grained alignment module (PFAM) leverages domain prior to facilitate dense cross-modal interaction and employs the SSM with a larger receptive field, thereby enabling fine-grained cross-modal alignment. In addition, we propose a prior-based cross-modal semantic consistency (CMSC) loss function to explicitly constrain the alignment relationship between images and text during the training process. Experiments conducted on the RefSegRS and RRSIS-D datasets demonstrate that our method achieves state-of-the-art performance.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: