A Multiscale Vision–Text Collaborative Dual Encoder for Referring RS Image Segmentation

Referring Remote Sensing Image Segmentation (RRSIS) aims to segment target objects in remote sensing images using natural language descriptions. Existing methods employing a single backbone and sequential fusion struggle to capture fine-grained semantics in remote sensing data due to their complex and multi-scale nature. Moreover, vision models pretrained on natural images often fail to generalize to remote sensing images due to domain-specific spatial and semantic discrepancies. To address these issues, we propose an innovative Multiscale Vision-Text Collaborative Dual-Encoder network, named MCD-Net. We first introduce a frozen SAM encoder as a structure-aware auxiliary branch to inject general-purpose spatial priors, enhancing the Swin Transformer’s ability to model fine-grained geometry and object-level information. To improve semantic consistency across scales and better align with referring expressions, we propose a Multi-scale Frequency-aware Alignment module that decomposes visual features into frequency components and modulates them via cross-scale textual attention. An Adaptive Deep Fusion module bridges the representation gap between dual visual branches while preserving spatial-semantic coherence. Extensive experiments on public RRSIS benchmarks demonstrate that our method outperforms state-of-the-art approaches, particularly in scenes with dense objects and ambiguous expressions. Our code will be released upon publication.