Referring Remote Sensing Image Segmentation With Positive-Incentive Visual Adaptation and Logical Reasoning

Referring Remote Sensing Image Segmentation (RRSIS) aims to segment target regions in aerial imagery according to natural language descriptions. Despite recent advances in vision-language foundation models, their application to RRSIS remains challenging due to severe domain shifts, cluttered back grounds, and complex spatial layouts. Existing methods mainly suffer from two limitations: insufficient cross-modal alignment caused by natural-image training bias, and limited reasoning ability when handling complex expressions involving multiple entities and spatial constraints. To address these issues, we propose a reasoning-enhanced decoupled framework for RRSIS with two task-oriented modules: a Positive-Incentive Cross-modal Alignment Module (PICAM) and a Referring Logic Reasoning Module (RLRM). PICAM introduces beneficial input-conditional noise into the visual representation, while using textual cues as conditioning signals, thereby alleviating the domain gap between remote sensing imagery and natural-image-pretrained vision language models and improving the robustness and discriminability of multimodal features. Built upon the alignment-enhanced features, RLRM explicitly introduces large-model reasoning into the localization stage by decomposing referring expressions into target semantics, attributes, anchor objects, and spatial relations, and converting the reasoning process into reliable dense and sparse prompts for downstream segmentation. Based on these two modules, our framework follows a clear decoupled paradigm: alignment, reasoning-guided localization, and prompt based mask generation. This design enables the model to resolve semantic ambiguity before precise boundary delineation, leading to improved interpretability and better adaptation to complex remote sensing scenarios.