SARShipRES: A Scattering-Aware Dataset and Structure-Regularized Framework for Referring SAR Ship Segmentation

Accurately localizing and segmenting user-specified ships in synthetic aperture radar (SAR) imagery is essential for maritime surveillance, traffic monitoring, and interactive ocean observation. Referring expression segmentation (RES) provides a flexible language-guided paradigm for target selection; however, existing RES methods are mainly developed for natural images and are difficult to directly transfer to SAR scenarios due to the scattering-dominated imaging mechanism, speckle noise, cluttered nearshore backgrounds, and ambiguous object boundaries. Moreover, most existing RES datasets assume a one-to-one correspondence between an expression and a single target, which is inconsistent with practical maritime interpretation, where a user may refer to multiple ships sharing similar spatial or scattering properties. To address these issues, we present SARShipRES, a SAR-specific referring ship segmentation dataset that integrates spatial, area, semantic, and scattering-aware descriptions and supports both one-to-one and one-to-many image-text-mask annotations. Based on this dataset, we propose a SAM-based multimodal segmentation framework tailored to SAR imaging characteristics. The proposed method introduces a decoder-level scatter-suppressed topological mask refinement strategy to reduce non-structural bright scatterers, enhance region connectivity, and alleviate fragmented predictions. In addition, an adaptive boundary-aware shape regularization scheme is designed to improve boundary smoothness and geometric consistency while avoiding over-regularization of small targets. Extensive experiments under both inshore and offshore conditions demonstrate that the proposed framework consistently outperforms representative RES baselines.