SAR image segmentation based on cross-domain transfer of frequency-spatial priors

Synthetic Aperture Radar (SAR) is widely used in Earth observation due to its all-weather imaging capability. However, its inherent speckle noise, strong background clutter, and lack of intuitive geometric texture make high-precision semantic segmentation in complex scattering scenes highly challenging. Existing deep learning methods mostly rely on isotropic convolutions, which are insufficient for effectively disentangling frequency-domain noise from spatial structures. Cross-domain auxiliary segmentation schemes based on optical priors are often limited by their dependence on strictly aligned optical-SAR paired data, which restricts their applicability. To address these issues, this paper proposes a Semantic Induction Dual-domain Anisotropic Reconstruction with Generative Adversarial Networks (SIDAR-GAN). The proposed method breaks the dependence on strictly aligned data and designs a cross-domain consistency learning mechanism based on a Generative Adversarial Network (GAN). By constructing a Semantic Induction GAN (SI-GAN) and a Dual-Domain Structural Alignment (DDSA) module, implicit optical geometric and texture priors are extracted from the generator encoder and adaptively injected into the SAR feature space through a cross-domain attention mechanism, enabling robust semantic induction under unregistered conditions. Furthermore, a Frequency-Spatial Disturbance Parsing (FSDP) module is designed. Its frequency-domain branch employs multi-scale wavelet decomposition to isolate and adaptively suppress noise, while its spatial anisotropic branch introduces shape-biased convolution kernels to enhance the feature representation of targets such as ships and buildings along their principal directions, thereby jointly modeling structural and noise characteristics in both frequency and spatial domains. Experiments on the SARBuD building dataset and the HRSID ship dataset demonstrate that the proposed method achieves Dice coefficients of 77.66% and 92.28%, respectively. Compared with 18 mainstream models, it significantly improves noise suppression capability and boundary extraction accuracy, providing a new cross-modal semantic-driven paradigm for fine-grained SAR interpretation under weak-texture conditions.