Noise-Resilient With Scattering-Aware Network for SAR Image Semantic Segmentation

Synthetic Aperture Radar (SAR) image semantic segmentation faces significant challenges, including severe speckle noise, intricate land cover patterns, weak feature background contrast, geometric distortions, and limited texture information. These factors obscure object boundaries and hinder accurate feature extraction. To tackle these challenges, we propose a Noise-Resilient with Scattering-Aware Network (NRSANet) for SAR image semantic segmentation. Specifically, the Scattering-Aware Dynamic Attention Module (SDAM) adaptively highlights key scattering regions to enhance feature discrimination in complex environments, while the Adaptive Noise-Aware Boundary Diffusion Module (NBDM) suppresses speckle noise and sharpens boundary clarity. The Adaptive Vision Enhanced Outlooker (AVEO) aggregates global and local features across multiple scales, enabling accurate segmentation of objects with diverse shapes and mitigating geometric distortion. Additionally, the Bidirectional Coordinate State Attention (BCSA) captures spatial correlations and enforces spatial consistency, facilitating the distinction of adjacent or overlapping features where texture cues are limited. The integration of these complementary components enables NRSANet to robustly address the challenges of SAR image segmentation in a unified framework. Extensive experiments on benchmark SAR datasets demonstrate that NRSANet consistently outperforms state-of-the-art methods, achieving more accurate and robust segmentation, especially in complex and noisy scenarios. The source code and dataset will be available on https://github.com/darkseid-arch/NRSANet.