Harmonized spatial-frequency domain synergy driven geospatial feature synthesis for enhanced SAR semantic segmentation

Abstract

The phase coherence of radar signals makes synthetic aperture radar (SAR) image analysis prone to significant challenges. Echo signal interference introduces speckle noise during imaging; noise appears as random fluctuations in pixel intensities. Besides, coherence exacerbates geometric distortions, complicating the accurate interpretation of spatial distributions within intricate geographic entities, thereby making it difficult to extract meaningful target information from SAR images. Addressing these challenges, this study introduces the DEcomposed-frequency PrOjection Network (Depo-Net), a segmentation-oriented model that mitigates SAR-specific interference through frequency-domain self-attention. It employs a dual-encoder structure for efficient semantic extraction and integrates Spatio-Frequency Synergistic Modulation (SFSM) to minimize speckle noise while maintaining structural integrity in the frequency domain. Additionally, the Harmonized Subspace Spectro-Temporal Attention (HSSTA) synthesizes Discrete Fourier and Wavelet Transform analyses to capture complex spatial correlations among geographic features. To mitigate noise amplification during decoding, the Pluri-frequency Mamba (purfMamba) module synergizes multi-dimensional spectral-spatial features, facilitating noise suppression during high-resolution restoration and maintaining a balance between global structure and local details. Results on three public SAR segmentation datasets demonstrate Depo-Net’s efficacy outperforming 22 previous State-of-the-Art (SOTA) methods while minimizing 95th Percentile Hausdorff Distance values. The complete code and model implementation is available on GitHub at https://github.com/IMOP-lab/Depo-Net.