VCINet: Visual Cue-Inspired Feature Learning Network for SAR Building Segmentation

Abstract

The extraction of building areas from synthetic aperture radar (SAR) imagery is of great importance for urbanization assessment and emergency response. However, SAR images are generated by active radar systems and exhibit characteristics, such as strong scattering effects, speckle noise, and high-contrast edges. General deep learning methods often struggle to effectively represent consistent building features due to insufficient exploitation of the characteristics and visual cues. In this article, we propose a visual cue-inspired feature learning network named VCINet, which considers both the interior and boundary regions of building areas to refine segmentation results. First, to improve intraclass variance and emphasize internal consistency, an elaborately designed pattern guidance module is introduced. We systematically study the stable spatial pattern of buildings and leverage the variogram function to model the characteristic to constrain deep network training. Second, to enhance subtle details and preserve semantic integrity, a boundary-aware detail-context fusion module is developed. The boundary characteristic of sharp contrast changes is observed and is utilized to generate boundary attention scores, guiding the adaptive aggregation of high-frequency detailed and low-frequency contextual information. Finally, extensive experiments show the effectiveness of VCINet compared to state-of-the-art methods. Our method achieves the best results on all three datasets, with mFscore of 89.66% on SARBud, 83.79% on AIR-PolSAR-Seg, and 84.15% on DFC23.

Keywords

• Building segmentation
• deep learning
• synthetic aperture radar (SAR)
• visual cues