FSC-Net: Synergizing Frequency Domain Analysis and Mamba for Structure-Aware Road Extraction Network

Accurately extracting structurally complete road networks from remote sensing images has significant application value in numerous fields such as autonomous driving. However, roads exhibit slender, linear geometries, and their features are highly coupled in the spatial domain with surrounding complex and heterogeneous background textures. Existing methods struggle to effectively capture long-range dependencies while suppressing noise, resulting in extracted road networks that commonly suffer from fragmentation and structural incompleteness. To address this issue, this paper proposes FSC-Net, a road extraction network that integrates frequency-domain and spatial-domain features through collaborative modeling. The network achieves high-precision road extraction via the cooperative operation of a global context branch and a local detail branch. Within the global branch, an Adaptive Frequency Attention Module (AFAM) was designed to fully exploit the slender structural features of roads and suppress complex background noise through frequency-domain feature learning while suppressing complex background noise. The structural features purified by the AFAM were then fed into an improved Mamba module, which can reliably model the global structural dependencies unique to roads with linear computational complexity. The local detail branch, composed of stacked multi-layer residual convolutional blocks, is responsible for extracting local detail features such as road edges and contours. Finally, a Bidirectional Guided Fusion Module (BGFM) was employed to fuse the complementary features of the two branches, ensuring that the output road segmentation results possess both complete topological structures and precise edge contours. Extensive experiments conducted on the public Massachusetts and DeepGlobe datasets verify the effectiveness of the FSC-Net model. Compared with other mainstream representative algorithms, FSC-Net can more robustly capture the long-range dependency features of roads, significantly enhancing the continuity and completeness of the road network extraction results.