Abstract:

Accurately extracting structurally complete road networks from remote sensing imagery is crucial in numerous fields, such as autonomous driving. However, the linear, slender structure of roads is highly coupled with complex and heterogeneous background textures in the spatial domain, making it difficult for existing methods to effectively capture long-range dependencies while suppressing noise, resulting in fragmented and structurally incomplete road networks.To address this, we propose FSC-Net, a network that collaboratively enhances features in both the frequency and spatial domains, achieving precise road network extraction through a global context branch and a local detail branch. In the global branch, an Adaptive Frequency-Domain Attention Module (AFAM) is designed to effectively extract the elongated structural features of roads and suppress background-related noise through frequency-domain learning. The refined structural features are then fed into a modified Mamba module to capture reliable global structural dependencies specific to roads with linear complexity. The detail branch, composed of residual convolutional blocks, is used to extract local detail features of roads. Finally, a Bidirectional Guidance Fusion Module (BGFM) aggregates the complementary features from the two branches, ensuring the generated road network possesses both structural integrity and edge precision.A series of experiments on the Massachusetts and DeepGlobe datasets validate the effectiveness of FSC-Net. Compared with other representative methods, FSC-Net can more robustly capture the long-range dependencies of road networks, thereby significantly improving the continuity and completeness of the extracted road network.

Key words: road extraction, remote sensing imagery, frequency domain enhancement, structure-aware, long-range dependency