RFSNet: Rotational Frequency-Aware Structural Network for Road Segmentation via Frequency–Geometry Collaboration

Road segmentation from high-resolution remote sensing imagery remains a formidable challenge, primarily due to the thin and elongated topology, arbitrary orientations, and complex background clutter. These factors often result in fragmented extractions and boundary ambiguity. To address orientation diversity and clutter-induced spectral interference, we present a rotational frequency-aware structural network (RFSNet) that couples geometry-aware representation learning with frequency-aware refinement for robust road extraction. In the encoder, a lightweight rotated depthwise convolution (R-DWConv) aggregates multidirectional responses and fuses them with angle-aware attention, producing orientation-sensitive features consistent with curvilinear road geometry. At the bottleneck, we propose a multidimensional frequency enhancement module (MFEM) with radial band attention, where the full-spectrum representation is embedded as a query and band-limited spectra are used as keys/values to form a competitive weighting mechanism, enabling content-adaptive spectral energy redistribution for denoising while preserving structure-relevant components to ensure road continuity. Furthermore, the decoder is driven by a frequency-domain recurrent linear attention module (FRLAM) that refines fused encoder–decoder features through gated recurrent linear attention (RLA) and frequency-domain residual refinement, thereby efficiently capturing long-range dependencies under high-resolution inputs. Extensive experiments on multiple public road benchmarks demonstrate that RFSNet consistently outperforms the state-of-the-art (SOTA) methods while maintaining favorable efficiency, validating the effectiveness of the proposed frequency–geometry collaborative design. The source code of the proposed method is available athttps://github.com/Nuist-HSI-Group/RFSNet