FDMamba: frequency-enhanced deformable Mamba for topology-aware road extraction

Extracting topological road networks from high-resolution remote sensing imagery is a fundamental yet challenging task, often hindered by slender geometries and complex occlusions. Existing methods based on CNNs or Transformers frequently struggle to balance global context modeling with local connectivity preservation under stringent computational constraints. In this paper, we propose FDMamba, a dual-domain framework that effectively integrates frequency analysis with the State Space Model (SSM). To capture holistic road structures, we first introduce Frequency Mamba (FreqMamba) equipped with a novel Spiral Scan (SS) strategy, which projects features into the frequency domain to efficiently extract global contextual priors. These frequency-domain priors then guide Deformable Mamba (DMamba), a spatial-domain module that discards rigid scanning paths in favor of data-driven, adaptive sampling to precisely fit arbitrary road curves and complex intersections. To further refine boundary details, a plug-and-play Frequency Attention (FA) module is designed to fuse multi-frequency amplitude information across feature stages. Comprehensive evaluations on four benchmark datasets (DeepGlobe, CHN6-CUG, SpaceNet3, and CityScale) validate the superiority of FDMamba. The proposed model achieves significant improvements, surpassing SegFormer by 3.63% Intersection over Union (IoU) on DeepGlobe. Crucially, when utilized as a backbone for topological extractors, it improves Average Path Length Similarity (APLS) by up to 2.81%, effectively mitigating network fragmentation. This study not only provides a robust, efficient, and topology-aware paradigm for high-resolution road extraction, but also opens new avenues for coupling frequency-domain global perception with spatial-domain geometric adaptability under scalable computational complexity.