FSNet: Frequency-Spatial Joint Learning for Road Extraction from Remote Sensing Images

Road extraction from remote sensing images (RSIs) plays an important role in a wide range of real-world applications. The primary challenges stem from the occlusions, as well as the high visual similarity between road and background, which complicate the identification, particularly in complex scenes. Existing methods predominantly focus on enhancing the feature representations through sophisticated designs and tend to overlook the inherent sensitivity of spatial features. To alleviate this issue, we propose a novel network that jointly explores representations in both the frequency and spatial domains, termed FSNet. This network comprises two key components: the joint-domain enhancement module (JDEM) and the cross-domain hybrid parser module (CDPM). Specifically, the JDEM utilizes Mamba within the frequency domain to capture global relationships across different spectral bands, which helps alleviate road occlusions. Accordingly, the CDPM separately parses frequency and spatial features to fully leverage the strengths of each and then effectively integrates them to improve overall performance. Experimental results on publicly available datasets demonstrate that FSNet surpasses most previous methods in intersection over union (IoU) and F1 -score, which indicate that our FSNet can generate road results with superior connectivity and accuracy. The source code has been publicly released https://github.com/jaybryant1/FSNet.