Edge-Guided Spatial and Frequency Synergistic Perception Network for Remote Sensing Image Semantic Segmentation

Semantic segmentation of high-resolution remote sensing imagery plays a fundamental role in land-cover mapping and urban analysis. Unlike natural images, remote sensing scenes often exhibit pronounced cross-scale structural heterogeneity: roads usually appear as elongated and topologically continuous targets, building regions present compact yet complex geometric layouts, and agricultural areas tend to form large contiguous pat terns. However, existing segmentation networks mainly empha size pixel-level recognition and generic multi-scale fusion, which often leads to structural fragmentation, boundary discontinuity, and semantic-detail misalignment in remote sensing imagery. To address these issues, we propose an Edge-Guided Spatial and Frequency Synergistic Perception Network (ESFSPNet), a structure-aware segmentation framework designed to preserve geospatial continuity and boundary integrity in complex remote sensing scenes. Specifically, we introduce a Spatial-Frequency Structural Interaction (SFSI) module to jointly model local structural details and global morphological patterns from spatial and frequency perspectives. Within SFSI, the Spatial Multi Scale Kernel Modulation (SMKM) branch adaptively captures heterogeneous local structures, while the Frequency-Domain Enhancement (FDE) branch strengthens long-range structural dependencies and global consistency. Furthermore, an Edge Guided Cross-Attention Fusion (ECAF) module is developed in the decoder to align high-level semantics with low-level structural cues during upsampling, thereby alleviating object breakage and boundary drift. Extensive experiments on the ISPRS Vaihingen, ISPRS Potsdam, and LoveDA datasets show that ESFSPNet achieves more competitive performance than existing methods, particularly in maintaining structural integrity and boundary consistency of remote sensing targets.