PEDNet: Prior-Guided Multielliptical Frequency Masking and Dual-Domain Fusion for Remote Sensing Semantic Segmentation

Semantic segmentation of high-resolution remote sensing imagery (RSIs) remains challenging due to the coexistence of direction-sensitive structures, complex land-cover patterns, and insufficient class discriminability. Existing methods often emphasize either spatial-domain local details or global contextual aggregation, while paying limited attention to the anisotropic frequency characteristics induced by real ground objects, often resulting in the suboptimal learned representations. To address this issue, we propose a prior-guided multielliptical frequency masking and dual-domain fusion network (PEDNet), a unified prior-guided anisotropic spatial-frequency modeling framework for remote sensing semantic segmentation (RSSS). The core idea of PEDNet is to align frequency modeling with the directional geometry and semantic heterogeneity of land-cover objects, rather than relying on fixed isotropic decompositions. Specifically, a multielliptical frequency decomposition module (MEFDM) is introduced to perform anisotropy-aware frequency partitioning in the Fourier domain using learnable elliptical masks, thereby enhancing anisotropic spectral modeling and boundary-detail perception. In addition, a contextualized dual-attention fusion module (CDAFM) and a prior-guided mask adaptation module (PGMAM) are developed to realize coordinated spatial-frequency interaction and scene-adaptive frequency modeling, where CDAFM strengthens cross-domain feature integration through contextualized bidirectional attention and gated fusion, and PGMAM dynamically refines frequency partitioning under semantic prior guidance. Extensive experiments on multiple benchmark datasets demonstrate that PEDNet consistently achieves superior segmentation performance, especially in challenging scenarios involving fine boundaries, anisotropic structures, and densely distributed small objects.