Frequency–Spatial Dual-Path Modeling for Remote Sensing Image Semantic Segmentation
Abstract
Semantic segmentation of remote sensing images (RSIs) faces persistent challenges in high-frequency information degradation, boundary ambiguity, and structural inconsistency. To address these issues, we present a frequency–spatial dual-path modeling network (FSDPNet) paradigm that achieves holistic structural enhancement and fine-grained detail preservation. Specifically, we introduce a frequency-reduction Mamba mechanism (FRMM), which explicitly disentangles and adaptively retains critical high-frequency and directional representations during encoding. In the decoding stage, a multiscale frequency-driven reconstruction module (MFRM) leverages frequency-domain enhancement and dilated convolutions for progressive resolution recovery and precise texture compensation. Furthermore, a segment anything model (SAM)-constrained semantic consistency fusion module (SCFM) harnesses structural priors to reinforce boundary delineation and suppress semantic ambiguity across scales. The overall training is guided by a composite objective that jointly optimizes global semantic consistency and local structural fidelity. Extensive experiments on multiple remote sensing benchmarks demonstrate that our approach consistently outperforms state-of-the-art segmentation methods, setting a new standard for structural integrity and fine-detail preservation in RSI segmentation. All codes and segmentation results will be available at: https://github.com/darkseid-arch/FSDPNet