A Frequency-Spatial Segmentation Network for High-Resolution Remote Sensing Images

Semantic segmentation of high-resolution remote sensing images faces three major challenges in frequency-spatial feature fusion: background clutter mixed into high-frequency components, semantic discontinuities within large homogeneous regions, and loss of fine rigid boundaries caused by convolutional downsampling. Traditional dual-stream networks rely on complex attention mechanisms and directly concatenate heterogeneous features, which often leads to representation conflicts, information loss, and excessive edge smoothing. To address these issues, we propose FSOD-Net, a Frequency-Spatial Feature Decoupling Network built upon the SFFNet architecture. The feature mapping stage is reconstructed to incorporate physical edge priors and frequency-domain attention-based denoising. Specifically, a Spatial Attention Wavelet Transform Feature Decomposer (SA-WTFD) is introduced in the frequency branch to adaptively suppress high-frequency background clutter and enhance discriminative features. In the spatial global branch, a lightweight pyramid pooling module (L-PPM) captures macro-contextual information to fill semantic gaps within areal objects. In the local branch, a fixed Laplacian operator is embedded as an explicit physical edge prior to preserving micro-scale rigid boundaries during downsampling. These three modules operate along independent paths, enabling decoupled yet collaborative multi-domain feature learning. Extensive experiments on the ISPRS Vaihingen and Potsdam datasets demonstrate that FSOD-Net achieves mean Intersection over Union (mIoU) scores of 84.24±0.11% and 86.79±0.09%, respectively, indicating its potential in segmentation accuracy and generalization.