HFSC-Net: A Hierarchical-Feature-Guided Spatial-Frequency Dual-Domain Collaborative Network for Remote Sensing Image Semantic Segmentation

Abstract

With the rapid development of high-resolution remote sensing platforms and the growing demand for intelligent Earth observation, remote sensing image semantic segmentation has become a key technology for road extraction, urban–rural planning, disaster monitoring, and ecological assessment. However, complex remote sensing scenes commonly exhibit large-scale variation, complex background interference, intraclass variation, and interclass similarity, while boundary and texture information is prone to degradation during decoding, which continues to limit segmentation accuracy. Although convolutional neural networks effectively model local textures, their global semantic representation is limited; Transformer- and Mamba-based methods are better at capturing long-range dependences, yet still suffer from detail loss during progressive reconstruction. To address these problems, we propose HFSC-Net, a hierarchical-feature-guided spatial-frequency dual-domain collaborative network. The hierarchical feature extraction framework in the encoder learns multilevel representations that simultaneously preserve local details, structural context, and high-level semantics. Specifically, the structural context pyramid mitigates large-scale variation through multiscale structural support; the semantic knowledge projector injects semantic priors to mitigate intraclass variation and interclass similarity while suppressing scene-wide complex background interference. Spatial-frequency dual-domain collaborative decoding in the decoder coordinates frequency-domain scale-sensitive and category-discriminative cues with long-range background constraints in the spatial domain, while improving boundary and texture recovery. A lightweight bidirectional cross-branch module performs context-conditioned gain coupling between the two paths and fuses the retained multiscale frequency features with the spatial output before prediction. Experiments show that HFSC-Net achieves mean intersection-over-union (mIoU) scores of 56.26%, 87.33%, 85.69%, and 74.00% on LoveDA, Potsdam, Vaihingen, and UAVid, respectively; under the evaluation protocols adopted in this article, these results are the highest mIoU values among the corresponding comparison methods.