Skip to content
Open access

A Frequency-Guided Axial Refinement Network for High-Resolution Remote Sensing Image Semantic Segmentation

2026 · IEEE Transactions on Geoscience and Remote Sensing · Vol 64, pp. 5645815-5645815 · 0 citations · 71 references

Abstract

The remote sensing image semantic segmentation is fundamental to fine-grained land-cover interpretation. However, in high-resolution complex scenes, segmentation performance is still limited by strong interclass similarity, large-scale variation, and the frequent degradation of boundary details. To address these issues, we propose a frequency-guided axial refinement network (FGAR-Net). The FGAR-Net adopts an asymmetric stepwise encoder–decoder architecture. Specifically, a frequency-guided channel recalibration (FGCR) module introduces discrete cosine transform (DCT) priors into skip pathways to suppress background redundancy and highlight discriminative cues. At the bottleneck stage, an axial-window perception (AWP) module jointly captures regional semantics, long-range structural dependencies, and local details. During decoding, the adaptive weighted fusion and context refinement (AWCR) blocks coordinate deep and shallow features to reduce cross-layer semantic inconsistency, while a final context-guided feature refinement (CFR) module imposes global contextual constraints to accurately restore local high-resolution details. Extensive experiments demonstrate that FGAR-Net achieves mean intersection over unions (mIoUs) of 85.17%, 87.55%, and 65.75% on the International Society for Photogrammetry and Remote Sensing (ISPRS) Vaihingen, ISPRS Potsdam, and Wuhan high-resolution dense labeling dataset (WHDLD) datasets. These results indicate that FGAR-Net provides an effective solution for fine-grained segmentation in complex remote sensing scenes.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.