Frequency–spatial decoupled co-modeling transformer for fine-grained remote sensing image segmentation

Li X, Qian S, Lyu X, Song Y, Liu F, Fang Y, Xu Z, Kaup A (2026)

Publication Type: Journal article

Publication year: 2026

Journal

Information Sciences Elsevier

Book Volume: 747

Article Number: 123489

DOI: 10.1016/j.ins.2026.123489

Abstract

Fine-grained semantic segmentation of remote sensing images (RSIs) remains a long-standing challenge due to complex scene structures, diverse object scales, and blurred boundaries that hinder the precise delineation of land-cover classes. Existing Transformer-based approaches mainly focus on spatial attention while overlooking the complementary frequency-domain cues essential for capturing subtle textural and boundary details. To address this issue, we propose FDFormer, a frequency–spatial co-modeling transformer that incorporates frequency-domain representation learning into the segmentation pipeline. Specifically, the proposed frequency decoupling attention (FDA) adaptively separates features into high- and low-frequency bands to enhance structural perception and suppress redundant components. Moreover, a second-order cross-attention mechanism fuses the spatial stream with the frequency stream to strengthen global-local interactions and produce more discriminative context-aware representations. In addition, a hybrid boundary-aware loss combines pixel-wise cross-entropy with frequency-guided boundary supervision to balance region consistency and edge precision. Experiments conducted on two benchmark datasets, ISPRS Potsdam and LoveDA, demonstrate that FDFormer consistently surpasses state-of-the-art methods. Qualitative results further verify its effectiveness in preserving thin structures and intricate object contours.

Authors with CRIS profile

Andre Kaup Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS)

Involved external institutions

Hohai University CN China (CN) Curtin University AU Australia (AU)

How to cite

APA:

Li, X., Qian, S., Lyu, X., Song, Y., Liu, F., Fang, Y.,... Kaup, A. (2026). Frequency–spatial decoupled co-modeling transformer for fine-grained remote sensing image segmentation. Information Sciences, 747. https://doi.org/10.1016/j.ins.2026.123489

MLA:

Li, Xin, et al. "Frequency–spatial decoupled co-modeling transformer for fine-grained remote sensing image segmentation." Information Sciences 747 (2026).

BibTeX: Download