.png){kind=icon}
.png){kind=icon}
Return
FFCDNet: Remote sensing image change detection method based on fourier frequency domain feature enhancement
.png){kind=icon}
OA
AI
H
M
J
Q
DOI：10.1016/j.srs.2026.100447.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Remote sensing imagery contains abundant local structural details and complex global variation patterns. However, CNN-, Transformer-, and Mamba-based methods each exhibit inherent limitations — restricted receptive fields, high computational complexity, and directional bias, respectively — making it difficult to achieve a balanced modeling of local and global representations. To address these issues, this paper proposes a Fourier frequency-domain feature enhancement–based remote sensing image change detection method, aiming to efficiently and accurately capture both fine-grained spatial details and non-directional global structural information in bitemporal images. Specifically, an end-to-end Fourier encoder–decoder framework, termed FFCDNet, is constructed. First, a Frequency-domain Hybrid Enhancement Block (FHEB) is introduced, which integrates parallel local convolutional and global frequency-domain branches based on the Fast Fourier Transform. This design enables non-directional and holistic long-range dependency modeling while preserving spatial detail. In particular, FHEB decomposes complex frequency spectra into real and imaginary components and learns them separately. Instead of assuming a strict correspondence between these components and specific spatial or frequency characteristics, they are treated as complementary representations that jointly contribute to modeling global structural information and fine-grained spatial details. Second, an Adaptive Bitemporal Interaction Module is designed to perform dynamic cross-temporal fusion and difference enhancement through global-local guided feature integration and an adaptive gating mechanism, enabling the extraction of the most discriminative change features. By jointly modeling frequency-domain and temporal interactions, the proposed framework achieves unified perception of global context and local detail without relying on recurrent scanning structures. Extensive experiments demonstrate that the proposed FFCDNet consistently outperforms existing state-of-the-art models across multiple remote sensing change detection benchmark datasets, exhibiting superior accuracy, boundary fidelity, and generalization capability.
Keywords:
Change detection
Remote sensing
Fourier
Frequency domain
Mamba
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
5.2
Papers:
457
Citations:
980
Organization
Cited Papers