SGRA: A Spectro-Geometric Rectification Dual-Attention Transformer for Cross-Domain Building Change Detection

Abstract

Reliable building change detection (BCD) in high-resolution remote sensing imagery remains a critical challenge for urban monitoring, primarily due to the severe domain shift between available training datasets and specific target regions, as well as the loss of fine-grained structural details in deep feature extraction. This study introduces a cross-domain BCD framework: a spectro-geometric rectification attention (SGRA) network combined with a target-oriented synthetic data strategy. Structurally, we integrate residual refinement blocks and the novel SGRA module into the Swin-Unet decoder. Unlike standard attention mechanisms, SGRA combines a geometric branch (utilizing orthogonal strip convolutions) to preserve long-range structural integrity and a frequency branch (utilizing Fast Fourier Transform) to adaptively suppress background noise based on global spectral characteristics. This dual design is intended to support building-boundary delineation in complex semiurban and industrial backgrounds. Furthermore, to address the scarcity of labeled target data, we propose a “target-crop” synthesis pipeline that extracts building instances from the target domain via the segment anything model and blends them into diverse backgrounds using Poisson editing to reduce the domain gap. Comprehensive experiments show that our method achieves an F1 score of 83.11% on a local agricultural dataset, improving over the baseline F1 score of 79.91% and delivering clear gains on the target-domain task studied in this work. Moreover, on public benchmarks, it yields competitive F1 scores of 86.22% on LEVIR-CD and 93.19% on CDD. A separate corrected WHU-CD analysis is provided in the main text under the official protocol.