Unsupervised Building Change Detection in Remote Sensing Images via Joint Structural and Non-Structural Variation Generation

Remote sensing change detection plays a vital role in urban expansion analysis, ecological assessment, and disaster monitoring. While deep learning-based methods have significantly advanced performance, they often rely on large-scale multi-temporal datasets with pixel-level annotations. To alleviate this dependency, generating paired bi-temporal remote sensing images has been explored, but its effectiveness is limited in the absence of pixel-level semantic masks. Moreover, most generation-based change detection methods focus solely on modeling structural changes such as the construction or removal of buildings, while neglecting non-structural variations (i.e., changes caused by illumination shifts, atmospheric fluctuations, or seasonal dynamics that do not alter the physical structure of objects). In this paper, we propose SNV-GenCD, an unsupervised change detection framework for buildings in high-resolution optical remote sensing images, which generates pseudo bi-temporal remote sensing images by jointly modeling structural changes and non-structural variations. Within SNV-GenCD, a structural change generation module produces object-manipulated images along with corresponding pixel-level change masks, and a non-structural variation module injects radiometric changes that reflect real-world remote sensing conditions. Further, to eliminate temporal bias and enhance robustness against pseudo-changes in the generated data, we introduce two complementary constraints that guide the detector to learn robust, order-invariant representations. Experiments on the LEVIR-CD, WHU-CD, and GoogleCD datasets show that SNV-GenCD achieves F1 scores of 82.2%, 87.6%, and 76.7%, respectively. These results significantly outperform existing state-of-the-art unsupervised methods and are competitive with weakly supervised approaches.