An Unsupervised Domain Adversarial Wavelet Learning Method for registration of Optical and SAR Heterogeneous Remote Sensing Images

Optical images and synthetic aperture radar (SAR) remote sensing images are highly complementary information for Earth observation and applications. Accurate registration of these two data types is essential for subsequent collaboration processing and analysis. However, traditional methods struggle to achieve effective registration due to nonlinear radiometric disparities and geometric distortions. Meanwhile, supervised deep learning approaches require costly manual annotations. To address these issues, this paper proposes an unsupervised registration method based on domain adversarial wavelet learning. We innovatively design an inter-domain adversarial mutual transfer (IAT) module to alleviate cross-modal differences while preserving structural features. Additionally, wavelet transform is employed at different scales for feature matching, to mitigate radiometric disparities and enhance feature recognizability. By incorporating an attention mechanism, the network structure is optimized through spatial alignment and image similarity. The proposed method outperforms other state-of-the-art methods in both quantitative and qualitative aspects of benchmark datasets.