Frequency-Adaptive Disentanglement and Topological Graph Alignment for Cross-Modal Optical–SAR Ship Reidentification

Cross-modal optical and hybrid optical and synthetic aperture radar (SAR) ship reidentification is pivotal for all-weather, continuous maritime surveillance. However, establishing reliable identity associations across these heterogeneous sensors is fundamentally bottlenecked by two deeply coupled challenges: extreme multiplicative speckle noise at the signal level and nonlinear geometric distortions induced by SAR side-looking imaging. Existing spatial-domain alignment methods are inherently ill-equipped to handle such severe radiometric and structural discrepancies. To fundamentally overcome this, we propose the frequency-domain decoupling and topological graph alignment (FD-GA) method, which shifts the cross-modal matching paradigm from rigid spatial pixels to a robust synergistic frequency-topological space. Specifically, a frequency-domain adaptive augmentation module is designed to suppress modality-specific scattering interference via cross-modal amplitude blending while strictly preserving the structural phase spectrum. To eradicate the persistent entanglement between identity semantics and modal styles, a modality-disentangled feature extractor leverages a sequence-level orthogonality constraint to guarantee representational purity. Finally, circumventing the catastrophic failure of absolute spatial matching, a topology graph alignment module dynamically models local ship components as a non-Euclidean graph, effectively capturing the invariant topological connectivity of rigid structures against severe viewpoint deformations. Extensive experiments on the large-scale ship reidentification (HOSS-ReID) benchmark demonstrate that FD-GA significantly outperforms state-of-the-art methods, establishing a robust and reliable baseline for cross-modal maritime target tracking in unconstrained environments. Our source code will be available online.