New Dual-Domain Manifold Model Enhances Hyperspectral Image Fusion

Hyperspectral image fusion aims to combine rich spectral information with high spatial detail, but current methods often struggle with effectively modeling geometric constraints. Existing approaches face challenges in both spatial and spectral domains. In the spatial domain, weak interactions between spatial and spectral features limit geometry-aware learning, leading to structural degradation. In the spectral domain, the local manifold structures derived from spectral similarity are not fully utilized, hindering the modeling of intrinsic pixel relationships and fine-grained spectral reconstruction. To overcome these issues, a new Dual-Domain Manifold Modeling (DDMM) framework has been introduced. This framework incorporates a Topology-Aware Transformer (TPFormer) that combines global attention with neighborhood propagation to jointly model spatial topology and pixel-level feature manifold relationships. This approach aims to capture intrinsic spatial-spectral structures and enhance topology-aware representation learning. Additionally, the DDMM framework includes a Frequency-Decoupled Spatial-Spectral Collaborative Fusion (FDSCF) module. This module projects features into the frequency domain using discrete cosine transform, explicitly separating them into low- and high-frequency components. Guided by a low-rank structural prior and spectral-driven spatial enhancement, FDSCF selectively boosts geometry-aware high-frequency features, thereby strengthening spatial-spectral coupling and recovering sharper edges and finer textures. Extensive experiments on benchmark datasets reportedly show that DDMM outperforms state-of-the-art methods in preserving spatial structure and reconstructing spectral information.