Dual-Domain Manifold Modeling for Hyperspectral Image Fusion

A novel approach known as Dual-Domain Manifold Modeling (DDMM) has been introduced for the fusion of hyperspectral images. This method overcomes the shortcomings of current techniques by incorporating geometric constraints in both the spatial and spectral realms. Within the spatial domain, a Topology-Aware Transformer (TPFormer) merges global attention with neighborhood propagation to effectively model spatial topology alongside pixel-level feature manifold relationships. This significantly boosts geometry-aware feature learning while maintaining high-frequency structural details. Meanwhile, in the spectral domain, local manifold structures driven by spectral similarity are utilized to enhance the modeling of intrinsic pixel relationships and achieve precise spectral reconstruction. The goal is to seamlessly integrate the richness of spectral data with spatial accuracy.