2 December 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Hyperspectral-multispectral image fusion reconstructs high-spatial-resolution hyperspectral images by combining spectral information from hyperspectral imagery with spatial information from multispectral imagery. Although deep learning methods achieve promising results through nonlinear modeling, existing approaches fail to preserve complete target information despite global-local spatial feature extraction. To overcome this limitation, we propose a two-stage Global-Local Dual-Domain Latent Diffusion Network (GLDFN) by incorporating a collaborative spatial-frequency domain representation scheme. Stage 1 extracts latent space representations using reference hyperspectral images as ground truth. This stage trains three interdependent parameter sets: global-local spatial-frequency inverse transform parameters, local-global feature fusion parameters, and spatial-frequency reconstruction parameters. The resulting latent space enhances feature discriminability while optimizing computational efficiency for subsequent processing. Stage 2 first linearly fuses low-spatial-resolution hyperspectral and high-spatial-resolution multispectral inputs. Parameters from Stage 1 then generate spatial-frequency conditional matrices. Using these matrices with the ground truth latent space, we train spatial-frequency diffusion model parameters. Finally, Stage 1 reconstruction parameters synthesize the fused hyperspectral image from diffusion model outputs. Experimental results on synthetic and simulated datasets demonstrate GLDFN achieves robust, high-quality fusion while outperforming state-of-the-art methods.