DC-Diff: A Diffusion-Based Network for Panchromatic Sharpening With Cross-Domain Feature Guidance

Panchromatic image enhancement aims to overcome the resolution limitations of satellite sensors by fusing multispectral and panchromatic (PAN) data. However, existing methods often overemphasize certain information, leading to distortion of local details or loss of global structure. To effectively integrate both global and local structural features while preserving the richness of spectral information, this article proposes a self-supervised PAN sharpening network based on a denoising diffusion probabilistic model, named DC-Diff. The network introduces a cross-domain feature fusion module (CDL) that jointly learns spectral and spatial representations, serving as conditional guidance for the denoising prediction network. In addition, a novel dynamic frequency modulation module (DFMM) is proposed, which incorporates diffusion timestep factors to selectively process information from different frequency components. Experiments conducted on the GF-2, QuickBird, and WorldView-3 datasets demonstrate that the proposed method achieves excellent performance in spatial detail recovery, spectral fidelity, and fusion quality, showcasing its effectiveness and robustness across both reduced-resolution and full-resolution test sets. Ablation studies further validate the effectiveness of the proposed modules, highlighting the method’s potential in remote sensing image fusion.