RoFDiff: Robust Hyperpansharpening via a High-Low Frequency Conditional Diffusion Model

Noise of hyperspectral images is inevitable due to the factors of hyperspectral imaging systems, damage sensors and physical environments. However, the existing hyperpansharpening methods often ignore the noise when fusing low-resolution hyperspectral (LRHS) and panchromatic (PAN) images, since noise will destroy their correlations. Although the pretrained denoising networks can be used to remove noise, the limitation of the generalizability and the error accumulation between models will decrease the fused image quality. To address this issue, this paper proposes a robust hyperpansharpening method based on a high-low frequency conditional diffusion model (CDM), namely RoFDiff. Since noise primarily exists in the high-frequency regions of LRHS images, the multiscale filter (MSF) is deployed to obtain clean low-frequency features (CLFFs) that retain spectral information and avoid noise interference. Additionally, the spectral information of the CLFFs and the spatial information from the PAN images serve as two prior conditions to guide the accurate generation of the high-frequency feature (HFFs). To enhance the reconstructed HFFs, the spectral fusion module (SFM) leverages inter-spectral correlations to improve spectral fidelity, while the spectral-spatial reconstruction module (S2RM) utilizes multiscale spatial relationships to refine texture details. Extensive experiments on three public datasets demonstrate that the proposed RoFDiff can generate higher-quality HRHS images under various noise conditions in terms of the spatial and spectral fidelity compared with recent state-of-the-art methods.