Transformer-Based Diffusion and Spectral Priors Model for Hyperspectral Pansharpening

Hyperspectral pansharpening aims to fuse a high-resolution panchromatic image (HR-PCI) with a low-resolution hyperspectral image (LR-HSI) to produce a high-resolution hyperspectral image (HR-HSI). While recent deep learning-based methods have achieved promising results, their reliance on supervised learning or pretrained models with high-quality labeled datasets limits their practicality in real-world applications. We propose unsupervised transformer-based diffusion with spectral priors (uTDSP), an unsupervised framework that addresses these limitations by leveraging spectral priors learned directly from LR-HSIs. The learned spectral prior is incorporated as a regularization term to guide the fusion process, balancing the contributions of the diffusion model and spectral priors for accurate HR-HSI reconstruction. Comprehensive evaluations on real-world datasets are conducted, including ablation studies and comparative experiments, demonstrating that uTDSP consistently outperforms state-of-the-art methods in quantitative metrics (e.g., peak signal-to-noise ratio, spectral angle mapper) and visual quality. The results underscore its effectiveness and practical applicability.