SPE-diff: A Physical Prior Guided Conditional Diffusion Model for Unsupervised High-Spatial-Resolution HSI Reconstruction

Abstract

Hyperspectral images (HSIs) suffer from low spatial resolution, long revisit intervals, limited swath width, and scarce data, whereas multispectral images (MSIs) excel in these aspects. Spectral super-resolution (SSR) of MSIs has thus become a mainstream approach to obtain HSIs with high resolution and wide coverage. However, existing deep learning SSR methods overly rely on data-driven approaches and lack physical interpretability, resulting in blurred details and spectral distortion, while the absence of both real-world MSI and its spectral response function (SRF) in short-wave infrared (SWIR; 1000–2500 nm) further compounds the reconstruction uncertainty. To address the issues mentioned above, a conditional diffusion model for unsupervised SSR that integrates SRF masking, dictionary-constrained spectral priors, and physical feature encoding, named SPE-diff, is proposed in the manuscript. The mask-dictionary-constrained spectral diffusion prior module applies SRF random masking to improve generalization in unobservable bands and leverages a hyperspectral dictionary that is pretrained via alternating direction method of multipliers (ADMM)-based sparse coding to compute sparse coding loss, enabling accurate recovery in the SWIR range. We align and fuse the physical information, jointly encoded by MSIs and SRF, with the spectral priors. This step yields a unified feature that integrates physical guidance and spectral prior knowledge. Afterward, we inject this feature into a diffusion module using a multilevel conditional mechanism. The mechanism enhances the feature’s informativeness, while the U-Net architecture with spatial–spectral attention ensures collaborative modeling of both spatial and spectral dimensions. The SPE-diff ultimately requires the high-spatial-resolution MSI (HRMSI) and the corresponding SRF as input and outputs the high-spatial-resolution HSI (HRHSI). Experiments show that the proposed SPE-diff can achieve accurate SSR performance in the range of 400–2500 nm while maintaining high spectral consistency and has high practical application potential.