Low-Rank Denoising for Chinese Spaceborne Hyperspectral Imagery With Spectral Fidelity Preservation

Abstract

Chinese pushbroom spaceborne hyperspectral imagery is frequently affected by structural artifacts, such as stripes and dead lines, and by mixed noise that is nonindependent and identically distributed across spectral bands. Such degradation is particularly critical for shortwave-infrared (SWIR) mineral mapping. Following the FastHyDe paradigm, this article proposes a Spectral-Fidelity-aware FastHyDe framework, a training-free low-rank denoising pipeline. First, a lightweight structural artifact suppression stage is used to stabilize the statistical properties of spectral noise, enabling subsequent subspace identification. Then, a bidirectional subspace refinement strategy determines the retained dimension from a symmetric candidate set around the statistical estimate, using a joint statistical–spectral criterion that balances noise suppression against feature-level spectral fidelity. Finally, in the eigen-image domain, a plug-and-play nonlocal prior is used for denoising. This article uses real data from the Gaofen-5 Advanced Hyperspectral Imager in the Xuemisitan area in Xinjiang and validates the results with in situ analytical spectral devices’ field spectra resampled to the sensor response. The results show that, compared to other model-driven and deep-learning methods, this framework removes stripes and suppresses mixed noise more consistently, better preserves the position and shape of the SWIR diagnostic absorption features, and has the lowest computational cost among all nondeep-learning baseline methods.