Zero-Shot Multidegraded Hyperspectral Image Restoration via Intrinsic Image Decomposition

Abstract

Hyperspectral images (HSIs) are often simultaneously degraded by spatial downsampling, complex noise contamination, and pixel loss during acquisition, resulting in spatial blurring and spectral distortion that severely hinder quantitative analysis and downstream applications. Most existing HSI restoration methods are designed for single degradation scenarios and heavily rely on paired supervision, which limits their applicability in unseen extreme environments where ground truth is unavailable. To address these challenges, this article proposes an unsupervised zero-shot HSI restoration (ZSHR) method based on intrinsic image decomposition (IID), grounded in physical imaging principles. Specifically, an HSI is modeled as the multiplicative interaction between reflectance and shading components, and the restoration under multiple degradations is reformulated as a joint estimation problem of intrinsic components under physical consistency constraints. This physics-driven decoupling allows the model to isolate inherent material properties from environmental illumination factors, ensuring high spectral fidelity in extreme scenarios. Accordingly, an end-to-end optimization framework is developed to jointly perform denoising and super-resolution (SR) reconstruction. By incorporating intrinsic consistency constraints between reflectance and shading, frequency-guided feature modulation, and panchromatic-guided cross-modal spatial fidelity constraints, the proposed method achieves high-quality HSI restoration without any external supervision. Extensive experiments on multiple public remote sensing datasets and challenging lunar in situ hyperspectral data demonstrate that the proposed method consistently outperforms existing approaches in terms of spectral fidelity, spatial detail preservation, and downstream classification and unmixing performance.