Unified Guided Hyperspectral Image Denoising by Continuous Coupled Tucker Decomposition

Hyperspectral image (HSI) denoising is a critical preprocessing step for subsequent interpretation tasks. While guided denoising utilizing auxiliary high-quality images [e.g., panchromatic (PAN), multispectral (MSI), or synthetic aperture radar (SAR)] has shown promise, existing methods are modality-specific and fail to generalize across different guidance types. This article proposes a unified guided HSI denoising framework termed continuous coupled Tucker decomposition (CCTD). The proposed method performs a joint factorization of the HSI and guidance modalities via Tucker decomposition, sharing the spatial factor matrices and core tensor across modalities to implicitly align spatial structures while preserving modality-specific spectral characteristics through independent spectral factors. To further exploit internal spatial smoothness priors without the burden of cumbersome hyperparameter tuning, the factor matrices are parameterized as continuous functions of coordinates via implicit neural representations (INRs) with sine activations, which naturally encode smoothness through architectural inductive bias. An explicit nuclear norm regularizer on the spectral mode eliminates the need for manual Tucker rank selection. Extensive experiments on three datasets (Beijing, Wuhan, and Florence) with six noise cases and three guidance modalities (MSI, SAR, and PAN) demonstrate that CCTD generally outperforms state-of-the-art single-image and guided denoising methods, achieving superior peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), erreur relative globale adimensionnelle de synthése (ERGAS), and spectral angle mapper (SAM) metrics. Ablation studies validate the effectiveness of each component. The proposed framework provides a versatile solution for multimodal HSI restoration.