Combined subspace low rank learning and nonlocal low rank estimation for spectral super-resolution of multispectral remote sensing images

Le résumé fourni par la source

Spectral Super-Resolution (SR) of Multispectral Images (MSI) enhances spectral resolution in MSI's non-overlapping regions by utilizing the overlapping regions with Hyperspectral Images (HSI). This technique effectively addresses the trade-off between spectral resolution and spatial coverage in remote sensing imagery, attracting significant research attention. However, traditional spectral SR methods based on sparse representation and deep learning primarily exploit local similarities in HSI, neglecting prevalent non-local spatial patterns, thus limiting reconstruction performance. In this paper, we propose a Combined Subspace Low Rank Learning and NonLocal Low Rank Estimation (CSLNLE) method for MSI spectral SR, capturing both global spectral correlations and nonlocal spatial-spectral self-similarity. The CSLNLE method decomposes the target HSI into a low-rank dictionary of subspaces and associated coefficient matrices. Through low-rank learning, we derive dictionaries from overlapping HSI-MSI regions and estimate coefficients for non-overlapping MSI regions using a nonlocal tensor multi-rank prior. Specifically, images are partitioned into patches, clustered by similarity, and coefficient matrices are estimated via low-tensor multi-rank decomposition within each group. Spectral SR experiments on simulated and real datasets using seven current benchmark methods validate the superiority of our proposed method.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.