Hyperspectral Super-Resolution via Global-Local Low-Rank Matrix Estimation

Ruiyuan Wu, Wing-Kin Ma, Xiao Fu, Qiang Li

Abstract

Hyperspectral Super-Resolution (HSR) is a critical task in remote sensing that seeks to overcome the hardware trade-off between spatial and spectral resolution. This paper presents GLORIA, a novel algorithm that fuses low-spatial/high-spectral resolution images with high-spatial/low-spectral ones. Unlike traditional global models, GLORIA assumes the data possesses a low-rank structure both globally and within local patches, effectively modeling endmember variability. Utilizing a non-convex Schatten-p approximation and an efficient "inexact" optimization strategy, GLORIA outperforms existing methods in reconstruction accuracy and visual fidelity across multiple benchmark datasets.

Introduction

Hyperspectral (HS) sensors provide rich spectral information but often suffer from low spatial resolution. Conversely, multispectral (MS) sensors offer high spatial detail with limited spectral bands. HSR is the computational fusion of these sources to estimate a high-resolution "super-image." A significant hurdle in HSR is Endmember Variability (EV)—the spatial variation of spectral signatures due to lighting or material differences. Standard global models often fail to capture these nuances, leading to artifacts.

The HSR Model and Problem Formulation

The proposed framework models the degradation process where a high-resolution scene is observed through different sensors. The MS sensor captures high spatial resolution but degrades the spectral information, while the HS sensor captures high spectral resolution but degrades the spatial information through blurring and downsampling.

The HSR observation model illustrating spectral and spatial degradation paths.

Global-Local Low-Rank Estimation

The core innovation of this work is the Global-Local Low-Rank (GLLR) approach. While the entire image is assumed to be low-rank (global structure), the authors segment the image into local patches, each possessing its own low-rank structure. This allows the model to adapt to local spectral variations (EV) that a global model would smooth over.

Illustration of local patch segmentation and matricization for local low-rank modeling.

To solve the resulting rank-minimization problem, the authors use a non-convex Schatten-p approximation ($p < 1$), which provides a tighter rank estimate than the standard convex nuclear norm. The GLORIA (Global-Local lOw-Rank promotIng Algorithm) utilizes Majorization-Minimization (MM) and Accelerated Projected Gradient (APG). An "inexact" MM scheme is employed, performing only one APG iteration per MM step, which drastically reduces computational time without sacrificing convergence guarantees.

Experimental Results

The researchers evaluated GLORIA on several well-known datasets, including Chikusei, Indian Pines, and Washington DC Mall.

  • Superior Recovery: GLORIA consistently ranked 1st or 2nd across all performance metrics. On the Chikusei dataset, it achieved a PSNR of 37.73 dB, significantly outperforming the baseline NNM (32.63 dB) and the popular HySure (36.15 dB).
  • Visual Fidelity: Qualitative results showed that GLORIA produces smoother images with fewer artifacts. Competing methods often suffered from "pepper noise" or "strip noise," whereas GLORIA maintained high visual integrity.
  • Robustness: In synthetic tests with high noise (15dB SNR), GLORIA maintained a significant lead, proving its robustness in challenging sensing environments.

Conclusion

GLORIA demonstrates that incorporating local low-rank priors is essential for handling endmember variability in HSR. While the method is computationally efficient for its class, future work will focus on optimizing memory usage for large-scale eigendecompositions and exploring more adaptive patch segmentation strategies to further enhance reconstruction quality.

Summary updated

BibTeX
@article{wu2020hyperspectral,
  author = {Ruiyuan Wu and Wing-Kin Ma and Xiao Fu and Qiang Li},
  title = {{Hyperspectral Super-Resolution via Global-Local Low-Rank Matrix Estimation}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2020},
  doi = {10.1109/tgrs.2020.2979908},
}
Figure