Interpretable Landsat-to-Hyperspectral Dual Super-Resolution Without Large Matrix Inversion

Abstract

Direct acquisition of global hyperspectral images (HSIs) is infeasible given contemporary hardware facilities and limited resources, while global hyperspectral monitoring is critical for remote sensing applications. A more economical approach is to interpretably convert global Landsat-8/9 multispectral images into NASA's AVIRIS-level HSIs. This conversion involves both spatial super-resolution (SpaSR, 30-m to 15-m) and the highly ill-posed spectral super-resolution (SpeSR, 7-band to 172-band), collectively referred to as dual super-resolution (DualSR), whose duality between SpaSR and SpeSR has recently been established. Existing SpeSR methods, mostly designed to reconstruct CAVE-level HSIs with only 31 visible bands, are not applicable to the AVIRIS-level task involving 172 visible, near-infrared, and shortwave-infrared bands. This motivates us to customize an interpretable alternating direction method of multipliers network (ADMM-Net) using the Woodbury W-Lemma and a spectral continuity prior. Unlike conventional generative SpaSR, we employ a panchromatic sharpening strategy to recover physically grounded spatial details. However, this strategy induces very large matrix inversions (LMIs), with dimensionality proportional to the number of pixels, even after applying the W-Lemma. We resolve this issue by designing an LMI-free proximal gradient descent network (PGD-Net). Consequently, the proposed PGD-ADMM interpretable network (PAINT) achieves substantial improvements in both computational complexity and reconstruction performance. Beyond state-of-the-art reconstruction performance, PAINT improves Landsat classification from 78.98% accuracy and 76.06% kappa to 92.16% accuracy and 90.98% kappa.