Hyperspectral Super-Resolution via Interpretable Block-Term Tensor Modeling

This work revisits coupled tensor decomposition (CTD)-based hyperspectral super-resolution (HSR). HSR aims at fusing a pair of and multispectral images to recover image (SRI). The vast majority the approaches take low-rank matrix recovery perspective. challenge is that theoretical guarantees for recovering SRI using models are either elusive or derived under stringent conditions. A couple recent CTD-based methods ensure recoverability relatively mild conditions, leveraging on algebraic properties canonical polyadic (CPD) Tucker models, respectively. However, latent factors both CPD have no physical interpretations in context spectral analysis, which makes incorporating prior information challenging---but priors often essential enhancing performance noisy environments. employs an idea as tensors following block-term model with multilinear rank-$(L_r, L_r, 1)$ terms (i.e., LL1 model) formulates problem problem. Similar existing CTD approaches, shown More importantly, can be interpreted key constituents images, i.e., endmembers' signatures abundance maps. connection allows us easily incorporate enhancement. flexible algorithmic framework series structural proposed advantage interpretability. effectiveness showcased simulated real data.