AHSIG-SSR: Auxiliary Hyperspectral Image-Guided Model for Spectral Super-Resolution

Recently, the spectral super-resolution (SSR) method has attracted increasing attention, which generates a hyperspectral image (HSI) by using a corresponding multispectral image (MSI). However, existing SSR methods usually inevitably have some limitations. Prior-driven SSR methods are difficult to obtain satisfactory results because of the severe lack of spectral information of the MSI. Spectral library-based SSR methods are affected by the incompleteness and poor generalizability of existing spectral libraries. Additionally, deep learning-based SSR approaches, despite their impressive performance, demand large-scale paired datasets and often lack robustness across different sensors and scenes. To alleviate these limitations, we propose a novel SSR framework that uses an easily accessible HSI as auxiliary information to enhance the accuracy and convenience of the SSR method, which circumvents the need for accurate registration and extensive training data. Specifically, we design three models to capture the spectral information from this auxiliary HSI and take the learned spectral information into the spectral reconstruction of MSIs. (i) We design Adaptive Local Spectral Dictionary Learning (ALSDL) to learn the spectral dictionary from the auxiliary HSI, (ii) we design Low-Rank Sparse Spectral Correlation Modeling (LRS-SCM) to learn the spectral correlation from the auxiliary HSI, and (iii) we design Similar Spectral Fibers Filling (SSFF) model to extract the similar spectral fibers from the auxiliary HSI. Experimental results on multiple hyperspectral datasets demonstrate the efficacy of the proposed method.