A Sub-Pixel Coupled Dictionary Learning Method for Large-Scale and High-Spatial-Resolution Satellite Hyperspectral Image Reconstruction

Tianzhu Liu, Zitong Liu, Xianhao Zhang, Yanfeng Gu, Jian Liu

Abstract: Satellite hyperspectral images (HSIs) are often constrained by the limited spatial resolution and narrow imaging width, hindering precise classification in vast observation scenes despite boasting hundreds of approximately contiguous spectral channels and high spectral resolution. In contrast, satellite multispectral images (MSIs) lack spectral details but offer comprehensive spatial information in large scenes. To overcome the limitations of HSIs and MSIs, we proposed a novel MSI-HSI spatial-spectral collaborative representation model (Sub-CDL), leveraging sub-pixel representation and coupled dictionary learning. The model learns information consistency between MSI and HSI through the construction of a sub-pixel coupled dictionary, where spatial-resolution disparities between HSI and MSI are addressed by using the sub-pixel dictionary, spatial down-sampling matrix, and spatial blur matrix. Additionally, spectral-resolution differences are resolved by spectral response function (SRF). More specifically, the method utilizes spatial regularization constraint and sub-pixel coupled dictionary low-rank constraint as restrictive conditions, and the alternating direction method of multipliers (ADMM) is used to optimize the model. Ultimately, high-resolution reconstruction of HSIs with large scenes is enabled. By comparing with various state-of-the-art methods, experimental results demonstrate the proposed model’s superiority in both authenticity and classification accuracy of reconstructed images.