Enhancing hyperspectral image synthesis via the spectral super-resolution post-processing network

Abstract

Hyperspectral images (HSIs) provide rich spectral information essential for numerous remote sensing applications. However, the high cost and complexity of hyperspectral cameras make them difficult to deploy widely. A practical alternative is synthesizing HSI from readily available RGB images, addressing the limited accessibility of HSI data. In this paper, we propose a simple yet effective post-processing network for spectral super-resolution, which enhances the quality of initially generated hyperspectral images. Our method builds on existing CNN-based models, such as DenseUnet, CanNet, and SSDCN, to produce preliminary HSIs by capturing local spatial features. To improve spectral accuracy and overall image quality, we introduce a post-processing stage using the HyperSIGMA foundation model, pre-trained on a large remote sensing dataset, to refine the preliminary HSIs by leveraging global spatial and spectral relationships. Experiments on the DFC2018 dataset show that our post-processing network significantly improves both spatial and spectral fidelity. Quantitative evaluations using PSNR, ERGAS, and SAM metrics confirm the superiority of our two-stage framework, particularly in enhancing spectral reconstruction. These results highlight the potential of foundation models as post-processing networks for spectral super-resolution, providing an accessible and effective solution to HSI data scarcity in remote sensing applications.