Spectral Quality Evaluation of Reconstructed Hyperspectral Images

With the advance in imaging optics, hyperspectral images (or cubes) have become low-cost and real-Time for acquiring images in the field, specifically thanks to the recent development of different 'snapshot' hyperspectral imaging systems. However, cameras producing high resolutions in both the spectral domains and the spatial domains are still rare or considered to be high-cost. Algorithm-based pansharpening, or in general image reconstruction methods, are often used to create high spatial-resolution cubes by fusing high-spatial gray or color images and low spatial-resolution hyperspectral images. Moreover, most of these methods emphasized achieving high visual quality in spatial resolution but not considering the spectral accuracy in the reconstructed images. This paper aims to evaluate the spectral quality of reconstructed images from multiple methods. A commercial hyperspectral camera (Cubert S185) was used to conduct the research. Important conclusions include that spectral information is lost to different degrees per different reconstruction methods when the spatial resolution is raised too high. The trade-off between spatial sharpening and retaining spectral information is important for machine learning tasks.