Multi-Source image enhancement via Coupled Dictionary Learning

Motivation. Multi and Hyperspectral remote sensing imagery provide valuable insights regarding the composition of a scene and significantly facilitate tasks like object and material recognition, spectral unmixing and region clustering, among others [1], [2]. However, current remote sensing imaging architectures are unable to concurrently acquire high spatial and spectral resolution imagery, due to fact that the three dimensional hyperspectral data must be acquired using a single, a 1D array, or 2D plane detector. On the one hand, traditional “push-broom” sensors obtain a high spectral resolution profile of a very low spatial resolution area (single line) during each exposure and generate the complete hyperspectral cube by progressive scanning of the scene. At the other end, Snapshot Spectral Imaging architectures sample the full spatio-spectral cube in a single exposure, without any need for successive frame acquisition, by associating each pixel with a specific spectral band. Even for the same type of sensor, a different point in the spatio-spectral operational curve might be selected depending on the application. We report herein on a novel machine learning method for postacquisition enhancement of multi and hyperspectral imagery. Example applications include imagery acquired from legacy low spectral resolution satellites, which could be enhanced using images of the same region, acquired by high resolution spectrometers aboard newer platforms. Respectively, the limited spatial information acquired by hyperspectral instruments, could be enhanced using high-spatial resolution imagery extracted by sensors with higher spatial resolution.