Can infrared (IR) hyperspectral imagery be acquired from IR multispectral sensors?

Hyperspectral imaging systems are able to address critical challenges, ranging from detecting chemical, biological, radiological, nuclear, and explosives materials to identifying targets from remote distances. However, due to their complexity, these sensors are expensive to build, maintain, and operate. A promising solution that has been recently explored is to leverage the abundance of inexpensive and mature multispectral cameras to reconstruct hyperspectral images via machine-learning (ML) software.
This report provides a review of ML algorithms that leverage color (red, green, blue) data to reconstruct hyperspectral imagery. There are two classes of algorithms that are reviewed here. The first is the classical, prior-information-based methods that utilize dictionary and manifold-learning techniques to reconstruct hyperspectral signatures from three-channel color data. The second relies on deep-learning methods that use neural networks to learn the mapping from color to hyperspectral using training data and supervised learning. These approaches are summarized, and their relative advantages and disadvantages are discussed. Finally, a plan is outlined to extend current work from visible and near-infrared (IR) data to IR imagery.