The current state-of-the science approach to hyperspectral imagery collection and exploitation derives from algorithms and tools developed in early 1990s, which apply many simplifying assumptions or expedient processing steps. The most significant simplification occurs in atmospheric correction where a single geometric solution for all elements of the radiance equation is applied to every pixel of a spectral image. We know this solution to be expedient, but also, error inducing. Another significant shortcoming is the inability of current algorithms to exploit more than one spectral image at a time. Atmospheric characterization and correction, and target detection is performed one image at a time: every image is an island.  

These concepts and solutions have demonstrated operational success and their methods are ensconced in existing algorithms and software. While these existing solutions are based on sound physics, mathematics, and statistics, they lag behind the revolution in computational power, artificial intelligence, and agile sensors and platforms. Most importantly, they have been shown to fail under varied environmental conditions, obscuration, and target material conditions. It can be said that today's analysis is effective in performing material identification in optimal collection conditions. It can also be said that in non-optimal conditions, cloudy scenes, shadows, intimate mixtures of materials, liquid spills and residues, or any combination of the above, material identification of solids, liquids, or gases is ineffective, unrepeatable, or subject to unknown levels of uncertainty.  

The gap in effectiveness stems from the incomplete solution of the radiance equation that a single hyperspectral image and existing analysis methods afford.   This talk discusses new research that promises to greatly improve today’s state-of-the-art hyperspectral imagery analysis methods by expanding the spatial and temporal dimensions of the radiance equation solution and implementing Deep Learning.