Classification of Spectral Image Using Genetic Programming

This paper describes an initial use of genetic programming as a discovery engine that derives two sets of information from hyper-spectral imagery. The first consists of a set of classification algorithms learned from the data. The second consists of reduced subsets of the most germane bands for use in a given classification, since not all spectral bands are of use in deriving a particular classification algorithm. Currently, there are only a few techniques to discover which bands would be the most useful for a specific classification task. We describe the design of a prototype system and discuss its efficacy on a novel data set from an imaging system that uses an acoustically tuned optical filter. The data preprocessing, training data extraction, training data formatter, GP implementation, and classification image generation tasks are detailed.