Feature subset selection is a well studied problem in machine learning. One short-coming of many methods is the selection of highly correlated features; a characteristic of hyperspectral data. A novel stochastic feature selection method with three major components is presented. First, we present an optimized feature selection method that maximizes a heuristic using a simulated annealing search which increases the chance of avoiding locally optimum solutions. Second, we exploit local cross correlation pair-wise amongst classes of interest to select suitable features for class discrimination. Third, we adopt the concept of distributed spacing from the multi-objective optimization community to distribute features across the spectrum in order to select less correlated features. The classification performance of our semi-embedded feature selection and classification method is demonstrated on a 12-class textile hyperspectral classification problem under several noise realizations. These results are compared with a variety of feature selection methods that cover a broad range of approaches. Abstract © IEEE
Second Workshop on Hyperspectral Image and Signal processing: Evolution in Remote Sensing (IEEE WHISPERS), 2010, pp. 1-4.
J. D. Clark, M. J. Mendenhall and G. L. Peterson, "Stochastic feature selection with distributed feature spacing for hyperspectral data," 2010 2nd Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, Reykjavik, Iceland, 2010, pp. 1-4, doi: 10.1109/WHISPERS.2010.5594951.