Automatic feature selection in hyperspectral satellite imagery

The current generation of satellite sensors acquire imagery in only a small set of spectral bands (1-10). In the nearby future the first satellite-based imaging spectrometers will produce a large set of bands (>100). The enormous quantities of data in this hyperspectral imagery will require either analysis methods different from those applied today, or methods to reduce the data quantity. These problems will be especially evident in the global change efforts. One way to overcome the problem with the enormous data quantities is to reduce the data set to an optimal or near-optimal set of bands. An optimal data set will contain the necessary information within given margins determined by the required accuracy and the quantity of data that can be handled by the image analysis methods and computers. A feature selection method requires in most cases that a set of ground truth data is given. The areas of known class in the imagery determine the information characteristics that the user is interested in, and the imagery determines the spectral features of the given classes. A straightforward way to determine which channels that contain the most important information could be to classify all combinations of bands in sets containing from one to all channels. This method is, however, extremely computer demanding. Therefore, an apparently new and promising method, based on Markov chain theory, is developed and presented