Formation of a set of informative classification features for solving cloud classification problem using MODIS satellite data

An algorithm for the formation of a set of effective classification features, based on the truncated search concept and the use of the information about individual classification indicators in the granules selection, is proposed. Its computational efficiency is ensured by the use of simple comparison operations of classification results of individual classes when choosing the most informative granule at the next iteration and using the parallel computing technology on graphics processing units.
Known methods of the truncated selection for the formation of sets of effective classification features are considered. The results of the informative features search are discussed through the example of solving the cloud classification problem on the basis of the application of a probabilistic neural network and the texture information of MODIS satellite imagery. A description of the used classifier and the statistical approach to describing the texture of images is given.
The most effective cloud classification characteristics are determined by comparing the combinations of textural features obtained by truncated selection methods. The study results of the change dynamics in the correctly classified clouds estimation when performing various algorithms for informative features searching are shown. It is established that the method, developed in this paper, makes it possible to reduce the variance of probability values of the correct classification of individual classes.