An approach to feature selection and classification of remote sensing images based on the Bayes rule for minimum cost

Classification of remote-sensing images is usually carried out by using approaches aimed at minimizing the overall error affecting land-cover maps. However, in several remote-sensing problems, it could be useful to perform classification by taking into account the different consequences (and hence the different costs) associated with each kind of error. This allows one to obtain land-cover maps in which the total classification cost involved by errors is minimized, instead of the overall classification error. To this end an approach to feature selection and classification of remote-sensing images based on the Bayes rule for minimum cost (BRMC) is proposed. In particular a feature-selection criterion function is presented that permits one to select the features to be given as input to a classifier by taking into account the different cost associated with each confused pair of land-cover classes. Moreover, a classification technique based on the BRMC and implemented by using a neural network is described. The results of experiments carried out on a multisource data set concerning the Island of Elba (Italy) point out the ability of the proposed minimum cost approach to produce land-cover maps in which the consequences of each kind of error are considered.