Feature selection using rough set theory for object-oriented classification of remote sensing imagery

Abstract

In object-oriented remote sensing imagery classification, numerous spectral, texture, shape and contextual features can be derived and used to discriminate classes and produce finer map. The high-dimensional features may induce Hughes phenomenon that classification accuracy decreases with more features involved. To improve the classification accuracy and efficiency, a hybrid feature selection method combined the relative attribute reduction and the significance estimation of features is proposed. This method can efficiently select features and solve the problems of combination explosion. Object-oriented classification of Quickbird image shows the selected features can correctly distinguish most of the objects with an overall accuracy of 86%.