Classification of multispectral satellite land cover data by 3D local discriminant bases algorithm

Abstract

In this study, a 3D Local Discriminant Bases based algorithm is developed to extract the discriminative features from multispectral satellite/airborne data. The developed algorithm, first localizes the information in hyperspectral data by the trees generated both along the spectral and spatialfrequency axis. These trees are then automatically pruned to obtain the location of the discriminative features in data space. The extracted features are ranked by feature selection algorithms to eliminate the irrelevant ones for classification. This combination of feature extraction and selection algorithms also identifies the specification of the relevant spectral bands including center frequency and bandwidth in imaging. The algorithm is implemented on a multispectral airborne data set from Tippecanoe County, Indiana for classifying five vegetative species and an average classification error of 8.85% is achieved with three extracted features. © 2011 Springer Science+Business Media B.V.