I. Introduction
MULTI/HYPERSPECTRAL imagery has the potential to precisely discriminate different land-cover types using the relatively abundant spectral information. However, with the rapidly increasing spatial resolution of the available remote sensing images, it has been argued that the spectral information alone is not adequate for classification [1]. Consequently, various spatial feature extraction methods have been proposed for providing discriminative information in pattern classification recently, such as pixel shape index [2], object-based procedure [3], Markov random field [4], extended morphological profiles (EMPs) [5], and attribute profiles (APs) [6]. However, the aforementioned conventional 2-D spatial feature extraction methods are mainly based on a single spectral band or the first few principal components of the multispectral bands, which are not efficient for spectral–spatial representation of the hyperspectral remotely sensed imagery.