Iterative Hyperspectral Image Classification Using Spectral–Spatial Relational Features

Abstract

This paper describes the principles and implementation \nof an algorithm for the classification of hyperspectral remote \nsensing images. The proposed approach is novel and can \nbe included within the category of the spectral–spatial classification \nalgorithms. The elements of novelty of the algorithm are \nas follows: 1) the implementation of two classifiers that work \niteratively, each one exploiting the decision of the other to improve \nthe training phase, and 2) the use of relational features based \non the current labeling and on the spatial structure of the image. \nThe two classifiers are fed with the spectral features and with the \nspatial features, respectively. The spatial features are built using \nthe relative abundance of each class in a neighborhood of the pixel \n(homogeneity index), where the neighborhood is properly defined. \nAn important contribution to the success of the method is the \nadoption of a multiclass classifier, the multinomial logistic regression, \nand a proper use of the posterior probabilities to infer the \nclass labeling and build the relational data. The results of the two \nclassifiers are eventually combined by means of an ensemble decision. \nThe algorithm has been successfully tested on three standard \nhyperspectral images taken from the Airborne Visible–Infrared \nImaging Spectrometer and ROSIS airborne sensors and compared \nwith classification algorithms recently proposed in the literature.