Iterative Hyperspectral Image Classification Using Spectral–Spatial Relational Features

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