Novel Classification Technique for Hyperspectral Imaging using Multinomial Logistic Regression and Morphological Profiles with Composite Kernels

Abstract

Hyperspectral imaging (HI) is getting much more attention among researchers in different fields like agriculture, defense, medical, and geographical surveys. In this work, we have proposed a novel automated system for the classification and segmentation of landscapes using hyperspectral images. The proposed semi-supervised based approach has improved the extraction of spatial characteristics of the scene that has employed an extended multi-attribute profile (EMAP) by stacking of several attributes. The unlabeled data points located near the classifier boundaries are selected on the basis of entropy related to the corresponding class labels. In the next segmentation phase, MLR probabilities are computed against the output of classifier. Finally, maximum-a-posteriori segmentation is carried out on the multilevel logistic prior labels. The simulated results have obtained classification accuracy of 95.50% by comparing predicted labels with original ones. The segmentation accuracy, after developing regions on the output of classification, is 98.31%. A performance comparison of the proposed approach with several approaches has also been carried out.