Hyperspectral Image Classification Using Multiple Kernel Learning SVM with FA-KLD-LFDA for Multi Feature Selection

Abstract

A new integrated tremendously growing technology for remote sensing is a hyperspectral remote sensing technology which is applied in wide range of fields. By providing rich spectral information from its hyperspectral images (HSI), this technology offers reliable recognition and classification of the objects in the Earth. The present conventional technologies consider only the s spectral and spatial features of the captured images, however, hyperspectral remote sensing technology operates with the high dimensions of hyperspectral images that ultimately results in termination of information retrieval. This redundancy is often more higher is in spatial-spectral domain thus requiring large number of training data for accurate modeling of the classifier. Hence in this technology, to get an accurate classification rate, the complementary properties of the different features of the images are taken into account and joined. Moreover, a multiple feature selection algorithm is proposed in this work to choose the essential features thus avoiding the problem of feature domain selection problem thus overcoming the problem of the dimensionality reduction. The exact objective of this work is to propose a novel dimensionality and multiple feature selection algorithms called FA-KLD-LFDA. In this method, dimensionality reduction is effectively carried out in a raw spectral-spatial feature domain using Firefly Algorithm (FA) with Kullback-Leibler divergence (KLD) for multiple feature selection and Local-Fisher’s Discriminant Analysis (LFDA) for feature projection. Multiple kernel learning (MKL) with Support Vector Machine (SVM) classifier follows these analysis with specific constraints. The experimental data along with the classification results implies that this novel method works exceptionally well than that of the conventionally followed dimensionality reduction and classification algorithms. The challenges of the traditional methods like small training sample size and mixed pixel conditions are seemed to overcome by this proposed method thus achieving higher accuracy.
Copyright © 2014 Praise Worthy Prize - All rights reserved.