IOSR Journal of Electronics and Communication Engineering (IOSR-JECE) e-ISSN: 2278-2834,p- ISSN: 2278-8735.Volume 11, Issue 6, Ver. IV (Nov.-Dec .2016), PP 25-32 www.iosrjournals.org DOI: 10.9790/2834-1106042532 www.iosrjournals.org 25 | Page Analysis of Band Selection Algorithms for Endmember Extraction in Hyperspectral Images M.Rashini 1 , G.Veera Senthil Kumar 2 , S.Vasuki 3 Student, ECE Dept, VCET, Madurai, India 1 Assistant Professor, ECE Dept, VCET, Madurai, India 2 Professor and Head, ECE Dept, VCET, Madurai, India 3 Abstract: This paper presents a novel approach of band selection for dimensionality reduction in Hyperspectral images (HSI). There are several methods of dimensionality reduction which can be further categorized into two groups; feature extraction and feature or band selection. Due to transformation in feature extraction, the critical information may have been distorted. Hence feature selection is preferable for dimensionality reduction because it preserves the relevant original information. Despite many algorithms exist for dimensionality reduction; it is even now a challenging task of selecting informative bands from the large volume data. The number of bands is estimated with the concept of Virtual Dimensionality (VD), because it provides reliable estimate. Bands are selected from hyperspectral images using Exemplar Based Band Selection (EBBS). End members are extracted from the selected bands using Simplex Growing Algorithm(SGA). The performance of EBBS is compared with the existing band selection techniques such as Constrained Band Selection (CBS) and Similarity Based Band Selection (SBBS) using the spectral angle distance as a measure. Keywords: Hyperspectral images, Virtual Dimensionality, Simplex Growing Algorithm, Exemplar based band selection, Spectral angle distance. I. Introduction The hyperspectral images can now concurrently capture hundreds of image band with wavelength range from the visible spectrum to the infrared region, due to the great improvement in current decade [3]. Even though the large number of hyperspectral bands provides sufficient information to distinguish resources, they may bring some problems, like Hughes phenomena [4]. Besides, the development of the mass data also stress significant calculation power. As an outcome of the dimensionality reduction is one of the most essential preprocess steps in hyperspectral data analysis to deal with these issues. Band selection (BS) is an efficient approach for hyperspectral dimensionality reduction, which has been rewarded an increasing attention in current years. The existing BS method [1] [2] consist of two broad types, namely supervised and unsupervised methods respectively. The supervised methods have need of training samples that may be virtually not available [5]. Thus, this paper mainly focus on unsupervised BS method. The unsupervised technique can be implicited as the development of selecting a skilled division from a larger set of bands, without any prior knowledge. Various unsupervised BS methods are based on information estimate resources. Their aim is to decide the subset with huge information [9]–[11], low similarity [5]. In order to choose the representative instead of extreme bands, currently, the researchers take BS as a clustering problem, i.e., a process of partitioning the bands into group of similar clusters. In these conditions, the cluster centers are generally considered preferable to insignificant bands. Based on different distant measures, like interquartile range, correlation coefficient and covariance, Ahmad estimated cluster and select the bands by corresponding different k-means version. These clustering-based methods can be hidden as the direct application of the clustering methods to hyperspectral bands. Though the centers of the clusters seem to be a best option these methods undergo from large computation complexity. Moreover, these methods may be greatly inclined by many clusters. Finally, these clustering methods require spherical distribution of the data since a data point is assigned to the closest center. In this paper, we choose the cluster centers without the actual clustering. Particularly, for each band, we use a pointer termed exemplar score (ES) to measure the possibility of a band to be an exemplar. The ES utilizes two reasonable assumptions namely, the exemplars contains maximum local density and they are at a comparatively great distance from points of higher density. Based on ES, we here represent a fast BS method, i.e., Exemplar Based Band Selection (EBBS), which aims to select the bands with high chance to be exemplars (or high ES). EBBS does not involve actual clustering; as a substitute, it prioritizes the bands according to their ESs. EBBS has quite low computation complexity since it is actually a band-ranking method,. In accumulation, it is verified in the experiment, EBBS is able to identify nonspherical clusters. And also, EBBS has no distribution requirements of the data points.