Pecora 20 - Observing a Changing Earth; Science for Decisions— Monitoring, Assessment, and Projection November 13-16, 2017  Sioux Falls, SD COMPARISON OF DIFFERENT SIMILARITY MEASURES FOR SELECTION OF OPTIMAL, INFORMATION-CENTRIC BANDS OF HYPERSPECTRAL IMAGES Munmun Baisantry School of Computing and Electrical Engineering Dericks P Shukla School of Engineering Indian Institute of Technology Mandi Himachal Pradesh, India 175005 munmun_baisantry@students.iitmandi.ac.in dericks@iitmandi.ac.in ABSTRACT Hyperspectral images consisting of large number of spectral bands suffer from limitations like High data redundancy, curse of dimensionality (insufficient training samples), and high computational complexity. Therefore, dimensionality reduction & band-selection has become a common practice in the field of hyperspectral image processing. Graph-based band selection is a well-known technique which is based on spectral clustering on similarity matrix to select the optimal band set. Thus, choice of similarity/ affinity matrix is a vital decision in these methods. We have conducted a comparison of some well-known affinity matrices used in spectral clustering to divide graph to smaller sub-graphs (indicating subsets of bands). Comparison was done using various types of metrics like ACC, AIE and ARE. KEYWORDS: Hyperspectral, AVIRIS, Indian Pines, Band selection, Similarity matrices, Dimensionality reduction. INTRODUCTION Hyperspectral sensors, with their nanometric spectral resolution and large number of continuous bands, have proved to be an asset to the remote sensing community, gaining prominence in myriad applications like target detection, classification etc. Subtle spectral divergences between objects can be easily identified using hyperspectral images, helping to distinguish between them (Vorovencii, 2009), (Mohan et. al., 2015). For example, deep-water bodies, roads or shadows are often considered as same in multispectral images based on reflectance values but they can be very easily distinguished in hyperspectral images. This property is very useful in detecting targets such as minerals, small vehicles, military camps, buildings,and oil tanks etc. in images where there are a lot of background classes. With increasing dimensionality, the size of training samples required to estimate parameters also increases exponentially, a term characterized as curse of dimensionality (Kouiroukidis et. al., 2011 ).In high dimensional space, data is mostly clustered in small low-dimensional subspaces & structures instead of being uniformly distributed across the space. Most of the high dimensional space is sparse and data is far apart from each other (Li et. al.,2011). Thus, conventionally used distance measures are not indicative of the true distance between the data points and may not give factual clustering results. Moreover, hyperspectral bands are highly correlated so band selection and dimensionality reduction methods are applied to reduce redundancy and complexity associated with high dimensionality in various applications. Since hyperspectral bands are continuous in nature, it may seem that adjacent, contagious bands can be combined as one. However, band selection and dimensionality reduction is a data-dependent process and the assumption that the band similarity is location-dependent phenomena may not hold true in many cases. In our paper, we have discussed the popular, graph-based band selection method and demonstrated the effects of different similarity measures available on it. The band selection & dimensionality reduction of hyperspectral image using graph-based approach consists of threemajor steps: 1) Learning a similarity matrix 2) Estimating the number of optimal number of bands. 3) Clustering the similarity matrix using spectral clustering.