International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org

Volume 3, Issue 10, October 2014

ISSN 2319 - 4847

Dimensionality Reduction Techniques for

Hyperspectral Images
Shraddha P. Lodha1 and Prof. S. M. Kamlapur2
1

Student, M.E. Computer Engineering, KKWIEER, Savitribai Phule University of Pune, India

Department of Computer Engineering, KKWIEER, Savitribai Phule University of Pune, India

ABSTRACT
Hyperspectral Imaging produces an image where each pixel is having narrow spectral bands with plentiful spectral
information. Spectral bands refer to the large number of measured wavelengths bands of Electromagnetic Spectrum. The large
number of spectral bands in hyperspectral data increases the computational burden. So, dimensionality reduction through
spectral feature selection thoroughly affects the accuracy of the given task. A fuzzy rough set is an approach that deals with the
concepts of vagueness and indiscernibility. It finds feature subsets preserving the semantics of the given datasets. Therefore,
this paper proposes the applicability of Fuzzy-Rough Set Approach to select the most significant spectral features from the
hyperspectral data. Selected features are employed to build a more easy and understandable learning model in order to improve
the classification quality of hyperspectral images.

Keywords: Dimensionality Reduction, Fuzzy-Rough Sets, Hyperspectral Imaging, Spectral features

1. INTRODUCTION
Hyperspectral Imaging is a foremost research area in Remote Sensing. It combines the potentiality of digital imaging
and spectroscopy. Hyperspectral imaging collects information across the Electromagnetic Spectrum. This technology is
used largely in the Minerals, Coal, Oil and Gas Industry. Hyperspectral images provide researchers with a picture of the
surface of a surveyed area so they can deduce some conclusions. Our eyes only see visible band of the light spectrum.
Many more bands of non-visible light spectrum with higher and lower frequencies exist to the left and right of visible
spectrum. This is known as the Electromagnetic Spectrum and it includes the bands of ultra-violet, x-rays, radio waves,
micro waves and infra-red radiations as shown in fig. 1(a) [1]. Hyperspectral Imaging is a technology which divides
regions of those spectra which we cant see into many more bands and creates visible images from them. Earths
surface is overlaid with a variety of materials like soil, rocks, foliage and water, all composed of diversifying minerals
and chemical compounds. These materials all reflect only a chunk of the Electromagnetic Spectrum and therefore can
be identified by their unique Electromagnetic Signatures, see fig. 1(b). Hyperspectral sensors scan the sunlight
bouncing off the Earths surface as shown in fig. 1(c). By analyzing great portion of the Electromagnetic Spectrum
contained within that sunlight, the composition of the region can be determined by comparing against known spectral
signature of different minerals and compounds. Hyperspectral sensors take an advantage of passive sensors and hence,
only scan the available light bouncing off the surface of the surveyed region. The sensors are usually mounted on either
airplanes or low orbit satellites that scan and record area of the Earths surface, refer fig. 1(d). Results from
hyperspectral scans are compiled by computer software into data cubes. An example of Hyperspectral data cube (3-D
cube) which is further known as hyperspectral image data cube is shown in fig. 1(e) [1]. Thus, every pixel in the
hyperspectral image consists of a contiguous reflectance spectrum. That means each pixel contains contiguous bands of
electromagnetic spectrum from visible to infrared regions and can be used to identify the objects in the scene since it
provides much more detailed information about the scene than the normal color camera. In conjunction with other
imaging technologies, hyperspectral spectral imaging can yield much more detailed information about the scene or the
surveyed area. Thus, hyperspectral imaging leads to an extremely enhanced ability to classify the objects in the scene
based on their spectral properties [1]. However, the classification of hyperspectral image is a very challenging task in
recent days due to the huge number of band values for each pixel. Hence, dimensionality reduction becomes a crucial
step for successfully implementing hyperspectral image classification.

Volume 3, Issue 10, October 2014

Page 92

International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 10, October 2014

ISSN 2319 - 4847

Figure 1 The concept of hyperspectral imaging technology which consists of (a) the electromagnetic spectrum, (b)
reflectance of a chunk of the electromagnetic spectrum from the materials on the Earths surface, (c) hyperspectral
sensors which scan the sunlight bouncing off the Earths surface, (d) sensors mounted on airplanes or low orbit
satellites and (e) hyperspectral image cube.

2. DIMENSIONALITY REDUCTION IN HYPERSPECTRAL IMAGES

Hyperspectral image has very strong spectral correlation that means there is often a significant amount of redundant
and misleading information present in the image and such information needs to be removed before any further
processing can be carried out [2],[3],[4], and [6].For this reason, it is often useful to reduce the data dimensionality.
Dimensionality reduction techniques can be a transformation-based or a selection-based [5]. The key distinction
between these two is whether a dimensionality reduction technique transforms or preserves the meaning of the original
dataset in the process of reduction. The choice of dimensionality reduction technique is often guided by the particular
problem at hand.
2.1 Transformation-based reduction
Transformation-based reduction approach is a semantics destroying dimensionality reduction technique (referred to as
feature extraction) which in the process irreversibly transforms the original dataset features [5]. These methods are
commonly employed in the situations where the semantics of the original datasets are not needed by further process.
Many approaches such as principal component analysis [7], minimum noise fraction transform [10], discriminant
analysis [11], [15], non-parametric weighted feature extraction [16], [17], wavelet transform [19], and spectral mixture
analysis [18] have been used for feature extraction in hyperspectral images in order to reduce data redundancy or to
extract specific information. This section briefly discusses several such popular techniques that have been used for
hyperspectral images. Following Table 1 give a summary about popular feature extraction methods used for
hyperspectral images.
Table 1: A summary of feature extraction methods for hyperspectral images
Feature Extraction Methods
Characteristics
Principal Component Analysis (PCA)

Volume 3, Issue 10, October 2014

In hyperspectral imaging, PCA is based on the fact that neighboring

bands are highly correlated which often convey almost the same info
about the object. Hence, the analysis is used to transfer higher
dimensional hyperspectral data to lower dimensional hyperspectral
data removing the correlation among the bands. It employs statistic
properties of hyperspectral image [7].
Merits:
Most of the information may exclusively be contained in the
first few PCA bands
As number of PCA bands increases, the information content of
the PCA bands reduces
Remaining bands contain noisy information
Demerits:
Only the subsets of the original bands are used for the analysis

Page 93

International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 10, October 2014

ISSN 2319 - 4847

and PCA is performed on these subsets

PCA ca not eliminate anomalies that can be seen in one
arbitrary band
PCA transforms an image into a new coordinate system
without considering factors such as noise
Minimum Noise Fraction (MNF)

MNF is a data transformation technique [10]. In hyperspectral image

processing, MNF is commonly used to align the data along axes of
decreasing signal to noise ratio (SNR). Since, hyperspectral images
tend to contain more noise, MNF is a technique to address the noise
present in the image. MNF maximizes the estimated SNR. Few noise
filters are used like Adaptive Filters (AF), etc.
Merits:
MNF of an image (i.e. higher order bands in hyperspectral
image) can be truncated incorporating low SNR that
eliminates much of the noise without losing much
information
Demerits:
If truncation consists of too many bands, too much noise is left
in the image
If truncation consists of too few bands, useful signal may be
excluded from the resulting image

Discriminant Analysis

Linear Discriminant Analysis (LDA): Fishers linear discriminant

analysis is well suited when data is normally distributed. It is a
parametric feature extraction method. LDA is based on following three
scatter matrices: within-class, between class and mixture (or total)
scatter matrices which can be used to formulate criteria of class
separability. Transforms the data from a high dimensional space into
low dimensional subspace with maximized class separability. In
hyperspectral images, transformation is done using the ratio of number
of training samples and spectral features [10], [15].
Merits:
It is distribution-free
Fast and easy to apply
Gives better within class scatter matrices
Demerits:
The rank of between-class scatter matrix is the number of
classes minus one. Therefore, the maximum number of
features that can be extracted remains the same which may
not be sufficient for achieving better accuracy in practical
application
Cannot perform well for non normally distributed data
Sensitive to outliers
In hyperspectral image processing, the ratio of the number of
training samples and features is small due to the fact that
labeling is expensive and that the feature space has high
dimensionality. Hence, it suffers from the singularity
problem since within class covariance is singular

Non-parametric feature extraction

It is commonly used for high dimensional multiclass pattern

recognition problem. Based on non-parametric extension of scatter
matrices. To overcome the limitations of LDA non-parametric feature
extraction algorithms are developed like nonparametric discriminant
analysis (NDA) based on nonparametric between class scatter matrix,
nonparametric weighted feature extraction (NWFE) with fully
nonparametric within-class and between-class scatter matrices,
nonparametric weighted feature extraction (NWFE), some
regularization techniques to solve singularity problem, and
nonparametric fuzzy feature extraction (NFFE) which introduces fuzzy
membership grades to design its within class and between-class scatter
matrices [16], [17].
Merits:
Suitable for handling linear problems
employment of a suitable regularization

Volume 3, Issue 10, October 2014

Page 94

International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 10, October 2014

ISSN 2319 - 4847

Adjustment of the extracted features can be done

Demerits:
NDA becomes hardy on unequal covariance or complexly
distributed data
May not be suitable for non-linear problems. To extract the
non-linear features kernel version of the algorithms have
been used

Linear Spectral Mixture Analysis

(LSMA)

Linear Spectral Mixture Analysis is widely used to address the mixed

pixel problem. In hyperspectral image analysis, LSMA is commonly
used for subpixel detection and mixed pixel classification. It is based
on the assumption that the image pixel is linearly mixed by materials
(referred to as targets). It determines the component parts of mixed
pixels and identifies the proportion of a pixel which belongs to a
particular class or feature based on the spectral characteristics of the
endmembers where endmembers are the pure spectra corresponding
to each of the classes [19].
Demerits:
Sensitive to target knowledge and noise
Abundance fraction of targets present in an image are assumed
to be unknown that needs to be estimated
To address these issues some algorithms are modified and developed
like Independent component analysis (ICA) based linear spectral
random mixture analysis (LSRMA) assuming abundance fraction as
random.

Wavelet-based Methods

Wavelet-based reduction for hyperspectral imagery is based on the fact

that the spectral data from original feature space is transformed to a
reduced feature space. The main principle of this method is to apply
Discrete Wavelet Transform (DWT) to hyperspectral data in spectral
domain and at each pixel location. Here, wavelet reduced data
represent a spectral distribution similar to the original distribution but
in compressed form [18].
Demerits:
Lossy compression
Removed high frequency signal may contain useful
information for class separation and identification

Thus, two major issues arise from such transformation-based dimensionality reduction techniques. One is the number
of data dimensions required for dimensionality reduction to avoid significant loss of information. The other is that,
since the data are transformed, some crucial and critical information may have been compromised and distorted. Hence,
the transformation-based dimensionality reduction technique usually changes the physical meaning of the original data
[4], [2].
2.2 Selection-based reduction
Selection-based reduction approach is a semantics preserving technique (referred to as feature selection or band
selection in hyperspectral images) which attempt to retain the meaning of the original dataset. The main aim of
selection-based dimensionality reduction technique is to find a minimal subset of the original data without losing their
physical meaning [4], [2]. In hyperspectral images, to reduce the dimensionality of hyperspectral data, band selection
technique is commonly used. Band selection in hyperspectral images refer to select a subset of original bands
containing most useful information about the object, removing redundant, irrelevant or misleading spectral features
from the hyperspectral data. According to the availability of class information, band selection is categorized into two
types: unsupervised band selection method and supervised band selection method [6].
2.2.1 Unsupervised band selection
When class information is not known in advance, unsupervised band selection has to be adopted to find the most
informative and distinctive bands. Unsupervised band selection methods can be categorized based on the following
classification as shown in Table 2.

Volume 3, Issue 10, October 2014

Page 95

International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 10, October 2014
Categories

ISSN 2319 - 4847

Table 2: Unsupervised band selection methods for hyperspectral images

Characteristics

Methods based on Information Theory

It includes the use of information theory like

mutual information. Provides a methodology to
find generalized correlation among image bands.
Divergence measures can be used to compare
different probability distributions [14].

Methods based on Constrained Band Selection (CBS)

It is different from variance-based methods or

information theoretic criteria-based methods,
where this approach linearly constrains a band to
minimize the correlation or dependency of this
particular band from the other bands in a
hyperspectral image [13], [14].

Methods based on eigenanalysis and decorrelation of the input

data

Maximum Variance Principal Component

Analysis (MVPCA) is one of the technique based
on eigenanalysis and decorrelation of input data.
In this, a joint band- prioritization and banddecorrelation
approach
is
adopted
for
hyperspectral image classification. The band
prioritization depends on an eigenanalysis,
decomposing a matrix into an eigen form matrix
from which a loading factors matrix has been
constructed and used to prioritize-bands [12], [14].

Methods based on clustering strategy

Clustering process strategy refers to group bands

that minimizes the intra-cluster variance and
maximizes the inter-cluster variance in terms of
dissimilarity measures based on band information
content such as mutual information, etc. [9].

2.2.2 Supervised band selection

When class information is already known supervised band selection is applied to preserve the desired object
information. Supervised feature selection methods involve two major steps: subset generation and subset evaluation
[23]. Subset generation is nothing but a process of heuristic search, where each point in the search space, specifies a
candidate subset for evaluation. Band selection in hyperspectral images requires an efficient searching strategy [21],
[22]. Optimal search techniques such as Exhaustive Search or Branch and Bound are not applicable for hyperspectral
images due to the huge amount of complexity, since in these techniques all possible band combinations are tested and
then optimal one will be selected. Hence, to avoid this, different suboptimal techniques, such as SFS, SBS, SFFS and
SBFS are used [21]. Various stochastic optimization based feature (band) selection techniques for hyperspectral images
have been used such as Genetic Algorithms (GA) [25], Particle Swarm Optimization (PSO) [24], [26] and Ant Colony
Optimization (ACO) [20]. In subset evaluation, the fitness function (evaluation criteria) of each subset of feature is
calculated. Based on different evaluation criteria, supervised band selection methods regardless of their searching
strategy stage, can be classified into two approaches: filter approach and wrapper approach [3].
Filter Approach: Band selection method based on filter approach is dependent on general characteristics of the data to
evaluate and select feature (band) subsets without involving any classification algorithm so that undesirable features are
omitted before the classification process begins. It uses some statistical methods (i.e. discrimination measures) such as
the Mahalanobis Distance [26], Bhattacharyya Distance [26], Jeffreys-Matusita Distance [27] to identify an optimal
subset of bands in hyperspectral image.
Wrapper Approach: Band selection method based on wrapper approach is dependent on one predetermined
classification algorithm and uses its performance as an evaluation criterion. It determines features (bands) which better
suit the classification algorithm in terms of performance. Feature Shaving is an example of this type which removes the
whole band at once including extra parameters such as maximum size of smoothing window, a few runs of whole
procedure, an expert knowledge, etc. [8].

3. PROPOSED SYSTEM
Rough Set Theory (RST) introduced by Pawlack has been widely used as a selection tool to discover data dependencies
and to reduce number of attributes contained in a dataset using the data alone and no additional information. It deals
with incomplete, imperfect data, and data vagueness. A Rough Set is defined by two approximation sets called as lower

Volume 3, Issue 10, October 2014

Page 96

International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 10, October 2014

ISSN 2319 - 4847

approximation sets and upper approximation sets. Lower approximation sets consists of the set of objects that are
certainly belong to the subset of interest, while upper approximation sets consists of the objects that are possibly belong
to the subset of interest. Given a dataset with discretized attribute values, then using RST it is possible to find a subset
(termed a reduct) of the original attributes that are most informative and all other attributes can be removed from the
dataset with minimal information loss. However, the dataset containing that the values of attributes may be both crisp
and real-valued, and this is where traditional RST encounters a problem. Therefore, it is desirable to develop
techniques to provide the means of data reduction for crisp and real valued attribute datasets and this can be achieved
through the use of fuzzy-rough sets [5]. Fuzzy rough set approach combines the rough set and fuzzy set to tackle the
weakness of the classical rough set in order to deal with real value data without discretization. Fuzzy-rough sets abridge
the related but distinct concepts of vagueness and indiscernibility both of which occur as a result of uncertainty in
knowledge. Thus, it preserves the meaning of the data knowledge. In addition, it does not require additional
information for the process of feature selection. Feature selection based on fuzzy rough sets particularly depends on
defining a fuzzy-similarity relation between objects within a subset of features. Using fuzzy similarity relation, fuzzy
concepts can be approximated by defining the fuzzy lower approximation sets and fuzzy upper approximation sets [5],
[28]. Let, a universe U as a finite non empty set of objects each defined by a set of attributes, denoted by A. Here, the
pair (U, A) is known to be an information system (IS) where for every subset P A there exist an associated similarity
relation.
denotes the similarity between objects x and y induced by the subset of features p. Given
, can be
approximated by the information contained in P constructing the P-lower and P-upper approximations of as defined
in Equations (1) and (2) respectively [28].

(1)
(2)
Where I denotes the fuzzy similarity implicator, T is the t-norm, and
denotes the fuzzy similarity relation induced by
the subset of features P. The main issue of fuzzy rough set approach is to select a subset of features called reduct R that
represents the original set of features. So feature selection based on fuzzy rough set approach search for the optimal
subset of features that approximately has either the same certainty or uncertainty of the original set of features. It
depends on finding the degree of dependency between the objects which can be computed using the lower
approximations. Let degree of dependency of object Q on a subset of features P denoted as
and can be computed
using Equation (3) [28].
(3)
Thus, in the proposed system, fuzzy-rough set approach is used for spectral feature selection to reduce the
dimensionality of the original hyperspectral data. The objective of the proposed system is twofold. First, select spectral
features from the original hyperspectral data which in turn will reduce the number of bands in hyperspectral image.
Second, use selected features for hyperspectral image classification. Fig. 2 shows the flow chart for the proposed
system.

Figure 2 Flow chart of the proposed system

4. CONCLUSIONS
This paper provides a survey of various dimensionality reduction techniques in terms of feature extraction and band
selection methods for hyperspectral images. In feature extraction methods, original data are transformed and compact
set of original features are generated. Commonly used feature extraction methods used such as principal component

Volume 3, Issue 10, October 2014

Page 97

International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 10, October 2014

ISSN 2319 - 4847

analysis, minimum noise fraction transform, discriminant analysis, non-parametric weighted feature extraction, wavelet
transform, and spectral mixture analysis have been discussed for hyperspectral images. In band selection, a subset of
significant features is selected for preserving the meaning of the original dataset. Supervised and unsupervised
approaches for band selection have been carefully studied. A new framework for dimensionality reduction have been
proposed based on fuzzy rough sets which selects important features with no additional information.

References
[1] http://www.youtube.com/watch?v=17cRTkG31_w
[2] Jun Li, Prashanth Reddy Marpu, Antonio Plaza, Jose M. Bioucas-Dias, and Jon Atli Benediktsson,Generalized
Composite Kernel Framework for Hyperspectral Image Classification, IEEE transactions. on geoscience and
remote sensing, vol. 51, no. 9, September 2013.
[3] B. Mojaradia, h.emamib, m.varshosazc and s. Jamali,"A Novel Band -Selection Method for Hyperspectral Data
Analysis," The international archives of the photogrammetry, remote sensing and spatial information sciences.
Vol. Xxxvii. Part b7. Beijing 2008.
[4] Kun tan, Erzhu Li, Giandu, and Peijun Du,Hyperspectral Image Classification Using Band Selection And
Morphological Profiles, IEEE journal of selected topics in applied earth observations and remote sensing, vol. 7,
no. 1, January 2014.
[5] Richard Jensen and Giang Shen, Computational Intelligence and Feature Selection: Rough and Fuzzy Approaches.
[6] Qian du and He Yang,"Similarity-based Unsupervised Band Selection for Hyperspectral Image Analysis," IEEE
geoscience and remote sensing letters, vol. 5, no. 4, October 2008.
[7] Craig Rodarmel and Jieshan,"Principal Component Analysis for Hyperspectral Image Classification," Surveying
and Land Information Systems, vol. 62, no. 2, 2002.
[8] Sergueiverzakov, Pavelpaclk, and Robert P.W. Duin,"Feature Shaving for Spectroscopic Data," Springer-Verlag
Berlin Heidelberg 2004.
[9] Adolfo Martinez-Uso, Filibertopla, Jose MartinezSotoca and Pedro Garcia-Sevilla,"Clustering-based Hyperspectral
Band Selection Using Information Measures," IEEE transactions on geoscience and remote sensing, vol. 45, no.
12, December 2007.
[10] Rhonda D. Phillips and Layne T. Watson, Christine e. Blinn and Randolph H. Wynne,"An Adaptive Noise
Reduction Technique for Improving The Utility of Hyperspectral Data," Pecora17 The future of land imaging.
Going operational November 18 20, 2008.
[11] Qian Du, Chein-I Chang,"A Linear Constrained Distance-based Discriminant Analysis for Hyperspectral Image
Classification" Pattern Recognition 34 (2001).
[12] Chein-I chang, Qian Du, Tzu-Lung Sun, and Mark l. G. Althouse,"A Joint Band-Prioritization And BandDecorrelation Approach To Band Selection for Hyperspectral Image Classification" IEEE transactions on
geoscience and remote sensing, vol. 37, no. 6, November 1999.
[13] Chein-I Chang, and Su Wang,"Constrained Band Selection for Hyperspectral Imagery," IEEE transactions on
geoscience and remote sensing, vol. 44, no. 6, June 2006.
[14] Adolfo Martnez-Uso, Filibertopla, Jose M. Sotoca, and Pedro Garcsevilla,"Comparison of Unsupervised BandSelection Methods for Hyperspectral Imaging," Springer-Verlag Berlin Heidelberg, pp. 3038, 2007.
[15] Tatyana V. Bandos, Lorenzo Bruzzone and Gustavo Camps-Valls,"Classification of Hyperspectral Images With
Regularized Linear Discriminant Analysis," IEEE transactions on geoscience and remote sensing, vol. 47, no. 3,
March 2009.
[16] Jinn-Min Yang, Pao-Ta Yu, Bor-Chen Kuo, and Ming-Hsiang Su,"Nonparametric Fuzzy Feature Extraction for
Hyperspectral Image Classification," International Journal of Fuzzy Systems, vol. 12, no. 3, September 2010.
[17] Bor-Chen Kuo and David A. Landgrebe,"Hyperspectral Data Classification Using Nonparametric Weighted
Feature Extraction".
[18] Chein-I Chang, Shao-Shan Chiang, James A. Smith and Irving W. Ginsberg,"Linear Spectral Random Mixture
Analysis for Hyperspectral Imagery," IEEE transactions on geoscience and remote sensing, vol. 40, no. 2, February
2002.
[19] B.Salehia and M.J.Valadanzoej,"Wavele-based Reduction of Hyperspectral Imagery.
[20] Jianwei Gao, Gian Du, Lianru Gao, Xu Sun and Bing Zhang,"Ant colony optimization-based supervised and
unsupervised band selections for hyperspectral urban data classification," Journal of Applied Remote Sensing
085094-1 vol. 8, 2014.
[21] Aloke Datta, Susmita Ghosh and Asish Ghosh,"Wrapper-based Feature Selection In Hyperspectral Image Data
Using Self-adaptive Differential Evolution," 2011 International Conference On Image Information Processing
(iciip 2011).
[22] P. A. Devijver and J. Kittler,"Pattern Recognition: A Statistical Approach," Prentice-Hall int., first ed., 1982.

Volume 3, Issue 10, October 2014

Page 98

International Journal of Application or Innovation in Engineering & Management (IJAIEM)

Web Site: www.ijaiem.org Email: editor@ijaiem.org
Volume 3, Issue 10, October 2014

ISSN 2319 - 4847

[23] A. Jain and D. Zongker, Feature Selection: Evaluation, Application, and Small Sample Performance, IEEE
transaction on pattern analysis and machine intelligence, vol. 19, pp. 153--189, 1997.
[24] S. Ding and L. Chan,Classification of Hyperspectral Remote Sensing Images With Support Vector Machines And
Particle Swarm Optimization, in proc. Of the int. Conf. on information engineering and computer science
(iciecs09), pp. 1--5, 2009.
[25] J. P. Ma, Z. B. Zheng, Q. X. Tong, and L. F. Zheng,An Application of Genetic Algorithms On Band Selection
for Hyperspectral Image Classification, in proc. of the 2nd int. Conf. on machine learning and cybernetics, pp.
2810--2813, 2003.
[26] P. A. Devijver and J. Kittler, Pattern Recognition: A Statistical Approach. Prentice-Hall int., first ed., 1982.
[27] L. Bruzzone and S. B. Serpico, A Technique for Feature Selection in Multiclass Problems, int. Journal of
remote sensing, vol. 21, no.3, pp. 549--563, 2000.
[28] Basem O. Alijla, Ahamad Tajudin Khader, Lim Chee Peng, Mohammed Azmi Al-Betar and Wong Li Pei,
Fuzzy- rough set approach for selecting the most significant texture features in mammogram images, 2013
Palestinian International Conference on Information and Communication Technology.

AUTHOR
Shraddha P. Lodha received the B.E. degree in Computer Engineering from the K. K. Wagh Institute of Engineering
Education and Research, Savitribai Phule University of Pune, India in 2013 and she is currently pursuing the M.E.
degree in Computer Engineering at K. K. Wagh Institute of Engineering Education and Research, Savitribai Phule
University of Pune, India.

Volume 3, Issue 10, October 2014

Page 99