A Neural Network Approach to Identify Hyperspectral Image Content

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 8, No. 4, August 2018, pp. 2115~2125
ISSN: 2088-8708, DOI: 10.11591/ijece.v8i4.pp2115-2125  2115
Journal homepage: http://iaescore.com/journals/index.php/IJECE
A Neural Network Approach to Identify Hyperspectral
Image Content
Puttaswamy M. R.1
, Balamurugan P2
1
Research and Development Center, Bharathiar University, India
2
Department of Computer Science, Government Arts College, India
Article Info ABSTRACT
Article history:
Received Nov 26, 2017
Revised Feb 16, 2018
Accepted Mar 30, 2018
A Hyperspectral is the imaging technique that contains very large dimension
data with the hundreds of channels. Meanwhile, the Hyperspectral Images
(HISs) delivers the complete knowledge of imaging; therefore applying a
classification algorithm is very important tool for practical uses. The HSIs
are always having a large number of correlated and redundant feature, which
causes the decrement in the classification accuracy; moreover, the features
redundancy come up with some extra burden of computation that without
adding any beneficial information to the classification accuracy. In this
study, an unsupervised based Band Selection Algorithm (BSA) is considered
with the Linear Projection (LP) that depends upon the metric-band
similarities. Afterwards Monogenetic Binary Feature (MBF) has consider to
perform the „texture analysis‟ of the HSI, where three operational component
represents the monogenetic signal such as; phase, amplitude and orientation.
In post processing classification stage, feature-mapping function can provide
important information, which help to adopt the Kernel based Neural Network
(KNN) to optimize the generalization ability. However, an alternative
method of multiclass application can be adopt through KNN, if we consider
the multi-output nodes instead of taking single-output node.
Keyword:
Band selection algorithm (BSA)
Hyperspectral image (HSI)
Linear projection (LP)
Monogenetic binary feature
(MBF)
Neural network (NN)
Copyright © 2018 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Puttaswamy M. R.,
Research and Development Center,
Bharathiar University,
Coimbatore, Tamil Nadu 641046, India.
Email: mrp.gowda@gmail.com
1. INTRODUCTION
The HSI consist of very large dimension data with the hundreds of channels that ranging from the
short infrared wave to the visible region at „electro-magnetic (EM) spectrum‟ [1]. A hyperspectral is the
imaging technique, which acquires the objects information based upon their EM-spectrum with the
wavelength of 400nm to 2500nm. Meanwhile, the HIS delivers the complete knowledge of imaging, thus the
several of application like as; material identification, target detection and object discovering has reported
in [2]. Information extraction is very significant process in HSI; therefore applying a classification algorithm
is very important tool for practical uses.
However, the HSI are always having a large number of correlated and redundant feature, which
causes the decrement in the classification accuracy [3]; moreover, the features redundancy come up with
some extra burden of computation that without adding any beneficial information to the classification
accuracy. Hence, HSI data processing (that contain high volume data) become somewhat difficult, especially
with the supervised learning method because the accuracy of classification decreases with specific set of
training as increasing features number, this called as „Hughes Occurrence‟. In order to achieve higher
classification accuracy, dimensionality reduction approach becoming very beneficial and it also reduces the

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 8, No. 4, August 2018 : 2115 – 2125
2116
requirement of data storage and computational time [4]. In [5], they concluded that reducing in number of
features tends to achieve better classification accuracy.
In dimensionality reduction, the most commonly applied technique is subjected to this paper is
feature selection (FS), which chooses an important sub-set feature from the main feature set and dispose the
remaining features. Another approach is feature extraction (FE), which generally extracts the important
properties from a feature set and then transforms the „main data‟ to create that more separable. Considering
both the feature approach, a feature selection is most suitable for the „real dataset‟ with respect to feature
extraction. However, the HSI provides a comprehensive separation of analogous surface materials, this kind
of spectral feature are specially correlate with the adjacent bands, hence it provide a redundant
information [6]. A Hughes effect [7], is related to low generalization ability of the „classifier‟, which
frequently has encountered in several „pattern recognition‟ applications such as; object recognition, text
categorization, computer vision and gene expression data [8], [9]. There are several feature techniques such
as „MNF‟ (minimum-noise fraction [10]), „PCA‟ (principal component analysis [11]), „SPP‟ (sparsity
preserving projection [12]), „LPP‟ (local preserving projection [13]), „MSME‟ (multi-structure manifold
embedding [14]), and „SPA‟ (sparsity preserving analysis [15]), etc. Though the some important information
at feature approach has not obtained properly, therefore it causes the performance degradation in HSI
classification. Therefore, it is imperious to develop an efficient and new feature selection technique, which
integrates the spectral band selection and classification to remotely sense the HSI.
The dimensionality can be achieve through „band selection approach‟ that is BSA; there are two
types of BSA (i.e., supervised and unsupervised BSA). Here, we are performing unsupervised BSA, which is
used to get the superlative information bands without having any information of objectives. In this study, an
unsupervised based BSA is considered with the LP that depends upon the metric-band similarities.
Afterwards MBF has consider to perform the „texture analysis‟ of the HSI, where three operational
component represents the monogenetic signal such as; phase, amplitude and orientation. In post processing
classification stage, Neural Network (NN) performs the classification process after the feature extraction
from MBF.
The NN comprises one hidden layer and one output layer; the weight assignment is perform
randomly in between „Input of NN‟ and hidden layer. Feature mapping function can provide important
information, which help to adopt the kernel based NN (KNN) to optimize the generalization ability [16].
However, an alternative method of multiclass application can be provide through KNN, if we consider the
multi-output nodes instead of taking single-output node. The two HSI data set namely as Salinas-scene [17]
and Pavia University [17], we are going to use throughout this paper to validate our classification results with
respect other state-of-art methods. This paper is organized as follows; Section 2 provide the detailed survey
of state-of-the-art, Section 3 gives the proposed methodologies knowledge, Section 4 provides the
experimental result and analysis of classification. Finally, Section 5 concludes this paper.
2. LITRATURE SURVEY
Analysis of HSIs task is difficult because the data sets which are having extremely large
dimensionality, so choose some essential features it is most significant, and helpful for learning. In order to
select essential features of heterogeneous, in [18] they has proposed a method called sparse feature selection
method which is built on regularized regression model. However, in this presentation with the noise
information are contaminated because of imaging devices. That noise will affects the process of learning for
example hyperspectral images of high dimensional data analysis. Redundancy reducing as well as preserving
data these two critical problems, which are necessarily to be handled: at the feature selection. In [19], based
on a recently designed memetic procedure they proposed a feature selection technique for hyperspectral
image classification. They designed a suitable objective task inside the proposed technique, that able be
measure the enclosed redundancy info as well as essential info on the selected feature subsets. In [20], they
selected no redundant information plus Gabor features for hyperspectral image classification and proposed a
Markov blanket based symmetrical uncertainty approach.
A fast forward (FS) approach [21], which based upon Gaussian Mixture (GM) prototype classifier.
The GM-classifier has used for classifying the hyperspectral images, this approach choose the spectral
feature to increase the prediction rate of classification. This can execute through „k-fold‟ cross validation in
order to achieve efficient implementation and fast computing time. Initially, the GM can upgraded with
computed classification rate, instead of re-estimate overall model. Finally, GM marginalization can be sub-
model from the full learning model of spectral features. The investigation of the unsupervised-BSA [22] for
the spectral FS and the classification process based upon SVM (Support Vector Machine) [23], are mostly
used in morphological profiles („MP‟). The difficulty at high dimensionality can minimize by resulting
features and it is most usual activity where MP has extracted from PCA.

Int J Elec & Comp Eng ISSN: 2088-8708 
A Neural Network Approach to Identify Hyperspectral Image Content (Puttaswamy M. R.)
2117
The main challenge of classification problems is selection of a subset, which has efficient and
minimum from the enormous amount of features. The entire feature selection methods consist of a search
algorithm used for selection of high time consuming finest candidate from the possible answers. In [24],
proposed dual feature selection approaches without requirement of search algorithm. With using mean values
plus simple calculation of standard deviation effective features of subset are selected from the method. In
[25], [26], they expressed two class classification issue, where modified (KNN) k nearest neighbors classifier
method has been applied to create validity rating, maximal coherence, plus categorize the test samples by
k-ford validation. In [27], they emphasis on the HIS classification data, which captured from the
Environmental Program (EnP). Moreover, they considered the dataset that presented with EnP contest, a
standard has recently initialized with several classifying object and the land use classes depends upon
hyperspectral data. In this [28], they has considered „Radial Basis Function‟ (RBF), where full bandwidth-
RBF kernel has used as feature weights, whenever the feature values become rescaled through the z-scores
value.
The methodology of HIS allows recognizing materials through using photo-thermal „infrared
spectroscopy‟ [29]-[30]. In order to obtain the infrared spectroscopy, an infrared camera capture the hundreds
of images of an object at different channels wavelength simultaneously while a „QCL‟ (quantum cascade
laser) causes the subjected material to be brightened. The main problems that effecting the classification
arrives from the disparaging ratio of high dimension HIS and the small size „training data‟ [31]. These kind
of problems causes advancement in machine learning approaches such as; multiple learning systems,
transductive learning, semi-supervised learning and active learning [32]-[34] and several methods of learning
depend upon the unlabeled samples [35], [36]. To overcome from such kind of issues, some strategy has
considered enhancing new similarity measurement for the band reduction and dimension reduction of
HIS [37], [38]. However, the high correlation between the bands has misidentify for designing new
techniques to reduce the data dimensionality, which including some of the methods that has great acceptance
such as; PCA [39], „linear discriminant analysis‟ (LDA) [40] or MNF [41]. It also has been investigated in
preceding works that HIS classification after the preprocessing process dimensionality reduction „or‟ band
selection are usually outperforms better classification with respect to traditional HSI data classification [42],
[43], which in order to reduce the computational complexity.
3. PROPOSED HYPERSPECTRAL CLASSIFICATION ALGORITHM
The presence of high spectral correlation in the HSI causes difficulty in feature selection process that
is why it‟s necessary to perform dimensional reduction at spectral feature in order to obtain a useful „compact
set‟ spectral feature. The dimensionality can be achieve through „band selection approach‟ that is BSA; there
are two types of BSA (i.e., supervised and unsupervised BSA). Here, we are performing unsupervised BSA,
which is used to get the superlative information bands without having any information of objectives. In this
study, an unsupervised based BSA is considered with the linear projection that depends upon the metric-band
similarities. This methodology does not require any information regarding prior-class to categorize the level of
data analysis. Here, we select the two different bands (i.e., ) from LP-BSA [44] and the subset of bands is
defined as;
* + (1)
The new selected subset of band is
( ) ( ) (2)
where, is a third band and above Equation (2) has repeated until third band is more than the subset of first
two bands, when this criteria has satisfy then only the Equation (2) is updated.
( ) ( ) (3)
Here, denotes the prediction band of LP by selecting first two bands (i.e., ) and the error in LP can be
minimalize through selecting parameters,
‖ ‖ (4)

 ISSN: 2088-8708
2118
In order to calculate the minimal square solution,
( ) (5)
Furthermore,
( ( )⁄ ) (6)
Where, denotes the „input data‟ in matrix format (i.e., matrix form). Here, first column contain single
value, second column contains all pixels values of band and, third column contains all pixels values of
band . Additionally, the represents a vector matrix that include the all pixel values of , the
bands and are compared with band, if the value of third band found more than the subset of band ( ) is
updated.
The MBF [45] has considered performing the „texture analyses of the HSI, where three operational
components represents the monogenetic signal such as; phase, amplitude and orientation. The operators of
MBF can be given as „MBF-P‟ (MBF-Phase‟), „MBF-A‟ (MBF-Amplitude) and „MBF-O‟ (MBF-Orientation).
This can be computed through [45], where parameters of local variable and the image intensity are integrated
to provide individual operator of monogenetic signal.
- ( ) [ ( ) ( ) ( ) ] (7)
- ( ) [ ( ) ( ) ( ) ] (8)
- ( ) [ ( ) ( ) ( ) ] (9)
The MBF-X represents the pattern number of MBF ( - * +), here, is pixel of monogenetic
signal that can be encoded with 2-bits ( ( ) ( )) . Each of the MBF operator feature can be used as
individually for the classification process, otherwise the combination of this can be used for further
classification process. In order to provide the optimized HSI, the single-vector histogram for the MBF-P,
MBF-A and, MBF-O is;
( ) {
* +
(10)
where, ( ) represents the ( ) histogram vector represents is feature map on the scale and
denotes a level of sub-region. Neural Network (NN) performs the classification process after the feature
extraction from MBF. The NN comprises one hidden layer and one output layer; the weight assignment is
perform randomly in between „Input of NN‟ and hidden layer. Weights of the linear output layer can be
determined through LRA (Linear Regression Analysis) [46], which will reduce the computation time/cost.
Feature mapping function can provide important information, which help to adopt the kernel based NN (KNN)
to optimize the generalization ability [16].
An alternative method of multiclass application can be provide through KNN, if we consider the
multi-output nodes instead of taking single-output node. The classifier of -class have -number of output
nodes, if the original label of class is , then the output vector with -output nodes is;
, - (11)
If we consider only element of output vector ( ) is one, then output vector is;
[ ] (12)
where, rests of the components are adjusting to zero and the multi-output nodes for the classification process
of NN can be providing as;
‖ ‖ ∑ ‖ ‖ (13)

2119
where user-specified parameter (H) provides an adjustment distance between the training error and the
separating margin. The output „weight vector‟ between output layer and hidden layer is denote as . Hidden
layer of output that corresponds to input sample ( ) is ( ),
( ) , where (14)
The „training error vector‟ ( ) of -output nodes w.r.t the ( ) training sample is,
[ ] (15)
„Karush–Kuhn–Tucker‟ approach (KKT) [47] is considered for training the KNN, which consider the dual
optimization problem;
‖ ‖ ∑ ‖ ‖ ∑ ∑ ( ( ) ) (16)
where, is weight vector that linking the hidden layer to corresponding output node.
, - (17)
The optimize solution can be provide through corresponding KKT-Approach, such as;
∑ ( ( )) (18)
where, is Lagrange multiplier [48] that corresponds to the ( ) training samples.
(19)
( ) (20)
In Equations (14) to (18) shows for the specific case of the multiple-output nodes and it also can be used for a
single-output node via setting value as one. Where, the Lagrange multiplier [48] for nonlinearity function
can be given as;
[ ] (21)
, - (22)
Therefore, we will consider the multiple-output nodes for the multi-class classifier; in that the hidden layer
matrix (H) size can be decide through the -number of training samples. Q is number of the hidden nodes that
are unrelated to output nodes number (i.e., number of classes). Here, we consider the circumstance where
training samples number is very more than the feature space dimensionality. The number of training
samples , we have the alternative solution from Equations (18) and (19);
, - (23)
, - (24)
From Equation (20) we can write it as;
, -
( , - ) (25)
[ ] (26)

 ISSN: 2088-8708
2120
For this circumstances, a decision function of linear feature space „or‟ the output functionality of Kernel based
NN-classifier is,
( ) ( ) ( ) [ ] (27)
Where, is the „Moore Penrose‟ generalize inverse matrix [49], [50] and matrix , - , - . This
proposed model can be used for different size of applications, however, these approaches have various
computational costs, and their efficiency may vary for the different applications. It found though the KNN
implementation, that the KNN performance generalization is not much sensitive towards the feature space
dimensionality ( ) and the better performance can be achieved as long as „ ‟ is more enough. Therefore, if the
data-set of training are very more (i.e., ) then a preferable choice is to apply Equation (27) in order to
decrease the cost of computation.
4. EXPERIMENTAL RESULTS AND ANALYSIS
4.1. Data-set description
In this paper, we will use the real HSI dataset, which is available for general-purpose use. All the
experiments has performed using MATLAB simulation on an Intel-i5 2.80GHz machine, 64bit operating
system with 8GB RAM. The two HIS data-set namely as Salinas-scene [17] and Pavia University [17], we are
going to use throughout this paper to validate our classification results with respect other state-of-art methods.
At first Salinas scene dataset has collected of Salinas valley (which is situated at California) through the
„AVIRIS-sensor‟ with 224-bands. Moreover, it can be characterize by the high „spatial level resolution‟ that is
3.7mp; this dataset consist of many radiance data such as vineyard, bare soils, fields and vegetables. There are
total sixteen classes present in the ground truth of Salinas dataset.
The second consider dataset is PaviaU, which dataset has collected of Pavia University at Northern
Italy (which is during flight campaign) through the „ROSIS-sensor‟ with 103-bands. Moreover, it can be
characterize by the high „geometric resolution that is 1.3m with 610×610 pixels; there are total nine classes
present in the ground truth of PaviaU dataset. Table 1 show the total number of samples at each class of
Salinas Scene dataset [17] and it comprises of 54129 total samples. Similarly, Table 2 shows the total number
of samples at each class of PaviaU dataset [17] and it comprises of 42776 total samples.
Table 1. Total Samples at each Class of Salinas
Scene dataset [17]
CLASSES TOTAL SAMPLES
Class1 „Brocoli-weeds_1‟ 2009.0
Class2 „Brocoli-weeds_2‟ 3726.0
Class3 „Fallow‟ 1976.0
Class4 „Rough-Fallow‟ 1394.0
Class5 „Smooth-Fallow‟ 2678.0
Class6 „Stubble‟ 3959.0
Class7 „Celery‟ 3579.0
Class8 „Grapes-untrained‟ 11271.0
Class9 „ - ‟ 6203.0
Class10 „Corn-green-weeds‟ 3278.0
Class11 „Lettuce_romaine-4wk‟ 10683.0
Class15 „ - ‟ 7268.0
Class16 „ - ‟ 1807.0
TOTAL SAMPLES 54129.0
Table 2. Total Samples at each Class of Pavia
University dataset [17]
CLASSES TOTAL SAMPLES
Class1 „Asphalt‟ 6631.0
Class2 „Meadows‟ 18649.0
Class2 „Gravel‟ 2099.0
Class2 „Trees‟ 3064.0
Class2 „Metal sheets‟ 1345.0
Class2 „Bare Soil‟ 5029.0
Class2 „Bitumen‟ 1330.0
Class2 „Self-Blocking-Bricks‟ 3682.0
Class2 „Shadows‟ 947.0
TOTAL SAMPLES 42776.0
4.2. Comparative study
Here we compare the classification performance of several state of art with respect to our proposed
model. In this section, a proposed preprocessing MBF-Algorithm will compare with some existing popular
feature extraction methods such as MNF [10], PCA [11], SPP [12], LPP [13], MSME [14], and SPA [15]. The
kappa coefficient of classification (Kappa) and overall accuracy (OA) has used to show the accuracy after
classification.

2121
Taking a SalinasS dataset, the classification result of different method has compared with proposed
MBF has shown in Table 3. The MBF (PA) is compared with respect to MNF [10], PCA [11], SPP [12],
LPP [13], MSME [14], and SPA [15] method, based upon the Kappa (%), OA (%) and their individual class.
In Table 3, five classes got accuracy of 100% at MBF algorithm (classes number are 3, 4, 9, 11 and 16).
The overall accuracy of proposed MBF model has compared with respect to MNF [10], PCA [11], SPP [12],
LPP [13], MSME [14], and SPA [15] and we got improved classification results of 6.23%, 7.8%, 9.4%, 7.9%,
4.9%, and 9.4%. The Kappa accuracy of proposed MBF model has compared with respect to MNF [10],
PCA [11], SPP [12], LPP [13], MSME [14], SPA [15] and we got improved classification results of 7.03%,
8.75%, 10.57%, 8.8%, 5.5%, and 10.57%.
Table 3. Class wise Comparison on the dataset ( )
Class Number MNF [10] PCA [11] SPP [12] LPP [13] MSME [14} SPA [15] MBF (PA)
1 99.80 99.30 98.90 99.40 99.90 88.70 99.74
2 100.0 99.60 100.0 100.0 100.0 98.60 99.89
3 100.0 99.30 97.20 100.0 100.0 98.40 100.0
4 99.80 99.70 99.30 99.60 99.90 97.50 100.0
5 97.70 96.70 92.70 97.80 98.40 94.60 96.60
6 99.80 99.70 99.90 99.90 99.90 97.00 96.70
7 99.60 98.40 99.20 99.60 99.70 98.50 99.97
8 85.00 77.40 78.90 83.10 86.30 75.20 98.73
9 99.40 98.10 96.20 99.60 99.90 96.40 100.0
10 95.60 91.10 91.60 93.00 97.30 82.80 98.17
11 99.00 98.10 96.90 98.60 99.00 93.80 100.0
12 100.0 100.0 99.90 100.0 99.90 92.60 99.67
13 98.20 99.50 98.00 98.20 100.0 87.40 99.50
14 95.30 93.10 93.90 94.60 98.50 96.80 98.97
15 74.80 81.90 69.10 66.80 80.50 70.40 98.60
16 99.00 98.70 98.50 98.90 97.60 96.30 100.0
OA (%) 92.80 91.20 89.60 91.10 94.10 89.60 98.97
Kappa (%) 91.90 90.20 88.40 90.10 93.40 88.40 98.85
Table 4 shows the class wise classification results on a PaviaU dataset. Out of nine classes, our
proposed model MBF has more accuracy in five classes (class number 2,3,6,7 and 8). The overall accuracy of
proposed MBF model has compared with respect to MNF [10], PCA [11], SPP [12], LPP [13], MSME [14],
SPA [15] and we got improved classification results of 4.89%, 1.8%, 14.2%, 8.4%, 0.3%, and 15.8%.
The Kappa accuracy of proposed MBF model has compared with respect to MNF [10], PCA [11], SPP [12],
LPP [13], MSME [14], SPA [15] and we got improved classification results of 6.4%, 2.38%, 18.38%, 10.8%,
0.42%, and 20.7%.
Table 4. Class wise Comparison on the dataset ( )
Class Number MNF [10] PCA [11] SPP [12] LPP [13] MSME [14} SPA [15] MBF (PA)
1 87.30 85.50 76.80 84.10 92.30 69.20 90.47
2 86.70 93.20 76.70 81.30 94.40 82.40 99.58
3 84.30 89.00 79.60 80.80 78.10 71.10 93.88
4 92.70 93.40 93.20 92.60 94.00 88.70 66.68
5 100.0 99.80 99.80 100.0 100.0 92.50 90.80
6 92.80 91.70 83.90 92.50 93.90 78.70 99.90
7 97.40 97.50 87.10 96.90 96.90 82.50 99.80
8 91.90 93.90 77.90 89.50 94.30 65.50 96.28
9 99.90 100.0 100.0 100.0 99.90 96.50 48.97
OA (%) 89.30 92.20 80.50 86.00 93.60 79.00 93.90
Kappa (%) 86.00 89.70 75.00 81.90 91.50 72.80 91.89
Classification results at 10-sample size of SalinasS-dataset has shown in Figure 1, where top line
mark at each bar represent the error bar. Proposed model is compared in terms of OA and Kappa (KA) with
MNF [10], PCA [11], SPP [12], LPP [13], MSME [14], and SPA [15]. As per Figure 1, MBF is compared in
terms OA (%) and we got 11.43%, 9.9%, 15.89%, 15.69%, 13.16% and 11.13% improvement w.r.t considered
state of art. Similarly, MBF is compared in terms KA (%) and we got 12.7%, 10.97%, 17.68%, 17.47%, 14.3%
and 12.39% improvement w.r.t considered state of art. Figure 2 shows the classification results at 20-sample
size (SalinasS-dataset), here MBF is compared in terms OA (%) and we got 11.61%, 7.96%, 12.11%, 13%,
10% and 7.9% improvement w.r.t considered state of art. Similarly, MBF is compared in terms KA (%) and

 ISSN: 2088-8708
2122
we got 12.8%, 8.8%, 13.41%, 14.4%, 11.2% and 8.65% improvement w.r.t considered state of art. In Figure 3,
a proposed MBF algorithm performs 10.74%, 6.7%, 9.9%, 10.4%, 9%, 7.3% better overall accuracy than the
existing feature algorithm MNF [10], PCA [11], SPP [12], LPP [13], MSME [14], and SPA [15]; in terms KA
(%), we got 11.9%, 7.5%, 11%, 11.5%, 10% and 8.1% improvement. At 40-sampling size that shows in
Figure 4, MBF performs 10.4%, 5.3 %, 8%, 9.5%, 6.4%, 6.6% better overall accuracy than the considered
feature algorithm and in terms KA (%), we got 11.6%, 6.7%, 10%, 11.38%, 8.3% and 8.14% improvement.
As given in Figure 1 to Figure 4 for -dataset, similarly in Figure 5 to Figure 8 has given for
-dataset. Figure 5, shows the classification results at 10-sample size ( -dataset), here MBF is
compared in terms OA (%) and we got 25%, 14%, 41%, 34%, 19.5% and 33.6% improvement with respect to
considered state of art methods. Similarly, MBF is compared in terms KA (%) and we got 31%, 18%, 50%,
41.4%, 24% and 41.6% improvement w.r.t considered state of art. In Figure 6 and 7, the kappa coefficient of
MBF is 91.52 and 91.8, and the overall accuracy is 93.8 and 93.9. In Figure 8, a proposed MBF algorithm
performs 17%, 4.4%, 18.7%, 22%, 5.8%, 14.36% better KA (%) than the existing feature algorithm MNF [10],
PCA [11], SPP [12], LPP [13], MSME [14], and SPA [15]. The analysis of above result provides the
performance evaluation of MBF model, where the performance in every sampling size is increasing as per
increment in the training-sample-size.
Figure 1. Classification results at 10-sample size
( -dataset)
( -dataset)
( -dataset)
( -dataset)

2123
( -dataset)
( -dataset)
( -dataset)
( -dataset)
5. CONCLUSION
The HSI are always having a large number of correlated and redundant features, which causes the
decrement in the classification accuracy. Hence, HIS data processing becomes somewhat difficult, especially
with the supervised learning method because the accuracy of classification decreases with specific set of
training as increasing features number. In order to achieve higher classification accuracy, dimensionality
reduction approach becoming very beneficial and it also reduces the requirement of data storage and
computational time. In this paper, an unsupervised based BSA is considered with the LP that depends upon
the metric-band similarities. Subsequently MBF has considered performing the „texture analyses of the HSI,
where three operational components represents the monogenetic signal such as; phase, amplitude and
orientation. In post processing classification stage, the KNN has applied to optimize the generalization ability
by an alternative method of multiclass application through considering the multi-output nodes instead of
taking single-output node.
The proposed model classification accuracy has compared with the accuracy of several state-of-art
methods and from the above analysis of results our proposed model has perform considerably well in every
training set. Proposed model OA and Kappa values at 40-training samples is 98.97% and 98.85% at Salinas-S
dataset; similarly at Pavia-U dataset, we got 93.9% and 91.9% of overall and Kappa classification accuracy.
In future work, experiments should be implement on very large datasets in order to calculate the scalability of
several approaches.

 ISSN: 2088-8708
2124
REFERENCES
[1] A. Muntasa, I. A. Sirajudin, M. H. Purnomo, “Appearance global and local structure fusion for face image
recognition”, TELKOMNIKA (Telecommunication, Computing, Electronics and Control), 2011, vol. 9, no. 1,
pp. 125-132.
[2] H. Grahn and P. Geladi, “Techniques and applications of hyperspectral image analysis,” Chichester, England;
Hoboken, J. Wiley, 2007.
[3] K. Fukunaga and R. R. Hayes, “Effects of sample size in classiﬁer design,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 11, no. 8, pp. 873-885, 1989.
[4] M. Pal and G. M. Foody, “Feature Selection for Classification of Hyperspectral Data by SVM,” IEEE Transactions
on Geoscience and Remote Sensing, vol. 48, no. 5, pp. 2297-2307, 2010.
[5] A. Janecek, et al., “On the Relationship between Feature Selection and Classification Accuracy,” Journal of
Machine Learning Research - Proceedings Track, vol. 4, pp. 90-105, 2008.
[6] B. M. Shahshahani and D. A. Landgrebe, “The effect of unlabeled samples in reducing the small sample size
problem and mitigating the Hughes phenomenon,” IEEE Trans. Geosci. Remote Sens., vol. 32, no. 5,
pp. 1087-1095, 1994.
[7] G. Hughes, “On the mean accuracy of statistical pattern recognizers,” IEEE Trans. Inf. Theory, vol. 14, no. 1,
pp. 55-63, 1968.
[8] Z. Sun, et al., “Object detection using feature subset selection,” Pattern Recognit., vol. 37, no. 11, pp. 2165-2176,
2004.
[9] I. Guyon, et al., “Gene selection for cancer classiﬁcation using support vector machines,” Mach. Learn., vol. 46,
pp. 1-3, pp. 389-422, 2002.
[10] G. Lixin, et al., “Segmented minimum noise fraction transformation for efficient feature extraction of hyperspectral
images,” Pattern Recognit., vol. 48, no. 10, pp. 3216-3226, 2015.
[11] X. M. Cheng, et al., “A novel integrated PCA and FLD method on hyperspectral image feature extraction for
cucumber chilling damage inspection,”' Trans. ASAE, vol. 47, no. 4, pp. 1313-1320, 2004.
[12] L. Qiao, et al., “Sparsity preserving projections with applications to face recognition,” Pattern Recognit.,
vol. 43, no. 1, pp. 331-341, 2010.
[13] Z. Wang and B. He, “Locality perserving projections algorithm for hyperspectral image dimensionality reduction,”
in Proc. 19th Int. Conf. Geoinform., Shanghai, China, pp. 1-4, 2011.
[14] Y. Gan, et al., “Feature Extraction Based Multi-Structure Manifold Embedding for Hyperspectral Remote Sensing
Image Classification,” in IEEE Access, vol. 5, pp. 25069-25080, 2017.
[15] F. Luo, et al., “Fusion of graph embedding and sparse representation for feature extraction and classi_cation of
hyperspectral imagery,” Photogram. Eng. Remote Sens., vol. 83, no. 1, pp. 37-46, 2017.
[16] K. R. Muller, et al., “An introduction to kernel-based learning algorithms,” in IEEE Transactions on Neural
Networks, vol. 12, no. 2, pp. 181-201, 2001.
[17] http://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Scenes
[18] Q. Zhang, et al., “Automatic spatial– spectral feature selection for hyperspectral image via discriminative sparse
multimodal learning,” IEEE Transactions on Geoscience and Remote Sensing, vol. 53, no. 1, pp. 261-279, 2015.
[19] M. Zhang, et al., “Memetic algorithm based feature selection for hyperspectral images classification,” 2017 IEEE
Congress on Evolutionary Computation (CEC), San Sebastian, pp. 495-502, 2017.
[20] L. Shen, et al., “Discriminative Gabor Feature Selection for Hyperspectral Image Classification,” in IEEE
Geoscience and Remote Sensing Letters, vol. 10, no. 1, pp. 29-33, 2013.
[21] M. Fauvel, et al., “Fast Forward Feature Selection of Hyperspectral Images for Classification With Gaussian
Mixture Models,” in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 8,
no. 6, pp. 2824-2831, 2015.
[22] K. Tan, et al., “Hyperspectral Image Classification Using Band Selection and Morphological Profiles,” in IEEE
Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 7, no. 1, pp. 40-48, 2014.
[23] M. Pal and G. M. Foody, “Feature Selection for Classification of Hyperspectral Data by SVM,” in IEEE
Transactions on Geoscience and Remote Sensing, vol. 48, no. 5, pp. 2297-2307, 2010.
[24] M. Imani and H. Ghassemian, “Fast feature selection methods for classification of hyperspectral images,”
Telecommunications (IST), 2014 7th International Symposium on, Tehran, pp. 78-83, 2014.
[25] N. Zhang, et al., “An Enhanced K-Nearest Neighbor Classification Method Based on Maximal Coherence and
Validity Ratings,” 14th International Symposium, Japan, 2017.
[26] A. Woodley, et al., “Efficient Feature Selection and Nearest Neighbour Search for Hyperspectral Image
Classification,” 2016 International Conference on Digital Image Computing: Techniques and Applications
(DICTA), Gold Coast, QLD, pp. 1-8, 2016.
[27] S. Keller, et al., “Investigation of the impact of dimensionality reduction and feature selection on the classification
of hyperspectral EnMAP data,” 2016 8th Workshop on Hyperspectral Image and Signal Processing: Evolution in
Remote Sensing (WHISPERS), Los Angeles, CA, pp. 1-5, 2016.
[28] P. J. Hsieh, et al., “A nonlinear feature selection method based on kernel separability measure for hyperspectral
image classification,” 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Milan,
pp. 461-464, 2015.
[29] C. A. Kendziora, et al., “Advances in standoff detection of trace explosives by infrared photo-thermal imaging,”
Proceedings SPIE 7664, Detection and Sensing of Mines, Explosive Objects, and Obscured Targets XV, 2010.

2125
[30] F. Fuchs, et al., “Infrared hyperspectral standoff detection of explosives,” Proceedings SPIE 8710, Chemical,
Biological, Radiological, Nuclear, and Explosives (CBRNE) Sensing XIV, 2013.
[31] C. I. Chang, “Hyperspectral Imaging: Techniques for Spectral Detection and Classification,” 2003.
[32] J. Li, et al., “Semisupervised hyperspectral image classification using soft sparse multinomial logistic regression,”
Geoscience and Remote Sensing Letters, IEEE, vol. 10, no. 2, pp. 318-322, 2013.
[33] M. M. Crawford, et al., “Active learning: Any value for classification of remotely sensed data?” Proceedings of the
IEEE, vol. 101, no. 3, pp. 593-608, 2013.
[34] L. Bruzzone, et al., “A novel transductive SVM for semisupervised classification of remote-sensing images,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 44, no. 11, pp. 3363-3373, 2006.
[35] J. Li, et al., “Generalized composite kernel framework for hyperspectral image classification,” IEEE Transactions
on Geoscience and Remote Sensing, vol. 51, no. 9, pp. 4816-4829, 2013.
[36] D. Tuia, et al., “Active learning methods for remote sensing image classification,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 47, no. 7, pp. 2218-2232, 2009.
[37] C. M. Bachmann, et al., “Exploiting manifold geometry in hyperspectral imagery,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 43, no. 3, pp. 441-454, 2005.
[38] C. I. Chang, “Hyperspectral Data Processing: Algorithm Design and Analysis,” John Wiley & Sons, 2013.
[39] Q. Du and J. E. Fowler, “Hyperspectral image compression using jpeg2000 and principal component analysis,”
Geoscience and Remote Sensing Letters, IEEE, vol. 4, no. 2, pp. 201-205, 2007.
[40] G. Licciardi, et al., “Decision fusion for the classification of hyperspectral data: Outcome of the 2008 grs-s data
fusion contest,” IEEE Transactions on Geoscience and Remote Sensing, vol. 47, no. 11, pp. 3857-3865, 2009.
[41] J. Li and J. M. B. Dias, “Minimum volume simplex analysis: a fast algorithm to unmix hyperspectral data,” in
Geoscience and Remote Sensing Symposium, IGARSS 2008. IEEE International, vol. 3, pp. III250-III253, 2008.
[42] A. Plaza, et al., “Recent advances in techniques for hyperspectral image processing,” Remote Sensing of
Environment, vol. 113, pp. S110-S122, 2009.
[43] J. M. B. Dias, et al., “Hyperspectral remote sensing data analysis and future challenges,” Geoscience and Remote
Sensing Magazine, IEEE, vol. 1, no. 2, pp. 6-36, 2013.
[44] Q. Du and H. Yang, “Similarity-Based Unsupervised Band Selection for Hyperspectral Image Analysis,” in IEEE
Geoscience and Remote Sensing Letters, vol. 5, no. 4, pp. 564-568, 2008.
[45] M. Yang, et al., “Monogenic Binary Coding: An Efficient Local Feature Extraction Approach to Face
Recognition,” IEEE Transactions on Information Forensics and Security, vol. 7, no. 6, pp. 1738-1751, 2012.
[46] P. M. Arsad, et al., “Neural Network and Linear Regression methods for prediction of students' academic
achievement,” 2014 IEEE Global Engineering Education Conference (EDUCON), Istanbul, pp. 916-921, 2014.
[47] X. Hu and J. Wang, “A Recurrent Neural Network for Solving Nonconvex Optimization Problems,” The 2006
IEEE International Joint Conference on Neural Network Proceedings, Vancouver, BC, pp. 4522-4528, 2006.
[48] S. Zhang and A. G. Constantinides, “Lagrange programming neural networks,” in IEEE Transactions on Circuits
and Systems II: Analog and Digital Signal Processing, vol. 39, no. 7, pp. 441-452, 1992.
[49] D. Serre, “Matrices: Theory and Applications,” New York, Springer- Verlag, 2002.
[50] C. R. Rao and S. K. Mitra, “Generalized Inverse of Matrices and its Applications,” New York, Wiley, 1971.
BIOGRAPHIES OF AUTHORS
Puttaswamy M. R. received the BSc degree from the University of Mysore, India, in 2001 and the
MCA degree in Computer Applications from Visvesvaraya Technological University, India in 2005
and the M.Tech in Information Technology from Karnataka State Open University, India in 2010
and the PgCert in Higher Education Professional Practice from Coventry University, UK in 2014. He
is currently working toward Ph.D degree in the Research and Development Center, Bharthiyar
University, India. Puttaswamy is a Fellow of Higher Education Academy (UK) and also he is a
Senior Fellow of Higher Education Academy (UK). He is currently member of the Society of
Photographic Instrumentation Engineers (SPIE) and also member of IEEE. His current research
interests include image processing, hyperspectral Image Classification/ Extraction and automatic
target recognition.
Dr. P. Balamurugan is currently an Assistant Professor in the Department of Computer Science
Govt. Arts College, Coimbatore, India where teaches digital image processing. His research interests
are Digital Image Processing and Pattern recognition. He is the author (or coauthor) of reputed
journal papers, several book chapters. He is a referee of several international journals and has served
on the Scientific Committees of several national and international conferences. He is a member of
the IEEE.

A Neural Network Approach to Identify Hyperspectral Image Content

More Related Content

What's hot (18)

Similar to A Neural Network Approach to Identify Hyperspectral Image Content (20)

More from IJECEIAES (20)

Recently uploaded (20)

A Neural Network Approach to Identify Hyperspectral Image Content