International Journal of Engineering and Science Invention (IJESI)

International Journal of Engineering Science Invention
ISSN (Online): 2319 – 6734, ISSN (Print): 2319 – 6726
www.ijesi.org Volume 3 Issue 1ǁ January 2014 ǁ PP.39-47

Hybrid Clustering Algorithm and Feed-Forward Neural Network
for Satellite Image Classification
1,

S.Praveena , 2,Dr.S.P.Singh

1

(Asst.Prof,ECE Dept,M.G.I.T,Hyderababd,A.P)
(Prof&HOD,ECE Dept,M.G.I.T,Hyderababd,A.P)

2

ABSTRACT: This paper presents a hybrid clustering algorithm and feed-forward neural network classifier for
land-cover mapping of trees, shade, building and road. It starts with the single step preprocessing procedure to
make the image suitable for segmentation. The pre-processed image is segmented using the hybrid geneticArtificial Bee Colony(ABC) algorithm that is developed by hybridizing the ABC and genetic algorithm to obtain
the effective segmentation in satellite image and classified using feed-forward neural network classifier .
Classification accuracy of hybrid algorithm is found better than well known KFCM, Moving KFCM.

KEYWORDS : Segmentation, classification, feature extraction, feed-forward neural network classifier,
hybrid genetic ABC algorithm.

I. INTRODUCTION
Satellite image classification includes mainly two steps segmentation and classification step. The main
objective of image segmentation is to partition the image into parts of strong correlation with objects or areas of
the real world image . Mainly, owing to its realistic significance, such as in the treatment of images attained
from satellite prospection, the recognition of the types of objects comprises an important issue in pattern
recognition [1]. Actually, image segmentation can be known as the practice of conveying a label to each pixel in
an image, such that pixels with the similar label signify the same object, or its components. Based on
segmentation methods, Segmentation algorithms can be categorized into dissimilar categories used such as the
features thresholding [2], template matching [3], region based technique and clustering. Those methods have
their own restrictions and benefits in terms of appropriateness, presentation and computational cost.At present ,
in the Literature, an extensive variety of satellite image categorization methods are: Cluster, Statistic, Bayesian
Net, Artificial Neural Networks (ANN), etc [4]. By visual understanding of satellite imagery and aerial
photography, now, operators can manually remove cartographic features, such as buildings, roads, and trees.
Semi-automatic algorithms also help the operator to develop the process. General categorization
algorithms for low-resolution satellite imagery are excessively restricted to deal with complex high-resolution
satellite information and need novel algorithms. Some authors have used categorization techniques for satellite
images [5,6]. Based on computational intelligence diverse classifiers in the final direction, has been proposed
[7]. Taking benefit of the improved mapping capabilities of neural networks, conventionally, neural network
classifiers are proposed [8,9]. For building classifiers that are effortlessly interpretable by humans, Fuzzy theory
presents an eye-catching framework, whereas at the same time modeling the inexactness come across in nature.
Fuzzy classifiers are frequently consigned to as „„soft‟‟ classifiers, as opposed to the non-fuzzy „„hard‟‟
classifiers, with their applicability being of still larger significance under the survival of varied pixels.

II. REVIEW OF LITERATURE
In the image processing domain, Satellite image categorization is an extremely taxing task. In order to
progress precision, researchers have applied different kinds of classification techniques. According to the
database of skilled knowledge for a more focused satellite image classification, a hybrid biologically inspired
method was adapted that was presented by Lavika Goel [10]. As a competent land cover classifier for satellite
image, a hybrid FPAB/BBO based algorithm has been presented by Navdeep Kaur Johal et al. [11].A
FPAB/BFO based algorithm for the categorization of satellite image has been presented by Parminder Singh et
al. [12]. To choose the optimal and minimum set of fuzzy rules, the use of a genetic algorithm (GA) has been
presented by O. Gordo et al. [13] and to categorize remotely sensed images using a fuzzy classifier.

www.ijesi.org

39 | Page

Hybrid Clustering Algorithm And Feed…
A generalized multiple-kernel fuzzy C-means clustering (MKFCM) methodology for satellite image
segmentation has been presented by M. Ganesh and V. Palanisamy [14]. Aissam Bekkari et al. [15] have
progressed a methodology using a combination kernel that effortlessly combined multi-spectral features,
Haralick consistency characteristics and Hybrid Median Filter, with dissimilar window sizes. The result
demonstrated that the common use of spectral and texture data as one considerably enhanced the precision of
satellite image categorization.This paper, proposes a significant satellite image classification technique using
hybrid clustering algorithm and feed-forward neural network classifier. The proposed approach consists of
three steps
[1] Pre-processing
[2] Segmentation using genetic-ABC algorithm
[3] Classification using feed-forward neural network classifier.
Initially pre-processing is performed to make the image suitable for segmentation. In segmentation, the preprocessed image is segmented using hybrid genetic-ABC algorithm that is developed by hybridizing the ABC
algorithm [16] and genetic algorithm to obtain the effective segmentation in satellite images. Then, feature is
extracted and the classification of satellite image into four different labels (tree, shade, road and building) is
done using feed-forward neural network classifier.

III. HYBRID CLUSTERING ALGORITHM AND FEED-FORWARD NEURAL NETWORK
CLASSIFIER FOR SATELLITE IMAGE CLASSIFICATION
In remote sensing various techniques like Parallelepiped Classification, Minimum Distance to Mean
Classification, Maximum Likelihood Classification etc. [17] are used satellite image classification. These
techniques not only need high resolution image for processing, but also have limited accuracy in information
retrieval and insensitive to different degrees of variances in the spectral response data. To provide a solution to
the above problems, a new segmentation technique is introduced for image classification. Three major stages are
involved in proposed methodology:
(1)Pre-processing
(2)Segmentation using genetic-ABC algorithm
(3)Classification using Feed-forward neural network
Classifier

Fig(1).Flow Diagram of Proposed approach
3.1 Preprocessing
Satellite images cannot be given directly as the input for the proposed technique. Thus, it is
indispensable to perform pre-processing on the input image, so that the image gets transformed to be relevant for
the further processing. In proposed technique,
 A Nonlinear Median filter which is used to filter out the noise in the R, G and B layers for filtering noise.
It is used because, under certain conditions, it preserves edges while removing noise.
Image adjustment function is used and the range of pixel intensity in each of the R, G and B layers is made
[0 - 255] by stretching if necessary.

www.ijesi.org

40 | Page

3.2 Segmentation using Genetic-ABC Algorithm
ABC algorithm and Genetic algorithm are combined to obtain the effective segmentation of satellite
images. Image segmentation process using hybrid algorithm as under:
Step (a): The satellite images are converted into gray images and then these gray images are segmented into six
new segments using prposed algorithm.
Step (b): The gray image is converted into HSL layer. From converted HSL images, the H layer image is
segmented into six segments using Hybrid algorithm.
Step(c): From the above two steps , finally we have obtained thirty six new segments like as tree, shadow,
building, road, etc.
The detailed explanation of proposed algorithm is given below.
Step 1: Chromosome representation and Generation of an initial solution
In the initialization phase, we have generated the „m‟ number of chromosome. Each element of the vectors
representing a cluster centre is an integer that indicates the ranges between 0 to 1.

Fig( 2): Chromosome representation process
In clustering process, the centroids should be defined based on the number of cluster. Most of algorithm uses
random data points as an initial centroid for clustering. In this paper, a set of cluster centorid (solutions) is
randomly generated (range between 0 to 1) to the hybridizing algorithm. Every chromosome (cluster centroid) in
the initial population consists „k‟ number of vectors where every vector is having a length of „d‟ which is equal
to the dimensions of data. For example, a chromosome encoding with N number of clusters and corresponding
centroids for 10 chromosomes is given in Fig(3). Here, M is the centroid value.

Fig(3): Candidate solution encoding with N number of clusters and corresponding centroids for 10
chromosomes.

www.ijesi.org

41 | Page

Step 2: Fitness Computation
The fitness of each chromosome in the solution is evaluated with objective function presented by Eq(1).
For every solution Yi , the objective function should be computed to find the survival of the solution. The
objective function of the clustering process taken for this paper is intra cluster distance that is computed by the
following equation

J1 (Y ) 

1

 min  X
N

i 1

i

Xj



; j  1, 2,..., k ---- (1)

Where, X i  X j is a chosen distance measure between a data point X i and the cluster centre X j , „N‟ and „k‟
are the number of input data and the number of cluster centres, respectively. The finding of objective function of
clustering process is illustrated in Fig(4).
Step 3: Employed bee phase
After initialization, in the employed bees‟ phase, each employed bee is sent to the food source in its
memory and finds a neighboring food source. Here, 20% chromosomes are randomly generated for every new
chromosome generation process. Subsequently, we have determined best fitness from the employed bee phase.

Fig(4): Example diagram of clustering process
Step 4: Onlooker bee phase
In the onlooker bees‟ phase, the onlookers receive the information of the food sources shared by
employed bees. Then each chromosome has chosen a solution to exploit depending on a probability related to
the nectar amount of the solution (fitness values of the solution). That is to say, there may be more than one
onlooker bees choosing a same solution if the source has a higher fitness. The probability Pi is calculated
according to Eq(2) as followed:



0.25
Pi  
* fit  1
Max fitness



(2)

After solutions have been chosen, each onlooker bee finds a new solution in its neighbourhood following
Eq(3),
(3)
xi , j  x min  rand (0,1)( x max  x min )
j
j
j

www.ijesi.org

42 | Page

Step 5: Mutation
After onlooker bee phase, some selected chromosomes are considered for mutation. Mutation consists
of changing the value of a random bit, which is randomly chosen in the chosen vector.
Step 6: Scout bee phase
In scout bees‟ phase, if the value of trials counter of a solution is greater than a parameter, known as
„limit‟, the solution is abandoned and the bee becomes a scout bee. A new food solution has produced randomly
in the search space using Eq(3), as in the case of initialization phase. The employed, onlooker, mutation and
scout bees‟ phases will recycle until the termination condition is met. The best food source which presents the
best solution.
3.3 Feature Extraction
Feature extraction process for satellite image segmentation is discussed in this section. Four new
segments like as road, building, tree and shadow are obtained from the 36 segments. First, HSL image is
obtained from the median filtered training images. Then feature extraction step is formulated in the following
steps.








H, S and L layer obtained from HSL image
T layer obtained from TSL image
A and B layer obtained from LAB image
U and V layer obtained UVL image
Index of peak value in 256 bin histogram of each is noted.
Mean pixel value in each of these layers is obtained.
Values obtained in step 5 and 6 are used as features.

3.3 Classification using Neural Network
Classification step is to identify road, building, tree and shadow regions from original satellite image.
16 features have been used for the classification purpose. Here, feed forward neural network is used for
classification purpose. The neural network model is depicted in Fig (5).

Fig(5): Neural network model for proposed approach
The proposed neural network model comprises three layers of nodes - an input layer, hidden layer and
an output layer, as shown in Fig. 5. Each satellite image is represented by its features in separate columns of the
pattern (input) matrix P. The weight matrix W is the connection matrix for the input layer to the output layer.
The first step in simple network implementation is determination of number of output neurons. Before training a
feed forward network, the weight and biases is initialized. Once the network weights and biases have been
initialized, the network is ready for training. We used random numbers around zero to initialize weights and
biases in the proposed neural network. The training process needs a set of proper inputs (for our approach we
give 16 features) and targets as outputs (tree, shadow, building and road). During training, the weights and biases
of the network are iteratively adjusted to minimize the network performance function. The default performance
function for feed forward networks is mean square errors, the average square errors between the network outputs
and the target output. Finally, the four different labels (tree, shade, road and building) of original satellite image
are identified .

www.ijesi.org

43 | Page

IV. SIMULATION RESULTS
The proposed approach of satellite image segmentation is experimented with the satellite image dataset
and the result is evaluated with accuracy.
4.1 Experimental Setup and Dataset Description
Experimental results of proposed technique using different satellite images are experimented using MATLAB
7.12 with WINDOWS 7and 4GB RAM original satellite image (.lan format) have multispectral (R, G, B, NIR)
bands and these images converted to TIFF format. This is done using Multispec32 which is a freeware
multispectral image data analysis system. The
Fig(6), shows the input satellite image.

Fig.6 Input Satillite Image
4.2 Evaluation Metrics
The evaluation of proposed technique in different satellite images are carried out using the following
metrics as suggested by below equation.
number of true positives 

Accuracy =

number of true negatives
number of true positives  false negatives 
true negatives  false positives

4.3 Experimental Results
The proposed technique is designed for identify the four different labels (tree, shade, road and building)
of satellite images. The obtained experimental results from the proposed technique are given in figure 7 (a)(d).These figures shows the extracted regions (tree, shade, road and building) from the Input satellite images.

Fig.7(a) Building

www.ijesi.org

44 | Page


Fig.7(b) Road

Fig.7(c) Shade

Fig.7(d) Tree
4.4 Comparative Analysis
In this paper, the proposed algorithm results are compared with against Moving KFCM and KFCM
algorithm. The performance of this is analyzed with parameter accuracy. The accuracy value is computed by
dividing the total number of similar pixels identified as land use to the number of pixels in the tree, shade,
building and road region. The detailed results obtained for proposed technique is drawn as graph shown in
Fig((10). As shown in Fig(10) the proposed technique is achieved the maximum accuracy value of 74% for
shade region which is high compared with the accuracy of MKFCM (67%), KFCM (68%) . It is giving better
results for road, shade and build regions except tree. Hence, it is clear from result analysis that the proposed
algorithm shows better results as compared with existing techniques.

www.ijesi.org

45 | Page


Fig(10): Comparison graph of Accuracy

IV.

CONCLUSION

In this paper, optimization algorithms for segmentation with the intention of improving the
segmentation in satellite images using feed-forward neural network classifier is proposed. The overall steps
involved in the proposed technique in three steps such as, i) Pre-processing, ii) segmentation using genetic-ABC
algorithm, and iii) classification using feed-forward neural network classifier. Initially pre-processing is
performed to make the image suitable for segmentation. In segmentation, the pre-processed image is segmented
using hybrid genetic-ABC algorithm that is developed by hybridizing the ABC algorithm and genetic algorithm
to obtain the effective segmentation in satellite images. Then, feature is extracted and the classification of
satellite image into four different labels (tree, shade, road and building) is done using feed-forward neural
network classifier. Finally, classification accuracy of the proposed algorithm in satellite image classification is
calculated and the performance is compared with Moving KFCM and KFCM algorithm. Practical
implications: Green area: The detection of green areas could help scientist to study deforestation and changes
in vegetation. Building area: This detection can be useful for urban development applications. Road area: This
detection can be used in several industrial applications such as for auto recognition and for military purposes. It
is important to provide military and other groups with accurate, up-to-date maps of the road networks in any
region of the world. Overall, this work can be usefully to do the intelligent analysis to find the growth of every
region as well as the awareness of building, and green.

REFERENCES
[1]
[2]
[3]
[4]
[5]

[6]
[7]
[8]
[9]
[10]

S. A. Rojas., and D. Fernandez-Reyes., “Adapting Multiple Kernel Parameters for Support Vector Machines using Genetic
Algorithms,” In Proc. IEEE Congr. Evol. Comput., Edinburgh, U.K., 2005, vol. 1-3, pp. 626–631.
K. S. Chuang., H. L. Tzeng., S. Chen., J. Wu., and T. J. Chen., “Fuzzy C-Means Clustering with Spatial Information for image
Segmentation,” Comput. Med. Imaging Graph, vol. 30, no. 1, pp. 9-15, Jan. 2006.
R. Szeliski., D. Tonnesen, and D. Terzopoulos, “Modeling Surfaces of Arbitrary Topology with Dynamic Particles”, In:
Proceedings of CVPR, pp. 82–87, 1999
Susana Arias, Hector Gómez, Francisco Prieto, María Botón, Raúl Ramos, “Satellite Image Classification by Self-Organized
Maps on GRID Computing Infrastructures ,"Proceedings of the Second EELA-2 Conference,2009.
Bernardini, A., Malinverni, E. S., Zingaretti, P. and Mancini, A.”Automatic classification methods of high-resolution satellite
images: the principal component analysis applied to the sample training set,” In: The International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences, Beijing, Vol. XXXVII. Part B7, pp. 701-706, 2008.
Alonso, M.C., Sanz, A. and Malpica, J.A,”Classification of high resolution satellite images using texture,” In Bebis G. et al.
(Eds.): Advances in Visual Computing ISVC07.Part II LNCS 4842, pp. 499-508. Springer-Verlag, 2007.
Stathakis, D., Vasilakos, A., “Comparison of computational intelligence based classification techniques for remotely sensed
optical image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol.44, no.8, 2305-2318, 2008
Benediktsson, J.A., Swain, P.H., Esroy, O.K., “Conjugate-gradient neural networks in classification of multisource and veryhigh-dimensional remote sensing data,” International Journal of Remote Sensing, vol.14, no.15, pp.2883–2903, 1993.
Liu, Z., Liu, A., Wang, C., Niu, Z.,” Evolving neural network using real coded genetic algorithm (GA) for multispectral image
classification,” Future Generation Computer Systems, vol. 20,no.7, pp.1119-1129, 2004.
Lavika Goel,"Land Cover Feature Extraction using Hybrid Swarm Intelligence Techniques - A Remote Sensing Perspective,"
ACEEE Int. J. on Signal & Image Processing, Vol. 01, No. 03, Dec 2010

www.ijesi.org

46 | Page

[11]
[12]
[13]
[14]
[15]

[16]
[17]

Navdeep Kaur Johal, Samandeep Singh, Harish Kundra, “Hybrid FPAB/BBO Algorithm for Satellite Image Classification,”
International Journal of Computer Applications, pp.0975- 8887, Vol.6, No.5, September 2010
Parminder Singh, Navdeep Kaur, Loveleen Kaur,"Satellite Image Classification by Hybridization of FPAB Algorithm and
Bacterial Chemotaxis,"International Journal of Computer Technology and Electronics Engineering (IJCTEE), Vol.1, No.3, 2011.
O. Gordo, E. Martínez, C. Gonzalo and A. Arquero,"Classification of Satellite Images by means of Fuzzy Rules generated by a
Genetic Algorithm,"IEEE Latin America Transactions, Vol. 9, No. 1, March 2011
M. Ganesh, V. Palanisamy,"Multiple-Kernel Fuzzy C-Means Algorithm for Satellite Image Segmentation, “European Journal of
Scientific Research, ISSN 1450-216X,Vol.83 No.2, pp.255-263,2012.
Aissam Bekkari, Soufian Idbraim,Azeddine Elhassouny, Driss Mammass and Mostapha El yassa,Danielle Ducrot,"Classification
of High Resolution Urban satellites Images using SVM and Haralick Features with a Hybrid Median Filter, “Special Issue of
International Journal of Computer Applications on Software Engineering, Databases and Expert Systems-SEDEXS, pp. 09758887,September 2012.
Changsheng Zhang, Dantong Ouyang, Jiaxu Ning, "An artificial bee colony approach for clustering", Expert Systems with
Applications, vol. 37, pp. 4761–4767, 2010.
Lillesand, T.M., Kiefer, R.W. and Chipman J.W., “Remote Sensing and Image Interpretation”, Fifth Edition, Wiley & Sons Ltd.,
England, pp.586-592, 2003.

www.ijesi.org

47 | Page

International Journal of Engineering and Science Invention (IJESI)

More Related Content

What's hot (19)

Viewers also liked (8)

Similar to International Journal of Engineering and Science Invention (IJESI) (20)

Recently uploaded (20)

International Journal of Engineering and Science Invention (IJESI)