Image Classification For SAR Images using Modified ANN

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 07 | July -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 5.181 | ISO 9001:2008 Certified Journal | Page 2924
Image Classification For SAR Images Using Modified ANN
Prateek Priyadarshini1, Col. Dr. O.P Malik2
1 P.G Student, Dept. of Electronics & communication, Al-Falah University, Haryana, India
2 Professor, Dept. of Electical & Electronics, Al-Falah University, Haryana, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Classification of polarimetric SAR images has
become a very important topic after the availability of
Polarimetric SAR images through different sensors like SIR-C,
ALOS-PALSAR etc. An analyst attempts to classify features in
an SAR image by using the elements of visual interpretation to
identify homogeneous groups of pixels that represents various
features or land cover classes of interest. There is need to
devise accurate methods for classification of SAR images. The
combinations of different polarizations from L- and C- band
helps to improve the classification accuracy. It was found that
the combinations of channels gavethebestoverallaccuracies..
The proposed classifier is examined in Matlab with the help of
Modified Artificial Neural Network using feed forward back
propagation technique. The method finds 9 different land
cover and sites.
Key Words: Synthetic Aperture Radar(SAR), Artificial
Neural Networks(ANN), Real Aperture Radar (RAR),
Classification, Polarimetry
1. INTRODUCTION
Synthetic Aperture Radar is a radar technology that is used
from satellite or airplane. It produceshighresolutionimages
of earth‘s surface by using special signal processing
techniques. Synthetic aperture radar has important role in
gathering information about earth‘s surface because it can
operate under all kinds of weather condition (whether it is
cloudy, hazy or dark). Polarimetric SAR (PolSAR) image
classification is arguably one of the most important
applications in remote sensing. Classification is the process
of assigning a set of given data elements to a given set of
labels or classes such thatvariousparameterofassigning the
data element to a class is optimized. Radar polarimetry is a
technique for classification of land use features. Various
research work have reported the use of polarimetric data to
map earth terrain types and land covers ([1], [2], [3], [4],
[5]). Image classification can be mainly divided into
supervised and unsupervised classification techniques. An
unsupervised classification technique, classifies the image
automatically by finding the clusters based on certain
criterion. On the other hand in supervised classification
technique the location and the identity of some cover type
and terrain type , for example urban, forest, and water are
known prior to us.The data is collected by a field work,
maps,
and personal experience. The analyst tries to locate these
areas on the remotely sensed data.Theseareasareknownas
“training sites”. An analyst can guide a classifier with the
help of these training sites to learn the relationship between
the data and the classes. This manual technique of selecting
training sets could be difficult when ground truth is not
available. In this paper a new technique is proposed using
modified ANN. It is a supervisedclassificationtechnique.The
proposed method is tested and analyzed in MATLAB.
1.1 Literature Survey
Both visual interpretation and automatic analysis of data
from imaging radarsare complicated by a fading effectcalled
speckle, which manifests itself as a strong granularity in
detected images (amplitude or intensity). For example,
simple classification methods based on thresholding of gray
levels are generally inefficient when applied to speckled
images, due to the high degree of overlap between the
distributions of the different classes. Speckle is causedbythe
constructive and destructive interference between waves
returned from elementary scatterers within each resolution
cell. It is generally modelled as a multiplicative randomnoise
. Compared with optical image, SAR image has more legible
outline, better contrast and more plentiful texture
information. The objects of different shape and physical
feature take on different texture character, which is a critical
technique of identifying objects by radar. At present, there
are many approaches to image classification, but there is not
an approach to suit all kindsof images. During the pastyears,
different methods were employed for classification of
synthetic aperture radar (SAR) data, based on the Maximum
Likelihood (ML), artificial Neural Networks (ANN) fuzzy
methods or other approaches . The NN classifier depends
only on the training dataand the discrimination powerofthe
features. Fukuda and Hirosawa applied wavelet-based
texture feature sets for classification of multi frequency
polarimetricSARimages.TheClassificationaccuracydepends
on quality of features and the employed classification
algorithm. For a high resolution SAR image classification,
there is a strong need for statistical models of scattering to
take into account multiplicativenoiseandhighdynamics.For
instance, the classification process needs to be based on the
use of statistics. ClutterinSARimagesbecomesnon-Gaussian
when the resolution is high or when the area is man-made.
Many models have been proposed to fit with non-Gaussian
Scattering statistics (Weibull, Log normal, Nakagami Rice,
etc.), but none of them is flexible enough to modelall kindsof
surfaces in our context.
For SAR image classification problem many fuzzy models
have been proposed, Fuzzy c-means clustering (FCM)
algorithm is widely applied in various areas such as image

processing and pattern recognition. Co-occurrence matrix
and entropy calculationsare used toextracttransitionregion
for an image. This transition region approach is used to
classify the SAR images.However most of the work remains
restricted to maximum 6 training sites
1.2 Image Processing in SAR
The raw data received from the imaging sensors on the
satellite platforms or aircrafts contains flaws and
deficiencies. To overcome these flaws and deficiencies in
order to get the originality of the data, it needs to undergo
several steps of processing. This will vary from image to
image depending on the type of image format, initial
condition of the image and the information of interest and
the composition of the image scenes. . Digital Image
Processing undergoes three general steps:
i. Pre-processing
ii. Display and enhancement
iii. Information extraction
Pre-processing consists of those operations that prepare
data for subsequent analysis that attempts to correct or
compensate for systematic errors. The digital imageries are
subjected to several corrections such as geometric,
radiometric and atmospheric, though all these corrections
might not be necessarily be applied in all cases. Theseerrors
are systematic and can be removed before they reach the
user. The investigator should decide which pre-processing
techniques are relevant on the basis of the nature of the
information to be extractedfromremotelysenseddata.After
pre-processing is complete, the analyst may use feature
extraction to reduce the dimensionality of the data. Thus
feature extraction is the process of isolating the most useful
components of the data for further study while discarding
the less useful aspects (errors, noise etc).
Image Enhancement operations are carried outtoimprove
the interpretability of the image by increasing apparent
contrast among various features in the scene. As an image
enhancement technique often drastically alters the original
numeric data, it is normally used only for visual (manual)
interpretation andnotforfurther numeric analysis.Common
enhancements includes transect extraction, contrast
adjustments, spatial filtering, Fourier transformations, etc.
Information Extraction is the last step toward the final
output of the image analysis. After pre-processingandimage
enhancement the remotely sensed data is subjected to
quantitative analysis to assign individual pixels to specific
classes. Classification of the image is based on the known
and unknown identity to classify the remainder of theimage
consisting of those pixels of unknown identity. After
classification is complete, it is necessary to evaluate its
accuracy by comparing the categories on the classified
images with the areas of known identity on the ground. The
final result of the analysis consists of maps (or images), data
and a report. These three components of the result provide
the user with full information concerning the source data,
the method of analysis and the outcome and its reliability.
2. Classification Using ANN
Supervised classification methods for the polarimetric SAR
data can be divided into statistical and neural network
approaches. Neural network techniques(Hara,1994;Chenet
al, 1996)[6][7] have also been applied using the complete
polarimetric information as input, and iterative training was
normally necessary; Chen etal.(Chenetal,1996)[7]applieda
dynamic learning neural network and fuzzy neural network
to classify multi frequency POLSAR. Ito et al. (1998) [8] have
proposed a classification method using a competitive neural
network trained by only two Learning. Vector Quantization
(LVQ) algorithms. A method which selects a suitable feature
vector using the JM distance is proposed. In addition, they
introduce a pseudo-relative phase between polarimetries in
order to obtain higher classification accuracy. Hellmann
(1999) [9] has proposed a classification based on H-alpha
decomposition theorem extended by the useofthefirsteigen
value of the coherency matrix. Fuzzy logic as well as ANN
strategies is used to improve the classification accuracy.
Lorenzo Bruzzone (2004) [10] integrates an advanced
pattern recognitionmethodology(basedonmachinelearning
techniques) with an accuratefeatureextractionphase(bases
on the SAR signal physics analysis) for better classification
accuracy. To classify a pattern, certain attribute values from
that pattern are input into the directed graph at the
corresponding source nodes. There is one sink node for each
class. The output value that is generated indicates the
probability that the corresponding input pattern belongs to
that class. The pattern will then be assigned to the class with
the highest probability of membership. The learning process
modifies the labeling of the arcs to better classify patterns.
After the classification is done for the training set the results
are compared with the actual classification and the accuracy
is computed. Learning process continues with different
weights and with all the training data or until the
classification accuracy is adequate.
3. Proposed Method
Data for classification problems can very often have textual
or non-numeric information. In our case, classes are non-
numeric (Light beach/Marshy/Flower
Fields/Reeds/vegetation/Houses/River/Growing
vegetation/Farmland). Neural networks however cannot be
trained with non-numeric data. Hence there is a need to
translate the textual data into a numeric form. There are
several ways to translate textual or symbolic data into
numeric data. Some of the common symbol translation
techniques used is unary encoding, binary encoding and
numbering classes. So unary encoding is used in this code to
perform symbol translation.

[1] Data is prep-processed into a form that can be used
with ANN. The neural network object in the Matlab toolbox
recognizes its features along rows and samples along
columns. Hence we work on transpose.
[2] The next step is to create a neural network (feedforward
back propagation network) that will learn to identify the
classes.The neural network starts with random initial
weights , the results varies every time it is run. The random
seed or twister is set to avoid the randomness.
[3] With 16 neurons in the hidden layer A 1-hidden layer
feed forward network is created.
[4] Now the network can to be trained. The samples are
automatically divided into training, validation and test sets.
The training set is used to teach the network. Training will
continue as long as the network continues improving on the
validation set.
[5] Testing the classifier.
The trained ANN can now be tested with the testing samples
of training sites. This will give us a sense of how well the
network will do when applied to data from the real world.
[6] Calculation of classification accuracy using confusion
matrix.
The network response can now be compared against the
desired target response to build the classification matrix
which will provide a comprehensive picture of a classifiers
performance
3.1 Back Propagation Algorithm
Back propagation is a form of supervised learning for
multilayer nets, also known as the generalizeddelta rule. the
back propagation algorithm has been widely used as a
learning algorithm in feed forward multilayer neural
networks. Error data at the output layer is back propagated
to earlier ones, allowing incoming weights to these layers to
be updated. It is most often used as training algorithm in
current neural network applications. The method calculates
the gradient of loss function with respects to all the weights
in the network. The gradient is fed to the optimization
method which in turn uses it to update the weights, in an
attempt to minimize the loss function. It requires a known,
desired output for each input value in order to calculate the
loss gradient function.
Fig -1: SAR Image
Fig -2: Flowchart
START
COLLECTING
PALSAR SIR-C DATA
PRE-PROCESSING
FOR 3 BAND
IMAGES
FEATURE
EXTRACTION USING
EXTRACTION
FEATURE
CLASSIFICATION
USING ANN
STOP

a)
b)
Fig -3 a & b: Areas for classification
Fig -4 : Classification of Areas in Pie chart
Fig -5 Light beach Fig -6 Marshy
Fig -7 Flower fields Fig -8 Reeds
3. CONCLUSIONS
A modified algorithm basedontheArtificial Neural Network
was designed for SAR images classification. The new
algorithm is successfully applied for classification of SAR
images. The experimental results consistentlyshowthatthe
proposed algorithm has high classification precision. When
compared with other two classifiers, K-means, and FCM, the
average performance of ANN is better than them. The
performance is evaluated by accuracy assessment. In terms
of image classification accuracy evaluation, attention was
mainly focused on subjective methods of evaluation.
Confusion matrix, overall accuracy andKappa coefficientare
the main considerations in organizing and running
subjective tests for method of image classification accuracy
evaluation. The effectiveness of ANN algorithm was
evaluated by accuracy assessment. We achieved more
accuracy with ANN because it is a global searching
technique. The performance of ANN technique isfoundto be
satisfactory and the performance of outperformed FCM and
K-means technique. The classification results are validated
with various SAR images.
classifie
d image
vegeta
tion
Mars
hy
Flowe
r
fields
ree
ds
∑ User‘s
accura
cy
vegetation 75.2 10.8 8.4 0.3 94.
7
79.40 %
Marshy 8.4 84.2 2.0 11.
8
106
.4
79.13 %
Flower
fields
6.4 1.2 88.1 8.7 104
.4
84.38 %
reeds 10 3.8 1.5 79.
2
94.
5
83.80
∑ 100 100 100
Procedu
re‘s
accuracy
75.2 84.2 88.1 79.
2
Table -1: confusion Matrix

ClassificationTechnique Overall accuracy
ANN 81.67
FCM 78.3
Table -2: Comparison Table
REFERENCES
[1] R. Touzi, W.M. Boerner, J.S. Lee, and E. Lueneburg , “A
review of polarimetry in the context of synthetic
aperture radar: concepts and information extraction”,
Can. J. Remote Sensing, Vol. 30, No. 3, pp. 380–407, 2004.
[2] S.Cloude, E. Pottier “A Review of Target Decomposition
Theorems in Radar Polarimetry”, IEEE Transactions of
Geoscience and Remote Sensing, Vol.34, No.2, pp. 498-
518, March 1996.
[3] Van Zyl, J.J., “Unsupervised classification of scattering
behavior using radar polarimetry data,” IEEE
Transactions on Geoscience and Remote Sensing, Vol. 27,
No. 1, pp. 37–45, 1989.
[4] M Ouarzeddine, and B Souissi, “Unsupervised
Classification Using Wishart Classifier”, USTHB, F.E.I, BP
No 32 EI Alia Bab Ezzouar, Alger.
[5] S. R. Cloude and E. Pottier, “An Entropy based
classification scheme for land applications of
polarimetric SAR,” IEEE IGRS, vol.35, no.1, pp. 68-78,
Jan.1997.
[6] Hara, Y., Atkins, R. G., Yueh, S. H., Shin, R. T., and Kong, J.
A., “ Application of neural networks to radar image
classification,” IEEE T ransactions on Geoscience and
Remote Sensing, vol. 32, pp100-109.1994.
[7] Chen, K. S., Huang, W. P., Tsay, D. H., and Amar, F.,
“Classification of multifrequency polarimetric SAR
imagery using a dynamic learning neural network,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 34,
pp 814- 820, 1996.
[8] Yosuke Ito and SigeruOmatutt,“APolarimetric SARData
Classification Method Using Neural Networks”,
INT.J.Remtoe Sensting, Vol.19, No. 14, pp 2665-2684,
1998.
[9] M. Hellmann, G. Jager, E. Kratzschmar,M. Habermeyer,
“Classification of h11 Polarimetric SAR-Data using
Artificial Neural Networks and Fuzzy Algorithms”,IEEE
Transactions, pp 1995-1997, 1999.
[10] Lorenzo Bruzzone, “An advanced system for the
automatic classification of multitemporal SAR images,”
IEEE Transactions on Geoscience and Remote Sensing,
Vol. 42, No. 6, June 2004.

Image Classification For SAR Images using Modified ANN

More Related Content

What's hot (20)

Similar to Image Classification For SAR Images using Modified ANN (20)

More from IRJET Journal (20)

Recently uploaded (20)

Image Classification For SAR Images using Modified ANN