108

CHAPTER 5

IMAGE CLASSIFICATION

Image classification is a complex process that may be affected by

many factors. This chapter details supervised and unsupervised classification
techniques. The emphasis is placed on the support vector machine
classification approach and how this technique is used for improving
classification accuracy.

5.1 INTRODUCTION

The amount of image data that is received from satellite is

constantly increasing. For example, nearly 3 terabytes of data are being sent
to Earth by NASA’s satellites every day. Advances in satellite technology and
computing power have enabled the study of multi-modal, multi-spectral,
multi-resolution and multi-temporal data sets for applications such as urban
land use monitoring and management, Geographic Information System (GIS)
and mapping, environmental change, site suitability, agricultural and
ecological studies. Automatic content extraction, classification and content-
based retrieval have become highly desired goals for developing intelligent
systems for effective and efficient processing of remotely sensed data sets.

Gong and Howarth (1992) discussed the classification of remote

sensing data is a complex process and requires consideration of many factors.
The user’s need, scale of the study area, economic condition, and analyst’s
skills are important factors influencing the selection of remotely sensed data,
 109

the design of the classification procedure and the quality of the classification
results. The major steps of image classification may include image
preprocessing, feature extraction, selection of training samples, selection of
suitable classification approaches, post-classification processing, and
accuracy assessment.

Figure 5.1 Classification of remote sensing data

Phinn et al (2000) described classification of remote sensing data is

used to assign corresponding labels with respect to homogeneous
characteristics of groups. The main aim of classification is to discriminate
multiple objects from each other within the image. Classification will be
executed on the base of spectral or spectrally defined features, such as
density, texture etc., in the feature space. It can be said that classification
divides the feature space into several classes based on a decision rule. Figure
5.1 shows the concept of classification of remote sensing data. In many cases,
classification will be undertaken using a computer, with the use of
 110

mathematical classification techniques. Classification will be made according

to the following procedures as shown in Figure 5.2.

Definition of classes

Selection of features

Sampling of training
data

Finding of proper
decision rule

Classification

Verification of results

Figure 5.2 Classification procedure

Step 1: Definition of Classification Classes

Depending on the objective and the characteristics of the image

data, the classification classes should be clearly defined.

Step 2: Selection of Features

Features to discriminate between the classes should be established

using multi-spectral or multi-temporal characteristics, colour, textures etc.

Step 3: Sampling of Training Data

Training data should be sampled in order to determine appropriate

decision rules. Classification techniques such as supervised or unsupervised
learning will then be selected on the basis of the training data sets.
 111

Step 4: Finding of proper decision rule

Various classification techniques will be compared with the training

data, so that an appropriate decision rule is selected for subsequent
classification.

Step 5: Classification

Depending upon the decision rule, all pixels are classified in a single
class. There are two methods of pixel by pixel classification and per-field
classification, with respect to segmented areas.

Step 6: Verification of Results

The classified results should be checked and verified for their

accuracy and reliability.

5.2 CLASSIFICATION TECHNIQUES

The learning algorithms are broadly classified into supervised and

unsupervised learning techniques. The distinction is drawn from how the
learner classifies data. In supervised learning, the classes are predetermined.
These classes can be conceived of as a finite set, previously arrived at by a
human. In practice, a certain classes of data will be labeled with these
classifications. Althausen (2002) reviewed the classes are then evaluated
based on their predictive capacity in relation to measures of variance in the
data itself. Some of the examples of supervised classification techniques are
Back Propagation Network (BPN), Learning Vector Quantization (LVQ), Self
Organizing Map (SOM), Support Vector Machine (SVM), etc.,

The basic task of unsupervised learning is to develop classification

labels automatically. Unsupervised algorithms seek out similarity between
 112

pieces of data in order to determine whether that can be characterized as

forming a group. These groups are termed clusters. Unsupervised
classification, often called as clustering, the system is not informed how the
pixels are grouped. The task of clustering is to arrive at some grouping of the
data. One of the very common of cluster analysis is K-means clustering.

5.2.1 Classification of Remote Sensing Data

Lefsky and Cohen (2003) described the intent of the classification

process is to categorize all pixels in an image into one of several land cover
classes, or themes. This categorized data may then be used to produce
thematic maps of the land cover present in an image. Normally, multispectral
data are used to perform the classification and, indeed, the spectral pattern
present within the data for each pixel is used as the numerical basis for
categorization. The objective of image classification is to identify and portray,
as a unique gray level (or colour), the features occurring in an image in terms
of the object or type of land cover these features actually represent on the
ground.

5.3 THE UNSUPERVISED CLUSTERING

The unsupervised clustering is a kind of clustering which takes

place with minimum input from the operator; no training sample is available
and part of the feature space is achieved by identifying natural groupings of
the given data. In unsupervised clustering technique, an individual pixel is
compared to each cluster to see the closest pixel. Finally, a map of all pixels
in the image, classified as to different clusters, each pixel is most likely to
belong, is produced. This then must be interpreted by the user as to what the
colour patterns may mean in terms of classes, etc. that are actually present in
the real world scene; this requires some knowledge of the scene's feature or
 113

cluster content from general experience or personal familiarity with the area
imaged.

The objective here is to group multi-band spectral response patterns

into clusters that are statistically separable. The cluster numbers are initially
set and the total number can be varied arbitrarily. Generally, in an area within
a SAR image, multiple pixels in the same cluster correspond to some ground
feature or cluster so that patterns of gray levels result in a new image
depicting the spatial distribution of the clusters. These levels can then be
assigned to produce a cluster map. This can be done by either being
adequately familiar with the major classes expected in the scene, or, where
feasible, by visiting ground truth and visually correlating map patterns to their
ground counterparts. Since the classes are not selected beforehand, this
method is termed as unsupervised classification. This section discusses three
unsupervised classification techniques namely K-means clustering, PCA
based K-means clustering and Fuzzy C-Means clustering (FCM).

5.3.1 K-Means Clustering

Mao and Jain et al (1996) defined K-means clustering is used to

cluster the pixel intensities in an image into K number of clusters. This offers
the segmentation of an image into K segments similar intensities and it
provides more automation than manual thresholding of an image.

Assume the size of an image is m n , convert to a vector with

( m n ) rows and 1 column. With the images, a 3-dimensional feature vector
per pixel is used. K-means puts pixels into K number of clusters based on
similarities of intensity values. Finally got a set of different segmented image,
where each segment is relatively homogeneous in terms of pixel intensities.
K-means clustering is used for a simple technique for image classification.
 114

The following are the main three steps of K-Means clustering until
convergence.

Repeat the following steps until convergence.

Determine the centroid coordinate.

Determine the distance of each object to the centroids.

Cluster the object based on minimum distance.

Finally, to minimize the following objective function the squared

error function is used.

m n
2
A x ij ci (5.1)
i 1 j 1

2
where x ij ci is a distance measure between a data point x ij and the cluster

center ci .

The design for implementing K-means clustering technique is

shown in Figure 5.3.
Acquire the
Landsat image

Calculate the
Centroid

Minimize the
objective
function

Water body and

non-water body
cluster

Figure 5.3 K-means clustering – system design

 115

5.3.2 PCA based K-Means Clustering

Thanh et al (2005) defined PCA based K-means clustering is a

statistical technique widely used for dimension reduction. Usually the
clustering methods are developed for different purposes. Clustering
algorithms used for unsupervised classification of remote sensing data vary
according to the efficiency with which clustering takes place. The
unsupervised clustering provides the cluster information about the water body
and non-water body in a relatively quick manner.

Unsupervised
classification through Calculation of
PCA based K-means clustering in the Water body and non-
clustering coverage area water body

Figure 5.4 Clustering modules

The most important aspect of image classification is finding groups

in data. Pixels with similar intensities forms clusters in images. Gray values
are specifically used for this purpose. The K-means algorithm has some
further refinements for its applications like change detection, land cover
mapping, etc., by splitting and merging of clusters. This is illustrated in
Figure 5.4.

First, the given image values are preprocessed by using PCA. Using
the most important components of PCA, the image information is mapped
into the new feature space. Then, the K-means algorithm is applied to the data
in the feature space. The final objective is to distinguish the different clusters
using eigen values. Clusters are grouped if its standard deviation exceeds a
threshold and the number of pixels is twice for the minimum number of pixels.
The main intention of K-means algorithm is to reduce the cluster variability.
The objective function is to find the sum of square distance between cluster
centre and its assigned pixel value as defined in Equation (5.2).
 116

n
2
F xi C xi (5.2)
i 0

where, xi is the pixel value assigned to mean value of the cluster C xi . Next
to determine the error as defined in Equation (5.3). Minimizing the error is
equivalent to minimizing the sum of squared distances.

n
2
xi C xi
i 0
Error (5.3)
N C

where, N indicates number of pixels, C specifies the expected number of

clusters. K-means is very responsive to initial values. It is often not obvious
that the clustering with the lesser Mean Squared Error (MSE) is truthfully the
better classification. Thus, the results obtained have aided in attaining this
objective and thereby PCA based K-means clustering has been efficiently
applied for the classification analysis.

5.3.3 Fuzzy C-Means Clustering

Rui and Donald (2005) discussed Fuzzy C-Means (FCM) clustering

algorithm permits a data to belong to more than two clusters. This method is
frequently used in change detection, pattern recognition and classification. It
is mainly aimed to minimize the objective function as defined in the
Equation (5.4).

N C
2
Am aijm x j ci , 1 m (5.4)
i 1 j 1

where m 1, aij is the degree of membership of x in the cluster j, x j is the j th

dimensional measured data, ci is the cluster center, and x j ci expressing

the similarity between measured data and the cluster center.

 117

The objective function of the optimization yields the fuzzy

partitioning shown above, with the updation of membership function a ij,
defined in Equation (5.5) and the cluster centers, cj.

1
aij 2
(5.5)
m 1
c xj ci
m 1 xj cm

The iteration will stop until max ij aijk 1

aijk ,0 1 , k is the

number of steps to repeat the iteration. The following Figure 5.5 depicts the
system design for implementing Fuzzy C-Means clustering.

Acquire the
Landsat image

Minimize the
objective function

Fuzzy partitioning
by optimization

Water body and

non-water body
cluster

Figure 5.5 Fuzzy C-Means Clustering – system design

5.3.3.1 Fuzzy C-means implementation steps

The FCM function is called upon which takes the pixel array
and the number of clusters as input.
 118

The FCM function returns the objective function values,

membership grades and the cluster centers as output.

Based on the degree of membership the data points are moved

into different clusters.

Given a data set, there are n training samples. The values for
the parameter X x1 , x 2 ,...x n denoted as Y sort X y1 , y 2 ,... y n . The

parameter Y indicates the ascending order of the parameter X.

Once the X values are sorted it is easy to find out the
association between adjacent values.

Next determine the difference between the set of training data

as mentioned in Equation (5.6).

di Yi 1 Yi , i 1, 2,....n 1. (5.6)

where Yi and Yi 1 are adjacent values.

Calculate the adjacent value similarities by using Equation (5.7).

di
1 , di CP s
Si CP s
(5.7)
0 , otherwise

where d i is the difference, s is the standard deviation of d i and

CP is the control parameter.

A threshold value s divides adjacent values into classes.

Determination of clusters can be recapitulated by a rule. This
can be expressed in Equation (5.8) as,
 119

IF s i THEN Yi , Yi 1 CPi
(5.8)
ELSE Yi CPi , Yi 1 CPi 1

where CPi and CPi 1 denote two dissimilar classes for the input or output
parameter.

5.4 THE SUPERVISED CLASSIFICATION

Huang et al (2009) presented supervised classification identifies the

specific area of interest i.e., water body and non-water body in the image.
The supervised classification is much more accurate for mapping classes, but
depends heavily on the cognition and skills of the image specialist. The
strategy is simple: the specialist must recognize conventional classes or
meaningful classes in a scene from prior knowledge, such as personal
experience with what's present in the scene, or more generally, the region it's
located in, by experience with thematic maps, or by on-site visits. This
familiarity allows the individual making the classification to choose and set
up discrete classes and then, assign them category names.

As a rule, the classifying person also locates specific training sites

on the image to identify the classes. The resulting training sites are areas
representing each known land cover category that appear fairly homogeneous
on the image. For each class thus outlined, mean values and variances of the
each band used to classify them are calculated from all the pixels enclosed in
each site. More than one polygon is usually drawn for any class. The
classification program then acts to classify the data representing each class.
When the classification for a class is plotted as a function of the band the
result is a spectral signature or spectral response curve for that class. The
multiple spectral signatures so obtained are for all of the materials within the
site that interact with the incoming radiation.
 120

Classification now proceeds by statistical processing in which every

pixel is compared with the various signatures and assigned to the class whose
signature comes closest. Many of the classes for the satellite images are
almost self-evident in portraying ocean water, waves, beach, marsh, shadows.
For example, difference between ocean and bay waters, but their gross
similarities in spectral properties would probably make separation difficult.
Some classes are broad-based, representing two or more related surface
materials that might be separable at high resolution but are inexactly
expressed in the image. Thus, the supervised classification has an edge over
the unsupervised methodology. Learning Vector Quantization (LVQ) and the
Support Vector Machines (SVM) are the two supervised classification
methodologies implemented to perform the analysis for the water body and
non-water body classification.

5.4.1 Learning Vector Quantization

Zhou Kaili (2005) Learning Vector Quantization (LVQ) is also used

in feature vector dimension reduction. Here is the chance to optimize training
data through reducing the number of samples for this analysis. LVQ is a
variation of the well-known Self Organizing Map (SOM) architecture. The
SOM system is known as a Kohonen network. This has a feed-forward
structure with a single computational layer of neurons arranged in rows and
columns. Each neuron is fully connected to all the source units in the input
layer. The architecture is being depicted in the Figure 5.6. A one dimensional
map will just have a single row or column in the computational layer. The
feature map is being derived from the spatially continuous input space, in
which input the vectors live, to the low dimensional spatially discrete output
space, which is formed by arranging the computational neurons into a grid.
 121

Computational
layer

Input layer

Figure 5.6 Arrangement of neurons in multi-dimensional SOM

architecture

5.4.1.1 Stages of LVQ

The different stages of LVQ are

Step 1: Initialization – Assign the initial weight vector w j by selecting the

random values.

Step 2: Sampling – Describe a sample input vector x from the input space.
Here the network chooses five texture features as input to classify
the two classes.

Step 3: Matching – Find weight vector closest to the input vector named
as winning neuron I (x ) , which is defined as in Equation (5.9).

n
2
xi w ji (5.9)
i 1
 122

Step 4: Updating – Apply the following weight update function to update

the weights as shown in Equation (5.10).

w ji T j ,I x xi w ji (5.10)

where T j , I x
is a Gaussian neighbourhood and is the learning

rate.

Step 5: Continuation – Repeat the steps 2 to 5 to reach the best accuracy of

input space.

5.4.1.2 Properties of the feature map

Once the SOM algorithm has converged, the feature map displays
important statistical characteristics of the input space. Given an input
vector x , the feature map provides a winning neuron I (x ) in the output
space and the weight vector WI ( x ) provides the coordinates of the image of that

neuron in the input space as shown in Figure 5.7.

Feature map
Continuous
input space
WI(x)

Discrete space

Figure 5.7 Topological ordering and density matching of the feature map
 123

5.4.1.3 Basics of vector quantization

The main aim of using a self organizing map is to encode a large set
of input vectors x by finding a smaller set of “representatives” or

“prototypes” or “code-book vectors” WI ( x ) that provide a good

approximation to the original input space. This is the basic idea of vector
quantization theory. The motivation of which is dimensionality reduction or
data compression. In effect, the error of the vector quantization approximation
is the total squared distance between the input vectors x and their
representatives WI ( x ) as defined in Equation (5.11).

D || x w I ( x ) || 2 (5.11)
x

A gradient descent style minimization of D does lead to the SOM

weight update algorithm, which confirms that it is generating the best possible
discrete low dimensional approximation to the input space.

5.4.2 Support Vector Machine classifier

Support Vector Machines were introduced by Vladimir Vapnik

(1995). A support vector machine (SVM) is a concept for a set of
related supervised learning methods that analyze data, recognize patterns and
then used for classification and regression analysis. The classification
problem can be restricted to consideration of the two-class problem without
loss of generality. The goal is to separate the two classes by a function which
is induced from available training data sets.
 124

Nonlinear Problem Linear Problem

Maps

SVM Hyperplane

Figure 5.8 Working of SVM

For each given input, the SVM determines if the input is a member
of water body or non-water body class. This makes SVM as a linear classifier.
The goal is to separate the two classes without loss of generality by a function
which is induced from available training data sets. The task is to produce a
classifier that will work in a generalized manner. The application of SVM for
the desired problem is minimizing the error through maximizing the margin
which means that it maximizes the distance between it and the nearest data
point of each class. Since SVMs are known to generalize well even in high
 125

dimensional spaces under small training samples, this linear classifier is

termed as the optimal separating hyper plane as shown in Figure 5.8.

The SVM takes a set of input data and then predicts, for each given
input, which of two possible classes. Given a set of training data, each marked
as belonging to one of two categories, an SVM training algorithm builds a
model that assigns new data into one category or the other.

SVM is used to construct a decision plane called the hyper plane

which separates the water body class from non-water body class. Figure 5.9
illustrates the SVM classifier for two classes: water body and non-water body.

Water body

Figure 5.9 SVM Classifier

5.4.2.1 SVM -Background

Given l training examples {xi , yi }, i = 1, , l , where each

example has d inputs (xi Rd ), and a class label with one of two
 126

values (yi { 1, 1}). Now, all hyperplanes in Rd are parameterized by a

vector (w),and a constant (b), expressed in the equation (5.12) as follows

w x+b=0 (5.12)

Given such a hyperplane (w,b) that separates the data, this gives
the function defined in equation (5.13) as

f (x) = sign(w x + b) (5.13)

which correctly classifies the training data (and hopefully other “testing”
data it hasn’t seen yet). However, a given hyperplane is represented by
(w,b)is equally expressed by all pairs { w, b} for R+ . So define the
canonical hyperplane to be that which separates the data from the hyper-
plane by a “distance” of atleast 1. That is, consider those that satisfy
equation (5.14) and (5.15):

xi w+b +1 when yi = +1 (5.14)

xi w+b 1 when yi = 1 (5.15)

or more compactly:

yi (xi w + b) 1 i (5.16)

All such hyperplanes have a functional distance 1 (quite literally,

the function’s value is 1). This shouldn’t be confused with the geometric or
Euclidean distance (also known as the margin). For a given hyperplane
(w,b), all pairs { w, b} define the exact same hyperplane, but each has a
different functional distance to a given data point. To obtain the geometric
distance from the hyperplane to a data point, it can be normalized by the
 127

magnitude of w.

Intuitively, the hyperplane that maximizes the geometric

distance to the closest data points.

5.5 ACCURACY CALCULATION

Defries and Chan (2000) described the accuracy of SVM

classification can be calculated in terms of classified rate, correct rate and
error rate. The classified rate is the ratio of classified samples to the total
number of samples. The correct rate is the ratio of correctly classified samples
to the classified samples and the error rate is the ratio of incorrectly classified
samples to the classified samples. These are defined as in Equations (5.17) –
(5.19) as follows:

Classified Samples
ClassifiedRate (5.17)
Total no of samples

Correctly Classified Samples

Correct Rate (5.18)
Classified Samples

Incorrectly Classified Samples

Error Rate (5.19)
Classified Samples

Powell et al (2004) and Congalton et al (2002) have discussed the

accuracy of the classification can also be calculated in terms of Receiver
Operating Characteristics (ROC) terminology, true positive (TP), false
positive (FP), true negative (TN) and false negative (FN). True positive rate is
the proportion of positive cases that were correctly identified and false
positive rate is the proportion of negatives cases that were incorrectly
 128

classified as positive. Classification accuracy is the proportion of the total

number of predictions that were correct.

Sensitivity is the true positive, the region of water body which is

correctly classified as positive. The true negative is speci city, the region of
water body who actually are without correctly classified as negative. The
implication of this is false positive and false negative can also be measured.
This is an important assumption of ROC analysis, making ROC useful for
highly prevalent classification. The typical ROC curve places FP on the
horizontal axis and TP on the vertical axis.

5.6 EXPERIMENTAL RESULTS AND DISCUSSIONS

5.6.1 Data Sets

This work deals with the Land Remote-Sensing Satellite

(LANDSAT) images taken from different time frames of Kochi,
Kanyakumari, Kolkata Visakhapatnam and Sydney regions.

5.6.2 Texture Features to be Subjected

The Figure 5.10 shows the texture features such as energy, entropy,
contrast, inverse difference moment and directional moment extracted using
GLCM along with the corresponding classes such as water body and non-
water body.
 129

Figure 5.10 Input data consisting of five texture features with the
corresponding classes

5.6.3 Training using Neural Network

Then the input samples are trained using neural network. The
numbers of input training data sets are 1026. The network is designed by 5
input layers, 30 hidden layers and 1 output layer. The activation function of
each output neuron is a radial function, i.e., a monotonously decreasing
function.
 130

Figure 5.11 Training using neural network

The term ‘Performance goal met’ indicates that the training data
satisfies the necessary and sufficient conditions in order to achieve the
maximum accuracy. Thus, there has been a successful result meeting 100%
accuracy which is shown in Figure 5.11 and Figure 5.12.
 131

Figure 5.12 MSE and accuracy in percent of the classified results

5.6.4 Results of Classification

The support vector machine classifies the images into water body
and non-water body and this shows that the resolution enhanced image gives
better accuracy than the denoised image. Figure 5.13 to Figure 5.17 show the
water body and non-water classified result of Kochi, Kanyakumari, Kolkata,
Visakhapatnam and Sydney region. In the output image, white colour shows
the water body and black colour shows the non-water body region.
 132

(a) Resolution enhanced input image

(b) Classified output

Figure 5.13 Experimental results of (a) resolution enhanced input image

and (b) Water body and non-water body classified result of
Kochi region
 133

(a) Resolution enhanced input image

(b) Classified output

Figure 5.14 Experimental results of (a) resolution enhanced input image

and (b) water body and non-water body classified result of
Kanyakumari region
 134

(a) Resolution enhanced input image

(b) Classified output

Figure 5.15 Experimental results of (a) resolution enhanced input image

and (b) water body and non-water body classified result of
Kolkata region
 135

(a) Resolution enhanced input image

(b) Classified output

Figure 5.16 Experimental results of (a) resolution enhanced input image

and (b) water body and non-water body classified result of
Visakhapatnam region
 136

(a) Resolution enhanced input image

(b) Classified output

Figure 5.17 Experimental results of (a) resolution enhanced input image

and (b) water body and non-water body classified result of
Sydney region
 137

5.6.5 Accuracy Calculation

The support vector machine classifies the images into water body
and non-water body and this shows that the resolution enhanced image gives
better accuracy than the denoised image. The classification accuracy of
denoised images and resolution enhanced images are shown in Table 5.1.

Table 5.1 Classification accuracy of denoised and DWT based

interpolation technique resolution enhanced images

Resolution Enhanced
HDL Denoised Image
Sl. Image
Region Name
No Classified Correct Error Classified Correct Error
Rate Rate Rate Rate Rate Rate

1 Kochi 100 45 55 100 65 35

2 Kanyakumari 100 60 40 100 85 15

3 Kolkata 100 40 60 100 55 45

4 Vishakhapatnam 100 55 45 100 60 40

5 Sydney 100 60 40 100 65 35

5.6.5.1 ROC Analysis

The Receiver Operating Characteristics (ROC) analysis is a

technique for visualizing, organizing and selecting classifiers based on their
performance. ROC curve shows true positive (TP) rate versus false positive
(FP) rate (sensitivity versus specificity) for different thresholds of the
classifier output. This analysis can be used to find the threshold value that
 138

maximizes the classification accuracy or to assess how the classifier performs

in the regions of high sensitivity and high specificity.

The ROC curve can compute values for various criteria to plot either
on the X-axis or on the Y-axis. This returns values of specificity, or false
positive rate in the X-axis and sensitivity, or true positive rate in the Y-axis.
All such criteria are described by a confusion matrix.

The confusion matrix, C is defined as

TP FN
C (5.15)
FP TN

where TP- True Positive, FP - False Positive, FN - False Negative and

TN - True Negative.

The first row of the confusion matrix defines how the classifier
identifies instances of the positive class: C 1,1 is the count of correctly

identified positive instances and C 1,2 is the count of positive instances

misidentified as negative.

The performance of the SVM classifier is also evaluated using ROC

curve. The ROC curve and confusion matrix of various regions are depicted
in Figures 5.18 to 5.22.
 139

(a) Confusion matrix

(b) ROC curve

Figure 5.18 Performance analysis (a) confusion matrix and (b) ROC
curve of Kochi region
 140

(a) Confusion matrix

(b) ROC curve

Figure 5.19 Performance analysis (a) confusion matrix and (b) ROC
curve of Kanyakumari Region
 141

(a) Confusion Matrix

(b) ROC curve

Figure 5.20 Performance analysis (a) confusion matrix and (b) ROC
curve of Kolkata Region
 142

(a) Confusion Matrix

(b) ROC curve

Figure 5.21 Performance analysis (a) confusion matrix and (b) ROC
curve of Vishakhapatnam region
 143

(a) Confusion matrix

(b) ROC curve

Figure 5.22 Performance analysis (a) confusion matrix and (b) ROC
curve of Sydney region
 144

The ROC analysis results show that the Kochi region has a sensitivity
of 80% and specificity of 86.7% and the accuracy is 85%, Kanyakumari region
has a sensitivity of 62.5% and specificity of 91.7% and the accuracy is 80%,
Kolkata and Vishakhapatnam region have a sensitivity of 71.4% and specificity
of 92.3% and the accuracy is 85% and Sydney region has a sensitivity of
83.3% and specificity of 92.9% and the accuracy is 90%.

5.6.6 CPU Time Calculation

The complexity of the proposed techniques, Hybrid Directional

Lifting and DWT based Interpolation techniques are analyzed based on the
CPU time or process time. The CPU time of a function is the amount of time
(in seconds) for which a central processing unit is used to process the
instructions of a computer program. Figure 5.23 shows the computation time
of HDL and DWT based resolution enhancement techniques.

Analysis of CPU Tim e

0.08
0.06
0.04
0.02
0
Kochi Kanyakuma Kolkata Vishakhapa Sydney

HDL 0.03125 0.0625 0.0625 0.046875 0.03125

DWT 0.015625 0.046875 0.0325 0.015625 0.015625
Region Names

Figure 5.23 Computation time of HDL and DWT based resolution

enhancement technique
 145

From the results of computation time, it is inferred that HDL and

DWT based interpolation techniques are efficient to execute the algorithm.

5.7 APPLICATION – CHANGE DETECTION

Change detection is the art of quantifying the changes in synthetic

aperture radar images occurring over a period of time. Remote sensing has
been instrumental in performing change detection analysis.

Remote sensing and related techniques have a persistent impact on

the conduct of real time work. With satellite instruments, it is possible to
observe a target repeatedly, thereby contributing effectively to perform
change detection in areas of interest. This application mainly focuses on
change detection, which happens because of many possible environmental
and human actions. SAR images enable direct observation of the land surface
at repeated intervals, allowing mapping, monitoring and assessment. Change
detection analysis is very effective in long-term planning.

5.7.1 Data Sets

This deals with the Land Remote-Sensing Satellite (LANDSAT)

images taken from different time frames of the coastal regions of the world.
Some of the sample input study area imageries are shown in Figure 5.24 and
Figure 5.25.
 146

Kochi Region Indonesia Region Kanyakumari Region

Figure 5.24 Input landsat images of coastal landscape taken during 2005

Kochi Region Indonesia Region Kanyakumari Region

Figure 5.25 Input landsat images of coastal landscape taken during 2010

5.7.2 Results of Classification through PCA based K-Means Clustering

The input SAR images have been subjected to unsupervised

classification performed using a PCA based K-means classifier. Figure 5.26
and Figure 5.27 depict the result of classification which facilitates
differentiation of water body from non-water body region. These results have
been further utilized in calculating the percentage of coverage area to detect
changes over a period of time.
 147

5.7.3 Results of Supervised Classification

The texture features are extracted using GLCM and subjected to

LVQ to satisfy the necessary and sufficient conditions to achieve maximum
accuracy. The training data has led to results with 100% accuracy.

Subjecting the combination of extracted texture features onto the

SVM classifier resulted in 98% accuracy for classifying input data. The
classified results are depicted in Figure 5.28.

Kochi Region Indonesia Region Kanyakumari Region

Figure 5.26 Clustering performed on 2005 images

Kochi Region Indonesia Region Kanyakumari Region

Figure 5.27 Clustering performed on 2010 images

 148

Water body
Figure 5.28 Classification Using SVM

So it is possible to quantifiably depict the amount of increase of

water body in 2010 when compared to 2005 as shown in Figure 5.29.

Kochi Region Indonesia Region Kanyakumari Region

Figure 5.29 Change depicted over five years

 149

Table 5.2 Coverage areas of water body and non-water body of

different landscapes in percentage

2005 2005 2010 2010 Changes

detected
Area Non- Non- in the
Water water
water water period of
5 years
Kochi 26.1200 73.88 29.2 70.8 3.08%
Indonesia 46.6476 53.3524 46.6599 53.3401 0.01%
Kanyakumari 46.4966 53.5034 44.2856 55.7144 0.16%
Vishakhapatnam 62.8494 37.1506 63.5956 36.4044 0.74%
Sydney 56.3477 43.6523 60.0342 39.9658 3.68%
Kolkata 27.4445 72.555 31.7780 68.2220 4.333%

5.7.4 Accuracy Calculation

The percentage of changes in water body and non-water body can be

interpreted from the results shown in Table 5.2. The count has been
materialised by taking pixel values of the classified image into account as the
input, and then pixels of the water region are counted first which leads to the
count of the non-water body. The change in coverage area of water body has
recorded a significant increase over five years in the study region, as shown in
Table 5.2. The increase in the coverage areas of water bodies due to global
warming has been substantially proven through these results.

5.8 APPLICATION – GLACIER CLASSIFICATION

The advent of global warming and thus the glacier meltdown has
been a serious issue amongst the various environmental bodies of the world.
The efficient analysis of LANDSAT images can provide valuable results and
 150

information which can help strategize the preventive steps taken by the
environmental bodies. The main aim of this is to compare the K-Means and
Fuzzy C-Means clustering technique and find out the change detection in
glacier classification by processing images taken over different time frames.

The classification of glaciers and determining the percentage of

change in glaciers over a time frame can give an accurate idea of the
repercussions of global warming and thus strategize environmental protection
plans for the future. Usage of unsupervised glacier classification over
supervised glacier classification can lead to reduction in manual labor and
automation of research.

5.8.1 Data Sets

The LANDSAT images correspond to the Himachal Pradesh region,

one dated June 2005 and the other dated June 2010 are taken. The Figure 5.30
and Figure 5.31 depict the input LANDSAT images of Himachal Pradesh
region during June 2005 and June 2010.

Figure 5.30 June 2005 image of Himachal Pradesh region

 151

Figure 5.31 June 2010 image of Himachal Pradesh region

5.8.2 Results of K-Means Clustering

The input images have been subjected to K-Means clustering. The

Figure 5.32 and Figure 5.33 show the results of K-Means clustering that gives
the change of glacier from non-glacier in the given input image. Using the
outputs it can be very easy to find out the percentage of change of glacier
from non-glacier.

Figure 5.32 Objects identified by K-Means clustering during June 2005

 152

Figure 5.33 Objects identified by K-Means clustering during June 2010

The percentage of glacier and non-glacier in June 2005 image is

36% and 63% and June 2010 is 29% and 71%. So, the change in coverage
area of glacier is a significant decrease over five years in the given image by
7% and vice-versa.

5.8.3 Results of Fuzzy C-Means Clustering

The Figure 5.34 and Figure 5.35 portray the results of June 2005
and June 2010 image after applying Fuzzy C-Means clustering. The
percentage of change of glacier in June 2005 image over June 2010 image is
8%. In accordance with the real time data, it gives a confidence to indicate
that the work to be performed will yield a fruitful result.
 153

Figure 5.34 June 2005 image after applying Fuzzy C-Means clustering

Figure 5.35 June 2010 image after applying Fuzzy C-Means clustering

The glacier classification using the Fuzzy C-Means approach yields

a superior result compared to K-Means approach. Using the K-Means
approach there are a number of unclassified glacier and so the classification
accuracy is low compared to the Fuzzy C-Means approach. According to the
ground truth the Fuzzy C-Means clustering derives the better result compared
to the K-Means clustering.
 154

5.9 SUMMARY

This chapter explained how the supervised and unsupervised

classification techniques are applied for change detection, glacier and water-
body classification approaches. The accuracy of the classification techniques
are evaluated by receiver operating characteristics. The next chapter provides
a comparative study of how all the techniques discussed so far are used for
improving classification accuracy.