Anda di halaman 1dari 53

A THESIS ON

CLASSIFICATION ON SATELLITE IMAGE


USING A SOFT CLASSIFICATION ALGORITHM
Submitted in partial fulfillment of the requirement for the award of degree of
Master of Technology
In
Computer Science & Engineering

By
SHIVANGI SHARMA
(150530705007 )

Under the supervision of:


Mr Manish Sharma
Professor

Department of Computer Science & Engineering


JB INSTITUTE OF TECHNOLOGY, Dehradun
UTTRAKHAND TECHNICAL UNIVERSTY
DEHRADUN(U.K)

1
Department of Computer Science & Engineering

CERTIFICATE

I hereby certify that the work which is being presented in M. Tech assessment entitled
CLASSIFICATION ON SATELLITE IMAGE USING A SOFT
CLASSIFICATION ALGORITHM in partial fulfillment of the requirement for the
award of degree of the Master of Technology in computer science and engineering and
submitted to the department of Computer Science and Engineering, J B Institute of
Technology, Dehradun(U.K.) is an authentic record of my own work carried out during
period from August 2016 to December 2016 under the supervision of Mrs. Ranjana
Sharma Assist Professor, TMU, Moradabad.

The matter presented in this assessment has not been submitted by me or anybody else
for the award of any other degree elsewhere.

Candidate Name: Shivangi Sharma


Enrolment Number: 150530705007

This is to certify that the above statement made by the candidate is correct to the best of
my knowledge.

Date

Mrs. Ranjana Sharma Mr Diwaker


(Supervisor) M.Tech

2
DECLARARATION

I hereby declare that the thesis entitled CLASSIFICATION ON SATELLITE IMAGE


USING A SOFT CLASSIFICATION ALGORITHM submitted for the award of degree of
Master of Technology in Computer Science and Engineering from J.B Institute of
Technology, Dehradun embodies results of original work and studies carried out by me and the
contents of the thesis do not form the basis for the award of any other degree to me or to anybody
else from this or any other University/Institute.

Shivangi Sharma
M.Tech(CSE).
Enrolment No: 150530705007
J.B Institute of Technology, Dehradun

3
ACKNOWLEDGEMENT

I would like to take this opportunity to express my deep sense of gratitude to all who helped me
directly or indirectly during this thesis work.

First of all I would like to express my deepest gratitude to my thesis supervisor Mr Manish
Sharma and his enormous help and advice and for providing inspiration which cannot be
expressed with words. I would not have accomplished this thesis without his patient care,
understanding and encouragement. His advice, encouragement and critics are source of
innovative ideas, inspiration and causes behind the successful completion of this thesis work.
The confidence shown on me by her was the biggest source of inspiration for me.

I would like to express my deep gratitude to Mr. Diwaker Maurya on behalf of the head of
department computer science and engineering department of Jb Institute of Technology, for
providing access to the good integration of intellectual properties, technical support and
facilities.

I am deeply thankful to JBIT Management for providing facilities for accomplishment of this
assessment.

Date.. Shivangi `Sharma


(150530705007)

4
ABSTRACT

Remote sensing has gained increasing attention from researchers and scholars. Remote sensing is
the science of obtaining information about objects and areas from a distance, typically from
aircraft or satellite. One of the major areas in remote sensing is image classification. Image
classification is done using the reflectance statistics for individual pixels. Classification is a
widely used technique for image processing and is used to extract thematic data for preparing
maps in remote sensing application. This assessment presents a study of image classification
through soft classifiers. Soft classifier helps in the development of more robust methods for
remote sensing application as compared to the hard classifier. In this two supervised soft
classifier are used. These classifiers are FCM and PCM. Supervised classifiers are human guided
classifiers. FCM is known as Fuzzy c-means. PCM is known as Possibilistic c-means.

5
Table of Contents
CERTIFICATE2
DECLARATION..3
ACKNOWLEDGEMENT...4
ABSTRACT..5

LIST OF FIGURES.7
LIST OF TABLE.8

CHAPTER 1. INTRODUCTION9-14
1.1 Introduction10
1.2 Research objective..13
1.3 Soft Classifiers Investigated...13
1.4 Mixed Pixel In Training Stage...13
1.5 Pre requisite for the development of software13
1.5.1 Software requirement14
1.5.2 Hardware requirement.14

CHAPTER 2. REVIEW15-20
2.1 Introduction16
2.2 Problem of mixed pixels.17
2.3 A review of soft and hard classification from remote sensing data17
2.4 Summary19

CHAPTER 3. SOFT CLASSIFICATION METHODS..21-27


3.1 Introduction.22
3.2 Fuzzy set based methods.23
3.2.1 Fuzzy c-means(FCM) clustering.23
3.2.2 Possibilistic c-means(PCM) clustering 26

CHAPTER 4. METHODOLOGY AND DATA28-32


4.1 Methodology..29
4.2 Study area..30
4.3 Data requirement30
6
4.4 Ancillary data32

CHAPTER 5. CONCLUSION AND RESULT ANALYSIS.33-42


5.1 Introduction..34
5.2 Result and discussion.34
5.3 Conclusion.42

APPENDIX.43-49

REFRENCE.50

7
LIST OF FIGURE

Fig 1: IFOV of different sensors of Resource Sat-1 (IRS-P6) satellite with


respect to single pixel19
Fig 2: Methodology29
Fig 3: Graph plotted on the basis of table 4...36
Fig 4: Graph plotted on the basis of table 537
Fig 5: Graph plotted on the basis of table 638
Fig 6: Graph plotted on the basis of table 7...39
Fig 7: Graph plotted on the basis of table 840
Fig 8: Graph plotted on the basis of table 941

8
LIST OF TABLE

Table 1: Land use and land cover classes present in the study area30
Table 2: Sensor specification of Resource Sat-1 (P-6).31
Table 3: Resource Sat-1 payload characteristic..31
Table 4: AWIFS entropy of various land cover classes from FCM classification output35
Table 5: AWIFS entropy of various land cover classes from PCM classification output.36
Table 6: LISS III entropy of various land cover classes from FCM classification output37
Table 7: LISS III entropy of various land cover classes from PCM classification output38
Table 8: LISS IV entropy of various land cover classes from FCM classification output39
Table 9: LISS IV entropy of various land cover classes from PCM classification output40

9
CHAPTER 1

INTRODUCTION

10
1.1 INTRODUCTION

Remote sensing is the scanning of the earth by satellite or aircraft to obtain information about it.
Remote sensing is a method for getting information about of different objects on the planet,
without any physical contacts with it. Remote sensing uses a part or several parts of the
electromagnetic spectrum. Its records the electromagnetic energy reflected or emitted by the
earths surface. Humans accomplish this task with aid of eyes or by the sense of smell or hearing;
so, remote sensing is day-today business for people. Reading the newspaper, watching cars
driving in front of you are all remote sensing activities. Most sensing devices record information
about an object by measuring an objects transmission of electromagnetic energy from reflecting
and radiating surfaces.

Remote sensing may be split into "active" remote sensing which means when a signal is emitted
by a satellite or aircraft and its reflection by the object is detected by the sensor and "passive"
remote sensing which means when the reflection of sunlight is detected by the sensor. Passive
sensors gather radiation that is emitted or reflected by the object or surrounding areas.
Reflected sunlight is the most common source of radiation measured by passive sensors.
Examples of passive remote sensors include film photography, infrared, charge-coupled devices,
and radiometers. Active collection, on the other hand, emits energy in order to scan objects and
areas whereupon a sensor then detects and measures the radiation that is reflected or
backscattered from the target. RADAR and LIDAR are examples of active remote sensing where
the time delay between emission and return is measured, establishing the location, speed and
direction of an object.

Remote sensing makes it possible to collect data of dangerous or inaccessible areas. Remote
sensing applications include monitoring deforestation in areas such as the Amazon
Basin, glacial features in Arctic and Antarctic regions, and depth sounding of coastal and ocean
depths. Remote sensing also replaces costly and slow data collection on the ground, ensuring in
the process that areas or objects are not disturbed.

The clustering is a method in analyzing remote sensing images, which generates a direct
overview of objects. Clustering is a grouping of data with similar characteristics. This
"similarity" in a given set may vary according to data, because clustering is used in various fields
such as numerical taxonomy, morph metric, systematic, etc. Thus, a clustering algorithm that fits
the numerical measure of optimization in a data may not optimize another set of data (for
example, depending on the units selected). There are many algorithms to solve a clustering
problem. The algorithms used in our applet concentrate on "joining", "splitting", and "switching"
search methods (also called bottom up, top down, and interchange, respectively). The clustering
method uses a three-dimensional model for demonstration purposes. However the algorithms can
calculate clusters in dimensions. We are planning to an algorithm based on principal component
analysis, projecting the dimensional data onto a three dimensional space. There are many

11
methods of clustering developed for wide variety of purposes. Clustering algorithms used for
unsupervised classification of remote sensing data vary according to the efficiency with which
clustering takes place. Clustering is divided into hierarchical clustering and non-hierarchical
clustering. One of the most popular and fastest clustering techniques is the K-means technique.
The results of the K-means technique depend on different factors such as a method of
determination of initial cluster centers.

Remote sensing is a popular and fast way to acquire information on geographic surface. The
multispectral images are main source of information on geo-surface. Classification is used to
obtain this data and constructing maps. Classification is an important part of image processing
and by this process we can use to gathering thematic data and maps. But this is Classification is
not meaningful or incomplete without assessment of accuracy. By the definition of Accuracy
assessment of a classification if focus on how accurately the classification process was done for
the data. In general, the classification is crisp and assigns a single class to each image elements
or pixel, known as hard classification. However, there is another consequence that some pixels
may not purely belongs to single class. This problem treated as mixed pixel problem. For course
resolution remote sensing images, the probability of mixed pixel is high as the ground coverage
for a pixel is considerably enough to contain multiple lands cover classes in it (Bastin, 1997;
Watanachaturaporn et al., 2006). This problem has been taken into account by soft or sub-pixel
classification. Scientists and practitioners have made great efforts in developing advanced
classification approaches and techniques for improving classification accuracy.

The pixel of data in soft classification deepened according to area or their belongingness. The
image in soft classification is equal to no of land cover classes. The fuzzy based algorithm has
received growing interest for their particular value in situations where the geographical
environment.

From amongst a number of soft classifiers, in this two statistical classifier- maximum likelihood
classifier(MLC) and linear mixture model(LMM), two fuzzy set theory based classifiers- fuzzy
c-means(FCM) and possibilistic c- means(PCM). Due to its wide acceptance MLC has been used
as a bench mark to evaluate other soft classifier studies. A new fuzzy set clustering algorithm
PCM has been introduced in this to get over the limitation of FCM.

The use of the soft classifier however partially resolves the problem of mixed pixels as these
have been accounted for in the allocation stage only. For images dominated with mixed pixels,
their incorporation in other stages may also be necessary. Suitable modifications have been made
in the training stage of all supervised classifier in order to accommodate mixed pixels in this
stage, which is one of the main objectives of this research. In the testing stage, mixed pixels have
been incorporated via the use of distance, entropy and fuzzy set based measure. The hypothesis
of fuzzy error matrix has been promoted to assess the accuracy of soft classification. Whatever
be the origin of mixed pixels, these may create problems in the image classification. For
instance, a mixed pixel displays a composite spectral response that may be dissimilar to the
spectral response of each of its component classes and, therefore, the pixel may not be allocated
to any of its component classes (Zhang and Foody, 2001). Hence error is likely to occur in the
classification methods, which assume that the pixels are pure, force the mixed pixels to be

12
allocated to one and only one class. This may result into a loss of pertinent information present in
a pixel. Mixed pixels may thus be treated as error, noise or uncertainty in class allocation for
hard classification method. These methods may therefore tend to.

Other soft classification method can be used as Artificial Neural Network algorithm both as
supervised as well as unsupervised Artificial Neural Network is a learning algorithm but does not
always generalize well. Other emerging classifiers are nonparametric learning algorithm known
as Support Vector Machine (SVM) Support Vector Machines (SVM) is a group of theoretically
superior machine learning algorithms. It is found to be competitive with the best available
machine learning algorithms in classifying high dimensional data sets (Huang et al. 2002).SVM
can work well with a small training data set yielding high classification accuracy. By this soft
classification techniques for Remotely Sensed Data.

The utilization of soft classification methods is still an active research area, which can be gauged
from a number of research papers. A review on this subject as provided in chapter 2 indicates
that although the problem of mixed pixels has been investigated for some time now but it has got
renewed interests with the launch of a number of satellites carrying sensors at varied spatial,
spectral and temporal resolutions. It has been observed that,

i. The soft classification methods are largely in their exploratory stage. The research needs
to be conducted to examine these methods on different remote sensing data products
acquired in complex and uncertain environments.

ii. Most of the studies have focused on generation of soft classification output(i.e. the
allocation stage) only. The images acquired over uncertain environments shall contain a
large number of mixed pixels. Therefore, their incorporation in training and testing stages
of a supervised classification process becomes mandatory.

iii. Many of the classifier is sensitive to a number of parameters. The determination of


optimum number of parameters and their values for each classifier have to be
investigated further.

13
1.2 RESEARCH OBJECTIVES

In view of a number of gaps, research needs to be targeted to significantly advance the work on
classification of images dominated by mixed pixels, which is vital to the current needs of the
society. The major research objectives may therefore be stated as

i. We are introducing soft classifier for mixed pixel data.

ii. To consider the mixed pixels in the testing stage of classification for better accuracy.

iii. To develop a dedicated, user-friendly and interactive software package as there is no


software available to fulfill the stated research needs.

iv. To assess the suitability of fuzzy set theory based algorithm (like, fuzzy c-means,
probabilistic c-means) for sub-pixel classification.

1.3 SOFT CLASSIFIERS INVESTIGATED

This research has focused its focus is on two statistical classifiers, two fuzzy set theory based
algorithms and two neural network classifiers. The statistical classifier is maximum likelihood
classifier (MLC). It has widely been utilized to perform hard classification of remote sensing
data. This classifier has been modified suitably to produce soft classification. Due to its wide
acceptance, it has also been used as a bench mark to evaluate other soft classification methods
evaluated in this. Another statistical method is linear mixture model (LMM). It has been adopted
here to perform soft classification. Based on the assumption that spectral response of pixel is the
linear sum of spectral response of classes weighted by their corresponding proportional area.

1.4 MIXED PIXELS IN TRAINING STAGE

The outputs from the soft classifiers investigated here may appropriately represent land use land
cover within a mixed pixel. The effect of mixed pixels has not been taken into account in the rest
of the classification process. The methods are assumed that training data are essentially derived
from pure pixels only. Mixed pixels are incorporated in allocation stage only in soft
classification methods which is just one of the stages of a classification process. Since in an
image, a large proportional of pixels may be mixed, these may need to be incorporated in the
training stage of classification. Therefore, suitable modification has been made in the training
stage of all supervised classifiers used here to accommodate mixed pixels in this stage.

1.5 PRE-REQUISITE FOR THE DEVELOPMENT OF SOFTWARE

Following are the features that java programming language uses


Platform independent

14
A program or technology is said to be platform independent if which can run on all
available operating systems with respect to its development and compilation.

Robust
Simply means of robust are strong. It is robust or strong programming language because
of its capability to handle run time error, automatic garbage collection, the lack of
pointer concept, exception handling. All these points make it robust language.

Portable
If any language supports platform independent and architectural neutral feature known
as portable. The language like C, CPP, Pascal are treated as non-portable language. It is
a portal language.

Dynamic
It supports Dynamic memory allocation due to this memory wastage is reduce and
improve performance of the application. The process of allocating the memory space to
the input of the program at a run time is known as dynamic memory allocation.

Secure
It is a more secure language compared to other language. In this language, all code is
covered in byte code after compilation which is not readable by human.

Interpreted
It is one of the highly interpreted programming languages.

1.5.1 SOFTWARE REQUIREMENT

Programming language used is java and the platform used is JDK1.6.

The development environment has been taken Note-pad++ where no requirement of backend
because the concept of file handling is used for image retrieval.

1.5.2 HARDWARE REQUIREMENT

Hard disk 80 ghz


RAM 512 MB

15
CHAPTER 2

REVIEW

16
2.1 INTRODUCTION

The extraction of land cover from remote sensing Images has traditionally been viewed as a
classification problem where each pixel in the image is allocated to one of the possible classes.
However, in reality, different amounts of land cover mixing within a pixel can be found due to
the continuum of variation in landscape and intrinsic mixed nature of most classes. Therefore
pixel with mixture of classes (i.e. mixed pixel) may be problematic.

Remote sensed data of the earth may be analyzed to extract useful thematic information. It is
observed that data can be transformed into information. Multi spectral classification is one of the
most often used methods of information extraction. This procedure assumes that imaginary of a
specific geographic area is collated in multiple regions of the electromagnetic spectrum and that
the images are in good geometric registration. The multi spectral classification may be performed
using a variety of algorithms, including parametric and non- parametric. Traditional image
classification techniques assume that pixels are pure, however, in reality, different amount of
mixing of land cover within a pixel can be found due to the continuum of variation in landscape
and intrinsic mixed nature of most classes. Therefore, pixels with mixture of classes (i.e. mixed
pixel) may be problematic.

2.2 PROBLEM OF MIXED PIXELS

Remote sensing images are dominated by mixed pixels, which do not represent a single land
cover class but contain two or more classes. For instance, in India, within a small stretch, one
may find forest land, agricultural land, residential areas and water bodies. As a result, the land
cover classes are generally mixed in nature and inter-grade gradually in an area. The mixed
pixels also occur at the boundaries of two land cover classes.

Furthermore, mapping from remote sensing is generally carried out at regional and global scales,
which requires coarse spatial resolution images where the chances of occurrence of mixed pixels
are high. Error is likely to occur in the classification of image dominated by mixed pixels. The
conventional use of hard classification methods that allocate one class to a pixel may tend to over
and under estimate the actual areal extents of the classes on ground and thus may provide
erroneous result. Other methods are, therefore, desired to solve the problem of mixed pixel.
Further, in images contaminated with a large number of mixed pixels, sufficient number of pure
pixels may not be obtained to accurately estimate the training data statistics. Therefore, mixed
pixels need to be incorporated in all the stages of supervised classification so as to produce fully
fuzzy classification.

Remote sensing images contain a mixture of pure and mixed pixels. While digital image
classification, however, a pixel is frequently considered as a unit belonging to a single land cover

17
class. However, due to limited image resolution, pixels often represent ground areas, which
comprise two or more discrete land cover classes. For this reason it has been proposed that
fuzziness should be accommodated in the classification procedure so that pixels may have
multiple or partial class membership. In this case, a measure of the strength of membership for
each class in output by the classifier, resulting in a soft classification technique. Also recent
advances in supervised image classification have shown that conventional hard classification
techniques, which allocate each pixel to a specific class, are often inappropriate where mixed
pixels are abundant in the image.

Mixed pixels are assigned to the class with the highest proportion of coverage to yield a hard
classification. Due to which a considerable amount of information is lost. To overcome this loss,
soft classification was introduced. A soft classification assigns a pixel to different classes
according to the area it represents inside the pixel. This soft classification yields a number of
fraction images equal to the number of land cover classes. Several researchers have addressed
this soft mixture problem. Among the most popular techniques for soft classification are artificial
neural networks, mixture modeling and supervised fuzzy c means classification.

2.3 A REVIEW OF SOFT AND HARD CLASSIFICATION FROM REMOTE SENSING


DATA

While research on classification continues and many usual and potentially useful methods are
being proposed, it is interesting to note that the soft classification process and methods have not
found the place in the manner those may have. For instance, in this research, review of a sample
of 40 research papers published recently in refereed journals revealed that most of the remote
sensing studies have focused on hard classification. Limited studies have been implemented for
soft classification whereas only a few have shown the utility of incorporating mixed pixels in
various stages of supervised classification. Only one study has incorporated mixed pixels in all
stages of LMM, whereas three other studies have incorporated mixed pixels in the training stage
of BPNN and MLC. Some salient features of these studies are now reported in a chronological
order.

In an attempt to increase the accuracy of MLC in the absence of appropriate reference hence a
suitable training data, Bruzzone and Prito(2001) proposed an unsupervised retaining technique
by approximately estimating the parameters that characterize the classes. Thus a kind of hybrid
approach was suggested, which resulted in high classification accuracies without relying on the
corresponding training data sets. This study shows how the performance of a classifier can be
improved when training data is either not available or is insufficient. Later Brozzone et al., 2002
implemented this concept of unsupervised retraining on RBF neural networks also and improved
accuracies were obtained.

18
Eastman and Laney(2002) examined the assumption and the procedure of the commonly cite
Bayesian soft classification procedure for sub pixel classification and test its ability to uncover
mixture proportions. Mixed pixels were used in training stage through fuzzy mean and fuzzy
variance covariance matrices to estimate the underlying class signatures. The sub pixel
classifications were evaluated by cross tabulating the actual and predicted class proportions in
manner similar to traditional error matrix. After critical evaluation, they recommended that the
use of fuzzy mean and variance covariance may not be effective in accurately recovering class
proportions in pixels unless there exists substantial overlap in the distributions of constituent
classes. Therefore, MLC for soft classification should be used with caution.

In order to discriminated clouds and shadows from built-up areas and water bodies Foody (2002)
recommended the use of fuzzy classification technique over the hard ISODATA clustering
augmented with temperature and NDVI information.

AWIFS 56m

Picture
LISS-III 23.5m

LISS-IV 5.8m

Figure 1: IFOV of different Sensors of Resource Sat-1 (IRS-P6) Satellite with respect to single pixel
coverage of AWIFS payload

Laba et. al., 2002 utilized to concept of fuzzy operators (Gopal and Woodcok, 1994) to evaluate
the accuracy of regional scale land cover maps produced from remote sensing data. The results
showed that the assessment of soft classification using fuzzy operators resulted in an
improvement in map accuracy by about 19% to 23%. This study also emphasized the need of
using mixed pixels in testing stage for the accuracy assessment of soft classification.

19
Shan et. al. (2003) presents a case study on the use of fully fuzzy classification to map land cover
from IRS LISS III imagery. Two classifier namely FCM and MLC were used to produce soft
classification. The soft classifications were evaluated with soft reference data using cross
entropy, Euclidean distance and correlation coefficient. The values of cross entropy were 0.262
and 0.287 for MLC and FCM whereas corresponding Euclidean distances were 0.057 and 0.060.
Although, the results show that MLC is more accurate than FCM, it is not clearly reflected by the
magnitudes of either cross entropy or distance measure. Therefore, other accuracy measures for
soft classification need to be devised.

2.4 SUMMAERY

Digital image classification is perhaps the key task to derive accurate and efficient information
from remote sensing images. The usefulness of remote sensing images can be justified when the
mapping is carried out at the regional level. This requires classification of images at coarse
spatial resolution from sensors onboard a number of satellites. However, coarse resolution
images suffer with the presence of mixed pixels that represent more than one ground cover class
especially in the areas where the classes inter-grade gradually (Foody and Cox, 1994). The
conventional hard classification methods, which assume that the pixels belong to only one class
(i.e. pure pixels), will allocate the mixed pixels to one and only one class. Mixed pixels may
therefore be treated as noisy or uncertain pixels and their presence is notorious to the
classification process. A brief review of the recent studies in this chapter has clearly highlighted
the problem of mixed pixels in classification and their incorporation in various stages of
supervised classification (i.e. training, allocation and testing). A few attempts have been made to
tackle the problem of mixed pixels but a lot more is to be done before the solution can be
accepted at an operational level. Following specific conclusion may be drawn from this review,

i. A number of soft classifier has been devised for soft classification but most of them are
based on the training data consisting of pure pixels only.

ii. Inclusion of a priori probabilities in training the classification improves the accuracy.

iii. Mixed pixels in the testing stage of classification have been incorporated through
measures such as, root mean square error (Powell et. al., 2004; Ju et. al., 2003; Shalan et.
al., 2003; Foody, 2000a), Euclidean and L distance (Foody and Arora, 1996), cross-
entropy(Foody, 1995a) and correlation coefficients (Yanget. Al., 2003; Maselii et. Al.,
1996).

To further summaries the review presented in this chapter, it can be stated that,

20
i. The soft classification methods are largely in their exploratory stage.

ii. Most of the studies have focused on generation of soft classification output (i.e. the
allocation stage) only.

iii. There is small amount of work done on devising appropriate accuracy measures for
evaluation of soft classifier for remote sensing images.

In the next chapter, mathematical detail of soft classifiers investigated in this research are
provided.

21
CHAPTER-3

SOFT CLASSIFICATION
METHODS

22
3.1 INTRODUCTION

Both supervised and unsupervised classification may be applied to perform hard and soft
classification. In hard classification, pixel is allocated to one and only one class which may
produce erroneous results particularly in classifying coarse spatial resolution images. It is thus
important that soft classification is used to produce class proportion within a pixel ill order to
increase the classification accuracy and to produce meaningful and appropriate land cover
composition. The difference between hard and soft classification outputs can be depicted in
Figure 3.1. In this figure, a set of four pixels having mixture of classes tiled roof and grass are
shown. The hard classification will allocate pixel 1 to tiled roof and rest of the pixels to grass. On
the other hand, soft classification will provide us the class composition in the pixel at scale from
0 to 1, which may be advantageous extraction of land cover from remote sensing data.

One of the most popular fuzzy clustering methods is the fuzzy c-means (FCM) (Bezdek et.
al.1981, 1984) which is an unsupervised classifier that in an iterative process assigns class
membership values to pixels of an image by minimizing an objective function. Although a few
studies on the use of FCM have been reported, the major limitations of are the probabilistic sum
to one constraint. Therefore, besides using this classifier, another fuzzy set clustering method,
namely possibilistic c-mean (PCM) (Krishnaparam and Keller, 1993, 1996), which relaxes this
constraint so as to be robust to the noise (i.e. pixels with high degree of class mixtures) present in
the dataset, has also been implemented.

Another group of classification methods used here comes from artificial neural network (ANN)
family, which has its origin in machine learning theory. One supervised ANN method based on
back propagation learning algorithm (BPNN) (Foody, 1996), and one unsupervised ANN method
based on competitive leaning (CLNN) fowolled by learning vector quantizers (LVQ) have been
investigated. BPNN has been widely used in remote sensing but only for hard classification in
majority of the studies. There appears to be a couple of studies reported on the use of LVQ and
that too for hard classification only. Here, there two neural network classifier have been
implemented to produce soft classification.

3.2 FUZZY SET BASED METHODS

The use of fuzzy set base classification methods in remote sensing has received growing interests
for their particular value in situations where the geographical phenomena are inherently fuzzy
(Zhang and Foody, 1998). Although a number of fuzzy set based methods may be adopted,
mathematical details of only two unsupervised methods used here namely FCM and PCM are
provided here.

23
3.2.1 FUZZY C-MEANS (FCM) CLUSTERING

FCM, is an iterative method that may be employed to partition pixels of remote sensing image
into different class membership values, the key is to represent the similarly that a pixel shares
with each cluster with a function (membership function) whose values lie between zero and one.
Each pixel will have membership in every cluster. Membership close to unity signifies a high
degree of similarity between the pixel and that cluster (Bezdek, et.al., 1984). The net effect of
such a function for clustering is to produce fuzzy c-partition (U) of a given data. A fuzzy c-
partition of the data is the one which characterizes the membership of each pixel in the entire
cluster by a membership function which ranges from zero to one. Additionally, the sum of the
memberships for each pixel must be unity. This is achieved by minimizing the generalized least-
square error objective function.


N C
m 2
j m (U, V) = ij xi v j (3.1)
A
i 1 j 1

Subject to constraints

j 1
ij 1 For all I (3.2)

j 1
ij 0 For all j (3.3)

0 ij 1 For all i,j (3.4)

Where xi is the vector denoting spectral response I (i.e. a vector of spectral response of a pixel in
various band of a multi-spectral remote sensing image), V is the collection of vector of cluster
centers, vj, ij are class membership values of a pixel (members of fuzzy c-partition matrix), c
and n are number of cluster and pixels respectively, m is a weighting exponent ( 1 m ),
2
which control the degree of fuzziness, xi v j A
is the squared distance (dij) between xi and vj,
and is given by,

2
dij2 xi v j ( xi v j )T A( xi v j ) (3.5)
A

Where A is the weight matrix.

24
The role played by most of the variable exhibited in equationis clear. Two parameters
of jm are m and A. Weighted exponent m controls the relative weight placed on each of the
square error d2ij At m=1, partitions that minimize jm become increasingly hard and as m tends to
reach , partition s that minimize jm become increasingly soft or completely fuzzy.
Consequently, increasing m tends to degrade the membership toward the fuzziest state. Each
choice of m will generate one FCM solution keeping all other parameters fixed. No theoretical or
computational evidence provides an optimal m. if a test is available for the process under
investigation, the best strategy for selecting m at present seem to be experimental (Bezek et.
1984) and shall be adopted in this research.

Another parameter in jm that deserves a special mention is the weight matrix a. this matrix
control the shape of the optimal clusters. Amongst a number of A-norms. Three namely
Euclidean, Diagonal and Mahalonobis norm each induced by specific weight matrix, are widely
used. The formulations of each norm are given as

A=1 Euclidean Norm (3.6)

A=D-1 Diagonal Norm (3.7)

A=C-1 Mahalonobis Norm (3.8)

Where I is the identify matrix, Dj is the diagonal matrix having diagonal elements as the Eigen
values of the variance covariance matrix cj given by

N
C j ( xi c j )( xi c j )T (3.9)
i 1

N
Where C j xi / N
i 1

(3.10)

When the diagonal norm is used, each dimension is effectively scaled via the Eigen values. The
Euclidean norm is the only choice for which extension with the data is available. The type of
norm may have to before decide before carrying out classification using fuzzy c-mean clustering
method.

Now that all Parameters in Equations have been defined, the fuzzy c-partition is obtained
through an iterative process of optimization using the following two equation. The cluster centers
are updated by,

25
N

ijm xi
vj i 1
N
(3.11)
i 1
m
ij

And the class membership matrix ij is obtained by,

1
ij 1
(3.12)
c d 2
( m 1)


ij
2
k 1 d ik

c
Where dik2 dij2 (3.13)
j 1

The class membership values of a pixel denote the class proportions, which in turn may represent
the soft classified output for a pixel.

The general procedure of FCM algorithm may be formalized in the following steps,

Step 1. Fix values of m and c and select the type of A-norm and maximum number of iterations.
Step 2. Choose an initial membership matrix, U, with its element selected as random start,
random non-fuzzy start and uniform start. In random start, each membership value is assigned a
number between 0 and 1 to all classes. In random non-fuzzy start, a value of 1is assigned to a
randomly chosen class and 0 to all other classes. In uniform start, a membership value of (lle
a) is assigned to all classes where a is a small random number.
Step 3. Repeat
Step 5. Compute the distance based on the selected A-norm using equation 3.17. Update U
matrix for the next iteration.
Step 6. Update U matrix for the next iteration Ui+1 using equationuntil ( U i 1 U i ) where is
the user defined limiting error.
Step 7. Write final U to file, which represent class proportions.

26
3.2.2 POSSIBILISTICS C-MEANS (PCM) CLUSTERING

The main motivation behind the use of PCM relates to the relaxation of the probabilistic
constraints of FCM. Therefore, the formulation of PCM is based on a modified FCM objective
function whereby an additional item called is regularizing term is also include (i.e. the second
term in equation). This forces ij to be as large as possible such that pixels with high degree of
typically pixels with high degree of typicality (i.e. the membership value of pixel in a cluster
representing the possibility of the pixel belonging to that cluster) with respect to cluster may
have value of ij and the pixels with less typicality may have low ij values in all clusters. Thus
similar to FCM, PCM clustering is also an iterative process where the class membership values
are obtained by minimizing the generalized least-square error objective function (Krishnapuram
and Keller, 1993), given by,

N C c
jm (U ,V ) (1 ij )m
2
( ij )m xi v j (3.14)
A
i 1 j 1 i 1

Subject to constraints

maxj ij > 0 for all i (3.15)

i 1
ij > 0 for all j (3.16)

0 ij 1 for all i , j (3.17)

Where j is a parameter that depends on the distribution of pixels in the cluster j and is assumed
to be proportional to the mean of intra cluster distance. For clusters with similar distribution, j.

May be set of the same value for each cluster (Massone et. al, 2000)

Generally, j depends on the shape and average size of the cluster j and its values may be
computed as,

ijm dij2
K i 1
N
(3.18)
i 1
m
ij

27
Where K is a constant and is generally kept as 1. The class memberships ij, are obtained from

1
ij 1
(3.19)
d 2
( m 1)
1 ij


j

The cluster centers are updated in the similar fashion as in FCM.

The general procedure of PCM algorithm may be formalized in the following steps,

Step 1. Fix values of m and c and select the type of A-norm and maximum number of iterations.
Step 2. Choose an initial membership matrix Ui , as described in FCM.
Step 3. Estimate j using equation 3.18
Repeat
Step 4. Input the image file.
Step 5. Compute the vector of cluster centers vj.
Step 6. Compute the distance based on the selected A-norm.
Step 7. Update U matrix for the next iteration Ui+1 until ( U i 1 U i ) where is the user
defined limiting error.
Step 8. Estimate the final value on j.
Step 9. Compute element of U matrix using final value of j.
Step 10. Write final U to file which represent class proportions when scaled from 0 to 1.

It may be mentioned that the value of m and its interpretation is different in the PCM than in
FCM (Krishnapuram and Keller, 1996). The Weighting Exponent m in PCM determines the rate
of decay of the membership values. It is true that in both PCM and FCM, as m approaches 1,
partitions that minimize Jm become increasingly hard whereas when m approaches , partitions
that minimize Jm become increasingly soft or complete fuzziness. However, in FCM, as m
increase, it represents increase in sharing of pixel in all clusters. Where in PCM, increase in
value of m represents increased possibility of all pixels in the dataset completely belonging to a
given cluster.

28
CHAPTER 4

METHODOLOGY AND
DATA

29
4.1 METHODOLOGY

In this assessment area will be used of multi-spectral remote sensing data of Resource Sat-1.
At sub pixel level different classification algorithms will be investigated for extracting
information of interest. These algorithms will be used in supervised classification mode; it
means reference data will be generated from input multi spectral image. The reference data
will be used for generating statistical parameters of different algorithms. The output
generated from these algorithms in the form of land use and land cover map will be
evaluated.

Remote Sensing Multi Spectral Data


AWIFS Data\ LISS-III Data\ LISS-IV Data Resource-Sat
Data

Geo-coding of data Topo sheet/GPS Data

for Image
Creation of Training Data Registration

Classification
(Soft/Hard)
Soft Classification

SVM ANN FCM PCM

(WORKING) (WORKING)

Outcome

Figure 2: Methodology

4.2 STUDY AREA

Study area for this research work has been identified, Sitarganj Tehsil (Udham Singh Nagar
District, Uttrakhand).

The study area has the following land use and land cover classes (Table 1)

30
Table 1. Land use and Land cover classes present in the study area.

Sl. No. Land cover categories

1. Dry fellow Agricultural Land

2. Dense Forest

3. Eucalyptus Forest

4. Moist without Crop

5. Dry without Crop

6. Water Body

4.3 Data Requirement

In this assessment work soft classification approach will be applied while preparing land use and
land cover map using multi-spectral remote sensing data sets (Resource Sat data). The data sets
to be used in this will be of two types. One data set will be fine spatial resolution data, to check
classifier algorithms towards coarser spatial resolution data set and second data with coarser to
medium coarser resolution for classification.

The AWIFS versus LISS III data set have been used for the purpose of classification whereby
LISS III versus LISS IV data set has been used to generate reference data and AWIFS versus
LISS IV data is used for the purpose of similarity analysis.

The characteristic of Resource Sat-1 (P6) sensors from which data will be used in this
assessment are given in Table 3.

Table 2: Sensor specification of Resource Sat-1 (P6)

31
Sensors

AWIFS LISS III LISS IV

Specifications

IGFOV 56 m (nadir) 23.5 m 5.8 m


70m (at field edge)

Spectral Bands B2: 0.52-0.59 B2: 0.52-0.59 B2: 0.52-0.59


B3: 0.62-0.68 B3: 0.62-0.68 B3: 0.62-0.68
B4: 0.77-0.86 B4: 0.76-0.86 B4: 0.76-0.86
B5: 1.55-1.70 (SWIR) B5: 1.55-1.70
(SWIR)

Swath 370 km each head 141 km 23.9 km in MS mode


740 km (combined) 70 km in mono mode

Data Quantization 10 bit 7 bit (VNIR), 10 10 bit (selected 7 bits will


bit (SWIR) be transmitted by the
data handling system)

Table 3: Resource Sat-1 payload characteristic

Sensor Bands Resolution [m] Swath [km] Quantization


[bits]

LISS-IV Mono red 5.8 70.3 7


mode
LISS-IV MX green red NIR 5.8 70.3 7
mode
LISS-III green red NIR 23 141 10
SWIR
AWiFS green red NIR 56 (nadir) 740 10
SWIR 70 (edge)

4.4 Ancillary Data

In this assessment approach will be used is supervised classification. In supervised classification


training data is required. To generating training data, existing topographic map, literature

32
information and existing land use and land cover maps will be used as well as GPS surveyed
data.

33
CHAPTER 5

CONCLUSION AND
RESULT ANALYSIS

34
5.1 INTRODUCTION

In remote sensing data, the uncertainty is a significant issue in the classification. It is important
and necessary to evaluate the classifier performance. In this, we addressed the evaluation of
EFCM classifier while estimating uncertainty in fuzzy with varying spatial resolution of
classification and reference sub-pixel outputs. This assessment addresses the evaluation of
entropy, based on FCM and PCM classifier which estimate uncertainty in classification results.

5.2 RESULT AND DISCUSSION

Euclidean norm has been chosen to investigate the effect of uncertain pixels in FCM and PCM
classifier for both the classifier whereas the values of weighting exponent . M varies from 1.1
to 4.0 for Sal forest, Eucalyptus plantation, water bodies, agriculture land with crop, agriculture
moist land without crop, and agriculture dry land without crop.

It is observed from the result table 1, 2, 3, 4, 5, 6, 7and 8, that uncertainty ratio is almost equal to
referential value 2.585, for FCM and PCM classifier using Euclidean norm. This reflects that
fuzzy based soft classifier FCM and PCM are producing higher classification accuracy with
minimum level of uncertainty. The computation of entropy is an absolute reflector of an
uncertainty and this study identifies that entropy criterion provides stable results for FCM and
PCM, classifier for optimized value of m. For setting the optimized value of m, a number of
experiments have been conducted individually for both classifiers by varying m from 1.1 to 4.0.
It has been observed from the resultant tables 1, 2, 3, 4, 5 and 6 that for homogenous classes like
agriculture land with crop, agriculture dry land without crop, agriculture moist land without crop
and water body for FCM classifiers the optimized value 2.9 and PCM classifiers this optimized
value of m is 3.2. Similarly for heterogeneous classes like Sal forest and Eucalyptus plantation,
the optimized value of m for FCM classifier is 2.7 and for PCM classifier is 3.0. These finding
suggest that using these optimized values of m for FCM and PCM classifiers on homogeneous
land and heterogeneous land cover classes the range of the computed entropy varies between the
range of [0,3] as shown in resultant tables 1 to 6. This in turn state that the information
uncertainty is not exceeding more than 3%. Measuring the spatial statistics of a satellite image
using an entropy of six land cover classes can be measured using equation (10) i.e. 6*(-
1/6*log21/6)=2.585 [21]. This states that if the computed entropy values of classified images are
lying within this range; then indirectly this reflects better classification results. It is shown i8n
table 1, 2, 3, 4, 5 and 6 where AWIFS, LISS III and LISS IV entropy of FCM and PCM
classifiers for six land cover classes have been computed and, found that the entropy values are
approximately lying within the specified range where in the value of weighting exponent is lying
from 1.1 to 4.0.

35
Table 4. AWIFS entropy of various land cover classes from FCM classification output

Value of Agriculture Sal forest Eucalyptus Agriculture Agriculture Water Body


weighting land with crop plantation dry land moist land
exponent m without crop without crop
1.1 0 0 0.005647 0 0.005647 0
1.2 0 0 0.005647 0 0.042621 0
1.3 0 0.005647 0.011271 0 0.07731 0
1.4 0.005647 0.005647 0.011271 0.005647 0.108948 0
1.5 0.005647 0.005647 0.011271 0.005647 0.201273 0
1.6 0.005647 0.005647 0.114217 0.011271 0.374569 0
1.7 0.011271 0.005647 0.219443 0.048223 0.607945 0.005647
1.8 0.042621 0.005647 0.372583 0.085153 0.827369 0.005647
1.9 0.048223 0.005647 0.50455 0.153411 1.059939 0.005647
2.0 0.07731 0.005647 0.696694 0.198918 1.321087 0.005647
2.1 0.108948 0.005647 0.889023 0.311619 1.49718 0.011271
2.2 0.155768 0.005647 1.025828 0.421707 1.631763 0.011271
2.3 0.229191 0.005647 1.189574 0.592181 1.740952 0.011271
2.4 0.2914 0.011271 1.323116 0.745557 1.849475 0.016873
2.5 0.349979 0.048223 1.450184 0.902941 1.955882 0.016873
2.6 0.429351 0.053802 1.536682 1.020398 2.01337 0.053802
2.7 0.499022 0.082867 1.628828 1.131156 2.081574 0.142274
2.8 0.528254 0.108948 1.718866 1.24783 2.12721 0.164456
2.9 0.640882 0.133084 1.800708 1.371942 2.171966 0.30079
3.0 0.751242 0.192584 1.883666 1.498588 2.228732 0.311653
3.1 0.800653 0.243009 1.940604 1.579978 2.263866 0.34832
3.2 0.891978 0.268888 1.994867 1.648406 2.283853 0.468001
3.3 0.979401 0.314408 2.046533 1.717386 2.319002 0.519514
3.4 1.060584 0.343834 2.094014 1.786385 2.335892 0.564537
3.5 1.157506 0.448557 2.134659 1.845735 2.353021 0.65343
3.6 1.221567 0.54316 2.149877 1.903476 2.38049 0.706779
3.7 1.30389 0.587018 2.176498 1.954026 2.388224 0.807428
3.8 1.340089 0.678095 2.216589 2.002838 2.407193 0.887616
3.9 1.431974 0.713617 2.250574 2.054258 2.425069 0.993425
4.0 1.477623 0.790351 2.252167 2.094103 2.424216 1.020262

36
14
Agriculture moist land
12 without crop
Agriculture dry land
10
without crop
8 Eucalyptus plantation

6 Sal forest

4
Agriculture land with
crop
2
Value of weighting
0 exponent m
1 3 5 7 9 1113151719212325272931333537

Figure 3: Graph plotted on the basis of table 1

Table 5. AWIFS entropy of various land cover classes from PCM classification output

Value Of Agriculture Sal forest Eucalyptus Agriculture Agriculture Water Body


weighting land with crop plantation dry land moist land
exponent m without crop without crop
1.1 0.005647 0.081051 0.005647 0.005647 0.005647 0
1.2 0.167979 0.386942 0.131026 0.005647 0.011271 0
1.3 0.399329 0.589088 0.549369 0.005647 0.08136 0
1.4 0.575347 0.698132 1.014802 0.005647 0.384783 0
1.5 0.812344 0.815264 1.384835 0.042621 0.743906 0.03135
1.6 1.019893 0.991494 1.671674 0.173603 1.131458 0.180259
1.7 1.224674 1.190383 1.869674 0.33201 1.495765 0.356848
1.8 1.408777 1.356825 2.032567 0.559391 1.798901 0.623284
1.9 1.567515 1.502973 2.176595 0.78944 2.032905 0.867066
2.0 1.703025 1.642977 2.312661 1.006198 2.230467 1.08376
2.1 1.79037 1.731664 2.410349 1.176979 2.365987 1.303749
2.2 1.856408 1.810626 2.481765 1.32949 2.471655 1.499127
2.3 1.920841 1.869183 2.534982 1.477669 2.558114 1.655716
2.4 1.967894 1.932716 2.589675 1.617271 2.618009 1.805544
2.5 2.016916 1.990429 2.633355 1.759798 2.66985 1.929757
2.6 2.06177 2.031302 2.668384 1.880278 2.70955 2.0543
2.7 2.096596 2.071814 2.699778 1.978764 2.74581 2.157313
2.8 2.13173 2.099389 2.728832 2.06977 2.786231 2.2456
2.9 2.172963 2.133255 2.747897 2.148217 2.826221 2.319515
3.0 2.222914 2.168053 2.777059 2.222914 2.846706 2.39804
3.1 2.254409 2.194899 2.803922 2.303073 2.807276 2.471089
3.2 2.28588 2.21303 2.820989 2.367444 2.825149 2.533843
3.3 2.313177 2.233169 2.834812 2.415572 2.841427 2.580015
3.4 2.333332 2.253543 2.8479 2.460965 2.857763 2.621677
3.5 2.359567 2.269023 2.860937 2.507254 2.863999 2.663224
3.6 2.378925 2.290968 2.87349 2.530244 2.877346 2.706939
3.7 2.402141 2.316157 2.885775 2.566049 2.881051 2.737986
3.8 2.425806 2.334049 2.894825 2.604207 2.891998 2.769042
3.9 2.444881 2.352773 2.902729 2.636352 2.902443 2.799337

37
4.0 2.467058 2.376502 2.91509 2.66271 2.902668 2.829192

4.5

4 Value Of weighting
exponent m
3.5
Agriculture land with
3 crop

2.5 Sal forest

2
Eucalyptus plantation
1.5

1 Agriculture dry land


without crop
0.5
Agriculture moist land
0 without crop
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31

Figure 4: graph plotted on the basis of table 2

Table 6: LISS III entropy of various land cover classes from FCM classification output

Value Of Agriculture Sal Eucalyptus Agriculture dry Agriculture moist Water


weighting land with crop forest plantation land without crop land without crop Body
exponent m
1.1 0.005647 0.005647 0.005647 0 0.005647 0
1.2 0.005647 0.005647 0.005647 0 0.073972 0
1.3 0.005647 0.122038 0.005647 0 0.294907 0
1.4 0.07173 0.261892 0.005647 0.005647 0.653681 0
1.5 0.144767 0.438321 0.005647 0.005647 1.05517 0
1.6 0.244104 0.585606 0.011271 0.005647 1.35955 0.005647
1.7 0.364689 0.714277 0.085153 0.005647 1.603583 0.005647
1.8 0.472074 0.850389 0.163806 0.011271 1.783335 0.011271
1.9 0.573062 0.980095 0.242081 0.053802 1.926197 0.016873
2.0 0.688359 1.154796 0.313409 0.187964 2.024459 0.059359
2.1 0.812223 1.271646 0.468849 0.308693 2.094082 0.162102
2.2 0.919235 1.385334 0.577848 0.436872 2.149881 0.277342
2.3 1.079375 1.466096 0.737814 0.558686 2.190752 0.382014
2.4 1.21204 1.562146 0.854526 0.697262 2.231481 0.531827
2.5 1.318114 1.653353 0.971561 0.826206 2.264671 0.64054
2.6 1.40474 1.730904 1.127387 0.947186 2.295833 0.765352
2.7 1.504222 1.801376 1.2389 1.061286 2.325257 0.870188
2.8 1.591054 1.867524 1.343727 1.208763 2.348919 0.995889
2.9 1.661768 1.918553 1.442624 1.335706 2.371786 1.104937
3.0 1.73146 1.968721 1.542677 1.42612 2.397596 1.225059
3.1 1.794947 2.016358 1.617287 1.492632 2.418438 1.331407
3.2 1.85302 2.05725 1.688421 1.584244 2.438308 1.416639
3.3 1.907959 2.09954 1.756235 1.664608 2.453257 1.491422
3.4 1.960115 2.139969 1.815458 1.732818 2.471248 1.566883
3.5 2.007936 2.179333 1.872325 1.807625 2.482828 1.63579
3.6 2.053551 2.213982 1.915535 1.867013 2.479834 1.698745

38
3.7 2.095712 2.244192 1.957011 1.91358 2.486553 1.765685
3.8 2.114165 2.262875 1.996795 1.960827 2.492935 1.843503
3.9 2.122742 2.26965 2.040745 2.011581 2.498988 1.897505
4.0 2.124947 2.269793 2.061909 2.070317 2.504722 1.942477

4.5
Value Of weighting
4 exponent m

3.5 Agriculture land with


crop
3
Sal forest
2.5

2 Eucalyptus plantation

1.5
Agriculture dry land
1 without crop

0.5 Agriculture moist land


without crop
0
Water Body
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Figure 5 graph plotted on the basis of table 3

Table 7 LISS III entropy land cover classes from PCM classifications output

Value Of Agriculture Sal Eucalyptus Agriculture dry Agriculture moist Water


weighting land with crop forest plantation land without crop land without crop Body
exponent m
1.1 0.005647 0.005647 0 0.005647 0.005647 0
1.2 0.005647 0.005647 0.005647 0.005647 0.011271 0
1.3 0.005647 0.005647 0.005647 0.005647 0.016873 0
1.4 0.005647 0.005647 0.005647 0.005647 0.022452 0
1.5 0.005647 0.005647 0.005647 0.005647 0.053802 0
1.6 0.042621 0.042621 0.005647 0.011271 0.23809 0
1.7 0.097857 0.07173 0.005647 0.053802 0.415129 0.005647
1.8 0.144767 0.127572 0.005647 0.193452 0.653195 0.005647
1.9 0.192398 0.192398 0.005647 0.330244 0.864063 0.005647
2.0 0.286224 0.280851 0.005647 0.530687 1.068206 0.005647
2.1 0.382887 0.371918 0.011271 0.726706 1.260592 0.011271
2.2 0.469436 0.450103 0.016873 0.902965 1.412865 0.022452
2.3 0.540047 0.517542 0.022452 1.068946 1.54625 0.033543
2.4 0.637522 0.580463 0.096244 1.218001 1.684216 0.11271
2.5 0.790682 0.639519 0.191075 1.360741 1.798627 0.238685
2.6 0.874402 0.731231 0.248428 1.48356 1.901779 0.368867
2.7 0.9576 0.811808 0.325234 1.581556 1.984885 0.449749
2.8 1.063631 0.949098 0.375461 1.668634 2.049721 0.548482
2.9 1.140056 1.099411 0.475092 1.740039 2.100816 0.642982
3.0 1.234354 1.211541 0.627877 1.825913 2.147537 0.735558
3.1 1.318565 1.315615 0.725852 1.838011 2.191935 0.840068
3.2 1.389581 1.413052 0.815275 1.898131 2.227696 0.967095

39
3.3 1.453251 1.500534 0.914436 1.955376 2.265943 1.077958
3.4 1.514214 1.583281 0.986734 2.003771 2.302689 1.178255
3.5 1.588201 1.653662 1.076588 2.058301 2.32652 1.273366
3.6 1.655488 1.717781 1.161803 2.103118 2.346336 1.353272
3.7 1.722504 1.766239 1.243087 2.15266 2.356008 1.430058
3.8 1.783701 1.800336 1.307944 2.193785 2.383625 1.487738
3.9 1.842241 1.814725 1.380038 2.230366 2.402686 1.555798
4.0 1.857169 1.824889 1.423223 2.259975 2.402588 1.589166

25
Water Body

20
Agriculture moist land
without crop
15 Agriculture dry land
without crop

10 Eucalyptus plantation

Sal forest
5

Agriculture land with


0 crop
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31

Figure 6 graph plotted on the basis of table 4

Table 8 LISS IV entropy of various land cover classes from FCM classification output

Value Of Agriculture Sal Eucalyptus Agriculture dry Agriculture moist Water


weighting land with crop forest plantation land without crop land without crop Body
exponent m
1.1 0.005647 0.005647 0.515133 0.005647 0.005647 0
1.2 0.159305 0.148035 0.710069 0.005647 0.086675 0
1.3 0.458672 0.430657 1.318766 0.036997 0.224373 0.03135
1.4 0.738771 0.688545 1.577286 0.066128 0.484666 0.054858
1.5 1.009651 0.954215 1.693985 0.221962 0.858429 0.164277
1.6 1.264439 1.169679 1.831426 0.408527 1.227814 0.432578
1.7 1.44293 1.368441 1.953749 0.629618 1.528898 0.700922
1.8 1.581536 1.536031 2.042543 0.889166 1.792705 0.973154
1.9 1.711045 1.656994 2.120191 1.116041 2.024335 1.210289
2.0 1.819654 1.759571 2.178887 1.321267 2.207148 1.4135
2.1 1.896176 1.856853 2.228556 1.500895 2.339992 1.618863
2.2 1.95878 1.925165 2.279108 1.64986 2.430146 1.798492
2.3 2.015419 1.995637 2.3135 1.784406 2.507934 1.959073
2.4 2.062725 2.062419 2.341868 1.903918 2.580296 2.08944
2.5 2.111727 2.120333 2.364183 2.006134 2.639264 2.215204
2.6 2.167633 2.174009 2.386591 2.095079 2.688398 2.324193
2.7 2.228392 2.220678 2.41137 2.177226 2.732176 2.423656
2.8 2.268899 2.266992 2.441783 2.26451 2.766897 2.50298

40
2.9 2.31079 2.316862 2.459165 2.343095 2.780304 2.576516
3.0 2.34609 2.355302 2.486786 2.407492 2.813747 2.638276
3.1 2.384577 2.392002 2.499265 2.470531 2.840047 2.691409
3.2 2.415658 2.417195 2.511218 2.529309 2.864499 2.737175
3.3 2.445378 2.441422 2.52796 2.577478 2.887159 2.777622
3.4 2.467903 2.465294 2.546082 2.622474 2.908079 2.811794
3.5 2.495519 2.487401 2.568461 2.656918 2.923335 2.843927
3.6 2.51675 2.509808 2.58812 2.682804 2.936682 2.874073
3.7 2.538504 2.53187 2.606685 2.710862 2.948797 2.902011
3.8 2.559757 2.55343 2.626424 2.734912 2.956965 2.930187
3.9 2.580511 2.57636 2.644617 2.764053 2.964735 2.950723
4.0 2.60552 2.593914 2.658435 2.784166 2.975491 2.973742

4.5

3.5

2.5

1.5

0.5

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29

Figure 7 graph plotted on the basis of table 5

Table 9 LISS IV entropy of various land cover classes from PCM classification output

Value Of Agriculture Sal Eucalyptus Agriculture dry Agriculture moist Water


weighting land with crop forest plantation land without crop land without crop Body
exponent m
1.1 0.171728 0.105319 0.171266 0.005647 0.109033 0
1.2 0.280358 0.27128 1.448655 0.005647 0.157189 0
1.3 0.541512 0.539848 1.301282 0.036997 0.391546 0.054858
1.4 0.868812 0.77269 1.438427 0.179385 1.020216 0.188036
1.5 1.176479 1.126859 1.64499 0.397338 1.560379 0.48939
1.6 1.43719 1.402599 1.838701 0.62751 1.907367 0.816543
1.7 1.636362 1.634862 1.981752 0.836575 2.183758 1.094187
1.8 1.794736 2.121273 2.113864 1.031655 2.376412 1.329639
1.9 1.913594 2.252041 2.201014 1.203467 2.503874 1.555592
2.0 2.008797 2.094087 2.273105 1.355153 2.61074 1.748845

41
2.1 2.097794 2.196847 2.339495 1.489767 2.691004 1.921628
2.2 2.177642 2.296202 2.390455 1.630617 2.759321 2.078709
2.3 2.235052 2.382956 2.437824 1.739274 2.818155 2.223974
2.4 2.285679 2.453322 2.480808 1.852672 2.853843 2.343828
2.5 2.342261 2.513818 2.52711 1.958228 2.890597 2.441456
2.6 2.395457 2.570749 2.566236 2.046349 2.923324 2.522726
2.7 2.444017 2.620091 2.601243 2.126834 2.947076 2.600851
2.8 2.486534 2.661133 2.630872 2.206592 2.970705 2.661939
2.9 2.523838 2.695787 2.661505 2.280857 2.990447 2.728514
3.0 2.558951 2.728568 2.684531 2.35048 3.00583 2.774441
3.1 2.592976 2.759445 2.707278 2.414746 3.019891 2.812137
3.2 2.617675 2.785684 2.733101 2.466663 3.032396 2.844095
3.3 2.644031 2.806858 2.757805 2.513756 3.043829 2.874048
3.4 2.673763 2.822548 2.781144 2.558382 3.056219 2.902142
3.5 2.702388 2.84003 2.792381 2.596201 3.068167 2.9285
3.6 2.727067 2.856929 2.808265 2.626507 3.073613 2.953157
3.7 2.746486 2.870459 2.818729 2.655341 3.079396 2.976148
3.8 2.764919 2.880206 2.828862 2.681666 3.084805 2.993132
3.9 2.782202 2.893205 2.842368 2.711857 3.088626 3.009168
4.0 2.799091 2.899399 2.855209 2.738527 3.093259 3.024005

25
Water Body

20
Agriculture moist land
without crop
15 Agriculture dry land
without crop

10 Eucalyptus plantation

Sal forest
5

Agriculture land with


0 crop
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31

Figure 8 graph plotted on the basis of table 6

These findings suggest that using these optimized values of m for FCM and PCM classifiers on
homogenous and heterogeneous land cover classes the range of the computed entropy varies
between the range of [0,3] as shown in resultant table 1 to 6.

42
5.3 CONCLUSION

In this assessment it has been tried to generate fraction outputs from FCM and PCM classifier
using Euclidean norm. These outputs have been generated from AWIFS, LISS III and LISS IV
images of IRS P6 data. Entropy has been used as assessment parameters of accuracy for various
land cover classes i.e. water bodies, Sal forest, Eucalyptus plantation, agriculture land with crop,
agriculture moist land without crop, agriculture dry land without crop. Uncertainty is intrinsic in
spatial data and this generally refer to error, inexactness, fuzziness and ambiguity. The objective
of this assessment on spatial data is to investigate, how uncertainties arise, or are created and
propagated in the spatial data.

43
APPENDIX

44
APPENDIX
Snapshot of reading the image

Snapshot of reading a satellite dataset

import java.io.*;

import javax.swing.*;

import java.awt.event.*;

import javax.imageio.*;

import java.awt.*;

import java.awt.image.*;

class MyFrame extends JFrame implements ActionListener

JMenuBar mb;

45
JMenu file;

JMenu edit;

JMenuItem reset, result;

JMenuItem fopen;

JLabel lab;

BufferedImage img;

static int pixVal[], clickCount;

public MyFrame()

pixVal = new int[16];

clickCount = 0;

img = null;

mb = new JMenuBar();

file = new JMenu("File");

fopen = new JMenuItem("Open");

edit = new JMenu("Edit");

reset = new JMenuItem("Reset");

result = new JMenuItem("Show Results");

file.add(fopen);

mb.add(file);

edit.add(reset);

edit.add(result);

mb.add(edit);

fopen.addActionListener(this);

reset.addActionListener(this);

46
result.addActionListener(this);

setJMenuBar(mb);

lab = new JLabel();

add(lab);

setSize(400,300);

setVisible(true);

setDefaultCloseOperation(JFrame.EXIT_ON_CLOSE);

addMouseListener(new MyMouse());

public void actionPerformed(ActionEvent ae)

if(ae.getSource() == fopen)

JFileChooser f = new JFileChooser();

f.showOpenDialog(null);

File fname = f.getSelectedFile();

String imgPath = fname.getPath();

System.out.println(imgPath);

lab.setIcon(new ImageIcon(imgPath));

try

img = ImageIO.read(new File(imgPath));

catch(Exception e)

47
System.out.println(e);

repaint();

System.out.println("hi");

else if(ae.getSource() == reset)

clickCount = 0;

else if(ae.getSource() == result)

int i, sum = 0;

if(clickCount < 16)

JOptionPane.showMessageDialog(null, "Please Click Some More


Samples");

return;

for(i=0; i<16; i++)

sum += pixVal[i];

int range = (int)(((double)sum / 16.0) % 255);

//JOptionPane.showMessageDialog(null, "Result : " + range);

if(range >=0 && range <= 50)

JOptionPane.showMessageDialog(null, "Dry fellow Agricultural Land");

if(range >50 && range <= 100)

48
JOptionPane.showMessageDialog(null, "Dense Forest");

if(range >100 && range <= 150)

JOptionPane.showMessageDialog(null, "Eucalyptus Forest");

if(range >150 && range <= 200)

JOptionPane.showMessageDialog(null, "Moist without Crop");

if(range >200)

JOptionPane.showMessageDialog(null, "Dry without Crop");

if(range >200)

JOptionPane.showMessageDialog(null, "Dry without Crop");

class MyMouse extends MouseAdapter

int col, r, g, b, pix;

public void mouseClicked(MouseEvent me)

System.out.println("X: " + me.getX() + ", Y: " + me.getY());

System.out.println("Color: " + img.getRGB(me.getX(), me.getY()));

col = img.getRGB(me.getX(), me.getY());

r = (col & 0x00ff0000) >> 16;

49
g = (col & 0x0000ff00) >> 8;

b = col & 0x000000ff;

pix = col & 0x00ffffff;

System.out.println("Color Codes at (" + me.getX() + ", " + me.getY() + ") is : ("


+ r + ", " + g + ", " + b +")");

System.out.println("Pixel Value : " + pix);

if(clickCount < 16)

pixVal[clickCount++] = pix;

else

JOptionPane.showMessageDialog(null, "No More Clicks Will Be


Accepted. Please Select Reset");

public static void main(String ar[])

new MyFrame();

50
REFERENCES

51
REFERENCE
ABE, S and THAWONMAS R, 1997, Fuzzy classifier with classifier with ellipsoidal regions.
IEEE Transaction on Fuzzy System, 5, 358-367.

ADAMS, J.B, SABOL, D.E, KAPOS, V.FILHO, R.A.ROBERTS, D.A.SMITH, M.O and
GILLESPIE, A.R, 1995, Classification of multi spectral images based on fractions of end
members: application to land cover changes in the Brazilian Amazon, Remote Sensing of
Environment.52, 137-145.

ARAI , K., 1993, A classification method with spatial spectral variability. International Journal
of Remote Sensing, 14, 699-709.

ARORA, M.K., 1996, An evaluation of factors affecting the accuracy of digital multi spectral
land cover classification, unpublished Ph.D. Thesis University of Wales Swansea, Uk.

ARORA, M.K, 2002, Land cover classification from remote sensing data. GIS@development,6,
24-25 and 30-31.

BARALDI, A, BINAGHI, E, BLONDA, P, BRIVIO, P.A, and RAMPINI,A, 2001. Comparison


of the multilayer perceptron with neuro-fuzzy techniques in estimation of cover class mixture in
remotely sensed data, IEEE Transaction on Geoscience and Remote Sensing 39, 994-1005.

BASTIN, L., 1997, Comparison of fuzzy c-means classification, linear mixture modeling and
MLC probabilities as tools for unmixing coarse pixels, International Journal of Remote
Sensing,18, 3629-3648.

BEZDEK.J.C., 1981, Pattern Recognition with Fuzzy Objective Function Algorithms, Plenum,
New York,USA.

BEZDEK, J.C.,EHRLICH, R, and FULL W., 1984, FCM: The fuzzy c-means clustering
algorithm, Computers and Geosciences. 10, 191-203.

BINAGHI, E, MADELLA , P. MONTESSANO, M.G., and RAMPINI, A, 1997, Fuzzy


contextual classification of multisource remote sensing images, IEEE Transaction on Geoscience
and Remote sensing, 35, 326-340.

BRUZZONE, L and PRIETO, D.F, 2001, Unsupervised retraining of a maximum likelihood


classifier for the analysis of multi temporal remote sensing images, IEEE Transaction on
geosciences and remote sensing, 39,456-460

BRUZZONE, L., COSSU, R and VERNAZZA, G.,2002, Combining parametric and non-
parametric algorithm for a partially unsupervised classification of multitemporal remote sensing
images, Information Fusion, 3,289-297.

52
BUTTNER G, HAJOS and KORANDI, M, 1989, Improvements to the effectiveness of
supervised training procedures, International Journal of remote sensing.

CHEN D, and STOW D, 2003, Strategies for integrating information from multiple spectral
resolution into land use/land cover classification routines, Photogrammetric Engineering and
remote sensing, 69, 1279-1287

DUNN, 1.C, 1974, A fuzzy relative of the ISODATA process and its use in detecting compact
well-separated cluster, Journal of Cybernet, 3, 22-57.
EASTMAN, J. R, and LANEY, R.M, 2002, Bayesian soft classification for sub pixel analysis a
critical evaluation, Photogrammetric Engineering and Remote Sensing , 68, 1149-1154
EMRAHOGLU, N., YEGINGIL, I., PESTEMALCI, SENKAI.,O., and KANDIRMAZ, H, M.,
2003Comparison of a new algorithm with supervised classification, International journal of
remote sensing, 24, 649-655

53

Anda mungkin juga menyukai