republic of iraqenahrainuniv.edu.iq/sites/default/files/saif kamil.pdf · 2018-01-25 · known...
TRANSCRIPT
Republic of Iraq
Ministry of Higher Education & Scientific Research
Al-Nahrain University
College of Science
Effect of Band Ratios and Indices on Classification Accuracy of Multispectral Satellite
Images
A Thesis
Submitted to the College of Science of Al-Nahrain University as a Partial Fulfillment of the Requirements for the Degree of Master of
Science in Physics
By
Saif Kamil Shnain
(B.Sc. 2005)
Supervised by
Professor Dr. Ayad A. Al-Ani Assistant Professor Dr. Salah A. Saleh
March 2008 Safar 1429
�‹È �u �Ü÷] �‡ �∑ �Ü÷] ��!] �‹ �ä�e ‡�⁄ �|Ê �Ü÷] �ÿ�Œ �|Ê �Ü÷] �‡¬ ‘fiÁ÷ˇ�äËÊ � È�◊Œ ��c �‹�◊�√�÷] ‡�⁄ �‹�jÈ�iÊ�_ ^⁄Ê Í�eÖ �Ü⁄_
‹Èø√÷] !] —Çë
DTQE]Üâ˜]
Acknowledgments
Praise to our almighty GOD to the most merciful and gracious
for enabling me to complete this thesis.
I would like to express my sincere thanks and deep gratitude
to my supervisors Professor Dr. Ayad A. Al-Ani and Assistant
Professor Dr. Salah A. Saleh for supervising the present work and
for their support, encouragement and suggestions throughout the
work.
Special thanks and deep affection to my family especially my
parents for their love, help, and supports through the years.
Finally, I would like to extend my thanks and gratitude to all
who had ever assisted me during the period of this research.
ft|y ^tÅ|Ä
Abstract Multispectral satellite images are holding whole information about the area
under investigation. Many of digital processing techniques are used to extract
most of the possible information from these images, such these techniques are
spectral band ratios, Tasseled Cap transformation and Principle Component
Analysis (PCA). These techniques are play good part for features extraction and
reducing the number of spectral bands with no loses in information.
The research aims to apply classifications on band ratios and image indices
which include PCA and Tasseled Cap transformations, of Enhanced Thematic
Mapper Plus (ETM+) satellite images, and to show how the accuracies of the
classification are differ from the accuracy of classification of the raw ETM+ bands
images.
The results show that the best bands ratios that adopted to represent five
selected land features are 5/7, 4/5, 2/5, 1/4, and 2/4. Optimum Index Factor
method (OIF) has been adopted on raw bands combinations, and on the bands
ratio combinations to select which band combination that contains much
information, where the results show that the best combination for raw bands is
3-4-7, and best combination for ratios band is 4/5-1/4-2/5.
Iterative Self Organizing Data Analysis (ISODATA) method has been
adopted as unsupervised classification, and Maximum Likelihood method as
supervised classification. The overall accuracies are 87.18%, 89.45%, 89.26%,
and 89.52% for raw bands, band ratios, Tasseled Cap, and PCA respectively,
which showed that transformation techniques give good spectral enhancements
and feature extraction which are undistinguishable in raw image, and represent
good tools to increase the accuracy of classification.
i
Contents
Title Page
Contents………….…………….…………….……………………….……..i
List of figures………………………………………………………………iv
List of tables…….….…………………….………………………………...vi
Abbreviations…………………………………………………….....…….vii
Chapter One: Introduction
1.1 Introduction……………………………………………………….1
1.2 Previous Researches …………………………...…………………2
1.3 The Study Area…………………………………………………...4
1.4 Aim of the Thesis…………………………………………………6
1.5 Thesis Organization………………………………………………6
Chapter Two: Basic Principle of Remote Sensing
2.1 Introduction……………………………………………………….7
2.2 Electromagnetic Energy…………………………………………..7
2.3 Spectral Reflectance……………………………………………..10
2.4 Sensor Resolution……………………………………………….12
2.5 Remote Sensing Sensors………………………………………...14
2.6 Landsat Systems ………………………………………………...16
2.6.1 Landsat 1, 2, and 3……………………………………...17
2.6.2 Landsat 4 and 5…………………………………………17
2.6.3 Landsat 6………………………………………………..18
2.6.4 Landsat 7………………………………………………..18
ii
2.6.5 Characteristics of ETM+ Spectral Bands………………22
Chapter Three: Bands Transformation Techniques & Classification
3.1 Introduction……………………………………………………...24
3.2 Multispectral Remote Sensing Images………………………......24
3.3 Band Ratios……………………………………………………...25
3.4 Vegetation Indices……………………………………………….29
3.5 Spectral Indices............................................................................30
3.6 Separating Soil Reflectance from Vegetation Reflectance……...31
3.7 Tasseled Cap Transformation…………………………………...33
3.8 Principle Component Analysis (PCA)…………………………..36
3.8.1 Mathematical Representation…………………………..38
3.9 False Color Composite (FCC)…………………………………...40
3.9.1 Selecting the Best Color Composites…………………..40
3.10 Multispectral Classification……………………………………42
3.10.1 Unsupervised Classification…………………………..42
3.10.2 Supervised Classification…..………………………….44
Chapter Four: Results & Discussion
4.1 Introduction……………………………………………………...48
4.2 Band Ratios……………………………………………………...51
4.3 Tasseled Cap Transformation…………………………………...58
4.4 Vegetation indices……………………………………………….60
4.5 Principle Component Analysis…………………………………..63
4.6 False Color Composites (FCC)………………………………….68
4.7 Images Classification……………………………………………73
4.7.1 Unsupervised Classification……………………………73
iii
4.7.2 Supervised Classification……………………………….76
4.7.3 Classification Accuracy Assessment…………………...80
Chapter Five: Conclusions & Recommendations
5.1 Conclusions……………………………………………………...83
5.2 Recommendations……………………………………………….84
References…………………………………………………………………85
Appendix…………………………………………………………………..90
iv
List of figures
Figure Title Page
(1-1) Location of study area 5
(2-1) Spectral reflectance curve 11
(2-2) Examples of remote sensing systems with spectral resolution 13
(3-1) Reduction of scene illumination effects through spectral ratioing 27
(3-2) Soil brightness line 32
(3-3) Seasonal variation of a field is defined by Greenness and 34
Brightness
(3-4) 2D scatter plot of two Tasseled Cap components 35
(3-5) PCA reduce the dimensionality 37
(4-1) 6 non thermal ETM+ band images with their histograms 48
(4-2) The scheme of the work 50
(4-3) 5/7 band ratio for water 53
(4-4) 4/5 band ratio for vegetations 54
(4-5) 2/5 band ratio for bare lands 55
(4-6) 1/4 band ratio for urban areas 56
(4-7) 2/4 band ratio for crop lands 56
(4-8) The first three Tasseled Cap transformation indices 59
(4-9) NDVI for the study area 60
(4-10) Spectral indices images 62
(4-11) PCA images with their histograms 65
(4-12) FCC images for different combinations 72
v
(4-13) Unsupervised classification of different images 75
(4-14) Supervised classification 78
vi
List of tables
Table Title Page
(2-1) Electromagnetic spectral regions 9
(2-2) Sensor systems 15
(2-3) Characteristics of Landsat 1 to 7 missions 20
(2-4) Selected current remote sensing systems and their major 21
characteristics
(4-1) The best band ratios for certain features 52
(4-2) Multivariate statistics of the band ratios 57
(4-3) Correlation between band ratios 57
(4-4) Multivariate statistics of the original bands 67
(4-5) Covariance-variance matrix 67
(4-6) Correlation between bands 68
(4-7) Ranked OIF values of raw band combinations 70
(4-8) Ranked OIF values of ratio bands combinations 71
(4-9) results of supervised classification of different band combinations 77
(4-10) Confusion Matrices 81
vii
Abbreviations
BSI: Bare Soil Index
DN: Digital Number
E: Electric field
EMR: Electromagnetic Radiation
ENVI: Environment for Visual Images
ERTS: Earth Resources Technology Satellite
ETM+: Enhanced Thematic Mapper plus
FCC: False Color Composite
GIS: Geographic Information System
H: Magnetic field
IR: Infra Red
ISODATA: Iterative Self Organizing Data Analysis
Landsat: Land Satellite
MIR: Mid Infra Red
MSS: Multispectral Scanner
NASA: National Aeronautics and Space Administration
NDVI: Normalized Difference Vegetation Index
NDWI: Normalized Difference Water Index
NIR: Near Infra Red
OIF: Optimum Index Factor
PCA: Principle Component Analysis
R: Red
viii
RBV: Return Beam Vidicon
RGB: Red-Green-Blue (color image)
SAVI: Soil Adjust Vegetation Index
SWIR: Short Wave Infra Red
TC: Tasseled Cap
TM: Thematic Mapper
UI: Urban Index
UV: Ultra Violet
CHAPTER ONE
INTRODUCTION
1
1.1 Introduction
Remote sensing is defined as the acquisition of information about an
object without being in physical contact with it. Information is acquired by
detecting and measuring changes that the object imposes on the surrounding
field [1]. The advantage of gathering data about objects remotely are that the
object is usually not disturbed objects in inaccessible areas can be examined,
and a large amount of information over any spatial area can be collected [2].
The detection and recording instruments for this special technology are
known collectively as remote sensors and include photographic cameras,
mechanical scanners, and radar systems. Remote sensors are typically carried
on aircraft and earth-orbiting spacecraft, which have led to the familiar phrase
“eye in the sky” [3], so that a remotely sensed image is a spatial representation
of surface of the earth.
The development of image sensor technology has made it possible to
capture image data in multi bands covering a broad spectrum of wavelength
range [4], i.e. the same scene was imaged simultaneously in several bands,
although satellite and airborne multispectral sensors provide data in the form of
several images of the same area of the earth’s surface, but taken through
different spectral windows or bands. The number of spectral bands varies, but
typically ranges from 4, as with Landsat 4 to 350 bands or more for many
satellite sensors [5].
The rich information available in multispectral imagery has posed
significant opportunities and challenges for feature extraction and classification.
So an alternative way is to use simple features that are physically meaningful,
one such feature that has received much attention in the remote sensing
community is the band ratios, the ratio of spectral values between two different
2
bands. Band ratios have been used for many years in the remote sensing
community to identify terrain cover types [4].
Many algorithms have been proposed for classification purpose, such as
Principle Component Analysis (PCA), which is not just a good tool for data
compression purposes but also as a good tool for feature extraction which, in
turn, play an important role in image classification and recognition techniques
[2]. The other algorithms are spectral indices (NDVI, BSI, NDWI, and UI) and
Tasseled Cap transformation. The Tasseled Cap transformation is a useful tool
for compressing spectral data into a few bands associated with physical scene
characteristics [6].
1.2 Previous Researches
Band ratios and indices are important data transformation techniques, used
in remote sensing work with multispectral data, which are represent the good
tool for classification, image enhancement and data extraction.
In 1989, Chaves [2] applied selective PCA for extracting spectral contrast
in Landsat TM image data. The result shows that selective PCA can be used to
enhance and map the spectral differences or contrast between different spectral
regions.
In 1996, Al-Ani [7] applied principle component analysis on six TM
satellite images for enhancement and color image composition. The researcher
applied maximum likelihood method to classify the TM satellite images. The
results show that the over all classification accuracy is 83%.
3
In 1999, Frankovich [8] used ISODATA unsupervised classification
method on PCA and Tasseled Cap transformation of the six TM satellite images
when he merged them layers into one image). The overall classification
accuracy is 88%.
In 2002, Al-Sepahe [2] used supervised and unsupervised classification to
classify selective original TM image bands and principle component images,
the researcher showed how the selective images investigate high classification
accuracy with minimum selected number of images, and suggested a new
method to select the type and the number of bands which verify high
classification accuracy based on maximum likelihood.
Abdeen et al [9] used ASTER band ratio images in geological mapping,
and they compared the results with established Landsat ETM+ band
combinations and ratio images. They used OIF method to select the best R-G-B
ratio combination which has the most spectral information.
In 2003, Inzana et al [10] used supervised classification of Landsat TM
band ratio images, and Landsat TM band ratio image with radar images. The
overall accuracy of classification of the radar images is 91.23% which is better
than the accuracies of the traditional TM and radar fused image products, which
their accuracies are 89.34% and 89.03% respectively.
In 2005, Feely et al [11] applied spectral indices (NDVI, IRI, and MIRI)
on Landsat ETM+ bands to examine the utility of these indices for measuring
various forest attributes in the semi-deciduous tropical dry forests on study area.
Fu et al [4] used band ratio on hyperspectral images, and they showed the
potential of spectral band ratio featured for the accurate pixel classification
when they adopted the boosting frame work for the selection of multiple ratio
features.
4
Thapa et al [12] used SAVI, BSI, NDWI, and UI indexes, were computed
in multi-temporal Landsat images; to analyze the environmental changes in
respect of vegetation, agriculture, water, and urban activities of Vietnam.
In 2006 Fadhil [13] applied NDVI, BSI, NDWI, and Tasseled Cap indices
on multi-temporal Landsat images; his study demonstrates the effectiveness of
the remote sensing and GIS technologies in detecting, assessing, mapping, and
monitoring the environmental changes.
In 2007, Buhe et al [14] indicated that R-G-B color overly using
atmospheric corrected ASTER original bands 2,3 and 6 which has the highest
OIF, but when they considered NDVI as one ASTER band, the highest OIF was
carrying out bands 3,4 and NDVI which was the best three-band combinations
for supervised and unsupervised classification.
1.3 The Study Area
The study area is Al-Kut city and its neighboring areas. The centre of
Al-Kut city is located at latitude 30º 30′ N and longitude 45 º 49′ E, it is the city
in the eastern Iraq , on Tigris river on the junction between the river and the
distribution canal Shatt Al-Gharraf, about 100 miles south east of Baghdad.
Close to the city lies the Kut Barrage, this distributes river water into irrigation
canals, figure (1-1).
The old town of Al-Kut is within a sharp “U” bend of the river, almost
making it an island but for a narrow connection to the shore. As of 2003 the
estimated population is about 400.000 people, the economy of Al-Kut is based
upon the rich agriculture of the region, of which Al-Kut is the main trade and
administrative centre [15].
5
The study was performed using Landsat ETM+ images (six no thermal
bands) of study area (path 167, row 38) acquired in 15th September 2002.
Figure (1-1): Location of study area
6
1.4 Aim of the Thesis
The aims of this thesis are to apply classification on original bands, band
ratios, and image indices (PCA and Tasseled Cap transformation) of Enhanced
Thematic Mapper plus (ETM+) satellite images, and to show how the
accuracies of this classification are differ from the accuracy of the classification
of the original ETM+ bands. Supervised and unsupervised methods will adopt
for the classification purpose.
The problem has been come from selecting best band combination for
classification process, so that the Optimum Index Factor method (OIF) is used,
such that high OIF values indicate bands that contain much information.
Some spectral indices NDVI, BSI, NDWI, and UI will be used which will
derive from ETM+ bands for monitoring the land cover in the study area.
1.5 Thesis Organization
In addition to this chapter the thesis consists of other four chapters
outlined as follows:
Chapter Two describes the theoretical concepts and fundamentals of remote
sensing techniques.
Chapter Three present band transformation techniques and describes the
classification methods, supervised and unsupervised classification methods
have been adopted.
Chapter Four illustrates the results and their discussion.
Chapter Five includes the conclusions and recommendations for the future
studies.
CHAPTER TWO
BASIC PRINCIPLE OF REMOTE SENSING
7
2.1 Introduction
Remote sensing became possible with the invention of the camera in the
nineteenth century. Astronomy was one of the first fields of the science to
exploit this technique, and to this day much of astronomy is inextricably linked
with remote sensing [16].
The advent of the space age opened a whole new dimension in our ability
to observe, study, and monitor planetary (include Earth) surfaces and
atmospheres on a global and continuous scale. This led to major developments
in the field of remote sensing, both in its scientific and technical aspects [1].
2.2 Electromagnetic Energy
The term remote sensing is restricted to methods that employ
electromagnetic energy as the means of detecting and measuring target
characteristics. Electromagnetic energy includes light, heat, and radio waves
[17].
Energy recorded by remote sensing systems undergoes fundamental
interactions that should be understood to properly interpret the remotely sensed
data. For example, if the energy being remotely sensed comes from the sun, the
energy [18]:
• Is radiated by atomic particles at the source (the Sun),
• Propagates through the vacuum of space at the speed of light,
• Interacts with the Earth’s atmosphere,
• Interacts with the Earth’s surface,
8
• Interacts with the Earth’s atmosphere once again, and
• Finally reaches the remote sensor, where it interacts with various optical
systems, filters, film emulsions, or detectors.
Electromagnetic Radiation (EMR) consists of an electrical field (E) that
varies in magnitude in a direction perpendicular to the direction of propagation.
In addition, a magnetic field (H), oriented at right angles to the electrical field,
is propagated in phase with electrical field [19].
The electromagnetic spectrum is the continuum of energy the speed of
light. Several regions of the EM spectrum are of particular interest for remote
sensing visible, infrared, and microwave regions. Table (2-1) illustrates a
simple description to each part in electromagnetic spectrum [17].
9
Region Wavelength Remarks
Gamma ray < 0.03 nm Incoming radiation is completely absorbed by the upper atmosphere
and it’s not available for remote sensing.
X-ray (0.03-3.0) nm Completely absorbed by atmosphere. Not employed in remote sensing.
Ultraviolet (0.03-0.4) μm Transmitted through atmosphere. Detectable with film and
photodetectors, but atmospheric scattering is severe.
Photographic
UV band (0.3-0.4) μm
Transmitted through atmosphere. Detectable with film and
photodetectors, but atmospheric scattering is severe.
Visible (0.4-0.7) μm Imaged with film and photodetectors. Includes reflected energy peak of
earth at 0.5 μm.
Infrared (0.7-100) μm Interaction with matter varies with wavelength. Atmospheric
transmission windows are separated by absorption bands.
Reflected IR
band (0.7-3.0) μm
Reflected solar radiation that contains no information about thermal
properties of materials. The band from 0.7 to 0.9 μmis detectable with
film and is called the photographic IR band.
Thermal IR
band
(3-5) μm
(8-14) μm
Principle atmospheric windows in the thermal region. Images at these
wavelengths are acquired by optical-mechanical scanners and special
vidicon systems but not film.
Microwave (0.1-30) cm Longer wavelengths can penetrate clouds, fog and rain. Images may be
acquired in the active or passive mode.
Radar (0.1-30) cm Active of microwave remote sensing. Radar images are acquired at
various wavelength bands.
Radio >30 cm Longest wavelength portion of electromagnetic spectrum. Some
classified radars with very long wavelength operate in this region.
Table (2-1): Electromagnetic spectral regions [17]
10
2.3 Spectral Reflectance
The reflectance characteristics of earth surface features may be quantified
by measuring the portion of incident energy that is reflected. This is measured
as a function of wavelength and is called spectral reflectance ρ λ where is
expressed as percentage. It is mathematically defined as [20]:
(2-1)
A graph of the spectral reflectance of an object as a function of
wavelength is termed a spectral reflectance curve. The configuration of spectral
reflectance curves gives us insight into the spectral characteristics of an object
and has a strong influence on the choice of wavelength region(s) in which
remote sensing data are acquired for a particular application [20]. In principle,
material can be identified from its spectral reflectance signature if the sensing
system has sufficient spectral resolution to distinguish its spectrum from those
of other materials. This premise provides the basis for multispectral remote
sensing [21]. Figure (2-1) shows typical spectral reflectance curves for different
characteristic of earth surface materials [20].
11
The reflectance of clear water is generally low. However, the reflectance is
maximum at the blue end of the spectrum and decreases as wavelength
increases. The reflectance of bare soil generally depends on its composition,
where increases monotonically with increasing wavelength. Whereas vegetation
has a unique spectral signature that enables it to be distinguish readily from
other types of land cover. The reflectance is low in both the blue and red
regions of the spectrum and high in the near infrared [16].
Figure (2-1): Spectral reflectance curve [20]
12
2.4 Sensor Resolution
Each remote sensing system has four major resolutions associated with it.
These resolutions should be understood by the scientist in order to extract
meaningful information from the remotely sensed imagery. Resolution (or
resolving power) is defined as a measure of the ability of an optical system to
distinguish between signals that are spatially near or spectrally similar. The
types of resolution are as follows [18]:
1. Spectral Resolution
This refers to the number and dimension of specific wavelength intervals
in the electromagnetic spectrum to which a remote sensing instrument is
sensitive [18]. Most remote sensing systems are multispectral and obtain data at
a number of wave bands.
Figure (2-2) shows examples of systems that obtain data in one, three, and
six bands. Each layer of data is referred to as a band, for examples Landsat
Multispectral Scanner (MSS) obtained data in four bands, and each pixel
consequently has four digital numbers associated with it, whereas the Landsat
Thematic Mapper (TM) system is a seven band system. Thus Landsat (TM) has
a better spectral resolution than Landsat (MSS) [16].
13
2. Spatial Resolution
There is a relationship between the size of a feature to be identified and
the spatial resolution of the remote sensing system. Spatial resolution is a
measure of the smallest linear separation between two objects that can be
resolved by the sensor [18].
3. Temporal Resolution
Temporal resolution of a remote sensing system refers to how often it
records imagery of a particular area [18]. It is generally surveys which tend to
take place at infrequent intervals, usually for specific projects [16].
Figure (2-2): Examples of remote sensing systems with spectral resolution [16]
14
4. Radiometric Resolution
Radiometric resolution of a remote sensing system is a measure of how
many grey levels are measured between pure black (which could represent no
reflectance from the surface) and pure white. The radiometric resolution is
measured in ‘bits’ [16].
2.5 Remote Sensing Sensors
In remote sensing, the sensor measured energy, whereby the
distinguishing between passive and active techniques must be explained.
Passive remote sensing techniques employ natural sources of energy, such as
the Sun. Passive sensor systems based on reflection of the Sun’s energy can
only work during daylight. Passive sensor systems that measure the longer
wavelengths related to the Earth’s temperature does not depend on the Sun as
the source of illumination and can be operated at any time. Active remote
sensing techniques, for example radar and laser, have their own source of
energy. Active sensors emit a controlled beam of energy to the surface and
measure the amount of energy reflected back to the sensor. The main advantage
of active sensor systems is that they can be operated day and night, have a
controlled illuminating signal, and are typically not affected by the atmosphere
[22]. Table (2-2) shows the classification of sensors.
15
PASIVE SENSORS
(Detect the reflected or emitted radiation from natural sources)
ACTIVE SENSORS
(Detect reflected responses from objects that are irradiated from
artificially generated energy sources such as radar)
Non-Scanning • Non-Imaging: (they are types of
profile recorder, examples: Microwave sensor, Gravimeter, Spectrometer)
• Imaging: (Examples of this are the cameras which can be monochromatic, Natural Color, Infrared etc.)
Scanning • Imaging: 1- Image plane scanning, example,
TV camera solid scanner 2- Object plate scanning, example,
and Optical Mechanical scanner.
Non-Scanning • Non-Imaging: (they are types of
profile recorder, examples: Microwave Radiometer, Microwave Altimeter, Laser Water Depth, Laser Distance Meter)
• Imaging: (it is Radar, examples, object plane scanning
1- Real Aperture Radar 2- Synthetic Aperture Radar
Scanning • Image plane scanning: 1-Passive Array Radar.
Table (2-2): Sensor systems [1]
16
2.6 Landsat Systems
In 1967, the National Aeronautics and Space Administration (NASA),
encouraged by the U.S. Department of the Interior, initiated the Earth Resource
Technology Satellite (ERTS) program [18].
On 23 July 1972, the first civilian satellite dedicated to obtaining repetitive
remote sensing data of the Earth with a spatial resolution of less than 100m was
launched [16]. The first two satellites launched were not called Landsat. Their
original designations were: ERTS–A and ERTS–B (Earth Resources
Technology Satellite). Prior to the launch of ERTS–B the name Landsat was
adopted. ERTS–A was renamed as Landsat 1 and ERTS–B left the launch pad
as Landsat 2 [23].
Satellite systems have a number of advantages and disadvantages
compared with aerial remote sensing. A single Landsat scene covering an area
of 185×185 km2 is the equivalent of around 1000 photographs. To form all
these photographs which may have been taken under varying lighting
conditions into a mosaic is impracticable. However, satellite images allow large
areas that have been imaged in very short time period to be analysed
synoptically. Scale distortions are minimised because topographic variations are
very low compared with the latitude at which the satellite revolves, and only a
small rotation in the mirror angle is required to image a wide area [16].
17
2.6.1 Landsat 1, 2, and 3
The first Landsat scene consisted of three satellites that were launched
between 1972-1978 at a nominal altitude of 918 km, this orbit did not prevent
obtaining images for a 185 km wide swathand global coverage every 18 days.
The orbital plane for a Landsat satellite carried two types of imaging
systems, these are [16]:
• A Return Beam Vidicon (RBV): it obtained data in 3-bands with 80m
resolution; it has not been used in the second Landsat series.
• Multispectral Scanner (MSS): that obtained data in four wavebands
corresponding to the green (band 4), red (band5), infrared (band 6 and 7),
and thermal infrared (band 8).
2.6.2 Landsat 4 and 5
Landsat 4 and 5 are the second generation of Landsat series launched in
July 16, 1984 and March 1, 1984, respectively [17]. Landsat 4 and 5 carried a
replacement for MSS known as the Thematic Mapper (TM), which can be
considered as an upgraded MSS. In these satellites, both the TM and MSS were
carried on an improved platform that maintained a high degree of stability in
orientation as a means of improving geometric qualities of the imagery [19].
18
2.6.3 Landsat 6
The Enhanced Thematic Mapper (ETM), planned as a replacement for
Landsat 6, was designed to acquire imagery over a 185 km swath, with a 16 day
coverage cycle. It was designed to acquire seven channels of data in the visible,
near infrared, shortwave infrared and thermal regions. The thermal band would
have a resolution of 120 m, the others 30 m, Plus an additional panchromatic
band at 15 m. The ETM was lost with the failed launch of Landsat 6 in
September 1993 [19].
2.6.4 Landsat 7
The LANDSAT 7 satellite was successfully launched in April 15, 1999.
LANDSAT 7 is a 5000 pound-class satellite, designed with a 16-day repeat
cycle. The payload is a single nadir-pointing instrument, the Enhanced
Thematic Mapper Plus (ETM+) [24].
Landsat 7 has a unique and essential role in the realm of earth observing
satellites in orbit. No other systems match Landsat's high spatial resolution,
spectral range and radiometric calibration. In addition, the Landsat Program is
committed to provide Landsat digital data to the user community in greater
quantities, more quickly and at lower cost than at any previous time in the
history of the program.
The earth observing instrument on Landsat 7, the Enhanced Thematic
Mapper Plus (ETM+), replicates the capabilities of the highly successful
Thematic Mapper instruments on Landsats 4 and 5. The ETM+ also includes
new features that make it a more versatile and efficient instrument for global
19
change studies, land cover monitoring and assessment, and large area mapping
than its design forebears [25]. The primary new features on Landsat 7 are:
• a panchromatic band with 15m spatial resolution
• on board, full aperture, 5% absolute radiometric calibration
• a thermal IR channel with 60m spatial resolution
Table (2-3) shows the characteristics of Landsat 1 to 7 missions, and table
(2-4) shows current remote sensing systems and their major characteristics.
20
Landsat 1-3 Multispectral scanner
Landsat 4-5 Thematic Mapper
Landsat 7 Enhanced Thematic Mapper+
Operation 1972-1982 since 1982 since 1999
Altitude 915 Km 705 Km 705 Km
Revisit Period 18 days 16 days 16 days
Swath Width 185 Km 185 Km 185 Km
Ground Resolution 79 × 79 m2 30 × 30 m2 30 × 30 m2
Spectral Bands
1 = 0.45 - 0.52 µm 1 = 0.45 - 0.52 µm BLUE (B)
4 = 0.50 - 0.60 µm 2 = 0.52 - 0.60 µm 2 = 0.52 - 0.60 µm GREEN (G)
5 = 0.60 - 0.70 µm 3 = 0.63 - 0.69 µm 3 = 0.63 - 0.69 µm RED (R)
6 = 0.70 - 0.80 µm 4 = 0.76 - 0.90 µm 4 = 0.76 - 0.90 µm NEAR INFRARED
(NIR) 7 = 0.80 - 1.10 µm
5 = 1.55 - 1.73 µm 5 = 1.55 - 1.73 µmMIDDLE
INFRARED (SWIR1)
7 = 2.08 - 2.35 µm 7 = 2.08 - 2.35 µmMIDDLE
INFRARED (SWIR2)
8 = 10.4 - 12.5 µm (237 × 237 m2)
6 = 10.4 - 12.5 µm (120 × 120 m2)
6 = 10.4 - 12.5 µm (60 × 60 m2)
THERMAL INFRARED
(TIR)
8 = 0.52 - 0.9 µm (15 × 15 m2)
PANCHROMATIC(Pan.)
Table (2-3): Characteristics of Landsat 1 to 7 missions [26]
21
Remote Sensing Systems No. of Bands
Spatial Resolution
(meters)
Temporal Resolution
(days) GOES Series (east
and west) 4, Panchromatic 700 0.5/hr
SeaWiFS 8 1130 1
ASTER 14 15-30-90 5-16
NOAA-9 AVHRR LAC 5 1100 14.5/day
NOAA- K, L, M 6 1100 14.5/day
ATLAS 14 2.5 to 25 Variable
AVIRIS 224 2.5 or 20 Variable
OMEGA 351 300-5000 Variable
Table (2-4): Selected current remote sensing systems and their major characteristics [18]
22
2.6.5 Characteristics of ETM+ Spectral Bands
Landsat 7 is new generation of Landsat satellite that replaces the Landsat
5. The functions of each band in the sensor to the wavelength like ETM+ are
follows [27]:
1. Band 1 (0.45μm-0.52μm), the energy that receive by the band can make a
deep penetration to water column and use to support the land use characteristic
analysis, soil and vegetation.
2. Band 2 (0.52μm -0.60μm), especially are built to sense the top reflection of
vegetation in green spectrum that situated in the two chlorophyll absorption
spectral channels. The observation in this band is means to notice the vegetation
differences and the fertilization level grading.
3. Band 3 (0.63μm -0.69μm), are an important channel to separate vegetation.
The channel is in one part of the chlorophyll absorption and strengthens the
visible contras between vegetation and not vegetation, and also to sharpen the
contras in vegetation class.
4. Band 4 (0.76μm -0.90μm), are plan to be perceptive to the amount of
vegetation. This can help identify plant and can strengthen contras between
plant-soil and land-water.
5. Band 5 (1.55μm-1.73μm), are the important channel in identifying plant
species, water resistance in plant and the soil humidity condition.
6. Band 6 (10.40 μm-12.50μm), are important channel to observed the earth
heat surface.
7. Band 7 (2.08μm-2.35μm), are an infrared channel that are known to classify
and analyze vegetation, soil humidity separation and other effect that are
involve with heat.
23
8. Band 8 (0.52 μm-0.9μm), include portion of the visible and the near infrared
bands [27].
CHAPTER THREE
BANDS TRANSFORMATION TECHNIQUES &
CLASSIFICATION
24
3.1 Introduction
The multispectral nature of most remote sensing data makes it possible
for spectral transformations to generate new sets of image components or
bands. The transformed image may make evident features not discernable in the
original data or alternatively, it might possibly preserve the essential
information content of the image with a reduced number of the transformed
dimensions [28].
In the context of image processing, the term feature extraction (or feature
selection) are not only geographical features, visible on an image but, rather,
statistical characteristics of image data-individual bands, or combinations of
band values, that carry information concerning systematic variation within the
scene [19].
3.2 Multispectral Remote Sensing Images
Digital images are representation of an object obtained by digital sensor
device. The sensor records digital numbers (DNs) which are related to a
property of the object such as its reflectance. Digital remotely sensed image is
typically composed of picture elements (pixels) [29].
Satellite and airborne multispectral sensors provide data in the form of
several images of the same area of the Earth’s surface, but taken through
different spectral windows or bands [5]. A band wavelength range measured by
a remote sensing system often used to designate how many layers of data are
recorded by a system, e.g. Landsat TM is a seven band system. The term
25
channel is occasionally used instead of band. Multispectral imagery provides
more information that data collected in any single spectral band image.
Digital Number value DN assigned to a pixel which is related to the
parameter being measured by a remote sensing system, a low value of DN may
represent a low reflectance in particular waveband while high value of DN
represent a high reflectance. DN values must be integers and common values
between 0-255, for an n-band dataset each pixel is associated with n digital
numbers [16].
A significant advantage of multispectral imagery is the ability to detect
important differences between surface materials by combination spectral bands.
Different materials may appear virtually the same within a single band [30].
Color images can be modeled as three band monochrome image data,
where each band of data corresponds to a different color. The actual
information stored in the digital image data is the brightness value in each
spectral band. Whenever the image is displayed, the corresponding brightness
information is displayed on the screen by picture elements that emit light
energy corresponding to that particular color.
Typical color images are represented as red, green, and blue, a display of
three band image in which one band is applied to the red region, one to the
green and one to the blue, it called RGB display [31].
3.3 Band Ratios
Each object has its own spectral reflectance pattern in different wavelength
portion. Spectral reflectance curve is a kind of fingerprint of the object. The
objects may have high reflectance value in some spectral portion, however, it
26
may absorb in another spectral region. According to this information, the main
concept of the band ratios technique was developed [32].
A band ratio is very simple and powerful technique in the remote sensing.
Basic idea of this technique is to emphasize or exaggerate the anomaly of the
target object [32].
Ratio is accomplished by dividing the data base DNs in one spectral band,
by the data base DNs in a second spectral band for each spatially registered
pixel pair. The ratio algorithm can be expressed in the form [3]:
DNo , =(DN ,
DN ,) (3-1)
Where: DNo(j,s)= output digital number at line j, sample s,
DNx(j,s)=input digital number of band x at line j, sample s,
DNy(j,s)=input digital number of band y at line j, sample s.
On band ratios image the extreme black-and-white tones of the gray scale
represent the maximum difference in spectral reflectivity between two bands.
The darkest tones are targets for which denominator of the ratio is greater than
the numerator. Conversely the numerator is greater than the denominator for the
lightest tones [17].
Ratio techniques are usually used to enhance the spectral differences
between surface materials that difficult to detect in raw images. Moreover, such
techniques may suppress the effect of variable illumination resulting from
topographic variations, slope shadows, seasonal changes, and either differences
in sunlight angle or intensity may also be eliminated. Information gained by
27
ratio transforms is considered to be almost new and cannot be obtained from
either of the bands independently [33].
Every pixel is one of the commonest arithmetic operations performed on
digital remote sensing data. If any area of forest-land on either side of a hill is
imaged in two bands, see figure (3-1), it will not have a uniform signature on
either image because the digital numbers will be consistently lower for both
bands in shadow. So classification of this scene may correspondingly assign
this part to different classes even though they should be considered as
belonging to a single class thus rationing effectively suppresses the topographic
effect and enhances gradient changes in the spectral reflection and reflectivity
curves of different materials, which may accentuate the difference between
them. This example was chosen deliberately because ratioing is particularly
relevant to vegetation studies [20].
Figure (3-1): Reduction of scene illumination effects through spectral ratioing [20]
28
The utility of any given spectral ratio depends upon the particular
reflectance characteristics of the features involved and the application at hand.
The number of possible ratios that can be developed from n bands of the data is
n(n-1). Thus, for the six non thermal bands of Landsat TM or ETM+ data there
are 6(6-1), or 30, possible combinations [20].
Ratio images can often provide information that cannot be obtained from
the band singly; there may be occasions where ratio image provides less
information. Thus, ratioing may result in two surfaces that could be
distinguished on two separates bands being indistinguishable on a ratio image.
For explanation, two surfaces have been imaged by two bands as follows:
Bands (1) DN (1) =60 DN (1) =80
Bands (2) DN (2) =30 DN (2) =40
For the above example the two surfaces can distinguished on either band
(1) or band (2), where in both cases surface B has a higher DN and will
consequently appear pale (brighter). However both surfaces have the same ratio
value =2 and would be indistinguishable on a ratio image of bands 1 and 2 [21].
Ratio=2 Ratio=2
Surface A Surface B
29
3.4 Vegetation Indices
Vegetation indices are quantitative measures, based on digital values that
attempt to measure biomass or vegetation vigor. Usually a vegetation index is
formed from combinations of several spectral values that are added, divided, or
multiplied in a manner designed to yield a single value that indicates the
amount or vigor of vegetation within a pixel. High values of the vegetation
index identify pixels covered by substantial proportions of healthy vegetation.
The simplest form of vegetation index is simply a ratio between digital values
from separate spectral bands [19].
The green vegetation would have slightly higher reflectance in the green
portion of the visible spectrum than in blue or the red. The dips in the blue and
the red are due to photosynthetic absorption at those wavelengths. To take
advantage of this particular spectral feature, researchers have developed a
metric known as the Normalized Difference Vegetation Index (NDVI) which is
equal to the signal received in the near infrared minus the signal in the red
divided by the sum of the signals in those bands. (This normalization is
performed to reduce intra-class variation due to lighting and topographic
effects) [34].
NDVI NIR RNIR R
(3-2)
Where NIR: Near Infrared band (0.76 - 0.90 µm),
R: Red band (0.63 - 0.69 µm).
30
This index largely compensates for changing illumination conditions,
surface slope, and viewing aspects. Healthy vegetation reflects strongly in the
near infrared portion of the spectrum while absorbing strongly in the visible red.
Other surface types, such as soil and water, show near equal reflectance in both
near infrared and red portions. Thus, the NDVI image would significantly
enhance the discrimination of vegetation from other surface cover types.
In general, vegetation yields high positive NDVI values. Clouds, water,
and snow yield negative values due to larger red reflectance than near infrared.
The NDVI values for rock and bare soil are near zero due to their similar
reflectance in both bands. Therefore, in a NDVI image the lighter tones are
associated with dense coverage of healthy vegetation.
3.5 Spectral Indices
There are other indices which are used in environmental monitoring and
image enhancement, there as follow:
1. Bare Soil Index (BSI) Bare Soil Index used to identify the bare soils which include bare area,
where [12]:
BSI SWIR R _ NIR BSWIR R NIR B
+1 (3-3)
Where SWIR1: is Short Wave Infrared one (1.55 - 1.73 µm),
B: is blue band (0.45 - 0.52 µm),
R: is Red band (0.63 - 0.69 µm).
31
2. Normalized Differential Water Index (NDWI) This index is used to oversee the situation of water in the study area. The
ratio between red and SWIR1 spectral region dearly enhanced water bodies to
the brighter pixels, where [12]:
NDWI R SWIRR SWIR
1 (3-4)
3. Urban Index (UI) At Most enhanced the urban activities such as housing, road, industrial
complex and so on, where [12]:
UI SWIR RSWIR R
1 (3-5)
Where SWIR2: Short Wave Infrared two (2.08 - 2.35 µm).
Originally all these equations produce relative value ranges from -1 to +1.
Number 1 has been added in each equation to avoid the negative value in
further analysis. Therefore, the entire resulted images will have the value
between 0-2, where higher value represents better existences of the selected
environmental parameters (BSI, NDWI, and UI) [12].
3.6 Separating Soil Reflectance from Vegetation Reflectance
In agricultural scene, however, reflections from individual plants, or
individual rows of plants, are closely intermingled with the bare soil between
plants and between rows of plants, so that reflectance are mixed even at the
32
finest resolution. But even after leaves have fully emerged, soil can still
contribute to reflectance because of penetration of some wavelengths through
the vegetation canopy.
In many spectral observations are made of surfaces of bare soil, the values
of red and near infrared brightness resemble those in figure (3-2) [20]. The soil
brightness tends to fall on a straight line as a soil becomes brighter in the near
infrared, so does it tend to get brighter in the red. Dry soils tend to be brighter in
both spectral regions and appear at the high end of the line C; wet soils tend to
be dark, and are positioned at the low end B.
In 1977 Richardson and Wiegand defined this relationship, known as the
soil brightness line, and recognized that spectral response of living vegetation
will always have a consistent relationship to the line. Soils typically have high
or modest response in the red and infrared regions, whereas living vegetation
must display low values the red (due to the absorption spectra of chlorophyll)
and high values in near infrared.
Figure (3-2): Soil brightness line [19]
33
Thus, point X typifies a “pure” vegetation pixel, with a spectral response
determined by vegetation alone, with no spectral contribution from soil. In
contrast, point Y typifies a response from a partially vegetation pixel, it is
brighter in the red and darker in the near infrared than is X [19].
3.7 Tasseled Cap Transformation
In 1976, Kauth and Thomas produced a transformation of the original
Landsat MSS data space to a new four dimensional spectral space. It was called
the Tasseled Cap or Kauth-Thomas transformation. The transformation consists
of linear combinations of the four MSS bands to produce a set of four new
variables [18]:
TC1=+0.332MSS1+0.603MSS2+0.675MSS3+0.262MSS4 (3-6)
TC2=-0.283MSS1-0.66MSS2+0.577MSS3+0.388MSS4 (3-7)
TC3=-0.899MSS1+o.428MSS2+0.076MSS3-0.041MSS4 (3-8)
TC4=-0.016MSS1+0.131MSS2-0.452MSS3+0.882MSS4 (3-9)
These new bands are designated as, for example, “TC1”, “for Tasseled
Cap band 1”. Although these four new bands do not match directly to
observable spectral bands, they do carry specific information concerning
agricultural scenes [19]. It can therefore in corporate more information than
ratio or modified ratio indices (like most vegetation indices) [35].
TC1 interpreted as brightness, a weighted sum of all four bands. TC2 is
designated as greenness, a band that conveys information concerning the
34
abundance and vigor of living vegetation. TC3 depicts yellowness, derived
from the contrast between red and green bands. Finally, TC4 is referred to as
nonsuch because it cannot be clearly matched to observable land space features
and is likely to carry system noise and atmospheric information.
The first two bands (TC1 and TC2) usually convey almost the information
present in an agricultural scene often 95% or more. Therefore, the essential
components of an agricultural landscape are conveyed by a two-dimensional
diagram, using TC1 and TC2 figure (3-3) [19].
Figure (3-3) shows a plot of data which is physically-based on crop
growth. The growing cycle of crop started from bare soil, then to green
vegetation and then to crop maturation with crops turning yellow. This different
stage of vegetation growth has made the data distribution in the two
Figure (3-3): Seasonal variation of a field in defined by Greenness and Brightness [19]
35
dimensional multispectral space as shown in figure (3-4) which appear in a
shape of cap [28].
The Tasseled Cap transformation on six non-thermal bands of Landsat TM
or ETM+ data can be performed using the following formulas [19]:
Brightness = 0.3037(TM1) + 0.2793(TM2) + 0.4743(TM3)
+ 0.5585(TM4) + 0.5082(TM5) + 0.1863(TM7) (3-10)
Greenness = -0.2848(TM1) - 0.2435(TM2) - 0.5436(TM3)
+ 0.7243(TM4) + 0.0840(TM5) - 0.1800(TM7) (3-11)
Wetness = 0.1509(TM1) + 0.1973(TM2) + 0.3279(TM3)
+ 0.3406(TM4) - 0.7112(TM5) - 0.4572(TM7) (3-12)
Figure (3-4): 2D scatter plot of two Tasseled Cap components [35]
36
3.8 Principle Component Analysis (PCA)
In satellite imagery, it is not uncommon to find that a strong degree of
correlation exists between the multi-spectral bands. Such correlation indicates
that if reflectance is high at a particular location on one band, they are also
likely to be high on the other band. In the extreme case, if two bands were
perfectly correlated they would essentially describe the same information. It is
not unusual to find that an image with 7 bands, such as Landsat Thematic
Mapper, actually contains far fewer than 7 bands of true information.
Principal Components Analysis (PCA) can be used to transform a set of
image bands such that the new bands (called components) are uncorrelated with
one another and are ordered in terms of the amount of image variation they can
explain. The components are thus a statistical abstraction of the variability
inherent in the original band set. Since each of the components produced by this
transformation is uncorrelated with the other, each carries new information.
Also, because they are ordered in terms of the amount of information they
carry, the first few components will tend to carry most of the real information in
the original band set, while the later components will describe only minor
variations [36].
The techniques of PCs analysis which provides means of generating set of
m-images with particular properties (zero correlation between m-bands and
maximum variance) from set of m-correlated images. It is usually found that, as
a result of the maximum variance property, of the PCs information in m
correlated bands is expressible in term of P (PCs), where p is less than m [37].
37
PCA has several advantages, they are [38]:
1. Most of the variance in a multispectral data set is compressed. The
information contents of a number of bands, for example, (six TM bands) into
just two or three transformed PC images. The transformation provides
uncorrelated variables with the first two or three components generally
containing (90 to 97) percent of the variance of the original six TM bands.
2. Noise may be relegated to the less correlated PC images.
3. Spectral differences between materials may be more apparent in PC images
than in individual bands.
However the PCA can be used for the following applications [2]:
• Effective classification of land use with multispectral data.
• Color representation or visual interpretation with multispectral data.
• Change detection with multi-temporal data.
• Compression of multispectral data.
Figure (3-5): PCA reduce the dimensionality
38
3.8.1 Mathematical Representation [5]
Let us consider two dimensional image represented by an array:
(3-13)
And the mean is defined by
(3-14)
Where is the expectation value. So the covariance matrix could be
defined as follow:
(3-15)
(3-16)
Where T indicates to the transpose.
The diagonal elements are the variance of each image, i.e. .
where is the standard deviation. The eigenvector and eigenvalue of the
covariance matrix denoted by and respectively.
39
The transformation matrix consists of the eigenvector of the covariance
matrix:
(3-17)
Where is unity matrix such that . In order to compute the
principle component we should diagonalized the covariance matrix using the
following expression:
0 00 0 0 0
(3-18)
So each eigenvalue corresponds to the variance of a new PC image, the
variance being related to the amount of contrast. The new matrix is diagonal
and the components have been chosen to be uncorrelated and
.
40
3.9 False Color Composite (FCC)
Combining data from a number of bands into color images allows more
information to be extracted. The image must be displayed in those colors
irrespective of the wavelength at which the image was obtained, (the human eye
responds to radiation between wavelength of approximately 0.38 and 0.72) [39].
The aims of false color composites are to [16]:
• Discriminate different bands combination to yield information on
different surfaces and in a single image.
• Investigate how surface appears different in visible, infrared and
microwave electromagnetic spectrum.
The large number of bands for the ETM+ system allows the production of
much great numbers of false color composite.
3.9.1 Selecting the Best Color Composites
The number of different color composites that can be produced by
combining single band black-and-white images in groups of three (red, green,
and blue components) can be easily determined for any multispectral imaging
system by use of the following equation [3]:
N !! !
(3-19)
Where: N=number of different color composites,
n=number of single band images available,
3=number of color assignments (red, green, and blue components).
41
Selecting the three band combination for color compositing that will allow
for the optimum discrimination of material classes in a scene can be both
difficult and time consuming because of the large number of possible
combinations. For example there are 20 different three band color composites
produced from the six non thermal ETM+ images [3]. To help overcome the
selection problem, various quantitative criteria have been developed to assist in
selecting which band combinations to include in color composites. The
Optimum Index Factor (OIF) is one such criterion. It ranks all possible three
band combinations based on the total variance present in each band and the
degree of correlation between bands [20]. High OIF values indicate bands that
contain much information (e.g. high standard deviation) with little duplication
(e.g. low correlation between the bands). By using the OIF technique, three
band color composites can be evaluated on their effectiveness for display. OIF
calculated using the following equation [14]:
OIF ∑∑ | | (3-20)
Where σ(i) is the standard deviation of band i, and r(j) is the correlation
coefficient between any two bands in the combination[24].
In this project the OIF factor technique has been adopted on the original
ETM+ bands and on the bands ratios to select best three bands combination for
classification.
42
3.10 Multispectral Classification
In remote sensing data provide the most significant clues about what is on
the ground, the spectral data in a scene can be recorded with multispectral
images. Each pixel in a multispectral image has spatial coordinates x & y and
spectral coordinates λ. It has characterized by its spectral signature, which is
determined by the relative reflectance in different wavelength bands.
Multispectral classification is a data extraction process that analyzes these
spectral signatures then assigns pixels to categories based on similar signatures
[40]. Pixels that do not fall within spectral class (cluster) are considered
unclassified.
The results from the classification process are typically in the form of a
thematic map from which data can be used to solve a particular problem or to
provide important data unobtainable from other sources.
There are two major approaches to multispectral classification of remote
sensing data [41]:
• Unsupervised classification technique (clustering).
• Supervised classification technique (human assisted).
3.10.1 Unsupervised Classification
Unsupervised classifiers do not utilize training data as the basis for
classification. Rather, this family of classifiers involves algorithms that examine
the unknown pixel in an image and aggregate those into a number of classes
based on natural grouping or clusters presented in the image values. This
method is called unsupervised classification because the data analyst has little
43
control over the establishment of the decision regions, so it is more computers
automated. It allows specifying some parameters, which the computer uses to
uncover statistical pattern that are inherent in the data [20]. Clustering
algorithms are used for unsupervised classification of remotely sensed data.
Many clustering algorithms have been developed for a wide variety of purposes
methods of unsupervised classification attempt to find clusters in the
distribution of the pixels [42].
The aim of clustering methods is to partition a set of data points into a
given number of clusters.
Clustering algorithms used for the unsupervised classification of remotely
sensed data generally vary according to the efficiency with which the clustering
takes place [29].
The advantages of clustering algorithm:
• No pre-knowledge required.
• Suitable for the application with large database.
There are more common methods of the determination of the clusters in the
data such as K-means algorithm, and Iterative Self Organizing Data Analysis
(ISODATA) algorithm. The first step of ISODATA algorithm, an initial mean
vector is arbitrarily specified for each class. Each pixel in the scene is assigned
to the class whose mean vector is closest to the pixel vector forming the first set
of decision boundaries. A new set of class mean vectors is then calculated from
the results of the previous classification and the pixels are resigned to the
classes. The procedure continues until there is no significant change in pixel
assignment from iteration to the next [40].
44
3.10.2 Supervised Classification
Supervised classification can be defined informally as the process of
using samples of known identity (i.e. pixels already assigned to informational
classes) to classify pixels of unknown identity (i.e. to assign unclassified pixels
to one of several informational classes) samples of known identity are those
pixels located within training areas, the selection of these training data is a key
step in supervised classification [19].
The classification technique will be as follows:
• Training site selection,
• Supervised classification,
• Output stage,
• Accuracy assessment.
a. Training Stage
Training area is an area on an image which the operator, using prior
knowledge, knows it belongs to one specific class which will be used in a
supervised classification procedure; it is considered the first step of any
classification procedure to train the computer program to recognize the class
signature of interest. The main problem is the selection of real representative
sample areas for the selected classes [16].
The selecting of training data has the most effect on the accuracy of the
classification results. Training stage requires close interaction between the
image analyst and image data, it also requires substantial reference data,
45
therefore, the analyst uses prior knowledge, derived from field surveys, photo
interpretation and other sources about small region of the image, to be classified
to identify these pixels that belong to the classes of interest [41].
An important statistical step of selecting training data is that, a sufficient
number of pixels may be needed to define the class signature properties
accurately [20]. Sometimes, to obtain a good estimation of classes statistics it
may be necessary to choose several training fields for the one cover type,
located in different regions of the image, a sufficient number of pixels must be
used to define a class signature propertied accurately [41].
b. Classification Stage
Supervised classification is the procedure most often used for extraction
of quantitative information of remote sensing image data; it rests upon using
suitable algorithm to label the pixels in an image representing particular cover
type, or class.
Many classification methods are available such as Minimum Distance to
Mean classifier, the Parallelepiped classifier and the Maximum Likelihood
classifier. The Maximum Likelihood procedure is unquestionably the most
commonly used procedure for classification in remote sensing. The foundation
for this approach expresses the relationship between evidence, prior Knowledge
and the likelihood that a specific hypothesis is true [36]. In this method the
computer calculates the probability of a pixel value occurring in the distribution
of the class to be a member of a particular of a certain class. The class with
maximum probability is recorded as the correct class of that pixel. After
evaluating the probability in each category, the pixel would be assigned to the
46
most likely class (highest probability value) or be labeled “unknown”. The
Maximum Likelihood classification is the most accurate in results; its accuracy
depends on the skill in selection of training areas [20]. The equation for the
Maximum Likelihood classifier is as follow [39]:
log log |∑ | ∑ (3-21)
Where:
= Weighted distance (Likelihood),
= Data vector,
= The mean vector of the sample of class (i),
∑ = The covariance matrix,
|∑ |= Determinate of ∑ , = percent probability that any pixel is a member of class (i),
log = natural logarithm function.
c. Output Stage
The results of image classification can be introduced in several forms.
The three general forms are [43]:
• Tabular data: These are tables listing necessary statistics of the whole
image or any desired part.
• Graphical products: Display of classified image (the classes) in different
colors, tones or characters, each cell represents its related ground cover
classes.
47
• Digital information files: Output results may be recorded on computer
disks for further analysis and processing.
d. Accuracy Assessment
As the classification process is preformed, there must be some errors. To
correctly performed classification accuracy assessments, it is necessary to
compare two sources of information: (1) the remote-sensing-derived
classification map and (2) what we will call reference test information (which
may in fact contain error). The relationship between these two sets of
information is commonly summarized in an error matrix.
Accuracy assessment of the classification represented by error matrix,
which is a square array of numbers laid out in row and columns that express the
number of sample units assigned to a particular category relative to the actual
category as verified in the field. The columns normally represent the reference
data, while the rows indicate the classification generated from the remotely
sensed data. An error matrix is a very effective way to represent accuracy
because the accuracy of each category is clearly described [29].
The overall classification accuracy can be calculated using the equation
[44]:
P(%)= ∑ N∑ T
*(100%) (3-22)
Where:
∑ N = the total number of correctly classified pixels in all classes.
∑ T = the total number of test pixels in all classes.
CHAPTER FOUR
RESULTS & DISCUSSION
48
4.1 Introduction
The results of band transformation techniques and image classification of
the study area are illustrated in this chapter. All images are adopted by using
ENVI version 4.1, a digital image processing software package, for more detail
see the appendix. The type of data (images) which are used in transformation
techniques is 6 ETM+ band (non-thermal band) images acquired in 15/9/2002.
Figure (4-1) shows the 6 band images, with their histograms, and figure (4-2)
shows the scheme of the work.
Band 1
Band 2
Figure (4-1): 6 non thermal ETM+ band images with their histograms (continue)
49
Band 3
Band 4
Band 5
Band 7
Figure (4-1): 6 non thermal ETM+ band images with their histograms
50
Raw data (6 nonthermal ETM+ satellite images acquired in 15/9/2002)
Tasseled Cap Band Ratios PCA
30 different ratios
Visual interpretation
5/7, 4/5, 2/5, 1/4, 2/4
Optimum Index Factor
False Color Composites
4/5-1/4-2/5
Classification
Supervised Classification
Accuracy
False Color Composites False Color Composites
Optimum Index Factor
3-4-7
False Color Composites
Figure (4-2): The scheme of the work
Unsupervised Classification
51
4.2 Band Ratios
Remotely sensed data acquired from high altitude platforms have
strongly developed during the last decades and became a very important tool for
many applications. Image enhancements are the techniques by which the image
becomes more informative. Band ratios represent one of the spectral
enhancement techniques which represent the ratio between digital numbers for
two or more different bands. The whole information that content in
multispectral images gives an opportunity to use such technique.
The number of possible ratios has been computed using n (n-1); where n
is the number of spectral bands, so there are 30 different ratios, 15 original and
15 reciprocal. Our goal is focused on getting the best ratios for land cover and
land use features (water, vegetations, bare lands, crop lands, and urban areas),
this done by visual interpretation for the 30 different ratios.
The images contain detailed record of features on the ground at the time of
data acquisition. The images have been examined, and frequently, other
supporting material such as maps and reports of the field observations. In this
work, visual interpretation is made as to the physical nature of features and how
much details of this features appeared in the 30 different ratios. According to
this, five different ratios have been chosen to be the best ratios represent land
features based on visual interpretation. These ratios are given in table (4-1).
52
Land Feature Band Ratio
Water 5/7
Vegetations 4/5
Bare lands 2/5
Urban areas 1/4
Crop lands 2/4
The general characteristics of major land cover and land use features that
have been seen on the five ratios images are discussed below:
1. Water:
The best ratio which identifies water is 5/7 (SWIR1/SWIR2). The
existence of water in different locations of study area, it includes Tigris River
and the canals that came from it. Water appeared as black in this ratio, this due
to the low reflectance of it in the IR bands. The small canals were merged to
the other land features because the limitation of the resolution of the image,
figure (4-3).
Table (4-1): The best band ratios for certain features
53
2. Vegetations:
Healthy vegetations have high reflectance in the range of infrared
wavelength, and low in the range of visible red wavelength. The best ratio
which identified vegetations in study area is 4/5 (NIR/SWIR1), where they
appeared as bright area due to the high reflectance at NIR band, figure (4-4).
Vegetations in study area are represented by agriculture and field lands which
are located at the both sides of the river and canals.
Figure (4-3): 5/7 band ratio for water
54
3. Bare lands:
The existence of bare lands in study area is out of city which is not used
for residence or agricultures and its surface are predominantly thin soils or
sands. The best ratio which identified barren areas is 2/5 (Green/SWIR1), which
its tone appeared as medium gray to little dark tones from other land features as
shown in figure (4-5). The appearance of bare lands is different due to different
surface roughness, content, moisture content and the presence of salts.
Figure (4-4): 4/5 band ratio for vegetations
55
4. Urban areas:
Urban areas represented the areas of intensive land use where much of
these areas consist of buildings, streets, and highways which connected the
cities. The 1/4 (Blue/NIR) ratio has been adopted to be the best ratio for
identifying the urban areas, which separated this lands from the other features,
and appears as bright tone, see figure (4-6). Some roads are indistinguishable
because the limitation of resolution.
5. Crop lands:
The ratio which identified these lands is 2/4 (Green/NIR), which appears
as lighter gray due to the high reflectance at visible band; figure (4-7).
Figure (4-5): 2/5 band ratio for bare lands
56
Figure (4-6): 1/4 band ratio for urban areas
Figure (4-7): 2/4 band ratio for crop lands
57
Tables (4-2) and (4-3) show the basic statistics and correlation of band
ratios respectively.
Bands MIN. MAX. MEAN STD
Band 5/7 0 28.5 0.96 0.61
Band 4/5 0 85 1.44 3.22
Band 1/4 0 36.25 0.82 0.66
Band 2/4 0 29.5 0.84 0.57
Band 2/5 0 58.33 0.81 0.59
Band 5/7 Band 4/5 Band 1/4 Band 2/4 Band 2/5
Band 5/7 1 0.046 0.075 0.071 0.017
Band 4/5 0.046 1 0.190 0.221 0.073
Band 1/4 0.075 0.190 1 0.964 0.211
Band 2/4 0.071 0.221 0.964 1 0.248
Band 2/5 0.017 0.073 0.211 0.248 1
Table (4-2): Multivariate statistics of the band ratios
Table (4-3): Correlation between band ratios
58
4.3 Tasseled Cap Transformation
Tasseled Cap transformation is one of the available methods for
enhancing spectral information content of Landsat ETM+ data. It attempts to
reduce the amount of data layers (dimensionality) needed for interpretation or
analysis by using mathematical equations to transform the original bands into
new dimensional space.
Tasseled Cap transformation has been adopted on the six non-thermal
ETM+ bands of study area. The production of Tasseled Cap consists of three
main indices: brightness, which contains the most information, greenness,
which conveys information about living vegetation since it displayed areas with
vegetation as the lightest tone (white areas), and wetness which is concern with
surface wetness where the river and canals appear as white as shown in figure
(4-8).
Tasseled Cap transformation showed high ability to improve
interpretability and extract information from the data which is not readily
visible in the raw form of ETM+ data of the study scene.
If the greenness and brightness components of a typical scene are plotted
perpendicular to one another on a graph to display the distribution of the data in
spectral plot, the shape of data plot is look like a cap, so that the name of
Tasseled Cap comes from this fact.
59
Brightness
Greenness
Wetness
Figure (4-8): The first three Tasseled Cap transformation indices
60
4.4 Vegetation indices
Vegetation indices kind of bands transformation techniques, one of them
is NDVI which is a good tool to transform multispectral data into a single
image band representing vegetation areas. The higher value of NDVI indicate
that the higher probability that the corresponding area on the ground has a dense
coverage of healthy green vegetation. The reflectance of vegetation is different
and depends on the type, situation and the water content of it.
NDVI has been adopted on the raw bands of the study area see figure
(4-9). The white area indicated dense healthy vegetations. The existence of
vegetations in study area is around the river and the canals.
Figure (4-9): NDVI for the study area
61
There are other indices which are used for feature extraction and
monitoring the land cover and land use, such as Bare Soil Index (BSI),
Normalized Difference Water Index (NDWI), and Urban Index (UI), where
computed in multispectral ETM+ images and tried to show their abilities for
monitoring the environment in respect of bare soil lands, water and urban
activities as shown in figure (4-10).
From figure (4-10) BSI shows the reflectance of bare lands as lighter gray,
this mean that there is high reflectance of bare lands at the 7. The urban areas
appeared as bare lands, this occurred because some urban features have similar
spectral signatures as bare lands and vice versa.
The result of NDWI shows that the river appeared as black tone but the
other canals appeared as white tone. The reflectance of water depends on the
turbidity and the depth of water and reaches its maximum value of reflectance
at the blue end of the spectrum and decreases as wavelength increases.
Unfortunately, the urban areas are undistinguishable in UI image and
appeared as mid gray tone. The appearance of vegetation is good in UI which
appeared as white.
62
BSI NDWI
UI
Figure (4-10): Spectral indices images
63
4.5 Principle Component Analysis
Principle Component Analysis (PCA) is a spectral enhancement which
can be used to compress the information content of a multispectral data set.
PCA uses mathematical algorithms to transform n bands of correlated data into
principle components which are uncorrelated.
PCA has been applied to a six non-thermal bands of this study. A set of
six principle components where obtained, and they are shown with their
histograms in figure (4-11).
Table (4-4) shows the multivariate statistics of the original bands of the
study area. Band 7 has the largest standard deviation (STD), this indicates that
the variance is large and the information in this band is very high. While band 1
has the smallest standard deviation, this indicates that the information in this
band is little.
Table (4-5) shows the covariance-variance matrix. The result of
this table shows that the largest value of variance is in band 7, and the
smallest is in band 1. This is obvious from table (4-4) for the standard
deviation.
In table (4-6), there are correlations between the bands. The largest
correlation is between band 1 and band 2; this means that there is common
information in these bands, i.e. these correlation values suggest that there is a
substantial amount of redundancy in the information contents among these
bands. The lowest correlation is between band 4 and band 7, this means that the
relation between them is very small, i.e. there is some information in band 7
that are not found in band 4.
64
Figure (4-11) shows that the first three principle components contain
most of the information in the original bands, and from histograms one can see
that the histogram of PC1 have wide range of brightness values, this mean that
PC1 image contain most of the information of the scene. However, it has been
observed that each principle component image can be used individually to
enhance and interpret the land features, for example, PC1 image shows the
ability to enhance the water regions which appears as black and the bare lands
which appears as light gray tone area, in PC2 image, it has been found to yield
enhanced the vegetation areas which appears as white areas around the river
and the canals and in PC3 image, it has been found that is contain some noise
rather than the previous PC’s, and it is much able to separate the river and the
canals from the other features.
65
PC1
PC2
PC3
Figure (4-11): PCA images with their histograms (continue)
66
PC4
PC5
PC6
Figure (4-11): PCA images with their histograms
67
Bands MIN. MAX. Mean STD
Band 1 0 255 85.10 44.46
Band 2 0 255 88.79 48.58
Band 3 0 255 97.74 51.70
Band 4 0 255 108.72 46.49
Band 5 0 255 102.98 50.66
Band 7 0 255 106.39 53.02
Band 1 Band 2 Band 3 Band 4 Band 5 Band 7
Band 1 1977.55 2131.12 2218.43 1120.73 1971.29 2062.94
Band 2 2131.12 2360.06 2472.43 1330.01 2233.43 2314.54
Band 3 2218.43 2472.43 2673.78 1354.04 2465.58 2558.82
Band 4 1120.73 1330.01 1354.04 2161.73 1373.42 1142.21
Band 5 1971.29 2233.43 2465.58 1373.42 2567.12 2606.89
Band 7 2062.94 2314.54 2558.82 1142.21 2606.89
2812.01
Table (4-4): Multivariate statistics of the original bands
Table (4-5): Covariance-variance matrix
68
4.6 False Color Composites (FCC)
From the results of band ratios, Tasseled Cap, and PCA can be used to
generate false color composites by combining three monochromatic (black and
white) images as color images.
From the result of Tasseled Cap, there are three monochromatic bands
(Brightness, Greenness, and Wetness). These three images have been used to
create FCC. From the result of PCA, the best three components which hold the
most information about the scene are PC1, PC2, and PC3. These three
components have been also used to create FCC.
Band 1 Band 2 Band 3 Band 4 Band 5 Band 7
Band 1 1 0.986 0.964 0.542 0.874 0.874
Band 2 0.986 1 0.984 0.588 0.907 0.898
Band 3 0.964 0.984 1 0.563 0.941 0.933
Band 4 0.542 0.588 0.563 1 0.583 0.463
Band 5 0.874 0.907 0.941 0.583 1 0.970
Band 7 0.874 0.898 0.933 0.463 0.970 1
Table (4-6): Correlation between bands
69
The problem has been come from selecting the best original bands and best
ratios combinations which hold the whole information about the scene.
Optimum Index Factor (OIF) has been adopted to solve this problem,
which it ranks all possible combinations based on the total variance present in
each band and the degree of correlations. The high OIF values indicated bands
that contained much information. OIF method has been adopted on the raw
ETM+ band combinations and on the ratios combinations, which interpreted
land cover and land use of the study scene. For the six raw ETM+ bands, there
are 20 different band combinations which are ranked in table (4-7) with their
OIF values. It is found that the combination of ETM+ raw bands 3, 4, and 7
gives the satisfactory highest OIF value. This indicated that there is low
correlation between bands. For the five band ratios there are 10 different color
combinations which are ranked in table (4-8) with their OIF values. High OIF
value is represented by the combination of band ratios 4/5, 1/4, and 2/5.
Figure (4-12) shows the result of the FCC of the adopted raw bands and
the transformed bands, which are hold the most information about the scene and
represented the good tools towards the classification process.
70
OIF
Order of combination of ETM+ bands
77.172 7 4 3
76.586 7 4 1
75.928 7 4 2
74.478 7 5 4
71.321 5 4 3
70.816 5 4 1
70.094 5 4 2
68.923 4 3 1
68.710 4 3 2
65.905 4 2 1
54.852 7 5 2
54.632 7 5 3
54.472 7 5 1
54.447 7 3 2
53.811 7 3 1
53.290 5 3 2
52.932 7 2 1
52.807 5 3 1
51.906 5 2 1
49.313 3 2 1
Table (4-7): Ranked OIF values of raw band combinations
71
OIF
Order of combination of ratio bands
47.486 2/5 1/4 4/5
44.111 2/5 2/4 4/5
43.337 2/5 4/5 5/7
20.581 1/4 4/5 5/7
17.925 2/4 4/5 5/7
15.926 2/5 1/4 5/7
11.211 2/5 2/4 5/7
8.081 2/4 1/4 4/5
2.272 2/4 1/4 5/7
1.288 2/5 2/4 1/4
Table (4-8): Ranked OIF values of ratio bands combinations
72
FCC image for raw bands 3, 4, 7 FCC image for ratio bands 4/5, 1/4, 2/5
FCC image for Tasseled Cap indices FCC image for PC
Figure (4-12): FCC images for different combinations
73
4.7 Images Classification
The classification process based on the information obtained from band
transformation techniques, visual interpretations, FCC images, and OIF
calculations, the classification process applied on:
1. The raw bands combination ranking of OIF calculation (band 3, band 4,
and band 7).
2. The ratio bands combination ranking of OIF calculation (4/5, 1/4, and 2/5).
3. The indices combination of Tasseled Cap transformation (brightness,
greenness, and wetness).
4. The first three components combination of PCA (PC1, PC2, and PC3).
4.7.1 Unsupervised Classification
Unsupervised classification method has been adopted, were similar pixels
are classified according to natural groupings called clusters. This method
considered an attempt of preprocessing to the supervised classification, in order
to get better understanding about the spatial structure of image data.
Unsupervised classification has been performed using the Iterative Self
Organizing Data Technique (ISODATA) with five clusters.
From the resulted unsupervised classification maps, five different classes
have been identified corresponding to: water (blue), vegetations (green), crop
lands (red), urban areas (cyan) and bare lands (yellow). Figure (4-13) shows the
classification maps for different color combinations. The classified images
74
show good representation of some classes and merges among the other classes,
especially the classified image of the raw combination, where the bare lands
classified as crop lands and some vegetation areas classified as urban areas. The
classification image of PCA combination (PC1, PC2 and PC3) gives a good
representation of the land features of study area, and pure result in the
classification of band ratios combination.
An advantage of unsupervised classification is that no prior knowledge of
the scene is required, so the setup time is much shorter which generally requires
a minimum input from the operator.
75
Classification of raw bands Classification of band ratios
Classification of PCA Classification of Tasseled Cap
Figure (4-13): Unsupervised classification of different images
76
4.7.2 Supervised Classification
A supervised classification method has been adopted on the images using
Maximum Likelihood classification. The training samples (region of interests)
have been selected and they are based on land cover and terrain. Five different
land cover and land use classes were selected including water, vegetations,
urban areas, crop lands, and bare lands. The collection of data about land cover
and land use classes of interests is fed to the computer as measurement of each
class to set statistical parameters (min., max., mean and STD).
The selection of training areas was repeated several times to overcome
the cases were some cover material could not be classified accurately and also
to avoid as much as possible the overlapping between the classes.
The classified images gave a good representation of the scene rather than
unsupervised classification images. Some errors are occurred in the
classification, where some of the crop lands are classified as bare lands and
some bare areas are classified as urban areas this done in all the four classified
images. Table (4-9) and figure (4-14) show the results of supervised
classification of different band combinations.
77
Classes Raw bands Band ratios Tasseled Cap PCA
Class name Color No. of pixels
Area (km2)
No. of pixels
Area (km2)
No. of pixels
Area (km2)
No. of pixels
Area (km2)
Bare lands 148665 120.75 188184 152.85 154114 125.17 154927 125.84
Urban areas
151521 123.07 85680 69.59 108486 88.11 109535 88.97
Crop lands
125113 101.62 147096 119.47 161421 131.11 166638 135.35
Water 20029 16.26 28511 23.15 17700 14.37 16714 13.58
Vegetations 531136 43.15 48993 39.79 56743 46.08 50650 41.14
Table (4-9): Results of supervised classification of different band combinations
78
-A-
-B-
Figure (4-14): Supervised classification of
A-raw bands
B-band ratios (continue)
79
-C-
-D-
Figure (4-14): Supervised classification of
C- Tasseled Cap
D- PCA
80
4.7.3 Classification Accuracy Assessment
The classification process is not complete until its accuracy is assessed.
The most common means of expressing classification accuracy is a
classification error matrix (sometimes called confusion matrix or contingency
table). Error matrix compares, on category-by-category basis the relationship
between known reference data (ground truth region of interests) and the
corresponding results of an automated classification.
The overall accuracy is calculated by summing the number of pixels
classified correctly and dividing by the total number of pixels. The pixels
classified correctly are found along the diagonal of the confusion matrix table
which lists the number of pixels that were classified into the correct ground
truth class. The total number of pixels is the sum of all the pixels in all the
ground truth classes. The accuracies of each class are illustrated in tables
(4-10a, b, c, and d). From the results of the classification accuracies, we see that
the accuracies of the classification of the transformed bands including band
ratios, Tasseled Cap transformation, and PC’s combinations are 89.45%,
89.26%, and 89.52% respectively, which are greater than the accuracy of the
raw band of ETM+ of the study scene which is equal 87.18%. This means that
the transformation techniques give a good spectral enhancements and feature
extraction of the raw bands of satellite images and show features that
undistinguishable in a raw images and which represent a good tools for
classification and increasing the accuracy.
81
1 2 3 4 5 Total
Unclassified 0 0 0 0 0 0
Bare lands 2884 101 82 0 1 3068
Urban areas 98 505 32 3 7 645
Crop lands 159 49 507 3 55 773
Water 0 3 0 226 19 248
Vegetations 0 1 2 3 1105 1109
Total 3141 659 621 235 1187 5843
Overall Accuracy = (5227/5843) 89.4575%
1 2 3 4 5 Total
Unclassified 0 0 0 0 0 0
Bare lands 2676 36 42 1 1 2756
Urban areas 243 592 103 7 16 961
Crop lands 222 23 474 0 41 760
Water 0 6 0 224 1 231
Vegetations 0 2 2 3 1128 1135
Total 3141 659 621 235 1187 5843
Overall Accuracy = (5094/5843) 87.1812%
Table (4-10a): Confusion Matrix of raw band image classification
Table (4-10b): Confusion Matrix of band ratios image classification
82
1 2 3 4 5 Total
Unclassified 0 0 0 0 0 0
Bare lands 2723 43 43 0 0 2809
Urban areas 139 594 30 4 5 772
Crop lands 279 11 547 3 53 893
Water 0 10 0 224 1 235
Vegetations 0 1 1 4 1128 1134
Total 3141 659 621 235 1187 5843
Overall Accuracy = (5216/5843) 89.2692%
1 2 3 4 5 Total
Unclassified 0 0 0 0 0 0
Bare lands 2720 34 37 0 0 2791
Urban areas 133 606 24 2 4 769
Crop lands 288 11 557 3 60 919
Water 0 8 0 226 1 235
Vegetations 0 0 3 4 1122 1129
Total 3141 659 621 235 1187 5843
Overall Accuracy = (5231/5843) 89.5259%
Table (4-10c): Confusion Matrix of Tasseled Cap bands image classification
Table (4-10d): Confusion Matrix of PCA bands image classification
CHAPTER FIVE
CONCLUSIONS & RECOMMENDATIONS
83
5.1 Conclusions
1. Band transformation techniques generated new sets of image components or
bands and made evident features not discernable in the original data, with a
reduced number of the original bands. Band ratios images gives a good
representation for land features, where the best bands ratios of study area are
5/7, 4/5, 2/5, 1/4, and 2/4 for water, vegetations, bare lands, urban areas, and
crop lands respectively.
2. Optimum Index Factor (OIF) represent optimum method in selecting the best
band combination for more information extracted for the study area, and
adopted for raw bands and bands ratios where:
a. 3-4-7 is the raw band combination that has the highest value of OIF
among 20 different raw band combinations.
b. 4/5-1/4-2/5 is the bands ratio combination that has the highest value of
OIF among 10 different bands ratio combinations.
3. Spectral indices such as NDVI, BSI, and NDWI gives good for monitoring
some land features. While UI gives a poor result for the urban area.
4. The results show that the Principle Components and Tasseled Cap
transformations represented good tools for image enhancement, feature
extraction and reducing the number of bands with no loses in information.
5. Digital image classification of the scene has been applied on:
a. raw band combination (3-4-7)
b. ratios band combination (4/5-1/4-2/5)
c. indices combination of Tasseled Cap (Brightness-Greenness-Wetness)
84
d. PCA combination (PC1-PC2-PC3).
Where the result of classification show that:
a. Unsupervised classification using ISODATA method is a useful technique, to
prepare a primitive map for the scene. The results of this classification showed
good representation of some classes and merges among the other.
b. Supervised classification using Maximum Likelihood method gives good
representation of the classes with overall accuracies 87.18%, 89.45%, 89.26%,
and 89.52% for raw bands, bands ratio, Tasseled Cap, and PCA images
respectively. This indicated that band transformation techniques increase the
accuracy of the classification and extracted the features that are
undistinguishable in raw bands.
5.2 Recommendations
1. Using multi-temporal images for the same scene with the same
transformation techniques to study the changes in the study area.
2. Using spectral indices such as NDVI, BSI, and NDWI …etc, with the
original bands or transformed bands to generate FCC images, and show how the
classification results appeared.
3. Using Hybrid method for combining between bands and for choosing the
type of image for the classification process, where some of topographic detail is
discriminate between certain features by using this method.
REFERENCES
85
References
1. Elachi, Charles (1987), “Introduction to the Physics and Techniques of Remote Sensing”, John Wiely and Sons, New York.
2. AL-Sepahe, A. H. (2002), “Classification of Digital Satellite Images Using Selective Bands and Principle Component Analysis”, M.Sc. Thesis, Al-Nahrain University, Iraq.
3. Avery, T. E. and Berlin, G. L. (1992), “Fundamentals of Remote Sensing and Air photo Interpretation”, 5th ed., Prentice Hall, New Jersey.
4. Fu, Z., Caelli, T., Liu, N. and Robles-Kelly (2007), “Boosted Band Ratio Feature Selection for Hyperspectyral Image Classification”,
http://users.rsise.anu.edu.au/~arobkell/papers/icpr06.pdf
5. Gonzalez, R. C. and Poul, W. R. (1987), “Digital Image Processing”, 2nd ed., Addison-Wesley Publishing Company, Inc.
6. Huang, C., Wylie, B., Yang, L., Homer, C. and Zylstra, G., “Derivation of a Tasseled Cap Transformation Based on Landsat 7 at Satellite Reflectance”,
USGS EROS Data Center http://landcover.usgs.gov/pdf/tasseled.pdf
7. Al-Ani, L. A. (1996), “Classification of Digital Satellite Images”,
Ph.D. Thesis, Al-Nahrain University, Iraq.
8. Frankovich, J. S. (1999), “Unsupervised Classification of Spectrally
Enhanced Landsat TM Data of Midland, MI”,
http://www.earthsensing.com/drs/stud-proj-f99/final.html
9. Abdeen, M. M., Thurmond, A., Abdelsalam, M. and Stern, R. (2002),
“Application of Aster Band-Ratio Images for Geological Mapping in Arid Regions: the Neoproterozoic Allaqi Suture, Egypt”, Remote Sensing (Syria), 14, 19-38.
86
10. Inzana, J., Kusky, T., Higgs, G. and Tucker, R. (2003), “Supervised Classification of Landsat TM band ratio images and Landsat TM band ratio image with radar for geological interpretations of cenral Madagascar”, Journal of African Earth Sciences, 37, 59-72.
11. Feeley, K., Gillespie, T. and Terborgh, J. (2005), “The Utility of Spectral Indices from Landsat ETM+ for Measuring the Structure and Composition of Tropical Dry Forests”, BIOTROPICA, 37, 508-519.
12. Thapa, R., Borne, F., Cu, P. and Porphyre, V. (2005), “Environmental change analysis using satellite imageries: case study of Thai Binh province, Vietnam”,
MapAsia 2005, http://pigtrop.cirad.frlen/worldwide/amerique.htm
13. Fadhil, A. M. (2006), “Environmental Change Monitoring by Geoinformation Technology for Baghdad and its Neighboring Areas”,
http://www.GISdevelopment.net/Application/Environment.pdf
14. Buhe, A., Tsuchiya, K., Kaneko, M., Ohtaishi, N. and Halik, M. (2007), “Land cover of oases and forest in XinJiang, China retrieved from ASTER data”, Advances in Space Research, 39(1), 39-45.
15. Free enecyclopedia (Kut), http://en.wikipedia.org/wiki/Kut
16. Gibson, P. J. and Power, C. H. (2000), “Introductory Remote Sensing: Digital Image Processing and Applications”, Routledge, Taylor and Francis Group, London.
17. Sabins, F. F. (1978), “Remote Sensing Principles and Interpretation”, 1st ed., Remote Sensing Enterprise, Inc., New York.
18. Jensen, J. R. (2000), “Remote Sensing of the Environment An Earth Resource Perspective”, Prentice Hall, New Jersey.
19. Campbell, J. B. (1996), “Introduction to Remote Sensing”, 2nd ed., Taylor and Francis, New York
87
20. Lillesand, T. M. and Kiefer, R. W. (2000), “Remote Sensing and Image Interpretation”, 4th ed., John Wiley and Sons, New York.
21. Abdul Latiff, R. K. (2005), “Change Detection of Environmental system using satellite images in Shatt AL-Arab region”, M.SC. Thesis, University of Technology, Iraq.
22. Bakker, W. H., Gieske, A. S., and Gort, B. G. (2000), “Principles of
Remote Sensing”, Longman, London.
23. Landsat History (2006), Landsat Newsletter, USGS 2006.
24. LANDSAT 7 +ETM Satellite Imagery,
http://www.satimagecrop.com/satellite-sensors.html
25. Landsat 7, http://geo.arc.nasa.gov/sge/landsat/17.html
26. Description of Landsat data,
http://www.eduspace.esa.int/eduspace/main.asp
27. Winarso, G., Judijanto, and Budhiman, S. (2001), “The Potential
Application Remote Sensing Data for Coastal Study”, Paper presented at
22nd Asian Conference on Remote Sensing, 5-9 Nov. 2001, Singapore.
28. Remote Sensing and Image Analysis,
http://www.cnr.berkeley.edu/~gong/textbook/chapter6/html/sect64.htm
29. Jensen, J. R. (1986), “Introductory Digital Image Processing, A Remote
Sensing Perspective”, Prentice Hall, Englewood Cliffs, New Jersey.
30. Yousif, B. F. (2004), “The Use of Remote Sensing Techniques in the
Classification of Al-Najaf Soil”, M.SC., Thesis, University of Technology,
Iraq.
31. Niblak, W. (1986), “An Introduction to Digital Image Processing”,
Prentice Hall International.
88
32. San, B. T., Sumer, E. O., and Gurcay, B. (2001), “Comparison of
Band Ratioing and Spectral Indices Methods for Detecting Alunite and
Kaolinite Minerals Using ASTER Data in Biga Region, Turkey”, General
Directorate of Mineral Research and Exploration, RS Center.
33. Mohammad, M., El-Sobky, H., Sedeik, K., and El-Raey, M. (2001),
“Application of Band Ratios Identified by HHRR for Recognition of Surface
Units Using TM Data in SIWA Depression, Western Desert, Egypt”,
Paper presented at 22nd Asian Conference on Remote Sensing, 5-9 Nov.
2001, Singapore.
34. Hultgren, T. L. (2005), “The Role of Resolution in Dasymetric
Population Mapping”, M.A., Thesis, University of Colorado.
35. Paris, J. F. (2005), “Generalized Mapping of Biophysical Properties”,
http://www.microimages.com/freestuf/ScriptsByJack.htm
36. Eastman, J. R. (2001), “Idrisi32 Release 2 Tutorial”, Clark Labs, Clark
University, USA.
37. Ali, A. H. (1998), “Studying the Flooded Area by Principle Component
Analysis of Multi-Temporal Landsat Thematic Mapper Data”, M.Sc.,
Thesis, Al-Nahrain University, Iraq.
38. Smith, W. L. (1977), “Remote Sensing Applications for Mineral Exploration”, Dowden, Hutchinson & Ross. 39. Swain, P. H., and Davis, S. M. (1978), “Remote Sensing: The
Quantitative Approach”, McGrew-Hall, Inc., New York.
40. Schowengerdt, R. A. (1983), “Techniques for Image Processing and
Classification in Remote Sensing”, Academic Press, Inc., New York.
41. Richards, J. A. & Jia, X. (1999), “Remote Sensing Digital Image
Analysis”, Springer-Verlag, New York.
89
42. Sania, S. A. (2000), “Image Segmentation using Genetic Algorithm”
M.Sc., Thesis, Al-Nahrain University, Iraq.
43. Green, W. B. (1989), “Digital Image Processing: A System Approach”
Van Nostrand Reinhold, New York.
44. Janza, F. J. (1983), “Manual of Remote Sensing”, American Society of
Photogrammetry, Vol. 1, Virginia.
APPENDIX
90
Appendix
ENVI (the Environment for Visualizing Images) is a revolutionary
image processing system. From its inception, ENVI was designed to address
the numerous and specific needs of those who regularly use satellite and
aircraft remote sensing data. ENVI provides comprehensive data
visualization and analysis for images of any size and any type all from
within an innovative and user-friendly environment.
ENVI includes all the basic image processing functions within a
friendly, interactive point-and-click graphical user interface. The software
includes essential tools required for image processing across multiple
disciplines, and has the flexibility to allow the user to implement his or her
own analysis strategies. ENVI simplifies comprehensive interactive
processing of large multiband data sets, screen-sized images, spectral plots
and libraries, and image regions-of-interest, all while providing flexible
display capabilities and geographic-based image browsing.
ةـالصـالخ
تحمل معلومات آبيرة عن المناطق قمار الصناعية المتعددة االطياف،صور االولغرض استخالص اآثر آمية ممكنة من المعلومات من تلك الصور، . المراد دراستها
نسب الحزم الطيفية، : استخدمت العديد من تقنيات المعالجة الرقمية، ومن هذه الطرق Principle Component) تحليل المرآبات االساسية، و Tasseled Capتحويل
Analysis) .و تقليل عدد استخالص المعلومات، ان هذه التقنيات تلعب دور مهم في .الحزم الطيفية المستخدمة بدون خسارة في تلك المعلومات
البحث يهدف الى تطبيق عملية التصنيف على نسب الحزم وعلى المعامالت لصور ،Tasseled Capيل المرآبات االساسية و تحويل الصورية المتضمنة تحل
ولبيان آم ،(+ETM)القمر الصناعي من نوع راسم الخرائط الغرضي المحسن .تختلف دقات هذه التصانيف عن دقة تصنيف الحزم االصلية
ان افضل نسب حزم تم اختيارها لتمثيل خمسة خصائص ارضية النتائج بينت تم تطبيق طريقة الحد االقصى للعامل الكشفي . ٤\٢و ،٤\١ ،٥\٢ ،٥\٤ ،٧\٥هي
(Optimum Index Factor) نسب على مجاميع الصور االصلية و على مجاميع الحيث ،الختيار افضل مجموعة والتي تحتوي على معلومات آبيرة للمنطقة الحزمية،
وافضل مجموعة للنسب ،٧-٤-٣بينت النتائج ان افضل مجموعة للحزم االصلية هي .٥\٢-٤\١- ٥\٤هي
تم اختيارها لتمثيل التصنيف الغير ISODATAطريقة Maximum)وتم اختيار طريقة االحتمالية العظمى ،(unsupervised)موجه
Likelihood) لتمثيل التصنيف الموجه(supervised).
و ،٪٨٩٫٢٦ ،٪٨٩٫٤٥ ،٪٨٧٫١٨ :الدقات الكلية للتصانيف آانت آاالتينتائج و حزم المرآبات ،Tasseled Capحزم للحزم االصلية، نسب الحزم، ٪٨٩٫٥٢ التحويل اعطت تحسينا طيفيا جيدا، ان تقنيات النتائج حيث بينت ،على التوالي االساسية
وتعتبر ن يصعب تمييزها في الصور االصلية،استخالص معلومات آا وعملت على .ادوات جيدة لزيادة دقة التصنيف
—]Ü√÷] ÌËÖÁ„∂
Í€◊√÷] ovf÷]Ê Í÷^√÷] ‹È◊√j÷] ÏÖ]áÊ
‡ËÜ„fl÷] Ì√⁄^q
›Á◊√÷] ÌÈ◊“
›à�£] l¯�⁄^√⁄Ê gäfi 4m`i ÏÅÇ√jπ] ÌȬ^flí÷] Ö^€Œ˘] ÖÁë ÃÈflíi ̌ŠÓ◊¬ Õ^Èõ˘]
Ì÷^âÖ
ÌqÖÅ Ó◊¬ ŸÁí£] l^f◊�j⁄ ‡⁄ àr“ ‡ËÜ„fl÷] Ì√⁄^q ›Á◊√÷] ÌÈ◊“ ±] ‰⁄Çœ⁄
^ËàÈÀ÷] ª ›Á◊¬ 4jäq^⁄
ÿfŒ ‡⁄
∞fl{ç ÿ⁄^“ Ã{È{{â DNLLQ åÁËÖÁ÷^”eE Õ]Üç`e
Ífi^√÷] àËà√÷] Çf¬ Å^Ë_ ÖÁj“Ç÷] É^jâ˜_
0^ë ÇÈ€£] Çf¬ |¯ë ÖÁj“Ç÷] Ǭ^äπ] É^jâ˜_
NLLT Ö]É] MPNUÜ�Àë