CHAPTER - III DIGITAL CLASSIFICATION OF …shodhganga.inflibnet.ac.in/bitstream/10603/30266/4/chapter 3.pdfCHAPTER - III DIGITAL CLASSIFICATION OF SATELLITE DATA ... (KNPC) have worked

CHAPTER - III

DIGITAL CLASSIFICATION OF SATELLITE DATA

AND OIL POLLUTION

3.1 GENERAL

Remote Sensing has great potential in the change detection of land cover.

Land covers can be identified and differentiated from each other by their unique

spectral responses patterns. In land cover changes, classifications are important

elements (Dickinson, 1995; Hall et al, 1995). Several studies demonstrated the

effectiveness of using satellite-derived data in producing land cover maps as well as

detecting the landscape changes over time (Sabins, 1997; El-Magd and Tanton, 2003;

Cardille and Foley, 2003; Lobo et al., 2004; Musaoglu et. al., 2005).

The multi-spectral remote sensing instruments measure the amount of energy

reflected and emitted in different bands of the electromagnetic spectrum of the various

features of the earth such as water body, wet land, settlement, vegetation, polluted

area, desert soil etc. The electromagnetic spectrum has 255 gray colors which is

reflected from the surface and stored in a remote sensing instrument (Thematic

Mapper). In this case, the spectral pattern of each pixel was used as the numerical

basis for the categorization of oil pollution and land cover areas.

Land cover oil pollution classification was mainly based on the oil spread in

the study area. The change detection in land cover classification can be done by the

reflection and emitting of the electromagnetic energy in different periods. Hence

multi-spectral Landsat images can be used for land cover classification.

3.2 METHODOLOGY

Image classification procedures automatically categorize all the pixels in an

image into land cover classes. TM multi-spectral 30 meter resolution data was used to

CChhaapptteerr –– II II II DDiiggii ttaall CCllaassssii ffiiccaattiioonn ooff SSaatteell ll ii ttee DDaattaa…………

Space and Spatial Technologies for Oil Pollution Evaluation in Kuwait

54

perform oil pollution classification in this study. Oil lakes, tarmats, contaminated soil

and polluted soil were the four main possible land cover oil polluted classes. Sand sheet

was added as the last class because the polluted zones were later covered by sand.

In 1993 satellite data, the oil polluted areas and also the boundaries between

different classes of pollution are vividly seen; hence, the 1993 data is taken for

classification. Mid-infrared band 7 has long wave length and it is highly useful for

pollution hazards. The stretched band7 was used for single band cluster analysis. Both

fine and broad models have been tried separately; the resulted image gave more

classes.

Similarly, cluster analysis was also performed with color composite image

bands 3, 4 & 5 for the same fine and broad classification. The resulted images were

reduced into three and five, respectively. Finally, the isocluster model was adopted

using same color composite image and all 6 bands (except thermal band) were

processed and the resulted image had 3 classes which are more general classes.

None of the unsupervised classifiers distinguished oil polluted classes;

supervised classifier was used to classify oil polluted land cover class. Four training

areas were chosen with enough pixels in each training areas in and around Burgan

area and the isoclust and minimum distance classifier was used. These two methods

had failed to indicating unknown pixels as sand sheet class.

Stretched 1993 band 7 image was again stretched using density slice method.

Here, it was seperated into 5 classes. The resulted 5 classes were similar to aerial

photo classes. But the first oil lake class had more number of pixel values than the

other classes because the DN values of oil lakes and sea water remained same. So, to

get exact classification, the same processes were used to subset the Burgan oil field.

The resulted classified image was checked with aerial photo of the same year of

Burgan area.

3.3 POSSIBLE LAND COVER OIL POLLUTION CLASSES

Ever since oil wells started burning from the Kuwait oil fields, dynamic

changes of oil polluted land cover classes have become uncontrolled from 1991 to

2001.



55

Three major types of oil pollution classes are observed in this region.

I. Pollution Due to Fire and Smoke

1. Oil fire and smoke

II. Pollution Due to Oil Flow

1. Oil Lakes & Oil River

2. Tarmats

III. Pollution of Soils Due to Oil

1. Contaminated Soil

2. Polluted Soil

3. Desert Soil

The first type of oil pollution class was mainly due to smoke plume formed

during oil fired. The land covered areas of fire and smoke which come this class.

The second type of pollution is due to gushing oil flow on the surface. In this

type again, if the oil was accumulated in low lying areas they became oil lakes or if

dried up became tarmats.

The third type of pollution was oil mixed in soil. It was again divided in to

three classes based on propotion of oil in the soil. Contaminated soil has oil sprayed

on soil. Polluted soil was directly derived from weathering of tarmats or by leaching

of contaminated soil or due to less concentration of oil in the soil.

The last class was desert soil which was nothing but the original desert sand.

If the polluted soil became desert soil it was the last stage of pollution change

detection in this series.

3.3.1 Pollution Due to Fire and Smoke

Landsat TM image of 15th February 1991 was the first image showing the oil

well fire with smoke on Burgan oil field. 25th February image showed that all

Kuwait’s oil field wells were set ablaze. Crude oil is nothing but a highly flammable

hydrocarbon which has enough static pressure to flow on the surface. The total

Kuwait oil fields were burnt over 7 months in this region.

Table 1-1 indicates 605 oil wells were fired in March 1991 and dark smoke

covered over the Kuwait and surrounding areas. Kuwait had been completely covered



56

by thick dark smoke so that sun light could not penetrate to ground surface for few

months. The thick dark smoke, literally called as wild smoke, spread many

kilometers along the wind direction around the Kuwait area. These oil fires and wild

smoke covered areas on Kuwait land were considered as class of “Fire and Smoke”

but the fire and smoke had been reduced within short period.

3.3.2 Pollution Due to Oil Flow

Kuwait oil resource aquifers have enough static pressure to flow oil on to the

surface. As per the damaged oil well statistical report, there were 46 wells gushed

and 108 wells damaged in March 1991 (Table1-1). These wells have released

enormous oil to flow on the surface. Fire fighting rescue teams did their excellent job

to control the oil well flow and fire. The Kuwait Oil Company (KOC) and Kuwait

National Petroleum Corporation (KNPC) have worked well in reducing oil fire by

stopping the oil wells flew.

The non burnt oil flew as stream on the hot siliceous desert and accumulated

in low lying areas as “Oil Lakes”. There were 300 small and big oil lakes found

(Kwarteng, 1998) in this region. The rescue teams recovered most of the oil from

these oil lakes. A big oil flow on the southern part of Burgan oil field was called as

“Oil River”. The dried oil became hard black oil layer called as “Tarmats”.

3.3.3 Pollution of Soils Due to Oil

The soils were polluted due to the flow of oil over them and based on the ratio

of oil mixed in soil. This soil pollution was again classified into two classes based on

the percentage of oil in the soil, namely contaminated soil and polluted soil. The

contaminated soil might form due to tarmats disintegrating into pieces resulting black

soil. Over oil lakes/tarmat and due to wind action fresh sand might get overcastted

thereby leading to reduction of gray values indicating contaminated soil. Sometimes,

instead of oil flow, oil sprayed over the fresh sand which happened even few hundred

meters away from the gushing well leading to less admixture of oil to the soil and the

same is also called as contaminated soil. Accordingly, it is evident that contaminated

soil might be formed due to many processes.



57

Consequently, if the mixture contained more oil in soil, it is formed as

contaminated soil or if oil was less, it is called as polluted soil. Polluted soil was

formed by either light spray of oil on the soil or due to transportation of weathered

loose particles of tarmats leaving subtle residues in the soil itself. The local sand

encroachment natural hazard is one of the main sources of these classes. The thick

sand deposited over the oil lakes and tarmats or the thin layer even over the

contaminated soil become polluted soil. Total polluted areas were covered by the

fresh sand and the black tone became white tone in the image that is sand sheets. It is

nothing but normal desert of the study area.

Finally, there were six possible dynamic land cover oil polluted classes.

1. Fire and Smoke 2. Oil lakes & oil river 3. Tarmats 4. Contaminated soil 5. Polluted Soil 6. Desert Soil

3.4 IMAGE CLASSIFICATION TECHNIQUES

A set of multi-temporal TM data have been acquired for a detailed investigation

of land cover oil pollution changes of Kuwait region. The pre-war TM image of 1990

represented non oil pollution at that date. Five temporal TM images of 1991 were used

for temporal changes of dark smoke over Kuwait area. Fire and smoke could be

identified easily by its red & dark color in the image. Tracing the dynamic changes of

smoke was clear on hard copy of the images. But oil lakes and tarmats needed image

processing to distinguish the boundaries between them because tarmats were nothing

but the dried oil. Oil pollution land-cover classifications process was approached with

medium (30m) spatial resolution remotely sensed data (Franklin and Wulder, 2002).

Six multi-temporal, multi-spectral TM and ETM+ data (1990, 1993, 1995, 1997,

1999 and 2001) satellite images have been used to distinguish temporal land cover

changes. Classification algorithms were used to map land cover pollution (Tso and

Mather, 2001; Landgrebe, 2003). The main goal of the image classification process was

to automatically classify all pixels in the image into one of a certain number of land



58

cover types, based on their brightness values. Unsupervised and supervised

classifications were the main two types of techniques used here.

3.4.1 Unsupervised Classification

In the unsupervised classification the dominant spectral response patterns

occurring within an image were extracted and desired information classes were

identified from known area (Goldberg and Shlien, 1976, 1978; Justice and

Townshend, 1982; Palylyk and Crown, 1984; and Khotanzad and Bouarfa, 1990). It

involved algorithms examining the unknown pixels in an image and aggregating them

into a number of classes based on the natural grouping or clusters present in the image

values. They were mainly spectral classes because they were based solely on the

natural groupings in the image DN values.

The acquired image had mainly Arabian Gulf and desert. In the land part,

desert covered most of the places followed by settlement, wet land (near the coastal

area), vegetation and oil polluted areas. The desert sand covered most part of the

image followed by city area developments along with vegetation.

3.4.1.1 Single Band Cluster Analysis

To find oil pollution land cover classes, first; an individual spectral band

cluster model was adopted. The spectral reflection of each band was varying with

reflection areas and its wave length. Band 7 of 1993 a winter season data was chosen

for single band cluster analysis. Band 7 has long wave length and usually used for

geological and pollution propose.

Cluster model using histogram peak selection techniques is defined as a value

taken as frequency. It can classify according to number of peaks (crest) in histogram

is taken as one class if broad classification is chosen. If it is fine classification, values

from each crest to trough. Idrisi-32 Release2 remote sensing software was used for

this classification processes. The generalization level was chosen for both broad and

fine classes; and “Drop least significant clusters” was used as clustering rule. But fine

mode gave 23 clusters whereas in broad mode 13 cluster (Figs. 3.1A, B).



59

Fig. 3.1 Unsupervised Fine (A) and Broad Cluster Classification (B) of TM Band7 1993

3.4.1.2 8Bit Color Composite Image Cluster Classification

Clustering land cover classification using a color composite image was better

than single band image classification because sample pixel had combination of three

color values rather than single value. 1993 TM bands 3, 4 & 5 were used for color

composite image. The band 3 was exposed under blue filter; the band 4 was under

green and band 5 under red filter. 8 bit pixel values color composite has gray pixel

values (0 to 255) (fig.3.2A).

Isoclust analysis classification mainly used spectral values of color composite

image. Bands 3, 4 & 5 color composite 8 bit image (fig. 3.3A) was used with all 6

multi-spectral TM bands of 1993 data. The spectral value of each pixel of color

composite image values was taken as sample values and classifying all 6 bands;

resulting values were given as an output image. Figure 3.3B is the isoclust 6 bands of

1993 image classification showing 3 classes which were almost same as previous

classification.

Fine Classes Broad Classes

B A



60

Fig. 3.2 Unsupervised 8 Bit Color Composite Cluster Classification of Band 3, 4 & 5(A)

TM Image (1993), [Fine Class (B), Broad Class (C) & Burgan Area(1)]

Both fine and broad options were adopted with “drop least significant clusters”

as its clustering rule. The fine cluster model image gave only 3 classes (Fig. 3.2B) in

green, blue and yellow colors. In broad mode, five classes were seen (Fig. 3.2C). It

shows more oil lakes in Burgan oil field (1), and other oil polluted classes could not

be disproved.

Fig. 3.3 8 Bit TM Color Composite of 1993 (A); All Bands Isoclust (B)

and Burgan (1)

A B C

A B

1

1

1 1

1



61

3.4.1.3 Isoclust Using 8 Bit Color Composite

The green color showed sea water (background color of this image). Most of

land side was covered by yellow color of low value and middle value was blue

covering Kuwait city and polluted oil field. There was no distinct color for oil

pollution land cover classification. The blue colored Burgan oil field (1, Fig. 3.3B)

was surrounded by desert sand in yellow color. Once again this method also did not

give proper polluted classes. So, supervised classification was attempted next.

3.4.2 Supervised Classification

Supervised Classification is widely used to identify the land cover classes

(Reeves et. al., 1975; Dymond, 1993). Land cover may be identified and

differentiated from each other by their unique spectral response patterns. Sample

pixels in each known cover type were selected as training sets. Location of each

training sets in the two-dimensional matrix of brightness values was entered into the

computer where a group of training set statistics was computed for each spectral band

and a covariance matrix for the data for all spectral bands. Taken together, these

statistics represent “spectral signatures” for each of the cover types to be identified.

From the spectral signature of known categories, the software assigns each pixel in

the image to the cover type to which its signature was most similar.

There are four basic steps involved in a typical supervised classification

procedure.

1. Local representation examples of each cover type that can be identified

in the image -Training sites.

2. Digitizing polygons around each training site, assigning a unique

identifier to each cover type.

3. Analyse the pixels within the training sites and create spectral

signatures for each of the cover types.

4. Classification of images.

Each pixel in the image data set was categorized into the land cover class

most closely resembling. If the pixel was insufficiently similar to any training data

set, it was usually labeled “unknown”. The category label assigned to each pixel in



62

this process was then recorded in the corresponding cell of an interpreted data set. The

mean or average spectral values in each band for each category were determined.

These values comprise the mean vector for each category.

For land cover change classification, the supervised multi-date classification

determines the direct transition of pixels from one class to another by the use of a

trained classify model; this method shows the time difference correlation between the

two images used in the analysis. The disadvantage of this method is the training pixels

used must be the same points on the ground in the two different dates.

3.4.2.1 Training Sets

The fields chosen as training areas for classifications are:

1. Oil lakes

2. Tarmats

3. Contaminated Soil

4. Polluted Soil

Training samples were collected from satellite image. Different collection

strategies, such as single pixel or polygon were used, but they would influence classification

results, especially for classification with fine spatial resolution image data. When the

landscape of a study area is complex and heterogeneous, selecting sufficient training

samples becomes difficult. This problem would be complicated if medium or coarse spatial

resolution data are used for classification, because a large volume of mixed pixels may

occur.

The TM satellite image of 1993 band 7 was chosen for homogeneous training

set. The image was zoomed as much as needed. A sufficient number of training

sample were used for image classification. The known set was chosen and digitized as

vector polygon of above sequence. Each training set was taken containing more than

70 pixels. There were many polygons in one class; some were two to three polygons

only. Digitized polygons of 4 training classes were saved with respective name (Fig.

3.4A).



63

4 training set Isoclust Minimum Distance

Fig. 3.4 Supervised Isoclust and Minimum Distance Classified TM 1993 Image

Burgan (1) and Ahamadi (2)

To create signature files, the digitized vector file from training set was taken to

further processing. The signature file name of all polluted class list was entered

sequentially. A separate signature file was created for each class in the training set

vector file. Makesig model was run to make a separate signature file of each class.

The sequence of training set classes was entered and TM satellite image of the

respective time of all 6 bands were chosen to run this mode. It produced 4 individual

signature files with respective to 6 bands module. Makesig automatically created a

signature group file that contains all signature filenames.

3.4.2.2 Isoclust Using Training Signature Files

Isoclust using training set model consisted of an image classification using

training signature files. The training classes of 4 signature files were used in this

model to classify oil polluted areas. The signature files of four training site file were

used as know signature data and all 6 bands were chosen as input data. The resultant

image was 6 bands checked with signature of training files. The resulting (Fig. 3.4B)

Isoclust image had 4 classes. But these four classes did not clearly define the polluted

classes.

A B C

1

2



64

3.4.2.3 Minimum Distance Normalized Classification

The minimum distance to mean classification was used to calculate the distance

of a pixel’s reflectance to the spectral mean of each signature file and then assigning the

pixel to the category with the closest mean. In this case, the classifier was evaluated with

standard deviations of reflectance values about the mean creating contours of standard

deviations. It then assigned a given pixel to the closest category in terms of standard

deviation.

The signature files of all four polluted classes were to be taken as input

signature files. Each pixel in the study area had values in each of the seven bands of

the imagery. These unique signature values could be compared to each of the

signature files. Pixel was then assigned to the land cover type. In this image resulted

(Fig. 3.4C) the pollution classes could not be deciphered.

3.4.3 Histogram Based Classification

None of the above classification was well suited for land cover oil pollution in

this area. Subset Burgan area and their individual band of 6 years histogram were

studied in detail. Their nature of peaks and division were noticed. Almost all

histograms had 4 peaks indicating 5 types of classes (Figs. 3.5B). Khotanzad and

Bouarfa (1990) and Kartikeyan et al., (1998) have implemented the clustering in the

histogram textural features for texture dominant images.

The histogram-based clustering method has only one underlying assumption

that the N-dimensional probability density function (pdf) of pixels of each region is

unimodal in nature. No assumption is made on the specific form of the pdf. The

independence of the form of the pdf is especially useful in dealing with remotely-

sensed images of the Earth’s surface where often the pdf of different land covers are

different.

3.4.3.1 Histogram Based Linear Stretched Classification

This classification can be done by simple linear stretching. Stretching is an

image processing usually done in 255 gray scales. Figure3.5B shows the 1993 band7

stretched graphs. The DN values between 0 to10 remained are almost flat in

horizontal axis except there was a small peak at 200 values. 255 gray scale values

were arranged according to the histogram distribution values and assigned into 5



65

Fig. 3.5 Histogram Based Linear Stretched Classification of Band 7 1993 (A)

Histogram (B) and Five Classes of Histogram(C)

According to the DN values, there are 5 categories in Cluster Analysis.

1. Dark Blue - Sea, water body and Oil Lakes, 2. Pink tone - wetland and

Tarmats. 3. Green tone – a mixture of Settlement, thick vegetation and

Contaminated Soil. 4. Yellow tone –Light Vegetation & polluted Soil and

5. Light yellow tone – Desert Soil. B. Non classified Histogram and

C. Classified Histogram.

classes. Figure 3.5C shows the five classes of band7 1993. These 5 classes are well

suited for land cover five classes in oil polluted region (Fig. 3.5A). But the number of

pixels in oil lake class was more than other 4 classes. It was inferred that sea water

had the same DN values.

To classify by this method, it needed an exact area that is a detailed subset of

Burgan oil field (Figs. 3.6A, B & C). The range of DN values in oil lakes class including,

water body and oil lakes were dark blue in color. They were different in sizes from big

(above 1km2) to small (less than one pixel (30X30m2).

Tarmats including wet soils and settlements were in pink color. The central part of

Burgan oil field was mostly covered with pink color indicating that the tarmat was more

in this time. Third class was a mixture of tarmats and polluted soil along with thick oily

B

C

A



66

Fig. 3.6 Five Classes of Burgan Oil Field of 1993 TM Image (A); Its Histogram (B) and Five Classes Histogram (C)

vegetations called as contaminated soil in green color. Contaminated soil intercalates with

tarmats indicating that tarmats became contaminated soil. The fourth class, polluted

soil with disturbed soil, was also in dark yellow color. The final, light yellow color is

almost desert sand that predominates in this region (Fig. 3.6A).

3.4.3.2 Cross Checking Using Aerial Photographs

Kuwait oil fields were highly damaged during Gulf War. The main economic

source of these fields was highly restricted to allow mapping. The accuracy of this study

was mainly based on aerial photos. Acquired in 1993 the southern part of Burgan oil

field’s colored aerial photo with scale 1: 60,000 was used for cross checking of oil

pollution land cover classes. It was bounded between latitude 280.51’N & 280.54’N and

longitude 470.58’E & 480.02’E of southwest of Burgan oil field.

Figure3.7A shows clearly 4 classes of oil pollution in colour aerial photo.

1. Oil lake – dark black tone

2. Tarmats – bluish dark

3. Contaminated soil – grey

4. Polluted soil – light grey

C

B A



67

For the same location, satellite image of 1993 band 7 TM data is given in

Figure 3.7B. This stretched image showed same pattern of oil pollution classes. Some

small oil lakes were not visible because of scale difference. The tarmats in this class

were same as aerial photos but some high DN values of contaminated soil were also

included. Contaminated soil was in gray tone. The polluted soil Figure 3.7B image is

in light grey tone around oil field.

48°2'E

48°2'E

48°0'E

48°0'E

47°58'E

47°58'E

28°52'N

28°52'N

28°50'N

28°50'N

47°58'0"E

47°58'0"E

28°54'0"N

28°54'0"N

28°52'0"N

28°52'0"N

Fig. 3.7 Shows Cross Checking of Oil Pollution Classification Using 1993 Arial Data

A- Aerial color photo of Southern part of Burgan oil field 1993.

B- TM Landsat image of Burgan oil field 1993.

C- Color image show five classes in the same location.

1. Oil lakes, 2. Tarmats, 3. Contaminated Soil and 4. Polluted Soil

In TM band 7 of 1993 images five colors are assigned to 5 classes (Fig. 3.7C).

Dark blue, pink, green, yellow and purple colors were related to oil lakes, tarmats,

contaminated soil, polluted soil and desert soil, respectively.

A

C B

3

1 1

4

3

2

2

1

2

2

3

4

2

2

3

4

3



68

3.5 SYNTHESIS

Subsequent to the digital image processing (Chapter-II), digital classification

of satellite data was carried out for oil pollution mapping. There were possibly six

classes of oil pollution:

i. Fire & Smoke

ii. Oil Lakes

iii. Tarmats

iv. Contaminated Soil

v. Polluted Soil

vi. Desert Soil

But only the classes ii to v could be precisely mapped as the date of the image

was after one year from the war; by then all Fire & Smoke had disappeared.

Then, all possible types of classification were performed:

Unsupervised single band cluster

Unsupervised color composite cluster

Isoclust color composite

Supervised classification – Isoclust

Supervised classification – Minimum distance

Histogram based linear stretched classification

_________@@@@@@____________

CHAPTER - III DIGITAL CLASSIFICATION OF …shodhganga.inflibnet.ac.in/bitstream/10603/30266/4/chapter 3.pdfCHAPTER - III DIGITAL CLASSIFICATION OF SATELLITE DATA ... (KNPC) have worked

Documents