-
An easy to use ArcMap based texture analysis program for
extractionof ooded areas from TerraSAR-X satellite image
Biswajeet Pradhan a,n, Ulrike Hagemann b, Mahyat Shafapour
Tehrany a, Nikolas Prechtel b
a Faculty of Engineering, Department of Civil Engineering,
University Putra Malaysia, Serdang, Malaysiab Institute for
Cartography, Faculty of Forestry, Geo and Hydro-Science, Dresden
University of Technology, 01062 Dresden, Germany
a r t i c l e i n f o
Article history:Received 27 May 2013Received in revised form13
October 2013Accepted 17 October 2013Available online 6 November
2013
Keywords:Texture analysisFeature extractionRemote
sensingTerraSAR-XPixel basedMalaysia
a b s t r a c t
Extraction of the ooded areas from synthetic aperture radar
(SAR) and especially TerraSAR-X data is oneof the most challenging
tasks in the ood management and planning. SAR data due to its high
spatialresolution and its capability of all weather conditions
makes a proper choice for tropical countries.Texture is considered
as an effective factor in distinguishing the classes especially in
SAR imagery whichrecords the backscatters that carry information of
kind, direction, heterogeneity and relationship of thefeatures.
This paper put forward a computer program for texture analysis for
high resolution radar data.Texture analysis program is introduced
and discussed using the gray-level co-occurrence matrix (GLCM).To
demonstrate the ability and correctness of this program, a test
subset of TerraSAR-X imagery fromTerengganu area, Malaysia was
analyzed and pixel-based and object-based classication were
attempted.The thematic maps derived by pixel-based method could not
achieve acceptable visual interpretationand for that reason no
accuracy assessment was performed on them. The overall accuracy
achieved byobject-based method was 83.63% with kappa coefcient of
0.8. Results on image texture classicationshowed that the proposed
program is capable for texture analysis in TerraSAR-X image and the
obtainedtextural analysis resulted in high classication accuracy.
The proposed texture analysis program can beused in many
applications such as land use/cover (LULC) mapping, hazard studies
and many otherapplications.
& 2013 Elsevier Ltd. All rights reserved.
1. Introduction
Texture is considered as an important characteristic in
imageprocessing which is useful in radar remote sensing and other
eldswhere it is necessary to interpret gray value images like in
themedical sector (Treitz et al., 1996; Mahmoud et al., 2011). Due
to thesurface properties such as roughness, humidity and
orientationevery object in SAR scene can have its own unique
backscatteringproperties (Haack and Bechdol, 2000). The recognition
of textureand object classication are the most challenging problems
in theremotely sensed data processing (Zhang et al., 2007;
Hamedianfarand Shafri, 2013). The main aim in image processing is
to convertthe remote sensing (RS) imagery information into tangible
informa-tion which can be understandable and possibly be used in
combi-nation with other data in Geographic Information System
(GIS)environment (Blaschke, 2010). Therefore precision of the data
andthe texture analysis method are two main factors that have
directimpact on the level of accuracy and information that can
beachieved (Chica-Olmo and Abarca-Hernandez, 2000).
SAR images such as TerraSAR-X with high spatial resolution
areoptimal solution for texture analysis from which
meaningfultexture parameters can be deduced (DLR-Deutsches Zentrum
frLuft- und Raumfahrt, 2009; Sousa et al., 2013). TerraSAR-X
satelliteis Germany's rst active space borne remote sensing
satellite whichhas been in orbit since 15 June 2007 (Biro et al.,
2012). SAR sensorscan provide its own illumination source and it
can record dataindependent of day and night time (Gibson, 2000;
Buckreuss, et al.,2006). Another advantage is the ability of this
data to penetrateeven cloud cover, making the image recording
independent of allweather conditions (Christina Herzfeld and
Zahner, 2001; Pradhanet al., 2009; Pradhan and Shae, 2009). The
penetration of radarbeams depends on the wavelength, humidity and
roughness of thesurface (Haack and Bechdol, 2000). SAR has a wide
range ofapplications such as land use and regional/urban
planning(Fugura et al., 2011; Mahmoud et al., 2011), change
detection(Vidal and Moreno, 2011), disaster management (Elbialy et
al.,2013; Hassaballa et al., 2013; Tehrany et al., 2013b; Pradhan
andShae, 2009; Pradhan et al., 2009; Pradhan and Youssef,
2011),snow and glacier monitoring (Floricioiu et al., 2008) and sea
as wellas sea ice and windmonitoring (Ren et al., 2012). Therefore,
there isa need to perform the classication and extract the
environmentalproperties of the ground (DellAcqua and Gamba,
2006).
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/cageo
Computers & Geosciences
0098-3004/$ - see front matter & 2013 Elsevier Ltd. All
rights reserved.http://dx.doi.org/10.1016/j.cageo.2013.10.011
n Corresponding author. Tel.: 60 3 89466383.E-mail addresses:
[email protected], [email protected] (B. Pradhan).
Computers & Geosciences 63 (2014) 3443
-
Extraction of the ooded areas, using SAR data through thetexture
analysis is one of the highly demandable areas of research.Usually
in tropical countries during the rainy season the area ismostly
covered by clouds (Al Fugura et al., 2011; Kia et al.,
2012;Pradhan, 2009; Pradhan and Youssef, 2011; Tehrany et al.,
2013b;Youssef et al., 2011). In this situation optical RS cannot be
used. Astexture analysis plays an important role in the visual
interpretationand recognition, it is necessary to develop efcient
method to extractinformation precisely (Haack and Bechdol, 2000).
In the recentdevelopment of object-based methods for classication
of RS data,texture is one of the most important factors in the
process ofsegmentation (Ouma et al., 2008). In hazard mapping using
radarimagery, the most difcult task is extracting the texture of
the areawhich is not easy to perform. Some methods have been
developedbut still there exist no user-friendly and simple method
for textureanalysis (Kiema, 2002; Luo et al., 2012).
The current study aimed to produce an easy to use ArcMap9.2
application to solve the problem and complexity of
texturerecognition in ood mapping using TerraSAR-X imagery. Also
theproposed program can be used in other applications such
asagriculture studies, landslide mapping, oil spill monitoring
etc.At the beginning of this article mathematical background
oftexture analysis will be explained. A novel application for
textureanalysis using the gray-level co-occurrence matrix (GLCM)
will beintroduced and discussed as well. Finally, to demonstrate
theability and correctness of this application, a test subset of
aTerraSAR-X image will be analyzed using simple pixel-based
andobject-based classication approaches.
2. Texture analysis: a preview
Tamura et al. (1978) considered texture as a repetitive
patternthat denes small areas. Haralick (1979) described it as
pixels of acertain type and number that have a spatial organization
eitherrandomly or with a dependency of two or more pixels. He
alsosuggested that the term texture should not stand alone as it is
alwaysconnected to the tone. Both require each other in a kind
ofinterrelationship with one always dominating the other. Thus
anarea with little gray value variation is dominated by tone; a
highvariation, however, refers to texture as the dominant
property.Haralick (1979) also stated that it is important to
examine theproperties of single gray value pixels as well as their
spatial relation-ship with each other. Therefore it is necessary to
use rst and secondorder statistics. The rst order describes the
properties of each grayvalue pixel separately, whereas the second
order statistics examinethe relationship of pixels and thus their
organization. Especially thesecond order calculation is important
because texture is a quality ofnot a single pixel but of a whole
area.
In a recent paper, Mahmoud et al. (2011) performed
textureanalysis using TerraSAR-X data in order to enhance the land
use/cover (LULC) classication near Pirna, Saxony, Germany.
Theyapplied both separability and threshold (SEaTH) method to
extracttextural information from the TSX image in order to assess
theenhancement of the classication accuracy. Their results proved
theefciency of TSX imagery and texture-based analysis in
LULCmapping which leaded to acquire an overall accuracy of 95%
withkappa coefcient 93%. Another study by Lee et al. (2012a),
utilizedGLCM to detect the landslides on levees using SAR data.
Theproposed method was applied on L-band SAR data collected
fromNASA's UAVSAR of the Mississippi River levee system
betweenVicksburg, MS and Clarksdale, MS, USA. All known levee
landslidesin their study area could be detected with a low number
of falsepositives. In a related paper by Wei et al. (2012), oil
spill wasmonitored from ERS-2 SAR image using GLCM texture analysis
inBohai Sea, China. Through their analysis, they discovered
that
variance, contrast, dissimilarity and correlation are four
texturecharacteristics suitable for classication of oil spill. So
based on theliterature it can be said that texture analysis is a
proper tool whichassist hazard mapping and other applications.
Various approaches are available to examine texture. Amongthose
there are different structural, statistic, model-based andtransform
solutions (Pant et al., 2010). The structural approachconcentrates
on the hierarchical structure of texture. Hereby, therelationship
between micro-texture elements (primitives) respec-tive to their
spatial arrangement is of prime importance. Thestatistic approach
uses non-deterministic properties. Image grayvalues distribution
and relationships which are governed by theseproperties indirectly
represent texture in this case. Often secondorder statistics are of
importance (Lee et al., 2012b).
Another possibility to calculate texture is the model
basedoption (Kim and Kang, 2007). Both generative image and
stochas-tically based models are used in this approach. The image
analysesare done by estimating parameters for the respective
model.However, it has a high computational complexity due to the
needto estimate parameters of statistic models (Pradhan, 2010).
Thelast approach uses transform methods like Gabor, Fourier
orwavelet transforms. In this case, the image is transformed
intoanother space where the respective coordinate system
representscharacteristics of the respective texture such as its
frequency orsize (Materka and Strzelecki, 1998).
In thecurrent research, a statistical approachwasused for
textureanalysis. The concentrationwill rely on the gray level
co-occurrencematrix (GLCM) as this method was used for the
implementation ofthe texture analysis tool. The co-occurrence
matrix that was rstintroduced by Haralick (1979), is among the most
widely used forderivation of these statistic features. The
following section willintroduce the program which can be used in
texture analysis ofTerraSAR-X and the efciency of this programwill
be tested.
3. Methodology
The practical and theoretical aspects of texture analysis
imple-mented in this study involve several steps as shown in Fig.
1.
3.1. General information on design and functionality
The texture analysis program was implemented using ArcGIS
ArcObjects 9.2 and Visual Basic for Application (VBA). Fig. 2
showsthe interface of proposed program.
The program can be divided into three parts: raster
denition,matrix framework creation and formula application. The
graphicinterface of the program consists of one window containing
threeframes. The rst frame of the starting window allows the user
toselect a raster le for the texture analysis either by selecting
analready opened le in ArcMap 9.2 or by searching through
thedirectory. The image le will always be saved as an .img in
thefolder of the original image. Depending on chosen direction
andtype of analysis, an appendix will be added to the le name,
thusgiving it an easy to recognize the description.
Thesecondandthirdframemakesvarious textureoptions for
fourdifferent directions available to the user. The texture options
whichcan be found in the second frame are divided into the three
groupssuch as contrast group, orderliness group and statistics
group. Toachieve these three groups, GLCM should be calculated. The
GLCM,also called gray tonespatial dependencymatrix,wasrst
introducedby Haralick et al. (1973). To dene this matrix, one has
to picture arectangular imagewithNr
rows,Nccolumnsandwithaquantizationof Ng gray levels from the tone
of each single pixel. Lr{0,1,2,.,Nr}and Lc{0,1,2,.,Nc} will be the
number of rows or spatial domainsand dened as the set of Ng
quantied gray levels. If we create the
B. Pradhan et al. / Computers & Geosciences 63 (2014) 3443
35
-
matrix framework Lr Lc, then a set of pixels can be received.
Theseare ordered according to row and column respectively.
Lr Lc :
0;0 0;1 0;21;0 1;1 1;22;0 2;1 2;2
Nr ;0 Nr ;1 Nr ;2
0;Nc 1;Nc
2;Nc
Nr ;Nc
26666664
37777775
1
Lr f0;1;2;:;Nrg
Lc f0;1;2;:;Ncg
Each pixel is dened as a pair of coordinates and has a gray
level Gassigned to itself by a function that can be described as
the imageLr Lc; I : Lr Lc-G. Now, the relative frequency of
occurrence of apair of neighboring pixels or resolution cells that
have a distance dbetween themcanbeexpressedas amatrixPij. Hereby
the variables iand j describe respective gray levels. This matrix
Pij or is called theGLCM. It is symmetric and denes distance and
direction, the socalled angular relationship, of two pixel
neighbors as a function(Haralick, 1979).
Fig. 2. Texture analysis program interface.
Fig. 1. Methodology ow chart.
B. Pradhan et al. / Computers & Geosciences 63 (2014)
344336
-
Here is the brief explanation of how the GLCM is calculated.This
will lead to the formulas implemented in the texture analysistool
(Hall-Beyer, 2007). As mentioned previously, the GLCM has tobe
calculated prior to the implementation of the formulas.
Thisrequires the following steps as shown in Fig. 3. Matrix
number(1) in Fig. 3 will be a 33 subset of an image with the
respectivegray values shown. The undertaken calculation will
produce ahorizontal GLCM in a 33 window.
First, the matrix framework needs to be prepared. It will
containthe number of occurrences of each pair of gray value
combination.The pair itself consists of a neighbor pixel and a
reference pixelhaving both a certain distance d away from each
other and a specicspatial relationship. This can be either a
horizontal, vertical ordiagonal relationship.
The offset has to be constant during the entire calculation. In
theexample a distance of d1 and a horizontal relationship
wasselected, thus the relationship of pixel lying next to each
other isbeing examined. For instance, it can be noted that the
pixelcombination (0, 1) occurs twice in the image, therefore the
numbertwo will be placed at its corresponding position. After the
number ofoccurrences has been determined the matrix is made
symmetrical.The reason is, that both opposed directions are taken
into account. Itdoes not matter whether a pair of pixels has the
combination (1,0) or(0,1) as this will change when the opposite
direction is considered.The symmetrical matrix V can be realized by
adding the matrixframework F to its transposed matrix FT :
FFT 0 2 01 1 01 0 1
264
375
0 1 12 1 00 0 1
264
375
0 3 13 2 01 0 2
264
375 V 2
as a nal step, the symmetrical matrix V has to be
normalized.Therefore each value of this matrix is divided by the
sum of all valuesin V.
Pi;j Vij
N1i;j 0Vij3
Now that the GLCM has been calculated, the formulas can
beapplied. As mentioned they can be divided into three
groups:contrast group, orderliness group and statistics group. The
pre-viously used symmetric normalized GLCM example P will be usedin
the following paragraphs as a demonstration for calculation.
Thesecond matrix will be containing the respective weights
resultingfrom the given formula. The scalar product of these two
matriceswill be the sought for result.
3.1.1. Contrast groupThe Eqs. (4)(6) belong to the contrast
group. In this group the
contrast or difference in gray value between related pixels
isemphasized (Eq. 4). As it can be seen in the Fig. 2 the
contrastgroup has three subdivisions of contrast, similarity and
homo-geneity. This calculation emphasizes higher gray value
differences:
N1
i;j 0Piji j2 4
Similar to contrast the dissimilarity is calculated (Eq. 5). The
imagecan display values from zero in areas of equal tone to x,
where x isa positive real number depending on the radiometric
resolution.For an 8 bit input image the highest value would thus be
255.
N1
i;j 0Pijji jj 5
The last part of the contrast group is homogeneity (Eq.
6).Contrary to the two aforementioned methods it uses an
inversedifference approach. To avoid an error in calculations the
enumeratorwas increased by one so that no zero can exist under the
fraction bar:
N1
i;j 0
Pij1i j2
6
3.1.2. Orderliness groupThe second group concentrates on the
orderliness of values in a
kernel. This means that it looks at their regularity (Yenugu et
al.,2010). Likewise the contrast group weighted values are
used.Angular second moment (ASM), and energy are the two
optionsthat are clubbed together. Hereby the ASM uses its own
co-occurrence matrix as weight. Thus values from one to almost
zeroare possible. When close to zero a value will depend on the
size ofthe searching window and direction of the GLCM. The formula
1/napplies with n N1i;j 0Vij. The value one can only be reached, if
anarea displays a homogenous tone with only one gray value. Thesame
applies to the energy, also referred to as uniformity, iscalculated
by applying the square root to the ASM result:
ASM N1
i;j 0Pi;j
2 EnergyASM
p7
The maximum probability, short MAX is retrieved through
thesimplest calculation. The highest value of the GLCM is
searchedand then applied. Similar to the ASM, the result can lie
betweenone and 1/n, where n is nN1i;j 0Vij. In the example the
largest Pijvalue in the window is one fourth or 0.25. This might
indicate aslightly higher occurrence of one or two combinations.
However,no dominance is visible.
Entropy, which is another measure for order, refers to theamount
of chaos or disorder in a window. Its value ranges fromzero to n,
with n being dependent on the size of the searchingwindow and the
direction of the neighbor pixel relationship. Inthis case n2.485.
The smaller the value is, the higher the degreeof order in an image
will be.
N1
i;j 0Pi;j ln Pi;j 8
3.1.3. Statistics groupThe last group that should be measured is
statistics group. The
mean value in this calculation can be derived by either using
thereference pixel i or j. In the case of the symmetric GLCM, this
willyield an equal result because values in this matrix will appear
to bemirrored on the diagonal. Therefore i and j need not be
calculatedseparately. As a result values from zero to n, where n is
the number of
Fig. 3. Development of GLCM (example): kernel (1), matrix
framework Fij (2), symmetrical matrix Vij (3), normalized matrix
Pij (4).
B. Pradhan et al. / Computers & Geosciences 63 (2014) 3443
37
-
different gray values possible in the original image, can be
calculated.The mean itself displays a value around which most
likely thedominant reference pixel value can be found. In areas of
high contrastthis value, however, will be further away from the
reality:
N1
i;j 0i Pi;j 9
The variance is closely connected to the mean that describes
thedeviation from the previous statistic element. Similar to the
meancalculation, it is possible to determine both the s2i and the
value s
2i .
Due to the symmetry of the matrix, however, it will not be
necessary,as both results will be equal. Variance may have an
advantage overthe mean as its results have a higher range of
positive numbers, thusa greater stretch might be visible on output
images.
The standard deviation is similar to the variance. Its
onlydifference lies in another range of values as it is dened as
thesquare root of the variance:
s2 N1
i;j 0Pi;j i2 Standard deviation : s
s2
p10
The last member of this statistics group is the correlation: it
is anindication of how much a gray value depends on its
neighbors.Results may range from zero to one, where zero points to
anabsolutely uncorrelated image subset and one refers to fully
corre-lated areas. It can be seen that an error may be received as
a result, ifthe variance of an area is zero because the variance
can be found inthe enumerator of the formula and thus a division by
zero can occur.This might only happen if only one gray value exists
in the respectivearea. This subset would be fully correlated. Thus,
when programmingthe correlation the result can be set to one in
this case:
N1
i;j 0Pi;j
ijjjs2 is2 j
q
264
375 -simplified :
N1
i;j 0Pi;j
ijs2
11
3.1.4. Program usabilityCertain steps have to be undertaken in a
program, in order to
make it work smoothly and user friendly. Although they are
notnecessary themselves for the main program, but they can be
usefulfor the user as they give further options and also they can
givewarning of errors. Among those it can be mentioned to
thepossibility of opening the les that have already been opened
inArcMap 9.2, to be informed about existing les of the same
name.Also it can give the opportunity to overwrite these les and
tocreate multiband images for four directions of the same
textureoption. Also using le selection and drop down menu the
mistakeof accidentally selecting a non-raster le can be prevented.
Thesefunctions are also added to the main program.
3.2. Analysis
3.2.1. Data and study areaThe study is carried out in Terengganu
which is situated in
north-eastern part of Peninsular Malaysia, and is bordered in
thenorthwest by Kelantan, the southwest by Pahang, and the east
bythe South China Sea (Zaleha et al., 2006). The data used in
thiswork was recorded by TerraSAR-X satellite, short TSX-1, from
27thNovember 2009. Data was single look, stripmap modus, with
threemeters spatial resolution and HH polarization. TerraSAR-X
dataused with radiometric resolution of 16 bit thus providing
65,536different gray values.
3.2.2. Pre-processingIn order to guarantee good results, the
images have to be pre-
processed (Albinet et al., 2012). Therefore the following steps
are
followed. First, the speckle effect needs to be removed
(Idreeset al., 2013). ERDAS IMAGINE provides a few lters which
willsuppress and smooth out this effect (Haack and Bechdol,
2000).However, not all of them are equally suitable. Dong et al.
(2000)discussed this issue, comparing different lters regarding
theirmean, edge and textural information preservation as well as
theirreduction of the standard deviation. They also indicated that
aspeckle reducing lter should not distort and degrade the
inherenttexture if it is intended for texture preservation. After
evaluatingthe Lee, Kuan, Frost, mean, median and edge-sharpening
lter theyconcluded, that the median lter was not suitable as a
specklelter for SAR data as it strongly distorts the texture.
Comparison has been done between the outputs of these ltersusing
signal to noise ratio (SNR). Visual interpretation concludedthat
frost lter performed better than the rest in this study. TheFrost
lter displays this area very well, with low noise levels. Forthis
purpose, 55 window Frost lter was used to remove thespeckle and the
results showed features are still visible withoutappearing blur.
Moreover, line and point features are clearly visibleand the
texture context seems to have been least altered.
In the next step the radiometric resolution has to be rescaled
fromunsigned 16 bit to unsigned 8 bit which means that the number
ofgray values available has to be downscaled from 65,536 to 256.
Thereason therefore is an incompatibility of a method used
withArcObjects. It does not matter which method is used,
information willalways be lost when an image is rescaled (Zan et
al., 2008). Thequestion that should now be answered is: which
method does theleast damage to the original texture information of
the image? Threedifferent methods using ERDAS IMAGINE were used and
analyzed.Finally the method which selected was the one that can be
found asrescale option through ERDAS Interpreter menu under the
optionutilities. The reason for this selection is it has neither
the problem of astrongly altered contrast situation nor does it
provide oating grayvalues.
Finally, suitable subset has to be found. Fig. 4 shows theproper
scene. The Terengganu river can be recognized as well asa
mountainous area and different agricultural elds with
differenttypes of crops. Also rectangular structure of different
buildings andthe linear, dark structure of streets can be seen.
Fig. 4. Subset of area of the Terengganu River, Malaysia.
B. Pradhan et al. / Computers & Geosciences 63 (2014)
344338
-
Fig. 5. Original image (upper left) and contrast group results:
contrast (upper right), dissimilarity (lower left), homogeneity
(lower right).
Fig. 6. Orderliness group results: ASM (upper left), energy
(upper right), entropy (lower left), MAX (lower right).
B. Pradhan et al. / Computers & Geosciences 63 (2014) 3443
39
-
3.2.3. Texture analysis and classicationAfter all these steps
have been undertaken, the texture analysis
could be applied to the subset image using the proposed program.
Inrecent years, the object-based classication methods proved to
bemore efcient than pixel based due to the spatial information that
isused in these methods (Al Fugara et al., 2009; Blaschke et al.,
2008;Mahmoud et al., 2011). Results of the texture analysis were
used inpixel-based classication and four classes could be
recognized by theERDAS program: water body, settlement,
agriculture/elds andforest. Also advanced object-based classication
method was appliedon the outputs of the texture analysis and nally
compared with the
pixel-based result. As we know segmentation is the basis of
theobject-based classication that divides the image into the
homo-geneous objects and classies these objects based on spectral,
spatial,textural, relational and contextual information
(Petropoulos et al.,2012). These regions are homogenous in some way
but also differfrom their adjacent regions (Morris et al., 1986).
Thus this segmenta-tion corresponds in a way with the human
perception of areas.
For object-based classication Deniens eCognition 7.0 wasused
which is very popular in optical and radar remote sensing(Tehrany
et al., 2013a). However, before the actual segmentationprocess it
was necessary to combine all texture results which
Fig. 7. Statistics group result: mean (upper left), variance
(upper right), Standard deviation (lower left), correlation (lower
right).
Fig. 8. Exemplary result of a pixel based classication with
ERDAS Imagine.
B. Pradhan et al. / Computers & Geosciences 63 (2014)
344340
-
achieved from texture analysis into one image. This was
doneusing Composite Bands of ArcMaps Data Management Tools.
Thiscreated an image le with 44 bands. After a recalculation of
itsstatistics and the assignment of the no data value 9999,
theimage could be used for segmentation. Scale and color values
wereselected, depending on the properties of the image used.
Herebythe resolution of the image object level is dened by the
scaleparameter. Thus low values indicate that smaller objects will
becreated. The second parameter, color, signies the importance
ofimage color homogeneity in the segmentation process. For
thiswork, scale received a parameter value of ten and 0.9 was
assignedas the color parameter. After the execution of the
segmentationprocess an image containing segmentation polygons was
calcu-lated and returned. Nearest neighbor was chosen as the
classica-tion method. Color and mutual embedding were both taken
intoconsideration. Training areas could be dened by assigning
certainsegments to either a positive or negative predened output
class.This pre-denition had to be done by the user and consisted of
aname and a color for the later class. For example, the
rstclassication cycle contained the positive output class water
(blue)and the negative output class land (green). This way an
evaluationof the texture analysis program and its results can be
conducted.
4. Results and discussion
The results of the structural texture analysis are as follows.
Theimage contains four bands to display the four possible
directions.Small part of the study area has been chosen to
represent here, in
order to show the impact of each group of texture analysis on
datamore clearly. As can be seen in Fig. 5 results of the contrast
groupproduced concordant results. Settlement areas had high
contrastvalues, and thus contained low homogeneity values,
whereaswater bodies and ooded areas displayed a low contrast and
thushigh homogeneity.
The surrounding area displays lower contrast values. Also
theresults of orderliness group and statistics group can be seen
inFigs. 6 and 7 respectively. Classication scheme was applied
usingthe texture analysis results to examine the efciency of
twoclassiers and to assess the impacts of the texture informationon
the precision of the results of each classication. The result ofthe
pixel-based classication is shown in Fig. 8.
Visual interpretation of the thematic map derived by
ERDASimagery has illustrated the weakness of pixel-based to extract
thefeature classes. Due to the high oat value variation in
thedifferent images a pixel based approach is not advisable.
Thiswould simply yield inhomogeneous areas with a high amount
ofnoise as well as misclassication. As can be seen in Fig. 8,
waterwas not recognized by the program. Although training areas
wereused and a supervised classication was conducted, it was
inter-preted as forest or agriculture area. Also clearly visible
are theinhomogeneous areas where forest or agricultural areas would
besuspected. Therefore a pattern recognition approach using
seg-mentation was examined. Fig. 9 shows the image before
applyingthe segmentation. It was possible to differentiate water
from theland. Also a settlement can be recognized. Fig. 10 shows a
part ofthe segmented image containing the biggest settlement as
well aselds and ooded area.
Fig. 9. Six layer false color representation of the texture band
composite in eCognition with possible elements for later
classication (rotated 901 clockwise). (Forinterpretation of the
references to color in this gure legend, the reader is referred to
the web version of this article.).
Fig. 10. Result of the multiresolution segmentation process.
B. Pradhan et al. / Computers & Geosciences 63 (2014) 3443
41
-
As can be seen in the Fig. 10, different classes could
beseparated very well through the segmentation and it shows
theefciency of the object-based classication and the strangeness
ofeCognition software in precise segmentation. Fig. 11 shows
thethematic map acquired by object-based method and ve landcover
classication results were obtained.
As can be seen in Fig. 11, the thematic map has acceptable
andrepresentable appearance which proves the efciency of
theobject-based classication. In some areas some
misclassicationsare obvious but accuracy assessment should be done
because it isnot proper to do any judgments without considering and
evaluat-ing the statistics. For that reason, the level of
correctness for theresults of classications was evaluated by
accuracy assessmentthrough confusion matrix (Foody, 2002; Magee,
2011). Table 1shows the results of the accuracy assessment. The
results indicatedthat the proposed object-based method produced an
acceptableproducer and user accuracies for all the classes except
the class ofsettlement and wood. The reason could be related to the
mixtureof the spectral information of two classes of settlement and
wood,
so the probability of having errors is very high. Also due to
spectralsimilarity of the class of settlement and wood and the
class ofwood, their accuracies are lower than other classes. This
impliesthat it was difcult for the program to differentiate between
thesetwo classes.
5. Conclusion and recommendation
In this paper, we have developed an easy to use ArcMapprogram
which is made in visual basic environment and it is ableto perform
the texture analysis of features using TerraSAR-Ximagery. Secondly,
ERDAS and eCognition software were used forpixel-based and
object-based classication respectively in order toexamine the
precision of both methods in extracting the features.Results showed
that pixel-based approach was not that successful,whereas
object-based approach showed very good results. Object-based by
83.63% overall accuracy and 0.8 kappa coefcient provedits efciency
in detecting the features over traditional pixel-basedmethod.
Results on image texture classication showed that theproposed
program is capable in the analysis of TerraSAR-Ximagery. Hence the
developed program can be used in many earthobservation applications
such as LULC mapping, change detection,and hazard studies. The
program could differentiate between theobjects which lead to
enhance their results. The program itselfperformed well on the task
but the computation process took littlelonger time. If the
algorithm is enhanced that might improve theprocessing time.
However, the most time consuming aspect of thisprogram is the
creation of pixel blocks, therefore improvement ofalgorithm does
not have signicant impact. Thus to really reducethe runtime it is
advisable to use smaller signicant subsets of ascene. The currently
available version is a robustly running betaversion. However, some
features have not yet been implemented.The two most important
features are, are the use of different sizedkernels in order to
assess a greater area of texture and thequestion of the no-data
value handling. Further, improvementsin usability and user options
might be implemented. Also usingdata fusion can improve the
accuracy of the texture analysis.
Acknowledgments
The German Aerospace Center (DLR) provided Terra-SAR-X dataunder
the Science proposal ID: HYD0326. Thanks to Thomas Hah-mann for his
valuable inputs on the processing of Terra-SAR-X data.
References
Al Fugura, A., Billa, L., Pradhan, B., Mohamed, T.A., Rawashdeh,
S., 2011. Coupling ofhydrodynamic model and aerial
photogrammetry-derived digital surface modelfor ood simulation
scenarios using GIS: Kuala Lumpur ood. Malaysia. DisasterAdv. 4
(4), 2028.Fig. 11. Result derived by object-based method.
Table 1Object-based classication accuracy assessment.
Settlement Water Settlement and wood Wood Woodless Sum User
accuracy %
Settlement 46 1 2 0 0 49 94Water 0 51 0 0 0 0 100Settlement and
wood 0 2 29 7 12 50 58Wood 1 4 4 33 1 43 77Woodless 0 2 0 1 30 33
91Sum 50 37 51 50 50 238Producer accuracy % 98 85 83 80 70
Overall accuracy 83.6%Kappa coefcient 0.8
B. Pradhan et al. / Computers & Geosciences 63 (2014)
344342
-
Al Fugara, A.M., Pradhan, B., Mohamed, T.A., 2009. Improvement
of land-useclassication using object-oriented and fuzzy logic
approach. Appl. Geomat. 1(4), 111120.
Albinet, C., Borderies, P., Fabre, S., 2012. Radar and optical
parallel modelling offorest remote sensing data. Int. J. Remote
Sens. Appl. 2, 4.
Biro, K., Pradhan, B., Sulieman, H., Buchroithner, M., 2012.
Evaluation of TerraSAR-Xdata for land use/land cover analysis using
object-oriented classicationapproach in the African Sahel Area,
Sudan. J. Indian Soc. Remote Sens. ,
http://dx.doi.org/10.1007/s12524-012-0230-7.
Blaschke, T., Lang, S., Hay, G.J. (Eds.), 2008. Object-Based
Image Analysis SpatialConcepts for Knowledge-Driven Remote Sensing
Applications, rst ed. Springer,New York.
Blaschke, T., 2010. Object based image analysis for remote
sensing. ISPRS J.Photogramm. Remote Sens. 65 (1), 216.
Buckreuss, S., Mhlbauer, P., Mittermayer, J., Balzer, W.,
Werninghaus, R., 2006. TheTerraSAR-X Ground Segment. In:
Proceedings of the European Conference onSynthetic Aperture Radar
(EUSAR). Dresden, Germany.
Chica-Olmo, M., Abarca-Hernandez, F., 2000. Computing
geostatistical imagetexture for remotely sensed data classication.
Comput. Geosci. 26 (4),373383.
Christina Herzfeld, U., Zahner, O., 2001. A
connectionist-geostatistical approach toautomated image
classication, applied to the analysis of crevasse patterns
insurging ice. Comput. Geosci. 27 (5), 499512.
DellAcqua, F., Gamba, P., 2006. Discriminating urban
environments using multi-scale texture and multiple SAR images.
Int. J. Remote. Sens. 27 (18), 37973812.
DLR-Deutsches Zentrum fr Luft- und Raumfahrt, 2009. DLR Portal-
Anwendungenund Nutzung. DLR Portal. [Online] Deutsches Zentrum fr
Luft- und Raumfahrt,20. March 2009. (Zitat vom: 21. February 2010)
http://www.dlr.de/desktopdefault.aspx/tabid-4219/8885_read-16087/.
Dong, Y., Milne, A.K., Forster, B.C., 2000. A review of SAR
speckle lters: texturerestoration and preservation. In: Proceedings
of the Geoscience and RemoteSensing Symposium 2000 (IGARSS 2000),
IEEE 2000 International, vol. 2,pp. 633635.
Elbialy, S., Mahmoud, A., Pradhan, B., Buchroithner, M., 2013.
Application ofspaceborne synthetic aperture radar data for
extraction of soil moisture andits use in hydrological modelling at
Gottleuba Catchment, Saxony, Germany. J.Flood Risk Manag. ,
http://dx.doi.org/10.1111/jfr3.12037.
Floricioiu, D., Eineder, M., Rott, H., Nagler, T., 2008.
Velocities of major outletglaciers of the Patagonia Iceeld observed
by TerraSAR-X. In: Proceedings of theGeoscience and Remote Sensing
Symposium 2008 (IGARSS 2008). IEEE Inter-national, vol. 4, pp.
IV347.
Foody, G.M., 2002. Status of land cover classication accuracy
assessment. RemoteSens. Environ. 80 (1), 185201.
Fugura, A., Billa, L., Pradhan, B., 2011. Semi-automated
procedures for shorelineextraction using single RADARSAT-1 SAR
image. Estuar., Coast. Shelf Sci. 95 (4),395400.
Gibson, P., 2000. Introductory remote sensing principals and
concepts. Taylor andFrancis, New York.
Haack, B., Bechdol, M., 2000. Integrating multisensor data and
RADAR texturemeasures for land cover mapping. Comput. Geosci. 26
(4), 411421.
Hall-Beyer, M. GLCM Texture Tutorial, 2007. University of
Calgary. (Online)University of Calgary, 21. February 2007. (Zitat
vom: 2. February
2010)http://www.fp.ucalgary.ca/mhallbey/tutorial.htm.
Hamedianfar, A., Shafri, H.Z., 2013. Development of fuzzy
rule-based parameters forurban object-oriented classication using
very high resolution imagery. Geo-carto. Int. ,
http://dx.doi.org/10.1080/10106049.2012.760006.
Haralick, R.M., Shanmugam, K., Dinstein, I.H., 1973. Textural
features for imageclassication. IEEE Trans. Syst. Man. Cybern. 6,
610621.
Haralick, R.M., 1979. Statistical and structural approaches to
texture. Proc. IEEE 67(5), 786804.
Hassaballa, A.A., Althuwaynee, O.F., Pradhan, B., 2013.
Extraction of soil moisturefrom RADARSAT-1 and its role in the
formation of the 6 December 2008landslide at Bukit Antarabangsa,
Kuala Lumpur. Arab. J. Geosci., 110,
http://dx.doi.org/10.1007/s12517-013-0990-6.
Idrees, M.O., Saeidi, V., Yusuf, Y.A., Pradhan, B., 2013.
Advanced differentialinterferometry SAR techniques for deformation
monitoring: A review onsensors and recent research development.
Geocarto. Int. ,
http://dx.doi.org/10.1080/10106049.2013.807305.
Kia M.B., Pirasteh S., Pradhan B., Mahmud A.M., Sulaiman W.N.A.,
Moraddi A., 2012.An articial neural network model for ood
simulation using GIS: Johor riverbasin, Malaysia. Environ. Earth
Sci. 67 (1), 251264,
http://dx.doi.org/10.1007/s12665-011-1504-z.
Kiema, J., 2002. Texture analysis and data fusion in the
extraction of topographicobjects from satellite imagery. Int. J.
Remote Sens. 23 (4), 767776.
Kim, S.C., Kang, T.J., 2007. Texture classication and
segmentation using waveletpacket frame and Gaussian mixture model.
Pattern Recognit. 40 (4), 12071221.
Lee, M.A., Aanstoos, J.V., Bruce, L.M., Prasad, S., 2012a.
Application of omni-directional texture analysis to SAR images for
levee landslide detection. In:Proceedings of 2012 IEEE
International Geoscience and Remote SensingSymposium (IGARSS),
IEEE, pp. 18051808.
Lee, S.J., Serre, M.L., van Donkelaar, A., Martin, R.V.,
Burnett, R.T., Jerrett, M., 2012b.Comparison of geostatistical
interpolation and remote sensing techniques forestimating long-term
exposure to ambient PM2. 5 Concentrations across theContinental
United States. Environ. Health Perspect. 120 (12), 17271732.
Luo, Y., Zhao, S., Zhou, H., Wang, A., He, K., Tan, L., 2012. A
novel classicationmethod based on texture analysis using
high-resolution SAR and optical data.In: Proceedings of the 2012
Second International Workshop on the EarthObservation and Remote
Sensing Applications (EORSA), IEEE, pp. 186190.
Magee, K.S., 2011. Segmentation, object-oriented applications
for remote sensingland cover and land use classication (Ph.D.
dissertation). University ofCincinnati, USA. (131pp).
Mahmoud, A., Elbialy, S., Pradhan, B., Buchroithner, M., 2011.
Field-based landcoverclassication using TerraSAR-X texture
analysis. Adv. Space Res. 48 (5),799805.
Materka, A. Strzelecki, M., 1998. Texture Analyisis Methods A
Review.Morris, O.J., Lee, M.D.J., Constantinides, A.G., 1986. Graph
theory for image analysis:
an approach based on the shortest spanning tree. Communications,
Radar andSignal Processing. IEE Proceedings, F 133 (2), 146152.
Ouma, Y.O., Josaphat, S.S., Tateishi, R., 2008. Multiscale
remote sensing datasegmentation and post-segmentation change
detection based on logical mod-eling: theoretical exposition and
experimental results for forestland coverchange analysis. Comput.
Geosci. 34 (7), 715737.
Pant, T., Singh, D., Srivastava, T., 2010. Advanced fractal
approach for unsupervisedclassication of SAR images. Adv. Space
Res. 45 (11), 13381349.
Petropoulos, G.P., Kalaitzidis, C., Prasad Vadrevu, K., 2012.
Support vector machinesand object-based classication for obtaining
land-use/cover cartography fromHyperion hyperspectral imagery.
Comput. Geosci. 41, 99107.
Pradhan, B., 2009. Flood susceptible mapping and risk area
estimation using logisticregression, GIS and remote sensing. J.
Spat. Hydrol. 9 (2), 118.
Pradhan, B., Shae, M., 2009. Flood hazard assessment for cloud
prone rainy areasin a typical tropical environment. Disaster Adv. 2
(2), 715.
Pradhan, B., Shae, M., Pirasteh, S., 2009. Maximum ood prone
area mappingusing RADARSAT images and GIS: Kelantan river basin.
Int. J. Geoinf. 5 (2),1123.
Pradhan, B., 2010. Remote sensing and GIS-based landslide hazard
analysis andcross-validation using multivariate logistic regression
model on three test areasin Malaysia. Adv. Space Res. 45 (10),
12441256.
Pradhan, B., Youssef, A.M., 2011. A100-year maximum oood
susceptibility map-ping using integrated hydrological and
hydrodynamic models: Kelantan RiverCorridor. Malaysia. J. Flood
Risk Manag. 4 (3), 189202,
http://dx.doi.org/10.1111/j.1753-318X.2011.01103.x.
Ren, Y., Lehner, S., Brusch, S., Li, X., He, M., 2012. An
algorithm for the retrieval of seasurface wind elds using X-band
TerraSAR-X data. Int. J. Remote Sens. 33 (23),73107336.
Sousa, J.J., Magalhes, L.M., Ruiz, A.M., A.M.R., Cardoso, G.,
2013. The viStaMPS toolfor visualization and manipulation of time
series interferometric results, seriesinterferometric results.
Comput. Geosci. 52, 409421,
http://dx.doi.org/10.1016/j.cageo.2012.11.012.
Tamura, H., Mori, S., Yamawaki, T., 1978. Textural features
corresponding to visualperception. IEEE Trans. Syst. Man Cybern. 8
(6), 460473.
Tehrany, M.S., Pradhan, B., Jebu, M.N., 2013a. A comparative
assessment betweenobject and pixel-based classication approaches
for land use/land covermapping using SPOT 5 imagery. Geocarto.
Int., http://dx.doi.org/10.1080/10106049.2013.768300.
Tehrany, M.S., Pradhan, B., Jebur, M.N., 2013b. Spatial
prediction of ood susceptibleareas using rule based decision tree
(DT) and ensemble bivariate andmultivariate statistical models. J.
Hydrol. 504, 6979,
http://dx.doi.org/10.1016/j.jhydrol.2013.09.034.
Treitz, P.M., Soulis, E.D., Howarth, P.J., Kouwen, N., 1996.
Texture processing ofsynthetic aperture radar data using
second-order spatial statistics. Comput.Geosci. 22 (1), 2734.
Vidal, A., Moreno, M.R., 2011. Change detection of isolated
housing using a newhybrid approach based on object classication
with optical and TerraSAR-Xdata. Int. J. Remote Sens. 32 (24),
96219635.
Wei, L., Hu, Z., Guo, M., Jiang, M., Zhang, S., 2012. Texture
feature analysis in oil spillmonitoring by SAR image. In:
Proceedings of IEEE 20th International Con-ference on
Geoinformatics (GEOINFORMATICS), 2012, pp. 16.
Yenugu, M.M., Marfurt, K.J., Matson, S., 2010. Seismic texture
analysis for reservoirprediction and characterization. Lead. Edge
29 (9), 11161121.
Youssef, A., Pradhan, B., Hassan, A.M., 2011. Flash ood risk
estimation along the St.Katherine road, southern Sinai, Egypt using
GIS based morphometry andsatellite imagery. Environ. Earth Sci. 62,
611623.
Zaleha, K., Roswati, M.A., Iwasaki, N., 2006. Distribution of
some species ofharpacticoid copepods in east coast of Peninsular
Malaysia. Coast. Mar. Sci.30 (1), 141145.
Zan, F.D., Guarnieri, A.M., Rocca, F., Tebaldini, S., 2008.
Sentinel-1 radar interfero-metry applications. In: Proceedings of
2008 7th European Conference onSynthetic Aperture Radar (EUSAR),
VDE, 2008, pp. 14.
Zhang, J., Marszaek, M., Lazebnik, S., Schmid, C., 2007. Local
features and kernelsfor classication of texture and object
categories: a comprehensive study. Int. J.Comput. Vis. 73 (2),
213238.
B. Pradhan et al. / Computers & Geosciences 63 (2014) 3443
43
An easy to use ArcMap based texture analysis program for
extraction of flooded areas from TerraSAR-X satellite
imageIntroductionTexture analysis: a previewMethodologyGeneral
information on design and functionalityContrast groupOrderliness
groupStatistics groupProgram usability
AnalysisData and study areaPre-processingTexture analysis and
classification
Results and discussionConclusion and
recommendationAcknowledgmentsReferences