This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1

A New Semiautomated Detection Mapping of Flood

Extent From TerraSAR-X Satellite Image Using
Rule-Based Classification and Taguchi
Optimization Techniques
Biswajeet Pradhan, Mahyat Shafapour Tehrany, and Mustafa Neamah Jebur

Abstract—Floods are among the most destructive natural di- Therefore, the proposed method can assist researchers and plan-
sasters worldwide. In flood disaster management programs, flood ners in implementing and expediting flood inventory mapping.
mapping is an initial step. This research proposes an efficient
methodology to recognize and map flooded areas by using Index Terms—Flood detection, GIS, Landsat, remote sensing,
TerraSAR-X imagery. First, a TerraSAR-X satellite image was rule-based classification, Taguchi, TerraSAR-X.
captured during a flood event in Kuala Terengganu, Malaysia, to
map the inundated areas. Multispectral Landsat imagery was then I. I NTRODUCTION
used to detect water bodies prior to the flooding. In synthetic aper-
ture radar (SAR) imagery, the water bodies and flood locations
appear in black; thus, both objects were classified as one. To over-
come this drawback, the class of the water bodies was extracted
A PPROXIMATELY 4% of the land on the planet is covered
by wetland ecosystems [1]. Most of these wetlands are
floodplain and are located in tropical countries [2]. Every year,
from the Landsat image and then subtracted from that extracted floods occur with increasing frequency and heavily damage
from the TerraSAR-X image. The remaining water bodies rep- lives and properties [3]. Floods can be managed efficiently
resented the flooded locations. Object-oriented classification and through flood susceptibility, hazard, and risk mapping [4], [5]
Taguchi method were implemented for both images. The Landsat
images were categorized into three classes, namely, urban, veg- which is based on the identification of flooded areas [6].
etation, and water bodies. By contrast, only water bodies were Moreover, the reliability of a flood inventory map directly
extracted from the TerraSAR-X image. The classification results influences the generation of susceptibility and hazard maps [7].
were then evaluated using a confusion matrix. To examine the Therefore, the method used to determine flood locations should
efficiency of the proposed method, iterative self-organizing data be accurate. Flood detection analysis should also be rapid [8]
analysis technique (ISODATA) classification method was applied
on TerraSAR-X after employing the segmentation process during because floods can subside quickly in an inundated area. Thus,
object-oriented–rule-based method, and the results were com- researchers have limited time with which to map all of the
pared. The overall accuracy values of the classified maps derived locations. Fieldwork is unsuitable for such analysis given on-
from TerraSAR-X using the rule-based method and Landsat im- site challenges and difficulties [9]. It is time-consuming and
agery were 86.18 and 93.04, respectively. Consequently, the flooded costly, which is not practical for real-time studies [10]. Flood
locations were recognized and mapped by subtracting the two
classes of water bodies from these images. The acquired overall happens over a large area, thus making it difficult to reach
accuracy for TerraSAR-X using ISODATA was considerably low all those areas as they will not stay for long duration [11].
at only 57.98. The current research combined the methods and Furthermore, traditional hydrological methods, such as gauge
the optimization technique used as an innovative flood detection and discharge measurements, have some weak points to moni-
application. The successful production of a reliable and accurate tor and map flood locations because of the temporal and spatial
flood inventory map confirmed the efficiency of the methodology.
heterogeneity of large wetlands [1]. The launch of various satel-
lites and sensors has revolutionized the monitoring, evaluation,
and prediction of natural disasters [12]. Moreover, flood detec-
Manuscript received September 21, 2015; revised January 7, 2016; accepted tion has improved considerably because of the availability of
March 1, 2016. This work was supported by the Ministry of Higher Edu- multitemporal data, the increased spatial resolution, the devel-
cation Research Project under Flood Disaster Management (project number:
FRGS/1/2015/STWN06/UPM/02/1) with vote number 5524687 and also by opment of change detection algorithms, and the incorporation
the UPM University Research Grant (05-01-11-1283RU) to stimulate research of remote sensing techniques [13].
under the RUGS scheme with project number 9344100. Visual interpretation is a popular method to detect flood loca-
B. Pradhan is with the Department of Civil Engineering, Geospatial In-
formation Science Research Center, Faculty of Engineering, Universiti Putra tions. Reference [14] investigated the efficiency of this method
Malaysia, 43400, Serdang Selangor Malaysia and also with the Department using European remote sensing satellite data. The flood bound-
of Geoinformation Engineering, Choongmu-gwan, Sejong University, Seoul ary was derived using automatic classification techniques. The
05006, South Korea (e-mail: [email protected]; [email protected]).
M. S. Tehrany and M. Neamah Jebur are with the Department of Civil results showed that visual interpretation is more accurate than
Engineering, Geospatial Information Science Research Center, Faculty of satellite analysis. However, this finding is attributed to the
Engineering, Universiti Putra Malaysia, 43400, Serdang Selangor Malaysia. poor resolution of the satellite data and the limitation of the
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. technology and of computing software at the time. The recent
Digital Object Identifier 10.1109/TGRS.2016.2539957 access to a wide range of software, very high resolution satellite
0196-2892 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

imagery, and active and passive sensors facilitate the collection their difficulty and by the significant amount of time required
of data and the analysis and mapping of flood events within a to classify at least two SAR data [25]. SAR interferometry
few hours [15]. By contrast, the visual interpretation of satellite should be conducted to produce a coherence map; however,
images is a time-consuming, inaccurate, and costly method. It is this technique is often difficult to understand and interpret [26].
based on expert knowledge; therefore, it can be erroneous [16]. The generation of a coherence map is also complex and
All of the optical images are unsuitable for flood detec- disadvantageous; for instance, it requires ground data and two
tion applications [6], [17] because clouds usually cover the precisely coregistered SAR images [27]. The ground data dis-
sky during a flood event, thereby limiting the observational tinguish flooded areas from other low-coherence zones.
capability of these optical sensors. Some of these sensors are The SAR backscatter model from a river flood assumes
incapable of penetrating the cloud cover, and they are highly that the water surface is smoother than the adjacent land and
affected by weather conditions. Thus, they have been replaced is a specular reflector that reflects radiation from a sensor.
by active sensors, which are unaffected by sun illumination Therefore, the water appears darker than the land [23]. Paddy
and atmospheric conditions [18]. Moreover, synthetic aperture fields, part of the mountain facing away from the SAR sensor,
radar (SAR) signals can penetrate vegetation and forest [19]. and all water bodies also appear dark or black in radar imagery
These sensors can operate both day and night and can highlight [6]. Therefore, we must develop a method that can discriminate
different aspects of a single terrain because of their single- or between flooded areas and other objects. The current research
multipolarized capability. Furthermore, the flooded areas under aims to overcome some of the drawbacks of the existing meth-
vegetation can also be detected using specific SAR imageries ods and to establish a reliable and precise technique to detect
[20]. Therefore, SAR imagery has much potential application flood locations. We aim to apply change detection without
in flood studies [21]. requiring two SAR images; instead, we utilize the SAR imagery
Researchers have assessed various techniques to map flood- captured during flooding because it can penetrate cloud cover
ing events, and each technique has its pros and cons [18], [22] and record all of the objects on the terrain surface.
and [23]. For instance, the threshold segmentation algorithm or Furthermore, a cloud-free Landsat image recorded prior to
histogram thresholding is a simple but widely used and effective flooding was used, as the free Landsat data can reduce the
method to generate a binary image [22]. Reference [24] utilized required budget for the project and provide the required in-
this algorithm to map the extent of flooding in the Dongting formation under proper weather conditions. It is a well-known
Area of Hunan Province based on RADARSAT-1 imagery. fact that various features appear dark in SAR imagery, such as
Reference [1] mapped flood locations by a split-based auto- building roofs, paddy fields, and water bodies. These features
matic thresholding procedure on high-resolution TerraSAR-X are discriminated using the object-oriented rule-based classifi-
data of southwest England, particularly at the River Severn, U.K. cation method. This method considers other object characteris-
They stated that object-based context-sensitive thresholding is tics, such as texture and shape [28]. Therefore, this procedure
proven superior to pixel-based context-insensitive procedures involves two steps: 1) recognizing water versus nonwater re-
gions before and during flooding and 2) comparing the regions
owing to the addition of spatial information to the pure spectral
categorized as water or nonwater before and during flooding to
information derived from histogram thresholding. The effec-
detect flooded areas.
tiveness of thresholding procedures for floodplain recognition
with SAR sensors depends on the contrast between the flooded
and nonflooded regions. Therefore, thresholding is sensitive to II. S TUDY A REA AND DATA U SED
low-contrast images. However, this method is limited because In Malaysia, the Kuala Terengganu area has undergone much
it is tailored to each satellite scene, i.e., it is usually based on flooding over the last decade. Therefore, it was selected as the
visual interpretation. Moreover, its procedure is manual and study area, in which the efficiency of the proposed method can
time-consuming [22]. The extent of flooding in an area can be evaluated with respect to flood location detection (Fig. 1).
also be mapped by active contour modeling. Reference [23] Terengganu is situated in Peninsular Malaysia and is bordered
applied this method to single-frequency and single-polarization by the South China Sea in the east. On November 27, 2009,
SAR images to map the flood locations in Thames, which is a disastrous flood struck this area as a result of heavy pre-
west of Oxford, U.K. This method is advantageous because it cipitation. Therefore, this research utilized two data sources:
limits the noise caused by SAR speckle. However, this method TerraSAR-X imagery captured during flooding and Landsat
can only be used by a researcher with a priori knowledge of imagery recorded during a nonflood instance. The applied SAR
the statistical properties of images. Moreover, the method is data were collected by a TerraSAR-X satellite on November 27,
hindered by local minima and is inaccurate when the initial 2009, using HH polarization, single look, and 3 m of spatial
selected contour is simple or is far from the object boundary. resolution. Furthermore, the data were composed of stripmap
Flood areas can be extracted from multipass SAR data modus and short TSX-1 images with 16-b radiometric resolu-
through amplitude change detection techniques or the gener- tion. HH polarization data are less affected by the variations in
ation of a coherence map [25]. The amplitude change detection the roughness of water surfaces as caused by wind or vegetation
method compares two SAR images of the same scene: one than other polarization types [1]. On October 21, 2009, Landsat
captured prior to flooding and the other during or immediately imagery with a spatial resolution of 30 m was acquired. The
after the event. Subsequently, the water-filled zones can be image was obtained from path 126 and row 56. However, its
detected by determining the regions with reduced backscatter. spatial resolution was enhanced after pan-sharpening, which is
However, amplitude change detection techniques are limited by described in the section on preprocessing.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

PRADHAN et al.: NEW SEMIAUTOMATED DETECTION MAPPING OF FLOOD EXTENT 3

Fig. 1. Study area. (Left image) Malaysian states and (right image) Landsat imagery.

Fig. 2. Methodology flowchart.

III. M ETHODOLOGY resolution of this data. In order to perform object-oriented rule-

The general methodology implemented in this study is shown based classification schemes, two main steps should be done.
in Fig. 2. In the first stage, preprocessing was done for both The first stage is segmentation, which defines the boundary of
images of Landsat and TerraSAR-X. Speckles were removed the objects. Here, Taguchi optimization technique was used in
from TerraSAR-X using ISODATA method. Gap filling and order to find the optimum segmentation combination. In the
pan-sharpening were applied for Landsat imagery in order to second stage, rules were defined, and classification was done
solve the problem of gaps in Landsat and increase the spatial based on the derived rules. Three classes of vegetation, water,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

and urban were produced using the Landsat image. Similarly, to classify the TerraSAR-X data. The rule-based classification
TerraSAR-X was classified into two classes of water and non- method was also applied in Landsat imagery classification
water bodies. By subtracting the two classes of water bodies because of its efficiency [34]. Object-oriented classification is
from Landsat and TerraSAR-X, flooded areas were extracted. based on objects rather than on pixels using additional infor-
As a last step, validation was done using a confusion matrix, mation, such as the texture and color of the objects [35], [36].
and reliability of flooded area map was assessed. Each stage is Using this additional information, object-oriented classification
detailed in the succeeding sections. recognizes features such as floods more effectively than pixel-
based techniques. This method applies expert rules in classifi-
A. Preprocessing cation and suitably extracts the spectral family of signatures for
a specific class and the spectral overlap among classes (e.g.,
The original images must be preprocessed to generate reli-
floods have spectral characteristics that are similar to rivers
able and precise outcomes. The gaps in Landsat imagery must
and paddy fields), which is induced by restrictions in spec-
be filled, and the image must undergo pan-sharpening. How-
tral resolution and bandwidth [37]. Nonetheless, a researcher
ever, the Landsat 7 scan line corrector tool, which was designed
can set comprehensive rules using additional information on
to correct the undersampling of the primary scan mirror, failed
spectral, spatial, textural, and contextual factors [38]. Object-
to work in 2003. This increased scan gap induced a loss of ap-
oriented classification can also be implemented using different
proximately 22% of Landsat ETM+ information. Nonetheless,
software, including eCognition, ERDAS Objective, and Envi
the gaps can be filled through numerous methods. The current
Zoom. The current research utilized ENVI software because it
study applied the local linear histogram matching method be-
is proficient in such applications and contains the appropriate
cause it is often used by various researchers due to its highly
analysis tools. Moreover, this method involves two main steps,
accurate results [29], [30]. Reference [29] stated that the local
namely, segmentation and proper definition of rules [38].
linear histogram method is very simple and easy to implement
1) Segmentation Using the Taguchi Technique: In object-
and can resolve many of the missing-data problems. As stated
oriented classification, parameters such as scale, color, shape,
in Section II, the Landsat imagery used in the current research
and segments should be defined properly to recognize flooded
was captured on October 21, 2009. The local linear histogram
areas. Segmentation is the first stage in object-oriented analysis,
matching method filled the scan gap in the Landsat imagery
and it partitions an image into nonoverlapping regions [39].
obtained on August 18, 2009. To do so, the precise information
SAR imagery often varies little in terms of mean amplitude
on the available pixels and the pixels that should be filled must
among different types of land use/cover (LULC) [40]. However,
be determined. A scan gap mask was produced for each band
SAR amplitude fails to differentiate among different features
that displays existing data as 1 and that denotes the missing
and LULC types strongly. As a result, texture has been con-
data in the scan gap and areas to be filled by 0. Once the gaps
sidered as the main segmentation parameter in SAR imagery
were recognized, the linear histogram matching detected linear
classification under numerous applications [41]. Segmentation
transformation in the images. Moreover, the Landsat imagery
precision significantly influences the quality of the final clas-
was pan-sharpened using the Gram-Schmidt (GS) spectral
sified map. Therefore, this study used the multiresolution seg-
sharpening method [31]. The spatial resolution of the Landsat
mentation algorithm. It began with one pixel and progressed
image was 30 m; however, this method can improve this spatial
until all of the criteria were fulfilled [39]. This type of segmen-
resolution by merging the high-resolution pan image with the
tation was achieved through parameters such as scale, color,
bands of low spatial resolution [32]. Therefore, the spatial res-
and shape, which generate 243 combinations for segmentation.
olution of Landsat was enhanced to 15 m after pan-sharpening.
However, the effects of each combination are time-consuming
Speckles should be removed from the TerraSAR-X image
to evaluate. Hence, an appropriate optimization technique
using appropriate filters [33]. Filters such as Lee, Frost, and
should be developed to reduce the number of examinations
mean can suppress and smooth out the speckle effect [6].
and thus accelerate segmentation and classification. In line
However, each filter performs differently; consequently, not all
with this requirement, the Taguchi technique can obtain the
are equally appropriate. A filter should not distort and degrade
optimum combination of segmentation parameters [39]. Refer-
the inherent texture of the image. Thus, some of the filters
ence [39] used this method to optimize pixel-based and object-
were evaluated based on signal-to-noise ratio (SNR). Based
oriented classification in mapping the landslide locations in
on the visual interpretation and SNR values acquired in the
Kermanshah City, Iran. Their research investigated and con-
current research, the Frost filter was better than the other filters.
firmed the efficiency of this method. Therefore, the current
It accurately displayed the study area with low noise and no
study aims to utilize this technique to optimize the segmentation
blurring effect. A 4 × 4 window Frost filter was therefore
parameters. Taguchi tables facilitate easy and stable experimen-
utilized to remove the speckle from TerraSAR-X imagery [6].
tal designs. Therefore, only 25 experiments are selected for
assessment by the Taguchi method in terms of the three seg-
B. Rule-Based Classification
mentation parameters. Moreover, the plateau objective function
The classification techniques that are used to classify very (POF) was measured for each test to evaluate segmentation pre-
high resolution optical images, including those of urban areas, cision in each of the 25 experiments. POF is a combination of
are often unadaptable for SAR data. This unadaptability may be a spatial autocorrelation index and a variance indicator. Precise
attributed to the spatial heterogeneity of urban areas. Therefore, segmentation is represented by a high POF. More information
the more advanced object-oriented rule-based method was used about POF can be obtained from [39].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

PRADHAN et al.: NEW SEMIAUTOMATED DETECTION MAPPING OF FLOOD EXTENT 5

TABLE I TABLE II
D EFINED RULES FOR T ERRA SAR-X D EFINED RULES FOR L ANDSAT

To apply the general steps in the Taguchi method, the follow-

ing steps were performed. 1) The process objective was deter-
mined in the beginning. This step entails defining the possible
values of a particular parameter for the process. 2) The param- C. Iterative Self-Organizing Data Analysis Technique
eters that can influence the process were then defined. These (ISODATA) Classification
parameters (scale, color, and shape) exhibit variable values that
can affect performance; thus, the level should be defined by the Another classification method was used for comparison with
user depending on the parameter’s effect on the process. For the acquired results to evaluate the efficiency of the proposed
instance, a scale value can vary from 0.1 to 1.0. When the level rule-based method. ISODATA is one of the most popular
increases, the number of experiments to be conducted will classification algorithms in image processing that requires few
be increased as well. 3) An orthogonal array was created to parameters, such as the number of classes to be derived, the
design the condition and determine the number of experiments. number of iterations, and the threshold value [44]. In the
Selecting the orthogonal array depends on the number of levels current research, the algorithm was set to 10–20 classes, with
a threshold of 5% and 1 iteration. ISODATA is an unsupervised
and the number of parameters. 4) The experiments were then
statistical classification technique that produces a well-defined
applied after the appropriate array had been selected. The effect
pixel or object classification. It entails assigning each pixel or
of each parameter on the performance was then measured.
segmented object to a specific class. Supervised classification
The loss function can be calculated as follows [42]:
methods require the user to define the characteristics of the
l(y) = kc (y − T )2 (1) classes prior to classification; unsupervised methods rely on the
classification process itself to define the classes [45]. Unsuper-
where T is the target value of y and the measured value y is a vised classification techniques create natural groupings of data
loss function. Kc is the constant in the loss function that can be in attribute space that can be used to obtain insight into the data
calculated by considering the acceptable interval as follows: structure. The ISODATA method is mostly used for satellite
image classification. Spectral reflectance from different bands
C
kc = (2) is used to find classes in multidimensional attribute space. The
Δ2 classification outcome should be analyzed by the user who has
where C is the loss associated with sp limit and Δ is the knowledge and experience of the features. In certain cases,
deviation of the specification from the target value. When the program assigns data to the classes that they do not really
whole parameters that affect the process are defined, the level belong to [46]. The reason is that ISODATA is not very robust
of each parameter should also be defined. The level refers to compared to other advanced classification techniques such as
the probable value of each parameter in terms of maximum, the rule-based method. It is unsupervised, and all of the decision
minimum, and current values. In the case of a big gap between was made based on spectral information. Hence, it is possible
the minimum and maximum values of a specific parameter, ad- that some pixels are classified wrongly.
ditional levels are added to that parameter. The proper array was
selected after the number of parameters and levels have been D. Validation
defined. A constant array was found for the Taguchi method. The accuracy of the generated flood inventory map can be
Each array can be selected depending on the parameters and determined based on the completeness of the map and the
levels. Table I was created using an algorithm of Taguchi. In precision of the information shown [47]. The efficiency and
the case of three parameters and two levels, L25 array was the quality of the generated inventory map can be assessed
selected. The array assumes that the number of levels is equal using various methods. A popular method involves the devel-
for each parameter. Otherwise, the assumption will be based on opment of a confusion matrix [48], [49]. The ground truth
the highest value. data for TerraSAR-X were collected from field survey from
2) Rule Definition: Rule-based classification aims to en- November 27 to 29, 2009. The number of locations that were
hance feature classification precision. This method can re- used to perform accuracy assessment was 105, which were
duce confusion errors among spectrally similar categories [43]. collected by GPS during field surveying. During the surveying,
Moreover, a user can create and flexibly manipulate rules until the boundaries of the flooded areas were measured. The number
the best classification results are obtained. The necessary rules of the observed locations varies based on the area of the flood,
were manually defined according to the segment attributes of where bigger areas had more observation compared to the small
both images. Tables I and II list the rules defined for the Landsat ones. However, the 105 sites represent the flood events rather
and TerraSAR-X imagery in the current research. than the number of the observation. Regarding the Landsat
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

TABLE III
L25 O RTHOGONAL A RRAY AND POF FOR S EGMENTATION P ROCESS W HERE L R EPRESENTS
THE L EVEL W HICH R EFLECTS THE N UMBER OF THE E XPERIMENTS

accuracy assessment, samples were collected from field sur-

veying in 2009 and reference maps from Jabatan Ukur Dan
Pemetaan Malaysia (JUPEM). Confusion matrix can be used
to conduct four types of accuracy assessments that are related
to different aspects, and it can determine the overall accuracy,
producer accuracy, user accuracy, and kappa coefficients [34].
Overall accuracy is measured as the total number of correctly
classified pixels divided by the total number of test pixels.
Producer accuracy is related to the probability that the specific
land cover of an area is classified as such. User accuracy is
associated with the probability that a pixel is accurately labeled
as a certain land cover class in the map. Kappa coefficient is the Fig. 3. Segmented (a) TerraSAR-X and (b) Landsat imagery.
difference of the actual agreement between the reference data
and an automated classifier from the chance agreement between
40, 0.1, and 0.3 for scale, color, and shape, respectively. By
the reference data and a random classifier [50].
contrast, the maximum POF of 1.380 for the Landsat image
combined parameter values of 80, 0.1, and 0.5 for scale,
IV. R ESULTS color, and shape, respectively. Thus, these two combinations
are the optimum segmentation parameters for the two data
A. Optimization and Segmentation Results sets. The generated segmentation maps are depicted in Fig. 3.
As mentioned in the methodology section, 243 combinations Fig. 3(a) displays the segmentation of the entire Landsat
can be defined using three parameters, namely, scale, color, and scene; however, Fig. 3(b) displays only a small portion of
shape. The Taguchi optimization technique reduces the number the segmented TerraSAR-X image given its very high spatial
of experiments reduced to 25, as shown in Table III. This table resolution. This small region is presented for enhanced visual-
represents the experiments selected (available combinations) ization because the segmented areas were very small and could
given the segmentation parameters and the measured POFs for not be observed from afar.
each combination. The optimum segmentation parameters determined using the
Table III suggests that the maximum POF (1.310) obtained Taguchi technique were used to recognize object boundaries
for the TerraSAR-X image combined parameter values of with reasonable accuracy. The precision of this segmentation
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

PRADHAN et al.: NEW SEMIAUTOMATED DETECTION MAPPING OF FLOOD EXTENT 7

Fig. 4. Classified maps. (a) TerraSAR-X using the rule-based method. (b) Landsat imagery. (c) TerraSAR-X using the ISODATA method.

was assessed visually as the boundaries of most of the objects image using the rule-based method, which contains two classes,
were detected accurately. The results also confirmed the effi- namely, water and nonwater bodies. Fig. 4(b) depicts the clas-
ciency of the multiresolution segmentation approach. sified map of Landsat imagery, which consists of three classes,
i.e., urban, vegetation, and water. Furthermore, the classified
TerraSAR-X map using ISODATA can be seen in Fig. 4(c).
B. Classified Maps
The extent of the water bodies visualized in the classified
The two images of TerraSAR-X and Landsat were classified Landsat image [Fig. 4(b)] was less than the amount of water
according to the segmentation results and the defined rules. bodies depicted in the TerraSAR-X image [Fig. 4(a)]. The
Fig. 4(a) illustrates the classified map of the TerraSAR-X water bodies detected in the classified TerraSAR-X map are
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

Fig. 5. (a) Hill-shaded map of the study area with flood locations that was detected by subtracting the classes of water bodies derived from both images.
(b) Flooded areas without hill-shaded map.

TABLE IV
C ONFUSION M ATRIX FOR T ERRA SAR-X U SING RULE BASED IN
O BJECT-BASED C LASSIFICATION

TABLE V
C ONFUSION M ATRIX FOR T ERRA SAR-X
U SING ISODATA C LASSIFICATION

TABLE VI
C ONFUSION M ATRIX FOR L ANDSAT

Fig. 6. Masked DEM using flooded locations.

flooded regions; by subtracting the two classified water bodies, in the result of ISODATA analysis. This observation shows
we determined the locations of the flooded areas. As can be the weakness of unsupervised methods in classification. The
seen in Fig. 4(c), considerable misclassifications are evident confusion matrix gives information about the precision of the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

PRADHAN et al.: NEW SEMIAUTOMATED DETECTION MAPPING OF FLOOD EXTENT 9

results. Fig. 5 illustrates the flood locations detected in the study has been proven by current and various researches [55]. Mis-
area by the rule-based method as shown in Tables I and II. classifications were visible in the Landsat image because of the
A confusion matrix was generated to assess the efficiency segmentation process. For example, the existence of wetlands
of the proposed method and to evaluate the generated flood near the river or in the boundary of the river could be segmented
inventory map. Tables IV and V display the confusion ma- as a water object. Moreover, Landsat showed misclassification
trix results for the classified maps of TerraSAR-X using the in high-contrast areas such as a city with vegetation areas
rule-based and ISODATA classification methods, respectively. in the midst of urban spaces [56], [57]. TerraSAR-X works
Moreover, Table VI shows the confusion matrix results for well in less dense vegetation areas because the radar’s beam
Landsat images. cannot penetrate through the vegetation [58], [59]. TerraSAR-X
The overall accuracies of the classified maps of the showed misclassification in urban and vegetation areas where
TerraSAR-X and Landsat images were 86.18 and 93.04, re- buildings and long trees coexist [60].
spectively, thus indicating that the rule-based method efficiently The proposed method works reasonably well in open
discriminates between objects. As a result, an accurate clas- areas or regions without tall structures. Therefore, accurate
sification map is produced. Moreover, the kappa coefficients classification can be obtained in rural areas, but accuracy will be
were 0.72 and 0.77 for the TerraSAR-X and Landsat classified reduced in urban areas partly because of the restricted visibility
maps, respectively. Both accuracy assessment results suggest of TerraSAR-X of the ground surface owing to shadow and
that all of the user and producer accuracy values are reasonably layover [33], [61]. The algorithm proposed by [62] proved that
high, thereby suggesting that the generated classes are reliable. flooding in rural areas can be detected by TerraSAR-X with
The flood location map was constructed by subtracting the good accuracy and in urban areas with reasonable accuracy.
two derived water bodies; thus, their accuracy values directly The accuracy was reduced in urban areas partly because
affect map precision. Statistically, 69% of the flood took place of TerraSAR-X’s restricted visibility of the ground surface
in the vegetation areas, and the rest happened in the urban attributed to radar shadow and layover. This problem can be
areas. The confusion matrix indicated that both water bodies solved using another data source (e.g., such as airborne laser
had high producer and user accuracy values, thus confirming the scanning data, i.e., LiDAR), where the area behind the tall
reliability of the final flood location map generated. However, feature can be modeled [33], [63], as proposed by [33]. How-
the overall accuracy achieved by ISODATA was 57.98, which ever, obtaining airborne laser scanning data is a very expensive
is considerably less than the acquired accuracy from the rule- affair and not easily available for less developed countries.
based method. Another limitation of this method shows when the flooded
areas are wetlands, such as farmlands where the differentiation
between the flooded areas and the water for planting is difficult
[64], [65].
V. D ISCUSSION
The TerraSAR-X data used in the current research were HH
This study has presented an approach to overcome the dif- polarized. Such polarization data are less affected by varia-
ficulties in flood detection by combining Landsat (medium tions in water surface roughness caused by wind or vegetation
spatial resolution) and TerraSAR-X (high spatial resolution) compared to other polarization types [66]. Moreover, the
imageries. The results of our study correspond with other HH-polarized backscattered coefficient generally presents a
works suggesting that Landsat data can be useful for wa- higher contrast between water and land surfaces [67]. Accuracy
ter area extraction and mapping [51], [52]. Current research assessment showed that the rule-based method is significantly
then extends previous studies on combination of various stronger than the unsupervised ISODATA method. Therefore,
sensors for flood detection using optical and active sensors the proposed methodology is an easy, rapid, reliable, and
[53], [54]. low-cost procedure to map flood locations. Researchers may
The main rationale of using Landsat data in current study was use this method to construct a flood inventory that will serve
to extract the water bodies before flooding. This can be done as basis for flood susceptibility, hazard, and risk analyses.
using many ways; however, the aim was to use purely space- It has been shown that the accuracy of Landsat is limited if
borne remote-sensing-based methods in a cost-effective way. the object has less than two pixel size. This may cause some
The difference in the spatial resolution can be an issue. How- problems covering the small streams where it is the case in
ever, in this study, the accuracies acquired from both Landsat some parts of this study. Moreover, some other streams were
and TerraSAR-X classifications were reasonably similar, which too narrow to not be detected by both sensors. This can be one
proved that both datasets were applicable for such application. of the sources for some of the misclassification [15].
In Fig. 5(a), several of the flooded locations were in hilly On the other hand, when comparing Landsat and
areas. Overlay analysis was conducted to assess the precision TerraSAR-X classified results visually, the active sensor has
of the proposed flood detection method and the derived flood less misclassification mostly due to the high spatial resolution
inventory. Elevation was masked using the detected flooded and the high capability of the active sensor in detecting the
locations (Fig. 6). Most of the flooded areas were located in water bodies. This statement has been proven by many studies
an elevation of 2–280 m. The elevation of the whole study area which utilized SAR data to recognize the flood locations
ranged between 0 and 1445 m. [68]–[71]. The shadow did not have a significant influence as
TerraSAR-X could perform better than Landsat in classifying most of the area is an open area. Moreover, even in the urban
areas near rivers. The efficiency of TerraSAR-X in flood studies areas, there are no tall buildings that exist to create the problem
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

of shadow, which could enhance the result. This issue may be map the TerraSAR-X for comparison purposes. The precision
resolved by user interaction to correct some of the parts where and reliability of the proposed method were assessed by a
it is a well-known area, but it is time-consuming. confusion matrix, and the acquired accuracy values proved
When comparing the proposed method in this study to other the applicability of the proposed method in flood inventory
studies, two published papers were considered. The first one is mapping. ISODATA was incapable of mapping the water bodies
“Flood detection in urban areas using TerraSAR-X” authored with acceptable accuracy. Thus, this method is not applicable
by [33], and the second one is “The accuracy of sequential in flood detection studies. Unsupervised methods do not use
aerial photography and SAR data for observing urban flood the characteristics of the object, such as texture and shape, in
dynamics, a case study of the UK summer 2007 floods” au- classification. The proposed rule-based method can be used in
thored by [63]. In both of the aforementioned studies, the flood detection in tropical and nontropical areas with acceptable
authors aimed to estimate the regions of the image in which accuracy and reduced budget. Planners and researchers can
water would not be visible due to shadow or layover caused by therefore use the derived maps to study flood susceptibility,
buildings. However, in the current case study, most of the area hazard, and risk mapping further.
is flat (plain), and the buildings are not tall enough to make a
considerable shadow effect.
Finally, in the current research, three issues can be high- ACKNOWLEDGMENT
lighted: the complimentary use of freely downloadable opti-
cal Landsat data in flood studies, the segmentation process The authors would like to thank the National Mapping
using the Taguchi algorithm, and the rule-based–object-based Agency (JUPEM), Malaysia, for providing various datasets
method. It is proved that Landsat has significant capability in used in this paper. The German Aerospace Center (DLR)
this kind of research; however, the differences in the spatial provided the TerraSAR-X data under the Science proposal
resolutions made few misclassifications in the boundary of the ID:HYD0326.
water bodies such as river. On the other hand, the Taguchi
method was very helpful in reducing the time required to
R EFERENCES
employ all possible segmentation combinations as it reduced
the number of possible combinations from 243 into 25 exper- [1] J. M. Martinez and T. Le Toan, “Mapping of flood dynamics and spatial
distribution of vegetation in the Amazon floodplain using multitemporal
iments. Finally, the rule-based object-based method performed SAR data,” Remote Sens. Environ., vol. 108, no. 3, pp. 209–223, 2007.
better than the unsupervised classification ISODATA algorithm, [2] A. M. Youssef, B. Pradhan, and A. M. Hassan, “Flash flood risk estimation
although both were applied on the segmented data. along the St. Katherine road, southern Sinai, Egypt using GIS based
morphometry and satellite imagery,” Environ. Earth. Sci., vol. 62, no. 3,
pp. 611–623, 2011.
[3] M. S. Tehrany, B. Pradhan, and M. N. Jebur, “Spatial prediction of flood
VI. C ONCLUSION susceptible areas using rule based decision tree (DT) and a novel ensemble
bivariate and multivariate statistical models in GIS,” J. Hydrol., vol. 504,
Flood detection and inventory map production are very no. 4, pp. 69–79, 2013.
important in the assessment of natural hazards and in the [4] M. S. Tehrany, B. Pradhan, and M. N. Jebur, “Flood susceptibility map-
prevention of flood in tropical areas. Some of the optical sensors ping using a novel ensemble weights-of-evidence and support vector
machine models in GIS,” J. Hydrol., vol. 512, no. 2, pp. 332–343, 2014.
cannot capture data effectively during flooding because of cloud [5] F. Ip et al., “Flood detection and monitoring with the Autonomous Sci-
cover. This limitation can be overcome by the introduction of encecraft Experiment onboard EO-1,” Remote Sens. Environ., vol. 101,
very high resolution active sensors, which have revolutionized no. 4, pp. 463–481, 2006.
[6] B. Pradhan, U. Hagemann, M. Shafapour Tehrany, and N. Prechtel,
hazard studies. The current research developed an efficient “An easy to use ArcMap based texture analysis program for extraction
flood mapping method to facilitate and enhance the application of flooded areas from TerraSAR-X satellite image,” Comput. Geosci.,
of traditional flood detection methods, some of which are time- vol. 63, no. 2, pp. 34–43, 2014.
[7] M. S. Tehrany, M. J. Lee, B. Pradhan, M. N. Jebur, and S. Lee, “Flood sus-
consuming, are costly, and require much work. This study ceptibility mapping using integrated bivariate and multivariate statistical
detected flooded areas precisely and rapidly using only two models,” Environ. Earth. Sci., vol. 72, no. 10, pp. 1–15, 2014. [Online].
data sets and through a simple analysis. The state of Kuala Available: http://dx.doi.org/10.1007/s12665-014-3289-3.
[8] G. R. Brakenridge, E. Anderson, S. V. Nghiem, S. Caquard, and
Terengganu in Malaysia was selected as the test bed site be- T. B. Shabaneh, “Flood warnings, flood disaster assessments, and flood
cause it frequently experiences flooding. The object-oriented hazard reduction: The roles of orbital remote sensing,” in Proc. 30th Int.
rule-based classification method was used to extract the water Symp. Remote Sens. Environ., 2003, pp. 1–6.
[9] G. Schumann, P. D. Bates, M. S. Horritt, P. Matgen, and F. Pappenberger,
bodies from TerraSAR-X data obtained during a flood event. “Progress in integration of remote sensing-derived flood extent and stage
Furthermore, Landsat imagery was captured prior to flood sea- data and hydraulic models,” Rev. Geophys., vol. 47, no. 4, pp. 18–29, 2009.
son, which denotes the period in which flooding did not occur [10] P. Farina, D. Colombo, A. Fumagalli, F. Marks, and S. Moretti,
“Permanent scatterers for landslide investigations: Outcomes from the
in the study area. The water bodies from Landsat imagery were ESA-SLAM project,” Eng. Geol., vol. 88, no. 3, pp. 200–217, Dec. 2006.
mapped using the same classification technique and were subse- [11] E. Opolot, “Application of remote sensing and geographical information
quently subtracted from the TerraSAR-X-derived water bodies. systems in flood management: A review,” Res. J. Appl. Sci. Eng. Technol.,
vol. 5, no. 10, pp. 1884–1894, 2013.
The remaining water bodies represented the flooded locations [12] F. A. Hirpa et al., “Upstream satellite remote sensing for river discharge
in the study area. The quality of the study was enhanced by the forecasting: Application to major rivers in South Asia,” Remote Sens.
Taguchi optimization technique, multiresolution segmentation, Environ., vol. 131, no. 2, pp. 140–151, 2013.
[13] T. W. Gillespie, J. Chu, E. Frankenberg, and D. Thomas, “Assessment and
and rule-based classification. These techniques also expedited prediction of natural hazards from satellite imagery,” Prog. Phys. Geog.,
the process. The ISODATA classification method was used to vol. 31, no. 5, pp. 459–470, 2007.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

PRADHAN et al.: NEW SEMIAUTOMATED DETECTION MAPPING OF FLOOD EXTENT 11

[14] R. Oberstadler, H. Hönsch, and D. Huth, “Assessment of the map- pp. 1–15, 2013. [Online]. Available: http://dx.doi.org/10.1080/10106049.
ping capabilities of ERS-1 SAR data for flood mapping: A case 2013.848944.
study in Germany,” Hydrol. Process., vol. 11, no. 10, pp. 1415–1425, [36] M. S. Tehrany, B. Pradhan, and M. N. Jebur, “Remote sensing data reveals
1997. eco-environmental changes in urban areas of Klang Valley, Malaysia:
[15] K. Auynirundronkool et al., “Flood detection and mapping of the Contribution from object based analysis,” J. Indian. Soc. Remote Sens.,
Thailand Central plain using RADARSAT and MODIS under a vol. 41, no. 4, pp. 981–991, 2013.
sensor Web environment,” Int. J. Appl. Earth. Observ., vol. 14, no. 1, [37] T. R. Martha, N. Kerle, C. J. van Westen, V. Jetten, and K. V. Kumar,
pp. 245–255, 2012. “Segment optimization and data-driven thresholding for knowledge-based
[16] Y. Chambenoit, N. Classeau, E. Trouvé, and J. P. Rudant, “Performance landslide detection by object-based image analysis,” IEEE. Trans. Geosci.
assessment of multitemporal SAR images’ visual interpretation,” in Proc. Remote Sens., vol. 49, no. 12, pp. 4928–4943, Dec. 2011.
IEEE Int. Geosci. Remote Sens. Symp., 2003, pp. 3911–3913. [38] S. Siyahghalati, A. K. Saraf, B. Pradhan, M. N. Jebur, and M. S. Tehrany,
[17] J. Sanyal and X. Lu, “Application of remote sensing in flood management “Rule-based semi-automated approach for the detection of landslides in-
with special reference to monsoon Asia: A review,” Nat. Hazards., duced by 18 September 2011 Sikkim, Himalaya, earthquake using IRS
vol. 33, no. 2, pp. 283–301, 2004. LISS3 satellite images,” Geomat. Nat. Haz. Risk., vol. 10, no. 2, pp. 1–19,
[18] M. Horritt, D. Mason, and A. Luckman, “Flood boundary delineation 2014.
from synthetic aperture radar imagery using a statistical active contour [39] V. Moosavi, A. Talebi, and B. Shirmohammadi, “Producing a landslide
model,” Int. J. Remote Sens., vol. 22, no. 13, pp. 2489–2507, 2001. inventory map using pixel-based and object-oriented approaches opti-
[19] M. Karjalainen, V. Kankare, M. Vastaranta, M. Holopainen, and mized by Taguchi method,” Geomorphology, vol. 204, no. 1, pp. 646–656,
J. Hyyppä, “Prediction of plot-level forest variables using TerraSAR-X 2014.
stereo SAR data,” Remote Sens. Environ., vol. 117, no. 1, pp. 338–347, [40] M. Horritt, “A statistical active contour model for SAR image segmenta-
2012. tion,” Image. Vis. Comput., vol. 17, no. 3, pp. 213–224, 1999.
[20] M. Horritt, D. Mason, D. Cobby, I. Davenport, and P. Bates, “Waterline [41] A. Voisin, V. A. Krylov, G. Moser, S. B. Serpico, and J. Zerubia, “Classi-
mapping in flooded vegetation from airborne SAR imagery,” Remote Sens. fication of very high resolution SAR images of urban areas using copulas
Environ., vol. 85, no. 3, pp. 271–281, 2003. and texture in a hierarchical Markov random field model,” IEEE Geosci.
[21] S. Elbialy, A. Mahmoud, B. Pradhan, and M. Buchroithner, “Applica- Remote Sens. Lett., vol. 10, no. 1, pp. 96–100, Jan. 2013.
tion of spaceborne synthetic aperture radar data for extraction of soil [42] G. Taguchi, Introduction to Quality Engineering: Designing Quality Into
moisture and its use in hydrological modelling at Gottleuba Catchment, Products and Processes. White Plains, NY, USA: Quality Resources,
Saxony, Germany,” J. Flood. Risk. Manag., vol. 7, no. 2, pp. 159–175, 2013. 1986.
[22] L. Pulvirenti, M. Chini, N. Pierdicca, L. Guerriero, and P. Ferrazzoli, [43] C. Baker, R. Lawrence, C. Montagne, and D. Patten, “Mapping wetlands
“Flood monitoring using multi-temporal COSMO-SkyMed data: Image and riparian areas using Landsat ETM+ imagery and decision-tree-based
segmentation and signature interpretation,” Remote Sens. Environ., models,” Wetlands, vol. 26, no. 2, pp. 465–474, 2006.
vol. 115, no. 4, pp. 990–1002, 2011. [44] F. Zambon et al., “Application of classification methods for mapping
[23] D. C. Mason, M. S. Horritt, J. T. Dall’Amico, T. R. Scott, and P. D. Bates, Mercury’s surface composition: Analysis on Rudaki’s area,” in Proc.
“Improving river flood extent delineation from synthetic aperture radar EPSC-DPS Joint Meet., 2011, p. 782.
using airborne laser altimetry,” IEEE Trans. Geosci. Remote Sens., vol. 45, [45] D. Chakraborty, V. Chowdhary, D. Dutta, and J. Sharma, “Classification
no. 12, no. 12, pp. 3932–3943, Dec. 2007. of high spatial resolution image using multi circular local binary pattern
[24] Y. Cunjian, W. Siyuan, Z. Zengxiang, and H. Shifeng, “Extracting the and variance,” Int. J. Electron. Commun. Comput. Eng., vol. 4, no. 6,
flood extent from satellite SAR image with the support of topographic pp. 1648–1654, 2013.
data,” in Proc. Int. Conf. Info-tech. Info-net, 2001, pp. 87–92. [46] B. J. Irvin, S. J. Ventura, and B. K. Slater, “Fuzzy and ISODATA classi-
[25] G. Nico, M. Pappalepore, G. Pasquariello, A. Refice, and S. Samarelli, fication of landform elements from digital terrain data in Pleasant Valley,
“Comparison of SAR amplitude vs. coherence flood detection Wisconsin,” Geoderma, vol. 77, no. 2, pp. 137–154, 1997.
methods—A GIS application,” Int. J. Remote Sens., vol. 21, no. 8, [47] M. Galli, F. Ardizzone, M. Cardinali, F. Guzzetti, and P. Reichenbach,
pp. 1619–1631, 2000. “Comparing landslide inventory maps,” Geomorphology, vol. 94, no. 3,
[26] M. N. Jebur, B. Pradhan, and M. S. Tehrany, “Detection of vertical slope pp. 268–289, 2008.
movement in highly vegetated tropical area of Gunung pass landslide, [48] D. C. Duro, S. E. Franklin, and M. G. Dubé, “A comparison of pixel-
Malaysia, using L-band InSAR technique,” Geosci. J., vol. 18, no. 1, based and object-based image analysis with selected machine learning
pp. 61–68, 2013. algorithms for the classification of agricultural landscapes using SPOT-5
[27] B. Brisco, A. Schmitt, K. Murnaghan, S. Kaya, and A. Roth, “SAR HRG imagery,” Remote Sens. Environ., vol. 118, no. 1, pp. 259–272,
polarimetric change detection for flooded vegetation,” Int. J. Digit. Earth., 2012.
vol. 6, no. 2, pp. 103–114, 2013. [49] S. W. Myint, P. Gober, A. Brazel, S. Grossman-Clarke, and Q. Weng,
[28] M. S. Tehrany, B. Pradhan, and M. N. Jebu, “A comparative assessment “Per-pixel vs. object-based classification of urban land cover extraction
between object and pixel-based classification approaches for land use/land using high spatial resolution imagery,” Remote Sens. Environ., vol. 115,
cover mapping using SPOT 5 imagery,” Geocarto. Int., vol. 12, no. 2, no. 5, pp. 1145–1161, 2011.
pp. 1–19, 2013. [Online]. Available: http://dx.doi.org/10.1080/10106049. [50] W. Zhou, A. Troy, and M. Grove, “Object-based land cover classifi-
2013.768300. cation and change analysis in the Baltimore metropolitan area using
[29] J. Chen, X. Zhu, J. E. Vogelmann, F. Gao, and S. Jin, “A simple and multitemporal high resolution remote sensing data,” Sensors, vol. 8,
effective method for filling gaps in Landsat ETM+ SLC-off images,” no. 3, pp. 1613–1636, 2008.
Remote Sens. Environ., vol. 115, no. 4, pp. 1053–1064, 2011. [51] Y. Wang, J. Colby, and K. Mulcahy, “An efficient method for mapping
[30] C. Zhang, W. Li, and D. Travis, “Gaps-fill of SLC-off Landsat ETM+ flood extent in a coastal floodplain using Landsat TM and DEM data,”
satellite image using a geostatistical approach,” Int. J. Remote Sens., Int. J. Remote Sens., vol. 23, no. 18, pp. 3681–3696, 2002.
vol. 28, no. 22, pp. 5103–5122, 2007. [52] Y. Wang, “Using Landsat 7 TM data acquired days after a flood event
[31] V. Karathanassi, P. Kolokousis, and S. Ioannidou, “A comparison study to delineate the maximum flood extent on a coastal floodplain,” Int. J.
on fusion methods using evaluation indicators,” Int. J. Remote Sens., Remote Sens., vol. 25, no. 5, pp. 959–974, 2004.
vol. 28, no. 10, pp. 2309–2341, 2007. [53] N. Kussul, A. Shelestov, and S. Skakun, “Grid system for flood ex-
[32] B. Aiazzi, S. Baronti, M. Selva, and L. Alparone, “MS + pan im- tent extraction from satellite images,” Earth. Sci. Inf., vol. 1, no. 3/4,
age fusion by an enhanced Gram-Schmidt spectral sharpening,” in pp. 105–117, 2008.
Proc. 26th EARSeL Symp. “New Strategies Eur. Remote Sens.,” 2007, [54] A. Demirkesen, F. Evrendilek, S. Berberoglu, and S. Kilic, “Coastal
pp. 113–120. flood risk analysis using Landsat-7 ETM+ imagery and SRTM DEM: A
[33] D. C. Mason et al., “Flood detection in urban areas using TerraSAR- case study of Izmir, Turkey,” Environ. Monit. Assess., vol. 131, no. 1–3,
X,” IEEE. Trans. Geosci. Remote Sens., vol. 48, no. 2, pp. 882–894, pp. 293–300, 2007.
Feb. 2010. [55] L. Giustarini et al., “A change detection approach to flood mapping
[34] A. Hamedianfar and H. Z. Shafri, “Development of fuzzy rule-based pa- in urban areas using TerraSAR-X,” IEEE Trans. Geosci. Remote Sens.,
rameters for urban object-oriented classification using very high resolu- vol. 51, no. 4, pp. 2417–2430, Apr. 2013.
tion imagery,” Geocarto. Int., vol. 29, no. 1, pp. 1–25, 2013. [56] F. Yuan, K. E. Sawaya, B. C. Loeffelholz, and M. E. Bauer, “Land
[35] M. N. Jebur, H. Z. Mohd Shafri, B. Pradhan, and M. S. Tehrany, “Per- cover classification and change analysis of the Twin Cities (Minnesota)
pixel and object-oriented classification methods for mapping urban land Metropolitan Area by multitemporal Landsat remote sensing,” Remote
cover extraction using SPOT 5 imagery,” Geocarto. Int., vol. 12, no. 2, Sens. Environ., vol. 98, no. 2, pp. 317–328, 2005.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

[57] W. L. Stefanov, M. S. Ramsey, and P. R. Christensen, “Monitoring urban Biswajeet Pradhan received the B.Sc. degree (with
land cover change: An expert system approach to land cover classification honors) from Berhampur University, Berhampur,
of semiarid to arid urban centers,” Remote Sens. Environ., vol. 77, no. 2, India, the M.Sc. degree from the Indian Institute
pp. 173–185, 2001. of Technology (IIT), Bombay, India, the M.Tech.
[58] Y. Yajima, Y. Yamaguchi, R. Sato, H. Yamada, and W.-M. Boerner, degree in civil engineering from the IIT, Kanpur,
“POLSAR image analysis of wetlands using a modified four-component India, and the Dresden University of Technology,
scattering power decomposition,” IEEE Trans. Geosci. Remote Sens., Dresden, Germany, the Ph.D. degree in GIS and
vol. 46, no. 6, pp. 1667–1673, 2008. geomatics engineering from the Universiti Putra
[59] N. Short et al., “A comparison of TerraSAR-X, RADARSAT-2 and Malaysia, Serdang, Malaysia, and the Habilitation
ALOS-PALSAR interferometry for monitoring permafrost environments, in “remote sensing” from the Dresden University of
case study from Herschel Island, Canada,” Remote Sens. Environ., Technology, in 2011.
vol. 115, no. 12, pp. 3491–3506, 2011. From 2008 to 2010, he was a recipient of the Alexander von Humboldt
[60] D. Brunner, L. Bruzzone, and G. Lemoine, “Change detection for earth- Research Fellowship from Germany. He is currently a faculty member of the
quake damage assessment in built-up areas using very high resolution Department of Civil Engineering, Universiti Putra Malaysia. He has more than
optical and SAR imagery,” in Proc. IEEE IGARSS, 2010, pp. 3210–3213. 16 years of teaching, research, consultancy, and industrial experience. Out of
[61] D. C. Mason, L. Giustarini, J. Garcia-Pintado, and H. L. Cloke, “Detection his more than 330 articles, more than 240 have been published in Science
of flooded urban areas in high resolution synthetic aperture radar images Citation Index (SCI/SCIE) technical journals. He has written 2 books on GIS
using double scattering,” Int. J. Appl. Earth. Observ., vol. 28, no. 2, data compression and disaster management, edited 3 volumes, and written 12
pp. 150–159, 2014. book chapters. He is on the editorial board of many ISI journals. He specializes
[62] D. C. Mason, I. J. Davenport, J. C. Neal, G. P. Schumann, and in remote sensing, GIS application, and soft-computing techniques in natural
P. D. Bates, “Near real-time flood detection in urban and rural areas using hazard and environmental problems.
high-resolution synthetic aperture radar images,” IEEE Trans. Geosci. Dr. Pradhan was the recipient of the prestigious German Academic Ex-
Remote Sens., vol. 50, no. 8, pp. 3041–3052, Aug. 2012. change Research (DAAD) fellowship award, Saxony State Fellowship from
[63] G. J. P. Schumann, J. C. Neal, D. C. Mason, and P. D. Bates, “The accuracy 1999 to 2002, Keith Aurtherton Research Award, and Georg Forster Research
of sequential aerial photography and SAR data for observing urban flood Award from the German Government. Since March 2016, Professor Pradhan
dynamics, a case study of the UK summer 2007 floods,” Remote Sens. is serving as “Ambassador Scientist” for Alexander von Humboldt Foundation,
Environ., vol. 115, no. 10, pp. 2536–2546, 2011. Germany.
[64] F. M. Henderson and A. J. Lewis, “Radar detection of wetland ecosys-
tems: A review,” Int. J. Remote Sens., vol. 29, no. 20, pp. 5809–5835,
2008.
[65] P. Townsend, “Relationships between forest structure and the detection of Mahyat Shafapour Tehrany was born in Tehran,
flood inundation in forested wetlands using C-band SAR,” Int. J. Remote Iran, on July 29, 1985. She received the B.Sc. degree
from the Allameh Mohaddes Nouri, Mazandaran,
Sens., vol. 23, no. 3, pp. 443–460, 2002.
[66] V. Herrera-Cruz, F. Koudogbo, and V. Herrera, “TerraSAR-X rapid map- Iran, in 2007, the M.Sc. degree from the Universiti
ping for flood events,” in Proc. Int. Soc. Photogramm. Rem. Sens., Earth Putra Malaysia, Serdang, Malaysia, in 2013, and the
Imaging Geospatial Inf., 2009, pp. 170–175. Ph.D. degree in GIS and geomatics engineering from
the Universiti Putra Malaysia, in 2015.
[67] J. B. Henry, P. Chastanet, K. Fellah, and Y. L. Desnos, “Envisat multi-
polarized ASAR data for flood mapping,” Int. J. Remote Sens., vol. 27, Her areas of interest are flood, landslides, and
no. 10, pp. 1921–1929, 2006. ensemble modeling applications.
[68] R. Hostache et al., “Water level estimation and reduction of hy-
draulic model calibration uncertainties using satellite SAR images of
floods,” IEEE Trans. Geosci. Remote Sens., vol. 47, no. 2, pp. 431–441,
Feb. 2009.
Mustafa Neamah Jebur was born in Misan, Iraq, on
[69] S. Martinis et al., “Comparing four operational SAR-based water and
flood detection approaches,” Int. J. Remote Sens., vol. 36, no. 13, January 11, 1988. He received the B.Sc. degree from
pp. 3519–3543, 2015. the Allameh Mohaddes Nouri, Mazandaran, Iran, in
2007, the M.Sc. degree from the Universiti Putra
[70] C. Kuenzer et al., “Varying scaule and capability of Envisat ASAR-WSM,
TerraSAR-X ScanSAR and TerraSAR-X stripmap data to assess urban Malaysia, Serdang, Malaysia, in 2013, and the Ph.D.
flood situations: A case study of the Mekong delta in Can Tho province,” degree in GIS and geomatics engineering from the
Remote Sens., vol. 5, no. 10, pp. 5122–5142, 2013. Universiti Putra Malaysia in 2015.
His areas of interest are flooding, landslides, and
[71] C. Kuenzer, J. Huth, S. Martinis, L. Lu, and S. Dech, SAR Time Series
for the Analysis of Inundation Patterns in the Yellow River Delta, China. ensemble modeling applications.
New York, NY, USA: Springer-Verlag, 2015, pp. 427–441.