Available online at www.sciencedirect.

com

ScienceDirect
Journal of Taibah University for Science 11 (2017) 381–391

Water body extraction and change detection using time series:

A case study of Lake Burdur, Turkey
Gulcan Sarp a,∗ , Mehmet Ozcelik b
a Department of Geography, Suleyman Demirel University, 32260 Isparta, Turkey
b Department of Geological Engineering, Suleyman Demirel University, 32260 Isparta, Turkey
Received 2 February 2016; received in revised form 14 April 2016; accepted 25 April 2016
Available online 13 May 2016

Abstract
In this study, spatiotemporal changes in Lake Burdur from 1987 to 2011 were evaluated using multi-temporal Landsat TM and
ETM+ images. Support Vector Machine (SVM) classification and spectral water indexing, including the Normalized Difference
Water Index (NDWI), Modified NDWI (MNDWI) and Automated Water Extraction Index (AWEI), were used for extraction of
surface water from image data. The spectral and spatial performance of each classifier was compared using Pearson’s r, the
Structural Similarity Index Measure (SSIM) and the Root Mean Square Error (RMSE). The accuracies of the SVM and satellite-
derived indexes were tested using the RMSE. Overall, SVM followed by the MNDWI, NDWI and AWEI yielded the best result
among all the techniques in terms of their spectral and spatial quality.
Spatiotemporal changes of the lake based on the applied method reveal an intense decreasing trend in surface area between 1987
and 2011, especially from 1987 to 2000, when the lake lost approximately one fifth of its surface area compared to that in 1987.
The results show the effectiveness of SVM and MNDWI-based surface water change detection, particularly in identifying changes
between specified time intervals.
© 2016 The Authors. Production and hosting by Elsevier B.V. on behalf of Taibah University. This is an open access article under
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: Support vector machine; Normalized difference water index; Modified NDWI; Automated water extraction index; Change detection

1. Introduction coastline change and erosion monitoring, flood pre-

diction and evaluation of water resources [1]. Timely
Water body extraction is an important task in differ- monitoring of surface water and delivering data on the
ent disciplines, such as lake coastal zone management, dynamics of surface water are essential for policy and
decision-making processes [2]. In recent years, integra-
tion of remote sensing data with Geographic Information
∗ Systems (GIS) has been used in automatic or semi-
Corresponding author. Tel.: +90 2462114332.
E-mail address: gulcansarp@gmail.com (G. Sarp). automatic water body extraction and mapping [3]. [4]
Peer review under responsibility of Taibah University. Automatically extracted shorelines from Landsat TM
and ETM+ multi-temporal images with subpixel pre-
cision techniques. [5] Developed an approach called
the GeoCoverTM Water bodies Extraction Method that
combines remote sensing and GIS to extract water bodies
http://dx.doi.org/10.1016/j.jtusci.2016.04.005
1658-3655 © 2016 The Authors. Production and hosting by Elsevier B.V. on behalf of Taibah University. This is an open access article under the
CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
382 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391

and study their abundance and morphometry. However, index and SVM classification. The spectral and spatial
automatic coastline extraction is a complex process due performances of the applied satellite-derived indexes and
to water saturated land transition zones at the land- SVM were evaluated with Pearson’s r and the Structural
water boundary [6,7]. To determine the spatially accurate Similarity Index Measure (SSIM). In the literature, many
coastline position, two methods have been explored: studies were performed to extract water bodies based on
image classification and spectral water indexing. Multi- satellite-derived indexes and evaluate the effectiveness
class support vector machine (SVM) classification for of the satellite-derived indexes. Until now there have
water body extraction and coastline detection has been been no spatial performance analysis applied to satellite-
commonly used by many researchers because it suc- derived indexes based on SSIM. Our study contributes to
cessfully minimizes errors and maximizes the geometric the effectiveness of the SSIM-based quality evaluation
characteristics of edge areas [8,9]. Additionally, it has of satellite-derived indexes. The SSIM analysis provides
shown considerable potential in the supervised classifi- a simple quantitative interpretation by comparing the
cation of remotely sensed data, requiring very limited correlations of luminance, contrast and structure locally
training [10]. However, several water-indexing meth- between images and averaging these quantities over the
ods for the extraction of water bodies from remotely entire image.
sensed data have been introduced by researchers. [11] Lake Burdur, which is located in SW Turkey, has
Introduced the Normalized Difference Water Index shrunk abruptly in recent decades. Therefore, regular
(NDWI) to extract water features from Landsat TM using and reliable measurements of the lake area are nec-
band 2 and band 4. [12] Introduced another NDWI for essary to monitor the dynamic changes of lake water
water extraction from Landsat TM using bands 3 and area for water resource balance analysis. Previous stud-
5. [11] Proposed a threshold value of zero to extract ies of the lake area were based on visual interpretation
surface water bodies from the raw digital Landsat val- and manual digitization of satellite data [20,21]. In this
ues, where all positive NDWI values were classified as study, the spatiotemporal changes of Lake Burdur from
water and negative values as non-water. However, this 1987 to 2011 are investigated based on SVM classifica-
threshold does not enable discrimination between built- tion and satellite-derived water body extraction indexes,
up surfaces and water pixels. Thus, [13] introduced the including NDWI, MNDWI and AWEI using Landsat
Modified Normalized Difference Water (MNDWI) for TM and ETM+ data. The performances of the applied
Landsat TM using bands 2 and 5. [14] Introduced the indexes were tested using Pearson’s r, the SSIM and the
Automated Water Extraction Index (AWEI) to improve Root Mean Square Error (RMSE). Overall, the SVM
water extraction accuracy in areas that include shadows and NDWI were found superior to other indexes. The
and dark surfaces. [15] Introduced a simple Enhanced approach is highly significant for time-series analyses
Water Index (EWI) based on the Modified Normalized of extracted shorelines using any number of Landsat
Difference Water Index (MNDWI). It can effectively dis- satellite images taken in different time intervals, and it
tinguish water surfaces from background information provides an important comparison that can be used to
such as desert, soil and vegetation. [16] Investigated investigate shoreline changes.
NDWI, MNDWI, NDMI, WRI, NDVI, and AWEI for
the extraction of surface water from Landsat data and 2. The study area and data
used a novel surface water change detection process
based on the principal components of multi-temporal Lake Burdur is located in southwest Turkey
NDWI. In the study, surface water was extracted from (Fig. 1a, b). The southern part of the lake is bound by
the indexes using the thresholding technique based on the tectonically active Burdur fault zone. Tectonically
the trial and error method. The performance of each influenced half graben morphology controls the amount
water body extraction process was tested using the over- and type of sediment supply and turbidite systems of the
all accuracy and kappa coefficient, and NDWI was found lake [21].
superior to other indexes. These indexes have also been The image data used in this study were taken from
previously tested in several applications, including sur- 28 August 1987 for Landsat TM, 1 August 2000 for
face water mapping [17,14], land use and land cover ETM+ and 19 August 2011 for Landsat TM+ (path 179,
change analyses [18] and ecological research [19]. row 034). The sub-scenes of the data were all free of
In this study, water body extraction techniques were clouds, except Landsat TM-2011.
applied to Lake Burdur to determine decreasing trends in For accuracy, high spatial resolution Google Earth
the lake surface area in specified time intervals. The study images were used for reference. The acquisition dates
focuses on the performance of each satellite-derived of the Google Earth reference data and Landsat TM and
 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391 383

Fig. 1. Location of Lake Burdur (Landsat TM 1987 false colour composites 4/3/2).

ETM images were matched to minimize errors in the using CC and the SSIM. The accuracies of the SVM and
lake surface water. satellite-derived index method were tested based on the
RMSE.
3. Methodology
3.1. Support Vector Machine classification
The histograms of TM images display a board range
of grey levels and display two different peaks for land SVM is a supervised learning system and is based
and water areas. Based on the histogram observations, on recent improvements in statistical learning theory
images were classified using the SVM and two classes [22]. [23] Developed an SVM for binary classification.
to differentiate the land–water boundary. In the spectral A number of studies have focused on the mathematical
water indexing process, a single number was derived formulation of SVMs [23,24].
from two or more spectral bands using an arithmetic An SVM splits classes with a decision surface that
operation. Based on the spectral characteristics, a suit- maximizes the boundaries between the classes. The sur-
able threshold of the index was then applied to image face is called the ideal hyperplane, and the data points
data to separate these two classes from each other. The closest to the hyperplane are deemed support vectors
pixels representing the coastline were converted into a (Fig. 2). The support vectors are the important elements
vector layer to determine the coastline boundary and of the training set [25,23,26].
enable calculations of the area and perimeter of the To execute an SVM, training data are needed. These
lake. The performance of each satellite-derived index data optimize the separation of the classes rather than
classifier was compared with those of other classifiers describing the classes themselves [27]. Using a Radial
384 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391

Fig. 2. Linear support vector machine example (modified from Burges (1998)).

Basis Function (RBF), class distributions with non-linear kernel; d is the polynomial degree term in the kernel
boundaries can be mapped in a high dimensional space function of the polynomial kernel; r is the bias term in
for linear separation [28]. Training the SVM with a the kernel functions of the polynomial and sigmoid ker-
Gaussian RBF requires setting two parameters: a reg- nels; and γ, d and r are user controlled parameters, as
ularization parameter that controls the trade-off between their correct definition significantly increases the SVM
maximizing the margin and minimizing the training error accuracy.
and kernel width. A small regularization parameter tends
to emphasize the margin and ignore the outliers in the 3.2. Spectral water indexes
training data. A large regularization parameter may over-
fit the training data. A comprehensive description of the Automatic coastline delineation is a complicated pro-
SVM parameters can be found in [24,22]. cess due to the presence of the water-saturated zone at
An SVM classifier includes four different types of the land–water boundary [6,7]. Several spectral water
kernels: linear, polynomial, RBF and sigmoid. The RBF indexes have been developed to extract water bodies
kernel works fine in most cases [29]. The mathematical from remotely sensed imagery, usually by calculating
illustration of each kernel is listed in Eqs. (1)–(4): the normalized difference between two image bands and
Linear : K(xi , xj ) = xit , xj (1) then applying an appropriate threshold to segment the
results into two classes (water and non-water features).
d In this study, satellite imagery-derived NDWI, MNDWI
Polynomial : K(xi , xj ) = (γxit xj + r) , γ>0
and AWEI are used to extract lake water bodies from TM
(2) and ETM images.

3.2.1. Normalized Difference Water Index (NDWI)

RBF : K(xi , xj ) = exp(−γ||xi − xj ||2 ), γ>0 The Normalized Difference Water Index (NDWI) was
(3) first suggested by [11] to detect surface waters in wetland
environments and measure surface water dimensions.
The NDWI for TM and ETM sensors is defined by
Sigmoid : K(xi , xj ) = tanh(γxit xj + r) Eq. (5).

(4) band2 − band4

NDWI = (5)
band2 + band4
where xi is ith support vector; xj is the jth training data
point; t is the smoothing parameter; K is the kernel func- As a result, water features have positive values and are
tion; || is the Euclidean norm; γ is kernel width in the enhanced. Vegetation and soil features usually have zero
kernel functions of all kernel types, except the linear or negative values and are suppressed [11].
 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391 385

3.2.2. Modified Difference Water Index (MNDWI) The Pearson’s r value between two images is defined in
The MNDWI method suggested by [13] has been Eq. (9):
commonly used and is a powerful index that can extract 
(Am,n − μA )(Bm,n − μB )
water bodies [30,31]. It is expressed by Eq. (6). r(A, B) =  m,n (9)
2
m,n (Am,n − μA ) Bm,n − μB )2
band2 − band5
MNDWI = (6)
band2 + band5 where μA and μB are the mean values of the two images
(A and B), respectively. Pearson’s r should be as close to
The resulting values representing the water features
one as possible. The difference between Pearson’s r val-
have positive values because of their higher reflectance in
ues will show how well the spatial quality is maintained
band 2 than in band 5, and non-water features have neg-
[32].
ative NDWI values [13]. A threshold value for MNDWI
(e.g., simply a value of zero) can be set to segment the
MNDWI results into two classes (water and non-water 3.3.2. Performance evaluation using the Structural
features). Similarity Index Measure
The SSIM determines the similarity between two
images by comparing the correlations of luminance,
3.2.3. Automated Water Extraction Index (AWEI) contrast and structure locally between the images and
The main aim of the AWEI is to maximize the averaging these quantities over the entire image.
separability of water and non-water pixels using band The luminance between the two signals is determined
differencing, addition and application of different using the mean intensity of the signals given in Eq. (10).
coefficients. Accordingly, two separate equations are The contrast is determined using the standard deviation
proposed to effectively suppress non-water pixels and presented in Eq. (11). Finally, the structure is determined
extract surface water with improved accuracy [14]. The using the correlation presented in Eq. (12). This index
mathematical definition of AWEI is given in Eqs. (7) was proposed by [33]. The SSIM values vary from zero
and (8): to one. Values close to one show the highest similarity
AWEInsh = 4x(ρband2 − ρband5 ) − (0.25xρband4 to the original images.
2μx μy + C1
+ 2.75xρband7 ) (7) l(x, y) = (10)
μ2x + μ2y + C1
2σx σy + C2
C(x, y) = (11)
AWEIsh = ρband 1 + 2.5xρband2 − 1.5x(ρband4 σx2 + σy2 + C2
+ ρband5 ) − 0.25xρband7 (8) σxy + C3
S(x, y) = (12)
σx σy + C 3
where ρ variables are the reflectance values of spectral
bands of Landsat 5 TM: band 1, band 2, band 4, band 5 In Eqs. (10)–(12), μx and μy are the sample means
and band 7. AWEInsh is formulated to effectively elimi- of x and y, respectively; σ x and σ y represent the sample
nate non-water pixels, including dark, built-up surfaces variances of x and y, respectively; and σ xy is the sample
in areas with urban backgrounds, and AWEIsh further correlation coefficient between x and y. x and y refer to
improves the accuracy by removing shadow pixels that local windows in images X and Y, respectively. Constants
AWEInsh may not effectively eliminate [14]. C1 , C2 and C3 are used to stabilize the algorithm when
the denominators approach zero. The SSIM (x, y) is a
3.3. Performance evaluations of spectral water multiplication of these three components, as presented
indexes in Eq. (13).
(2μx μy + C1 )x(2σxy + C2 )
In this study, the performances of spectral water SSIM(x, y) = (13)
(μ2x + μ2y + C1 )x(σx0 + σy2 + C2 )
indexes for water body extraction were tested using Pear-
son’s r and the SSIM. 3.4. Validation of results using Root Mean Square
Error
3.3.1. Performance evaluation using Pearson’s r
Pearson’s r is a statistical measure of the strength and The RMSE is used to measure the difference between
direction of a linear association between two images. values predicted by a model and actual values. These
386 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391

Fig. 3. LANDSAT TM image from 1987 (a); LANDSAT-7 ETM+ image from 2000 (b); LANDSAT TM image from 2011; SVM-based extracted
water body from the 1987 LANDSAT TM image (d), LANDSAT-7 ETM+ image from 2000 (e); LANDSAT TM image from 2011 (f).

individual differences are also called residuals, and the the resultant images because the study focuses on water
RMSE serves to aggregate them into a single measure of body areas (Fig. 3d–f).
predictive power.
The RMSE of a model prediction with respect to the
4.2. Water body extraction using spectral water
estimated variable X model is defined as the square root
indexes
of the mean squared error (14):

n The three spectral water indexes (NDWI, MNDWI
i=1 (Xobs,i − Xmodel,i )
2
RMSE = (14) and AWEI) are applied to the lake water area to high-
n light the differences between water and non-water areas
where Xobs,i represents the observed values of the ith (Fig. 4a–i). The NDWI separates water and non-water
observation and Xmodel,i represents the predicted values objects well, with water areas generally having values
at location i. greater than zero and vegetation areas having strong
negative values. The NDWI and MNDWI images are
classified into water and non-water using a threshold
4. Results and analysis
of zero [11]. The RMSE of the AWEI depends on the
applied threshold value. The optimal threshold value for
4.1. Water body extraction using SVM
the AWEI, as recommended by [14], varies from −0.15
to 0.045. In this study, the threshold value was set to zero
The image given below, which was acquired using
to provide consistency between all applied indexes.
the Landsat satellite (Fig. 3a–c), shows the changes in
the lake from 1987 to 2011. These images were classi-
fied into two classes: water and land. The classification 4.3. Testing performance of shoreline extraction
training samples were collected randomly from the rep-
resentative homogeneous areas. The RBF was selected as 4.3.1. Pearson’s r
the kernel method for SVM classification. This function The spectral qualities of the indexes were measured
works well in most cases and can handle linearly non- with Pearson’s r, as depicted in Table 1. The best corre-
separable problems [29]. γ was determined as the inverse lation between the indexes shows the highest Pearson’s
of the number of bands in the input image, and 1000 was r value. The highest Pearson’s r was observed between
taken as the value of the regularization parameter. After the NDWI and MNDWI in 1987, 2000 and 2011, with
SVM classification, non-water areas were masked from values of 0.96, 0.91 and 0.96, respectively. The lowest
 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391 387

Fig. 4. Image-derived spectral water indexes of the (a–c) NDWI; (d–f) MNDWI; and (g–i) AWEI for Lake Burdur.

Table 1
Pearson’s r between the NDWI, MNDWI and AWEI in 1987, 2000 and 2011.
Pearson’s r in 1987 Pearson’s r in 2000 Pearson’s r in 2011

NDWI MNDWI AWEI NDWI MNDWI AWEI NDWI MNDWI AWEI

NDWI 1 0.96 0.94 1 0.91 0.90 1 0.96 0.87

MNDWI – 1 0.95 – 1 0.91 – 1 0.93
AWEI – – 1 – – 1 – – 1

Pearson’s r values were observed between the NDWI structural similarities, with values of 96%, 87% and 95%
and AWEI in 1987, 2000 and 2011, with values of 0.94, in 1987, 2000 and 2011, respectively. The SSIM was
0.90 and 0.87, respectively. the lowest between NDWI and AWEI, with values of
96%, 82% and 94% in 1987, 2000 and 2011, respectively.
4.3.2. Structural Similarity Index (SSIM) The differences between the applied indexes are given in
The structural similarity among spectral water index- Fig. 5. In this figure, white pixels represent no difference
ing was measured with SSIM. Values close to one show and black pixels indicate maximum difference.
the highest similarity between indexes. All the applied
indexes had SSIM values higher than 82%, verifying the 4.3.3. Geometric accuracy assessment and
fact that there was high similarity among the applied comparison
water indexes. As seen in Table 2, the SSIM rate indi- To compare the derived lake water surface areas,
cates that the NDWI and MNDWI provide the highest the lake water area was digitized manually on-screen
388 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391

Table 2
SSIM between the NDWI, MNDWI and AWEI in 1987, 2000 and 2011.
SSIM in 1987 SSIM in 2000 SSIM in 2011

NDWI MNDWI AWEI NDWI MNDWI AWEI NDWI MNDWI AWEI

NDWI 1 0.96 0.96 1 0.87 0.82 1 0.95 0.94

MNDWI 1 0.99 1 0.85 1 0.94
AWEI 1 1 1

Fig. 5. SSIM map of the spectral water indexes derived from a LANDSAT TM image from 1987 (a–c); LANDSAT-7 ETM+ image from 2000 (d–f);
LANDSAT TM image from 2011 (g–i).

from the Landsat images. High-resolution Google Earth digitizing the lake perimeter using the Landsat images
images were used as references to help differentiate con- and Google Earth images. The RMSE was calculated
fusing water pixels from background features. Many of using the reference data, as shown in Table 3. The accu-
the pixel values of land and water areas were mixed racies achieved by the SVM and MNDWI in 1987, 2000
near the lake shoreline. To compare the Landsat data and 2011 are higher than those of the NDWI and AWEI
with data obtained from Google Earth, image-to-image classifiers. For the SVM classifier, the RMSE ranges
registration was performed to geometrically align two between 33.14 m and 50.48 m. For the NDWI classi-
images. The image-derived coastlines were compared fier, the RMSE ranges between 51.94 m and 61.62 m.
to reference data, which was produced by manually The water body was extracted from satellite images with
 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391 389

Fig. 6. Changes in the area of Lake Burdur in 1987, 2000 and 2011 generated using (a) SVM; (b) NDWI; (c) MNDWI; (d) AWEI.

Table 3 Table 4
RMSE between reference data and image-derived shorelines. Water body area change of Lake Burdur in 1987, 2000 and 2011.
SWM NDWI MNDWI AWEI SVM NDWI MNDWI AWEI

RMSE (m) 1987 50.48 61.62 45.67 64.36 Area in 1987 (km2 ) 203.55 203.10 203.17 203.59
RMSE (m) 2000 33.14 51.94 50.44 109.81 Area in 2000 (km2 ) 156.19 157.12 157.49 161.76
RMSE (m) 2011 48.16 55.77 45.99 60.95 Area in 2011 (km2 ) 142.18 141.40 141.68 142.32
Change 1987–2000 (%) 23.27 22.64 22.48 20.55
Change 2000–2011 (%) 8.97 10.01 10.04 12.02
Total change 1987–2011 (%) 32.24 32.65 32.52 32.57
a 30 m spatial resolution. Therefore, these errors corre-
spond to approximately one and two pixels in the satellite
image.
1987. However, from 2000 to 2011, the lake lost about
4.3.4. Evaluation of the change one tenth of its surface area compared to that in 2000
The lake water area was extracted using SVM clas- (Table 4). According to the results given in Fig. 6a–d,
sification and spectral water indexing in 1987–2000 and the highest rate of water area change was observed in
2011, as shown in Fig. 6. The shoreline shrunk in the the NE part of the lake.
twenty-four year span of images. According to the SVM,
NDWI, MNWI and AWEI, between 1987 and 2000 5. Discussion
there were dramatic changes in lake water area, and the
changes between 2000 and 2011 were compared. The Water body extraction accuracy problems may be
SVM, NDWI, MNWI and AWEI results reveal that the particularly noticeable in areas where the background
surface area of Lake Burdur in August 1987 was approx- land cover includes low albedo surfaces (e.g., buildings,
imately 203 km2 (Table 4). By August 2000, based on asphalts, shadows and clouds). The presence of shadows
the NDWI and MNWI, the surface area had decreased in the images may cause misclassification due to simi-
to approximately 157 km2 . From 1987 to 2000, the lake lar spectral reflectance patterns as water body areas, and
lost about one fifth of its surface area compared to that in this similarity may decrease the accuracy of extracted
390 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391

surface water areas and vary the analysis between spec- to extract information regarding lake water area change
ified time intervals [13,3,34]. In environments with low is faster and more accurate than other observation meth-
spectral reflectance where non-water dark surfaces are ods, particularly in identifying changes between two and
found, simple classification methods may not adequately three different time intervals. The approach is based on
and accurately distinguish water pixels from non-water SVM classification and spectral water indexing (NDWI,
pixels, particularly in shadows [34]. Thus, three differ- MDWI and AWEI). According to the results of Pear-
ent spectral water indexes were used in this study. In son’s r, the SSIM and RMSE of the SVM classification
the NDWI, MNDWI and AWEI images, slight differ- and spectral water indexing, SVM and NDWI performed
ences cannot be manually detected by visual inspection. significantly better than did other indexes for mapping
Therefore, Pearson’s r and SSIM are measured among the lake water surface using Landsat data.
the three indexes. The SSIM analysis provides a simple Depending on these outcomes, from 1987 to 2000,
quantitative interpretation by comparing the correla- the lake lost about one fifth of its surface area, as com-
tions between luminance, contrast and structure locally pared to the surface area in 1987. From 2000 to 2011, the
between the images and averaging these quantities over lake lost about one tenth of its surface area compared to
the entire image. The major advantage of SSIM is that that in 2000. This study demonstrated the feasibility of
it is a simple and straightforward method for compar- estimating lake water area variations using only freely
ing two or more maps. The results reveal that SSIM and available satellite data.
Pearson’s r provide quality scores that are correlated to
RMSE values.
The results of the applied methodology suggest that Acknowledgements
no existing water index was able to automatically dif-
The authors would like to thank the anonymous
ferentiate water surfaces from shadow surfaces and low
albedo urban surfaces. The MNDWI is more appro- reviewers for their helpful and constructive feedback.
priate for differentiating water in many built-up areas
compared to using NDWI. The applied threshold values References
of the MNDWI used to obtain the best water extrac-
tion outcome are usually much less than those of the [1] Y.O. Ouma, R. Tateishi, A water index for rapid mapping of shore-
NDWI. Using zero as a default threshold value can line changes of five East African Rift Valley lakes: an empirical
analysis using Landsat TM and ETM+ data, Int. J. Remote Sens.
produce better water body separation accuracy using
27 (2006) 3153–3181.
the MNDWI. This would be very useful for more [2] C. Giardino, M. Bresciani, P. Villa, A. Martinelli, Applica-
accurately extraction of water bodies from image data tion of remote sensing in water resource management: the case
using the MNDWI. Additionally, the MNDWI gives study of Lake Trasimeno, Italy, Water Resour. Manag. 24 (2010)
more detailed information regarding open water than 3885–3899.
[3] H. Frey, C. Huggel, F. Paul, W. Haeberli, Automated detection of
does the NDWI. This property is also useful for the
glacier lakes based on remote sensing in view of assessing asso-
detection of water quality differences in specified time ciated hazard potentials, Grazer Schriften Geogr. Raumforsch. 45
intervals. (2010) 261–272.
The most critical aspect of the applied method is the [4] J. Pardo-Pascual, J. Almonacid-Caballer, L. Ruiz, J. Palomar-
accuracy assessment. For the accuracy test, a reference Vázquez, Automatic extraction of shorelines from Landsat TM
and ETM+ multi-temporal images with subpixel precision,
lake water area was digitized manually on-screen from
Remote Sens. Environ. 123 (2012) 1–11.
the Landsat images. This step is totally user-dependent [5] C. Verpoorter, T. Kutser, L. Tranvik, Automated mapping of water
and subjective. The literature suggests that the accuracy bodies using Landsat multispectral data, Limnol. Oceanogr.:
of the reference map is directly related to the experience Methods 10 (2012) 1037–1050.
of the user. The advantage of the extraction processes [6] J. Ryu, J. Won, K. Min, Waterline extraction from Landsat TM
data in a tidal flat: a case study in Gosmo Bay, Korea, Remote
in this study is that the study area is well known by the
Sens. Environ. 83 (2002) 442–456.
authors, who are familiar with the lake morphology of [7] S. Maiti, A. Bhattacharya, Shoreline change analysis and its
the area. application to prediction: a remote sensing and statistics based
approach, Mar. Geol. 257 (2009) 11–23.
6. Conclusions [8] R.K. Nath, S.K. Deb, Water-body area extraction from high reso-
lution satellite images – an introduction, review, and comparison,
Int. J. Image Process. 3 (2010) 353–372.
This study used satellite image interpretation and GIS [9] Z. Hannv, J. Qigang, X. Jiang, Coastline extraction using sup-
to detect and analyze the spatial changes and quantify the port vector machine from remote sensing image, J. Multimed.
water area change of Lake Burdur. Using satellite images 8 (2013) 2.
 G. Sarp, M. Ozcelik / Journal of Taibah University for Science 11 (2017) 381–391 391

[10] M. Dalponte, L. Bruzzone, D. Gianelle, Tree species classifi- [22] N. Cristianini, J. Shawe-Taylor, An Introduction to Support
cation in the Southern Alps based on the fusion of very high Vector Machines and Other Kernelbased Learning Methods,
geometrical resolution multispectral/hyperspectral images and Cambridge University Press, 2000, http://dx.doi.org/10.1017/
lidar data, Remote Sens. Environ. 123 (2012) 258–270. CBO9780511801389.
[11] S.K. McFeeters, The use of the normalized difference water index [23] C. Cortes, V. Vapnik, Support-vector network, Mach. Learn. 20
(NDWI) in the delineation of open water features, Int. J. Remote (1995) 273–297.
Sens. 17 (1996) 1425–1432. [24] C.J.C. Burges, A tutorial on support vector machines for pattern
[12] A.S. Rogers, M.S. Kearney, Reducing signature variability in recognition, Data Mining Knowl. Discov. 2 (1998) 121–167.
unmixing coastal marsh Thematic Mapper scenes using spectral [25] B.E. Boser, I.M. Guyon, V.N. Vapnik, A training algorithm for
indices, Int. J. Remote Sens. 25 (2004) 2317–2335. optimal margin classifiers, in: Proceedings of the Fifth Annual
[13] H. Xu, Modification of normalised difference water index ACM Conference on Computational Learning Theory, Pittsburgh,
(NDWI) to enhance open water features in remotely sensed 1992, pp. 144–152.
imagery, Int. J. Remote Sens. 27 (2006) 3025–3033. [26] G.M. Foody, D.S. Boyd, C. Sanchez-Hernandez, Mapping a spe-
[14] G.L. Feyisa, H. Meilby, R. Fensholt, S.R. Proud, Automated cific class with an ensemble of classifiers, Int. J. Remote Sens. 28
Water Extraction Index: a new technique for surface water map- (2007) 1733–1746.
ping using Landsat imagery, Remote Sens. Environ. 140 (2014) [27] G.M. Foody, A. Mathur, The use of small training sets containing
23–35. mixed pixels for accurate hard image classification: training on
[15] S. Wang, M.H.A. Baig, L. Zhang, H. Jiang, Y. Ji, H. Zhao, J. mixed spectral responses for classification by a SVM, Remote
Tian, A Simple Enhanced Water Index (EWI) for percent surface Sens. Environ. 103 (2006) 179–189.
water estimation using Landsat data, IEEE J. Sel. Top. Appl. Earth [28] C. Huang, L.S. Davis, J.R.G. Townshend, An assessment of sup-
Observ. Remote Sens. 8 (2015) 90–97. port vector machines for land cover classification, Int. J. Remote
[16] K. Rokni, A. Ahmad, A. Selamat, S.H. Rokni, Water feature Sens. 23 (2002) 725–749.
extraction and change detection using multitemporal landsat [29] ENVI Manual, 2004, Available at: http://aviris.gl.fcen.uba.ar/
imagery, Remote Sens. 6 (2014) 4173–4189. Curso SR/biblio sr/ENVI userguid.pdf.
[17] Z. Duan, W.G.M. Bastiaanssen, Estimating water volume vari- [30] L. Ji, L. Zhang, B. Wylie, Analysis of dynamic thresholds for the
ations in lakes and reservoirs from four operational satellite normalized difference water index, Photogramm. Eng. Remote
altimetry databases and satellite imagery data, Remote Sens. Sens. 7 (2009) 1307–1317.
Environ. 134 (2013) 403–416. [31] S. Lu, B. Wu, N. Yan, H. Wang, Water body mapping method with
[18] A. Davranche, G. Lefebvre, B. Poulin, Wetland monitoring using HJ-1A/B satellite imagery, Int. J. Appl. Earth Observ. Geoinform.
classification trees and SPOT-5 seasonal time series, Remote 13 (2011) 428–434.
Sens. Environ. 114 (2010) 552–562. [32] G. Sarp, Spectral and spatial quality analysis of pan-sharpening
[19] B. Poulin, A. Davranche, G. Lefebvre, Ecological assessment of algorithms: a case study in Istanbul, Eur. J. Remote Sens. 47
Phragmites australis wetlands using multi-season SPOT-5 scenes, (2014) 19–28.
Remote Sens. Environ. 114 (2010) 1602–1609. [33] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Simoncelli, Image quality
[20] M. Ataol, Burdur Gölü’nde Seviye Değişimleri, Coğrafi Bilimler assessment: from error visibility to structural similarity, IEEE
Dergisi CBD 8 (2010) 77–92 (in Turkish). Trans. Image Process. 13 (2004) 600–612.
[21] E. Şener, A. Davraz, T. İsmailov, Burdur Gölü Seviye [34] C. Verpoorter, T. Kutser, L. Tranvik, Automated mapping of water
Değişimlerinin Çok Zamanlı Uydu Görüntüleri ile İzlenmesi, in: bodies using Landsat multispectral data, Limnol. Oceanogr.-
Türkiye Kuvaterner Sempozyumu-V, İstanbul, 2005 (in Turkish). Methods 10 (2012) 1037–1050.