Pol. J. Environ. Stud. Vol. 29, No.

2 (2020), 1759-1769
DOI: 10.15244/pjoes/110447 ONLINE PUBLICATION DATE: 2019-07-17

Original Research
Water Body Detection Analysis Using NDWI
Indices Derived from Landsat-8 OLI

Emre Özelkan*

Department of Urban and Regional Planning, Faculty of Architecture and Design,

Çanakkale Onsekiz Mart University, Çanakkale, Turkey

Received: 24 February 2019

Accepted: 2 July 2019

Abstract

Normalized different water indices (NDWIs) derived from satellite images are commonly and
successfully utilized in surface water body detection and mapping. In this study, the water body detection
capability of three NDWI models (NDWI(Green, NIR), NDWI(Green, SWIR1) and NDWI(Green, SWIR2)) generated
using 28 multitemporal Landsat-8 OLI multispectral satellite images was analyzed for Atikhisar Dam
Lake, the only water source of Çanakkalecity’s central district in Turkey between 2013 and 2017. This
study focused on two important open research questions: (i) Which NDWI model produces the most
superior results? and (ii) How much does accuracy change in the use of 15 m and 30 m spatial resolution
satellite data? For the accuracy analysis, area values extracted from the NDWI models were compared
with in-situ lake area values as measured by the General Directorate of State Hydraulic Works (DSI).
The results of this study show that as the lake area grows, discrimination of water from other classes
is better with NIR region, and that the performance of NDWI(Green, NIR) is relatively better in terms of
lake expansion effect. Results also indicate that hydrometeorological factors such as precipitation and
evaporation and anthropogenic factors such as irrigation and daily consumption are decisive in lake area
variations. The order of accuracy from high to low was found to be NDWI(Green, NIR), NDWI(Green, SWIR1) and
NDWI(Green, SWIR2), and 15 m spatial resolution data generated better results than 30 m resolution.

Keywords: water body detection, remote sensing, NDWI, Landsat-8 OLI

Introduction precision water resources management is of much

greater importance – especially in the organization
Water is the most important resource for sustaining of agricultural irrigation activities [4, 5]. However,
life since it keeps the whole ecosystem alive and shapes effective water resource management is only possible
human civilization [1, 2]. Water resource management with continuous monitoring [6, 7]. Satellite remote
will be the most important issue in the future as it is sensing is functional technology for monitoring natural
today, especially for arid and semi-arid regions [3]. Due resources such as water bodies and enables a time-
to the increasing population and consequent increase in and cost-effective monitoring of water resources with
food requirements, global warming and climate change, reliable data [8-10].
Several satellite remote sensing methods such as
image classification, linear unmixing, single-band
thresholding and water index are available to determine
*e-mail: [email protected] water bodies [11-14]. Since linear unmixing and image
1760 Özelkan E.

classification depend on human expertise and comprise In terms of spatial resolution, many studies have
high computation and single-band thresholding based produced NDWI from either 15 m [12, 21] or 30 m
on limited information, water indices that can produce [20, 23] resolution data. For accuracy assessment of
more accurate, faster and easier information than methods, the general approach is to calculate the user’s
others are better at detecting water bodies [12]. The accuracy, producer’s accuracy, overall accuracy and
normalized difference water index (NDWI) introduced kappa coefficient of produced data using high spatial
by McFeeters [15] is one of the most commonly used resolution images of the corresponding water body or
water indices to detect open surface water bodies GPS measurement for ground control point selection, as
and was firstcreated by the green and near-infrared performed after image classification [17, 20].
(NIR) spectral bands of Landsat TM. Modified NDWI In this study, the water body detection capability of
(MNDWI) comprising the green and short-wave infrared the NDWI of McFeeters (NDWI(Green, NIR)) and MNDWI
(SWIR) bands of Landsat TM was proposed by Xu [16] of Xu (NDWI(Green, SWIR))were analyzed by using 28
and is another commonly used index for water body multitemporal Landsat-8 Operational Land Imager
detection. Another prominent water index was presented (OLI) multispectral satellite images of Atikhisar Dam
by Feyisa [17], whose automated water extraction index Lake in Çanakkale Province, Turkey between 2013
(AWEI) comprises five bands of Landsat TM (Blue, and 2017. Since Landsat-8 OLI has two SWIR bands,
Green, NIR, SWIR1 and SWIR2) and is composed NDWI(Green, SWIR1) and NDWI(Green, SWIR2) were examined
of two sub-indices (AVEIsh and AVEInsh). Many other individually. At the same time, the examined NDWIs
studies proposing new methods and evaluation of water were produced from both 15 and 30 m resolution data.
body detection indices are available in the literature [18- Unlike the general approach in accuracy assessment
20]. The common feature in the design of the generated (i.e., using high spatial resolution images, GPS, etc.),
water indices is the absorption of water in the infrared the lake area values extracted by NDWI models were
region [11]. NDWI and MNDWI produced from tested with in-situ measured lake area values. Lastly, the
Landsat images are at the top of the most-compared results were correlated with hydrometeorological factors
water indices in the literature [8, 12, 16]. While NDWI such as precipitation and evaporation and anthropogenic
benefits from the high reflectance in NIR of vegetation factors such as irrigation and daily consumption. Based
and soil features [19], MNDWI can better separate on the above assessments, in this study two important
built-up features from water [12]. While many under-researched issues were addressed: (i) Which
researchers have found NDWI(Green, SWIR) to besuperior to NDWI model produces the most superior results? and
NDWI(Green, NIR) [11, 12], others have found the opposite (ii) How much does the accuracy change in the use
[2,8]. This is because quality may vary depending on of 15 and 30 m spatial resolution satellite data? In the
the color, content and depth of the investigated water remainder of this paper the methods followed, including
body [23]. data, will first be presented. Thereafter, the paper will

Fig. 1. Study area.

Water Body Detection Analysis Using... 1761

proceed with a detailed discussion of the results. Lastly, south and east of the basin with maximum values of
the paper will be completed with a summary of the 908 m height and 45.7º slope. The area with the lowest
major findings and recommendations. height and slope is the plain where Sarıçay Creek flows
freely into Çanakkale Strait (Dardanelles) and reaches
its widest border between Atikhisar Dam Lake and the
Material and Methods sea. The longest stream line comprises Sarıçay Creek
within the basin, which is approximately 43 km, and 8.5
Study Area km of it lies within the dam.

The study area is Atikhisar Dam Lake within the Remote Sensing Data and Pre-Processing
borders of Çanakkale province in western Turkey
located between 26°31’2.22”-26°33’10.30” eastern Different dated 28 Level 1 U.S. Geological Survey
meridians and 40°7’36.31”-40°3’49.67” northern (USGS) Landsat-8 Operational Land Imager (OLI)
parallels (Fig. 1). Atikhisar Dam was built 1971-1975 as images were downloaded between 2013 and 2017 over
an earthfill body on Sarıçay Creek [24] at a height of the study area from the USGS’s Earth Explorer data
60 m above sea level and 11 km from Çanakkale city portal (Table 1). In selecting the remote sensing data,
center, with a maximum area of 3.8 km 2 and volume special attention was given to images with no cloud or
of 53.5 hm3 according to the General Directorate of fog over the study area. Since NDWI from reflectance
State Hydraulic Works (DSI). Atikhisar Dam is a multi- images can generate more accurate results than NDWI
purpose dam (supplying drinking water, irrigation, flood from DN value images [8], surface reflectance images
protection, etc.) and serves as the only water source of were utilized in this study. As an initial step, radiometric
the central district of Çanakkale [6]. The study area is corrections, including the top-of-atmosphere (TOA)
located in the subtropical Mediterranean climatic region reflectance transformation of images were performed
[25] and under the influence of the Marmara climate, using ENVI software [21]. After that, the atmospheric
which is a transition zone between the Black Sea and correction of images was performed using Quick
Mediterranean [26]. Especially in winter, winds from Atmospheric Correction (QUAC) of ENVI in order to
the north cause temperatures to fall and winds blowing obtain surface reflectance data [28]. Images were pan-
from the south bring rain to the region [27]. According sharpened in the 15m spatial resolution panchromatic
to the Turkish State Meteorological Service long term band using the nearest-neighbor diffusion-based
data, the total annual precipitation and mean monthly (NNDiffuse) pan sharpening algorithm [29]. Since
temperature are 616 mm and 15ºC. The rainiest month is this research was also concerned with the comparison
December, with 106.8 mm precipitation, and the driest of 15 m or 30 m spatial resolution data for water body
is August with 6.4 mm. The hottest month is July (25ºC) detection, three NDWI models were employed for both
and the coldest month is January (6.1ºC). Hydrologically, spatial resolutions. 15 and 30 m resolutions of NDWI
according to Aster GDEM (global digital elevation models will be referred as NDWI (15 m) (i.e. NDWI(Green,
map), the area of ​​Sarıçay Basin where Atikhisar Dam NIR)
(15 m), NDWI(Green, SWIR1) (15 m) and NDWI(Green, SWIR2)
Lake is located is 473.1 km 2. Sarıçay Basin does not (15 m)), and NDWI (30 m) (i.e., NDWI(Green, NIR) (30 m),
have a dominant aspect direction due to its basin NDWI(Green, SWIR1) (30 m) and NDWI(Green, SWIR2) (30 m))
characteristics. High and sloping areas are situated respectively inthe remainder of the paper. As a result,

Fig. 2. Thresholding of 2013.05.02 NDWI models.

1762 Özelkan E.

2 (15 and 30 m resolution data) × 3 (NDWI models) Table 1. List of Landsat-8 OLI images used.
× 28 images = 168 water index data were produced. Image Day Difference
The examined NDWI models based on McFeeters [15] No Path/Row Acquisition from In-Situ
and Xu [16] are as follows: Date Measurement Date
1 181/32 2013.05.02 1
(1)
2 182/32 2013.06.26 5
(2) 3 182/32 2013.08.29 3
4 182/32 2013.09.30 1
(3)
5 182/32 2013.10.25 7
6 181/32 2014.04.03 2
Green is the 3rd band, NIR is the 5th band, SWIR1
is the 6th band and SWIR2 is the 7th band of Landsat-8 7 182/32 2014.05.28 4
OLI. For distinguishing water and non-water areas, 8 182/32 2014.06.29 2
NDWI threshold values can be successfully designated
both manually and automatically [8, 11, 16, 18, 30]. 9 181/32 2014.08.25 7
In this study, manual designation was preferred in 10 182/32 2014.11.04 3
consideration of the water and non-water classes. For
11 182/32 2015.07.02 1
NDWI(Green, NIR), NDWI(Green, SWIR1) and NDWI(Green, SWIR2),
threshold values for resolution were found as 0, 0.04 12 181/32 2015.07.27 5
and 0.12, respectively (Fig. 2). The threshold values 13 181/32 2015.08.28 4
are indicated by an arrow in the histograms (Fig. 2).
On the histograms, the dark greypart on the right side 14 182/32 2015.12.25 7
of the threshold is water, and on the left is the 15 181/32 2016.03.07 6
distribution of non-water pixels. If there were still mixed
16 182/32 2016.03.30 2
pixels over the classes out of water, they were manually
cleaned. 17 182/32 2016.06.02 1
18 181/32 2016.06.27 4
Validation Data
19 181/32 2016.08.30 2
The NDWI model area results were evaluated 20 181/32 2016.10.01 0
by comparison with in-situ water area measurement
values of the Turkish General Directorate of State 21 181/32 2016.11.02 1
Hydraulic Works. All measurements were performed 22 182/32 2017.01.28 4
on the first day of the month. The acquisition dates
23 182/32 2017.04.02 1
of the satellite images cannot always be the same as
the lake measurement days. Deviations of data 24 182/32 2017.05.04 3
acquisition dates from lake measurement days are 25 181/32 2017.06.30 1
shown in Table 1. Note that deviations of up to 7 days
were included in the study. Additionally, meteorological 26 181/32 2017.09.02 1
data from the Turkish State Meteorological Service 27 181/32 2017.10.04 3
were used to interpret the model results.
28 182/32 2017.10.27 5
Data Analysis

The performance of the NDWI models was tested Results and Discussion
utilizing root mean square error (RMSE) and Pearson’s
correlation coefficients (R) to correlate the computed The water body detection capability of
data (lake area values derived from NDWI models) NDWI(Green, NIR), NDWI(Green, SWIR1) and NDWI(Green, SWIR2)
with validation data (lake area values derived from derived from Landsat-8 OLI was examined in this
in-situ measurements). As previously mentioned, each study. NDWIs were produced from both 15 and 30 m
of 6 different data sets’ (3 NDWI models with 2 resolution data. Each of the 6 data sets (3 NDWI models
different resolutions each) of water body area values with 2 different spatial resolutions each) were compared
were compared with validation data separately. Each with in-situ measured lake area values. Consequently,
data set contained 28 differently dated values and was findings were evaluated by considering the water input-
compared with the corresponding in-situ measurement output parameters.
values.
Water Body Detection Analysis Using... 1763

In-situ Measurements rainy season, the main reason for the water body
of the lake being narrow on 2017.10.27 was that the
When the used in-situ measured values are precipitation in September, the first month of the rainy
examined, we found that the lake generally reaches its period, declined by about 70% (i.e., a meteorological
greatest limits at the end of the rainy period and its drought) compared to the long-term data. The second
narrowest limits at the end of the dry period. However, reason was water used for agricultural irrigation in
dates for the minimum and maximum areas may vary, September. Conversely, the largest limit was formed on
and the most decisive elements are hydrometeorological May 2, 2013, when it was not yet the end of the rainy
factors such as precipitation and evaporation, and season, with 3.692 km2. The most important factor in
anthropogenic factors such as irrigation and daily this early formation was the excessive precipitation in
consumption. In evaluating the values of the data set, January, February and April, which is about twice the
the smallest area was formed on October 27, 2017 with long-term average.
2.135 km2. Although October is normally within the

Fig. 3. Spectral signatures of different land cover classes for 2013.05.02 and 2017.10.27.
1764 Özelkan E.

Lake Area and Environment Performance of NDWI Models

While the adjacent classes to the lake area are The area results of all NDWI models have a
trees, soil and meadows, the lake area consists of water high positive correlation with in-situ measured area
and mud varying seasonally due to meteorological values (Fig. 4). Moreover, all NDWI models produced
conditions and water consumption. NDWI models, maximum and minimum values on the same dates as
which are formed from sample spectral curves in-situ measurements. Even though the NDWI(Green, SWIR1)
(i.e.,spectral signatures) in and adjacent to the lake (30 m) and NDWI(Green, SWIR2)(15 m) generated the lowest
area, successfully produced water classes as positive and highest correlated resul ts, respectively, R values
and other classes as negative (Fig. 3). All classes were may not be determinative for detecting the most superior
verified by using Google Earth’s high spatial resolution NDWI model since the difference between minimum
images. Deep Water1, Deep Water2, Shallow Water, and (R = 0.987) and maximum (R = 0.991) correlations is
Very Shallow Water were the classes examined. Deep just 0.004. According to significance probabilities (p) of
Water classes were from almost the deepest point of the ANOVA (analysis of variance) (i.e., significance F (SF)),
lake. Deep Water1 from 2013.05.02 had the maximum all correlations have almost 100% confidence level.
lake area and the color of the water body was blue. It can be seen that the differences between the
Deep Water2 from 2017.10.27 had the minimum lake performance of the NDWI models can be distinguished
area and the color of the water was greenish-blue. Blue by the RMSE results (Fig. 4, Table 2). The lowest
water has lower reflectance and greenish-blue water has RMSE values were obtained by NDWI(Green, NIR), then
higher reflectance in visible bands. Shallow water has NDWI(Green, SWIR1), and the highest RMSE values were
higher reflectance in the green and red bands due to obtained by NDWI(Green, SWIR2). In each NDWI model,
the bottom effect and can still be classified as water. 15 m resolution generated more accurate data than the
The spectral signature of very shallow water behaves 30 m. The 15 m resolution of NDWI(Green, NIR), NDWI(Green,
like that of mud, having a low reflectance in visible SWIR1)
, and NDWI(Green, SWIR2) were 13.433%, 9.774% and
bands and high in infrared bands, and cannot be 3.646% better than the 30 m models, respectively. 15 m
assigned as belonging in the water class by all NDWI was expected to be far superior to what 30 m had done
models (Fig. 3). especially for NDWI with SWIR bands. These ratios

Fig. 4. Correlations between in-situ measurements and NDWI model values.

Water Body Detection Analysis Using... 1765

Table 2. Mean, standard deviation (Stdv) and RMSE values of complete data setof NDWI and in-situ measurement areas.

Resolution NDWI(Green, NIR) NDWI(Green, SWIR1) NDWI(Green, SWIR2) In-situ

(m) 15 30 15 30 15 30 Measurement

RMSE (km2) 0.068 0.076 0.133 0.148 0.192 0.199 -

Mean (km )2
2.955 2.969 3.032 3.050 3.107 3.115 2.932
Stdv (km )
2
0.412 0.415 0.483 0.481 0.487 0.486 0.430

may show that utilizing the pan-sharpening algorithm As the results of Table 2 indicate, RMSE increases
NNDiffuse is better for NIR. with increasing water body area. The correlations
On the other hand, RMSE values have a positive between the NDWI models’ values and the absolute
correlation witht he mean and standard deviation values. difference of the NDWI models’ output values from
When the mean and standard deviations increase, RMSE measured values were examined. The results that are
also increases. RMSE decreases as the measurement shown in Fig. 5 seem to confirm the correlation between
values ​​approach the mean and standard deviation values RMSE and lake area, especially for NDWI(Green, SWIR1)
(Table 2). It was determined that the mean and standard and NDWI(Green, SWIR2), with a significant correlation of
deviation values ​​closest to the in-situ measurements around 0.7 R. However, the NDWI(Green, NIR) model results
were reached with McFeeter’s NDWI and the farthest were less influenced than other NDWI models by area
values ​​were reached with Xu’s NDWI with SWIR2. enlargement, as seen in Fig. 5. On the other hand, for
The 15 m spatial resolution of NDWI(Green, NIR) 2013.05.02, the gradual and consecutive enlargement
produced the closest result with 0.0784% and of the lake area as a result of the model performances
4.186% deviations from the mean and standard from NDWI(Green, NIR) to NDWI(Green, SWIR2) can be seen in
deviation values, while the 30 m spatial resolution of Fig. 6. While NDWI(Green, NIR) (15m) was found to be
NDWI(Green, SWIR2) generated the farthest result with the best model with 3.606 km2, the calculated areas
6.241% and 13.023% deviations. increase and diverge consecutively from the in-situ

Fig. 5. Correlations between absolute difference and NDWI model values.

1766 Özelkan E.

measured values of 3.692 km2 (Fig. 6), where the NDWI(Green, NIR), which normally cannot sufficiently
faultiest is NDWI(Green, SWIR2) but close to NDWI(Green, suppress the reflectance from built-up areas. NDWI(Green,
SWIR1)
. For the smaller areas, the success of the models SWIR1)
and NDWI(Green, SWIR2) perform better for smaller
may not be sequential (i.e., best: NDWI(Green, NIR), mid: lake areas owing to less interaction with classes adjacent
NDWI(Green, SWIR1), and worst: NDWI(Green, SWIR2)) as they to the water body (Fig. 6).
do in large lakes (Fig. 6). In Fig 6, since the areas of In addition, it may be thought that classification
NDWI(Green, SWIR1) and NDWI(Green, SWIR2) are so close, color errors arising from this interaction are caused by
discrimination can be recognized in the outermost areas spatial resolution. However, where the area is large and
and in small amounts. the interaction between classes is high, the accuracy
The results indicate that when the lake expands, difference between the 15 m and the 30 m spatial
water is in contact with other land cover classes such resolution is similar to that in small areas.
as trees, meadows and soil on its border. High reflection If there is a structure that creates shadows in the
of vegetation and soil in NIR allows NDWI(Green, NIR) to area and the images are cloudy, this could cause an error
be less affected by this interaction on its border and to in determining the water body [2]. Although a shadow
give more accurate results. The lack of built-up classes effect originating from trees could not be determined in
adjacent to the water mass increases the accuracy of this study area; after expansion of the lake, whether the

Fig. 6. Results from 2013.05.02 and 2017.10.27 Landsat-8 OLI image (Fig. 3): a1) 2013.05.02 dated 15 m spatial resolution NDWI
model results, a2) 2013.05.02 dated 30 m spatial resolution NDWI model results, b1) 2017.10.27 dated 15 m spatial resolution NDWI
model results, and b2) 2017.10.27 dated 30 m spatial resolution NDWI model results (light grey to black in order of NDWI(Green, NIR),
NDWI(Green, SWIR1) and NDWI(Green, SWIR2), respectively).
Water Body Detection Analysis Using... 1767

water enters the forest areas or not may be determined –– On 2015.07.27, 5 days before the measured day, the
by field studies. area determined by NDWI was 0.112 km2 larger than
the in-situ measured lake area. Although there was
Effect of Data Acquisition Date no precipitation for 5 days, a total of 0.81 hm3 open
and Water Input-Output water surface evaporation and 3.43 hm3 irrigation
usage occurred in July 2015.
When Table 2 is examined again, it is seen that –– On 2015.12.25, 7 days prior to the measured day,
RMSE values of the 6 NDWI models correspond to the NDWI result was 0.166 km2 higher than the
2.285%, 2.592%, 4.536%, 4.980%, 6.548% and 6.787% measured figure. This was the maximum error
of the mean of the lake area (2.932 km2). Although determined in this study. During this 7-day period
errors are not very high, when the size of the study area, there was no precipitation, but constant consumption
uncomplicated topographical and formal structure and of the available water by the city took place.
several land cover classes in and adjacent to the lake –– On 2016.03.07, due to 19.8 mm of precipitation within
are considered, the accuracy could be higher. So far, 6 days, the expansion of the lake area since the
the causes of these errors have been sought from the beginning of the month caused an NDWI mapping
NDWI model (i.e., remote sensing) framework. As a error of 0.130 km2.
next step, the time difference between the imaging and –– On 2016.06.02, just 1 day ahead of the measured
measurement dates and the water input-output occurring day, the NDWI result was 0.083 km2 smaller than
at this time are also evaluated. Water input-output data the measured value. Initially, this error was more
were gathered from the General Directorate of State than expected, but when the ongoing agricultural
Hydraulic Works and the Turkish State Meteorological irrigation and evaporation due to summer conditions
Service. Table 1 includes the day difference between are taken into account, the result makes sense.
in-situ measurement date and acquisition date of the –– On 2016.06.27, although there was 33.5 mm
image. As already mentioned, all lake measurements precipitation between 2016.06.27 (the imaging day)
were made on the first day of the month. Only one of and the beginning of the month, the difference
the image acquisitions is the same date as the in-situ was 0.096 km2 under the expected error. This may
measurement. The images with different dates were be due to the 0.7 hm3 of monthly evaporation in June
utilized up to 7 days, and the mean day difference from 2016.
the measurement date is 3 days. Sometimes the error The examples cited above, as well as the quality
may be high even though the time difference is low, of the remote sensing data and the preferred NDWI
while in some cases it is vice versa. model, not only emphasize the importance of the timing
Some of the remarkable results from the superior between satellite imaging and in-situ measurement,
model NDWI(Green, NIR) (15 m), whose absolute differences but also the enormous influence of water input-output
of the NDWI model’s output values from the measured that occurred in the possible time difference between
values were greater than 0.068 km2 RMSE, were satellite imaging and in-situ measurement of the lake.
examined as follows. Firstly, however, it is worth
remembering that this dynamic and rapidly changing
dam lake is the only water source of the city and its Conclusions
usage is multipurpose.
–– On 2013.05.02, although there was no precipitation Water indices derived from satellite data are
for 1 day, the area determined by NDWI was effectively used in water resource management. This
0.086 km2 lower than the in-situ measured lake case study analyzed the water body detection capabilities
area. Since the dam area was around its maximum, of three NDWI models: NDWI(Green, NIR), NDWI(Green, SWIR1)
water was probably released. According to the dam’s and NDWI(Green, SWIR2)of Landsat-8 OLI imagery at the
operation report, the water outlet from the bottom Atikhisar Dam Lake in Çanakkale Province, Turkey.
was 1.81 hm3 in May 2013. The effect of spatial resolution on NDWI performance
–– On 2013.10.25, 7 days before measurement day, the was tested by using both the 15 and the 30 m resolution
NDWI result was 0.098 km2 larger than the measured data. Unlike generally used methods, the accuracy
amount. This error is less than expected since this assessment of the detected water areas by NDWI was
7-day period had no precipitation and the evaporation not performed with common classification accuracy
was low due to autumn temperatures. assessment methods; rather, the detected area values
–– On 2014.11.04, NDWI produced a water area that were tested with in-situ measured values. Consequently,
was 0.129 km2 larger than the value measured 3 days the results were assessed by considering water input-
before. The precipitation was just 0.6 mm during this output parameters. The main findings of this study are
3-day period and insufficient for feeding the lake. as follows.
On the other hand, 111 mm of precipitation occurred –– NDWI(Green, NIR) was the best NDWI model for
in September and October 2014, which probably detecting the water body.
contributed to increasing underground water, a vital –– The fact that the structure of the lake and its border
source for the lake area. is completely natural (i.e., water, trees, meadows,
1768 Özelkan E.

mud and soil) was a major factor in the success of 4. GENC L., DEMİREL K., ÇAMOGLU G., ASIK S.,
NDWI(Green, NIR). SMITH S. Determination of plant water stress using
–– When the lake area increases and water interacts spectral reflectance measurements in watermelon (citrullus
with the lake boundary land cover classes, the vulgaris). American-Eurasian Journal of Agricultural &
Environmental Sciences, 11 (2), 296, 2011.
performances of NDWI(Green, SWIR1) and NDWI(Green,
5. ÇAMOĞLU G., DEMİREL K., GENC L. Use of
SWIR2)
were more affected than NDWI(Green, NIR). infrared thermography and hyperspectral data to detect
–– All data sets of the NDWI models with a spatial effects of water stress on pepper. Quantitative InfraRed
resolution of 15 m produced better results than 30 m. Thermography Journal, 15 (1), 81, 2018.
–– The time differences between remote sensing data 6. ÖZELKAN E., KARAMAN M. The analysis of the
acquisition and measurement dates can increase the effect of meteorological and hydrological drought on
water body area detection error in dynamic lakes dam lake via multitemporal satellite images: a case study
such as the one studied. in Atikhisar Dam Lake (Çanakkale). Omer Halisdemir
–– The water input-output, i.e., hydrometeorological University Journal of Engineering Sciences, 7 (2), 1023,
2018.
factors such as precipitation and evaporation and
7. KARAMAN M., BUDAKOGLU M., UCA AVCI Z.D.,
anthropogenic factors such as irrigation and daily ÖZELKAN E., BULBUL A., CIVAS M., TASDELEN
consumption, was found to be decisive in these S. Determination of seasonal changes in wetlands using
errors. CHRIS/Proba Hyperspectral satellite images: A case study
In this study, a comprehensive analysis of NDWI from Acigöl (Denizli), Turkey. Journal of Environmental
models was performed within an existing limited data Biology, 36, 73, 2015.
set. Future studies may be more comprehensive by 8. LIU Z., YAO Z., WANG R. Assessing methods of
identifying open water bodies using Landsat 8 OLI
utilizing more lake data from other parts of Turkey
Imagery. Environ Earth Sci, 75, 873, 2016.
and assessing new water indices. As a next step, the 9. KARAMAN M., ÖZELKAN E., TASDELEN S. Influence
effect of pan-sharpening and atmospheric correction on of basin hydrogeology in the detectability of narrow
spatial and spectral information and different algorithms rivers by Sentinel2-A satellite images: A case study in
may be tested for image processing. Additionally, Karamenderes (Çanakkale). Journal of Natural Hazards
statistical accuracy (i.e., comparing the NDWI’s area and Environment, 4, 140, 2018.
with in-situ measurement area) and thematic accuracy 10. KALE S., ACARLI D. Shoreline change monitoring in
(i.e., comparing the NDWI classification image with Atikhisar reservoir by using remote sensing and geographic
a high-resolution image) of the water area may be information system (GIS). Fresenius Environmental
Bulletin, 28 (5), 4329, 2019.
evaluated.
11. JI L., ZHANG L., WYLIE B. Analysis of dynamic
thresholds for the normalized difference water index.
Photogrammetric Engineering & Remote Sensing, 75 (11),
Acknowledgements 1307, 2009.
12. DU Z., LI W., ZHOU D., TIAN L., LING F., WANG H.,
The author would like to thank the United States GUI Y., SUN B. Analysis of Landsat-8 OLI imagery for
Geological Survey for the Landsat-8 OLI images, the land surface water mapping. Remote Sensing Letters, 5
General Directorate of State Hydraulic Works for the (7), 672, 2014.
hydrological data, and the Turkish State Meteorological 13. GÜRSOY Ö., ATUN R. Investigating surface water
pollution by integrated remotely sensed and field spectral
Service for the meteorological data.
measurement data: A case study. Polish Journal of
Environmental Studies, 28 (4), 2139, 2019.
14. GÜRSOY Ö., BIRDAL A., ÖZYONAR F., KASAKA
Conflict of Interest E. Determining and monitoring the water quality of
Kizilirmak River of Turkey: First results. ISPRS -
The authors declare no conflict of interest. International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences, XL-7/W3,
1469, 2015.
References 15. MCFEETERS S.K. The use of normalized difference water
index (NDWI) in the delineation of open water features.
International Journal of Remote Sensing, 17, 1425, 1996.
1. FENG M., SEXTON J.O., CHANNAN S., TOWNSHEND
16. XU H. Modification of normalised difference water index
J.R. A global, high-resolution (30-m) inland water body
(NDWI) to enhance open water features in remotely
dataset for 2000: first results of a topographic–spectral
sensed imagery. International Journal of Remote Sensing,
classification algorithm. International Journal of Digital
27 (14), 3025, 2006.
Earth, 9 (2), 113, 2016.
17. FEYISA G.L., MEILBY H., FENSHOLT R., PROUD S.R.
2. ACHARYA T.D., LEE D.H., YANG I.T., LEE J.K.
Automated water extraction index: A new technique for
Identification of water bodies in a Landsat 8 OLI image
surface water mapping using Landsat imagery. Remote
using a J48 decision tree. Sensors, 16 (7), 1075, 2016.
Sensing of Environment, 140, 23, 2014.
3. DEMIREL K., KAVDIR Y. Effect of soil water retention
18. JI L., GENG X., SUN K., ZHAO Y., GONG P. Target
barriers on turfgrass growth and soil water content.
detection method for water mapping using Landsat 8 OLI/
Irrigation Science, 31 (4), 689, 2013.
TIRS imagery. Water, 7 (2), 794, 2015.
Water Body Detection Analysis Using... 1769

19. KO B.C., KIM H.H., NAM J.Y. Classification of potential 25. ALTAN G, TÜRKEŞ M. Hydroclimatologic
water bodies using Landsat 8 OLI and a combination of Characteristics of the forest fires occurred at the Çanakkale
two boosted random forest classifiers. Sensors, 15 (6), district and relationship with climate variations. Aegean
13763, 2015. Geographical Journal, 20 (2), 1, 2015.
20. MISHRA K., PRASAD P.R.C. Automatic extraction of 26. SENSOY S., DEMIRCAN M., ULUPINAR Y., BALTA
water bodies from Landsat Imagery using perceptron Z. 2008. Climate of Turkey. Turkish State Meteorological
model. Journal of Computational Environmental Sciences, Service Report. Available online: https://www.mgm.gov.
2015 (903465), 1, 2015. tr/FILES/genel/makale/13_turkiye_iklimi.pdf (accessed on
21. YANG Y., LIU Y., ZHOU M., ZHANG S., ZHAN W., 12.10.2018) [In Turkish].
SUN C., DUAN Y. Landsat 8 OLI image based terrestrial 27. ILGAR R. Drought status and trends in the Dardanelles
water extraction from heterogeneous backgrounds using a and the standardized precipitation index determination.
reflectance homogenization approach. Remote Sensing of Marmara Geographical Review, 0 (22), 183, 2010.
Environment, 171, 14, 2015. 28. BERNSTEIN L.S., ADLER-GOLDENS.M., SUNDBERG
22. SHENG Y., SONG C., WANG J., LYONS E.A., KNOX R.L., LEVINE R.Y., PERKINS T.C., BERK A.,
B.R., COX J.S., GAO F. Representative lake water extent RATKOWSKI A.J., FELDE G., HOKE M.L. Validation
mapping at continental scales using multi-temporal of the QUick atmospheric correction (QUAC) algorithm
Landsat-8 imagery. Remote Sensing of Environment, 185, for VNIR-SWIR multi- and hyperspectral imagery. Proc.
129-, 2016. SPIE, 5806, 668, 2005.
23. FISHER A., FLOOD N., DANAHER T. Comparing 29. SUN W., CHEN B., MESSINGER D. Nearest-neighbor
Landsat water index methods for automated water diffusion-based pan-sharpening algorithm for spectral
classification in eastern Australia. Remote Sensing of images. Opt. Eng. 53 (1), 013107(1-11), 2014.
Environment, 175, 167, 2016. 30. OZTURK D., SESLI F.A. Determination of temporal
24. AKBULUT M., ODABASI D.A., KAYA H., CELIK E.S., changes in the sinuosity and braiding characteristics of the
YILDIRIM M.Z., ODABASI S., SELVI K. Changing of Kizilirmak River, Turkey. Pol. J. Environ. Studies, 24 (5),
mollusca fauna in comparison with water quality: Saricay 2095, 2015.
Creek and Atikhisarreservoir models (Canakkale-Turkey)”,
Journal of Animal and Veterinary Advances, 8 (12), 2699,
2009.