| 387 |
| 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
V
G 202
2
GEODETSKI VESTNIK | letn. / Vol. 66 | št. / No. 3 |
SI 
 |  
EN
KLJUČNE BESEDE KEY WORDS
remote sensing, image processing, water body, river,
corine data 
daljinsko zaznavanje, obdelava slik, vodno telo, reka, 
podatki Corine 
Water resources is crucial for the continuity of life. Therefore, 
mapping water resources is required. Successful analysis of 
remotely sensed images can provide reliable information 
for water researches. However, it is very complex process to 
ensure that the maps created are not affected by shadows, 
cloud or other noise. In addition, it is necessary to successfully 
map all water types in various geographies. It is important 
that the method used is practical so that scientists who are 
not image analysts can use the large data pool provided by 
satellite images.
In this paper, a novel algorithm for water body extraction 
from Landsat imagery is proposed. In this method, Corine 
data are used as auxiliary data to automatically generate 
training data. Four study areas with different characteristics, 
from different parts of the world, are used to test the proposed 
method. The results obtained are compared with other 
automatic classification methods. 
UDK: 004.652:528.8 
Klasifikacija prispevka po COBISS.SI: 1.01
Prispelo: 7. 2. 2022
Sprejeto: 8. 8. 2022
DOI: 10.15292/geodetski-vestnik.2022.03.387-402
SCIENTIFIC ARTICLE
Received: 7. 2. 2022
Accepted: 8. 8. 2022
AUTOMATIC DELINEATION 
OF WATER BODIES USING 
CORINE DATA FROM 
REMOTELY SENSED IMAGES
Dilek Küçük Matci
SAMODEJNA RAZMEJITEV 
VODNIH TELES Z UPORABO 
PODATKOV CORINE IZ 
DALJINSKO ZAZNANIH SLIK
Vodni viri so ključnega pomena za obstoj življenja, zaradi 
česar je pomembno, da jih kartiramo. Z uspešno analizo 
daljinsko zaznanih slik lahko pridobimo zanesljive 
informacije za raziskave vode, vendar je postopek zapleten 
in poskrbeti moramo, da na ustvarjene karte ne vplivajo 
sence, oblaki ali drugi šumi. Poleg tega je treba uspešno 
preslikati vse vrste voda na različnih geografskih območjih. 
Pomembno je tudi, da je uporabljena metoda praktična, da 
lahko znanstveniki, ki niso slikovni analitiki, uporabljajo 
širok nabor podatkov, ki jih zagotavljajo satelitske slike.
V prispevku je predlagan nov algoritem za ekstrakcijo 
vodnega telesa iz posnetkov Landsat. Pri tej metodi se 
podatki Corine uporabljajo kot pomožni podatki za 
samodejno ustvarjanje učnih podatkov. Za testiranje 
predlagane metode se uporabljajo štiri študijska področja z 
različnimi značilnostmi iz različnih delov sveta. Dobljene 
rezultate primerjamo z drugimi samodejnimi metodami 
razvrščanja.
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
ABSTRACT IZVLEČEK 
| 388 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
1 INTRODUCTION 
Climate change and rapid, uncontrolled urbanization in recent years have affected the distribution and 
quality of water areas (Franczyk and Chang, 2009; Huong and Pathirana, 2013). Sustaining a healthy 
environment, climate and agriculture depend on the protection of water resources. Regular monitoring 
of water resources and the identifiying changes are the first steps towards conservation of  resources. 
On the other hand, it is necessary to monitor water bodies in order to reveal floods and their effects on 
the environment as a result of the storms and excessive precipitation which are a result of the changing 
climate (Huong and Pathirana, 2013).
With the rapid development of technology, remotely sensed images have begun to provide data about 
environmental changes in many areas. In the past years, various water body extraction studies have 
been carried out using remotely sensed images (Chen et al., 2020; Feng et al., 2016; Monegaglia et 
al., 2018; Rad et al., 2021; Somasundaram et al., 2020). In one of these studies, an inland water body 
dataset was created using Landsat satellite images.  Landsat images are used in another water extraction 
study, surface water area extraction has been carried out by using learning vector quantization (LVQ) 
methods (Somasundaram et al., 2020).  Similarly, Landsat images were used to test theproposed method 
(Monegaglia et al., 2018). In another study a novel neural network method for water body extraction 
method was proposed. To evaluate this proposed method, ZY-3, GF-1, and GF-2 multispectral images 
were used (Chen et al., 2020). 
The water index, which is used to identify water areas, is a simple and effective method of identifying 
surface water bodies by separating the signals of water from other surface classes with the band ratio 
method. The normalized difference water index (NDWI) based on extracting green and near infrared 
(NIR) bands from multispectral bands has been proposed (McFeeters, 2013). The modified normalized 
difference water index (MNDWI) has been proposed to suppress background residential areas, while 
eliminating water areas (Hanqiu Xu, 2006). Recently, the Augmented Normalized Difference Water 
Index has been proposed to extract water bodies from Landsat images (Rad et al., 2021). Other water 
indices such as the automated water extraction index (AWEI), all band water index (ABWI) and multi-
band water index (MBWI) have been proposed using other bands in multispectral images (Feyisa et al., 
2014; Wang et al., 2018; Xiong et al., 2018).
One of the other water extraction approaches is supervised classification approaches. For example, a 
water and urban land use map was prepared is a random forest (RF) classifier (Chen et al., 2018). Also, 
two machine learning frameworks (pixel based and object based) have been proposed to determine water 
types (Huang et al., 2015). The monthly changes in the water surface area of  Lake were determined in 
another study using Landsat and Sentinel 1 images and the Otsu method (Jingzhe Wang et al., 2019). 
Using different satellite images, spectral indices, and thresholds, the spatial and volumetric changes of the 
Aral Lake between 1960 and 2018 have been determined in one of the recent studies (Yang et al., 2020). 
Unsupervised classification methods are frequently used in the literature for the extraction of water 
areas. In a study conducted in Sri Lanka, the Learning Vector Quantization (LVQ) method was used to 
extract the water area. The obtained results were compared with other state of the art methods such as 
support vector machine, K-nearest neighbor, discriminant analysis, combination of modified normalized 
difference water index and modified fuzzy clustering method, and K-means methods. The results of the 
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 389 |
GEODETSKI VESTNIK | 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
study indicated that LVQ performed well in  detecting all water types with fewer training samples (Soma-
sundaram et al., 2020). In another study, LVQ was used to extract water areas from Landsat 4 images 
(Wang and Zhu, 2003). In a study proposed to extract water regions in mountainous areas, the image 
was first classified with K-Means, and these classes were used as masks (Gao et al., 2016). High resolution 
Dem and K-Means were used in a study to extract valley and channel networks (Hooshyar et al., 2016).
When these studies are examined, supervised classification generally gives higher accuracy results, compa-
red to unsupervised classification or water indices (Jiang et al., 2018). However, creating separate training 
data for each classification in the supervised classification process is laborious and time-consuming. In 
addition, the accuracy of the training data directly affects the results (Foody and Mathur, 2004; Matci 
and Avdan, 2020). 
Corine data have been created for the purpose of monitoring natural resources, and prepared in line with 
the criteria determined by the European Environment Agency (Corine, 2019). The European Corine 
land cover mapping scheme has a 100m resolution and includes 44 land cover and land use classes. 
These data are used in various studies such as the determination of the changes in land cover classes, the 
creation of population density maps, the spatial interpolation of air pollution measurements, and the 
determination of urban heat islands  (Dabija et al., 2021; Dzieszko, 2014; Gallego and Peedell, 2001; 
Janssen et al., 2008; Kucsicsa et al., 2019; Stathopoulou and Cartalis, 2007; Waltner et al., 2020). In a 
study conducted using this data, the effectiveness of satellite image classification, according to the Corine 
land cover guidelines, was examined. Training data were collected by randomly selecting samples of pixels 
falling within the Corine land classes (Dabija et al., 2021). In another study, Corine data were used to 
reveal the main land-use and land-cover change (LUCC) patterns in the Poznań Lakeland mesoregion 
(Dzieszko, 2014). In a study to identify and predict land cover change in Romania, Corine data was used 
in conjunction with the existing land use/cover model(Kucsicsa et al., 2019). In a study conducted in 
Greece, an urban heat island extraction study was conducted and Corine data were used together with 
Landsat 7 images(Stathopoulou and Cartalis, 2007).
The Corine data were used in another study to improve the mapping of population density that the 
European Commission could use at the commune level (Gallego and Peedell, 2001). In another study, 
Corine data were used to model air pollution data(Janssen et al., 2008). Corine data were used in a study 
to evaluate changes in soil erosion potential caused by changes in land use and land cover based on Land 
Cover data in Hungary (Waltner et al., 2020).
This study uses a fully automate water extraction approach, from satellite images, using   data. The 
method uses a training data collection system based on the Corine data. With the training data prepared 
in this way, the water bodies in selected areas outside the Corine region were also successfully extracted 
from the Landsat Images. In order to test the method developed, images of regions in Canada, Guinea, 
Zhengzhou and Tibetan were used, and the water bodies in these areas were automatically mapped. The 
results obtained reveal the potential of the proposed method.
2 MATERIALS AND METHODS
The materials and methods used in the study are presented under the following sub-headings.
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 390 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
2.1 Satellite Imagery
Launched on February 11, 2013, Landsat-8 collection includes two sensors, the Operational Land Imager 
(OLI) and Thermal Infrared Sensors (TIRS). Landsat-8 OLI images were used in the study. The bands 
of the Landsat-8 satellite used in the study are given in Table1.
Table 1: Specifications of Images Used in the Study
Satellite Band Name Resolution (m) Wavelength(nm)
Landsat 8
Blue 30 483
Green 30 560
Red 30 660
NIR 30 865
SWIR 30 1650
SWIR 30 2220
In the study, 30m resolution Landsat images were used. The acquired dates of the images are given in 
Table2.
Table 2:  Acquisition Date of Used Images
Study Area Specification of the Area Landsat 8
Training Turkey, Greece, Swe-den, UK, France
Turkey: Sea, Lake, Urban areas, river, Agri-
cultural areas
Greece: Sea,
Fr: Muddy canal, 
Sweden: Surrounded by the trees.
Mean of im-
ages acquired 
between 01.01.2021-
09.12.2021
Classification
Loughborough Lake, 
Canada
It has different depths and is surrounded by 
trees. 24.08.2021
Guinea Region Muddy and the shores are covered with rhizophoraceae mangroves 03.06.2021
Lake yamzho Yumco, 
Tibet
The selected area is surrounded by snow-
capped mountains
28.01.2021
04.02.2021
Yellow River
Zhengzhou,China
The selected area has urban areas, water bod-
ies, mostly turbid rivers, artificial channels, 
artificial lakes, reservoirs and ponds.
02.10.2021
2.2 Study Areas
In order to test the proposed method, images of the regions given in Figure 1 were chosen. The reason 
for choosing these study regions is that the proposed method can be evaluated in various geographical 
conditions (mountainous, delta, urban, with and without vegetation).
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 391 |
GEODETSKI VESTNIK | 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Figure 1:  Study Areas a) Loughborough,Canada b)Dangara, Guinea d) Yamzho Yumco –Tibetan,China d) Yellow River Zhen-
gzhou,China
The first study area is Loughborough Lake, located in the Cataraqui River basin, 20 km north of Kingston, 
Canada. The size of Loughborough Lake is 2002.9 hectares. The lake is divided into two main basins with 
very different characteristics. The western basin is on limestone, is much deeper (maximum 38.4m), is 
colder and contains species such as trout. The eastern basin is on the Canadian Shield and is shallower 
(maximum 7.6m) and warmer, with many islands and shallows. Like most lakes in the Cataraqui region, 
Lake Loughborough mixes in spring and autumn as the lake water warms and cools. During mixing, 
nutrients are cycled throughout the lake and it may appear cloudy with a brown or green colour from 
algae using cyclic nutrients. The water temperature varies according to depth (Gehrke, 2021).
The second image used in the study is Dangara in the Boke region on the Atlantic coast of Guinea. The 
Dangara inlet is located 20 km west of the Fatala Estuary. A tropical monsoon climate is dominant in the 
region. The length of the Dangara branch is 30 km, and the width of its mouth is 1 km. The shores are 
covered with rhizophoraceae mangroves. The bottom is covered with mud. Seasonal flow variations are not 
very evident in Dangara. Fresh water inflow is minimal and marine influence is dominant (Baran, 1995).
The third image used in the study is Yamzho Yumco Lake, located in the fault basin of southern Tibet. 
The altitude of the lake surface is 4441m. The lake is branch-shaped, extending northwest to southeast, 
with a maximum length of 74 km, a maximum width of 33 km and an average width of 8.62 km. The 
structural lake contains brackish water. The lake is quite meandering, with many major branches, he-
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 392 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
adlands and bays. There are many islands and islets in the lake with a total area of 44 km2, the largest of 
which is 16.8 km2 (Shao et al., 2008).
The last image is the Yellow River, located in the Zhengzhou district in Yellow River, China. In Zhen-
gzhou the climate is warm, temperate, continental and monsoon, with four distinct seasons. The selected 
image has urban areas, water bodies, mostly turbid rivers (Yellow River), artificial channels, artificial 
lakes, reservoirs and ponds (Zhang et al., 2019).
2.3 Methods
The proposed method for automatic mapping of water bodies is given in Figure 2. The method has three 
stages: creation of the training pool and the training stage; image pre-processing stage; and classification 
stage. All operations in the method are carried out in the Google Earth Engine.
Figure 2: Workflow of the Proposed Method
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 393 |
GEODETSKI VESTNIK | 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
One factor that directly affects the accuracy of the classification process is the training data (Foody ve 
Mathur, 2004). Therefore, creating a strong training dataset is critical, which is why the training regions 
are chosen (Table 2). After filtering the Corine data of the areas, the land classes were determined, and 
the boundaries of these areas were vectorised. In the training phase of the algorithm all of the Corine 
classes are used. Ones the class borders are extracted from the Corine image, classes which had the id 
of 511 (Water courses), 512 (Water bodies), 521 (Coastal lagoons),522 (Estuaries) and 523 (Sea and 
ocean) were grouped into water class. All remaining classes were grouped into the non-water class.  Due 
to the resolution difference between the Corine data and Landsat data, a negative buffer of 200 m was 
applied, in order to get rid of the mixed classes at the edges of the Corine data (Figure 3).
Figure 3:  Created polygons from the Corine Data
Training data have been prepared in order to give successful results for water areas with different chara-
cteristics. The inclusion of classes found in the Corine data was accepted as a criterion when selecting 
the samples. In addition, versions of the same class in different geographies were also selected. Samples 
from Turkey, England, Greece, France and Sweden were selected as training data. Some of the selected 
sample regions are given in Figure 4.
At the end of the first stage, the satellite images in the training pool were filtered temporally and spati-
ally. The average of the images for the year containing the test date is taken as the training data. Because 
there are too many cloud problems in optical images. In addition, it is necessary to take into account 
the periodic changes of classes. For the test images, the mean of the images belonging to October 2021 
was used in this study.
 Then the cloud images were removed from the pool. The spectral indices were calculated and included 
in the training data in order to increase the classification accuracy. The spectral indexes NDVI, NDWI, 
MNDWI-1, MNDWI-2, AWEIsh and AWEInsh were used (Table 3).
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 394 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Figure 4: Training Areas a) Samples from the water bodies b) Samples from the other classes
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 395 |
GEODETSKI VESTNIK | 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Table 3:  Indices used in the study
Index Name Band math Feature extraction Reference
NDVI
Modified Normali-
zed Difference Vege-
tation Index
−
=
+
NIR REDNDVI
NIR RED
 Vegetation extraction
(DeFries and 
Townshend, 
1994) 
NDWI
Modified Normali-
zed Difference Water 
Index
−
=
+
GREEN NIRNDWI
GREEN NIR
Identifies water com-
ponents from satellite 
images 
(Gao, 1996; Mc-
Feeters, 1996) 
MNDWI-1
Modified Normali-
zed Difference Water 
Index- 1
11
1
−
− =
+
GREEN SWIRMNDWI
GREEN SWIR
Aims to separate water 
from the background (Xu, 2006) 
MNDWI-2
Modified Normali-
zed Difference Water 
Index-2
22
2
−
− =
+
GREEN SWIRMNDWI
GREEN SWIR
Aims to detect very low 
reflectance values for 
water features, shallow 
water and water with 
other features in the 
image  
(Xu, 2006) 
AWEIsh 
Automated Water 
Extraction In-
dex-Shadow
2.5
1.5 ( 1) 0.25 2
= + ×
− × + − ×
AWEIsh BLUE Green
NIR SWIR SWIR  
Effectively eliminates 
non-water pixels, includ-
ing dark surfaces, in 
areas with urban back-
grounds
 (Feyisa et al., 
2014)
AWEInsh 
Automated Water 
Extraction In-
dex-Non Shadow
   
4 ( 1)
(0.25 2.75 1)
= × × −
× + ×
AWEInsh Green SWIR
NIR SWIR
Improves the accuracy 
by removing shadow 
pixels that AWEIsh can-
not effectively remove 
 (Feyisa et al., 
2014)
After calculating spectral index values, the last step of the first stage was to filter the satellite images using 
the class boundaries obtained. Thus, the training pixels of the classes were determined. In the second 
stage of the proposed method, the satellite images were filtered temporally and spatially. Then, cloudy 
images were removed according to the Cloud rate data. In this study, images with a cloud cover of less 
than 0.5%, for training images and less than 10% for test images were selected. Then, the spectral indices 
were calculated. In the last stage, classification was carried out with the random forest algorithm. This 
algorithm has been preferred due to its ease of use and high accuracy rates in the literature.
The Random Forest classification algorithm is a machine learning method based on decision trees. In-
stead of branching each node using the best branch among all variables, RO branches each node using 
the randomly selected best variables at each node. Each dataset is generated by displacement from the 
original dataset. Trees are then developed using random feature selection. RO is also very fast, resistant 
to overfitting, and the the number of trees to work with can be adjusted (Breiman, 2001). 
The results obtained were compared with the results obtained by K-means and Learning Vector Quantiza-
tion (LVQ) methods. K-Means is one of the most preferred clustering algorithms. Clustering of data with 
statistically similar characteristics was performed. While performing clustering operations, separations 
are made so that a data is a member of only one cluster. The main purpose of the algorithm is to ensure 
that each cluster from the “K” clusters to be created is as different from each other as possible, so that 
the data of each cluster is close to each other (Hartigan and Wong, 1979). 
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 396 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
The LVQ model is an artificial neural network model with a supportive learning strategy. Unlike the 
supervised learning strategy, both the input values and the expected output values are not given to the 
network for training in the supportive learning strategy. It only tells the LVQ network whether the output 
values produced for the input values are true or false. In the LVQ model, the goal is to represent a multi-
dimensional vector with a set of vectors. That is, to represent a vector with a certain number of vectors. 
During training, the network determines with which vector set the input vectors will be expressed (Ito 
and Omatu, 1999; Kohonen, 1995). Accordingly, the number of classes for K-means was determined 
as 2. For LVQ, the number of classes = 2, the learning rate = 0.1 and the epoch value of 100 were used.
The accuracy of the classification results obtained was determined by the overall accuracy and kappa 
methods. One of the most commonly used metrics to evaluate the performance of classification results is 
overall accuracy. This value is calculated over the error matrix, where the estimates of the target attribute 
and the actual values are compared (Alberg et al., 2004). Kappa is a measure of how well the classifier 
actually performs. In other words, a model will have a high Kappa score if there is a large difference 
between accuracy and null error rate. Kappa can take a value between -1 and +1. A Kappa value of +1 
indicates perfect agreement between the two observers, while a value of -1 indicates perfect discord 
between the two observers. If the kappa value is found to be 0, it indicates that the agreement between 
the two observers is not different from the agreement that may be due to chance (Kraemer, 2014).
3 RESULTS AND DISCUSSION
In order to test the proposed method, the images of water bodies in Canada, Tibetan, China and Guinea 
were classified. In order to assess the classification accuracy, 110 points for Canada, 100 points for Guinea 
and 125 points for Tibetan were created randomly using the high-resolution data available from Google 
Earth (Figure 5). The obtained results were compared with the K-Means and LVQ methods, which are 
frequently used in the unsupervised classification of remote sensing images.
Figure 5: Points used for accuracy assessment
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 397 |
GEODETSKI VESTNIK | 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
In order to test the proposed method, the various study areas with different geographies from around the 
world are used, the first of which is the region in Canada. The classification results are given in Figure 6.
Figure 6:  Classification Results for Canada Region a) Landsat Image b) Result of Proposed Algorithm c) Result of K-means d) 
Result of LVQ
To test the accuracy of the algorithm, the overall accuracy processed and kappa values are calculated. 
The results are given in Table 4, Table 5. Table 6, Table 7 respectively.
Table 4:   Accuracy rates of obtained results
Method Overall Accuracy Kappa
Proposed Algorithm 0.94 0.87
LVQ 0.92 0.83
K-Means 0.92 0.83
The second classification process is carried out on the images of the Tibetan region. The classification 
results are given in Figure 7.
Figure 7:  Classification Results for Guinea Region a) Landsat Image b) Result of Proposed Algorithm c) Result of K-means d) 
Result of LVQ
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 398 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Table 5:   Accuracy rates of obtained results for Guinea Region
Method Overall Accuracy Kappa
Proposed Algorithm 0.97 0.93
LVQ 0.92 0.84
K-Means 0.92 0.86
A third classification process is carried out on the images of the Tibetan region. Results are given in Figure 8.
Figure 8: Classification Results for Tibetan Region a) Landsat Image b) Result of Proposed Algorithm c) Result of K-means d) Result 
of LVQ
Table 6:  Accuracy rates of obtained results for Tibetan Region
Method Overall Accuracy Kappa
Proposed Algorithm 0.98 0.97
LVQ 0.82 0.66
K-Means 0.9 0.78
The last classification process is carried out on the images of the Zhengzhou region. Results are given in Figure 9
.
Figure 9: Classification Results for Zhengzhou Region a) Landsat Image b) Result of Proposed Algorithm c) Result of K-means 
d) Result of LVQ
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 399 |
GEODETSKI VESTNIK | 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Table 7:   Accuracy rates of obtained results for Zhengzhou Region
Method Overall Accuracy Kappa
Proposed Algorithm 0.97 0.92
LVQ 0.83 0.53
K-Means 0.83 0.52
A new automatic water body extraction algorithm has been proposed in this study. The results were 
compared with other automatic classification methods, and the accuracy was higher. One of the biggest 
advantages of the proposed method is that it does not require any parameter or threshold values. It can 
work all over the world with a single training dataset. The methods used in the literature require either 
a parameter for classification or a specified threshold (Behnamian et al., 2017; Das et al., 2021). For 
example, a study using OpenStreetMap to remove water requires a filtering threshold to be determined 
(Zhang et al., 2019). 
Figure 10:  Classification Results for Zhengzhou Region a) Landsat Image b) Result of Proposed Algorithm c) Cloudy regions of 
the satellite image d) Classification results obtained by proposed algorithm e) Result of K-means f ) Result of LVQ g) 
Classification results obtained by K-means h) Classification results obtained by LVQ
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 400 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Another advantage of the proposed method is that it is not specific to any study area such as a city, 
mountainous region, plant, etc. In the Yellow River area used in this study, bare surfaces, urban area and 
turbid waters make water surface removal difficult. However, with the proposed method, various types 
of water area in the image are extracted successfully.
One of the biggest problems encountered in the extraction of water bodies is the presence of cloud-like 
noise in images. There is cloud in the image of the Zhengzhou area used in this study, but the results 
obtained with the proposed method show that it is much less affected by cloud than other methods 
(Figure 10). 
In the literature, water extraction methods are usually prepared for a single type of water body, such as a 
lake, river, etc. (Jin et al., 2021; Nugraha et al., 2018; Xu et al., 2021). The results show that the proposed 
method can successfully extract various types of water body from the Landsat Images. 
 4 CONCLUSIONS 
Water bodies play an important role in the life of living things and socioeconomic development. There-
fore, any changes or developments should be monitored. Rapidly developing remote sensing technology 
offers opportunities for automatic extraction and dynamic monitoring of surface water bodies. However, 
these images need to be analysed accurately and quickly.
This study proposes an automated method of surface water extraction using Landsat 8 images and Corine 
data, which are used to prepare the training data. Once prepared, the training data can be used all over 
the world. The method gives successful results compared to other automatic classification methods used 
in the literature. The results show that the proposed method can be used successfully for water body 
extraction and monitoring.
In future studies, the success of the proposed method on images with different resolutions will be tested. 
For this purpose, satellite images such as Sentinel-2 and Planetscope will be used.
 Literature and references
Alberg, A.J., Park, J.W., Hager, B.W., Brock, M.V., Diener-West, M., 2004. The use of 
“overall accuracy” to evaluate the validity of screening or diagnostic tests. Journal 
of general internal medicine, 19(5p1): 460-465. 
Baran, E., 1995. Dynamique spatio-temporelle des peuplements de poissons estuariens 
en Guinée: Relations avec le milieu abiotique. 
Behnamian, A. et al., 2017. Semi-automated surface water detection with synthetic 
aperture radar data: A wetland case study. Remote Sensing, 9(12): 1209. 
Breiman, L., 2001. Random forests. Machine learning, 45(1): 5-32. 
Chen, J., Du, P., Wu, C., Xia, J., Chanussot, J., 2018. Mapping urban land cover of a 
large area using multiple sensors multiple features. Remote Sensing, 10(6): 872. 
Chen, Y., Tang, L., Kan, Z., Bilal, M., Li, Q., 2020. A novel water body extraction neural 
network (WBE-NN) for optical high-resolution multispectral imagery. Journal 
of Hydrology, 588: 125092. 
Corine, 2019. CLC 2018 Copernicus Land Monitoring Service. 
Dabija, A. et al., 2021. Comparison of support vector machines and random forests 
for corine land cover mapping. Remote Sensing, 13(4): 777. 
Das, N. et al., 2021. Time series analysis of automated surface water extraction 
and thermal pattern variation over the Betwa river, India. Advances in Space 
Research, 68(4): 1761-1788. 
DeFries, R.S., Townshend, J., 1994. NDVI-derived land cover classifications at a global 
scale. International Journal of Remote Sensing, 15(17): 3567-3586. 
Dzieszko, P., 2014. Land-cover modelling using corine land cover data and multi-
layer perceptron. 
Feng, M., Sexton, J.O., Channan, S., Townshend, J.R., 2016. A global, high-resolution 
(30-m) inland water body dataset for 2000: first results of a topographic–
spectral classification algorithm. International Journal of Digital Earth, 9(2): 
113-133. 
Feyisa, G.L., Meilby, H., Fensholt, R., Proud, S.R., 2014. Automated Water Extraction 
Index: A new technique for surface water mapping using Landsat imagery. 
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 401 |
GEODETSKI VESTNIK | 66/3 |
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Remote Sensing of Environment, 140: 23-35. 
Foody, G.M., Mathur, A., 2004. A relative evaluation of multiclass image classification 
by support vector machines. Ieee T Geosci Remote, 42(6): 1335-1343. 
Franczyk, J., Chang, H., 2009. The effects of climate change and urbanization on the 
runoff of the Rock Creek basin in the Portland metropolitan area, Oregon, USA. 
Hydrological Processes: An International Journal, 23(6): 805-815. 
Gallego, J., Peedell, S., 2001. Using CORINE Land Cover to map population density. 
Towards Agri-environmental indicators, Topic report, 6(2001): 92-103. 
Gao, B.-C., 1996. NDWI—A normalized difference water index for remote sensing 
of vegetation liquid water from space. Remote sensing of environment, 58(3): 
257-266. 
Gao, H., Wang, L., Jing, L., Xu, J., 2016. An effective modified water extraction method 
for Landsat-8 OLI imagery of mountainous plateau regions, IOP conference series: 
earth and environmental science. IOP Publishing, pp. 012010. 
Gehrke, T., 2021. MODELING SPECIES ABUNDANCE WITH IMPERFECT DETECTION 
USING ANGLING. 
Hartigan, J.A., Wong, M.A., 1979. Algorithm AS 136: A k-means clustering algorithm. 
Journal of the Royal Statistical Society. Series C (Applied Statistics), 28(1): 
100-108. 
Hooshyar, M., Wang, D., Kim, S., Medeiros, S.C., Hagen, S.C., 2016. Valley and channel 
networks extraction based on local topographic curvature and k-means 
clustering of contours. Water Resources Research, 52(10): 8081-8102. 
Huang, X., Xie, C., Fang, X., Zhang, L., 2015. Combining pixel-and object-based 
machine learning for identification of water-body types from urban high-
resolution remote-sensing imagery. IEEE Journal of Selected Topics in Applied 
Earth Observations and Remote Sensing, 8(5): 2097-2110. 
Huong, H.T.L., Pathirana, A., 2013. Urbanization and climate change impacts on 
future urban flooding in Can Tho city, Vietnam. Hydrology and Earth System 
Sciences, 17(1): 379-394. 
Ito, Y., Omatu, S., 1999. Extended LVQ neural network approach to land cover mapping. 
Ieee T Geosci Remote, 37(1): 313-317. 
Janssen, S., Dumont, G., Fierens, F., Mensink, C., 2008. Spatial interpolation of 
air pollution measurements using CORINE land cover data. Atmospheric 
Environment, 42(20): 4884-4903. 
Jiang, W. et al., 2018. Multilayer perceptron neural network for surface water extraction 
in Landsat 8 OLI satellite images. Remote Sensing, 10(5): 755. 
Jin, S. et al., 2021. River body extraction from sentinel-2A/B MSI images based on an 
adaptive multi-scale region growth method. Remote Sensing of Environment, 
255: 112297. 
Kohonen, T., 1995. Learning vector quantization, Self-organizing maps. Springer, 
pp. 175-189. 
Kraemer, H.C., 2014. Kappa coefficient. Wiley StatsRef: statistics reference online: 1-4. 
Kucsicsa, G. et al., 2019. Future land use/cover changes in Romania: regional 
simulations based on CLUE-S model and CORINE land cover database. Landscape 
and ecological engineering, 15(1): 75-90. 
Matci, D.K., Avdan, U., 2020. Optimization-based automated unsupervised 
classification method: A novel approach. Expert Systems with Applications, 
160: 113735. 
McFeeters, S.K., 1996. The use of the Normalized Difference Water Index (NDWI) in 
the delineation of open water features. International journal of remote sensing, 
17(7): 1425-1432. 
McFeeters, S.K., 2013. Using the Normalized Difference Water Index (NDWI) within 
a Geographic Information System to Detect Swimming Pools for Mosquito 
Abatement: A Practical Approach. Remote Sensing, 5(7): 3544-3561. 
Monegaglia, F., Zolezzi, G., Güneralp, I., Henshaw, A.J., Tubino, M., 2018. Automated 
extraction of meandering river morphodynamics from multitemporal remotely 
sensed data. Environmental Modelling & Software, 105: 171-186. 
Nugraha, P.V.N., Wibirama, S., Hidayat, R., 2018. River body extraction and classification 
using enhanced models of modified normalized water difference index at 
Yeh Unda River Bali, 2018 International Conference on Information and 
Communications Technology (ICOIACT). IEEE, pp. 337-342. 
Rad, A.M., Kreitler, J., Sadegh, M., 2021. Augmented Normalized Difference Water 
Index for improved surface water monitoring. Environmental Modelling & 
Software, 140: 105030. 
Shao, Z. et al., 2008. Characteristics of the change of major lakes on the Qinghai-Tibet 
Plateau in the last 25 years. Frontiers of Earth Science in China, 2(3): 364-377. 
Somasundaram, D. et al., 2020. Learning vector quantization neural network for surface 
water extraction from Landsat OLI images. J Appl Remote Sens, 14(3): 032605. 
Stathopoulou, M., Cartalis, C., 2007. Daytime urban heat islands from Landsat ETM+ 
and Corine land cover data: An application to major cities in Greece. Solar 
Energy, 81(3): 358-368. 
Waltner, I. et al., 2020. Spatial Assessment of the Effects of Land Cover Change on Soil 
Erosion in Hungary from 1990 to 2018. Isprs Int J Geo-Inf, 9(11): 667. 
Wang, K., Zhu, Y., 2003. Recognition of water bodies from remotely sensed imagery 
by using neural network. 
Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance 
open water features in remotely sensed imagery. International journal of remote 
sensing, 27(14): 3025-3033. 
Xu, Y., Lin, J., Zhao, J., Zhu, X., 2021. New method improves extraction accuracy of lake 
water bodies in Central Asia. Journal of Hydrology: 127180. 
Zhang, Z. et al., 2019. Automated surface water extraction combining Sentinel-2 
imagery and openstreetmap using presence and background learning (PBL) 
algorithm. IEEE Journal of Selected Topics in Applied Earth Observations and 
Remote Sensing, 12(10): 3784-3798. 
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |
| 402 |
| 66/3 | GEODETSKI VESTNIK  
RE
CE
NZ
IRA
NI
 ČL
AN
KI 
| P
EE
R-
RE
VIE
W
ED
 AR
TIC
LE
S
SI 
| E
N
Dr. Dilek Küçük Matcı  
Institute of Earth and Space Sciences, Eskişehir Technical University
Iki Eylul Campus, 26470, Eskisehir / TURKEY  
e-mail: dkmatci@anadolu.edu.tr 
Matcı D. K. (2022). Automatic Delineation of  Water Bodies Using Corine Data from Remotely Sensed Images. 
Geodetski vestnik, 66 (3), 387-402. 
DOI: https://doi.org/10.15292/geodetski-vestnik.2022.03.387-402
Dilek Küçük Matci | AUTOMATIC DELINEATION OF WATER BODIES USING CORINE DATA FROM REMOTELY SENSED IMAGES | SAMODEJNA RAZMEJITEV VODNIH TELES Z UPORABO PODATKOV 
CORINE IZ DALJINSKO ZAZNANIH SLIK  |  387-402 |