Author’s Accepted Manuscript

Improved water extraction using Landsat

TM/ETM+ Images in Ebinur Lake, Xinjiang, China

Fei Zhang, Tashpolat Tiyip, Hsiang-te Kung,

Verner Carl Johnson, Juan Wang, Ilyas Nurmemet

www.elsevier.com/locate/rsase

PII: S2352-9385(16)30079-9
DOI: http://dx.doi.org/10.1016/j.rsase.2016.08.001
Reference: RSASE35
To appear in: Remote Sensing Applications: Society and Environment
Received date: 25 December 2015
Revised date: 7 August 2016
Accepted date: 9 August 2016
Cite this article as: Fei Zhang, Tashpolat Tiyip, Hsiang-te Kung, Verner Carl
Johnson, Juan Wang and Ilyas Nurmemet, Improved water extraction using
Landsat TM/ETM+ Images in Ebinur Lake, Xinjiang, China, Remote Sensing
Applications: Society and Environment,
http://dx.doi.org/10.1016/j.rsase.2016.08.001
This is a PDF file of an unedited manuscript that has been accepted for
publication. As a service to our customers we are providing this early version of
the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting galley proof before it is published in its final citable form.
Please note that during the production process errors may be discovered which
could affect the content, and all legal disclaimers that apply to the journal pertain.
Improved water extraction using Landsat TM/ETM+ Images in Ebinur
Lake, Xinjiang, China
Fei Zhang1,2*, Tashpolat Tiyip1,2, Hsiang-te Kung3, Verner Carl Johnson4, JuanWang 1,2, Ilyas Nurmemet1,2
1
Resources and Environment Department, Xinjiang University/Key Laboratory of Oasis Ecology, Ministry of Education, Urumqi 830046,
P.R.China
2
Key Laboratory of Oasis Ecology, Ministry of Education , Xinjiang University, Urumqi 830046, P.R.China
3
Department of Earth Sciences, the University of Memphis, Memphis, TN 38152, USA
4
Department of Physical and Environmental Sciences, Colorado Mesa University Grand Junction, CO 81501, USA
*
E-mail address: zhangfei3s@163.com
Abstract
Water is the most common and important resources on earth. In this paper, we tested and analyzed a variety of water indices for water
surface extraction using Landsat TM/ETM+ images, and evaluated extraction accuracies over Ebinur Lake area in Xinjiang Uyghur
Autonomous Region of China. Eleven algorithms found in literature on land surface water extraction, including normalized difference
water index (NDWI), modified normalized difference water index (MNDWI), automatic water extraction index with no shadow
(AWEInsh), automatic water extraction with shadow (AWEIsh), vegetation index 1 (VI-1), vegetation index 2 (VI-2), vegetation index 3
(VI-3), water index (LBV_B), national wetland inventory (NWI), enhanced water index (EWI), and revised normalized different water
index (RNDWI) were used. The results were validated by a maximum likelihood classification method, edge extraction accuracy
assessment and extracted the area of Ebinur Lake from higher-resolution Landsat ETM+ panchromatic band. The results showed that

these algorithms have different accuracies in extracting water information. We proposed VI-2 for KT3+ TM4＞TM2+ TM7 and VI-3 for

KT3+ TM2＞TM4+ TM3 as an optimized and comprehensive method for land surface water extraction from Landsat TM/ETM+ Images

(KT3, i.e. Wetness index from the Kauth-Thomas Transformation). The proposed VI-2 and VI-3 models demonstrated its potential in
water body extraction with 92.95% and 85.04% accuracies, outperforming other algorithms. Using the optimal mask water model, the two
models achieved up to 93.80% accuracy while minimizing the disturbance from vegetation and non-exploited land. Through comparison
with other commonly used methods, it shows that the performance of the proposed method is superior to the others. Therefore VI-2 and
VI-3 are the best indicators for water mapping using Landsat TM/ETM+ images. This study provided its great potential for quantitative
evaluating of temporal changes of Ebinur Lake in Xinjiang Uyghur Autonomous Region of China.
Keywords: Water information; spectral relationship; Ebinur Lake

1. Introduction
Water is one of the most important natural resources for human survival(Ridd and Liu, 1998; Komeil, 2015).
Surface waters such as lakes, rivers, artificial reservoirs, and seas are essential for climate equilibrium, hydrological
cycle as well as ecosystem balance, providing fundamental resources to terrestrial life (Tulbure and Broich, 2013;
Verpoorter, et al., 2014). Information accuracy about the spatial distribution of land surface water is imperative for
assessment of water resources, watershed changes, land surface water management as well as environmental
monitoring (NRC, 2008; Sun et al., 2012).
In recent decades, remote sensing (RS) technology has become an essential tool in monitoring water information
(Senthilnath et al., 2012; Kodikara et al., 2011; Muster, 2013; and Rokni, 2014). Lakes in arid climatic regions with a
valley/basin are the last link between natural water circulation and human economic activity systems. Water statistics
are supportive in understanding the ecological and environmental changes of the inland river basins (Bhasang et al.,
2012). These systems are likely to be the first reaction to disturbances from humans. The development of remote
sensing technology, monitoring techniques of lake waters have advanced from the traditional hydrometrical station to
more comprehensive monitoring, combined with stream-flow measuring and remote sensing imaging (Mcfeeters,
1996; Feyisa, 2014; Alanazi and Ghrefat, 2013; and Ma et al., 2014). Landsat images have frequently been used to
extract water bodies. The methods can be summarized into three classes: single wavelength threshold, multi-
wavelength spectral relationship, and the water index method (Rundquist, 1987; Alesheikh et al., 2007; and Zhang,
2008). The single wavelength threshold method is simple in operation, but creates noise which makes the
identification resolution low. The multi-wavelength spectral relationship and the water index methods are widely used
to obtain water information. A number of researchers have utilized various methods to study water body extraction
methods. For example, McFeeters (1996) utilized the normalized different water index (NDWI) to restrain the
vegetation and ground samples to extract water information. In the research of the Wagga Lake region of Australia,
Frazier (2000) pointed out TM5 threshold segmentation with a higher accuracy in water extracted samples, but failed
to include small water samples. Yang (2010) applied ETM+ defined thresholds and automated extraction methods for
small water bodies in the mountainous areas of south China. Guo (2012) took the relationship of various water index
models and investigated other methods to define the thresholds for water extraction of. Mohamed Hassani (2014)
developed an index of water surfaces (IWS) for separating the water surfaces from other types of land use by using the
images of Landsat 7 ETM+.
As a result of the dry riverbeds, sedimentation in the water, aquatic vegetation, and shallow water depth will
generate diverse spectral patterns. These produce certain disturbances for the extracted water samples. Figure 1
summarizes the reflectance characteristics of the samples. “The clear water reflectance is lower than that of the
vegetation, urban, mountain shadow, and mixed water categories. Because of the reflectance pattern for mountain
shadows was similar to the pattern for the water, which may lead to confusion between these categories during
automated image processing" (Feyisa et al., 2014). However, it is worth noting that the clear water reflectance is
higher than that of the mountain shadow in Band 1 through Band 3 and opposite is true for the longer bands. In
addition, the spectral absorption intensity ratio of the water body is larger than others on the near infrared (NIR) and
shortwave Infrared (SWIR) wavelengths of Landsat TM/ETM+ images. Even if the water is very shallow, reflection
energy is low, and other features, such as vegetation and bare soil, in these two band energy absorption is lesser,
reflection energy is higher, which makes the water body on NIR and SWIR wavelengths have the obvious difference
with other features. So the NIR and SWIR wavelength can be used to eliminate the interference of water bodies by
threshold segmentation method. The turbid water reflectance increased with the presence of hydrophytes and
sediments on the lake. (Jiang et al., 2014)
 Fig.1 Means and standard deviations of surface reflectance in the five lands cover types (Jiang et al., 2014)

Asphalt pavements and shadows cause errors in the identification of water bodies, which is common when used
optical remote sensing images. To effectively extract the water samples of enclosed lakes, we used Landsat
TM/ETM+ images from the United States Geological Survey (USGS) over Ebinur Lake and surrounding areas, the
largest salt-water lake in Xinjiang Uyghur Autonomous Region of China. Eleven water extraction indices including
single wavelength threshold method, and multi-band spectral methods, and new index was proposed in this study that
demonstrated an improved accuracy over the existing methods.

2. Study area
Ebinur Lake is located between 44º 43′N and 45º 12′N latitude and 82º 35′E and 83º 11′E longitude. The lake is
situated next to Gurbantunggut Desert in the east, close to Tuoli County in the north, and adjacent to Republic of
Kazakhstan in the northwest. Ebinur Lake Wetland National Nature Reserve has a high concentration of many types
of desert plants and also serves as an important habitat for birds in the northwestern part of Xinjiang (Qian et al., 2004)
(Fig. 2). In this area, the average temperature is 7.36 °C; the mean annual precipitation is 149 mm; and mean annual
potential evaporation is 2281 mm. It belongs to a typical temperate continental climate (Yuan and Zhang, 2010).
Within one year, about 72 days have the highest temperatures being between 30 to 35 °C, 15 days with 35 to 40 °C,
and about 2 days with greater than 40 °C. Also about 88 days have lowest temperatures between -20 to -10 °C, 39
days with -30 to -20 °C, and about 2 days with less than -30 °C. Three major rivers, Boertala, Jinghe, and Kuitun,
flow into Ebinur Lake. In the past several decades, the lake area has reduced and the exposed lakebed has grown
larger. As a result is the area around Ebinur Lake is a major source of dust, which impacts the ecological environment
of northern Xinjiang Province.
 Fig.2 Location of the study area

2 Data and methods

2.1 Data
Landsat ETM+ images collected on September 5, 2011, and Landsat TM collected on September 6 and 13, 2011
at path/row of 145/29 and 146/29, respectively were used for this study. The TM/ETM+ images were acquired from
USGS GLOVIS (http://glovis.usgs.gov/). All Landsat images used are corrected and cloud free. TM/ETM+ images
acquired in the form of digital number (DN) were calibrated to surface reflectance values (Wang et al., 2006).
Atmospheric correction was applied to all images using the Fast Line-of-Sight Atmospheric Analysis of Spectral
Hypercube (FLAASH) tool in ENVI 5.1 (Feyisa et al., 2014; Estoque and Murayama, 2015). The images were
mosaicked to generate new images covering the entire study area. For this reason, a mosaicking tool based on geo-
referenced images was used. For co-registration of the multi-temporal images, one of the images was used as the
reference to register the other images. The input images were co-registered with a root mean square error (RMSE) of
less than 0.5 pixels. Finally, the Nearest Neighbor method was used to resample the data (Rokni et al., 2015). In this
study, each image pixel is 30 m × 30 m, and study area has 7 391× 2 137 pixels.
2.2 Land surface water extraction
Landsat TM have seven bands, and Earth’s surface reflects, absorbs, and transmits a part of an incident radiation
differently with these spectral bands based on its structural, chemical, and chromatic features. Numerous methods
have been developed to delineate water information in remote sense imagery (Table 1). The most commonly
used methods include: (1) Spectral bands which identify water information by applying thresholds to one or more
spectral bands which is easy to implement, but often misclassify mountain shadows, urban areas and water body
(Rundquist et al., 1987); (2) Classification methods based on supervised or unsupervised machine-learning algorithms
to extract water bodies from multi-spectral imagery (Otukei and Blaschke, 2010); and (3) Water indices (WIs) derived
from two or more spectral bands that enhance the difference between water body and land use/cover types
(Townshend, 1986). However, the WIs are widely used because of their relatively high accuracy in water information
detection. For this paper the summary of the Landsat TM images of the Ebinur Lake for extracting water body sample
methods in remotely sensed imagery are specified in Table 1.

Table 1 Summary of land surface water information acquiring methods

Water information acquiring

Number methods
Formulation

1 NDWI (TM2-TM4) / (TM2+TM4)(McFeeters 1996; Brakenridge R,et al 2006)

2 MNDWI (TM2-TM5) / (TM2+TM5) (Tebbs et al.,2013)

3 AWEInsh 4×(TM2-TM5)-(0.25TM4+2.75×TM7) (Feyisa et al., 2014)

4 AWEIsh TM1+2.5×TM2-1.5×(TM4+TM5)-0.25TM7 (Feyisa et al., 2014)

5 Vegetation index 1 (VI-1) (TM1-TM7) /(TM1+TM7) (Wu et al.,2013)

6 Vegetation index 2 (VI-2) KT3+TM4>TM2+TM7 (Liu, et al., 2013)

7 Vegetation index 3 (VI-3) KT3+TM2>TM4+TM3(Liu, et al., 2013)

1.126971×TM2+0.673348×TM3+0.077966×TM4-
8 LBV_B
1.878287×TM5+159(Liu, et al., 2013)

(TM1-(TM4+TM5+TM7)) / (TM1+(TM4+TM5+TM7)) ×C
9 NWI (Liu, et al., 2013)
(C is constant, in this study, the C=10 )

10 EWI (TM2-TM4-TM5) / (TM2+TM4+TM5) (Yan et al.,2007)

11 RNDWI (TM5-TM3) / (TM5+TM3) (Wu et al.,2013)

Note: NDWI: Normalized Difference Water Index; MNDWI: Modified Normalized Difference Water Index; AWEInsh: Automated Water
Extraction Index with no shadow; AWEIsh: Automated Water Extraction Index with shadow; LBV_B: radiance Level, Balance and
Variation vector_Balance; NWI: New Water Index; EWI: Enhanced Water Index; RNDWI: Revised Normalized Difference Water
Index, the same as follows.

2.3 Quantitative analysis

To quantitatively compare the accuracy of the water information acquiring from eleven indices, the discrepancy
(D) between water bodies and non-water body is used in a specified equation (1) from Liu, et al. (2013)

| WI n  WI w |
D 100% (1)
| WI n  WI w |
 where WI n and WI w are the mean values index of the non-water body and water body respectively. D

represents the difference of the water body and non-water body. If D is greater than 80%, the separation degree of the
water/non-water body is high. If D is smaller than 30%, the separation degree is low.
2.3 Accuracy verification for classification
A quantitative assessment of the accuracy was conducted for the study area. The overall accuracy and the kappa
coefficient on the basis of the error matrix (Foody, 2002) were employed to evaluate the accuracy of maximum
likelihood classification map. The kappa coefficient can be calculated as:
n n
P  Pii   ( Pi   Pi )
Kappa  i 1
n
i 1
(1)
P   ( Pi   Pi )
2

i 1

where P is the total number of pixels from the reference data; Pii is the total number of correct pixels from the
category; Pi+ is the total number of pixels for the category derived from the classified data; P+i is the total number of
pixels for the category derived from the reference data; and n is the total number of categories.
2.4 Water area extraction
Visual interpretation and quantitatively statistical analysis methods are used to analyze and precisely evaluate the
areas of water body from various models. According to Baodong’s (2014) research, the water area can be calculated
as:
S  Sw  Sm (2)

where S is the total water area, Sw is the area of pure water; and Sm is the water area in the mixed pixels. In this
study, according to the condition of the study area, the area calculation of water body is based on equation (3) from
Jiang (2013)
N
S   pixelSize  pixelSize 106 (3)
i 1

Where S is the total water area of Ebinur Lake; N is the pixel numbers; and pixelSize is the resolution of images.
Water extraction results of various models were overlaid with the original images, through visual inspection and on
the basis of Maximum Likelihood Classification. Joint validation was conducted for water extraction accuracy of the
proposed method.
2.5 Accuracy of the edge for the extracted Ebinur Lake
Boundary extraction is a simple morphological image processing algorithm that extracts the outermost
pixels of features from a binary image (Gonzalez et al., 2009). Two metrics were applied to evaluate the results
from different perspectives:
EC  NC / N R 100% (4)
EO  NO / N R 100% (5)
Here, E represents the edge position error frequency (C for commission and O for omission), NR is the number
of edge pixels in the reference image; NC is the number of extracted edge pixels that lie outside the reference
water bodies (i.e., the number of commission errors); and NO is the number of extracted edge pixels that lie inside
the reference water bodies and that are not directly adjacent to the actual edge (i.e., omission errors). In this study,
each image pixel is 30 m × 30 m, so the equation (4) and (5) are valid for 30m pixel. Fig. 3 shows methods adopted in
this study to extract land surface water information.

Fig.3 Flowchart showing the overall methods adopted in this study

3. Results
3.1 Land surface water extraction
Using the Landsat TM data of Ebinur Lake, eleven common algorithms shown in Table 1 were used for
extracting water body. Results are shown in Fig.4. We found VI-2 and VI-3 methods performed well in extracting
water from the Landsat TM data and partially eliminating the impact of mountain shadows. Furthermore, 1000
samples of water bodies, vegetation cover, and non-exploited ground were selected for statistical analysis discrepancy
with the equation (1) and the results are shown in Table 2. From both Landsat ETM+ panchromatic band (Sep.5, 2011)
and random sampling comparison of partially ground conditions indicate VI-2 and VI-3 provided the best methods for
extracting water samples from the Landsat TM data.
 Fig.4 Comparison of different land surface water information map of various extraction methods in the study area

Water extraction ability of VI-2 and VI-3 is the best. The discrepancy between water body and vegetation with VI-
2 model is 87.56%, and that between water body and non-exploited land is 88.20%. While the discrepancy between
water body and vegetation with VI-3 is 89.13% and 86.11%, respectively. The water extracted information of VI-2
and VI-3 are more accurate than other methods. The lower accuracy of other methods might be attributed to the small
water-body features which is subject to greater error and middle spatial resolution images (e.g., Landsat).
Table 2 Comparison of water body with non-water body

Land surface water information

Number Objects Mean value Discrepancy /% Objects Mean value Discrepancy /%
extraction methods
Water body 0.525 Water body 0.525
43.64
1 NDWI 19.35 Unutilized
Vegetation -0.777 -0.206
land
 Water body 0.915 Water body 0.915
53.52
2 MNDWI 27.26 Unutilized
Vegetation -0.523 -0.277
land
Water body 0.501 Water body 0.501
45.01
3 AWEInsh 21.16 Unutilized
Vegetation -0.770 -1.321
land
Water body 0.358 Water body 0.358
3.47
4 AWEIsh 34.13 Unutilized
Vegetation -0.729 -0.334
land
Water body 0.824 Water body 0.824
25.32
5 Vegetation index 1 (VI-1) 13.58 Unutilized
Vegetation -0.627 -0.491
land
Water body 0.064 Water body 0.064
88.20
6 Vegetation index 1 (VI-2) 87.56 Unutilized
Vegetation 0.965 1.021
land
Water body 1.166 Water body 1.166
86.11
7 Vegetation index 1 (VI-3) 89.13 Unutilized
Vegetation -0.067 -0.087
land
Water body 159.220 Water body 159.220
0.13
8 LBV_B 0.12 Unutilized
Vegetation 158.824 158.809
land
Water body 2.026 Water body 2.026
58.67
9 NWI 61.99 Unutilized
Vegetation -9.401 -7.779
land
Water body 0.478 Water body 0.478
5.44
10 EWI 27.25 Unutilized
Vegetation -0.836 -0.533
land
Water body -0.870 Water body -0.870
60.37
11 RNDWI 25.81 Unutilized
Vegetation 0.513 0.215
land
Furthermore, the detailed comparison of water bodies, vegetation cover, and non-exploited ground from these
eleven algorithms is listed in the Fig.5. In the red elliptical area, VI-2 and VI-3 water extraction models suppressed
vegetation information more effectively, which increased the accuracy of the water body extraction by lowering the
mixing of water and non-water covered ground. Furthermore the images of other models have lower identification of
water, vegetation cover, and non-exploited ground. The water information could not be extracted exactly and therefore
the accuracy of water information acquiring was lower.
 Fig.5 Comparison between Landsat TM image and land surface water information map of various extraction models
a: Original image, b: NDWI, c: MNDWI, d: AWEInsh, e: AWEIsh, f: Vegetation index 1 (VI-1),g: Vegetation index 2 (VI-2), h:
Vegetation index 3 (VI-3), i: LBV_B, j: NWI, k:EWI, l:RNDWI
3.2 Maximum likelihood classification
Maximum likelihood classifier is one of the most commonly used methods in land use/cover classification.
(Feyisa et al., 2014; Frazier and Page, 2000). It was easy to find visual differences between water and shadows in
stretched water image derived using threshold technique in SWIR band, in which some shadow pixels were present.
Therefore, the method of maximum likelihood classification could be used to classify the water image (Du et al.,
2002). We classified the remote sensing images into two parts: water body and others (farmland, forestland, grassland,
etc.) (Fig.6). According to the equation (1), we obtained classification overall accuracy of 99.3% and the kappa
coefficient of 0.99. Also from calculation using equation (3) water body area was estimated at 55.35×107 km2. These
results indicated the classification was very accurate.

Fig.6 Maximum likelihood classification map

The threshold method is considered a common approach to extract water bodies because it is easier to use and
less computationally time-consuming than alternatives approaches (Ryu et al., 2002). In this study, to determine the
optimal threshold, a series of thresholds were tested and each possible threshold generates a water body extraction
result. Because a higher threshold may identify only a few water bodies, a lower threshold may not only identify many
water bodies, but also mistakenly identify land as water bodies. So, according to the tested thresholds for all water
indices, we obtained the optimal thresholds for each index, and the extraction results of eleven indices were presented
in Table 2.
Table 2 The results and accuracy evaluations of water information extraction models

Optimal Models area Classification area

Land surface water information Accuracy (%)
Number 2
extraction methods threshold (m ) (m2)

84.44
1 NDWI 0.00 65.55×107

67.64
2 MNDWI 0.32 81.83×107

63.55
3 AWEInsh 0.00 87.10×107 55.35×107

73.17
4 AWEIsh 0.20 75.65×107

64.14
5 Vegetation index 1 (VI-1) 0.30 86.30×107
 92.95
6 Vegetation index 2 (VI-2) 0.20 59.55×107

85.04
7 Vegetation index 3 (VI-3) 1.03 65.09×107

55.44
8 LBV_B 159.08 99.84×107

58.32
9 NWI -5.00 94.91×107

76.60
10 EWI -0.10 72.26×107

61.82
11 RNDWI -0.20 89.54×107

The results indicate the accuracy of water area from various water body extraction models. The accuracy for VI-2
was the highest at 92.95%. The second highest accuracy was VI-3 at 85.04%. These results for VI-2 and VI-3 were in
consistent with in-situ measurements and validation index highlighting the potential to provide more accurate water
information from Landsat data.
Nevertheless, the method of threshold for discriminating water body from other land covers is somewhat
dependent on human experience, thus influencing the accuracy of image classifications. To further evaluate the
accuracy of water indices, Landsat ETM+ panchromatic was used to identify the edges of the Ebinur Lake manually
and the lake edge image was used as reference image. We choose 200 points at the edges of Ebinur Lake to identify
the mixed pixels.
Table 3 Edge extraction accuracy assessment for Ebinur Lake. E: errors (C, commission; O, omission)

Number Land surface water information extraction methods Ec(%) Eo(%)

1 NDWI 3.0 35.0

2 MNDWI 5.0 50.5

3 AWEInsh 5.5 57.0

4 AWEIsh 9.5 36.0

5 Vegetation index 1 (VI-1) 2.5 56.5

6 Vegetation index 2 (VI-2) 4.0 22.5

7 Vegetation index 3 (VI-3) 4.5 26.5

8 LBV_B 2.5 60.0

9 NWI 5.0 58.5

10 EWI 11.5 34.0

11 RNDWI 5.0 61.5

Table 3 summarizes the edge extraction accuracy for Ebinur Lake. It can be seen that there still either high
commission or omission errors at the edges in some water index results regardless of a suitable threshold. According
to the results, the VI-2 and VI-3 showed smaller error of lake boundary extraction. The reason for this result is that the
VI-2 and VI-3 are more vulnerable to shadow pixels than the others.
 The area of Ebinur Lake was validated using the Landsat ETM+ panchromatic band which offers a higher spatial
resolution. The lake shoreline was delineated manually on-screen from the reference data. Furthermore, the result was
used to assess the accuracy of the indices of water body (Table 4). The accuracy for VI-2 was the highest at 95.12%.
The second highest accuracy was with the VI-3 at 94.46%. That result is the same as the maximum likelihood
classification validation results that the accuracy of the VI-2 and VI-3 methods are considerably higher than the others.
So, the maximum likelihood classification method is feasible in this study area. The image used in this analysis was
clipped from the image for study area and shown in Figure 7.
Table 4 Precision of extraction of the area of Ebinur Lake.
Reference area from
Optimal Models area
Land surface water information ETM+ Panchromatic Accuracy (%)
Number
extraction methods threshold (m2) 2
(m )
93.13
1 NDWI 0.00 4.82×108

84.48
2 MNDWI 0.32 5.21×108

81.60
3 AWEInsh 0.00 5.34×108

93.79
4 AWEIsh 0.20 4.79×108

83.15
5 Vegetation index 1 (VI-1) 0.30 5.27×108

95.12
6 Vegetation index 2 (VI-2) 0.20 4.73×108 4.51×108

94.46
7 Vegetation index 3 (VI-3) 1.03 4.76×108

79.60
8 LBV_B 159.08 5.43×108

82.70
9 NWI -5.0 5.29×108

94.01
10 EWI -0.10 4.78×108

74.28
11 RNDWI -0.20 5.67×108
 Fig.7 Comparison of the area of Ebinur Lake between Landsat ETM+ image and various extraction models

3.3 Comprehensive extraction method of land surface water information

There were no clear boundaries between vegetation cover, river-beds and water bodies in the study area. All
three of these features generated errors of thresholds which are difficult to differentiate the mixed pixels exactly. In
order to solve some of the problems of the traditional water extraction, model can hardly extract small and tiny water
body, according to discrepancy between water body and non-water body. We selected VI-2 and VI-3 to complement
the distinctions between water body and non-water-body, for example water from the vegetation cover (see flowchart
in Fig.8). Based on both of these methods, according to the scope of the research region, the accuracy of
comprehensive method was obtained by the estimated water areas divided by the water areas from the maximum
likelihood classification.
 Vegetation index 2 (VI2) Vegetation index 3 (VI3)

Take the upper limit threshold Take the lower limit threshold
t1 (t1<0.2) t2 (t2>1.03)

Mask 1 (B1) Mask 2 (B2)

B1*B2

The optimal mask water

information extraction model

Accuracy evaluation

Fig.8 Technique flow chart of the comprehensive method

Wu (2008) proposed a multiplication method to optimize water extraction model. Two image masks were
multiplied (i.e. B1×B2), if the water was presented in both masks, the value would be one; otherwise it would be zero.
This procedure reduced the spatial information and increased the water extraction accuracy. In this proposal, the space
information is reduced, which improves the accuracy of land surface water information (Fig.9). Using an optimized
mask, through equation (3), the calculated water body area is 59.01×107m2 and compared to the classification result,
the accuracy is 93.80%. It is concluded that the comprehensive water extracted model provided better information on
water and reduced the confusion of water and other ground object in the Landsat TM data. Thus accuracy was greatly
improved after adopting the comprehensive extracted method.

Fig.9 Water information extraction result of the comprehensive method

4. Discussion
Many researchers used water extracted methods based mainly on a single model using only spectral data and it is
not suitable for complicated areas such as Ebinur Lake. These methods do have inherent deficiencies. For example,
using differ extraction methods specified in Table 2, the results showed the overall accuracy of water information
extraction varied between 55.44% and 92.95%. These ranges of accuracy indicated that there were some
misclassifications in the land surface and water maps. This misclassification may have been caused by overestimation
of small water bodies and underestimation of the river channels with mixed pixels. This can be addressed by using
spectral linear unmixing methods by Keshava and Mustard (2002).
One factor that may significantly affect the efficiency of land surface water maps is the threshold selection which
is influenced by the subjective judgment of the researcher. Another major factor is the absence of ground-truth water
body surface distribution and area data. To solve this problem, high-resolution data should be used. Additionally, the
Landsat-8 Operational Land Imager (OLI) may provide an important data source for a better estimation of water
bodies due to its multiple spectral bands within water sensitive wavelengths as specified below:
(1) Due to the water extraction indices were developed on specific areas of interests, there were significant
differences exist among the results of various water extraction indices. Therefore, the selection of proper models or
the assembling of water models is extremely important (Song et al., 2014; Rokni et al., 2015)
(2) The salinized land, such as dry riverbeds with silt and aquatic plants/hydrophytes has certain disturbances for
water body spectrum. This disturbance generates the deviation of water information estimation. So, the application of
multi-temporal remote sensing information and the study of water identification models should provide guidance to
solve the problems.
(3) Estimation accuracy of small-water bodies is a current bottle-neck in remote sensing of water resources.
According to Liu (2013), there are certain spectral similarities between shadows and water features. The accuracy of
water identification is mainly influenced by shadow noise. Therefore, efforts should be made to reduce or eliminate
shadow noise from the water extraction indices.

5. Conclusion
In this study, we conducted a comparative study among eleven water indices by using visual interpretation and
quantitative statistical analysis. Our results showed that all eleven models can automatically extract water information
to some degree. The results were validated by using maximum likelihood classification, edge extraction accuracy
assessment and Ebinur Lake water body area derived from 15m Landsat ETM+ pan band. VI-2 and VI-3 models
proposed in this study, demonstrated the highest accuracy in the identification of water body. In model VI-2, the
identification of water body vs. vegetation cover and water body vs. non-exploited ground. vs. vegetation cover is
87.56%; and that of water body vs. non-exploited ground is 88.20%. In model VI-3, the identification of water body
vs. vegetation cover is 89.13%; and that of water body vs. non-exploited ground is 86.11%. So VI-2 and VI-3 models
are more suitable for the water information extraction of Ebinur Lake. These two models utilize the visible and near
infrared wavebands. It provides references for the water information automatic extraction from the images of National
China-Brazil Earth Resources Satellite and Environment & Disaster Monitoring Small Satellite.
 In the eco-tone areas, the complexity of no clear boundaries of vegetation cover, river-beds, and water bodies
often causes larger errors due to mixed pixel effect. It is difficult to exactly differentiate objects in mixed pixels. To
improve the extraction accuracy of small water bodies using the common water body models, we proposed VI-2 and
VI-3 as an optimized method of extracting water information. The calculated water area is 59.01×107 m2 and the
accuracy is 93.80% to compare with classification result. This method utilizes the comprehensive water extraction
model and provides much better water information.
Overall, the VI-2 and VI-3 algorithms can extract surface water information more accurately than the other
methods. Also these methods are suitable for the surface water extraction from Landsat TM images. The results of
other models are not very good because some small water bodies could not be effectively extracted from Landsat TM
images. Thus further study is needed in other areas for water information extraction, because the proposed approach
has proven to be effective in detecting the land surface water information of Ebinur Lake in our study. Accordingly,
the research project method may also be useful in studying other land surface water questions (such as flood
monitoring) in the world.

Acknowledgments
We are grateful to the financial support provided by the National Natural Science Foundation of China
(No.41361045), Xinjiang Local Outstanding Young Talent Cultivation Project of National Natural Science
Foundation of China (U1503302)，the Young Technology-Innovation Training Program Foundation for Talents from
the Xinjiang Uygur Autonomous Region (No.2013731002), the National International Scientific and Technological
Cooperation Projects (No. 2010DFA92720-12), the State Key Program of National Natural Science of China (No.
41130531), the Key Laboratory of Oasis Ecology in Xinjiang University (No.XJDX0201-2012-01). We wish to thank
Professor Bethany Iris LaGrone, Julia Shena Crutchfield, Spencer Colton, and other referees for providing helpful
suggestions to improve this manuscript.

References
Alesheikh A A, Ghorbanali A, Nouri N (2007). Coastline change detection using remote sensing. Int.J.Environ.Sci.
Technol.4:61–66.
Alanazi H A, Ghrefat H A(2013).Spectral analysis of multispectral Landsat 7 ETM+ and ASTER data for mapping
land cover at Qurayah Sabkha, Northern Saudi Arabia. Journal of the Indian Society of Remote Sensing, 41(4),
833-844.
Brakenridge R, Anderson E(2006). MODIS-based flood detection, mapping and measurement: The potential for
operational hydrological applications. Transboundary floods: reducing risks through flood management, 1-12.
Ma B D, Wu L X, Zhang X X, Li X C, Liu Y, Wang S L(2014). Locally adaptive unmixing method for lake-water
area extraction based on MODIS 250m bands. International Journal of Applied Earth Observation and
Geoinformation, 33:109-118.
Bhasang T, Liu J S, Niu J F, Dhawa T, Bian C (2012). Area Variation and Its Causes of Bamu Co Lake in the Central
Tibet. Journal of Natural Resources, 27(2):302-310. (in Chinese)
Du J K, Feng X Z, Wang Z L, Huang Y S, Ramadan E(2002). The methods of extracting water information from
SPOT image. Chinese geographical science,12(1):68-72.
El-Asmar H M, Hereher M E (2011).Change detection of the coastal zone east of the Nile Delta using remote sensing.
Environ Earth Sci, 62:769–777.
Estoque R C,Murayama Y J (2015). Classification and change detection of built-up lands from Landsat-7 ETM+ and
Landsat-8 OLI/TIRS imageries: A comparative assessment of various spectral indices. Ecological Indicators,56,
205-217.
Frazier P S,Page K J (2000).Water body detection and delineation with Landsat TM data.PE & RS-Photogrammetric
Engineering & Remote Sensing, 66(12):1461-1467.
Foody G M(2002). Status of land cover classification accuracy assessment. Remote Sens. Environ. 80:185-201.
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. Remote Sensing of Environment, 140:23-35.
Guo Z Y, Wang X Y, Wang CH, Gao C, Wu H Z (2012).The Research on Extraction Method of Water Body
Information in Chaohu Lake Basin. Remote Sensing Technology and Application, 27(3):443-448. (in Chinese)
Gonzalez R C, Woods R E, Eddins S L(2009). Digital Image Processing Using MATLAB; Gatesmark Publishing
Knoxville: Knoxville, TN, USA 2009.
Hassani M, Chabou M C, Hamoudi M, Guettouche M S(2015). Index of extraction of water surfaces from Landsat 7
ETM+ images. Arabian Journal of Geosciences, 8(6):3381-3389.
Huang J L, Tao H, Wang Y J, Bai Y G (2012). Analysis on relationship between water level and water area of lake
based on MODIS image. Transactions of the Chinese Society of Agricultural Engineering, 28(23):140-146. (in
Chinese)
Jiang L G,Yao Z J, Liu z F, Wu S S (2013). A review of lake dynamic change research based on remote sensing.
Remote Sensing Techology and Application, 28(5):807-814.
Jiang H, Fen M, Zhu Y Q, Lu N, Huang J X, Xiao T (2014). An Automated Method for Extracting Rivers and Lakes
from Landsat Imagery. Remote Sensing,6(6):5067-5089.
Keshava N and Mustard J F (2002).Spectral unmixing. IEEE Signal Process.Mag.19:44-57.
Kodikara R L, Woldai T, Ruitenbeek F J A, Kuria Z, Meer F, Shepherd K D,Hummel G J(2011). Hyperspectral
remote sensing of evaporate minerals and associated sediments in Lake Magadi Area, Kenya.Int J Appl Earth
Obs Geoinf,14(1):22–32.
Liu GL, Zhang LC Liu J, Li G Y (2013). Water body information Extraction based on Landsat TM remote sensing
imagery. Journal of University of Chinese Academy of Sciences, 30(5): 644-650. (in Chinese)
Ma B D, Wu L X,Zhang X X, Li X C.Liu Y, Wang S L (2014).Locally adaptive unmixing method for lake-water area
Extraction based on MODIS 250m bands. International Journal of Applied Earth Observation and
Geoinformation, 33:109-118.
Muster S; Heim B; Abnizova A; Boike J(2013). Water body distributions across scales: a remote sensing based
comparison of three arctic tundra wetlands. Remote Sens., 5:1498–1523.
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.
National Research Council(2008). Integrating Multiscale Observations of U.S. Waters; The National Academies Press:
Washington, DC, USA.
Otukei J, Blaschke T(2010). Land cover change assessment using decision trees, support vector machines and
maximum likelihood classification algorithms. Int. J. Appl. Earth Obs. Geoinf, 12: 27–31.
Qian Y B, Wu Z N, Jiang J, Yang Q (2004). Eco-environmental Change and Its Impact Factors in the Ebinur Lake
Catchments during the Past 50 Years. Journal of Glaciology and Geocryology, 26 (1):17-26. (in Chinese)
Rundquist D C, Lawson M P, Queen L P, Cerveny R S (1987).The relationship between summer-season rainfall
events and lake-surface area. J. Am. Water Resour. Assoc. 23: 493–508.
Ridd M K, Liu J (1998).A comparison of four algorithms for change detection in an urban environment. Remote
Sensing of Environment, 63:95-100.
Ryu J H, Won J S., Min K D(2002). Waterline extraction from Landsat TM data in a tidal flat: A case study in
Gomso Bay, Korea. Remote Sens. Environ. 83, 442–456.
Rokni K, Ahmad A, Selamat A, Hazini S (2014).Water feature Extraction and change detection using multi-temporal
Landsat imagery.Remote Sens, 6:4173-4189.
Rokni K, Ahmad A, Solaimani K, Hazini S (2015).A new approach for surface water change detection: Integration of
pixel level image fusion and image classification techniques. International Journal of Appplied Earth
Observation and Geoinformation, 34:226-234.
Senthilnath J, Bajpai S. Omkar S N ,Diwakar P G, Mani V(2012).An approach to multi-temporal MODIS image
analysis using image classification and segmentation. Adv.Space Res.50:1274–1287.
Song C Q, Huang B, Ke L H, Richards K S (2014). Remote sensing of alpine lake water environment changes on the
Tibetan Plateau and surroundings: A review.ISPRS Journal of Photogrammetry and Remote Sensing,92:26-37.
Sun F, Sun W, Chen J, Gong P (2012).Comparison and improvement of methods for identifying waterbodies in
remotely sensed imagery, Int. J.Remote Sens, 33:6854-6875.
Tebbs E J, Remedios J J, Avery, S T, Harper D M (2013).Remote sensing the hydrological variability of Tanzania’s
Lake Natron, a vital lesser flamingo breeding site under threat. Ecohydrology and Hydrobiology, 13: 148-158.
Townshend J R.,Justice C(1986). Analysis of the dynamics of African vegetation using the normalized difference
vegetation index. Int. J. Remote Sens, 7, 1435–1445.
United States Geological Survey (USGS) (2012). Landsat data archive. Global Visualization Viewer (GLOVIS).
Verpoorter C, Kutser, T,Seekell D A, Tranvik L J(2014). A global inventory of lakes based on high-resolution
satellite imagery. Geophysical Research Letters, 41:6396–6402.
Wu W Y, Shen X H, Zhou L J, Lu S L, Zhang G F (2008). A Integrated Method for Water Body Detection
Delineation Using Landsat ETM+ data. Bulletin of Science and Technology, 24(2):252-259. (in Chinese)
Wu J T, Tan W, Yu L F (2013). Comparative study of different water indexes based on TM/ETM+ imagery. Science
of Surveying and Mapping, 38(4):193-195. (in Chinese)
Wang F, Han L, Kung H T,Van Arsdale R N (2006).Applications of Landsat-5 TM imagery in assessing and mapping
water quality in Reelfoot Lake, Tennessee. International Journal of Remote Sensing, 27(23):5269-5283.
Yang S W, Xue C S, Liu T (2010). A Method of Small Water Information Automatic Extraction from TM Remote
Sensing Images. Acta Geodaeticaet Cartographica Sinica, 39(6):611-617. (in Chinese)
Yuan X C, Zhang L P (2010). Analysis of hydrological characteristics of Xinjiang River Basin in Jinghe. Modern
Agricultural Science and Technology, 6(1):289-294. (in Chinese)
Yan P, Zhang Y J, Zhang Y (2007). A Study on Information Extraction of Water System in Semi-arid Regions with
the Enhanced Water Index (EWI) and GIS Based Noise Remove Techniques. Application of remote sensing,
6:62-67. (in Chinese)
Zhang M H (2008). Extracting Water-Body Information with Improved of Spectral Relationship in a Higher Mountain
Area. Geography and Geo-Information Science, 24(2):14-16. (in Chinese)

Highlights
 Eleven algorithms found in literature on land surface water extraction were used.
 The results were validated by a maximum likelihood classification method, edge extraction accuracy assessment
and extracted the area of Ebinur Lake from higher-resolution Landsat ETM+ panchromatic band.

 We proposed VI-2 for KT3+ TM4 ＞TM2+ TM7 and VI-3 for KT3+ TM2 ＞TM4+ TM3 as an optimized and

comprehensive method for land surface water extraction from Landsat TM/ETM+ Images (KT3, i.e. Wetness
index from the Kauth-Thomas Transformation).
 This study provided its great potential for quantitative evaluating of temporal changes of Ebinur Lake in Xinjiang
Uyghur Autonomous Region of China.