Remote Sensing of Environment 140 (2014) 23–35

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Automated Water Extraction Index: A new technique for surface water

mapping using Landsat imagery
Gudina L. Feyisa a,⁎, Henrik Meilby a, Rasmus Fensholt b, Simon R. Proud b
a
Department of Food and Resource Economics, University of Copenhagen, Rolighedsvej 23, DK-1958 Frederiksberg C, Denmark
b
Department of Geosciences and Natural Resource Management, University of Copenhagen, Oester Voldgade 10, DK-1350 Copenhagen K, Denmark

a r t i c l e i n f o a b s t r a c t

Article history: Classifying surface cover types and analyzing changes are among the most common applications of remote
Received 16 April 2013 sensing. One of the most basic classification tasks is to distinguish water bodies from dry land surfaces. Landsat
Received in revised form 3 August 2013 imagery is among the most widely used sources of data in remote sensing of water resources; and although
Accepted 21 August 2013
several techniques of surface water extraction using Landsat data are described in the literature, their application
Available online xxxx
is constrained by low accuracy in various situations. Besides, with the use of techniques such as single band
Keywords:
thresholding and two-band indices, identifying an appropriate threshold yielding the highest possible accuracy
Classification accuracy is a challenging and time consuming task, as threshold values vary with location and time of image acquisition.
Threshold stability The purpose of this study was therefore to devise an index that consistently improves water extraction accuracy
Subpixel in the presence of various sorts of environmental noise and at the same time offers a stable threshold value. Thus
Mixed pixel we introduced a new Automated Water Extraction Index (AWEI) improving classification accuracy in areas that
include shadow and dark surfaces that other classification methods often fail to classify correctly. We tested the
accuracy and robustness of the new method using Landsat 5 TM images of several water bodies in Denmark,
Switzerland, Ethiopia, South Africa and New Zealand. Kappa coefficient, omission and commission errors were
calculated to evaluate accuracies. The performance of the classifier was compared with that of the Modified Nor-
malized Difference Water Index (MNDWI) and Maximum Likelihood (ML) classifiers. In four out of five test sites,
classification accuracy of AWEI was significantly higher than that of MNDWI and ML (P-value b 0.01). AWEI im-
proved accuracy by lessening commission and omission errors by 50% compared to those resulting from MNDWI
and about 25% compared to ML classifiers. Besides, the new method was shown to have a fairly stable optimal
threshold value. Therefore, AWEI can be used for extracting water with high accuracy, especially in mountainous
areas where deep shadow caused by the terrain is an important source of classification error.
© 2013 Elsevier Inc. All rights reserved.

1. Introduction the dynamics of surface water are, therefore, essential for policy and
decision making processes (Giardino, Bresciani, Villa, & Martinelli,
Environmental changes and their impacts on natural systems and 2010; Morss, Wilhelmi, Downton, & Gruntfest, 2005).
human societies are topics of research in a wide range of scientific fields. Remote sensing has become an important source of information in
Surface water is among the most vital earth resources undergoing analyzing and delivering data on changes in different earth resources,
changes in time and space as a consequence of land use/cover (LULC) and surface water in particular. Examples of studies applying remote
changes, climate change and other forms of environmental changes in sensing and GIS techniques for various applications in relation to water
many parts of the world. The ecological, social, health and economic resources include flood hazard/damage assessment and management
effects of surface water changes have been the subject of academic (Dewan, Islam, Kumamoto, & Nishigaki, 2007; Ji, Zhang, & Wylie, 2009;
study for many years (Alderman, Turner, & Tong, 2012; Bond, Lake, & Proud, Fensholt, Rasmussen, & Sandholt, 2011), change in surface
Arthington, 2008; Charron et al., 2004; Kondo et al., 2002; Lake, 2003; water resources (Gardelle, Hiernaux, Kergoat, & Grippa, 2009; Haas,
Li, Wu, Dai, & Xu, 2012); Sun, Sun, Chen, and Gong (2012). Changes in Bartholomé, & Combal, 2009; Prigent et al., 2012), water quality assess-
surface water may result in disasters such as flooding, outbreaks of ment and monitoring (Guttler, Niculescu, & Gohin, 2013; He et al.,
waterborne disease and water shortage in dry tropical areas, which 2012; Novoa et al., 2012), and water-related disease epidemiology
may involve loss of lives. Timely monitoring and delivery of data on (Charoenpanyanet & Chen, 2008; Dambach et al., 2012; Lacaux, Tourre,
Vignolles, Ndione, & Lafaye, 2007).
⁎ Corresponding author. Tel.: +45 91414185; fax: +45 353 31508.
Satellite sensors of varying spatial, temporal and spectral resolution
E-mail addresses: [email protected], [email protected] (G.L. Feyisa), have been used to extract and analyze information regarding surface
[email protected] (H. Meilby), [email protected] (R. Fensholt), [email protected] (S.R. Proud). water. Landsat satellites are among the most widely used optical

0034-4257/$ – see front matter © 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.rse.2013.08.029
24 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35

sensors in surface water and other environmental research. The use of Landsat TM data, we noted accuracy problems due to failure of existing
these remotely sensed data commonly starts with classification of water extraction methods in accurately distinguishing water from
land use/cover types. Common water classification methods for optical shadows and low albedo urban surfaces. Particularly, no existing
imagery could be categorized into four basic types (Ji et al., 2009): water index was able to automatically separate water and shadowed
(a) thematic classification (Lira, 2006), (b) linear unmixing (Sethre, surfaces. In this paper, therefore, we introduce a multiple-band index
Rundquist, & Todhunter, 2005), (c) single-band thresholding (Jain, called Automated Water Extraction Index (AWEI), with the objectives
Singh, Jain, & Lohani, 2005) and (d) two-band spectral water indices to: (a) improve accuracy of surface water mapping by automatically
(Jain, Saraf, Goswami, & Ahmad, 2006; McFeeters, 1996; Rogers & suppressing classification noise from shadow and other nonwater
Kearney, 2004; Xu, 2006). Combinations of various methods are also dark surfaces, and (b) test the robustness of the new method under
proposed to improve water extraction accuracies. Examples are, Jiang, different environmental conditions and evaluate its relative accuracy
Qi, Su, Zhang, and Wu (2012), Sheng, Shah, and Smith (2008), Sun in comparison with existing classification techniques.
et al. (2012) and Verpoorter, Kutser, and Tranvik (2012). Single band
thresholding and two-band indices are commonly used water extrac- 2. Study areas and data sources
tion methods because of ease of use and the fact that these methods
are computationally less time-consuming than alternative approaches 2.1. Test sites
(Ryu, Won, & Min, 2002).
McFeeters (1996) introduced the Normalized Difference Water The accuracy and robustness of the Automated Water Extraction
Index (NDWI) to delineate open water features using the green (band Index (AWEI) were tested considering several lakes and other water
2) and near-infrared (band 4) of Landsat TM. Rogers and Kearney bodies in different environmental conditions ranging from humid tem-
(2004) used another NDWI for water extraction where they applied perate through sub-tropical to tropical dry regions. The test water bod-
bands 3 and 5 of Landsat TM. McFeeters (1996) proposed a threshold ies were obtained from five different countries: Denmark, Switzerland,
of 0 for extracting surface water using the raw digital number of Ethiopia, South Africa and New Zealand. The water bodies that include
Landsat, where all positive NDWI values would be classified as water small freshwater reservoirs, large lakes, harbors and the sea differ
and negative values as nonwater. However, Xu (2006) found that the with regard to depth, turbidity, chemical composition and surface ap-
NDWI cannot efficiently suppress the signal from built-up surfaces pearance. A summary of the basic characteristics of the test sites is
and using an NDWI threshold of 0 does not accurately enable discrimi- shown in Table 1.
nating built-up surfaces from water pixels. Xu (2006) therefore pro- The test sites were deliberately selected so that the sub-scenes
posed another index, called Modified Normalized Difference Water consist of complex surface features, such as hill shade, built-up areas
Index (MNDWI), where McFeeters (1996) NDWI was modified by and other dark surfaces as background to the water bodies. The test
replacing band 4 by band 5 of Landsat 5 TM. The MNDWI of Xu (2006) sites in Switzerland, Ethiopia and South Africa are characterized by the
is one of the most widely used water indices for various applications, presence of built-up surfaces and shadows of mountains. The site in
including surface water mapping, land use/cover change analyses Denmark also consists primarily of urban background but with no
and ecological research (Davranche, Lefebvre, & Poulin, 2010; Duan major shadow problems since the terrain is predominantly flat and
& Bastiaanssen, 2013; Hui, Xu, Huang, Yu, & Gong, 2008; Poulin, tall buildings in the urban area are rare. The test site in New Zealand
Davranche, & Lefebvre, 2010). consists of mountain slopes with deep shadows, but no major urban
Even though a number of water extraction techniques are described surfaces are included.
in the literature, the choice between them is constrained by accuracy In addition to the five test sites for which detailed accuracy analyses
problems. Environmental monitoring and change detection techniques and comparisons were carried out, further validation of the robustness
such as post-classification comparison are likely to be less reliable of the new index was undertaken considering shadow-dominated
when classifiers of low accuracy are used (Congalton & Green, 2009; water bodies in Norway, rivers with urban surfaces and shadows from
Mucher, Steinnocher, Kressler, & Heunks, 2000). For instance, in a tall buildings in Shanghai, China, and several crater lakes with built-up
study focusing on water dynamics monitoring, Ji et al. (2009) faced background surfaces in Bishoftu, Ethiopia. However, these additional
two major problems in appropriately using water indices: first, the re- test sites were not analyzed in detail and classification output from
sults obtained using different indices were inconsistent and unreliable; these sites is not included in the Results section; instead, the classifica-
second, the threshold values applied to distinguish water from non- tion maps are included in Appendix A for visual inspection of classifica-
water were unstable, varying with scene and locations. These authors tion accuracy.
compared four different water indices using simulated datasets of
four satellite sensors: Landsat ETM+, Système Pour l'Observation de 2.2. Landsat images
la Terre (SPOT), the Advanced Space-borne Thermal Emission and
Reflection radiometer (ASTER), and the Moderate Resolution Imaging Landsat 5 TM images were acquired from USGS GLOVIS portal
Spectroradiometer (MODIS), aiming to identify the best method for (United States Geological Survey (USGS), 2012). All Landsat images
delineating water features. Among the four alternatives, they found used are of product type L1T and with a scene quality score of 9,
that the MNDWI performed best in delineating water, and featured which means perfect scenes with no errors detected. The images were
the most stable threshold. also georeferenced with precision better than 0.4 pixels (NASA, 2012).
Water classification accuracy problems may be especially pro- The sub-scenes were all free of clouds. Descriptions of the Landsat
nounced in areas where the background land cover includes low albedo images are presented in Table 2.
surfaces such as asphalt roads in urban areas, and shadows from
mountains, buildings and clouds. The presence of shadows may cause 2.3. Reference data
misclassification due to the similarity in reflectance patterns, and this
may lessen the accuracy of surface water mapping and change analysis Reference data used in accuracy assessment are described in Table 2.
(Frey, Huggel, Paul, & Haeberli, 2010; Verpoorter et al., 2012; Xu, 2006). For the test site in Denmark, colored Digital Orthophoto Quadrangles
In environments where nonwater dark surfaces are found, simple clas- (DOQs) from year 2010 were used as reference. These aerial photos
sification methods such as two-band water indices and single-band have a spatial resolution of 12.5 cm and location accuracy better than
thresholding may not sufficiently and accurately distinguish be- 0.5 m (COWI, 2010). For the four other test sites, high spatial resolution
tween water pixels and nonwater dark surfaces, particularly shadows images provided by Google Earth™ were used for reference. The acqui-
(Verpoorter et al., 2012). In a study of land cover dynamics using sition dates of the reference data and the Landsat 5 TM images were
 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35 25

Table 1
Characteristics of the study sites.
The source of climate information is: (http://www.climatedata.eu/).

Country and name of water bodies Center point coordinate Area General characteristics Mean alt. (m) Topography Climate
(UTM) (ha) of water bodies

Denmark
Several artificial lakes, a harbor and 6,172,085 m N, 12°34′57.42″E 2085 Shallow clear artificial 9 Predominantly flat Temperate
the sea (Øresund and Køge Bugt) lakes, clear seas
Switzerland
Lake Lauerz 5,209,030 m N, 469,608 mE 289 Clear lake 1100 Mountainous Temperate
Ågeri lake 5,218,774 m N, 471,530 mE 719 Clear lake
Sihl lake 5,218,191 m N, 484,028 mE 1034 Clear lake
Wägitaler lake 5,214,616 m N, 494,092 mE 402 Clear lake
Klöntaler lake 5,207,839 m N, 498,040 mE 309 Clear lake
Ethiopia
Gefersa 1,002,432 m N, 459,709 m E 144 Clear reservoir 2377 Mountainous Tropical dry
Dire 1,011,794 m N, 493,000 m E 106 Turbid reservoirs
Legedadi 1,002,374 m N, 497,446 m E 423 Turbid reservoirs
South Africa
Berg river 6,244,161 m S, 320,605 m E 426 Clear reservoir 600 Rugged hilly Subtropical semi-arid
Wemmershoek 6,255,473 m S, 323,355 m E 195 Clear reservoir
Brandvlei 6,265,857 m S, 354,413 m E 3097 Clear reservoir
New Zealand
Lake Te Anau 5,004,239 m S, 723,800 m E 6495 Large clear lake 800 Rugged hilly Humid temperate

closely matched to minimize bias in the surface water boundaries that Orians, & Wolfe, 1992)). Water vapor calibration constants for each of
could arise because of large differences in time. The dates of acquisition the Landsat images are summarized in Table 3. The overpass time of
of the Landsat images and reference data are shown in Table 2. Landsat TM and MODIS Terra at each test site was closely matched
The “true” boundaries of all the test water bodies were digitized (less than 2 h difference).
manually on-screen from the reference data. In the analysis, the manu-
ally digitized water map was used to assess the accuracy of the different 3:912
VIS ¼  Z: ð1Þ
water extraction methods when applied to the Landsat images. AOD

Image-to-image co-registration between the reference data and

3. Methods Landsat images was undertaken for the test sites in Denmark and
Ethiopia. Manual co-registration was performed with a Root Mean
3.1. Image preprocessing Square Error (RMSE) of less than 0.4 pixels. At least 25 control points
were used for co-registration of each image with the reference data.
Landsat 5 TM images acquired in the form of raw digital number The co-registration between Landsat and Google Earth™ images at the
were calibrated to surface reflectance values. Atmospheric correction rest of the test sites was already highly accurate and hence there
was applied to all images using the Fast Line-of-Sight Atmospheric Anal- was no need for manual co-registration. Since the L1T Landsat TM and
ysis of Spectral Hypercubes (FLAASH) module in ENVI v. 4.8 (Exelis ETM+ products are geometrically corrected (NASA, 2012), no such
Visual Information Solutions, 2010). Aerosol Optical Depth (AOD) and corrections were applied in the pre-processing.
total Column Water Vapor data to be used in the FLAASH atmospheric
correction module were retrieved from MODIS Terra atmospheric prod- 3.2. Pure-pixel selection
ucts (Jimenez-Munoz, Sobrino, Mattar, & Franch, 2010) for each of the
five test sites. The MODIS Terra products that had been acquired on An independent set of “pure” pixel reflectance values of nine major
the same date as the Landsat images were used to retrieve the calibra- land cover types was sampled from the six reflective bands of a Landsat
tion constants. Initial visibility was estimated using the aerosol optical 5 TM image of Addis Ababa, acquired on Dec. 9, 2010. The land cover
thickness obtained from MODIS data using Eq. (1), where, VIS = initial types are: water, vegetation (forest and non-forest), bright soil, dark
visibility, Z = mixing layer height and AOD = Aerosol Optical Depth. soil, brown soil, bright built, asphalt, other dark built and shadow.
The mixing layer height values for the test sites were not available. Spectral data from these pure pixels were used to examine reflectance
Therefore, a daytime mixing layer height of 1.5 km was used for all patterns and identify land cover types that affect water extraction accu-
test sites (typical values range from 1 to 2 km (Butcher, Charlson, racy, aiming to design a method that accurately discriminates between

Table 2
Description of Landsat TM scenes and corresponding reference data.

Test site Landsat scene Reference data and sources

Acquisition date Path Row

Water bodies in Copenhagen, Denmark June 29, 2010 195 21 Colored Digital Orthophoto quadrangles acquired between May and July 2010, ©COWI
Lakes in Switzerland Sept 30, 2009 195 27 Google Earth™ image acquired on Jul 1 2009
Reservoirs in Addis Ababa, Ethiopia Dec 9, 2010 168 54 Google Earth™ image acquired on Oct 13 and Dec 20, 2010, ©Digital globe, CNES/SPOT Image
Reservoirs in South Africa Mar 29, 2010 175 83 Google Earth™ image acquired on Jan 4, Jan 7 and Jan 18, 2010, ©GeoEye and AfriGIS
A lake in New Zealand Feb 3, 2010 76 91 Google Earth™ image acquired on Apr 09, 2010 and Feb 16, 2011, ©GeoEye
26 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35

Table 3
Calibration values used in atmospheric correction using FLAASH.

Test Site Date of Landsat/MODIS Average water vapor column Water vapor multiplier Average aerosol optical Visibility (km)
Terra overpass from MODIS (g/cm2) thick-from MODIS

Denmark Jun 29, 2010 3.00 1.03 0.33 17.8

Switzerland Sept 30, 2009 1.43 0.49 0.08 73.4
Ethiopia Dec 9, 2010 2.30 0.56 0.10 58.7
South Africa Mar 29, 2010 1.90 0.46 0.033 177.8
New Zealand Feb 3, 2010 2.67 0.91 0.12 48.9

such surfaces and water. The pure pixel data were intended neither for similarities in reflectance patterns. In addition to enhancing separability
classification nor accuracy assessment and therefore, only the image of of water and nonwater pixels, the choice of the coefficients also aimed
Addis Ababa and its surroundings was used for pure pixel extraction. to stabilize the threshold needed to distinguish water from nonwater
The reason for choosing Addis Ababa for pure pixel extraction was pixels by forcing nonwater pixels below 0 and water pixels above 0,
that this area includes all the major challenging features influencing implying that 0 could be used as a reasonable starting threshold for
water extraction accuracy: shadow, dark built-up surfaces and other classifying land cover into binary classes of water and nonwater under
low albedo surfaces such as black soil. a wide range of environmental conditions.
The methods used to extract pure pixels of the selected land cover
types include spectral feature space scatter plot from Minimum Noise AWEInsh ¼ 4  ðρband2 −ρband5 Þ−0:25  ρband4 þ 2:75  ρband7 ð2Þ
Fraction Transform (MNFT) images, Pixel Purity Index (PPI), manual
digitization from images accessed through Google Earth™, ground- AWEIsh ¼ ρband1 þ 2:5  ρband2 −1:5  ðρband4 þ ρband5 Þ−0:25
based land cover assessment and the familiarity of the first author  ρband7 ð3Þ
with the local area. Pure pixel samples for water were taken from the
middle of lakes to avoid mixed edge pixels. Similarly, high forest where ρ is the reflectance value of spectral bands of Landsat 5 TM:
with closed canopy from Menagesha national forest was applied for band 1 (blue), band 2 (green), band 4 (NIR), band 5 (SWIR) and band
sampling vegetation pure pixels. Pure pixels of built-up land cover 7 (SWIR).
were sampled from homogenous surfaces such as airport runways and AWEInsh is an index formulated to effectively eliminate nonwater
large warehouse roofs in Addis Ababa city. Since land cover types in pixels, including dark built surfaces in areas with urban background
Addis Ababa are highly heterogeneous, detection of pure built pixels and AWEIsh is primarily formulated for further improvement of accuracy
was assisted by PPI and spectral feature space in ENVI v.4.8 (Exelis by removing shadow pixels that AWEInsh may not effectively eliminate.
Visual Information Solutions, 2010). Shadow pixels were extracted by The subscript “nsh” in Eq. (2) is included to specify that the index is
thresholding hill-shade images derived from elevation data using the suited for situations where shadows are not a major problem. The sub-
ASTER Digital Elevation Model (DEM) in mountainous parts of the test script “sh” in Eq. (3) indicates that the equation is intended to effectively
site. Homogenous agricultural fields with exposed black, brown and eliminate shadow pixels and improve water extraction accuracy in
bright soils were also sampled from the outskirts of the city. areas with shadow and/or other dark surfaces. But in areas with highly
For each land cover type, 312 pure pixels were extracted from the six reflective surfaces such as ice, snow and reflective roofs in urban areas,
reflective bands of the Landsat 5 TM images. Average reflectance values Eq. (3) may misclassify such surfaces as water.
of the pure pixels are shown in Fig. 1. Separability of the spectral In Eq. (2), quadrupling the difference between the band 2 and band 5
signatures of the selected nine major land cover types was tested by results in large positive values for water pixels and negative values for
Jeffries–Matusita's pairwise separability measure (Richards, 1993) in most nonwater pixels. To help in discriminating water from other
ENVI v. 4.8. All pairs of land cover types were found to be separable surfaces that have similar spectral patterns, band 4 and band 7 are
with values ranging from 1.89 to 2.0. subtracted from the result and different weights are assigned to these
bands to force nonwater pixels to have even larger negative values;
this subtraction will not to any greater extent lead to negative values
3.3. Formulation of the Automated Water Extraction Index (AWEI) for water pixels because water has very low reflectance in the spectral
ranges of band 4 and band 7. The equation results in large negative
Five spectral bands of Landsat 5 TM were used in developing the values for pixels covered by vegetation, soil, bright built and other
new index (AWEI) to increase the contrast between water and other surfaces that have large reflectance for band 4 or 7. The equation is
dark surfaces. The primary aim of the formulation of AWEI was to also intended to enhance separability between water, dark surfaces
maximize separability of water and nonwater pixels through band and other nonwater surfaces. In many cases, water absorbs almost all
differencing, addition and applying different coefficients. Accordingly, of the incoming radiation in bands 4, 5 and 7 and achieves relatively
two separate equations are proposed to effectively suppress nonwater highest reflectance between bands 1 and 2 of Landsat 5 TM (Lillesand,
pixels and extract surface water with improved accuracy (Eqs. (2) and Kiefer, & Chipman, 2004). Shadowed surfaces also have low reflectance
(3)). The coefficients used in Eqs. (2) and (3) and the arithmetic combi- in all spectral bands, but the magnitude of reflectance varies due to var-
nations of the chosen spectral bands were determined based on critical iation in surface characteristics and the depth of shadow. Hence, Eq. (2)
examination of the reflectance properties of various land cover types. alone may not completely eliminate all types of shadows and other low
The coefficients of these equations are empirical results determined albedo surfaces. As shown in Fig. 1, for instance, subtracting band 5 from
based on reflectance patterns observed across the dataset of pure pixels band 2 could yield positive values for both water and shadows. Quadru-
of various land cover types. An iterative process was applied to identify pling the difference and subtracting bands 4 and 7 may result in some
parameters that maximize the separability of water and nonwater sur- shadow pixels obtaining similar values as that of water due to the
faces characterized by low reflectance. In the final index, the coefficients similarity in reflectance patterns within these bands, hence making it
were rounded for ease of use. Particular emphasis was given to the en- difficult to exclude shadow pixels from water class.
hancement of the separability of water and dark surfaces such as shad- Due to these limitations of Eq. (2), Eq. (3) was formulated to achieve
ow and built-up structures that are often difficult to distinguish due to enhanced separability of water and shadows and/or dark surfaces. It can
 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35 27

0.30
Blue band (b1) Green band (b2) Red band (b3)
0.25
Reflectance

0.20

0.15

0.10

0.05

0.00

0.4 NIR (b4) SWIR (b5) SWIR (b7)

0.3
Reflectance

0.2

0.1

0.0

1.0

0.5
Index value

0.0

-0.5

-1.0
Eq. 2, AWEInsh Eq. 3, AWEIsh MNDWI
-1.5
est ter ow en halt uilt uilt soil soil soil est ter ow een halt uilt uilt soil soil soil est ter ow een halt uilt uilt soil soil soil
For WaShador. gre Asp ark bight bright lack rown For W aShador. gr Asp ark bight bright lack rown For W aShador. gr Asp ark bight bright lack rown
n-f D Br B B B n-f D Br B B B n-f D Br B B B
No No No
Surface category

Fig. 1. Reflectance distributions of pure pixels of major land cover types. Each box plots shows the location of the 10th, 25th, 50th, 75th, and 90th percentiles using horizontal lines (boxes
and whiskers) and the circles are 5th and 95th percentiles.

be noted from Fig. 1 that the largest difference between the reflectance no shadowed areas, no dark urban backgrounds and no high-albedo
of water and shadow is found in bands 1 and 2. Therefore, adding these surfaces, either of the two can be used alone.
two bands, while at the same time multiplying band 2 by the specified
coefficient, enhances the separability between water and shadow 3.4. Classification, threshold optimization and per-pixel accuracy assessment
pixels, yielding relatively large positive values for water pixels com-
pared to shadow pixels. Subtracting bands 4, 5 and 7 forces nonwater At the test sites in Denmark and Ethiopia, urban background
pixels in the negative direction, and the net effect of this subtraction dominates the sub-scenes. Therefore, both equations of AWEI (AWEInsh
on water pixels is minimal compared to nonwater surfaces including and AWEIsh) were applied sequentially: first, AWEInsh was applied to the
shadows, which are forced considerably below zero. Band 3 was not image; next, AWEIsh was used to eliminate misclassified pixels with
used in Eqs. (2) and (3) because during the preliminary tests, including shadows and other dark surfaces. At the test sites in Switzerland, New
this band did not improve separability and accuracy. From the arithmet- Zealand and South Africa, only AWEIsh was applied because urban
ic formulation of Eq. (3), it may be noted that the addition of the short surfaces are rare in these sites.
wave bands (bands 1 and 2) may result in large positive values for To compare accuracy of the proposed water extraction technique
high albedo surfaces such as ice, cloud, and highly reflective building with other methods, we made preliminary tests of various water indices
roofs. Eq. (3) may therefore not be able to distinguish these high- including the Water Index (WI) of Ouma and Tateishi (2006), the
albedo surfaces from water. Normalized Difference Water Index (NDWI) of McFeeters (1996) and
Therefore, the intended use of the two AWEI equations is as follows: other indices that Ji et al. (2009) used in their studies. Based on this pre-
1) in situations where shadows are major sources of accuracy loss but liminary evaluation, it appeared that all indices, except the MNDWI,
surfaces such as snow, ice and high albedo built surfaces are not present, performed poorly at our test sites. We therefore only considered
AWEIsh alone is proposed to automatically enhance the separability of MNDWI for comparison with the new index proposed in this paper. A su-
pixels of water from nonwater (more importantly from shadow pixels) pervised maximum likelihood (ML) classifier was also included in our
so that application of a threshold close to 0 is suitable for the extraction comparison as this classifier is one of the most widely used methods in
of surface water; 2) in areas where shadows are not a major problem, land cover classification. For the ML classifier, water and nonwater train-
AWEInsh alone is proposed; 3) in conditions where both high albedo ing data were produced for each test site. The minimum size of reference
surfaces and shadow/dark surfaces are found, we propose using datasets for training was determined using the multinomial conservative
Eqs. (2) and (3) sequentially in a classification tree; 4) in areas with sample size equation described in Congalton and Green (2009). The
28 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35

reference data were generated by digitizing multiple polygons on the where, f12 and f21 denote the frequencies of cases that are correctly clas-
true-color composites of Landsat bands and evenly distributing the sam- sified by one classification method but wrongly classified by the other.
ples across all parts of the sub-scenes. It was easy to generate large refer-
ence data units since the classes considered are only water and nonwater 3.5. Sub-pixel accuracy assessment
and it is relatively easy to visually distinguish between water and
nonwater surfaces from high spatial resolution images retrieved through The sensitivity of different classifiers to various mixtures of water
Google Earth™. These images were used to differentiate nonwater dark and nonwater was evaluated using sub-pixel commission–omission
areas from water surfaces. No separate validation data were necessary errors and by plotting graphs showing the cumulative percentage of
for accuracy assessment of the ML classifier since the classification result edge pixels classified as water against the proportion of each individual
was compared against the true map of water. pixel covered by water for mixed edge pixels at test sites in Denmark,
Since the AWEI equations are formulated to enhance separability of Switzerland and Ethiopia. Detailed sub-pixel accuracy analysis and
water and nonwater pixels by applying coefficients that force nonwater comparisons were undertaken using the three reservoirs in Ethiopia
pixels below 0 and water pixels above 0, a threshold of 0 can be used as (Gefersa, Dire and Legedadi). The total number of mixed edge pixels
a default starting point. But due to variation in scene brightness and in the three reservoirs was 1819 (164 ha). In the sub-pixel accuracy as-
contrast with time and space, the default threshold may not always sessment, commission and omission errors brought about by edge
result in the highest possible water extraction accuracy that can be pixels were quantified by the use of an overlay analysis in ArcGIS. Any
achieved by application of the index. In order to determine the optimal pixels that included water and nonwater surfaces were considered to
threshold, multiple thresholds were considered, and for each threshold be mixed edge pixels (Fig. 2). If a mixed edge pixel was classified as
value corresponding commission errors (over-estimation) and omis- water, the fraction of it that fell outside the “true” boundary was consid-
sion errors (under-estimation) were calculated and the percentage ered to be sub-pixel commission error. Similarly, in cases where mixed
errors were plotted against threshold values. The intersection point of pixels are classified as nonwater, the fraction of these pixels that fell
commission and omission error graphs was then considered as the inside the “true” water body was considered to be an omission error
optimal threshold since it approximates the minimum possible sum of at the sub-pixel level. Mixed pixels consisting predominantly of water
the two error types. We evaluated the stability of optimal thresholds (N50% water) should ideally be classified as water and vice versa. In
of the new method and of MNDWI by examining the variation of the the sub-pixel accuracy assessment, influences of misregistration arti-
optimal threshold values for the two indices across the five test sites. facts and manual digitization of true water boundaries were assumed
Classification accuracy of the three methods, i.e. AWEI, MNDWI and to be insignificant.
ML, was assessed by calculating kappa coefficients and error matrices.
The accuracy comparison between AWEI and MNDWI was made at
their optimal thresholds. McNemar's statistical test was applied to 4. Results
examine whether the new water extraction method significantly
improves accuracy compared to MNDWI and ML in the test sites. 4.1. Water extraction maps
McNemar's continuity corrected chi-square statistic was computed as
shown in Eq. (4) (De Leeuw et al., 2006): The outputs of water extraction using the three classifiers at the five
test sites are presented in Fig. 3. Visual inspection of Fig. 3 indicated that
AWEI resulted in better accuracy of surface water mapping compared to
MNDWI and ML. Particularly at test sites in Switzerland, South Africa
2
2 ðj f 12 − f 21 j−1Þ and New Zealand, the new index (AWEI) was consistently better in sup-
X ¼ ð4Þ
f 12 þ f 21 pressing shadow and other nonwater surfaces. In most cases, MNDWI

Fig. 2. Edge pixels around Gefersa reservoir (Ethiopia) showing mixed pixels with different proportions of water (shown on high spatial resolution image accessed through Google Earth™).
 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35 29

Test site A WEI M NDWI ML

Denmark

Switzerland

Ethiopia

Site A WEI M NDWI ML

S. Africa

New
Zealand

Fig. 3. Comparison of water extraction results using three classifiers at the five test sites.

and especially ML produced noisy results. However, at test sites in and Shanghai where dark shadows were abundant, visual inspection
Denmark and Ethiopia, visual inspection of Fig. 3 indicated smallest clearly shows that AWEIsh suppressed shadowed surfaces more effec-
difference among the three classification methods. tively than MNDWI (shown in Appendix A).
Visual inspection of classification outputs at the three additional test
sites shown in Appendix A (Figs. A1–A3) also indicates that AWEI is 4.2. Classification accuracy and edge pixel effects
effective in extracting surface water in the presence of shadow and
urban surfaces. At Bishoftu lakes in Ethiopia, where no major shadow The results of mapping accuracy at each of the five main test sites are
surfaces were present, both AWEIsh and MNDWI resulted in (visually) summarized in Table 4. At all test sites the accuracy achieved by AWEI
similar classification outputs. By contrast, at the test sites in Norway was higher than that of the MNDWI and ML classifiers. Averaged over
30 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35

Table 4 Table 5
Summary of classification accuracy of the three classifiers by test site. Summary of McNemar's continuity corrected χ2 test for differences in classification
accuracy.
Classifier Denmark Switzerland Ethiopia S. Africa New Zealand
Test sites Classifier χ2 P-value
Kappa coeff. Kappa coeff. Kappa coeff. Kappa coeff. Kappa coeff.
MNDWI ML MNDWI ML
AWEI 0.93 0.95 0.97 0.98 0.98
MNDWI 0.92 0.89 0.95 0.94 0.90 Denmark AWEI 0.8 114 0.30 0.00
ML 0.89 0.81 0.93 0.62 0.97 Switzerland AWEI 408.0 619 0.00 0.00
Ethiopia AWEI 19.0 30 0.00 0.00

the five test sites the total omission and commission error of AWEI was
only about 50% of that of the MNDWI and 25% of that of the ML classi- Based on the sum of total overestimation and underestimation of edge
fiers (Fig. 4). Details of accuracy assessment including users' and pro- pixels, AWEI performed slightly better than MNDWI, and ML achieved
ducers' accuracy are shown in Appendix A (Table A1). Since visual the lowest accuracy in classifying mixed edge pixels.
inspection from Fig. 3 indicates small variation in accuracy at test sites
in Denmark and Ethiopia, McNemar's chi-square test of significance of
accuracy difference at the test sites in Denmark, Switzerland and 4.3. Optimal threshold and its variability
Ethiopia are included in Table 5. At these three test sites, significant ac-
curacy improvement was achieved by AWEI (P-value b 0.01) compared A comparison of the stability of the optimum thresholds of AWEI and
to ML. At the test site in Denmark, accuracy difference between AWEI MNDWI is shown in Fig. 6. It clearly appears that the optimal threshold
and MNDWI was insignificant (Table 5). ML performed worst at test of MNDWI at different test sites exhibited large variation compared
site in South Africa (kappa coefficient 0.62) and at this test site, the to AWEI. The optimal threshold of MNDWI ranged from 0.005 in
highest accuracy was achieved by AWEI, with a kappa coefficient of Denmark to 0.6 in South Africa, whereas for AWEI the optimal threshold
0.98 (Table A1 in Appendix A). only varied from −0.15 (AWEInsh in Denmark) to 0.045 (AWEIsh in
The sub-pixel accuracy analysis is presented in Fig. 5. The compari- South Africa), and in the three other sites the optimal threshold of
son shows the ability of the three classifiers in correctly classifying AWEIsh was 0 (Fig. 6).
edge pixels with various mixtures of water and nonwater components.
The vertical line in Fig. 5 indicates the 50% water–nonwater mixture and 5. Discussion and perspectives
the figure shows that among the edge pixels that AWEI classified as
water, only 13% were predominantly nonwater. Conversely, 87% of The new water extraction index introduced in this paper contributes
mixed edge pixels that were classified as water were correctly classified to the efforts being made to improve the accuracy of surface water
by AWEI. Using MNDWI, 81% of the mixed edge pixels were correctly mapping and change analysis for various environmental studies and
classified. Further analysis of mixed edge pixels at test sites in Addis applications. This method uses a simple and systematic technique of
Ababa showed that sub-pixel commission error of AWEI corresponded enhancing class separability without a need for additional data to
to an overestimation of 16.6 ha (total area of predominantly nonwater remove shadow and dark surface noises, which are often major causes
edge pixels classified as water), and omission error corresponded to of misclassification in surface water mapping. Using a simple classifica-
4.3 ha (total area of predominantly water edge pixels classified as tion tree approach, the AWEI was shown to extract surface water with
nonwater). For comparison, edge pixel commission and omission errors high accuracy, particularly in mountainous areas where hills cast
of MNDWI corresponded to overestimation and underestimation of shadows on background surfaces and in urban areas with complex
18.3 ha and 4 ha, respectively. At this site edge pixel omission and com- land cover. AWEI is not only a simple technique but was also shown
mission error of ML corresponded to 49.3 ha and 0.6 ha, respectively. to be robust under various environmental conditions and for different
types of water bodies.

60
Total error (% omission and commission)

South Africa

100
Cum. % of edge pixels classified as water

A
M
L
50

40 80 A AWEI
M MNDWI
New Zealand
L Max. Like.
M
30 A
60
L
M
20 New Zealand
Denmark
A
40 L
Denmark Ethiopia M
Denmark South Africa
10 New Zealand Ethiopia
A
M
Ethiopia
L
South Africa Switzerland 20 M A
L
Switzerland Switzerland
0 M L
A
L
M
L A
AWEI MNDWI ML M
L A
M
L A
0 M
A
L A
Classifier
0 20 40 60 80 100
Fig. 4. Total classification error (combined commission and omission error). The box plots Percentage of water in edge pixels
show the variability of classification errors among test sites. Each box plots shows the
location of the 10th, 25th, 50th, 75th, and 90th percentiles using horizontal lines Fig. 5. Cumulative frequency of mixed edge pixels classified as water (average of test sites
(boxes and whiskers) and the circles are 5th and 95th percentiles. in Denmark, Switzerland and Ethiopia).
 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35 31

1.0 waterline are of interest, the accuracy of classifying mixed edge pixels
may become an important issue.
As mentioned in the Results section, when applying the ML classifier
x to the reservoirs around Addis Ababa, a substantial number of edge
0.5
pixels that predominantly consist of water were classified as nonwater,
x thus obviously leading to underestimation of surface water extents. The
Index value

sh x x implication of this could be that even if the water boundary increases by

0.0
sh
x x
sh x x
sh x
certain distance, thus changing the proportion of water in mixed edge
nsh
x pixels, say from 40% to 60%, the ML method could still classify the pixels
as nonwater since it seems that the classifier is sensitive to nonwater
components of the mixture. The ability of different classifiers to classify
-0.5
such mixed pixels correctly into water and nonwater classes may
vary depending on the spectral bands and algorithms used. The
nonwater components of the mixture could be composed of many
-1.0 combinations of different land cover types. So, the reflectance values
ark rland iopia Africa aland ark rland iopia Africa aland of mixed pixels can vary considerably, even for pixels where the pro-
nm e Eth uth w Ze nm e Eth uth w Ze
De Switz De Switz
So Ne So Ne portion of water is similar (Ji et al., 2009). The relative improvement
AWEI MNDWI in sub-pixel accuracy achieved by AWEI may make it suitable for
consistent and reliable estimation of surface water dynamics using
Fig. 6. Threshold variability and distribution of index values for AWEI and MNDWI. Dashed Landsat data.
lines show mean optimal threshold of the five test sites, and symbol “x” shows optimal Though the new water extraction index was tested under wide
threshold for each site.
range of environmental conditions and water body types, several vari-
ables that were not considered at our test sites are likely to affect the ac-
curacy of water extraction methods. Seasonal and daily variation in the
angle of the sun, atmospheric composition, and changes in biophysical
and chemical properties of water bodies, such as changes in phyto-
In many water indices, the lack of stability of the threshold is a prob- plankton (Zhang et al., 2010) may influence the reflectance patterns of
lem (Ji et al., 2009), making it difficult to decide which value should be water bodies. The use of different atmospheric correction methods
used in classification trees. The lack of a reasonably stable threshold may also influence thresholds and accuracies. Therefore one may need
may make the classification more time-consuming and lead to a subjec- to consider the importance and type of atmospheric correction applied
tive choice of threshold which may also affect accuracy. In addition to in the image preprocessing stage in evaluating accuracies of different
accuracy improvement, our new index was also shown to have a rela- water extraction methods. AWEI was tested using Landsat TM data
tively stable optimal threshold which makes the use of the method only and its use may therefore need to be evaluated on data from
even simpler. It should be noted that in our study, images from all test other sensors.
sites were atmospherically corrected applying the FLAASH module in In our test cases, we did not consider the influence of seasonal vari-
ENVI v.4.8. In classifying images that are calibrated to TOA reflectance, ation in appearances of water bodies. Therefore, the robustness of the
but with no atmospheric corrections, the optimal thresholds may differ new method also needs to be tested in different seasons. In addition
slightly from what is observed in this study. to the five test sites examined in details, we included three additional
A number of authors contributed to previous research on the im- sites in Appendix A (Figs. A1–A3). Nevertheless, more sites may need
provement of surface water mapping accuracy using remotely sensed to be included for a thorough evaluation of the performance of the
data, including those that have emphasized the improvement of general index.
land cover classification accuracies (Aguirre-Gutierrez, Seijmonsbergen,
& Duivenvoorden, 2012; Rozenstein & Karnieli, 2011) and surface water
mapping in particular (Ji et al., 2009; Sun et al., 2012; Verpoorter et al., 6. Conclusion
2012; Xu, 2006). In a recent work of Verpoorter et al. (2012), a six-step
water extraction method called GeoCover Water Bodies Extraction The main purpose of this study was to devise a method that im-
Method (GWEM) was introduced. In GWEM, the authors proposed a proves water extraction accuracy by increasing spectral separability be-
combination of various classification techniques for improvement of tween water and nonwater surfaces, particularly in areas with shadows
accuracy. The same authors identified shadows of cloud and mountains and urban backgrounds that are often major causes of low classification
as major sources of accuracy problems and proposed a method where accuracy. Using Landsat 5 TM data, we introduced a new automated
elevation data were used to detect shadows and water surfaces that water extraction method (AWEI) and compared its per-pixel and sub-
overlap with shadow were removed from the classification dataset. pixel accuracy and threshold stability with that of the MNDWI and
Our new method which automatically suppresses shadow pixels with- ML classifiers. AWEI significantly improved accuracy in areas where
out the need for other data input or separate shadow detection proce- shadow and other dark surfaces were the main sources of classification
dures, may ease surface water mapping, particularly in situations errors.
where mapping, monitoring and change detection of surface water A sub-pixel analysis of errors at the edges of water bodies revealed
resources across multiple scenes or over regional and global scales are that the AWEI classifier was relatively more accurate in classifying
required. edge pixels compared to the MNDWI and ML classification methods. Be-
Despite a number of surface water mapping and accuracy improve- sides, the optimal threshold of AWEI was shown to be less variable with
ment methods reported in the literature, limited research has been un- images of different locations and times compared to that of MNDWI.
dertaken on accuracy assessment at sub-pixel level. This is particularly Therefore, AWEI is proposed as an alternative and improved water
important when satellite images such as Landsat are used. Because of index, especially in extracting water information from areas where
the limited spatial resolution of reflective bands of Landsat TM, edge noisy results are expected because of the presence of shadows and
pixels cover relatively large areas which likely consist of a mixture of built-up surfaces. This new method would also be suitable for surface
water and nonwater components. In the use of Landsat TM data for water change detection studies since it classifies edge pixels with high
environmental studies where monitoring and detecting changes in accuracy and with a stable threshold.
32 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35

Appendix A

AWEIsh at 0 threshold AWEIsh at 0.17 threshold

MNDWI at 0.0 threshold MNDWI at 0.7 threshold

Landsat true color composite Google 3D viewshowing water and mountain shadow

Fig. A1. Water extraction images applying AWEIsh (top row) and MNDWI (middle row) at test site in Norway using Landsat 5 TM acquired on Aug 13, 2011 (2382 by 2382 pixels, top-left
corner coordinate in UTM: 6,989,655 N, 371,295 E).
 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35 33

AWEIsh at 0.15 threshold AWEIsh at 0.2 threshold AWEIsh at 0.25 threshold

MNDWI at 0.2 threshold MNDWI at 0.4 threshold MNDWI at 0.5 t hreshold

Landsat true color composite Google 3D view showing building shadow and river
Fig. A2. Water extraction images applying AWEIsh (top row) and MNDWI (middle row) at test site in China Shanghai using Landsat ETM+ acquired on Nov 27, 2002 (400 by 400 pixels,
top-left corner coordinate UTM: 3,461,925 N, 350,415 E).

AWEIsh at 0.0 threhold AWEIsh at 0.1 threhold MNDWI at 0.0 threshold MNDWI at 0.15 threshold

X X

X
X X
X
X

Fig. A3. Water extraction images applying AWEIsh (top row) and MNDWI (middle row) at test site in Bishftu Ethiopia using Landsat ETM+ acquired on Nov 27, 2002 (400 by 400 pixels,
top-left corner coordinate UTM: 3,461,925 N, 350,415 E). Location of actual water bodies is shown by “X” mark on true color composite of the Landsat ETM+ image (bottom image).
34 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35

Table A1
Summary of accuracy assessments at the five main test sites showing various accuracy measures.

Test site Classification method Threshold Land cover class User accu. Produc accu. Kappa Comm. error % Omi. error % Total error %
a
Denmark AWEInsh 0.000 Water 97.08 91.43 0.93 2.92 8.57 11.49
Nonwater 98.57 99.54 1.43 0.46 1.89
−0.2 Water 96.35 92.89 0.94 3.65 7.11 10.77
Nonwater 98.81 99.41 1.19 0.59 1.79
−0.15 Water 98.30 92.58 0.95 1.70 7.42 9.12
Nonwater 98.75 99.72 1.25 0.28 1.53
MNDWI 0.00 Water 95.13 91.04 0.92 4.87 8.96 13.82
Nonwater 98.50 99.22 1.50 0.78 2.28
0.05 Water 97.10 89.89 0.92 2.90 10.11 13.02
Nonwater 98.32 99.55 1.68 0.45 2.14
0.1 Water 98.05 88.65 0.92 1.95 11.35 13.31
Nonwater 98.12 99.70 1.88 0.30 2.18
MaxLike Water 96.85 84.61 0.89 3.15 15.39 18.53
Nonwater 97.46 99.54 2.54 0.46 3.00
Switzerland AWEIsh −0.050 Water 52.89 96.77 0.66 47.1 3.2 50.3
Nonwater 99.84 95.81 0.2 4.2 4.4
0.100 Water 99.34 76.61 0.86 0.7 23.4 24.1
Nonwater 98.88 99.98 1.1 0.0 1.1
0.000 Water 99.01 90.96 0.95 1.0 9.0 10.0
Nonwater 99.56 99.96 0.4 0.0 0.5
MNDWI −0.005 Water 99.34 76.61 0.94 0.7 23.4 24.1
Nonwater 99.84 95.81 0.2 4.2 4.4
0.100 Water 73.92 95.64 0.82 26.1 4.4 30.4
Nonwater 99.79 98.36 0.2 1.6 1.9
0.300 Water 87.99 91.92 0.89 12.0 8.1 20.1
Nonwater 99.61 99.39 0.4 0.6 1.0
0.500 Water 96.76 78.95 0.86 3.2 21.1 24.3
Nonwater 98.99 99.87 1.0 0.1 1.1
MaxLike – Water 74.46 92.08 0.81 25.5 7.9 33.5
Nonwater 99.61 98.47 0.4 1.5 1.9
Ethiopia AWEIsh −0.050 Water 95.47 98.85 0.95 4.53 1.15 5.68
Nonwater 100.00 99.97 0.00 0.03 0.04
0.000 Water 95.47 98.85 0.97 4.53 1.15 5.68
Nonwater 100.00 99.98 0.00 0.02 0.02
0.100 Water 96.60 91.48 0.94 3.40 8.52 11.92
Nonwater 99.97 99.99 0.03 0.01 0.05
MNDWI 0.000 Water 92.57 96.66 0.95 7.43 3.34 10.77
Nonwater 99.99 99.97 0.01 0.03 0.04
0.100 Water 95.01 95.39 0.95 4.99 4.61 9.60
Nonwater 99.98 99.98 0.02 0.02 0.04
0.150 Water 97.53 92.66 0.95 2.47 7.34 9.81
Nonwater 99.97 99.99 0.03 0.01 0.04
MaxLike – Water 99.71 86.96 0.93 0.29 13.04 13.33
Nonwater 99.95 100.00 0.05 0.00 0.05
S. Africa AWEIsh 0.020 Water 83.23 98.86 0.90 16.77 1.14 17.91
Nonwater 99.97 99.50 0.03 0.50 0.53
0.045 Water 98.32 98.30 0.98 1.68 1.70 3.38
Nonwater 99.96 99.96 0.04 0.04 0.09
0.060 Water 98.43 97.51 0.98 1.57 2.49 4.06
Nonwater 99.94 99.96 0.06 0.04 0.10
MNDWI 0.300 Water 70.54 97.67 0.81 29.46 2.33 31.79
Nonwater 99.94 98.97 0.06 1.03 1.09
0.450 Water 89.07 96.37 0.92 10.93 3.63 14.56
Nonwater 99.91 99.70 0.09 0.30 0.39
0.600 Water 94.46 93.58 0.94 5.54 6.42 11.96
Nonwater 99.84 99.86 0.16 0.14 0.30
MaxLik – Water 46.74 97.12 0.62 53.26 2.88 56.14
Nonwater 99.93 97.21 0.07 2.79 2.87
N. Zealand AWEIsh −0.100 Water 98.74 99.87 0.96 1.26 0.13 1.39
Nonwater 99.29 93.65 0.71 6.35 7.06
0.000 Water 99.82 99.56 0.98 0.18 0.44 0.61
Nonwater 97.85 99.13 2.15 0.87 3.02
0.100 Water 99.90 99.45 0.98 0.10 0.55 0.65
Nonwater 97.33 99.51 2.67 0.49 3.17
MNDWI 0.000 Water 96.93 99.79 0.89 3.07 0.21 3.28
Nonwater 98.79 84.23 1.21 15.77 16.98
0.150 Water 97.39 99.44 0.90 2.61 0.56 3.16
Nonwater 96.91 86.72 3.09 13.28 16.37
0.200 Water 98.14 98.50 0.90 1.86 1.50 3.35
Nonwater 92.39 90.71 7.61 9.29 16.89
MaxLike – Water 99.84 99.18 0.97 0.16 0.82 0.98
Nonwater 96.05 99.22 3.95 0.78 4.73
a
At this test site, shadow is not a major source of classification noise but built-up surfaces are predominant land cover type. Therefore, the use of AWEInsh resulted in high accuracy of
water extraction (a combined use of both AWEIsh and AWEInsh did not improve accuracy).
 G.L. Feyisa et al. / Remote Sensing of Environment 140 (2014) 23–35 35

References Jimenez-Munoz, J. C., Sobrino, J. A., Mattar, C., & Franch, B. (2010). Atmospheric correction
of optical imagery from MODIS and reanalysis atmospheric products. Remote Sensing
Aguirre-Gutierrez, J., Seijmonsbergen, A.C., & Duivenvoorden, J. F. (2012). Optimizing land of Environment, 114, 2195–2210.
cover classification accuracy for change detection, a combined pixel-based and Kondo, H., Seo, N., Yasuda, T., Hasizume, M., Koido, Y., Ninomiya, N., et al. (2002).
object-based approach in a mountainous area in Mexico. Applied Geography, 34, Post-flood-infectious diseases in Mozambique. Prehospital and Disaster Medicine, 17,
29–37. 126–133.
Alderman, K., Turner, L. R., & Tong, S. L. (2012). Floods and human health: A systematic Lacaux, J. P., Tourre, Y. M., Vignolles, C., Ndione, J. A., & Lafaye, M. (2007). Classification of
review. Environment International, 47, 37–47. ponds from high-spatial resolution remote sensing: Application to Rift Valley Fever
Bond, N. R., Lake, P.S., & Arthington, A. H. (2008). The impacts of drought on freshwater epidemics in Senegal. Remote Sensing of Environment, 106, 66–74.
ecosystems: An Australian perspective. Hydrobiologia, 600, 3–16. Lake, P.S. (2003). Ecological effects of perturbation by drought in flowing waters.
Butcher, S. S., Charlson, R. J., Orians, G. H., & Wolfe, G. V. (Eds.). (1992). Global Freshwater Biology, 48, 1161–1172.
biogeochemical cycles. London: Academic Press Limited. Li, K. Z., Wu, S. H., Dai, E. F., & Xu, Z. C. (2012). Flood loss analysis and quantitative risk
Charoenpanyanet, A., & Chen, X. (2008). Satellite-based modeling of Anopheles mosquito assessment in China. Natural Hazards, 63, 737–760.
densities on heterogeneous land cover in Western Thailand. The International Archives Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2004). Remote sensing and image interpre-
of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 37, 159–164. tation (5th ed.): John Wiley & Sons, Inc.
Charron, D. F., Thomas, M. K., Waltner-Toews, D., Aramini, J. J., Edge, T., Kent, R. A., et al. Lira, J. (2006). Segmentation and morphology of open water bodies from multispectral
(2004). Vulnerability of waterborne diseases to climate change in Canada: A review. images. International Journal of Remote Sensing, 27, 4015–4038.
Journal of Toxicology and Environmental Health-Part a-Current Issues, 67, 1667–1677. McFeeters, S. K. (1996). The use of Normalized Difference Water Index (NDWI) in the
Congalton, R. G., & Green, K. (2009). Assessing the accuracy of remotely sensed data: delineation of open water features. International Journal of Remote Sensing, 17,
Principles and practices (2nd ed.)Boca Raton: Lewis Publishers. 1425–1432.
COWI (2010). DDOland2010: Danish National Digital Orthophoto, version 2010. http:// Morss, R. E., Wilhelmi, O. V., Downton, M. W., & Gruntfest, E. (2005). Flood risk, uncertain-
www.geodata-info.dk/Portal/ShowMetadata.aspx?id=d6a57343–c930–4d28–94ce- ty, and scientific information for decision making — Lessons from an interdisciplinary
85b5ca2847dc project. Bulletin of the American Meteorological Society, 86 (1593-+).
Dambach, P., Machault, V., Lacaux, J. -P., Vignolles, C., Sié, A., & Sauerborn, R. (2012). Uti- Mucher, C. A., Steinnocher, K. T., Kressler, F. P., & Heunks, C. (2000). Land cover char-
lization of combined remote sensing techniques to detect environmental variables acterization and change detection for environmental monitoring of pan-Europe.
influencing malaria vector densities in rural West Africa. International Journal of International Journal of Remote Sensing, 21, 1159–1181.
Health Geographics, 11. NASA (2012). Landsat 7 science data users handbook. Online.
Davranche, A., Lefebvre, G., & Poulin, B. (2010). Wetland monitoring using classification Novoa, S., Chust, G., Sagarminaga, Y., Revilla, M., Borja, A., & Franco, J. (2012). Water qual-
trees and SPOT-5 seasonal time series. Remote Sensing of Environment, 114, 552–562. ity assessment using satellite-derived chlorophyll-a within the European directives,
De Leeuw, J., Jia, H., Yang, L., Liu, X., Schmidt, K., & Skidmore, A. K. (2006). Comparing ac- in the southeastern Bay of Biscay. Marine Pollution Bulletin, 64, 739–750.
curacy assessments to infer superiority of image classification methods. International Ouma, Y. O., & Tateishi, R. (2006). A water index for rapid mapping of shoreline changes
Journal of Remote Sensing, 27, 223–232. of five East African Rift Valley lakes: An empirical analysis using Landsat TM and
Dewan, A.M., Islam, M. M., Kumamoto, T., & Nishigaki, M. (2007). Evaluating flood hazard ETM+ data. International Journal of Remote Sensing, 27, 3153–3181.
for land-use planning in Greater Dhaka of Bangladesh using remote sensing and GIS Poulin, B., Davranche, A., & Lefebvre, G. (2010). Ecological assessment of Phragmites
techniques. Water Resources Management, 21, 1601–1612. australis wetlands using multi-season SPOT-5 scenes. Remote Sensing of Environment,
Duan, Z., & Bastiaanssen, W. G. M. (2013). Estimating water volume variations in lakes 114, 1602–1609.
and reservoirs from four operational satellite altimetry databases and satellite imag- Prigent, C., Papa, F., Aires, F., Jimenez, C., Rossow, W. B., & Matthews, E. (2012). Changes in
ery data. Remote Sensing of Environment, 134, 403–416. land surface water dynamics since the 1990s and relation to population pressure.
Exelis Visual Information Solutions (2010). ENVI v. 4.8. Boulder, Colorado: EXELIS. Geophysical Research Letters, 39, L08403.
Frey, H., Huggel, C., Paul, F., & Haeberli, W. (2010). Automated detection of glacier lakes Proud, S. R., Fensholt, R., Rasmussen, L. Y., & Sandholt, I. (2011). Rapid response flood de-
based on remote sensing in view of assessing associated hazard potentials. Grazer tection using the MSG geostationary satellite. International Journal of Applied Earth
Schriften der Geographie und Raumforschung, 45, 261–272. Observation and Geoinformation, 13, 536–544.
Gardelle, J., Hiernaux, P., Kergoat, L., & Grippa, M. (2009). Less rain, more water in ponds: Richards, J. A. (1993). Remote sensing digital image analysis: An introduction (2nd ed.)
A remote sensing study of the dynamics of surface waters from 1950 to present Berlin: Springer-Verlag.
in pastoral Sahel (Gourma region, Mali). Hydrology and Earth System Sciences Rogers, A. S., & Kearney, M. S. (2004). Reducing signature variability in unmixing coastal
Discussions, 6, 5047–5083. marsh Thematic Mapper scenes using spectral indices. International Journal of Remote
Giardino, C., Bresciani, M., Villa, P., & Martinelli, A. (2010). Application of remote sensing Sensing, 25, 2317–2335.
in water resource management: The case study of Lake Trasimeno, Italy. Water Rozenstein, O., & Karnieli, A. (2011). Comparison of methods for land-use classification
Resources Management, 24, 3885–3899. incorporating remote sensing and GIS inputs. Applied Geography, 31, 533–544.
Guttler, F. N., Niculescu, S., & Gohin, F. (2013). Turbidity retrieval and monitoring of Ryu, J. H., Won, J. S., & Min, K. D. (2002). Waterline extraction from Landsat TM data in a
Danube Delta waters using multi-sensor optical remote sensing data: An integrated tidal flat — A case study in Gomso Bay, Korea. Remote Sensing of Environment, 83,
view from the delta plain lakes to the western-northwestern Black Sea coastal 442–456.
zone. Remote Sensing of Environment, 132, 86–101. Sethre, P. R., Rundquist, B. C., & Todhunter, P. E. (2005). Remote detection of prairie pot-
Haas, E. M., Bartholomé, E., & Combal, B. (2009). Time series analysis of optical remote hole ponds in the Devils Lake Basin, North Dakota. GIScience and Remote Sensing, 42,
sensing data for the mapping of temporary surface water bodies in sub-Saharan 277–296.
western Africa. Journal of Hydrology, 370, 52–63. Sheng, Y. W., Shah, C. A., & Smith, L. C. (2008). Automated image registration for hydro-
He, B., Oki, K., Wang, Y., Oki, T., Yamashiki, Y., Takara, K., et al. (2012). Analysis of stream logic change detection in the lake-rich Arctic. IEEE Geoscience and Remote Sensing
water quality and estimation of nutrient load with the aid of Quick Bird remote sens- Letters, 5, 414–418.
ing imagery. Hydrological Sciences JournalJournal Des Sciences Hydrologiques, 57, Sun, F., Sun, W., Chen, J., & Gong, P. (2012). Comparison and improvement of methods for
850–860. identifying waterbodies in remotely sensed imagery. International Journal of Remote
Hui, F., Xu, B., Huang, H., Yu, Q., & Gong, P. (2008). Modelling spatial–temporal change of Sensing, 33, 6854–6875.
Poyang Lake using multitemporal Landsat imagery. International Journal of Remote United States Geological Survey (USGS) (2012). Landsat data archive. Global Visualization
Sensing, 29, 5767–5784. Viewer (GLOVIS).
Jain, S. K., Saraf, A. K., Goswami, A., & Ahmad, T. (2006). Flood inundation mapping using Verpoorter, C., Kutser, T., & Tranvik, L. (2012). Automated mapping of water bodies using
NOAA AVHRR data. Water Resources Management, 20, 949–959. Landsat multispectral data. Limnology and Oceanography-Methods, 10, 1037–1050.
Jain, S. K., Singh, R. D., Jain, M. K., & Lohani, A. K. (2005). Delineation of flood-prone areas Xu, H. (2006). Modification of normalised difference water index (NDWI) to enhance
using remote sensing technique. Water Resources Management, 19, 337–347. open water features in remotely sensed imagery. International Journal of Remote
Ji, L., Zhang, L., & Wylie, B. (2009). Analysis of dynamic thresholds for the normalized dif- Sensing, 27, 3025–3033.
ference water index. Photogrammetric Engineering and Remote Sensing, 75, 1307–1317. Zhang, Y. L., Feng, L. Q., Li, J. S., Luo, L. C., Yin, Y., Liu, M. L., et al. (2010). Seasonal–spatial
Jiang, Z., Qi, J., Su, S., Zhang, Z., & Wu, J. (2012). Water body delineation using index compo- variation and remote sensing of phytoplankton absorption in Lake Taihu, a large eu-
sition and HIS transformation. International Journal of Remote Sensing, 33, 3402–3421. trophic and shallow lake in China. Journal of Plankton Research, 32, 1023–1037.