Content uploaded by Hanqiu Xu
Author content
All content in this area was uploaded by Hanqiu Xu
Content may be subject to copyright.
Abstract
The fast urban expansion has led to replacement of natural
vegetation-dominated land surfaces by various impervious
materials. This has a significant impact on the environment
due to modification of heat energy balance. Timely under-
standing of spatiotemporal information of impervious surface
has become more urgent as conventional methods for
estimating impervious surface are very limited. In response
to this need, this paper proposes a new index, normalized
difference impervious surface index (NDISI), for estimating
impervious surface. The application of the index to the
Landsat ETM⫹image of Fuzhou City and the ASTER image of
Xiamen City in China has shown that the new index can
efficiently enhance and extract impervious surfaces from
satellite imagery, and the normalized NDISI can represent the
real percentage of impervious surface. The index was further
used as an indicator to investigate the impact of impervious
surface on urban heat environment by examination of its
quantitative relationship with land surface temperature (LST),
vegetation, and water using multivariate statistical analysis.
The result reveals that impervious surface has a positive
exponential relationship with LST rather than a simple linear
one. This suggests that the areas with high percent impervi-
ous surface will accelerate LST rise and urban heat island
development.
Introduction
At present, urbanization is progressing worldwide with
accelerated speed. The world’s urban population grew
rapidly from 220 million to 2.8 billion over the twentieth
century. It was expected that more than 50 percent of the
world population would be living in urban areas in 2008,
while this figure is only 15 percent one hundred years ago
(http://www.unfpa.org/swp/2007/english/introduction.html).
The urbanization has brought about progress to the world,
but also brought with it the threat of environmental degrada-
tion and increased pressure on natural resources. Anthro-
pogenic land-cover totals to approximately 40 percent of the
Earth’s surface and the natural vegetation-dominated
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2010 557
College of Environment and Resources, Fuzhou University,
And Key Laboratory of Humid Subtropical Eco-geographical
Process at Fujian Normal University, Ministry of Education,
Fuzhou, Fujian 350108, China (fdy@public.fz.fj.cn,
hxu@fzu.edu.cn)
Photogrammetric Engineering & Remote Sensing
Vol. 76, No. 5, May 2010, pp. 557–565.
0099-1112/10/7605–0557/$3.00/0
© 2010 American Society for Photogrammetry
and Remote Sensing
Analysis of Impervious Surface and its
Impact on Urban Heat Environment
using the Normalized Difference
Impervious Surface Index (NDISI)
Hanqiu Xu
landscapes have been replaced by impervious surfaces
(Sterling and Ducharne, 2008). The direct impact of this is
the rise of urban temperature and thus the formation of the
urban heat island (UHI) (Gluch et al. 2006). Detecting
impervious surface in urban areas is critical to understand-
ing the impact, and hence has attracted increasing interest.
This requires a cost-effective method to fast obtain impervi-
ousness information over urban areas. Fortunately, the
advantages of the repetitive, synoptic, and real-time view of
the satellite observation over other methods offer a promis-
ing method to map large impervious surface areas. The
methods for estimating impervious surface area (ISA) with
remote sensing technology have been frequently discussed
in recent remote sensing literature, and can generally be
catalogued into four groups: (a) by manual or semi-automatic
method through visual interpretation or multispectral
classification (Chen et al., 2006; Gluch et al. 2006; Jennings
et al., 2004; Madhavan et al., 2001), (b) through integration
of classified imperviousness results with data derived from
other sources (Plunk et al., 1990; Sleavin et al., 2000), (c) by
spectral mixture analysis (SMA) (Ji and Jensen, 1999; Phinn
et al., 2002; Wu and Murray, 2003; Xian and Crane, 2005;
Yang et al., 2003), and (d) through the relationship of
impervious surface with other land-covers, e.g., vegetation
(Carlson and Arthur, 2000).
The first two methods are mainly applied to high-
resolution imagery, while the latter two are carried out in
moderate-resolution imagery mainly based on the vegetation-
impervious surface-soil (V-I-S) model of Ridd (1995) which
treats each pixel in urban imagery as a linear combination
of these three end members. Plunk et al. (1990) were able to
measure impervious surface in Fort Worth, Texas using a
supervised classification of Landsat TM data and noted that
total impervious surface area was often underestimated.
Ji and Jensen (1999) obtained ISA using the SMA and a
layered classification method and achieved an accuracy of 83
percent. Carlson and Arthur (2000) retrieved the fraction
impervious surface cover of Chester County based on the
inverse relationship between impervious surface and vegeta-
tion. Phinn et al. (2002) mapped the urban composition of
557-565_08-078.qxd 4/19/10 5:36 PM Page 557
558 May 2010
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Brisbane City and found that the SMA has advantage in
estimating urban ISA over the conventional classification
method. Wu and Murray (2003) used the SMA to estimate
the ISA of Columbia City from Landsat ETM⫹data with an
overall RMS error of 10.6 percent but had to mask out soil
and water features before processing impervious surface.
Yang et al. (2003) mapped a large area of impervious surface
using Landsat ETM⫹data with the assistance of high-
resolution imagery and a decision tree algorithm, and yielded
average errors ranging from 8.8 to 11.4 percent. Yang and Liu
(2005) estimated ISA of the Pensacola area by synergistic use
of Landsat ETM⫹imagery and DOQQ data through multivari-
ate statistical regression models, but pointed out that the
model was unable to discriminate between impervious
surface and soil. Chen et al. (2006) estimated the impervious-
ness of Nanjing City in China by a classification of Ikonos
data and a subsequent visual correction to improve accuracy.
Gluch et al. (2006) used a supervised classification approach
to identifying land-cover classes in the Salt Lake Valley
while investigating overall surface thermal pattern in the
area. The majority of the classification error is due to
spectral confusion between light impervious and light soil.
More recently, Powell et al. (2008) estimated the impervious
surfaces in the Snohomish area of Washington by using a
combined spectral- and spatial knowledge-based classifica-
tion method to overcome the problem raised by the spectral
confusion between impervious surface and soil. With the
technique the authors were able to reveal the urban and
residential development of the area over a 34 year time
period.
The above studies suggest that the present remote
sensing methods used in estimating ISA are rather compli-
cated with low degree of automation. Many methods need
pre-processing work to remove soil and water which
otherwise could be confused with impervious surface. Some
methods have to be assisted by high-resolution imagery. Up
to now, there is no automatic method to extract impervious
surface and the index-based method such as a NDXI (collec-
tively called for various normalized difference thematic
indices) has not yet been seen. Also, the detailed quantita-
tive relationship between ISA and temperature has not been
analyzed due to lack of a suitable index to represent ISA.
Therefore, this paper aims to demonstrate an approach for
facilitating estimation of impervious surface through a new
index, normalized difference impervious surface index
(NDISI). This was followed by the quantitative analysis of the
relationship of ISA with land surface temperature (LST) and
other urban biophysical components, such as vegetation and
water, to further examine the impact of impervious surface
on urban heat environment.
Methodology
Creating the NDISI
To create an index to enhance impervious surface without
the problems existing in the previous work, the following
two general rules have to be taken into account: (a) simple
calculation and independent use just like a NDXI, and (b) no
much image preprocessing work required.
Various normalized difference indices have been
developed and intensively used since the development of
the NDVI of Rouse et al. (1973). A rule for creating such
indices is trying to find out the strongest and weakest
reflectance bands, respectively, among multispectral bands
for the interested land-cover type to be enhanced. Using the
strongest reflectance band as numerator and the weakest one
as denominator, the normalized ratio of these two bands can
maximally enlarge the contrast between the interested cover
type and background noise. Therefore, the new index should
be developed based on this rule.
The urban area is covered dominantly by impervious
surfaces such as building roof, paved road, parking lot, etc.
Apart from the concrete-dominated impervious cover,
vegetation, water, soil, and sand may be locally developed.
Among them, the soil and sand have similar spectral
response features to impervious surface, and thus often
made noise mixed up with the enhanced impervious surface
information in previous studies (Gluch et al., 2006; Wu and
Murray, 2003; Yang and Liu, 2005). Therefore, the solution
of this problem is a main task of the new index.
Many present imperviousness mapping methods are
based on Ridd’s V-I-S model. However, ignoring water, one
of the important components of urban ecosystem, in the
model has brought problems to these methods (Xu, 2007).
As a result, water features had to be mask out before
processing impervious surface. The removal of water
information from imagery is not only time-consuming, but
also causes unexpected errors in the resultant imagery.
Therefore, the use of the new index should avoid this
water-related preprocessing work.
According to the aforementioned three basic considera-
tions for creating the new index, the spectral characteristics
of impervious surfaces were first examined to find out the
strongest and weakest reflectance bands, respectively, for the
impervious materials. Through inspection of spectral
characteristics of major urban impervious surface types, a
common feature can be found, i.e., the impervious materials
generally have high emittance in thermal band (TIR) and low
reflectance in near-infrared (NIR) band (Figure 1). This is
because the impervious materials such as concrete and
asphalt have a strong capability of emitting heat but are
impossible for vegetation to grow upon. Therefore, the ratio
of a thermal band to a NIR band should greatly enhance
impervious surface features. However, soil, sand, and water
also possess this spectral feature (Figure 1). Obviously, the
index simply using the ratio of a thermal band to a NIR band
cannot effectively enhance impervious surface features,
because the so-enhanced impervious surface information
could be mixed with soil, sand, and water noise just like the
problems occurring in previous work.
Further inspection can find that despite the spectral
similarity to impervious materials in thermal and NIR bands,
soil, sand, and water generally have a higher reflectance
Figure 1. Signatures of main land-cover classes of Fuzhou
City from a Landsat ETM⫹image (29 May 2003). The DN
values are derived from reflectance of multispectral bands
or emittance of thermal band (see text for details).
557-565_08-078.qxd 4/19/10 5:36 PM Page 558
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2010 559
than impervious materials in visible (VIS) bands, and soil
and sand have stronger reflectance than impervious
materials in middle infrared (MIR) band (Figure 1). Accord-
ingly, the VIS and MIR bands, together with the NIR band, can
form a weak-reflectance group of impervious surface, and
the ratio of thermal band to this group should be effective to
enhance impervious surface and suppress soil, sand, and
water noise. Therefore, the NDISI can be constructed as:
(1)
where NIR and MIR1 are a near-infrared band and a mid-
infrared band, respectively, such as ETM⫹band 4 and band
5; TIR is a thermal band such as ETM⫹band 6; VIS
1
is one
of visible bands such as ETM⫹bands 1, 2 or 3. Dividing the
sum by 3 is to avoid too small a value of the index. Conse-
quently, the index can have a value ranging from ⫺1 to 1.
This index is designed to (a) maximize the radiation
that is emitted from impervious surface in the form of heat
by using thermal wavelengths, (b) minimize the low
reflectance of NIR, MIR, and VIS bands by impervious surface,
and (c) take advantage of the high reflectance of MIR by sand
and soil. Obviously, the impervious surface features would
have positive values and thus be enhanced. While, the
unwanted non-impervious covers would be in negative
values and hence suppressed (Table 1).
However, the experiments showed that water could have
lower reflectance than impervious surface in visible wave-
lengths typically when water was clear. In this case, the
impervious surface features enhanced using Equation 1 can
be mixed with water noise. This is always a major problem
in previous work and seems impossible to be solved by using
original multispectral bands. Masking out water before using
Equation 1 appears to be the only method to solve the
problem. However, the seeking of the solution has found that
using a water index (WI)-derived band instead of the VIS band
in Equation 1 can significantly suppress the water noise, as it
can enlarge the contrast between water and impervious
surface. Therefore, Equation 1 can be re-written as:
(2)
where WI can be represented by the NDWI of McFeeters
(1996) or the MNDWI of Xu (2006):
(3)
(4)
where Green is a green band such as ETM⫹band 2.
NDWI ⫽ (Green ⫺NIR)
(Green ⫹NIR)
MNDWI ⫽ (Green ⫺MIR1)
(Green ⫹MIR1)
NDISI ⫽TIR ⫺[(WI ⫹NIR ⫹MIR1)/3]
TIR ⫹[(WI ⫹NIR ⫹MIR1)/3]
NDISI ⫽ TIR ⫺[(VIS1⫹NIR ⫹MIR1)/3]
TIR ⫹[(VIS1⫹NIR ⫹MIR1)/3]
Compared with accompanied multispectral bands, the
thermal band generally has a coarser resolution, i.e., 60 m
versus 30 m in ETM⫹imagery. Nevertheless, the computa-
tion of the thermal band with finer-resolution multispectral
bands using Equation 1 or 2 can reduce the pixel size of the
thermal band, due to being fused with 30 m resolution
multispectral bands. As a result, the produced NDISI image
can have a resolution similar to multispectral bands rather
than to the thermal band.
Image Processing
The new index has been tested in different cities. Among
them, the results of Fuzhou and Xiamen cities of Fujian
province in China are provided in this paper but with
Fuzhou as a main case to fully demonstrate the new
approach. Fuzhou is the capital city of Fujian province,
which is located in the coastal area of southeastern China
(Figure 2). A Landsat ETM⫹image of Fuzhou City acquired
on 29 May 2003 was used for the test. Even if a raw image
can be directly used for computing the NDISI, the image was
radiometrically corrected before the calculation to avoid a
dataset-specific result. The radiometric correction employed
the algorithm of Chander and Markham (2003) with the
addition of an atmospheric correction model suggested by
Chavez (1996). This converted the digital number (DN) of the
raw image to at-satellite reflectance. Thermal band 6 was
converted to at-satellite temperature, and then emissively
corrected to LST using the methods described by Weng et al.
(2004) and Nichol (2005). As the at-satellite reflectance
values range from 0 to 1, the values were multiplied by 400
and truncated to produce 8-bit DN values as suggested by the
USGS (2006). The LST was rescaled within 0 to 255 to
produce 8-bit data.
However, if a raw image is to be used, the only
preprocessing work should be done before computing the
NDISI is to rescale the DNs of thermal band between 0 and
255 to produce 8-bit data. In the case of using Equation 2,
the MNDWI or NDWI values also have to be rescaled within
0 to 255 for the same reason.
Application of the NDISI to Fuzhou’s ETM⫹image has
effectively enhanced the impervious surfaces of the city
(Figure 3a). Three visible-light bands (blue, green, and red)
were used respectively to generate different band combina-
tion modes to test Equation 1. A default value of 0 and
several manually adjusted threshold values were used to
extract impervious surface features from the NDISI images
generated with different band combination modes (Table 2).
The pixels with values greater than the threshold are
impervious surface and assigned a value of 1, while the
pixels with values equal or less than the threshold are non-
impervious surface and assigned a value of 0. Thus, the
resultant image is a binary image, only showing the
extracted impervious surfaces (Figure 3d). The extracted
impervious surface images were quantitatively assessed for
T
ABLE
1. STATISTICS OF FUZHOU CITY’S MAIN LAND-COVER CLASSES IN THE NDISI IMAGE
PRODUCED USING ETM⫹BANDS 2, 4, 5, AND 6
Impervious
surface Sand Soil Vegetation Water
Minimum 0.198 ⫺0.125 ⫺0.110 ⫺0.377 ⫺0.827
Maximum 0.474 0.053 ⫺0.010 ⫺0.029 ⫺0.384
Mean 0.365 ⫺0.031 ⫺0.067 ⫺0.214 ⫺0.573
Standard division 0.064 0.029 0.022 0.062 0.079
Mean difference with / 0.396 0.432 0.579 0.938
impervious surface
557-565_08-078.qxd 4/19/10 5:36 PM Page 559
560 May 2010
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
Figure 2. Map showing the location of Fuzhou City and the test area.
(d) (e)
(a)
(f)
(b) (c)
Figure 3. Impervious surface-enhanced images of Fuzhou City using (a) NDISI, (b) SMA, and (c) Carlson
methods; and extracted impervious surface images of Fuzhou City thresholded at (d) 0.1048, (e) 0, and
Xiamen City, thresholded at (f) 0.121.
557-565_08-078.qxd 4/19/10 5:36 PM Page 560
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2010 561
the accuracy as suggested by Congalton et al. (1983) and
Congalton (1991). A high-resolution SPOT5 image dated
14 December 2003 was employed as reference data for the
accuracy assessment. A simple random sampling scheme
was employed since only two categories (ISA versus non-ISA)
were assessed. A total of 663 pixels were sampled at random
coordinates mainly within urban-suburb areas (76,176
pixels) of the image. To investigate the relationship between
the NDISI and the real percentage of impervious surface
coverage within a pixel, a subpixel accuracy assessment was
further conducted using the method suggested by Wu and
Murray (2003). A total of 100 samples in the image were
randomly chosen for the assessment, as this sampling
number is thought to be a compromise between statistical
rigor and practical limitations (Wu and Murray, 2003). A 1
m resolution Ikonos Pan image of the study area, dated
22 June 2003, was used as reference data for the accuracy
assessment. The difference in acquisition dates between the
Ikonos and the test ETM⫹images is less than one month. For
every sampled NDISI pixel, the corresponding Ikonos image
was digitized and the percent impervious surface area
(%ISA) was calculated. In this way, each sampled NDISI
pixel can be related to real percentage of impervious surface
coverage within that pixel.
Multivariate Statistical Analysis
The analysis of the impact of impervious surface on urban
heat environment was carried out by investigating its
quantitative relationship with other urban biophysical
components, i.e., LST, open water, and vegetation. The
quantitative relationship between vegetation and tempera-
ture has been frequently discussed based on the related
thematic index such as NDVI (Owen et al., 1998; Weng et al.,
2004). However, as a key component of urban ecosystem,
the relationship of impervious surface with temperature and
other urban components has not been quantitatively
analyzed in detail due to lack of a representative index. The
creation of the new index would allow this analysis to be
implemented by using multivariate regression analysis with
a huge number of samples throughout whole image. The
sampling was carried out systematically with a 5 ⫻5 grid
within whole image, and a total of 25,445 pixels out of
328,329 pixels of the test image were sampled. The system-
atic sampling throughout whole image with a great number
of samples would allow examine the relationship in more
detailed and objectively, as this can avoid the shortcomings
of simple random sampling regarding the coverage of a
whole image with a few hundreds of samples (Jensen, 2005).
Before the multivariate analysis, values of all thematic-
oriented indices, NDISI for impervious surface, MNDWI for
water and NDVI for vegetation, were normalized between
0 and 1 and then multiplied by 100. This permits analysis
of each index as a continuous variable from 0 to 100 percent
(Goetz et al., 2003) and allows the use of the NDISI as an
indicator of %ISA.
Results and Discussion
Figure 3a shows that the impervious surfaces of Fuzhou City
have been greatly enhanced. Urban area is characterized by
a light gray to white tone, indicating high NDISI values and
thus high coverage of impervious surface. While, dark gray
and black shades mainly occur in suburban and rural areas,
suggesting low coverage of impervious surface. The accuracy
assessment of the extracted impervious surfaces yielded a
highest overall accuracy of 90.7 percent along with a kappa
coefficient of 0.812, when using bands 2, 4, 5, and 6 to
compute the NDISI and a threshold value of 0.1048 to extract
impervious surface features (Figure 3d; Table 2). In the case
of using a default threshold value of 0, the accuracy can still
reach to 85.3 percent when using a band combination of
1, 4, 5, and 6 (Figure 3e; Table 2). The outputs at manually
T
ABLE
2. ACCURACY ASSESSMENT OF EXTRACTED IMPERVIOUS SURFACES USING THE NDISI
Bands used: 1, 4, 5, & 6 Bands used: 2, 4, 5, & 6 Bands used: 3, 4, 5, & 6
Threshold value: 0.039 Threshold value: 0.1048 Threshold value: 0.1028
ISA Non- Line User’s ISA Non- Line User’s ISA Non- Line User’s
ISA total accuracy ISA total accuracy ISA total accuracy
ISA 325 38 363 89.5% 325 35 360 90.3% 324 42 366 88.5%
Non-ISA 27 273 300 91.0% 27 276 303 91.1% 28 269 297 90.6%
Column 352 311 663 352 311 663 352 311 663
total
Producer’s 92.3% 87.8% 92.3% 88.8% 92.1% 86.5%
accuracy
Overall 90.2% 90.7% 89.4%
accuracy
Kappa 0.803 0.812 0.788
Default threshold value: 0 Default threshold value: 0 Default threshold value: 0
ISA 347 92 439 79.0% 348 109 457 76.2% 344 117 461 74.6%
Non-ISA 5 219 224 97.8% 4 202 206 98.1% 8 194 202 96.0%
Column 352 311 663 352 311 663 352 311 663
total
Producer’s 98.6% 70.4% 98.9% 65.0% 97.7% 62.4%
accuracy
Overall 85.3% 83.0% 81.2%
accuracy
Kappa 0.701 0.651 0.616
557-565_08-078.qxd 4/19/10 5:36 PM Page 561
562 May 2010
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
T
ABLE
3. ACCURACY ASSESSMENT OF EXTRACTED IMPERVIOUS SURFACES USING THE SMA AND CARLSON METHODS
SMA method Carlson method
ISA Non-ISA Line total User’s accuracy ISA Non-ISA Line total User’s accuracy
ISA 339 119 458 74.02% 338 157 495 68.28%
Non-ISA 11 181 192 94.27% 12 143 155 92.26%
Column total 350 300 650 350 300 650
Producer’s accuracy 96.86% 60.33% 96.57% 47.67%
Overall accuracy 80.1% 74.0%
Kappa 0.587 0.458
*
Water areas were not sampled for assessment as the water features were masked out beforehand for these two methods.
Figure 4. Scatterplot of NDISI versus %ISA.
selected threshold values are better when compared with
those thresholded at a default value, as the latter overesti-
mates ISA, and therefore has user’s accuracy of less than
80 percent (Table 2). The test of the new index in Xiamen
City with an ASTER image acquired on 06 September 2002,
using Equation 2 and MNDWI, yielded an overall accuracy of
90.8 percent and a Kappa of 0.815 when thresholded at
0.121 (Figure 3f), or 83.54 percent and 0.677 when thresh-
olded at 0. The similarity in the results of accuracy assess-
ment between Fuzhou and Xiamen cities using two different
sensor data suggests that the two independent processes
reinforce each other and the results merit recognition.
The subpixel accuracy assessment shows that the
normalized NDISI and the real percentage of impervious
surface are highly correlated (R
2
⫽0.9054) with an overall
estimation RMS error of 0.0787 for all samples (Figure 4).
The relationship between NDISI-estimated ISA and %ISA
closely matches the 1:1 prediction line, and a consistent
linear relationship between the two measurements can be
examined in Figure 4. The subpixel accuracy assessment
indicates that the normalized NDISI values are close
enough to %ISA, and therefore can be used as continuous
variable to represent the real percentage of impervious
surface.
The impervious surfaces of the Fuzhou image were also
enhanced using the SMA and Carlson methods for the
comparison with the new approach, because they are most
widely used techniques (Figures 3b and 3c). Two major
problems can be found in their resultant images, i.e., the
enhanced impervious surface features are mixed with
sand/soil noise (see a sandy bar marked with “A” in the
figures) and slope shadow noise (marked with “B”). Conse-
quently, the ISA were over-estimated and non-ISA thus
underestimated (Table 3). Methodologically, the SMA method
extracts impervious surfaces simply by the addition of the
low-albedo and high-albedo fraction images. However,
besides bright impervious surface, sand and dry soil also
have high albedo, while wet soil, shadow, water as well as
dark impervious surface can all have low albedo characteris-
tics. Therefore, the enhanced impervious surface information
was unavoidably mixed with sand, soil, and shadow noise
and the previous study had to remove the sand, soil, and
water features before making low- and high-albedo images.
The Carlson method is mainly based on the relationship
between impervious surface and vegetation fraction cover,
and thus treats non-vegetation areas except water as imper-
vious surfaces (Carlson and Arthur, 2000). As a result, the
sand and soil are significantly mixed with impervious
surface. By contrast with these two methods, the NDISI can
remove shadow noise and greatly reduce the sand and soil
noise (Figure 3a).
In the study of the thermal properties of various
land-covers in the Salt Lake Valley, Gluch et al. (2006)
noted that soil thermal energy response was quite similar
to that of impervious surface. This caused spectral
confusion between impervious and soil in their classifica-
tion results (high omission errors in both impervious and
soil). Powell et al. (2008) also recognized the spectral
overlap between impervious surface and soil and had to
identify impervious surface with a combined spectral-
and spatial knowledge-based classification method.
Apparently, the similarity in both thermal energy and
spectral responses between imperviousness and soil can
also account for the relative low accuracy of the SMA and
Carlson methods. To overcome the problem raised by the
similarity, the creation of the new index takes full
consideration on the signatures of impervious surface and
soil, and uses multiple bands rather than only two bands
to construct the index. The much higher reflectance of
soil than impervious surface in the MIR wavelength range
(such as ETM⫹5) makes them separable each other
(Figure 1). In addition, the ratio-based index is helpful
for reducing topographic difference and thus the slope
shades. The introducing of a visible band or a WI-derived
band contributes to the differentiation between water and
impervious surface. This is another advantage of the NDISI
over the SMA and Carlson methods, as the NDISI method
needs not to mask out water feature before enhancing
imperviousness features.
However, because the construction of the NDISI needs
thermal data, the new index can only be used in the multi-
spectral imagery having the thermal band(s) such as Landsat
TM/ETM⫹and ASTER, and is not suitable for the data without
thermal band(s) such as QuickBird and Ikonos.
557-565_08-078.qxd 4/19/10 5:36 PM Page 562
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2010 563
T
ABLE
4. IMPERVIOUS SURFACE LEVELS WITH CORRESPONDING
LST COMPUTED FROM REGRESSION MODEL
ISA LST ⫽15.052e
0.0064NDISI
LST (°C) Increased LST (°C)
0 15.05 /
10 16.05 0.99
20 17.11 1.06
30 18.24 1.13
40 19.44 1.21
50 20.73 1.29
60 22.10 1.37
70 23.56 1.46
80 25.12 1.56
90 26.78 1.66
100 28.55 1.77
(a) (b)
Figure 5. Relationship of ISA to LST and NDVI: (a) and (b) 2D scatterplot, and
(c) 3D scatterplot.
Quantitative Relationships between Impervious Surface and
Other Urban Components
Relationships of Impervious Surface with LST and Vegetation
This study examined the relationship between impervious
surface and LST through different regression models such as
linear, exponential, polynomial, and power using LST as
dependent variable and NDISI as independent variable. The
results of all models show up a strong positive correlation
between impervious surface and LST. This proves that the
development of impervious surface has contributed to
temperature rise. Of the four regression models, the expo-
nential model achieves the highest correlation value
(Figure 5a, significant at the 0.01 level). This suggests that
the relationship between ISA and LST is not simple linear but
rather exponential, and the temperature rise in the high-
density impervious surface areas is faster than in low-
density impervious surface areas. Table 4 shows that in
high-density impervious surface area (ISA ⬎70), each
10 increment in ISA value could raise LST by more than
1.5°C, while this figure is around 1°C in low-density
impervious surface area (ISA ⬍30).
The relationship between impervious surface and
vegetation was also explored using the four regression
models mentioned above. A result of strong negative linear
relationship has been revealed (Figure 5b; significant at 0.01
level), which indicates that the increase in impervious
surface area is responsible for substantial decrease in
vegetation covers. A declined and sharp upper edge of the
scatterplot shown in Figure 5b may imply a more strong
negative correlation between impervious surface and
vegetation.
Finally, the relationship among impervious surface,
vegetation, and temperature can be vividly shown in
graphical form with a three-dimensional spectral feature
space (Figure 5c). The top of the scatterplot is composed of
impervious surface-dominated pixels with high LST and low
vegetation, while the root represents low LST, low ISA, but
highly-vegetated pixels. Because the impervious surface
dominates the urban area, the plot is of a generally
cuneiform shape, big in head but thin in foot.
Relationship of Impervious Surface with LST, Vegetation, and Water
Impervious surface, water, and vegetation are three most
important biophysical components of urban ecosystem. The
interaction among the three components can have a signifi-
cant impact on urban temperature. Therefore, the quantita-
tive relationship of the three components with LST was
analyzed by stepwise regression, with LST as dependent
variable and NDISI, NDVI, and MNDWI as independent
variables. This produced the following model (significant at
0.005 level):
557-565_08-078.qxd 4/19/10 5:36 PM Page 563
564 May 2010
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
T
ABLE
5. LST CHANGE AFTER ADJUSTING THE VALUE OF NDISI
TO NDVI OR TO MNDWI
Original LST after increment LST after increment
NDISI LST (°C) of NDVI value (°C) of MNDWI value (°C)
10 14.89 13.29 13.12
20 17.06 15.46 15.29
30 17.92 16.32 16.15
40 19.25 17.65 17.48
50 20.72 19.12 18.95
60 22.60 21.00 20.83
70 23.56 21.96 21.79
80 25.27 23.67 23.50
90 26.24 24.64 24.47
100 27.96 26.36 26.19
.
(5)
The three independent variables are all retained during the
stepwise regression procedure, which suggests that impervi-
ous surface, water, and vegetation are all important factors
in modulating urban temperature. Nevertheless, they work
differently as indicated by the regression model. Impervious
surface contributes positively to LST rise, while water and
vegetation work inversely. Impervious surface has low NDVI
and MNDWI but high LST values, and thus is lack of vegeta-
tion and water coverage and in favor of sensible heat
exchange, which forms one of the bases of the UHI effect. On
the other hand, the greater water and vegetation cover
suggests comparatively higher rates of evapotranspiration
and favoring of latent over sensible heat exchange between
surface and atmosphere as compared with impervious
surface areas (Wilson et al., 2003). Based on very strong
fitness of the regression model (Equation 5; R
2
⫽0.9396), we
may predict LST change at each 10 decrement of NIDSI value
by adjusting the proportion of the other two components.
Table 5 shows that each 10 decrement in NDISI value with
the increment in NDVI or MNDWI by the same value would
lower LST averagely by 1.6 °C or 1.8 °C, respectively. This
proves that the increase in vegetation and water coverage in
impervious surface area can considerably mitigate the UHI
effect.
Conclusions
One of the key challenges in urban remote sensing is to
properly discriminate between impervious surface and other
land-cover classes in a heterogeneous context. Therefore,
complicated methods have to be used in previous studies for
mapping impervious surfaces. To date, no index has been
created to enhance impervious surface feature, as a simple
two-band ratio index is not capable of doing so. Recognizing
this, the creation of the NDISI is based on the combination of
several spectral bands. This makes the new index be able to
enhance impervious surface features while suppressing
background noise such as sand, soil, and water, which
otherwise could be mixed with the extracted impervious
surface information and thus have to be masked out before-
hand. The NDISI images can delineate the distribution of
impervious surface features and represent real percentage of
impervious surface. It can be used straightforward, and
generally no preprocessing work is needed. Therefore, it will
greatly improve the efficiency of mapping large areas of
impervious surface and is a handy and fast tool for the
extraction of impervious surface.
⫹18.762, (R2⫽0.9396)
LST ⫽0.122NDISI ⫺0.038NDVI ⫺0.055MNDWI
The creation of the NDISI adds an important biophysical
input to the urban ecosystem model. This allows the quantita-
tive analysis of the relationship of impervious surface with
temperature as well as other components of urban ecosystem.
The regression model of Fuzhou City demonstrates that
impervious surface accounts for most of the variation in
urban heat environment dynamics. It contributes significantly
to the urban LST rise, while vegetation and water work
inversely. The reveal of the strong positive exponential
relationship between impervious surface and LST suggests that
the increase in built-up land percentage would exponentially
accelerate LST rise. Accordingly, increasing vegetation and
water covers in high-density impervious surface areas should
become a critical issue for mitigating the UHI effect.
Acknowledgments
This study is supported by the National Natural Science
Foundation of China (No. 40371107) and Fujian Provincial
Department of Science and Technology, China (No.
2005YZ1011, No. 2007J0132, and No. JK2009004).
References
Carlson, T.N., and S.T. Arthur, 2000. The impact of land use - land
cover changes due to urbanization on surface microclimate and
hydrology: a satellite perspective, Global and Planetary Change,
25(1):49–65.
Chander, G., and B. Markham, 2003. Revised Landsat-5 TM
radiometric calibration procedures and postcalibration dynamic
ranges, IEEE Transactions on Geoscience and Remote Sensing,
41(11):2674–2677.
Chavez, P.S., Jr., 1996. Image-based atmospheric corrections -
Revisited and revised, Photogrammetric Engineering & Remote
Sensing, 62(9):1025–1036.
Chen, S., X. Zhang, and L. Peng, 2006. Impervious surface coverage
in urban land use based on high resolution satellite images,
Resources Science, 28(2):42–46.
Congalton, R.G., R.G. Oderwald, and R.A. Mead, 1983. Assessing
Landsat classification accuracy using discrete multivariate
statistical techniques, Photogrammetric Engineering & Remote
Sensing, 49(12):1671–1678.
Congalton, R.G., 1991. A review of assessing the accuracy of
classifications of remotely sensed data, Remote Sensing of
Environment, 37(1):35–46.
Gluch, R., A.D. Quattrochi, and J.C. Luvall, 2006. A multi-scale
approach to urban thermal analysis, Remote Sensing of
Environment, 104(1):123–132.
Goetz, S.J., R.K. Wright, A.J. Smith, E. Zinecker, and E. Schaub,
2003. IKONOS imagery for resource management: Tree cover,
impervious surfaces, and riparian buffer analyses in the
mid-Atlantic region, Remote Sensing of Environment,
88(1–2):195–208.
Jennings, D.B., S.T. Jarnagin, and C.W. Ebert, 2004. A modeling
approach for estimating watershed impervious surface area from
National Land Cover Data 92, Photogrammetric Engineering
& Remote Sensing, 70(11):1295–1307.
Jensen, J.R., 2005. Introductory Digital Image Processing: A Remote
Sensing Perspective, Upper Saddle River, New Jersey: Prentice
Hall, 526 p.
Ji, M., and J.R. Jensen, 1999. Effectiveness of subpixel analysis in
detecting and quantifying urban imperviousness from Landsat
Thematic Mapper imagery, Geocarto International, 14(4):33–41.
Madhavan, B.B., S. Kubo, and N. Kurisaki, 2001. Appraising the
anatomy and spatial growth of the Bangkok metropolitan
area using a vegetation-impervious-soil model through
remote sensing, International Journal of Remote Sensing,
22(5):789–806.
McFeeters, S.K., 1996. The use of normalized difference water index
(NDWI) in the delineation of open water features, International
Journal of Remote Sensing, 17(7):1425–1432.
557-565_08-078.qxd 4/19/10 5:36 PM Page 564
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2010 565
Nichol, J., 2005. Remote sensing of urban heat islands by day and
night, Photogrammetric Engineering & Remote Sensing,
71(5):613–621.
Owen T.W., T.N. Carlson, and R.R. Gillies, 1998. Remotely-sensed
surface parameters governing urban climate change, Interna-
tional Journal of Remote Sensing, 19:1663–1681.
Phinn, S., M. Stanford, P. Scarth, A.T. Murray, and P.T. Shyy, 2002.
Monitoring the composition of urban environments based on
the vegetation-impervious surface-soil (VIS) model by subpixel
analysis techniques, International Journal of Remote Sensing,
23(20):4131–4153.
Powell, S.L., W.B. Cohen, Z. Yang, J.D. Pierce, and M. Alberti,
2008. Quantification of impervious surface in the Snohomish
Water Resources Inventory Area of western Washington
from 1972–2006, Remote Sensing of Environment,
112(14):1895–1908.
Plunk, D.E., Jr., K. Morgan, and L. Newland, 1990. Mapping
impervious cover using Landsat TM data, Journal of Soil and
Water Conservation, 45(5):589–591.
Ridd, M.K., 1995. Exploring a V-I-S (Vegetation-Impervious Surface-
Soil) model for urban ecosystem analysis through remote
sensing: Comparative anatomy for cities, International Journal
of Remote Sensing, 16(12):2165–2185.
Rouse, J.W., R.H. Haas, J.A. Schell, and D.W. Deering, 1973.
Monitoring vegetation systems in the Great Plains with ERTS,
Proceedings of the Third ERTS Symposium, Washington D.C.,
NASA SP-351, I: 309–317.
Sleavin, W.J., D.L. Civco, S. Prisloe, and L. Giannotti, 2000.
Measuring impervious surfaces for non-point source pollution
modeling, Proceedings of the ASPRS Annual Conference, 22–26
May, Washington, D.C., unpaginated CD-ROM.
Sterling, S., and A. Ducharne, 2008. Comprehensive data set of
global land cover change for land surface model applications,
Global Biogeochemical Cycles, 22(3),GB3017.
USGS, 2006. Multi-resolution land characteristics 2001 (MRLC2001)
image processing procedure, URL: http://landcover.usgs.gov/
pdf/image_preprocessing.pdf (last date accessed: 01 February
2010).
Weng, Q., D. Lu, and J. Schubring, 2004. Estimation of land surface
temperature – Vegetation abundance relationship for urban heat
island studies, Remote Sensing of Environment, 89(4):467–483.
Wilson, J.S., M. Clay, E. Martin, D. Stuckey, and K. Vedder-Risch,
2003. Evaluating environmental influences of zoning in urban
ecosystems with remote sensing, Remote Sensing of Environ-
ment, 86(3):303–321.
Wu, C., and A.T. Murray, 2003. Estimating impervious surface
distribution by spectral mixture analysis, Remote Sensing of
Environment, 84(4):493–505.
Xian, G., and M. Crane, 2005. Assessments of urban growth in the
Tampa Bay watershed using remote sensing data, Remote
Sensing of Environment, 97(2):203–215.
Xu, H.Q., 2006. Modification of normalised difference water index
(NDWI) to enhance open water features in remotely sensed imagery,
International Journal of Remote Sensing, 27(14): 3025–3033.
Xu, H.Q., 2007. Extraction of urban built-up land features from
Landsat imagery using a thematic-oriented index combination
technique, Photogrammetric Engineering & Remote Sensing,
73(12):1381–1391.
Yang, L., C. Huang, C.G. Homer, B.K. Wylie, and M.J. Coan, 2003.
An approach for mapping large-area impervious surfaces:
synergistic use of Landsat-7 ETM⫹and high spatial resolution
imagery, Canadian Journal of Remote Sensing, 29(2):230–240.
Yang, X., and Z. Liu, 2005. Use of satellite-derived landscape
imperviousness index to characterize urban spatial growth,
Computers, Environment and Urban Systems, 29(4):524–540.
(Received 26 September 2008; accepted 25 August 2009; final
version 03 September 2009)
557-565_08-078.qxd 4/19/10 5:36 PM Page 565
