Remote Sensing of Environment 102 (2006) 146 – 160

www.elsevier.com/locate/rse

Use of impervious surface in urban land-use classification

Dengsheng Lu a,⁎, Qihao Weng b,1
a
Center for the Study of Institutions, Population, and Environmental Change, Indiana University, Bloomington, IN 47408, USA
b
Department of Geography, Geology, and Anthropology, Indiana State University, Terre Haute, IN 47809, USA

Received 20 June 2005; received in revised form 8 February 2006; accepted 9 February 2006

Abstract

Impervious surface has been recognized as a key indicator in assessing urban environments. However, accurate impervious surface extraction
is still a challenge. Effectiveness of impervious surface in urban land-use classification has not been well addressed. This paper explored extraction
of impervious surface information from Landsat Enhanced Thematic Mapper data based on the integration of fraction images from linear spectral
mixture analysis and land surface temperature. A new approach for urban land-use classification, based on the combined use of impervious surface
and population density, was developed. Five urban land-use classes (i.e., low-, medium-, high-, and very-high-intensity residential areas, and
commercial/industrial/transportation uses) were developed in the city of Indianapolis, Indiana, USA. Results showed that the integration of
fraction images and surface temperature provided substantially improved impervious surface image. Accuracy assessment indicated that the root-
mean-square error and system error yielded 9.22% and 5.68%, respectively, for the impervious surface image. The overall classification accuracy
of 83.78% for five urban land-use classes was obtained.
© 2006 Elsevier Inc. All rights reserved.

Keywords: Impervious surface; Urban land-use classification; Linear spectral mixture analysis; Land surface temperature; Population density

1. Introduction network, contextual, object-oriented, and knowledge-based

classification approaches (Stuckens et al., 2000; Thomas et
Improving urban land-use/cover classification accuracy has al., 2003; Zhang & Foody, 2001; Zhang & Wang, 2003).
been an important issue in remote-sensing literature. Different However, urban land-use/cover classification is still a challenge
approaches have been applied, which include incorporation of with medium or coarse spatial resolution remotely sensed data
geographic data (Harris & Ventura, 1995), census data (Mesev, due to the large number of mixed pixels and the spectral
1998), texture features (Lu & Weng, 2005; Myint, 2001; Shaban confusions among different land-use/cover types.
& Dikshit, 2001), and structure or contextual information (Gong Mixed pixels have been recognized as a problem affecting
& Howarth, 1990; Stuckens et al., 2000) into remote sensing the effective use of remotely sensed data in urban land-use/
spectral data. Moreover, expert systems (Hung & Ridd, 2002; cover classification (Cracknell, 1998; Fisher, 1997). Traditional
Stefanov et al., 2001), fuzzy classification (Zhang & Foody, per-pixel classifiers, such as maximum-likelihood classifier,
2001), and merged multisensor data such as those between radar cannot effectively handle complex urban landscapes and the
and Thematic Mapper (TM) data (Haack et al., 2002), and mixed-pixel problem. One of the recent major advances in
between SPOT and TM data (Gluch, 2002) have been applied. urban land-use/cover analysis is Ridd's (1995) vegetation–
Much attention has recently been shifted to the development of impervious surface–soil (V–I–S) model. This model assumes
more advanced classification algorithms, including neural that the spectral signature of land cover in urban environments
is a linear combination of three components, namely vegetation,
impervious surface, and soil. The V–I–S model has demon-
⁎ Corresponding author. Tel.: +1 812 856 3102; fax: +1 812 855 2634.
strated to be an effective approach to cope with the mixed-pixel
E-mail addresses: [email protected] (D. Lu), [email protected]
problem and provides a guideline for decomposing low-
(Q. Weng). resolution images of urban landscapes and a link for these
1
Tel.: +1 812 237 2255; fax: +1 812 237 8029. components to spectral signatures.
0034-4257/$ - see front matter © 2006 Elsevier Inc. All rights reserved.
doi:10.1016/j.rse.2006.02.010
 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160 147

The conceptual V–I–S model may be implemented by using extraction (Dare, 2005). Because of the high inverse
the technique of linear spectral mixture analysis (LSMA), which correlation between vegetation cover and impervious surface
decomposes the spectral reflectance of a pixel into different areas in urban landscapes, one potential approach for
proportions. LSMA is regarded as a physically based image- impervious surface extraction is through the use of vegetation
processing tool that supports repeatable and accurate extraction cover, such as tasseled cap greenness (Bauer et al., 2004) and
of quantitative subpixel information (Smith et al., 1990). fractional vegetation cover from the normalized difference
Because of its effectiveness in handling the spectral mixture vegetation index (Gillies et al., 2003). However, different
problem, LSMA has been widely used in many fields, such as seasons may result in large variations in impervious surface
mapping of land-cover types (Adams et al., 1995; Aguiar et al., estimation. An alternative approach is to use the LSMA
1999; Cochrane & Souza, 1998; DeFries et al., 2000; Lu et al., approach. Impervious surfaces may be directly extracted from
2003). In urban studies, LSMA has shown the potential for multispectral images as one of the end members (Phinn et al.,
estimating impervious surface and vegetation abundance and 2002; Rashed et al., 2001). The impervious surface estimation
improving urban classification (Lu & Weng, 2004; Phinn et al., may be improved by the addition of low-albedo and high-
2002; Rashed et al., 2001; Small, 2001, 2002; Wu & Murray, albedo fraction images resulting from spectral mixture analysis
2003). Because impervious surface is closely related to urban (Wu, 2004; Wu & Murray, 2003). Because a low-albedo
land-use patterns, the use of impervious surface may provide fraction image may relate to different kinds of features,
new insights for urban land-use classification. including water, canopy shadows, building shadows, moisture
Impervious surfaces are anthropogenic features through in grass or crops, and dark impervious surface materials, it is
which water cannot infiltrate the soil, including roads, critical to remove other types of covers from the low-albedo
driveways, sidewalks, parking lots, rooftops, and so on. fraction image before it is used for the extraction of
Research on impervious surface extraction from remotely impervious surfaces. However, previous research has not
sensed data has attracted interest since the 1970s. Slonecker examined the advantage of this approach. Because of different
et al. (2001) grouped methods of impervious surface thermal responses between impervious surface and other cover
extraction from remotely sensed data into three categories: types, the use of thermal infrared data may be able to aid
interpretive applications, spectral applications, and modeling impervious surface extraction. Schueler (1994) noted that
applications, which were mainly based on the achievements impervious surface areas could yield temperature up to 10–12
in the 1970s and 1980s. Brabec et al. (2002) summarized four degrees higher than the fields and forests they replaced,
ways for evaluating impervious surface: using a planimeter to because impervious surfaces can hold and emit more heat than
measure impervious surface areas on an aerial photograph, natural features. Since the usefulness of thermal infrared
counting the number of intersections on the overlaid grid on images, such as those from Landsat TM/ETM+ or Terra
an aerial photograph, conducting image classification, and ASTER data, has not been examined for assisting impervious
estimating impervious surface through the percentage of surface extraction, our objectives of this paper are to test the
urbanization in a region. Recently, research in impervious feasibility and effectiveness of using land surface temperature
surface extraction has moved toward per-pixel image data to improve impervious surface mapping, and then to
classification (Dougherty et al., 2004; Hodgson et al., 2003; classify urban land-use classes based on the combined use of
Jennings et al., 2004), subpixel classification (Civico et al., impervious surface and population density, because land-use
2002; Ji & Jensen, 1999; Phinn et al., 2002; Rashed et al., distributions are closely related to both patterns of impervious
2003), and decision tree modeling (Goetz et al., 2004; Jantz surface and population density.
et al., 2005; Yang et al., 2003a, 2003b). Extraction of
impervious surfaces have also been conducted by the 2. Study area and data used
combination of high-albedo and low-albedo fraction images
(Wu, 2004; Wu & Murray, 2003), and by establishing the 2.1. Description of study area
relationship of impervious surfaces with vegetation cover
(Bauer et al., 2004; Gillies et al., 2003). Indianapolis/Marion County, Indiana, USA, was chosen as
Extraction of impervious surfaces is still a challenge the study area (Fig. 1). It possesses several advantages that
because of the complexity of urban and suburban landscapes make it an appropriate choice for such a study. It has a single
and the limitation of remotely sensed data in spectral and central city, and other large urban areas in the vicinity have not
spatial resolutions. The extractions based on an image influenced its growth. The city is located on a flat plain and is
classification are often underestimated due to the heterogene- relatively symmetrical, having possibilities of expansion in all
ity of urban landscapes. The impervious surfaces may be directions. Like many other American cities, Indianapolis is
mixed with other land cover types, such as trees, grasses, and rapidly increasing in population and in area. The areal
soils. The difficulty in selecting training samples for expansion occurs through encroachment into the adjacent
impervious surfaces is a major factor resulting in land-use/ agricultural and non-urban land. Certain decision-making
cover misclassification. When high spatial resolution data, forces, such as density of population, distance to work, property
such as IKONOS, are used for impervious surface mapping value, and income structure, encourage some sectors of
(Goetz et al., 2003), shadows caused by tall buildings or large metropolitan Indianapolis to expand faster than others. Extract-
tree crowns may create a severe problem in effective ing information of impervious surface from satellite images
148 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160

Fig. 1. Study area – Marion County (city of Indianapolis), Indiana, USA.

allows for monitoring urban changes over time, and integrating clear weather conditions, was used in this research. The ETM+
imperviousness into other spatial and non-spatial data for image has one panchromatic band with 15-m spatial resolution,
environmental management and urban planning activities. six reflective bands with 30-m spatial resolution, and one
thermal infrared band with 60-m spatial resolution. The ETM+
2.2. Data sets data were geometrically rectified with 1 : 24,000 topographic
maps. The root-mean-square error (RMSE) during image
A Landsat 7 ETM+ image (path 21/row 32) of Marion rectification was less than 0.2 pixels. A nearest-neighbor
County, Indiana, which was acquired on June 22, 2000 under algorithm was used to resample the ETM+ image (including
 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160 149

the six reflective bands and the thermal infrared band) to a potential to distinguish impervious surfaces from other cover
pixel size of 30 m × 30 m during image rectification. The types in the study area.
ETM+ reflective bands were used to develop fraction images The 2000 census data in a shapefile format were used for
with LSMA, and the thermal infrared band was used to assisting urban land-use classifications. Because of the
compute land surface temperature. The ETM+ panchromatic difference in coordinate systems between census data and the
band was not used for the research described in this paper. ETM+ image, the geographic coordinate of the census data
No atmospheric calibration was conducted for the ETM+ was converted into UTM to be consistent with the coordinate
image, because previous research had demonstrated that system of the ETM+ image. In the census data, population data
atmospheric calibration did not have an effect on fraction were organized at three levels: tract, block group, and block.
images when image end members were used (Lu et al., 2004; The block level provided the most detailed population
Small, 2004). information and was thus used in this research. Because the
A land surface temperature image (Fig. 2) was directly used size of each block depended on the residential pattern, it was
in this paper. A detailed description of the computation of land necessary to calculate population density (i.e., population/area
surface temperature from the ETM+ thermal infrared band was at the block level, persons/km2) in order to eliminate the
provided in Weng et al. (2004). The temperature gradually impact of different block sizes. Fig. 3 illustrated population
decreased from the highest values in commercial areas (such as density distribution in the study area. The population density in
A in Fig. 2), to medium values in high- and medium-intensity commercial/industrial areas and in agricultural areas was very
residential areas (such as B and C), to lowest values in non- low. The population was mainly distributed in residential areas
urban areas such as crops, forested lands, and water (such as D, of different intensities.
E, and F). The difference in land surface temperature between Orthophotographs were used for validation of impervious
impervious surfaces and other cover types provided the surface estimation results and for accuracy assessment of the

Fig. 2. Land surface temperature image derived from a Landsat 7 ETM+ thermal infrared band (land surface temperature ranges from 290 to 320K, with lowest values
in black and highest values in white): (A) commercial areas; (B) high-intensity residential areas; (C) medium-intensity residential areas; (D) agricultural lands; (E)
forested lands; (F) water.
150 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160

Fig. 3. A map of population density (persons/km2) based on 2000 Census data.

urban land-use classification image. The color orthophoto- of the area covered by distinct features on the ground (Adams
graphs were provided by the Indianapolis Mapping and et al., 1995; Mustard & Sunshine, 1999). In the LSMA
Geographic Infrastructure System, which was acquired in approach, end member selection is a key step, and many
April 2003 for the entire county. The orthophotographs have a approaches have been developed (Lu et al., 2003; Theseira et
spatial resolution of 0.14 m. The coordinate system belongs to al., 2003). In practice, image-based end member selection
Indiana State Plane East, Zone 1301, with North American methods are frequently used, because image end members can
Datum of 1983. The orthophotographs were reprojected and be easily obtained and they represent the spectra measured at
resampled to 1-m pixel size for the sake of quicker display and the same scale as the image data. Image end members can be
shorter computing time. derived from the extremes of the image feature space,
assuming that they represent the purest pixels in the images
3. Mapping of impervious surfaces (Mustard & Sunshine, 1999). In order to effectively identify
image end members and to achieve high-quality end
3.1. Linear spectral mixture analysis members, different image transform approaches, such as
principal component analysis and minimum noise fraction,
The LSMA approach assumes that the spectrum measured may be used to transform the multispectral images into a new
by a sensor is a linear combination of the spectra of all data set (Boardman & Kruse, 1994; Green et al., 1988). End
components (end members) within the pixel and that the members are then selected from the feature spaces of the
spectral proportions of the end members represent proportions transformed images (Cochrane & Souza, 1998; Garcia-Haro et
 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160 151

al., 1996; Small, 2001, 2002, 2004; Van der Meer & de Jong, In this research, the minimum noise fraction procedure was
2000). In this research, image end members were selected applied to transform the six ETM+ reflective bands into a new
from the feature spaces formed by the minimum noise coordinate set. The first three components accounted for the
fraction components. majority of the information (approximately 99%) and were used
for selection of end members. The scatterplots between
3.2. Development of fraction images minimum noise fraction components 1, 2, and 3 were illustrated
in Fig. 4, showing the potential end members. Four end
The minimum noise fraction transform contains two steps: members: vegetation, high-albedo, low-albedo, and soil, were
(1) de-correlation and rescaling of the noise in the data based on selected. A constrained least-squares solution was then applied
an estimated noise covariance matrix, producing transformed to unmix the six ETM+ reflective bands into four fraction
data in which the noise has unit variance and no band-to-band images, which were shown in Fig. 5. The high-albedo fraction
correlations; and (2) implementation of a standard principal image mainly related to impervious surface information in the
component analysis of the noise-whitened data. The result of urban area and some dry soils in agricultural areas. The
minimum noise fraction transform is a two-part data set, one commercial/industrial/transportation areas exhibited the highest
part associated with large eigenvalues and coherent eigen- values in the high-albedo fraction image. The value was
images, and a complementary part with near-unity eigenvalues gradually decreased from high-intensity residential to low-
and noise-dominated images (ENVI, 2000). In the minimum intensity residential areas. The low-albedo fraction image was
noise fraction transform, noise is separated from the data by more complex than other fraction images, because it contained
using only the coherent portions, in order to improve spectral different features, such as water, building shadows in the central
processing results. business district, vegetation canopy shadows in forested areas,

Fig. 4. Feature spaces between the three components from minimum noise fraction transform of the Landsat ETM+ image, illustrating four potential end members:
high-albedo, low-albedo, vegetation, and soil.
152 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160

Fig. 5. Four fraction images from spectral mixture analysis of six ETM+ reflective bands (the fraction values range from 0 to 1, with lowest values in black and highest
values in white in the fraction images): (A) high-albedo; (B) low-albedo; (C) soil; (D) vegetation.

and dark impervious surface materials. The soil fraction image hand, the low-albedo fraction image highlighted information of
highlighted the soil information located in the agricultural areas, water and shadows, such as water bodies, shadows from forest
and the vegetation fraction image detected high values in forest, canopy and tall buildings, and moistures in crops or pastures.
pasture, grass, and croplands. However, dark impervious surface information was also
included in the low-albedo fraction image. It is important to
3.3. Development of impervious surface image develop a suitable approach for removal of other cover types
from the low-albedo fraction image. This study found that the
Extraction of impervious surface information from remotely incorporation of land surface temperature provided a good
sensed data is a complex procedure. As illustrated in Fig. 5, premise to separate other cover types from impervious surface
some other cover information was also included in the high- covers on the low-albedo fraction image.
albedo and low-albedo fraction images. Direct extraction of The scatterplot between land surface temperature and low-
impervious surface image through the addition of high-albedo albedo fraction images can reveal the different characteristics of
and low-albedo fraction images is obviously unsuitable. land surface temperature and low-albedo fraction relationships
Although the high-albedo fraction image related mainly to among the land-use/cover classes (Fig. 6). The four corners A,
impervious surface information, such as buildings and roads, B, C, and D represent four land covers: corner A represents
other cover type information, such as dry soils, was also grass/crops with very low values in both low-albedo fraction
included due to the similar spectral responses between some and land surface temperature images; corner B represents water
bright impervious surface materials and dry soils. On the other with very high values in the low-albedo fraction image but very
 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160 153

Fig. 6. Feature space created from the low-albedo fraction and land surface temperature images, illustrating the conceptual relationships of land-use/cover distribution.

low values in the land surface temperature image; corner C surface from other cover type pixels, including water and
represents bright impervious surface objects with very low vegetation (forest, pasture, grass, and crops). The following
values in the low-albedo fraction image but very high values expert rules were used in this study:
in the land surface temperature image; and corner D represents
dark impervious surface objects with very high values in both (1) If the pixel values in the land surface temperature image
the low-albedo fraction and land surface temperature images. were less than or equal to t1, pixels in low-albedo and
Vegetation (crops, grass, and forest) and water possess lower high-albedo fraction images would be assigned to zero,
land surface temperature values than built-up lands. Land otherwise, pixel values in low-albedo and high-albedo
surface temperatures of shadow areas are influenced strongly fraction images would be kept; and
by the characteristics of the features generating the shadows (2) If the pixel values in the soil fraction image were greater
and by their interactions with the surrounding pixels. For than t2, pixels in high-albedo fraction image would be
example, shadows in vegetation canopy were found to be assigned to zero, otherwise, pixel values in the high-
generally cooler than shadows cast by tall buildings, because albedo fraction image would be kept.
vegetation offsets absorptive heating by transpiration. For dark
impervious surface objects with their absorptive features, land In these rules, the t1 and t2 thresholds were identified
surface temperature increased as the values of low-albedo based on reference data of water, vegetation, and soils. After
fraction increased. Shadow and moisture both generated a other cover type pixels were removed from the low-albedo
cooling effect, while impervious surface showed a heating and high-albedo fraction images, the impervious surface
effect. image was then developed by addition of the modified low-
For the high-albedo fraction image, impervious surface was albedo and high-albedo fraction images. Fig. 7 provided a
predominantly confused with dry soils. So the soil fraction comparison of the impervious surface images before and
image may be used to remove the pixels of soils in the high- after the adjustment procedure. The final impervious surface
albedo fraction image. For the low-albedo fraction image, dark image (B in Fig. 7) indicated that the other cover type
impervious surface was confused with water and shadows. information was successfully removed, comparing to the
Therefore, the critical process was to separate impervious direct addition of low-albedo and high-albedo fraction
154 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160

3.4. Evaluation of impervious surface image

Accuracy assessment is an important step to ensure the

quality of the impervious surface image. Selecting sufficient
reference data through a suitable sampling scheme is important.
Seventy-six sample plots with 10 × 10 pixel sizes (i.e.,
300m × 300 m) were initially selected on the impervious surface
image. The plots were selected on the intersections between
straight lines and cycles, with the distance of 100 pixels between
two cycles (Fig. 8). The selected sample plots on the impervious
surface image were linked to corresponding orthophotographs.
The impervious surface objects within each sample plot were
digitized on the orthophotograph. The proportion of impervious
surface in each plot was then calculated. Because of the 3-year
difference between ETM+ data (2000) and the referenced
orthophotographs (2003), a careful check of each sample plot
between a satellite color composite and orthophotographs was
conducted. We found that seven sample plots showed apparent
changes in landscape composition, and they were therefore
removed during the accuracy assessment. Sixty-nine samples
were finally used to assess the developed impervious surface
image quality. The RMSE and system error were calculated
(Wu, 2004). The accuracy assessment results were also
analyzed with graphs (Fig. 9). The samples with low impervious
surface (such as less than 0.6) tended to be overestimated, while
samples with high impervious surface (such as greater than 0.6)
were underestimated. The trend was in agreement with previous
work, but the residuals were smaller than previous, similar work
(Wu, 2004). The overall RMSE of 9.22% and a system error of
5.68% were obtained. For sample plots with less than 30%
impervious surface, the RMSE was 9.98% and system error
8.59%. For sample plots with greater than or equal to 30%
impervious surface, errors became smaller, with RMSE of
8.36% and system error of 2.77%.

4. Urban land-use classification

The classification approach used in this paper was based on

the expert systems on the integration of impervious surface and
population density data. Because of the unique feature of
comparing with classifications based on spectral signatures, a
suitable definition of land-use classes based on impervious
surface and population density was required.
Fig. 7. Comparison of impervious surface images developed from different
approaches (the image values range from 0 to 1, with lowest values in black and
4.1. The classification scheme
highest values in white): (A) based on the direct addition of high-albedo and
low-albedo fraction images; (B) based on the addition of modified high-albedo Although medium spatial resolution data (10–100m), such
and low-albedo fraction images, which other cover types were removed in the as Landsat TM/ETM+, were frequently used for land-cover
impervious surface image through the combinative use of land surface classifications, the heterogeneity of urban environments and the
temperature and fraction images.
large number of mixed pixels inherent images often induced
difficulty in urban land-use/cover classification based on
spectral signatures (Lu & Weng, 2004). Traditional per-pixel
images (see A in Fig. 7). The gradient impervious surface classification algorithms made it difficult to distinguish urban
change was obvious, from highest in commercial areas in the land-use/cover classes. Remote sensors mainly capture the
center of the study area, to medium in the high-intensity features of Earth's surface, i.e., land-cover information. Hence,
residential areas, to the lowest in the low-intensity residential remotely sensed data are more suitable for land-cover
areas. classification, instead of land-use classification. In urban
 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160 155

Fig. 8. Strategy for collection of reference data for the accuracy assessment of impervious surface image, illustrating the approach for allocating sample plots on the
impervious surface image and for digitizing the impervious surface areas on the orthophotograph.

studies, however, land-use is more useful than land-cover images was illustrated in Fig. 10. The critical issue was to
information, because land-use is directly related to social– develop a high-quality impervious surface image and to
economic activities. A uniform definition of urban land-use establish expert rules of population density and impervious
classes is not available yet. Most previous research classified surface for each urban land-use class. In this research, five
urban built-up areas into residential areas and commercial and land-use classes were classified with a rule-based approach.
industrial areas, or their subcategories when high-resolution Water and non-urban areas were classified based on low-
data were used (Epstein et al., 2002; Phinn et al., 2002; Van der albedo and high-albedo fraction images. The classification
Sande et al., 2003; Zha et al., 2003). The criteria for result from the 2000 ETM+ image (Fig. 11) demonstrated that
classification of urban land-use/cover classes were different, the land-use seemed to display a concentric pattern from the
and classification accuracies varied with the classification center to the peripheral, transiting from commercial lands to
approaches employed and ancillary data incorporated (Phinn et very-high-, high-, medium-, and finally low-intensity residen-
al., 2002; Zha et al., 2003). Because of the close relationship tial lands. Commercial lands were mainly distributed in the
between population density distribution and urban land-use center of the city. High- and very-high-intensity residential
pattern, our study classified the urban landscape into five land- lands were distributed around commercial lands, especially in
use classes (i.e., very-high-, high-, medium-, and low-intensity the eastern part of the city. Medium-intensity residential lands
residential lands, and commercial/industrial/transportation were located in the transition zone between high-intensity
lands) based on the combination of impervious surface and residential lands and non-urban areas. Low-intensity residen-
population density images. Other classes included water and tial lands were scattered in the vegetated and agricultural
non-urban areas (vegetated and agricultural lands). Table 1 areas.
provides the definition of urban land-use types used in this
paper. 4.3. Accuracy assessment of the land-use classification image

4.2. Urban classification with integration of impervious A common method for accuracy assessment of a classifica-
surface and population density tion image is through the use of an error matrix. Some important
measures, such as overall accuracy, producer's accuracy, and
The procedure of urban land-use classification based on the user's accuracy, can be calculated form the error matrix
combination of impervious surface and population density (Congalton, 1991; Congalton & Green, 1999; Foody, 2002;
156 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160

values than vegetation, especially in the visible and shortwave

infrared bands, exaggerating their area proportions accounted
for in the mixed pixels. Another factor was that dark
impervious surface objects and water or shadows had similar
spectral characteristics and similar values in the low-albedo
fraction image. Dark impervious surfaces from the low-albedo
fraction image may be over extracted based on the difference
in land surface temperature because relative lower spatial
resolution of Landsat ETM+ thermal band smoothed the
boundaries between impervious surfaces and other cover
types. In the well-developed areas, such as high-, and very-
high-intensity residential lands, similar spectral responses in
some bright impervious surfaces and dry soils created
confusion, as shown in Fig. 5C. Tree crowns that cover
portions of some impervious surfaces, such as roads, may be
another reason resulting in underestimation of impervious
surfaces.
Impervious surfaces are complex in terms of composition of
materials. The limitation in spectral bands of medium spatial
resolution images, such as Landsat TM/ETM+ data, and the
high correlation among the image bands restrain the use of a
large number of end members in the standard LSMA approach.
Moreover, the image-based end member selection method may
not effectively identify different impervious surface end
members. However, some potential methods may remedy this
problem. For example, hyperspectral imagery may be
employed to extract a large number of end members. Use of
reference end members, in conjunction with image end
members, might provide more impervious surface end
members. In addition, multiple end member LSMA, instead
Fig. 9. Accuracy assessment results of estimated impervious surface from the of standard LSMA, may be applied for development of fraction
Landsat ETM+ image. images, if a large number of end members are desirable. This
method has been applied to retrieve vegetation, soil, and snow
Smits et al., 1999). In this study, a total of 230 sample plots were cover, and has showed a better performance than the standard
randomly allocated on the classified image. Excluding the plots LSMA approach (Okin et al., 1999, 2001; Painter et al., 1998;
outside the study area, 206 plots were examined on the high Rashed et al., 2003; Roberts et al., 1998). The feasibility of
spatial resolution orthophotographs. Table 2 summarizes the applying a multiple end member LSMA approach to extract
accuracy assessment results. The overall classification accura-
cies of 87.38% for seven classes and of 83.78% for five urban Table 1
land-use classes were achieved. The classification of commer- Definitions of urban land-use types
cial/industrial/transportation lands and medium-intensity resi- Urban land-use type CODE Definition
dential lands were especially successful. The major errors were
Low-intensity LIRLImpervious surface is very low, usually less
from the confusions between low-intensity residential lands and residential lands than 30%; population density is less than
non-urban areas and between adjacent intensities of residential 500persons/km2
areas, such as low-intensity and medium-intensity residential Medium-intensity MIRL Impervious surface is usually greater than
lands. residential lands 20%, but often less than 50%; population
density is between 500 and 1500persons/km2
High-intensity HIRL Impervious surface is usually greater than
5. Discussion residential lands 40%; population density is between 1500 and
3000persons/km2
5.1. Issues related to estimation of impervious surface Very-high-intensity VIRL Impervious surface is usually greater than
residential lands 50%; population density is greater than
3000persons/km2
This research showed that impervious surface was over-
Commercial, industrial, CITL Impervious surface is usually greater than
estimated in the less-developed areas, but was underestimated and transportation 60%; population density is very low, usually
in the well-developed areas. In the less-developed areas, such lands less than 10persons/km2
as medium- and low-intensity residential lands, impervious Non-urban NURL Vegetated areas and agricultural lands
surfaces were often mixed with vegetation and soil. However, lands
Water WAT Water bodies
bright impervious surface objects had higher reflectance
 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160 157

Population density Impervious surface Low-albedo fraction High-albedo fraction

(PopD) (ImpS) (LAF) (HAF)

Other-cover-types
(OCT)
ImpS > 0.6 and
Yes CITL
PopD < 10?

Is OCT?
No Extraction OCT from
LAF and HAF

ImpS > 0.5 and Yes Yes

Yes VIRL
PopD > 3000?
Adjusted LAF Adjusted HAF
(ALAF) (AHAF)
No

ALAF > 0.7 and

ImpS > 0.4 and No
HIRL AHAF < 0.2?
PopD > 1500? Yes Urban land
use image
Yes
No

Water
ImpS > 0.2 and
Yes MIRL
PopD > 500?
Combination of Other land covers
No land use classes (e.g., forest,
agricultural lands)

ImpS < 0.3 and

LIRL Final urban land use
PopD ≤ 500? Yes
Classification image

No

Other-cover-types

Fig. 10. Procedure for urban land-use classification based on the integration of impervious surface and population density images. ImpS and PopD represent
impervious surface and population density (persons/km2); and CITL, VIRL, HIRL, MIRL, and LIRL represent commercial/industrial/transportation lands, very-high-,
high-, medium-, and low-intensity residential lands, respectively.

different impervious surface materials constitutes a promising research. This may be due to the differences in data
direction for future research. structures, spatial resolutions (scales), and coordinate systems
between population data and remotely sensed data. The sizes
5.2. Issues related to urban land-use classification of polygons for census units vary with urban land-use
patterns. Hence, GIS plays an important role in the integration
Image classification is usually based directly on spectral of population and remotely sensed data. This research
signatures that are captured by remote sensors. However, indicates that setting up thresholds for impervious surface
complex landscapes with the mixed-pixel problem would lead and population density for each urban land-use category is the
to the difficulties in classification. In urban environments, key for success with this approach.
classification algorithms based solely on spectral signatures
cannot effectively handle the confusions (1) among water, 6. Conclusions
dark impervious surface objects, and shadows from tall
buildings or forest canopy; (2) among dry soils, commercial/ Five urban land-use classes can be successfully classified
industrial/transportation lands, and very high intensity resi- using medium spatial resolution remotely sensed data with
dential lands; and (3) between forest and low-intensity an overall classification accuracy of 83.78%. The critical
residential lands. In particular, spectral-based classification issue is to develop a high-quality impervious surface image.
approaches are not good at separating different densities of The integration of land surface temperature and LSMA-
residential classes and not good at direct classification of derived fraction images has been demonstrated to be
urban land-use classes. Our research demonstrates that a effective for refining the impervious surface image, which
classification approach based on the combination of imper- has an overall RMSE of 9.22% and a system error of 5.68%.
vious surface and population density can successfully solve Because multispectral reflective images as well as thermal
these problems. Population data has not been extensively used infrared images, such as Landsat TM/ETM+ and Terra
for improvement of urban land-use classification in previous ASTER, are readily available, estimation of impervious
158 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160

Fig. 11. Urban land-use classification image with the rule-based approach based on the integration of impervious surface and population density images. CITL, VIRL,
HIRL, MIRL, and LIRL represent commercial/industrial/transportation lands, very-high-, high-, medium-, and low-intensity residential lands, respectively; NURL
represents non-urban lands including vegetated and agricultural lands; and WAT represents water.

surface from remotely sensed data should be improved population density provides a new insight for urban land-use
through the combinative use of land surface temperature and classification and the approach developed in this paper can
fraction images. The integration of impervious surface and be applied to other urban environments.

Table 2
Accuracy assessment of the urban land-use classification image
Reference data
LIRL MIRL HIRL VIRL CITL NURL WAT RT UA%
Classified data LIRL 6 5 11 54.55
MIRL 7 47 3 57 82.46
HIRL 4 12 16 75.00
VIRL 4 6 10 60.00
CITL 22 22 100.00
NURL 1 1 1 80 83 96.39
WAT 7 7 100.00
CT 14 52 19 6 23 85 7 206
PA% 42.86 90.38 63.16 100.00 95.65 94.12 100.00 87.38
(1) CITL, VIRL, HIRL, MIRL, and LIRL represent commercial/industrial/transportation lands, very-high-, high-, medium-, and low-intensity residential lands, respectively.
(2) NURL and WAT represent non-urban lands (vegetated and agricultural lands) and water.
(3) CT and RT mean column total and row total.
(4) PA and UA mean producer's accuracy and user's accuracy.
(5) The overall classification accuracies for seven classes and for five urban land-use classes are 87.38% and 83.78%.
 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160 159

Acknowledgments Fisher, P. (1997). The pixel: A snare and a delusion. International Journal of
Remote Sensing, 18, 679–685.
Foody, G. M. (2002). Status of land cover classification accuracy assessment.
The authors would like to thank three anonymous Remote Sensing of Environment, 80, 185–201.
reviewers for their constructive comments and suggestions. Garcia-Haro, F. J., Gilabert, M. A., & Melia, J. (1996). Linear spectral mixture
This research is supported by National Science Foundation modeling to estimate vegetation amount from optical spectral data.
(BCS-0521734) for a project entitled “Role of Urban Canopy International Journal of Remote Sensing, 17, 3373–3400.
Gillies, R. R., Box, J. B., Symanzik, J., & Rodemaker, E. J. (2003). Effects of
Composition and Structure in Determining Heat Islands: A
urbanization on the aquatic fauna of the Line Creek watershed, Atlanta – a
Synthesis of Remote Sensing and Landscape Ecology satellite perspective. Remote Sensing of Environment, 86, 411–422.
Approach”, and by the USGS IndianaView program and the Gluch, R. (2002). Urban growth detection using texture analysis on merged
NASA's Indiana Space Grant program for a project entitled Landsat TM and SPOT-P data. Photogrammetric Engineering and Remote
“Indiana Impervious Surface Mapping Initiative.” Dengsheng Sensing, 68, 1283–1288.
Lu also acknowledges the support from the Center for the Goetz, S. J., Jantz, C. A., Prince, S. D., Smith, A. J., Wright, R., & Varlyguin, D.
(2004). Integrated analysis of ecosystem interactions with land use change:
Study of Institutions, Population, and Environmental Change The Chesapeake Bay watershed. In R. S. DeFries, G. P. Asner, & R. A.
at Indiana University, through the National Science Founda- Houghton (Eds.), Ecosystems and land use change (pp. 263–275).
tion (grants 99-06826). Washington, DC: American Geophysical Union.
Goetz, S. J., Wright, R., Smith, A. J., Zinecker, E., & Schaub, E. (2003). Ikonos
imagery for resource management: Tree cover, impervious surface, and
References riparian buffer analyses in the mid-Atlantic region. Remote Sensing of
Environment, 88, 195–208.
Adams, J. B., Sabol, D. E., Kapos, V., Filho, R. A., Roberts, D. A., Smith, M. O., Gong, P., & Howarth, P. J. (1990). The use of structure information for
et al. (1995). Classification of multispectral images based on fractions of improving land-cover classification accuracies at the rural–urban fringe.
endmembers: Application to land cover change in the Brazilian Amazon. Photogrammetric Engineering and Remote Sensing, 56, 67–73.
Remote Sensing of Environment, 52, 137–154. Green, A. A., Berman, M., Switzer, P., & Craig, M. D. (1988). A transformation
Aguiar, A. P. D., Shimabukuro, Y. E., & Mascarenhas, N. D. A. (1999). Use of for ordering multispectral data in terms of image quality with implications
synthetic bands derived from mixture models in the multispectral for noise removal. IEEE Transactions on Geoscience and Remote Sensing,
classification of remote sensing images. International Journal of Remote 26, 65–74.
Sensing, 20, 647–657. Haack, B. N., Solomon, E. K., Bechdol, M. A., & Herold, N. D. (2002).
Bauer, M. E., Heinert, N. J., Doyle, J. K., & Yuan, F. (2004). Impervious surface Radar and optical data comparison/integration for urban delineation:
mapping and change monitoring using Landsat remote sensing. ASPRS A case study. Photogrammetric Engineering and Remote Sensing, 68,
annual conference proceedings, May 23–28, 2004, Denver, Colorado. 1289–1296.
Bethesda, MD: American Society for Photogrammetry and Remote Sensing. Harris, P. M., & Ventura, S. J. (1995). The integration of geographic data with
Boardman, J. W., & Kruse, F. A. (1994). Automated spectral analysis: A remotely sensed imagery to improve classification in an urban area.
geological example using AVIRIS data, north Grapevine Mountains, Photogrammetric Engineering and Remote Sensing, 61, 993–998.
Nevada. Proceedings, ERIM tenth thematic conference on geologic remote Hodgson, M. E., Jensen, J. R., Tullis, J. A., Riordan, K. D., & Archer, C. M.
sensing, Ann Arbor, MI (pp. 407–418). (2003). Synergistic use of LiDAR ad color aerial photography for mapping
Brabec, E., Schulte, S., & Richards, P. L. (2002). Impervious surface and water urban parcel imperviousness. Photogrammetric Engineering and Remote
quality: A review of current literature and its implications for watershed Sensing, 69, 973–980.
planning. Journal of Planning Literature, 16, 499–514. Hung, M., & Ridd, M. K. (2002). A subpixel classifier for urban land-cover
Civico, D. L., Hurd, J. D., Wilson, E. H., Arnold, C. L., & Prisloe Jr., M. P. mapping based on a maximum-likelihood approach and expert system rules.
(2002). Quantifying and describing urbanizing landscapes in the northeast Photogrammetric Engineering and Remote Sensing, 68, 1173–1180.
United States. Photogrammetric Engineering and Remote Sensing, 68, Jantz, P., Goetz, S. J., & Jantz, C. A. (2005). Urbanization and the loss of
1083–1090. resource lands within the Chesapeake Bay watershed. Environmental
Cochrane, M. A., & Souza Jr., C. M. (1998). Linear mixture model classification Management, 36, 808–825.
of burned forests in the eastern Amazon. International Journal of Remote Jennings, D. B., Jarnagin, S. T., & Ebert, C. W. (2004). A modeling approach for
Sensing, 19, 3433–3440. estimating watershed impervious surface area from national land cover data
Congalton, R. G. (1991). A review of assessing the accuracy of classifications of 92. Photogrammetric Engineering and Remote Sensing, 70, 1295–1307.
remotely sensed data. Remote Sensing of Environment, 37, 35–46. Ji, M., & Jensen, J. R. (1999). Effectiveness of subpixel analysis in detecting and
Congalton, R. G., & Green, K. (1999). Assessing the accuracy of remotely quantifying urban imperviousness from Landsat Thematic Mapper.
sensed data: Principles and practices. Boca Raton, FL: CRC/Lewis Press. Geocarto International, 14, 31–39.
Cracknell, A. P. (1998). Synergy in remote sensing – What's in a pixel? Lu, D., Batistella, M., Moran, E., & Mausel, P. (2004). Application of spectral
International Journal of Remote Sensing, 19, 2025–2047. mixture analysis to Amazonian land-use and land-cover classification.
Dare, P. M. (2005). Shadow analysis in high-resolution satellite imagery of International Journal of Remote Sensing, 25, 5345–5358.
urban areas. Photogrammetric Engineering and Remote Sensing, 71, Lu, D., Moran, E., & Batistella, M. (2003). Linear mixture model applied to
169–177. Amazonian vegetation classification. Remote Sensing of Environment, 87,
DeFries, R. S., Hansen, M. C., & Townshend, J. R. G. (2000). Global continuous 456–469.
fields of vegetation characteristics: A linear mixture model applied to multi- Lu, D., & Weng, Q. (2004). Spectral mixture analysis of the urban landscapes in
year 8km AVHRR data. International Journal of Remote Sensing, 21, Indianapolis with Landsat ETM+ imagery. Photogrammetric Engineering
1389–1414. and Remote Sensing, 70, 1053–1062.
Dougherty, M., Dymond, R. L., Goetz, S. J., Jantz, C. A., & Goulet, N. (2004). Lu, D., & Weng, Q. (2005). Urban classification using full spectral information
Evaluation of impervious surface estimates in a rapidly urbanizing of Landsat ETM+ imagery in Marion County, Indiana. Photogrammetric
watershed. Photogrammetric Engineering and Remote Sensing, 70, Engineering and Remote Sensing, 71, 1275–1284.
1275–1284. Mesev, V. (1998). The use of census data in urban image classification.
ENVI. (2000). ENVI user's guide. Boulder, CA: Research Systems Inc. Photogrammetric Engineering and Remote Sensing, 64, 431–438.
Epstein, J., Payne, K., & Kramer, E. (2002). Techniques for mapping suburban Mustard, J. F., & Sunshine, J. M. (1999). Spectral analysis for earth science:
sprawl. Photogrammetric Engineering and Remote Sensing, 63, 913–918. Investigations using remote sensing data. In A. N. Rencz (Ed.), Remote
160 D. Lu, Q. Weng / Remote Sensing of Environment 102 (2006) 146–160

sensing for the earth sciences: Manual of remote sensing, vol. 3 (3rd ed.) Smits, P. C., Dellepiane, S. G., & Schowengerdt, R. A. (1999). Quality
(pp. 251–307). New York, NY: John Wiley & Sons Inc. assessment of image classification algorithms for land-cover mapping: A
Myint, S. W. (2001). A robust texture analysis and classification approach for review and a proposal for a cost-based approach. International Journal of
urban land-use and land-cover feature discrimination. Geocarto Interna- Remote Sensing, 20, 1461–1486.
tional, 16, 27–38. Stefanov, W. L., Ramsey, M. S., & Christensen, P. R. (2001). Monitoring urban
Okin, W. J., Okin, G. S., Roberts, D. A., & Murray, B. (1999). Multiple land cover change: An expert system approach to land cover classification of
endmember spectral mixture analysis: Endmember choice in an arid semiarid to arid urban centers. Remote Sensing of Environment, 77,
shrubland. In R. O. Green (Ed.), The 1999 AVIRIS workshop. Pasadena, 173–185.
CA: Jet Propulsion Laboratory. Stuckens, J., Coppin, P. R., & Bauer, M. E. (2000). Integrating contextual
Okin, G. S., Roberts, D. A., Murray, B., & Okin, W. J. (2001). Practical limits on information with per-pixel classification for improved land cover classifi-
hyperspectral vegetation discrimination in arid and semiarid environments. cation. Remote Sensing of Environment, 71, 282–296.
Remote Sensing of Environment, 77, 212–225. Theseira, M. A., Thomas, G., Taylor, J. C., Gemmell, F., & Varjo, J. (2003).
Painter, T. H., Roberts, D. A., Green, R. O., & Dozier, J. (1998). The effects of Sensitivity of mixture modeling to endmember selection. International
grain size on spectral mixture analysis of snow-covered area from AVIRIS Journal of Remote Sensing, 24, 1559–1575.
data. Remote Sensing of Environment, 65, 320–332. Thomas, N., Hendrix, C., & Congalton, R. G. (2003). A comparison of urban
Phinn, S., Stanford, M., Scarth, P., Murray, A. T., & Shyy, P. T. (2002). mapping methods using high-resolution digital imagery. Photogrammetric
Monitoring the composition of urban environments based on the Engineering and Remote Sensing, 69, 963–972.
vegetation–impervious surface–soil (VIS) model by subpixel analysis Van der Meer, F., & de Jong, S. M. (2000). Improving the results of spectral
techniques. International Journal of Remote Sensing, 23, 4131–4153. unmixing of Landsat Thematic Mapper imagery by enhancing the
Rashed, T., Weeks, J. R., Gadalla, M. S., & Hill, A. G. (2001). Revealing the orthogonality of end-members. International Journal of Remote Sensing,
anatomy of cities through spectral mixture analysis of multispectral satellite 21, 2781–2797.
imagery: A case study of the Greater Cairo region, Egypt. Geocarto Van der Sande, C. J., de Jong, S. M., & de Roo, A. P. J. (2003). A segmentation
International, 16, 5–15. and classification approach of IKONOS-2 imagery for land cover mapping
Rashed, T., Weeks, J. R., Roberts, D., Rogan, J., & Powell, R. (2003). to assist flood risk and flood damage assessment. International Journal of
Measuring the physical composition of urban morphology using multiple Applied Earth Observation and Geoinformation, 4, 217–229.
endmember spectral mixture models. Photogrammetric Engineering and Weng, Q., Lu, D., & Schubring, J. (2004). Estimation of land surface
Remote Sensing, 69, 1011–1020. temperature–vegetation abundance relationship for urban heat island
Ridd, M. K. (1995). Exploring a V–I–S (Vegetation–Impervious Surface–Soil) studies. Remote Sensing of Environment, 89, 467–483.
model for urban ecosystem analysis through remote sensing: Comparative Wu, C. (2004). Normalized spectral mixture analysis for monitoring urban
anatomy for cities. International Journal of Remote Sensing, 16, composition using ETM+ imagery. Remote Sensing of Environment, 93,
2165–2185. 480–492.
Roberts, D. A., Gardner, M., Church, R., Ustin, S., Scheer, G., & Green, R. O. Wu, C., & Murray, A. T. (2003). Estimating impervious surface distribution by
(1998). Mapping chaparral in the Santa Monica mountains using multiple spectral mixture analysis. Remote Sensing of Environment, 84, 493–505.
endmember spectral mixture models. Remote Sensing of Environment, 65, Yang, L., Huang, C., Homer, C. G., Wylie, B. K., & Coan, M. J. (2003). An
267–279. approach for mapping large-area impervious surfaces: Synergistic use of
Schueler, T. R. (1994). The importance of imperviousness. Watershed Landsat-7 ETM+ and high spatial resolution imagery. Canadian Journal of
Protection Techniques, 1, 100–111. Remote Sensing, 29, 230–240.
Shaban, M. A., & Dikshit, O. (2001). Improvement of classification in urban Yang, L., Xian, G., Klaver, J. M., & Deal, B. (2003). Urban land cover change
areas by the use of textural features: The case study of Lucknow city, Uttar detection through sub-pixel imperviousness mapping using remotely
Pradesh. International Journal of Remote Sensing, 22, 565–593. sensed data. Photogrammetric Engineering and Remote Sensing, 69,
Slonecker, E. T., Jennings, D., & Garofalo, D. (2001). Remote sensing of 1003–1010.
impervious surface: A review. Remote Sensing Reviews, 20, 227–255. Zha, Y., Gao, J., & Ni, S. (2003). Use of normalized difference built-up index in
Small, C. (2001). Estimation of urban vegetation abundance by spectral mixture automatically mapping urban areas from TM imagery. International Journal
analysis. International Journal of Remote Sensing, 22, 1305–1334. of Remote Sensing, 24, 583–594.
Small, C. (2002). Multitemporal analysis of urban reflectance. Remote Sensing Zhang, J., & Foody, G. M. (2001). Fully-fuzzy supervised classification of sub-
of Environment, 81, 427–442. urban land cover from remotely sensed imagery: Statistical neural network
Small, C. (2004). The Landsat ETM+ spectral mixing space. Remote Sensing of approaches. International Journal of Remote Sensing, 22, 615–628.
Environment, 93, 1–17. Zhang, Q., & Wang, J. (2003). A rule-based urban land use inferring method for
Smith, M. O., Ustin, S. L., Adams, J. B., & Gillespie, A. R. (1990). Vegetation in fine-resolution multispectral imagery. Canadian Journal of Remote Sensing,
deserts: I. A regional measure of abundance from multispectral images. 29, 1–13.
Remote Sensing of Environment, 31, 1–26.