6_01-124 4/4/03 10:49 PM Page 555

Fractal Analysis of Satellite-Detected

Urban Heat Island Effect
Qihao Weng

Abstract 1988; Roth et al., 1989; Gallo et al., 1993). The 1.1-km spa-
Surface radiant temperatures derived from Landsat TM tial resolution of these data is found suitable only for small-
thermal infrared images of 13 December, 1989, 03 March, scale urban temperature mapping. The 120-m spatial reso-
1996, and 29 August 1997 were used to study the urban lution Landsat TM thermal infrared (TIR) data have also been
heat island (UHI) phenomenon in Guangzhou, China. To utilized to derive surface temperature. Carnahan and Larson
examine the spatial distribution of surface radiant temper- (1990) and Larson and Carnahan (1997) used TM TIR data to
atures, transects were drawn and analyzed from each tem- observe meso-scale temperature differences between the
perature image. Moreover, the fractal dimensions of these urban area and the rural area in Indianapolis, Indiana. Kim
transects were computed using the divider method, so that (1992) studied the phenomenon of urban heat island in
the spatial variability of surface radiant temperatures metropolitan Washington, D.C., and pinpointed the signifi-
caused by the thermal behavior of different land-cover cance of soil albedo and moisture availability to surface en-
types and landscape pattern characteristics can be better ergy balance. Nichol (1994) carried out a detailed study
understood. The effect of urban development on the geo- using TM thermal data to monitor microclimate for housing
graphical distribution of surface radiant temperatures and estates in Singapore. Weng (2001) examined the surface
thus on the UHI was also investigated. The results revealed temperature pattern and its relationship with land cover in
two major heat islands, one in the southwest and the other urban clusters in the Zhujiang Delta, China. More recently,
in the east of the city. The areal extent of the UHIs varied researchers have used very high spatial resolution (10-m)
as the season changed. The transact derived from the ATLAS data (specifically, channel 13: 9.60 to 10.2 mm) to as-
spring image had the lowest fractal dimension while that sess the UHI effect during daytime and nighttime in several
from the summer image the highest value. Urban develop- cities and found it effective (Quattrochi and Ridd, 1994;
ment increased the spatial variability of radiant tempera- Lo et al., 1997; Quattrochi and Luvall, 1999). Quattrochi
tures, resulting in higher fractal dimension values. The and Ridd (1998) applied 5-meter resolution TIMS (Thermal
thermal surfaces have become more spatially uneven and Infrared Multispectral Scanner) to study thermal energy
the textures more complex. responses for two broad classes (i.e., natural grass and
urban/residential trees) and ten subclasses of vegetation in
Salt Lake City, Utah. It is concluded that thermal responses
Introduction for vegetation can be highly varied depending on the bio-
Urban development usually gives rise to a dramatic change physical properties of the vegetation and other factors, and
of the Earth’s surface, as natural vegetation is removed and trees have a great effect on doming or mitigating the ther-
replaced by non-evaporating and non-transpiring surfaces mal radiation upwelling into the atmosphere. Satellite-
such as metal, asphalt, and concrete. This alteration will derived surface temperatures are believed to correspond
inevitably result in the redistribution of incoming solar more closely with the canopy layer heat islands, although a
radiation, and induce the urban-rural contrast in surface ra- precise transfer function between the ground surface tem-
diance and air temperatures. The difference in ambient air perature and the near ground air temperature is not yet
temperature between an urban area and its surrounding available (Nichol, 1994). Byrne (1979) has observed a dif-
rural area is known as the effect of urban heat island (UHI). ference of as much as 20°C between the air temperature
Given the relationship between air temperature, surface and the warmer surface temperature of dry ground.
radiant temperature, and the texture of land cover, the im- Fractals have two basic characteristics suitable for
pact of urban development on the UHI can be examined. modeling the topography and other spatial surfaces in the
Generally speaking, two types of UHI can be distinguished Earth’s surface: self-similarity and randomness. Self-
pertinent to the methods of temperature measurement: similarity refers to the well-known observation that the
(1) the canopy layer heat island and (2) the boundary layer Earth’s morphology appears similar across a range of scales
heat island (Oke, 1979). The former consists of air between (Malinverno, 1995). The concept of self-similarity also con-
the roughness elements, e.g., buildings and tree canopies, tains randomness, because the resemblance of the Earth’s
with an upper boundary just below roof level. The latter is morphology at different scales is not exact but statistical
situated above the former, with a lower boundary subject to (Malinverno, 1995). A fractal construction that includes
the influence of the urban surface. randomness is termed a self-affine fractal. Self-affine frac-
Studies on the UHI phenomenon using satellite remote tals, such as fractional Brownian motion, requires rescaling
sensing data have been conducted primarily using NOAA
AVHRR data (Kidder and Wu, 1987; Balling and Brazell,
Photogrammetric Engineering & Remote Sensing
Vol. 69, No. 5, May 2003, pp. 555–566.
Department of Geography, Geology, and Anthropology, 0099-1112/03/6905–555$3.00/0
Indiana State University, Terre Haute, IN 47809 (geweng@ © 2003 American Society for Photogrammetry
scifac.indstate.edu). and Remote Sensing

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2003 555

6_01-124 4/4/03 10:49 PM Page 556

by different coordinates for an enlargement to look like the ence of two navigable rivers of the Zhujiang River (literally
original, and have been used to model the topography of the the “Pearl River”) system (Figure 1). With a population
Earth’s surface (Mandelbrot, 1983). If remotely sensed im- of 3.99 million and total area of 1,444 km2 (Guangdong
agery is considered to be one form of a spatial surface, then Statistical Bureau, 1999), Guangzhou is the sixth most
the complexity of the spatial surface should be apt for de- populous city in China. It has been the most important
scription and measurement by a fractal model (Lam, 1990), political, economic, and cultural center in southern China.
particularly by self-affine fractals. Surface radiant emittance, The Guangzhou municipality comprises eight administra-
as recorded by thermal infrared sensors, includes both topo- tive districts in the city proper, and four rural counties
graphically and non-topographically induced high frequency (i.e., Huadu, Zengchen, Panyu, and Conghua). This study
variations such as roads and edges caused by different spec- focuses on the city proper of Guangzhou.
tral characteristics of different neighboring land covers Guangzhou has a history of urban development
(Lam, 1990). The spatial surfaces generated from thermal in- stretching across 2,000 years (Xu, 1990). It originated as
frared image data therefore have a fractal characteristic that a small village about 300 B.C. The city wall was first
mixes topographic and non-topographic frequencies. The built in the third century, after China was united in the
use of fractals for analyzing thermal infrared images will Qin Dynasty. During the Sui, Tang, and Song Dynasties,
improve our understanding of the thermal behavior of dif- Guangzhou developed into the largest seaport in the na-
ferent land-cover types and the effect of landscape pattern tion as international sea going trade developed rapidly.
on thermal environmental processes. The establishment of With economic development and population increase, the
the relationships among land use, land cover, and radiant city wall was expanded several times toward the east and
temperature makes it possible for a better scientific under- the north. During the Qing Dynasty, the city was further
standing of how the encroachment of urban and related uses expanded to the south, approaching the north bank of the
on the natural environment have motivated the generation Zhujiang. Following the establishment of the People’s
and growth pattern of UHIs. Moreover, because remote sens- Republic of China in 1949, Guangzhou’s locational ad-
ing analysis of UHIs often involve multiple sensors, it will vantages became more prominent, especially with the
be critical to know if fractal dimension is predictable with ending of the Vietnam conflict and the rising prosperity
changes in spatial and spectral resolutions. of Southeast Asia. Its urban development has progressed at
Previous studies have used fractals as a spatial mea- an unprecedented pace since the implementation of the
sure for describing and analyzing remotely sensed imagery economic reforms after 1978, with eastward and north-
(De Cola, 1989; Lam, 1990; Emerson et al. 1999; Qiu et al. ward expansion replacing the traditional southern and
1999). De Cola (1989) applied fractal analysis to describe western expansion.
the complexity of eight land-cover types in northwest
Vermont, while Lam (1990) compared the fractal dimension
of three land-cover types in the coastal Louisiana. Both Data Acquisition and Processing
studies employed Landsat TM images. Emerson et al. (1999)
further studied the complexity of Normalized Difference Computation of Surface Radiant Temperatures
Vegetation Index (NDVI) images derived from Landsat TM The data used in this study were three Landsat Thematic
images, and examined the effect of changing pixel size on Mapper (TM) images, dated on 13 December 1989, 03 March
fractal dimension. Qiu and his team (Qiu et al., 1999) used 1996, and 29 August 1997, respectively. Each Landsat
two AVIRIS (Airborne Visible Infra-Red Imaging Spectrome- image was rectified to a common UTM coordinate system
ter) images of the Los Angeles area to compare the image based on 1:50,000-scale topographic maps. These data were
complexity across the full spectral range of the hyperspec- resampled using the nearest-neighbor algorithm, with a
tral images (224 bands). It was found that urban areas have pixel size of 30 m by 30 m for all bands including the ther-
higher fractal dimension values than rural areas, and the mal band. The resultant root-mean-squared error was found
differences become more distinct in the visible bands. Most to be 0.77 pixel (23.1 m on the ground) for the 1989 image,
recently, Lam and her colleagues (Lam et al., 2002) evalu- and 0.58 pixel (17.4 m on the ground) for the 1996 and the
ated three fractal measures (i.e., isarithm, variogram, and a 1997 images.
modified version of triangular prism), and compared them Surface radiant temperatures were derived from geo-
with Moran’s I and Geary’s C. It was concluded that both metrically corrected TM thermal infrared data (band 6).
the isarithm and triangular prism methods could accurately The TM thermal band has a spatial resolution of 120 me-
measure a range of fractal surfaces, and that the variogram ters and a noise level equivalent to a temperature differ-
method was a relatively poor estimator particularly with ence of 0.5°C (Gibbons and Wukelic, 1989). The local
high fractal dimension surfaces. In spite of these significant time of satellite overpass was in the morning (approxi-
contributions, no substantial research has yet attempted to mately 1000 hours), so that the chance for detecting a
study a UHI pattern using fractal analysis, particularly to weak UHI is maximized. Because all the three images
detect the temporal changes of the UHI pattern using multi- were acquired at approximately the same time, a com-
date Landsat TM thermal infrared data. The objectives of parative study is feasible. According to Professor Yang
this paper are (1) to detect and analyze the spatial patterns at South China Normal University, who has been study-
of urban heat islands derived from satellite imagery, (2) to ing the UHI of Guangzhou for decades using weather sta-
examine the fractal characteristics of the urban heat island tion and in situ data, air temperature difference between
patterns, and (3) to model the UHI impact of urban develop- the urban and the rural peaks at 2100 hours, and goes
ment. While the focus of this study is to analyze satellite- down gradually until reaching a minimum at 1400 hours
detected UHI patterns, the results should also be of value (Yang et al., 1984). Although the impact of the diurnal
for environmental management and urban planning within heating cycle on the UHIs will be an interesting issue to
the study area. address, there has been no attempt to include it here
because TM does not provide day and night infrared
images at the same day. Moreover, because absolute
The Study Area: Guangzhou, China temperatures are not used for the purpose of computa-
Guangzhou (also known as Canton) is located at latitude tion, atmospheric correction was not carried out at this
23°08 N and longitude 113°17 E and lies at the conflu- stage. A quadratic model was used to convert the digital

556 May 2003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

 6_01-124 4/4/03 10:49 PM Page 557

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

May 2003
Figure 1. A map of the study area. The location, topography, and urban core area are shown as inset maps. W-E, S-N, and SW-NE profiles
(transaction lines) for Figure 4 are indicated in (c).

557
6_01-124 4/4/03 10:49 PM Page 558

number (DN) into radiant temperatures (Malaret et al., 1985): The resultant land-use and land-cover maps for 1989,
i.e., 1996, and 1997 are displayed in Plate 1. The overall accu-
racy of classification was determined to be 90.57 percent,
T(k)  209.831  0.834DN  0.00133 DN2 (1) 87.96 percent, and 85.43 percent, respectively. From these
maps, it is clear that there have been considerable changes
However, the temperature values obtained above are refer-
enced to a black body. Therefore, corrections for emissivity
(â) became necessary according to the nature of land cover.
Vegetated areas were given a value of 0.95 and non-vege-
tated areas 0.92 (Nichol, 1994). The emissivity corrected
surface temperatures were computed as follows (Artis and
Carnhan, 1982):

T(k)
Ts  (2)
1  (l* T(k) / r) lne
where l is the wavelength of emitted radiance (for which
the peak response and the average of the limiting wave-
lengths (l  11.5 mm) (Markham and Barker, 1985) will be
used), r  h * c/s (1.438 * 102 mK), s is the Stefan Bolz-
mann’s constant (5.67 * 108 W m2 K4  1.38 * 1023 J/K),
h is Planck’s constant (6.626 * 10(34 J sec), and c is the ve-
locity of light (2.998 * 108 m/sec).
The normalized difference vegetation index (NDVI)
image was computed for each year from visible (0.63 to
0.69 mm) and near-infrared (0.76 to 0.90 mm) data of TM im-
ages. The resultant NDVI image was thresholded into two
categories to represent vegetated and non-vegetated areas.
The thresholds were determined interactively by overlay-
ing and comparing the NDVI images with false color com-
posites. Equation 2 was then implemented with the aid of
the Spatial Modeler algorithm in ERDAS Imagine.

Derivation of Land-Use and Land-Cover Data

Land-use and land-cover patterns for 1989, 1996, and 1997
were mapped using Landsat TM data. A modified version of
the Anderson scheme of land-use/land-cover classification
was adopted (Anderson et al., 1976). The categories in-
clude (1) urban or built-up land, (2) barren land, (3) crop-
land (rice), (4) horticulture farms (primarily fruit trees),
(5) dike-pond land, (6) forest, and (7) water.
At this stage, detection of fine changes in surface re-
flectances was required, and radiometric correction became
necessary. Considering that no ancillary data on the atmos-
pheric conditions during the satellite overpasses were avail-
able to account for atmospheric differences, a relative radio-
metric correction method using image regression (Jensen,
1996) was employed. The brightness value of each pixel of
the subject scene (the 1989 and the 1996 scenes) was related
to that of the reference image (the 1997 scene) band by band
to produce a linear regression equation. This image normal-
ization method can minimize or eliminate the effects caused
by using historical remotely sensed images of non-anniver-
sary dates (Jensen, 1996).
A supervised signature extraction with the maximum-
likelihood algorithm was employed to classify the Landsat
images. Both statistical and graphical analyses of feature
selection were conducted. Bands 2 (green), 3 (red), and
4 (near infrared) were found to be most effective in dis-
criminating each class, and were therefore used for classi-
fication. The feature selection process reduced the number
of bands to be processed in the database, but should not
affect the classification accuracy (Jensen, 1996). The accu-
racy of the classification was verified by field checking
or by comparing with large-scale aerial photos and exist-
ing land-use and land-cover maps that have been field-
checked. The urban and built-up lands were extracted Plate 1. Land-use and land-cover maps for 1989, 1996,
from each land-use and land-cover image to create three and 1997. The boundary of districts is superimposed.
urban maps.

558 May 2003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

6_01-124 4/4/03 10:49 PM Page 559

in land use and land cover in Guangzhou during the 8-year The power spectrum of a series of elevation measure-
period. The urban and built-up area was 194.8 km2 in 1989 ments quantifies how its overall variance is partitioned
and 295.2 km2 in 1997, indicating an increase of 100.4 km2 in a corresponding series of harmonic components whose
(by 51.5 percent). wavelength varies between the length of a profile and
twice the sampling interval (Priestly, 1981). The divider
Data Analysis method (used in this paper), also known as ruler or com-
pass method, is implemented by using different “step
Urban Heat Island Detection and Analysis sizes” that represent the segments necessary to traverse a
To examine the spatial distribution of surface radiant tem- curve. For a comprehensive review of fractals, two texts
peratures, each surface radiant temperature image was re- on fractals (Mandelbrot, 1983; Feber, 1988) and review
classified into four temperature zones based on the classifi- articles in the field of Earth sciences (Brown, 1995; Mal-
cation scheme of standard deviation (Smith, 1986). Four inverno, 1995) are recommended.
classes were differentiated based on the mean and standard The triangular prism method has also been used to
deviation of the data distribution. Class one data are more calculate fractal dimensions of remotely sensed imagery
than one standard deviation below the mean, class two val- (Clarke, 1986; Jaggi et al.1993; Qiu et al.1999). This method
ues fall between the mean and one standard deviation calculates the surface areas, defined by triangular prisms,
below the mean, and so forth. The high and low tempera- using various sizes of measuring grids or cells. The four
ture zones in each year were thus identified. corners of each grid or cell on an image surface are denoted
Based on the spatial patterns of radiant temperatures, by pixel values of the four adjacent pixels. Each pixel value
three profiles (Figure 1b) were drawn to analyze the distri- is interpreted as a “z” value in creating a three-dimensional
bution of the urban heat islands. The west-east profile cube. Assigning the mean value of the four pixels as its “z”
starts with the city’s western boundary, from the Zhujiang value would generate the vertical line in the center of the
(the Pearl River), through the Liwan, Yuexiu, Dongshan, cell. This creates a prism with four triangular facets. The
and Tianhe Districts, and ends at the Nangang Township of surface area of this particular cell is obtained by summing
the Huangpu District. The line basically goes along Zhong- those of the four triangular facets based on the trigonomet-
shan Road in the urban core area, and passes Tianhe Park ric formula (Jaggi et al.1993).
and other rural areas in the suburb. Because the eastward With the divider method, the total length of a curve (a)
expansion has been the main direction of urban develop- is plotted against the step size (r) in a log-log form, and the
ment after the 1980s, an analysis of surface radiant temper- fractal dimension D can be estimated from the slope of the
ature along the west-east profile is conducive to the exami- resulting curve (B): i.e.,
nation of urban growth’s impact on the thermal behavior of
the environment. Log a  C  B * log r (3)
The south-north profile sets out at the Fangcun Dis-
trict, passes through the Haizhu and Yuexiu Districts, D  1B (4)
and arrives at the Renhe Township of the Baiyun District. Figure 2 illustrates a self-affine fractal with a nominal
This line matches Jiefan Road in Yuexiu District, and length l0 digitized at discrete intervals of r and a standard
also passes through the industrial area in the Haizhu Dis-
trict, Baiyun International Airport, and the newly devel-
oped areas around Xinshi Township. Therefore, an in-
quiry into this profile improves our understanding of
how industrial development and infrastructure construc-
tion have altered the city’s thermal characteristics.
The southwest-northeast profile begins at the Fangcun
District. After successively passing through the Liwan,
Yuexiu, and Tianhe Districts, it ends at the Jiufo Township
of the Baiyun District. On the way to Jiufo, the profile also
meets with Baiyun Hill, as well as a rapidly urbanized sub-
urb at the base of Baiyun Hill, namely, Jinxi Village. Be-
cause it crosses over both the urban core area and the vast
rural area in the northwestern part of the city, this profile
contributes to our understanding of the urban-rural thermal
difference.
Because the profiles cross the Pearl River, some ideas
of the thermal difference between water bodies and terres-
trial systems may be obtained.

Estimating the Fractal Dimension of Profiles

The fractal dimension of the spatial surfaces and profiles of
remotely sensed data may be estimated using one of four
main methods: (1) the box-counting method, (2) the spectral
method, (3) the divider method, and (4) the triangular prism
method. The box-counting method is implemented by lay-
ing a regular grid of boxes of characteristic size over the
profile and counting the number of boxes intersected by the Figure 2. Estimation of fractal dimension using the
profile. The process is repeated for different sizes of box, divider method. In the figure, d is the standard deviation
and the number of boxes filled is plotted against the total of surface height (average deviation from the mean) over
number of boxes in the grid in a log-log form. If linear, the a length scale r, and l0 is the total profile length (Brown,
slope of this curve is related to the fractal dimension. The 1995).
spectral method is based on the power spectrum.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2003 559

6_01-124 4/4/03 10:49 PM Page 560

deviation of heights equal to d. Then the total length of line in the same year. A mean temperature value (Tm) was then
a as a function of r is approximately calculated for each land-cover type by averaging the Ts val-
ues of the three years. Next, the land-use and land-cover
1 r2  d2 2
l0 1/2 maps were converted into the maps of simulated radiant
a r
. (5) emittance by assigning Tm values to each land-cover class.
Three profiles (W-E, S-N, and SW-NE) were drawn over
For a self-affine fractal over distance r, the standard devia- these simulated maps, and the fractal dimensions of the pro-
tion of heights is files were calculated. By comparing the fractal dimension
2D
values in 1989, 1996, and 1997, one can see the impact of
d  ba b
r land use and land cover changes in general and urban devel-
(6)
b opment in particular on the growth patterns of the UHI.

where b is the crossover length. It is interpreted as the Results

horizontal sampling interval, above which the divider
method breaks down. To obtain a meaningful result from Spatial Patterns of the Urban Heat Islands
the divider method, data must be digitized at a scale much The statistics of surface radiant temperatures of each image
smaller than the crossover length (Brown, 1995). Replace d by year are summarized in Table 1. On 13 December 1989,
with the above expression; then Equation 5 becomes the lowest radiant temperature was 16.98°C, the highest
temperature was 25.23°C, and the mean was 21.17°C, with
1/2
2(1D) a standard deviation of 1.72°C. A choropleth map (Figure
a  l0 £ 1  a b § .
r 3) was produced to show the spatial distribution of the
(7)
b surface radiant temperatures based on the classification
scheme of standard deviation. It is clear from the map that
For r V b, log (a) versus log (r) has slope of 1 2 D. When all the urban or built-up areas have a relatively high tem-
r W b, then a  l0. In this case, calculating D by using the perature, ranging from 21.17°C So 25.23°C. Some hot spots,
slope of the r-a curve gives D  1. or urban heat islands, could be identified. The most exten-
The algorithm used in this study was originally devel- sive UHI was distributed in the western part of the Haizhu
oped by Goodchild (1980) and Shelberg et al.(1983), and District, an industrial region in the city. Another noticeable
elaborated in Lam (1990), Lam and De Cola (1993), and Jaggi UHI was detected in the southeast corner of the city proper,
et al. (1993). This algorithm measures the fractal dimension where the Guangzhou Economical and Technological De-
of isarithm lines characterizing a spatial surface, and aver- velopment Zone (GETDZ) is located. Numerous factories
ages the D values of all lines as the surface’s final fractal di- have found their homes here after 1978, when the Chinese
mension. The isarithm lines of a remotely sensed image are government initiated a series of preferential policies to spe-
generated by dividing the range of the pixel values of the cial regions in order to attract foreign investment. There
image into equally spaced intervals. The length of each is- were also many small UHIs throughout the Tianhe District,
arithm, as represented by the number of edges, is then mea- which are related to the sparsely distributed industries in
sured at various step sizes. The logarithm of the number of the region. However, there did not exist an extensive UHI in
edges is regressed against the logarithm of the step sizes. the old urban areas such as Liwan, Yuexiu, and Dongshan
The slope of the regression is obtained using a best-fit linear Districts, in spite of their high construction density. Appar-
model, and thus a D value for the isarithm. A computer pro- ently, commercial and residential areas are less effective in
gram was written for implementing the procedure. promoting the development of UHI.
The lowest temperature class (16.98 to 19.45°C) ap-
Modeling Urban Development Impact on the UHI peared in the following three areas: (1) the eastern part of
Given the relationship between surface radiant temperature Baiyun District around Maofeng Mountain; (2) Baiyun Hill,
and the texture of land cover that is influenced by land use, and (3) the southeastern part of Haizhu District. These areas
changes in land use and land cover (e.g., urban develop- were substantially rural at one time, and mostly covered
ment) can have a profound effect on the surface radiant tem- by forest. Both the northwest Baiyun District and the west
perature and the UHI. The thermal signature of each land- Fangcun District have a moderate temperature ranging
cover type was obtained by overlaying a surface radiant from 19.45°C to 21.17°C, where cropland and dike-pond
temperature (Ts) image with a land-use and land-cover map land prevailed.

TABLE 1. SUMMARY STATISTICS OF THE SURFACE RADIANT TEMPERATURES (UNIT: ºC)

Coefficient
Standard of
Date Specification Mean Minimum Maximum Range Deviation Variation
13 Dec 1989 Overall TM Scene 21.17 16.98 25.23 8.25 1.72 0.08
W-E Profile 22.39 18.01 25.23 7.22 1.36 0.06
S-N Profile 21.61 17.67 25.23 7.56 1.26 0.06
SW-NE Profile 20.88 16.98 25.23 8.24 1.93 0.09
03 Mar 1996 Overall TM Scene 23.12 11.38 34.03 22.65 1.64 0.07
W-E Profile 23.84 15.49 30.09 14.60 1.29 0.05
S-N Profile 23.88 18.35 27.54 9.19 1.19 0.05
SW-NE Profile 22.51 18.80 27.02 8.22 1.60 0.07
29 Aug 1997 Overall TM Scene 31.93 27.62 39.62 12.00 2.54 0.08
W-E Profile 34.22 28.80 39.17 10.56 2.34 0.07
S-N Profile 33.57 28.11 38.26 10.15 2.83 0.08
SW-NE Profile 31.20 27.12 37.80 10.68 2.83 0.09

560 May 2003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

6_01-124 4/4/03 10:49 PM Page 561

state of vegetation, and atmospheric influences of remotely

sensed TM data set, but also changes in land use and land
cover. The distinctive thermal characteristics of land-cover
classes can affect the development of urban heat islands
(Lo et al., 1997; Weng, 2001). The mean radiant temperature
of 03 March 1996 was 23.12°C, with a standard deviation
of 1.64°C. However, a much higher range was observed for
the date (22.65°C), varying from 11.38°C (minimum) to
34.03°C (maximum). While some of the urban or built-
up areas fell into the highest temperature class (24.76 to
34.03°C), many others exhibited a moderate temperature
around the mean. This indicates that urban surfaces experi-
enced a much wider variation in radiant temperature. The
thermal difference between the urban and the rural areas
had been widened, reaching at 10°C, as compared to ap-
proximately 4°C on 13 December 1989. These different pat-
terns are primarily attributed to a seasonal variation in
solar illumination and the state of vegetation. The 1989
image was taken near winter solstice while the 1996 image
around the spring equinox. The difference in data acquisi-
tion season was also evidenced in the radiant tempera-
tures of water bodies. The radiant temperature of water
bodies was in the intermediate temperature Class 3 (19.45
to 21.17°C) in 1989, but in the lowest temperature Class 4
(11.38 to 21.48°C) in 1996.
There were noteworthy UHIs (Figure 3), including one
at west Haizhu, one at GETDZ of Huangpu District, and
several others in the Tianhe District. These heat islands,
however, were not spread out as extensively as those in
1989. The most significant difference between the two
maps is seen in the northwest part of the Baiyun District.
Urban development along the northward highways, partic-
ularly along the Guangzhou-Huaxian Highway, had created
several visible urban heat strips. The conversion of crop-
land to other uses (e.g., to horticulture farms and to dike-
pond land) has also altered the spatial patterns of surface
radiant temperature, because cropland registered a lower
temperature (mean  23.26°C) than horticultural farms
(23.61°C) but higher than dike-pond land (22.63°C).
The 1997 image (29 August) was taken in the hottest
month. The average temperature was 31.93°C, with a range
between 27.62°C and 39.62°C. The standard deviation was
also the largest (2.54°C) among the three years, indicating
that the surfaces experienced a wider variation in radiant
temperature. The urban and rural areas can easily be dis-
tinguished from Figure 3. The urban areas showed a high
temperature of over 34.47°C, while the rural settlements a
minimal temperature of 31.93°C. The UHI expanded east-
ward from the urban core areas of Liwan, Yuexiu, and
Dongshan to Huangpu, forming a heat island corridor. The
UHI to the south of the Pearl River seemed to stretch out
from the western Haizhu to the Fangcun District. Two large
UHIs emerged, one centered at the Guangzhou Railroad Sta-
tion and Baiyun International Airport, and the other to the
east of Baiyun Hill. Both areas had been under a rapid
urban sprawl since the 1990s. In additions, numerous strip-
shaped UHIs have developed along the Guangzhou-Huax-
ian, Guangzhou-Huandong, and Guangzhou-Conghua High-
Figure 3. Spatial distribution of surface radiant temper- ways. During the summer time, Baiyun Hill seemed to be
ature in 1989, 1996, and 1997. The boundary of dis- the only place where urban residents can escape from the
tricts is superimposed. W-E, S-N, and SW-NE profiles heat owing to its dense vegetated cover.
(transaction lines) for Figure 4 are indicated.
Fractal Characteristics of the uhi Patterns
Figure 4 displays the W-E, S-N, and SW-NE temperature
profiles for each year. From the city’s core urban area in
The spatial patterns of surface radiant temperature on the southwest outward, there exist numerous “peaks,” “val-
03 March 1996 were markedly different from those of 13 leys,” “plateaus,” and “basins,” indicating the heteroge-
December 1989, as seen from Figure 3. These differences neous nature of radiant temperature over the space. Factors
reflected not only the differences in solar illumination, the such as the spatial pattern of different land-cover classes,

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2003 561

 562
May 2003
6_01-124 4/4/03 10:49 PM Page 562

Figure 4. W-E, S-N, and SW-NE profiles across surface radiance temperature maps in Figure 3. Place names along the profiles are annotated in Figure 1b.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

6_01-124 4/4/03 10:49 PM Page 563

the occurrence of water bodies and parks, building and profile was the greatest (Figures 4c, 4f, and 4i). Table 1
population densities, and the division of the city’s func- shows that the temperature difference between the urban
tional districts, among many others, may have affected the and the rural area along this transect may reach 8.22°C to
development of UHIs. By visually comparing the profiles 10.68°C, depending on the season.
among the three rows of Figure 4, the seasonal variation of The relatively low values of fractal dimension for
the UHI pattern becomes visible. Table 1 indicates that the the W-E profiles suggest that the texture is less spatially
mean value of surface radiant temperature varies from pro- complex, i.e., the spectral responses to the thermal band
file to profile. The W-E (1989 scene: 22.39°C; 1996 scene: along the line tend not to vary drastically (Figures 4a, 4d,
23.84°C; 1997 scene: 34.22°C) and S-N profiles (1989 scene: and 4g). This is particularly true from the Zhujiang to the
21.61°C; 1996 scene: 23.88°C; 1997 scene: 33.57°C) gener- Tianhe District, where urban or built-up cover has occu-
ally have a higher value than the SW-NE profile (1989 pied the majority of the surface. A computation of coeffi-
scene: 20.88°C; 1996 scene: 22.51°C; 1997 scene: 31.20°C). cients of variation indicates that the data variability of ra-
This indicates that the W-E and S-N profiles cross more land diant temperatures is small in all the three years (0.06,
cover types that possess higher thermal signatures (such as 0.05, and 0.07, respectively). A closer look at Figure 4a,
urban or built-up) than the SW-NE profile. Column 7 of 4d, and 4g, however, reveals that it has a coarser texture
Table 1 shows that the standard deviation value of surface from Tianhe Park eastward, where different types of land
radiant temperature varies from 1.19°C to 2.83°C among the cover intermingle within a short distance.
nine profiles. Overall, the profiles of 03 March 1996 have The three S-N profiles were drawn to reveal different
the smallest standard deviation values (W-E profile: 1.29°C; thermal responses among industrial, commercial, residen-
S-N profile: 1.19° C; and SW-NE profile: 1.60°C), while those tial, airport, and rural areas. The resulting profiles for 13
of 29 August 1997 have the highest values (W-E profile: December 1989 (Figure 4b) and for 03 March 1996 (Figure
2.34°C; S-N profile: 2.83°C; and SW-NE profile: 2.83°C). It 4e) exhibit a smooth appearance, while the one for 29 Au-
appears that the effect of UHI became more prominent as gust 1997 (Figure 4h) has many peaks and pits, especially
the season changes from spring to summer. The calculated in the northern part of Baiyun District. The fractal dimen-
values of fractal dimension with their corresponding coeffi- sion values for these three profiles are 1.2256, 1.2088, and
cients of determination (r-squared, R2) confirm this observa- 1.5239, respectively, implying a fine texture for the winter
tion (Table 2). Spring possessed the lowest dimensions and the spring profiles, but a coarse texture for the summer
(W-E profile: 1.2327; S-N profile: 1.2088; and SW-NE pro- profile. However, water bodies such as the Zhujiang always
file: 1.3018), while summer possessed the highest dimen- create a “valley” regardless of the season, implying the
sions (W-E profile: 1.4484; S-N profile: 1.5239; and SW-NE importance of water bodies for damping the UHI effect. A
profile: 1.5219). Winter profiles exhibited a closer resem- careful examination of Figure 4h indicates that a “plateau”
blance to spring rather than to summer, with a bit higher occurs extending from the Zhujiang to Xinshi Township.
value of D than that of spring. (W-E profile: 1.2493; S-N This “plateau” apparently corresponds to the south-north
profile: 1.2256; and SW-NE profile: 1.3716). stretch of the UHI in the city, where residential, commer-
A comparison among the W-E, S-N, and SW-NE profiles cial, industrial, and infrastructure uses were predominant.
indicates that the SW-NE profiles yield the highest average Many newly developed areas in the northern part of Baiyun
dimension (mean  1.3984), followed by the S-N (mean  District in recent years have created a zone of “kotadesasi”
1.3194) and W-E profiles (mean  1.3103) (Table 2). This (meaning the “town-village” mixed zones) (McGee, 1989).
suggests that the SW-NE profiles are more spatially com- Whether the jig-jag appearance of the 1997 profile in this
plex in texture than others. In other words, a higher spec- portion is caused by this pattern of urban development
tral variability existed among the neighboring pixels be- remains to be examined.
cause of topographic changes along the line (see Figure 1b)
and the spatial arrangement and areal extent of different
Urban Development and the Urban Heat Islands
land-cover types. Indeed, the SW-NE profiles pass over all
In order to understand the impacts of land-use/land-cover
the seven land-cover types. Furthermore, a coefficient of
change on surface radiant temperature, the characteristics
variation (standard deviation/mean) is calculated, as an as-
of the thermal signatures of each land-cover type must be
patial statistical measure, to indicate the data variability of
studied first. The average values of radiant temperatures by
surface radiant temperatures. The SW-NE profiles also pos-
land-cover type are summarized in Table 3. It is clear that
sess the highest values, 0.09 for 1989, 0.07 for 1996, and
0.09 for 1997 (Table 1). This finding asserts that the spatial
variability of surface radiant temperature along the SW-NE TABLE 3. AVERAGE SURFACE RADIANT TEMPERATURE (ºC) BY LAND-COVER TYPE
13 Dec 1989 03 Mar 1996 29 Aug 1997
TABLE 2. FRACTAL DIMENSION VALUE AND R2 OF THE PROFILES (Standard (Standard (Standard
Land Cover Deviation) Deviation) Deviation) Average
Fractal Coefficient of
Date Profile Dimension Determination Urban or 22.50 23.95 34.75 27.07
Built-Up (1.23) (1.30) (2.09)
13 Dec 1989 W-E Profile 1.2493 0.8664
Land
S-N Profile 1.2256 0.8639
Barren Land 22.16 23.79 32.23 26.06
SW-NE Profile 1.3716 0.9026
(1.33) (1.38) (2.05)
03 Mar 1996 W-E Profile 1.2327 0.8587 Cropland 21.30 23.26 30.96 25.17
S-N Profile 1.2088 0.8549 (1.26) (1.37) (1.72)
SW-NE Profile 1.3018 0.8779 Horticulture 21.71 23.61 31.24 25.52
Farms (1.83) (1.52) (1.90)
29 Aug 1997 W-E Profile 1.4484 0.9228
Dike-pond 20.18 22.63 32.83 25.21
S-N Profile 1.5239 0.9429
Land (0.26) (1.62) (2.46)
SW-NE Profile 1.5219 0.9427
Forest 19.77 21.82 29.88 23.82
Mean W-E Profile 1.3103 (1.74) (1.40) (1.15)
S-N Profile 1.3194 Water 20.35 20.36 31.34 24.02
SW-NE Profile 1.3984 (0.72) (1.98) (1.45)

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2003 563

6_01-124 4/4/03 10:49 PM Page 564

urban or built-up land exhibited the highest radiant tem- That is to say, land-use and land-cover changes, especially
perature (22.50°C in 1989, 23.95°C in 1996, and 34.75°C in urban development, have caused a greater contrast in the
1997), followed by barren land (22.16°C in 1989, 23.79°C surface thermal properties in the S-N and W-E directions.
in 1996, and 32.94°C in 1997). This implies that urban de- The scope of the UHI has extended to the north and to the
velopment does bring up radiant temperature by replacing east direction as a result of urban development.
natural vegetation with non-evaporating, non-transpiring
surfaces such as stone, metal, and concrete. The standard
deviations of the radiant temperature values were relatively Discussions and Conclusions
small for urban cover, indicating that urban surfaces did not The utility of Landsat TM infrared data to detect urban heat
experience a wide variation in radiant temperature because islands in Guangzhou proved to be effective. Two major
of the dry nature of non-evapotransipirative materials. The heat islands were detected, one in the west Haizhu District
lowest radiant temperature in 1989 was observed in forest and one in the Guangzhou Economic and Technological
(19.77°C), followed by dike-pond land (20.18°C) and water Development Zone, and several others in the Tianhe Dis-
bodies (20.35°C). On 29 August 1997, the lowest radiant trict. The distribution of the UHIs was closely associated
temperature was found in forest (29.88°C), followed by with industrial land uses but not with residential land uses.
cropland (30.96°C) and horticultural farms (31.24°C). This The areal extent of the UHIs varied as the season changed.
pattern is in contrast with that in 1996, when the lowest ra- The profiles derived from the spring image had the lowest
diant temperature was detected in water (20.36°C), followed fractal dimension while from the summer image the high-
by forest (21.82°C) and dike-pond land (22.63°C). This dif- est value. Changes in fractal dimension in different seasons
ferent pattern is primarily attributed to the seasonal differ- were not only attributed to solar illumination and climato-
ences in solar illumination, the state of vegetation, and at- logic conditions relating to soil moisture and air tempera-
mospheric influences on the remotely sensed TM data set. ture, but also to the topographic variation and the spatial
Forests showed a considerably lower radiant temperature in arrangement and areal extent of different land cover types.
all years, because dense vegetation can reduce the amount Urban development increased the spatial variability of ra-
of heat stored in the soil and surface structures through diant temperatures, resulting in higher fractal dimension
transpiration. Water bodies tended to get warm slowly dur- values. The thermal surfaces have become more spatially
ing the day owning to convection. Cropland and horticul- uneven and the textures more complex.
ture farms had an intermediate level of radiant temperature, As has been pointed out by Lam (1990), the computa-
because they consisted of sparse vegetation and exposed tion results of fractal dimension are subject to the influ-
bare soil. ence of the method used. For the sake of simplicity, the
A mean temperature Tm value was calculated for all divider method was used in this study. The divider meth-
land-cover types by averaging the Ts values of the three ods give a correct D value for self-affine fractals only under
years. The land-use and land-cover maps of 1989, 1996, certain conditions (Mandelbrot, 1983), but always give a
and 1997 were converted into a map of mean temperature good approximation for self-similar fractals. Further stud-
by assigning Tm values to each land-cover type. The frac- ies should compare this method with the box and spectral
tal dimensions were calculated for the three profiles over methods. The spectral method is expected to be valid for
these mean temperature maps. Results indicate that all the self-affine fractals in the form of time or spatial series
three profiles have increased their D value over the time (Brown, 1995). The box method is also able to give a reli-
(Table 4). The W-E profile increased from 1.1761 in 1989 to able D value for self-affine fractals under certain circum-
1.1777 in 1996 and to 1.1922 in 1997; the S-N profile from stances. The triangular prism method can accurately mea-
1.1921 to 1.1937 and 1.2093; and the SW-NE profile from sure a range of fractal surfaces, but is sensitive to contrast
1.1902 to 1.1974 and 1.1984. The increase in D values im- stretching (Lam et al., 2002).
plies that the thermal surfaces became more spatially un- Many factors, such as solar illumination, the state of
even and the textures more complex due to the land-use vegetation, atmospheric effect, land-use and land-cover
and land-cover changes. Because the three profiles pass pattern, and topography, affect the fractal dimensionality
over both the urban and the rural areas, these results reflect from remotely sensed imagery. More research is needed to
well the impact of urban growth on the texture of thermal differentiate the impact of each element. Radar generated
surfaces. Urban development has increased the variability topography data (e.g., SRTM, lidar, IFSAR) may be desirable
of radiant temperatures over the space, resulting in higher to relate urban topography to UHI studies. Remotely sensed
fractal dimension values of the transects. Table 4 indicates data from different sensors should be acquired to examine
that, among the three profiles, the largest amount of change the effects of various spectral and spatial resolutions on
in D value occurs in the S-N profile (0.0172), followed by the resultant fractal dimensions (Lam, 1990). For example,
the W-E profile (0.0161), and the SW-NE profile (0.0082). application of fractals to data from the NASA Earth Obser-
vation System sensors (e.g., Landsat 7, ASTER, MODIS )
TABLE 4. FRACTICAL DIMENSION VALUE AND R2 FOR THE PROFILES
holds great potential for the analysis and characterization
OF SIMULATED TEMPERATURES
of urban thermal landscape, as well as for the planning
and development of more advanced thermal infrared
Fractal Coefficient of platforms.
Year Profile Dimension Determination It should also be noted that the computation of surface
1989 W-E Profile 1.1761 0.8324 radiant temperature can be far more complicated than the
S-N Profile 1.1921 0.8383 method presented in the current study. Quattrochi and Goel
SW-NE Profile 1.1902 0.8476 (1995) conducted a comprehensive review on major techni-
1996 W-E Profile 1.1777 0.8381 cal and theoretical difficulties in using thermal-infrared
S-N Profile 1.1937 0.8475 remotely sensed data obtained at different spatial and tem-
SW-NE Profile 1.1974 0.8521 poral scales. Factors such as the scale-dependent nature
of landscape characteristics, physiography, emissivity,
1997 W-E Profile 1.1922 0.8389
S-N Profile 1.2093 0.8417
atmospheric effects, and sensor-to-target noise have con-
SW-NE Profile 1.1984 0.8435 tributed to the difficulties in the determination of surface
temperatures (Quattrochi and Goel, 1995). Radiant temper-

564 May 2003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING

6_01-124 4/4/03 10:49 PM Page 565

atures may be higher than surface temperatures. Goetz Diablo Canyon, Photogrammetric Engineering & Remote Sens-
et al. (1993) found that Landsat 5 TM data consistently ing, 55(6):903–909.
overestimated surface temperatures by 1°C to 8°C (com- Goetz, S.J., F.G. Hall, B.L. Markham, and R.O. Dubayah, 1993.
pared with near-surface-level temperatures) due primarily Inter-comparison of retrieved surface temperature from multi-
to the sensor’s radiometric characteristics. The effects of resolution sensors at the FIFE site, Technical Papers of
ASPRS/ACSM 1993, 16–18 February, New Orleans, Louisiana,
surface roughness on surface temperature have not been pp. 108–117.
taken into account in this paper. Several authors (Kimes,
Goodchild, M.F., 1980. Fractals and the accuracy of geographical
1983; Cassels et al., 1992a; Cassels et al., 1992b) have measures, Journal of Mathematical Geology, 12(2):85–98.
elaborated on this issue, and suggest scrutinizing the tem-
Guangdong Statistical Bureau, 1999. Guangdong Statistical
peratures of each part of the vegetation-ground system and Yearbook, China Statistical Press, Beijing, China, 721 p. (in
examining the effects of different canopy structures. Effec- Chinese).
tive land surface temperature can be derived only after its Jaggi, S., D. Quattrochi, and N.S. Lam, 1993. Implementation of
relationship to the component temperatures have been operation of three fractal measurement algorithms for analysis
mathematically modeled. In addition, a more complicated of remote sensing data, Computers & Geosciences, 19(6):
emissivity correction scheme should be applied. Effective 745–767.
measurement of surface temperatures requires the analysis Jensen, J.R., 1996. Introductory Digital Image Processing: A Remote
of the significance of the nature of surface and its roughness Sensing Perspective, Second Edition, Prentice Hall, Upper
on emissivity. The emissivity values for urban areas are Saddle River, New Jersey, 316 p.
highly variable, and need to be refined in future studies. Kidder, S.Q., and H.T. Wu, 1987. A multispectral study of the
St. Louis area under snow-covered conditions using NOAA-7
AVHRR data, Remote Sensing of Environment, 22:159–172.
Acknowledgments Kim, H.H., 1992. Urban heat island, International Journal of Remote
The funded support of the National Geographic Society Sensing, 13(12):2319–2336.
(Grant number: 6811-00) is greatly acknowledged. The Kimes, D.J., 1983. Remote sensing of row crop structure and com-
comments and suggestions of three anonymous reviewers ponent temperatures using directional radiometric temperatures
helped improve this paper. and inversion techniques, Remote Sensing of Environment,
13:33–55.
Lam, N.S.N., 1990. Description and measurement of Landsat TM
References images using fractals, Photogrammetric Engineering & Remote
Anderson, J. R., E. E., Hardy, J. T., Roach, and R.E. Witmer, 1976. Sensing, 56(2):187–195.
A Land Use and Land Cover Classification Systems for Use Lam, N.S.N., and L. De Cola, 1993. Fractals in Geography, Prentice
with Remote Sensing Data, USGS Professional Paper 964, U.S. Hall, Englewood Cliffs, New Jersey, 308 p.
Government Printing Office, Washington, D.C., 27 p. Lam, N.S.N., H. Qiu, D.A. Quattrochi, and C.W. Emerson, 2002. An
Artis, D.A., and W.H. Carnahan, 1982. Survey of emissivity vari- evaluation of fractal methods for characterizing image com-
ability in thermography of urban areas, Remote Sensing of plexity, Cartography and Geographic Information Science,
Environment, 12:313–329. 29(1):25–35.
Balling, R.C., and S.W. Brazell, 1988. High resolution surface tem- Larson, R.C., and W.H. Carnahan, 1997. The influence of surface
perature patterns in a complex urban terrain, Photogrammetric characteristics on urban radiant temperatures, Geocarto Inter-
Engineering & Remote Sensing, 54:1289–1293. national, 12(3):5–16.
Brown, S.R., 1995. Measuring the dimension of self-affine fractals: Lo, C.P., D.A. Quattrochi, and J.C. Luvall, 1997. Application of
examples of rough surfaces, Fractals in the Earth Sciences high-resolution thermal infrared remote sensing and GIS to as-
(C.C. Barton and P.R. LaPointe, editors), Plenum Press, New sess the urban heat island effect, International Journal of Re-
York, N.Y., pp. 77–87. mote Sensing, 18:287–304.
Byrne, G.F., 1979. Remotely sensed land cover temperature and Malaret, E., L.A. Bartolucci, D.F. Lozano, P.E. Anuta, and C.D.
soil water status — A brief review, Remote Sensing of Environ- McGillem, 1985. Landsat-4 and Landsat-5 Thematic Mapper
ment, 8:291–305. data quality analysis, Photogrammetric Engineering & Remote
Sensing, 51:1407–1416.
Carnahan, W.H., and R.C. Larson, 1990. An analysis of an urban
heat sink, Remote Sensing of Environment, 33:65–71. Malinverno, A., 1995. Fractals and ocean floor topography: A re-
view and a model, Fractals in the Earth Sciences (C.C. Barton
Cassels, V., J.A. Sobrino, and C. Coll, 1992a. On the use of satellite and P.R. LaPointe, editors), Plenum Press, New York, N.Y.,
thermal data for determining evapotranspiration in partially pp. 107–130.
vegetated areas, International Journal of Remote Sensing,
13:2669–2682. Mandelbrot, B.B., 1983. The Fractal Geometry of Nature, W.H.
Freeman, San Francisco, California, 468 p.
———, 1992b. A physical model for interpreting the land surface
Markham, B.L., and J.K. Barker, 1985. Spectral characteristics of
temperature obtained by remote sensors over incomplete
the LANDSAT Thematic Mapper sensors, International Journal
canopies, Remote Sensing of Environment, 39:203–211.
of Remote Sensing, 6:697–716.
Clarke, K.C., 1986. Computation of the fractal dimension of topo-
McGee, T.G., 1989. Urbanisasi or kotadesasi? Evolving patterns of
graphic surfaces using the triangular prism surface area
urbanization in Asia, Urbanization in Asia: Spatial Dimen-
method, Computers and Geosciences, 12(5):713-722.
sions and Policy Issues (F. Coasta, A.K. Dutt, L.J.C. Ma and
De Cola, L., 1989. Fractal analysis of a classified Landsat scene, A.G. Nobel, editors), University of Hawaii Press, Honolulu,
Photogrammetric Engineering and Remote Sensing, Hawaii, pp. 93–108.
55(5):601–610. Nichol, J.E., 1994. A GIS-based approach to microclimate monitor-
Emerson, C.W., N.S.N. Lam, and D.A. Quattrochi, 1999. Multi- ing in Singapore’s high-rise housing estates, Photogrammetric
scale fractal analysis of image texture and pattern, Photogram- Engineering & Remote Sensing, 60:1225–1232.
metric Engineering & Remote Sensing, 65(1):51–61. Oke, T.R., 1979. Technical Note No. 169: Review of Urban Clima-
Feder, J., 1988. Fractals, Plenum Press, New York, N.Y., 283 p. tology, World Meteorological Organization, Geneva, Switzer-
Gallo, K.P., A.L. McNab, T.R. Karl, J.F. Brown, J.J. Hood, and J.D. land, 43 p.
Tarpley, 1993. The use of NOAA AVHRR data for assessment of the Priestly, M.B., 1981. Spectral Analysis and Time Series, Academic
urban heat island effect, Journal of Applied Meteorology, Press, London, United Kingdom, 890 p.
32:899–908. Qiu, H.L., N.S.N. Lam, D.A. Quattrochi, and J.A. Gamon, 1999.
Gibbons, D.E., and G.E. Wukelic, 1989. Application of Landsat Fractal characterization of hyperspectral imagery, Photogram-
Thematic Mapper data for coastal thermal plum analysis at metric Engineering & Remote Sensing, 65(1):63–71.

PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING May 2003 565

6_01-124 4/4/03 10:49 PM Page 566

Quattrochi, D.A., and N.S. Goel, 1995. Spatial and temporal scaling Shelberg, M.C., N.S.N. Lam, and H. Moellering, 1983. Measuring
of thermal remote sensing data, Remote Sensing Review, the fractal dimensions of surfaces, Proceedings of the Sixth In-
12:255–286. ternational Symposium on Automated Cartography (Auto-
Quattrochi, D.A., and J.C. Luvall, 1999. High Spatial Resolution Carto 6), 04–06 October, Ottawa, Canada, 2:319–328.
Airborne Multispectral Thermal Infrared Data to Support Smith, R.M., 1986. Comparing traditional methods for selecting
Analysis and Modeling Tasks in the EOS IDS Project class intervals on choropleth maps, Professional Geographer,
ATLANTA, URL: http://wwwghcc.msfc.nasa.gov/atlanta/, 38(1):62–67.
Global Hydrology and Climate Center, NASA, Huntsville, Weng, Q., 2001. A remote sensing-GIS evaluation of urban expan-
Alabama (date last accessed: 19 September 2002). sion and its impact on surface temperature in the Zhujiang
Quattrochi, D.A., and M.K. Ridd, 1994. Measurement and analysis Delta, China, International Journal of Remote Sensing,
of thermal energy responses from discrete urban surfaces using 22(10):1999–2014.
remote sensing data, International Journal of Remote Sensing, Xu, J., 1990. A Collection of Essays on Lingnan Historical Geogra-
15(10):1991–2022. phy (Lingnan Lishi Dili Lunji), Zhongshan University Press
———, 1998. Analysis of vegetation within a semi-arid urban (Zhongshan Daxue Xuebao Chubanshe), Guangzhou, China,
environment using high spatial resolution airborne thermal in- 408 p. (in Chinese).
frared remote sensing data, Atmospheric Environment, Yang, S., M. Zhang, and R. Zeng, 1984. The urban heat island ef-
32(1):19–33. fect of Guangzhou, Journal of South China Normal University,
Roth, M., T.R. Oke, and W.J. Emery, 1989. Satellite derived urban 1984(2):35–45 (in Chinese with English abstract).
heat islands from three coastal cities and the utilisation of
such data in urban climatology, International Journal of Re- (Received 12 October 2001; revised and accepted 10 September
mote Sensing, 10:1699–1720. 2002)

566 May 2003 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING