Landscape and Urban Planning 135 (2015) 1–10

Contents lists available at ScienceDirect

Landscape and Urban Planning

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

Research paper

Impacts of urban biophysical composition on land surface

temperature in urban heat island clusters
Guanhua Guo a,b,c , Zhifeng Wu d,∗ , Rongbo Xiao e , Yingbiao Chen d ,
Xiaonan Liu a,b,c , Xiaoshi Zhang d
a
Guangzhou Institute of Geochemistry, Chinese Academy of Sciences, Guangzhou 510640, China
b
University of Chinese Academy of Sciences, Beijing 100049, China
c
Guangdong Key Laboratory of Agricultural Environment Pollution Integrated Control, Guangdong Institute of Eco-environment and Soil Sciences,
Guangzhou 510650, China
d
School of Geographical Sciences, Guangzhou University, Guangzhou 510006, China
e
Guangdong Provincial Academy of Environment Science, Guangzhou 510045, China

h i g h l i g h t s

• We extracted urban heat island (UHI) clusters using an innovative approach.

• We examined the nonlinear relationships between the LST and urban biophysical composition in UHI clusters.
• The spatial heterogeneity of the landscape within UHI results in the complexity of LST.
• Both NDVI and NDBI are great indicators for predicting the variations of LST in UHI clusters.

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

Article history: The spatio-temporal pattern of biophysical composition significantly affects land surface temperature
Received 11 February 2014 (LST). Previous studies, however, mostly characterized urban heat island (UHI) clusters being spatially
Received in revised form 28 October 2014 homogeneous. The landscape spatial heterogeneity in urban across UHI clusters challenges us to more
Accepted 11 November 2014
accurately characterize the relationships between LST and corresponding urban biophysical composition.
In this study, we introduced an innovative integrated approach that combined object-oriented image
Keywords:
segmentation with local indicators of spatial autocorrelations (LISA) to extract UHI clusters from an LST
UHI clusters
image. We used a regression tree model to examine the nonlinear relationships between LST and each
Object-oriented segmentation
Nonlinear relationship
of three satellite-based indices within the UHI clusters: normalized differential vegetation index (NDVI),
Urban biophysical composition normalized differential build-up index (NDBI), and normalized difference bareness index (NDBaI). We
Land surface temperature found that both NDVI and NDBI are strongly correlated with the variations of LST whereas NDBaI has a
weaker correlation with LST. We also found that the regression tree model built in this study enabled us
to effectively detect the nonlinear relationship between LST and biophysical composition. Furthermore,
based on a set of rules derived from a regression tree analysis, we found that urban landscapes strongly
affect LST and its spatial heterogeneity within a UHI. These rules were used to detect the nonlinear impacts
of complex urban biophysical composition on LST. The results of this study provided insights into how
LST within UHI varies with urban surface characteristics at fine spatial scale and also a new method for
investigating effects of land surface composition on LST in urbanized areas.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction

Urbanization is considered to be one of the most evident aspects

∗ Corresponding author. Tel.: +86 13922448970.
E-mail addresses: gghgiser@qq.com (G. Guo), gzuwzf@163.com (Z. Wu),
ecoxiaorb@163.com (R. Xiao), 1064553130@qq.com (Y. Chen), 42447481@qq.com
(X. Liu), 285265827@qq.com (X. Zhang).
of human impacts on the earth in terms of greenhouse gas emis-
sions and land use changes (e.g., road and building construction),
both of which can in turn affect daily mean surface temperature
(Gallo & Owen, 1999; Owen, 1998). It is well documented that
urban centers experience higher temperature than surrounding

http://dx.doi.org/10.1016/j.landurbplan.2014.11.007
0169-2046/© 2014 Elsevier B.V. All rights reserved.
2 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10
suburban/rural areas, and this phenomenon is known as the urban
heat island (UHI) (Voogt & Oke, 2003). Traditionally, UHI was often
identified using in-situ weather station networks or mobile mea-
surements of the urban canopy layer (Li et al., 2011). With the
development of remote sensing (RS) technology, RS-based data
– such as surface radiative temperatures derived from remote
imagery thermal bands – are increasingly used to examine UHI.
RS-based technology enables us to thoroughly characterize LST dis-
tribution in urban areas and thus contribute us to explicit revealing
spatial patterns and variations of UHI although LST only provides a
snapshot of temperature during the day (Li et al., 2011). As a result,
RS-based LST have been widely used to examine the relationship
between UHI and surface biophysical parameters (e.g., Weng, 2009;
Arnfield, 2003).
Previous studies that focused on the relationship between urban
thermal behavior and landscape patterns indicated that LST is
strongly subject to both the composition and configuration of land
cover/use (Buyantuyev & Wu, 2010; Li et al., 2011; Weng, Liu, & Lu,
2007; Zhou, Huang, & Cadenasso, 2011). RS-based urban surface
biophysical composition – such as vegetation abundance, impervi-
ous surface and soil fractions – were found to be the best indicators
of LST variations across space (Deng & Wu, 2013; Xiao et al., 2007;
Zhang, Odeh, & Han, 2009). The relationship between biophysi-
cal composition and LST also varies with seasons and is thought
to be nonlinear (Chen, Zhao, Li, & Yin, 2006; Owen, Carlson, &
Gillies, 1998). Based on landscape metrics derived from different
LST zones, Liu and Weng (2008) found that the seasonal variations
in landscape pattern was correlated with LST. Buyantuyev and Wu
(2010) linked LST variations with changes in land covers as well
as socioeconomic features (i.e., family income, population density
and age). Nevertheless, the results from these studies were based on
urban-wide scale and ignored the spatial heterogeneity of compo-
sitions and configurations of urban landscapes at fine spatial scales.
The neglect of spatial heterogeneity of urban landscapes across
space is likely to disable these studies to accurately predict LST
Fig. 1. Geographic location of the Guangzhou Central Urban Districts.
variations through space and time. Given the spatial heterogeneity
and nonlinear interactions among urban ecological systems (Wu &
David, 2002) and their significance for urban landscape manage-
ment and planning, it is essential to improve our understanding of
the relationship between urban thermal behavior and its surface
biophysical composition at fine spatial scales.
Therefore, the objectives of this study are to (i) introduce an
integrated and object-oriented segmentation approach for extract-
ing the UHI clusters from LST image, (ii) quantify the relationship
between LST and each of three urban biophysical compositions,
and (iii) explore the spatial heterogeneity of LST in UHI clusters. To
achieve these ends, we selected the center of Guangzhou in China,
an area characterized by strong spatio-temporal variability in UHI,
as the study region. Specifically, we focused on UHI clusters in the
Central Urban Districts of Guangzhou City to investigate impacts
of three selected urban biophysical compositions (i.e., NDVI, NDBI,
and NDBaI) on LST in UHI clusters. The results from this study
will improve our understanding of how LST varies across urban
landscapes at fine spatial scale characterized by different surface
conditions.
2. The study site
Guangzhou is the capital of Guangdong province of China. It is
situated on the broad flat alluvial plain of the Pearl River Delta of
southern China, and has a total area of approximately 7434.40 km2
(Fig. 1). The climate condition in Guangzhou is characterized by
the typical subtropical monsoon maritime climate of Southern Asia,
with annual mean temperature about 23.2 ◦ C and maximum tem-
perature about 38.7 ◦ C (based on long-term climate records from
the Wushan weather station that is located at 23.17◦ N and 113.33◦ E
and operated by Guangzhou Meteorological Bureau). Guangzhou
has been one of the fastest urbanizing cities in China since the
1980s and its GDP reached 1.35 trillion RMB in 2012, a 10.5%
increase compared to that in 2011. The population in the core
 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10
Fig. 2. The variations in monthly mean air temperature and humidity in 2005. Cli-
mate data shown here were collected from Wushan weather station in Guangzhou.
Error bars indicate the standard deviation.
urban area of Guangzhou has increased from 4.09 million in 2000
to 7.72 million in 2010. Guangzhou and its Central Urban Districts,
therefore, have become the engine of urbanization in southern
China’s economy. The Central Urban Districts of Guangzhou con-
sists of 126 census tracts that cover approximately 1461.66 km2 and
encompasses diverse landscapes such as forested hills, farmland,
under-developing built-up areas, urban fringes, and high-density
built-up areas. The diversity and spatial heterogeneity of land sur-
face conditions make it an ideal study area for UHI research.
3. Methodology
3.1. Data and LST retrieval
A cloud-free Landsat TM image, acquired at approximately
10:43 am (Beijing time) on July 18, 2005 was utilized for this study.
This daytime image was selected because UHI are most remark-
able during the daytime of warm season when air temperatures
and humidity are both higher in this study area (Fig. 2). In addition,
UHI is generally stronger and spatially more variations during the
daytime (Arnfield, 2003; Nichol, Fung, Lam, & Wong, 2009; Roth,
Oke, & Emery, 1989).
The Landsat TM image used in this study has been pre-processed
using radiometric correction and geometric rectification, and thus
is of good quality. However, given that the spatial extent of the
study area is much smaller than that of the TM image, we further
processed the TM image as follows. First, the raw TM image was
masked with a boundary shapefile of the Central Urban Districts,
and then rectified with a universal transverse mercator (UTM) pro-
jection using a WGS84 datum and zone 49 UTM. Second, IKONOS
fusion imagery with a 1-m spatial resolution acquired in 2005 were
used to georeference the TM image based on 60 ground control
points. A nearest neighbor algorithm was used to resample the
rectified and georeferenced image of 120 m spatial resolution onto
30 m by 30 m grid points and this approach was applied to all bands
of the TM image. The re-gridding of the TM image aims to match the
spatial resolutions of RS indices derived from different bands. Third,
we calibrated the TM image through converting the digital num-
bers of (DNs) of all bands to at-satellite reflectance. The resultant
root mean squared errors were smaller than 0.5 pixel (15 m).
We calculated NDVI, NDBI and NDBaI for the study area based on
the TM image. These variables have proven useful in establishing
meaningful statistical associations with the UHI (Bechtel, Zakšek, &
Hoshyaripour, 2012; Chen et al., 2006; Li et al., 2011; Weng, Lu, &
Liang, 2006). A thermal band from Landsat TM was used to retrieve
the LST using a mono-windows algorithm (Qin, Karnieli, & Berliner,
3
2001a; Qin, Dall’Olmo, Karnieli, & Berliner, 2001b). The derived
LST was treated as an indicator of urban heat condition. Observed
surface air temperatures from the Wushan weather station dur-
ing the period when the TM image was acquired were collected to
determine the LST (Fig. 5A).
3.2. Object-oriented segmentation of the LST
Previous studies have attempted to carry out segmentation
process by simple density slicing based on pixel-by-pixel image
analysis, and assumed that one pixel is not related to its nearest
neighbors. Blaschke and Hay (2001) and Blaschke (2003,2010) rec-
ommended a method of “beyond pixels” during image analysis.
To analyze the spatial structure of LST, We used the object-
oriented image segmentation approach to build LST objects for
the study area. The segmentation processing of LST images
was accomplished using software known as eCognition Devel-
oper (http://www.ecognition.com/). Fig. 3 shows a comparison
of density-slicing and object-oriented-segmentation-derived LST
images. It is clear that the results generated with a density slicing
method will result in a so-called ‘salt and pepper effect’ phe-
nomenon (Fig. 3A) because this approach classifies LST levels based
on pixel-by-pixel values. In contrast, the LST based on objected-
oriented segmentation is spatially more homogeneous compared
that based on density-slicing. (Fig. 3B). The LST objects with geo-
graphic boundaries are relatively homogeneous and composed of
distinct groups of LST pixels. They can be recognized and trans-
ferred into meaningful objects for UHI management and mitigation
as well as for urban design and planning.
The modification of the segmented scale parameter will influ-
ence the consequent results in a way that higher values lead to
larger LST objects whereas smaller values yield smaller LST objects.
It is, however, often unable to determine the absolute correct scale
of segmentation in advance because different LST image charac-
teristics (i.e., ASTER and Landsat ETM+ thermal bands with various
spatial resolutions) require different segmentation scales. In many
cases, UHI processes emerge at different segmentation scales on
the same image due to the hierarchical network of image objects
and the urban landscape’s spatial structure (Burnett & Blaschke,
2003; Wu, 1999; Wu & Loucks, 1995). A visual comparison was
used to determine the appropriate scale parameter, and the results
are described in Section 4.1.
3.3. UHI clusters extraction
As a measurement of spatial associations, the Local Indicators of
Spatial Autocorrelations (LISA) statistics allow us to detect clusters
on the assumption that a spatial pattern is a non-random distribu-
tion of geospatial parameters (Anselin, 1995; Anselin, Syabri, & Kho,
2006). The LST objects that are located close to one another are more
likely to have similar characteristics and geographic homogene-
ity, compared to those located farther from each other. Therefore,
the application of LISA can be extended to UHI studies in urban
areas. This study adopted a local Moran’s I to determine the location
clusters of UHI after the segmentation of LST image, as mentioned
in Section 3.2. These potential UHI clusters allow for the evalua-
tion of the relationship between LST and related urban biophysical
composition at fine spatial scale.
The local Moran’s I is calculated as follows (Anselin, 1995;
Anselin et al., 2006):
(xi − x)  zi 
Ii = n ×  2
× wij (xi − x) =
mo
× wij zj , with :
i (xi
− x)
j j
Z
i
m0 = ,
n
i
4 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10
Fig. 3. Comparison of (A) density slicing and (B) object-oriented segmentation for a LST image.
where xi is the observation x, and x is the average across the units; n
is the total number of observation x; The spatial weights matrix Wij
was standardized by row to avoid scale dependence and to facilitate
the interpretation of the statistics (Anselin, 1995), and defined as
a local neighborhood with a threshold distance of 919 m around
each geographic unit and 9999 permutation tests by setting the
significance level as 0.01; Zi and ZZ are deviations from the mean. mo
is the average of total Zi deviations. The UHI clusters were obtained
using GeoDa software environment (Anselin et al., 2006) and shown
in Fig. 5B.
Four types of spatial patterns were defined on the basis of pre-
specified significances: high–high (H–H), low–low (L–L), low–high
(L–H), and high–low (H–L), a combination of the Local Moran’s I
(Anselin, 1995; Anselin et al., 2006) statistics calculated for the clus-
ter map. The specific meanings of the four patterns defined in this
study are described as follows:
H–H (hot spots): the LST object and its surroundings have a tem-
perature concentration level greatly higher than the average level,
suggesting UHI clusters from a central place.
L–L (cold spots): the temperature levels of both the LST object
and its surroundings are relatively lower than the average level of
all objects.
H–L (island): the LST object exhibits an obviously high con-
centration whereas the concentration of its neighboring objects is
lower than average.
L–H (Atoll): the LST object has a significantly low-level concen-
tration, while the concentration levels of its surrounding objects
are relatively high.
Furthermore, “non-significance” is also a type of pattern, indi-
cating no significant local spatial associations.
3.4. Statistical methods
A zonal analysis method with 250 × 250 m blocks was used to
evaluate the mean of LST and other three biophysical compositions.
Pearson’s correlation coefficients were first conducted to exam-
ine the strength of relationship between LST and each of the other
variables within UHI clusters. We further evaluated the mean LST
at 0.01 increments from the minimum and maximum NDVI, NDBI
and NDBaI, and plotted the relationships between them. Regression
tree was used to develop a model to investigate the nonlinear rela-
tionship between the LST and each of the biophysical composition.
Furthermore, a traditional multiple regression method was also
developed to exhibit the superiority of regression tree analysis dur-
ing LST prediction modeling. Based on various rules developed from
the regression tree model, the spatial distribution of rules in the
study area was also mapped. Therefore, a corresponding landscape
of these rules was identified from the higher spatial resolution of
IKONOS imagery, as well as from field surveys.
4. Results
4.1. Determination of the scale parameter for LST image
segmentation
According to the characteristics of the LST image, five layers of
LST objects were generated using the scale parameters of 2, 4, 6,
8 and 10. The resulting segment outputs were visually analyzed
by overlaying them with the high spatial resolution IKONOS image
(Fig. 4A–C list scale parameters of 6, 8 and 10). As shown in Fig. 4,
the scale parameter of 6 results in smaller LST objects. For exam-
ple, LiWan Lake Park was over-segmented into several LST objects
(Fig. 4A, the arrow). When the scale parameter increased to 10, the
boundary of built-up areas tended to be composed of various build-
ing densities with greater LST heterogeneity although LiWan Lake
Park keeps its integrity (Fig. 4C, the arrow). These results indicate
that the scale parameter of eight is the best option (Fig. 4B) because
it resulted in meaningful LST objects with ‘natural’ geographic
boundaries and avoided the production of scattered classification
resulting from over-segmentation. The total number of LST objects
in the study area amounted to 6776, with a mean temperature of
approximately 312.52 K and standard deviation of 2.56 K.
4.2. Correlation analysis of the LST with the NDVI, NDBI and
NDBaI
The Pearson’s correlation coefficients indicate that all of the
biophysical variables were significantly correlated with the LST
(Table 1). The NDBI and NDVI had stronger relationship with the
Table 1
Pearson’s correlation coefficients between the LST and the NDVI, NDBI and NDBaI.
LST NDVI NDBI NDBaI
LST 1.00
NDVI −0.67** 1.00
NDBI 0.66** 0.69** 1.00
NDBaI −0.11* 0.36** 0.29** 1.00
*
Correlation is significant at the 0.05 level (two-tailed).
**
Correlation is significant at the 0.01 level (two-tailed).
 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10 5

Fig. 4. The segmentation results of an LST image at various scales (from panel A to C, the scale was 6, 8, and 10, respectively).

LST than the NDBaI. The NDVI was negatively correlated with the
LST, whereas the NDBI had a positive effect on the LST. The relation-
ship between the LST and NDVI we observed in this study agreed
well with those reported by a number of previous studies (Gillies,
Kustas, & Humes, 1997; Weng, Lu, & Schubring, 2004; Yue, Xu,
Tan, & Xu, 2007). The NDBaI that represents soil distribution in
urban areas, also tended to have a negative relationship with LST
as reported by Chen et al. (2006). The unexpected negative rela-
tionship between LST and NDBaI may result from the complexity
of landscape in urban.
Previous studies have documented a “temperature edge” phe-
nomenon when investigating the relationship between the LST and
NDVI (Carlson, Capehart, & Gillies, 1995; Carlson & Traci Arthur,
2000; Gillies et al., 1997; Li et al., 2011). In this study, we found
that the scattergram of the LST and NDBI also clearly displayed
both a “warm edge” and “cool edge”, locating at the upper edge
and lower edge of the scattergram. This phenomenon also applies
to NDVI (Fig. 6A and B). The pixels with the same NDBI values are
warmer and drier near the warm edge compared to those near the
cool edge. Therefore, the potential impacts of the NDBI on the LST
needed further emphasizing during future UHI studies. Compared
to the scattergram of NDVI and NDBI, the shape of NDBaI’s scat-
tergram formed like a circle (Fig. 6C). This pattern may indicate a
random relationship between LST and NDBaI.
Although there is a negative relationship between NDVI and LST
and a positive relationship between NDBI and LST, the relation-
ship is subject to NDVI and NDBI levels, respectively. For example,
Figs. 6D and E indicate that the linear relationships between the
LST and NDVI/NDBI are strong but the relationships are not mono-
tonic across all NDVI values. For NDVI, the strong negative linear
relationship only exists when the NDVI values are approximately
greater than −0.06 (Fig. 6D), suggesting that the LST has a linear
relationship with NDVI when the urban surface is covered by veg-
etation. This finding was similar to that reported by Li et al. (2011)
and Yuan and Bauer (2007). Interestingly, we found that the lin-
ear relationship between the LST and NDBI no longer exist when
the NDBI values are greater than 0.2 or less than −0.5 (Fig. 6E).
A similar phenomena also applies to NDBaI (Fig. 6F), which is not
strongly correlated with the mean LST when its values are greater
than −0.2 or less than −0.45. This suggests that the relationship
between NDVI and LST or between NDBI and LST is not always linear
or can be nonlinear.

Fig. 5. The LST distribution (A) and spatial evolution of UHI clusters (B).
6 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10

Fig. 6. The LST (A–C) scattergrams and linear plots of mean LSTs (D–F) versus the NDVI, NDBI and NDBaI within the UHI clusters.

4.3. LST nonlinear modeling of UHI clusters the regression tree predicted LST variations better than the mul-
tiple linear regression as suggested by the higher R2 adj (0.61), the
Regression tree analysis was used to construct the LST nonlinear lower root mean square error (RMSE), and the higher correlation
model based on NDVI, NDBI and NDBaI. Table 2 shows the compari- coefficients (0.81, significant at the 0.01 level).
son of the LST predictions based on the regression tree analysis and The calculated R2 adj based on the regression tree analysis sug-
the traditional multiple linear regression method. It indicates that gested that approximately 61% of variations in the LST can be
explained jointly by the three biophysical compositions (Table 2).
Table 2 In addition, the regression tree analysis enabled us to differentiate
A comparison of the result from a regression tree analysis and a linear regression. the impacts of each of the three biophysical compositions on the
LST. The nonlinear associations between LST and each of the three
RTa LRb
biophysical compositions can be grouped into four categories as
R2 adj 0.61 0.54 follows:
RMSE 1.05 1.12
Correlation 0.81 0.73
a,b
Note that the RT and LR represent a regression tree and linear regression, respec- Rule 1: [108 cases, mean 316.275 K, range 310.481 K to 320.813 K, est
tively. err 1.3188 K]
 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10
Fig. 7. The predicted LST data based on different rules against the actual satellite-derived LST. The vertical black line represents a 1:1 relationship between the predicted and
actual LST values. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Rule: NDVI > −0.0592 and NDBI > 0.0216
Expression: LST = 313.8386 − 11.3NDBaI + 2.6NDBI − 0.6NDVI
Rule 2: [3885 cases, mean 314.8267 K, range 309.18 K to 320.36 K, est
err 0.8025 K]
Rule: NDBI ≤ 0.2106 and NDBaI > −0.3422
Expression: LST = 310.1869 + 21.7NDBI − 7.4NDBaI − 4.3NDVI
Rule 3: [2136 cases, mean 315.1658 K, range 309.907 K to 320.185 K,
est err 0.7864 K]
Rule: NDVI > −0.0592 and −0.4258 < NDBaI ≤ −0.3422
Expression: LST = 311.832 + 21.2NDBI − 3.7NDVI − 1.8NDBaI
Rule 4: [192 cases, mean 315.3750 K, range 310.555 K to 319.929 K,
est err 1.2039 K]
Rule: NDBI ≤ 0.2106 and NDBaI ≤ −0.4258
Expression: LST = 291.8386 + 50.2NDBI − 42.9NDBaI + 16.2NDVI
The relationships between the satellite-derived LST and the esti-
mated LST based on above four rules were examined (Fig. 7). The
results indicate that the variation in satellite-derived LST is strongly
correlated with each of the estimated LST based on the four rules,
respectively, as suggested by the corresponding coefficients that
range from 0.31 to 0.66. Among these four rules, the predictive
model based on Rule 2 was the best with an R2 adj of 0.66 (Fig. 7
A). This model can predict 61.47% of the total cases with a variance
of 0.8025 K. The number of cases in Rule 3 accounts for the second
proportion of the UHI clusters area size, with an R2 adj of 0.56 and a
variance of only 0.7864 K. It suggests a high level of homogeneity
existing within Rule 3 (Fig. 7B). Both Rules 1 and 4 can only correctly
predict relatively smaller cases. Rule 1 shows higher temperature
than the average value. However, the formula of Rule 1 has a weaker
capability to capture variations in the LST (R2 adj is 0.31). Although
Rule 4 produces relatively stronger heterogeneity when compared
to the other three rules with a variance of 1.2039 K, it only explains
approximately 55% of the variations in the LST within UHI clusters
(Fig. 7B).
To further determine if different rules represent different types
of urban surface characteristics, we mapped the spatial distribution
of rules using the same colors as in Fig. 7, and identified the specific
landscape structure characteristics of each rule using high spatial
resolution IKONOS imagery (Fig. 8). We found that the spatial pat-
tern of Rule 2 (yellow) tends to concentrate on the outlying areas
of the urban villages. For instance, the Central Business District
(CBD) of Zhujiang New Town in the Tianhe district was associated
with relatively abundant vegetation and open space. In contrast,
the distribution of Rule 3 was clearly similar to the old downtown
of Guangzhou’s core area, such as the whole areas of the LiWan and
7
YueXiu districts, and the western part of the Haizhu district, and
the eastern part of the Fangcun district along the Pearl River. Rule 4
accounted for a smaller area compared to Rules 2 and 3, and it was
distributed mainly in the Fangcun and Huangpu districts near the
Pearl River.
The contribution of built-up areas to the LST within UHI clus-
ters strongly depends on its morphology and vegetation conditions;
the former influences the aerodynamics, and the latter affects the
transpiration (Li et al., 2011; Nichol, 1996, 1998). The landscape
characteristics in Rule 2 (with lower LST values; average tempera-
ture of 314.8267 K) mainly consist of newly developed commercial
buildings and institutions with relatively lower building density
and inner courtyards composed of grass or trees (Fig. 8B). Wind
speeds in these open spaces around tall buildings in the lower
boundary layer increases and thus cools the surrounding environ-
ment. In a densely-built low-rise environment (i.e., urban villages,
Rule 3), although the closeness of low residential blocks can pre-
clude solar radiation from heating the ground or streets directly,
the roofing on these residential blocks is warmer due to the pri-
marily dark concrete material that it is made of (Fig. 8A). As a
result, a remote sensor can only “see” this part rather than the
narrow alleyways (Nichol, 1996). Furthermore, urban villages in
Guangzhou’s core areas tend to be surrounded by high buildings
and are formed like a “basin” (Fig. 9), which could also lower the
wind speed and weaken the environmental cooling effect, espe-
cially at night. Compared to Rules 2 and 3, the landscape features
in Rule 4 are largely factories or industrial-residential complexes,
encompassing very low buildings. For instance, warehouses and
heavy industry buildings have very dark iron tiles and asphalt
rooftops, and little vegetation along the streets (Fig. 8C).
5. Discussions
5.1. Implications of UHI clusters
Compared to previous UHI-related studies that selected a whole
city or region as the study area (Buyantuyev & Wu, 2010; Cao,
Onishi, Chen, & Imura, 2010; Nichol, 1996; Zhou et al., 2011), we
specifically focused on UHI clusters and investigated the impacts
of urban biophysical composition on the magnitude of LST at rel-
atively fine spatial scale. We found that the combination of the
integrated method of object-oriented segmentation and LISA sig-
nificantly improved the efficiency in differentiating the LST levels
and keeping the integrity of LST objects (Fig. 2). The UHI clusters
we derived for the Guangzhou city included a group of meaningful
8 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10

Fig. 8. The spatial distribution of various rules developed by a regression tree analysis and corresponding urban surface landscape characteristics identified by IKONOS
imagery. Panels A to C represent urban villages, CBD and heavy industry, respectively.

real objects with generally higher LST temperature values dur-
ing daytime of warm season. The ‘natural’ UHI boundaries were
an antithesis to the view that the ecological variables in urban
areas are continuum of mosaics. It provides important insights into
urban ecology design and management. Therefore, our studies of
UHI clusters may have important implications for urban planners
and natural resource managers. It also is necessary for spatial plan-
ning of the urban environment as well as for many monitoring and
simulation UHI study programmers.
The UHI is a typical phenomenon in urbanized areas and
can occur across various spatial scales or along a scale contin-
uum similar to visible landscape mosaics. Many studies suggested

Fig. 9. The morphology of typical urban villages in Guangzhou. They tend to be surrounded by high buildings and formed like a “basin”.
 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10
that object-oriented segmentation could build a critical bridge
between the spatial concepts applied to multi-scale landscape
analyses and Geographic Information Science (Blaschke, 2010;
Hay, Marceau, Dube, & Bouchard, 2001; Wu, 1999). Burnett and
Blaschke (2003) considered objected-oriented segmentation as “A
multi-scale segmentation/object relationship modeling methodol-
ogy for landscape analysis.” This helps explain why object-oriented
segmentation has recently been widely used for explaining the
complexity and heterogeneity of multi-scale and hierarchical UHI.
5.2. Relationship between urban compositions and LST
The LST-NDVI relationship has been a heated research topic
because of the physical significance of vegetation to the LST
(Carlson et al., 1995; Gutman & Ignatov, 1998). Although there is a
strong negative relationship between NDVI and LST with a “temper-
ature edge” (Table 1 and Fig. 6A), LST is positively related to NDVI
at regional scales (Fig. 6D). A similar finding has been reported by Li
et al. (2011), who studied the whole city of Shanghai. Interestingly,
we found the LST - NDBI relationship also has a “temperature edge”
with an association almost as strong as that between LST and NDVI
(Table 1 and Fig. 6B). Considering the seasonal variability of NDVI,
NDBI can provide a complementary metric for LST studies.
The nonlinearity and seasonal variations in the urban LST-
characteristic (i.e., LST–NDVI, LST–NDBI) relationship have been
documented (Li et al., 2011; Weng et al., 2004; Yuan & Bauer,
2007; Zhang et al., 2009). This challenges us to predict LST varia-
tions, especially in spatially heterogeneous landscapes as our study
area. The use of regression trees to predict LST has proven to be
a greater improvement with an overall accuracy of 61%, higher
than 54% from multiple regressions (Table 2). Furthermore, differ-
ent rules obtained from regression trees enable us to interpret the
relationship between LST and landscape characteristics in a more
meaningful way. For example, higher temperatures in urban areas
are usually associated with large amounts of built-up areas with
higher NDBIs or impervious surface area, as well as little vegeta-
tion (Chen et al., 2006; Xiao et al., 2007). Oke (1976) reported that
the density of UHI has a closer relationship than a city’s size (as
measured by its population) to surface structure characteristics. In
this study, we found that landscape characteristics with various
rules varies significantly, indicating that the LST and landscapes in
UHI clusters are of high spatial heterogeneity. Different rules tende
to link to a particular landscape structure in urban areas (Fig. 8).
This further suggests that UHI mitigation and management poli-
cies should consider explicitly the particular landscape structure
(i.e., CBD or urban villages).
5.3. Limitations
The results from this study are likely to be subject to several limi-
tations. First, the scale of segmentation is one of the most significant
parameters for segmentation processing. In this study, the scale of
segmentation was empirically determined based on visual analysis
and comparison. Therefore, study that can optimize the selection of
the segmented scale for LST image is necessary for future studies.
Second, the timing of Landsat TM (morning) is not ideal for study-
ing the UHI because UHI often peaks later during a day. However,
since the limitations of remote sensing technology, the paucity of
thermal sensors that can produce thermal images with high spatial
resolution hinders us to detect LST variations at fine spatial scales
among different time of a day in urban (Chun & Guldmann, 2014).
Third, because only one set of horizontal daytime LST data were
used for this study, our study thus ignored the variation of LST in
the vertical direction within a UHI. Previous studies have demon-
strated that diurnal and seasonal variations exist between LST and
land surface characteristics (Buyantuyev & Wu, 2010; Yuan & Bauer,
9
2007; Zhang et al., 2009; Zhou, Qian, Li, Li, & Han, 2014). The ecolog-
ical context and different climatic conditions also can significantly
influence the amplitude of summer daytime UHI (Imhoff, Zhang,
Wolfe, & Bounoua, 2010). Therefore, further studies that address
the temporal and spatial variation of the relationships between UHI
and the corresponding land cover features are necessary.
6. Conclusions
The LST plays an essential role in affecting the exchange and
interaction of energy fluxes between the land surface and atmo-
sphere. Understanding the linkage between LST and urban surface
characteristics is important for designing effective measures to mit-
igate the amplitude of UHI. In this study, an innovative approach
that combined object-oriented segmentation and LISA statistics
was used to extract UHI clusters. Then, we quantified the nonlin-
ear relationships between LST and each of NDVI, NDBI and NDBaI
in UHI clusters, respectively. We found that both NDVI and NDBaI
are significantly correlated with the LST: the former mitigates the
UHI whiles the latter strengthens it. As a result, both NDVI and
NDBaI are good indicators of UHI. In addition, we found that the
NDBI also has a “temperature edge” relationship to the LST, as does
the NDVI. Comparatively, the NDBaI has a weak negative relation-
ship to the LST in UHI clusters. Compared to traditional multiple
linear regressions, the tree regression models resulted in advanced
improvement for exploring nonlinear relationships between LST
and landscape characteristics. Furthermore, a set of rules derived
from regression tree analyses demonstrated the spatial hetero-
geneity of landscapes within UHI clusters. These rules were linked
to special landscape characteristics, such as urban villages and CBD.
Based on the type of landscape characteristics in urban areas, we
build a particular model that is able to predict the LST variations
efficiently and can be applied to other urban planning and UHI
mitigation efforts.
Acknowledgements
This study was supported by the National Natural Science Foun-
dation of China (41171446, 41201543 and 31170445), National
Science and Technology support project plan (2012BAC13B01). The
authors would like to thank the editor and the anonymous review-
ers for their helpful comments and suggestions. Especially, we are
grateful to Dr. Guoping Tang for his comments on the early version
of this manuscript.
References
Anselin, L. (1995). Local indicators of spatial association—LISA. Geographical Analysis,
27(2), 93–115.
Anselin, L., Syabri, I., & Kho, Y. (2006). GeoDa: An introduction to spatial data analysis.
Geographical Analysis, 38(1), 5–22.
Arnfield, A. J. (2003). Two decades of urban climate research: A review of turbulence,
exchanges of energy and water, and the urban heat island. International Journal
of Climatology, 23(1), 1–26.
Bechtel, B., Zakšek, K., & Hoshyaripour, G. (2012). Downscaling land surface tem-
perature in an urban area: A case study for Hamburg, Germany. Remote Sensing,
4(10), 3184–3200.
Blaschke, T. (2003). Object-based contextual image classification built on image seg-
mentation. In Advances in techniques for analysis of remotely sensed data 2003 IEEE
workshop on (pp. 113–119). IEEE.
Blaschke, T. (2010). Object based image analysis for remote sensing. ISPRS Journal of
Photogrammetry and Remote Sensing, 65(1), 2–16.
Blaschke, T., & Hay, G. J. (2001). Object-oriented image analysis and scale-space: The-
ory and methods for modeling and evaluating multiscale landscape structure.
International Archives of Photogrammetry and Remote Sensing, 34(4), 22–29.
Burnett, C., & Blaschke, T. (2003). A multi-scale segmentation/object relationship
modelling methodology for landscape analysis. Ecological Modelling, 168(3),
233–249.
Buyantuyev, A., & Wu, J. (2010). Urban heat islands and landscape heterogeneity:
Linking spatiotemporal variations in surface temperatures to land-cover and
socioeconomic patterns. Landscape Ecology, 25(1), 17–33.
10 G. Guo et al. / Landscape and Urban Planning 135 (2015) 1–10
Cao, X., Onishi, A., Chen, J., & Imura, H. (2010). Quantifying the cool island intensity
of urban parks using ASTER and IKONOS data. Landscape and Urban Planning,
96(4), 224–231.
Carlson, T. N., Capehart, W. J., & Gillies, R. R. (1995). A new look at the simpli-
fied method for remote sensing of daily evapotranspiration. Remote Sensing of
Environment, 54(2), 161–167.
Carlson, T. N., & Traci Arthur, S. (2000). The impact of land use—Land cover changes
due to urbanization on surface microclimate and hydrology: A satellite perspec-
tive. Global and Planetary Change, 25(1), 49–65.
Chen, X.-L., Zhao, H.-M., Li, P.-X., & Yin, Z.-Y. (2006). Remote sensing image-based
analysis of the relationship between urban heat island and land use/cover
changes. Remote Sensing of Environment, 104(2), 133–146.
Chun, B., & Guldmann, J.-M. (2014). Spatial statistical analysis and simulation of the
urban heat island in high-density central cities. Landscape and Urban Planning,
125, 76–88.
Deng, C., & Wu, C. (2013). Examining the impacts of urban biophysical compositions
on surface urban heat island: A spectral unmixing and thermal mixing approach.
Remote Sensing of Environment, 131, 262–274.
Gallo, K. P., & Owen, T. W. (1999). Satellite-based adjustments for the urban heat
island temperature bias. Journal of Applied Meteorology, 38(6), 806–813.
Gillies, R., Kustas, W., & Humes, K. (1997). A verification of the ‘triangle’ method
for obtaining surface soil water content and energy fluxes from remote mea-
surements of the Normalized Difference Vegetation Index (NDVI) and surface e.
International Journal of Remote Sensing, 18(15), 3145–3166.
Gutman, G., & Ignatov, A. (1998). The derivation of the green vegetation fraction from
NOAA/AVHRR data for use in numerical weather prediction models. International
Journal of Remote Sensing, 19(8), 1533–1543.
Hay, G., Marceau, D., Dube, P., & Bouchard, A. (2001). A multiscale framework for
landscape analysis: Object-specific analysis and upscaling. Landscape Ecology,
16(6), 471–490.
Imhoff, M. L., Zhang, P., Wolfe, R. E., & Bounoua, L. (2010). Remote sensing of the
urban heat island effect across biomes in the continental USA. Remote Sensing of
Environment, 114(3), 504–513.
Li, J., Song, C., Cao, L., Zhu, F., Meng, X., & Wu, J. (2011). Impacts of landscape structure
on surface urban heat islands: A case study of Shanghai, China. Remote Sensing
of Environment, 115(12), 3249–3263.
Liu, H., & Weng, Q. (2008). Seasonal variations in the relationship between land-
scape pattern and land surface temperature in Indianapolis, USA. Environmental
Monitoring and Assessment, 144(1–3), 199–219.
Nichol, J. (1998). Visualisation of urban surface temperatures derived from satellite
images. International Journal of Remote Sensing, 19(9), 1639–1649.
Nichol, J. E. (1996). High-resolution surface temperature patterns related to urban
morphology in a tropical city: A satellite-based study. Journal of Applied Meteo-
rology, 35(1), 135–146.
Nichol, J. E., Fung, W. Y., Lam, K.-s., & Wong, M. S. (2009). Urban heat island diagnosis
using ASTER satellite images and ‘in situ’ air temperature. Atmospheric Research,
94(2), 276–284.
Oke, T. (1976). The distinction between canopy and boundary-layer urban heat
islands. Atmosphere, 14(4), 268–277.
Owen, T. (1998). Using DMSP-OLS light frequency data to categorize urban envi-
ronments associated with US climate observing stations. International Journal of
Remote Sensing, 19(17), 3451–3456.
Owen, T., Carlson, T., & Gillies, R. (1998). An assessment of satellite remotely-sensed
land cover parameters in quantitatively describing the climatic effect of urban-
ization. International Journal of Remote Sensing, 19(9), 1663–1681.
Qin, Z.-h., Karnieli, A., & Berliner, P. (2001). A mono-window algorithm for retriev-
ing land surface temperature from Landsat TM data and its application to
the Israel–Egypt border region. International Journal of Remote Sensing, 22(18),
3719–3746.
Qin, Z., Dall’Olmo, G., Karnieli, A., & Berliner, P. (2001). Derivation of split window
algorithm and its sensitivity analysis for retrieving land surface temperature
from NOAA-advanced very high resolution radiometer data. Journal of Geophys-
ical Research: Atmospheres (1984–2012), 106(D19), 22655–22670.
Roth, M., Oke, T., & Emery, W. (1989). Satellite-derived urban heat islands from three
coastal cities and the utilization of such data in urban climatology. International
Journal of Remote Sensing, 10(11), 1699–1720.
Voogt, J. A., & Oke, T. R. (2003). Thermal remote sensing of urban climates. Remote
Sensing of Environment, 86(3), 370–384.
Weng, Q. (2009). Thermal infrared remote sensing for urban climate and
environmental studies: Methods, applications, and trends. ISPRS Journal of Pho-
togrammetry and Remote Sensing, 64(4), 335–344.
Weng, Q., Liu, H., & Lu, D. (2007). Assessing the effects of land use and land cover
patterns on thermal conditions using landscape metrics in city of Indianapolis,
United States. Urban Ecosystems, 10(2), 203–219.
Weng, Q., Lu, D., & Liang, B. (2006). Urban surface biophysical descriptors and land
surface temperature variations. Photogrammetric Engineering & Remote Sensing,
72(11), 1275–1286.
Weng, Q., Lu, D., & Schubring, J. (2004). Estimation of land surface
temperature–vegetation abundance relationship for urban heat island
studies. Remote sensing of Environment, 89(4), 467–483.
Wu, J. (1999). Hierarchy and scaling: Extrapolating information along a scaling lad-
der. Canadian Journal of Remote Sensing, 25(4), 367–380.
Wu, J., & David, J. L. (2002). A spatially explicit hierarchical approach to model-
ing complex ecological systems: Theory and applications. Ecological Modelling,
153(1), 7–26.
Wu, J., & Loucks, O. L. (1995). From balance of nature to hierarchical patch
dynamics: A paradigm shift in ecology. Quarterly Review of Biology, 70(4),
439–466.
Xiao, R.-b., Ouyang, Z.-y., Zheng, H., Li, W.-f., Schienke, E. W., & Wang, X.-k.
(2007). Spatial pattern of impervious surfaces and their impacts on land sur-
face temperature in Beijing, China. Journal of Environmental Sciences, 19(2),
250–256.
Yuan, F., & Bauer, M. E. (2007). Comparison of impervious surface area and
normalized difference vegetation index as indicators of surface urban heat
island effects in Landsat imagery. Remote Sensing of Environment, 106(3),
375–386.
Yue, W., Xu, J., Tan, W., & Xu, L. (2007). The relationship between land surface temper-
ature and NDVI with remote sensing: Application to Shanghai Landsat 7 ETM+
data. International Journal of Remote Sensing, 28(15), 3205–3226.
Zhang, Y., Odeh, I. O., & Han, C. (2009). Bi-temporal characterization of land surface
temperature in relation to impervious surface area, NDVI and NDBI, using a
sub-pixel image analysis. International Journal of Applied Earth Observation and
Geoinformation, 11(4), 256–264.
Zhou, W., Huang, G., & Cadenasso, M. L. (2011). Does spatial configuration matter?
Understanding the effects of land cover pattern on land surface temperature in
urban landscapes. Landscape and Urban Planning, 102(1), 54–63.
Zhou, W., Qian, Y., Li, X., Li, W., & Han, L. (2014). Relationships between land cover
and the surface urban heat island: Seasonal variability and effects of spatial and
thematic resolution of land cover data on predicting land surface temperatures.
Landscape Ecology, 29(1), 153–167.