Landscape and Urban Planning 163 (2017) 107–120

Contents lists available at ScienceDirect

Landscape and Urban Planning
journal homepage: www.elsevier.com/locate/landurbplan

Research Paper

On the association between land system architecture and land surface
temperatures: Evidence from a Desert Metropolis—Phoenix, Arizona,
U.S.A
Xiaoxiao Li a,∗ , Yiannis Kamarianakis b , Yun Ouyang c , Billie L. Turner II d , Anthony Brazel e
a
School of Geographical Sciences & Urban Planning, Julie Ann Wrigley Global Institute of Sustainability, Central Arizona-Phoenix Long-Term Ecological
Research, Arizona State University, Tempe, AZ 85287-5302, United States
b
School of Mathematics & Statistical Sciences, College of Liberal Arts and Sciences, Arizona State University, Tempe, AZ 85287-1804, United States
c
Enterprise Data and Analytics Office, USAA, 1 Norterra Drive, Phoenix, AZ 85085, United States
d
School of Geographical Sciences & Urban Planning, School of Sustainability, Arizona State University, Tempe, AZ 85287-5302, United States
e
School of Geographical Sciences & Urban Planning, Arizona State University, Tempe, AZ 85287-5302, United States

h i g h l i g h t s
•
•
•
•
•

Land system architecture affects land surface temperature (LST) of residential parcels.
Land-cover composition has the largest effect on LST but land-cover configuration is significant.
Compact and concentrated land-covers, foremost vegetation, improves nighttime cooling.
Large land-cover units of irregular shape improve daytime cooling.
Parcel level land architecture can be used to mitigate the LST of residences.

a r t i c l e

i n f o

Article history:
Received 23 March 2016
Received in revised form 3 February 2017
Accepted 14 February 2017
Available online 31 March 2017
Keywords:
Land system architecture
Linear mixed-effects models
MODIS/ASTER
Urban heat island effect
Land surface temperature
NAIP
Parcel scale

a b s t r a c t
The relationship between the characteristics of the urban land system and land surface temperature (LST)
has received increasing attention in urban heat island and sustainability research, especially for desert
cities. This research generally employs medium or coarser spatial resolution data and primarily focuses
on the effects of a few classes of land-cover composition and pattern at the neighborhood or larger level
using regression models. This study explores the effects of land system architecture—composition and
configuration, both pattern and shape, of fine-grain land-cover classes—on LST of single family residential
parcels in the Phoenix, Arizona (southwestern USA) metropolitan area. A 1 m resolution land-cover map
is used to calculate land architecture metrics at the parcel level, and 6.8 m resolution MODIS/ASTER data
are employed to retrieve LST. Linear mixed-effects models quantify the impacts of land configuration on
LST at the parcel scale, controlling for the effects of land composition and neighborhood characteristics.
Results indicate that parcel-level land-cover composition has the strongest association with daytime and
nighttime LST, but the configuration of this cover, foremost compactness and concentration, also affects
LST, with different associations between land architecture and LST at nighttime and daytime. Given
information on land system architecture at the parcel level, additional information based on geographic
and socioeconomic variables does not improve the generalization capability of the statistical models. The
results point the way towards parcel-level land-cover design that helps to mitigate the urban heat island
effect for warm desert cities, although tradeoffs with other sustainability indicators must be considered.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction

∗ Corresponding author.
E-mail addresses: xiaoxia4@asu.edu (X. Li), yiannis76@asu.edu
(Y. Kamarianakis), yun.ouyang.1@gmail.com (Y. Ouyang), Billie.L.Turner@asu.edu
(B.L. Turner II), abrazel@asu.edu (A. Brazel).
http://dx.doi.org/10.1016/j.landurbplan.2017.02.009
0169-2046/© 2017 Elsevier B.V. All rights reserved.

Over a quarter century ago, Forman (1990) called for ecological research on landscape mosaics (land-cover composition and
configuration) in regard to the environmental performance and
sustainability of landscapes (also Wu, 2013), and more recently has
championed this approach for urban ecology (Forman, 2014). Urban

108

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

climatology, in turn, has long considered the impacts of urban morphology or geometry on the urban heat island (UHI) effect. Much
attention has been given to the vertical dimensions and spacing
of buildings (e.g., urban canyon height-to-width ratios), sky-views,
and the density of general types of land covers, such as vegetation
and impervious surfaces (e.g., Stewart & Oke, 2012). In addition,
landscape architecture has long addressed the structure of outdoor
spaces, foremost parks and common and reclamation areas, with a
strong orientation towards their aesthetics attributes. It has drawn
on ecological principles in landscape design, however, more so than
examining the environmental dynamics of existing landscape conditions (Brown & Corry, 2011; Yu, Li, & Ji, 2001).
Joining these research interests and the approaches applied
to them, but differing somewhat from each, is the transdisciplinary subfield of land system (change) science, focused on the
causes and consequences of land change (Turner et al., 2007).
Land system science, partly a product of global environmental
change and sustainability interests, seeks to provide a sciencebased understanding of land-use and –cover change, increasingly
with an eye towards informing decision making. Land system architecture (henceforth, land architecture) is one dimension of this
effort (Janetos, Verbug, & Murray, 2013). It addresses the composition and configuration of land covers, including pattern, shape
and connectivity, and their impacts on social-environmental system performance (Turner, 2017). It differs slightly from landscape
mosaic approaches in its attention to fine-grain spatial analysis
(i.e., 1–30 m), and detailed categories of land covers (e.g., among
various categories of vegetation cover) and, in some cases, land
uses. These distinctions make the land architecture orientation
well fit to address the heterogeneous character of urban land
cover, and compatible with calls for land system science to address
urbanization-land relationships (e.g., Seto and Reenberg, 2014).
Land architecture explores configuration metrics derived from the
spatial sciences rather than relying solely on those developed in
ecology (see below). Urban land architecture has focused on the
UHI effect (Turner, 2017), but differs from urban morphology or
geometry in regard to the array of land-cover categories examined
and the details of configuration. Finally, the focus of land architecture (as defined here) on base research to inform decision making
about sustainability problems (Clark, 2007) distinguishes it from
the primary motivation of activities within the subfield of landscape architecture.
The UHI effect increasingly captures the attention of administrators and decision makers concerned with mitigating urban
temperature extremes and their impacts on energy and water use,
human health, and thermal comfort (e.g., City of Phoenix, 2010;
Kleerekoper, van Esch, & Baldiri Salcedo, 2012; Shashua-Bar, Pearlmutter, & Erell, 2009, Shashua-Bar, Pearlmutter, & Erell, 2011;
Solecki et al., 2005; Wentz, Rode, Li, & Tellman, 2016). This concern,
in turn, has drawn insights from UHI research undertaken in the
subfields and approaches discussed above as well as in remote sensing (e.g., Gago, Roldán, Pacheco-Torres, & Ordoñez, 2013; Hart &
Sailor, 2009; Oke, 1981; Taha, 1997). The problem and the research
given to it are particularly acute for urban areas worldwide situated
in warm and hot climates where the UHI effect is expected to be
significantly amplified by global climate warming (e.g., GrossmanClarke, Schubert, Clarke, & Harlan, 2014; McCarthy, Best, & Betts,
2010; Saha, Davis, & Hondula, 2014).
With these concerns in mind, the role of urban land composition and configuration on the UHI and its capacity to mitigate the
phenomenon has been examined in regard to both air and land
surface temperature, including the role of “smart buildings”, green
and white roofs, urban street canyons, green spaces, shade, impervious surfaces, bare soil, and sky-views (e.g., Coseo & Larsen, 2014;
Donovan & Butry, 2009; Erellet al., 2012; Giridharan, Lau, Ganesan,
& Givoni, 2007; Li, Bou-Zeid, & Oppenheimer, 2014; Oke, 1981;

Santamouris, 2013). Configuration has largely been addressed in
terms of the spatial distribution or pattern of general types of land
covers (Buyantuyev & Wu, 2010; Middel, Häb, Brazel, Martin, &
Guhathakurta, 2014; Myint, 2012; Myint, Wentz, Brazel, & Qualttrochi, 2013; Stewart and Oke, 2012), with abundant attention to
Chinese cities (e.g., Huang, Li, Zhao, & Zhu, 2008; Li, Wang, Wang,
Ma, & Zhang, 2009; Su, Gu, & Yang, 2010; Zhang et al., 2013).
Complementing recent research that treats land configuration
in rural, forest landscapes (Mitchell, Bennett, & Gonzalez, 2013;
Wu, Jenerette, Buyantuyev, & Redman, 2011) and consistent with
Forman’s (1990) call, nascent efforts are underway to determine
the role of land architecture, as defined here, on the UHI effect.
Most studies to date tend to employ 30 m or coarser resolution
satellite data to calculate daytime land surface temperature (LST)
of urban land units, address one or two land-covers at the neighborhood or larger level, use FRAGSTATS metrics (McGarigal and
Marks, 1995; McGarigal, Cushman, & Ene, 2012) to determine landcover configuration, and apply regression models to assess the land
cover-LST relationship (e.g., Turner, 2017). The results indicate that
the concentration of different land covers in Phoenix and Las Vegas
(USA) increased heating or cooling effects (Fan, Myint, & Zheng,
2015; Myint et al., 2015; Zheng, Myint, & Fan, 2014), whereas different elements of shape (e.g., edge density) have a strong effect
on land-cover temperature in Beijing, China (Li, Zhou, Ouyang, Xu,
& Zheng, 2012), Baltimore, USA (Zhou, Huang, & Cadenasso, 2011),
and Phoenix (Connors, Galletti, & Chow, 2013). Indeed, at least one
study in Phoenix found land-cover shape at the parcel level, as measured by the Normalized Moment of Inertia, trumped land-cover
pattern in terms of effects on LST at the neighborhood level (Li, Li,
Middel, Harlan, & Brazel, 2016).
In addition to the overall work on the UHI effect, land architecture approaches increasingly point to the significant role that both
the pattern and shape of small-size land covers have on temperature extremes. Here we extend this line of research by exploring
the role of land configuration on LST at the single-family residential (SFR) parcel level for the metropolitan area of Phoenix,
Arizona, U.S.A., using 2010–2011 data (Fig. 1). Consistent with
most other studies, FRAGSTATS serve as the metrics of composition and configuration. Following Li et al. (2016), multiple land
covers and their mosaics are examined. These land covers are
linked to 6.8 m MASTER data in order to match the fine-resolution,
heterogeneous land cover of the residential parcels. New to this
assessment, both daytime and nighttime LST are analyzed. In addition, the neighborhood effect (Cox, 1969; Johnston, Propper, Sarker,
Jones, & Bolster, 2005) on individual parcels is investigated using
linear mixed effect models (LMEs). LMEs allow us to identify significant predictors related to land configuration while accounting
for nested, neighborhood-specific data structures and controlling
for land composition, socioeconomic neighborhood characteristics and spatial correlation of neighboring parcels. Evaluation of
LMEs emphasized their generalization capability based on their
performance in predicting outcomes on neighborhoods not used to
estimate model parameters. Taken together, this study constitutes
a novel attempt to determine the role of residential land architecture on the UHI effect, specifically the surface urban heat island
(SUHI). The results have the potential to inform how a redesign or
reshaping of the land units could mitigate the extremes of the SUHI.

2. Study site, data, and methods
2.1. Study site
The Phoenix metropolitan area (Fig. 1) consists of 26 joined
cities and towns that housed about 4.2 million people in 2010
and covered over 7600 km2 of the northern Sonoran Desert of

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

109

Fig. 1. Location of the Phoenix, Arizona, USA, metropolitan area in the American Southwest and the Household Income of the Sampled Neighborhoods.

Maricopa County, AZ. Rapid urbanization, largely at the expense
of former irrigated farmlands (Keys, Wentz, & Redman, 2007;
Grossman-Clarke, Zehnder, Loridan, & Grimmond, 2010) and the
cooling effects that they render, has generated extreme heat problems, with maximum temperatures exacerbated by the SUHI effect
(Buyantuyev & Wu, 2009; Chow, Chuang, & Gober, 2012). The average summer air temperature exceeds 30 ◦ C, and the number of days
in excess of 43 ◦ C increased by 20 over the last decade (Connors
et al., 2013; Jenerette, Harlan, Stefanov, & Martin, 2011). Interestingly, the SUHI effect relative to the ambient rural temperature
is greater at night than day (Hawkins, Brazel, Stefanov, Bigler, &
Saffell, 2004). Overall the SUHI in the greater Phoenix area raises
a series of issues about water consumption—a highly constrained
resource, the largest urban uses of which are irrigated, parcel-level
vegetation and swimming pools—energy use, and public health
(Harlan, Brazel, Prashad, Stefanov, & Larsen, 2006; Silva, Phelan, &
Golden, 2010; Wentz et al., 2016). For these and other reasons, the
City of Phoenix (2010) plans to respond to the UHI, in part, through
greening its urban landscape.
2.2. Data collection and preparation
Three basic data sources are employed in this analysis (Fig. 2):
(1) remotely sensed imagery (1 m), augmented by parcel information; (2) locational and socio-economic data; and (3) LST taken from
remotely sensed imagery (6.8 m) matched to individual parcels.
2.2.1. Land-cover data
The data used to generate the land-cover classification and
metrics (independent variables; see 2.3.1) for SFR parcels were
generated from the 1 m resolution, 4-band orthophotography produced in 2010 by the National Agriculture Imagery Program (NAIP
imagery). The classification utilized an object-based approach with
expert knowledge decision rules, assisted by a cadastral vector

layer as ancillary data (Li, Myint, Zhang, Galletti, & Zhang, 2014).
Original NAIP imagery and cadastral shapefiles were pre-processed
by two types of spectral transformation and one type of spatial
enhancement. RGB to I transformation, Principal Components Analysis (PCA), Normalized Difference Vegetation Index (NDVI), and
morphological convolution operations were employed. The output images from the transformation functions were merged with
the original four-band images as the input data for an object-based
image analysis.
The object-based method sets up a hierarchical network in
which image objects at different levels are delineated according
to their spectral, spatial, contextual, and geometrical characteristics correspondingly. A number of segmentation algorithms
(multi-resolution, multi-threshold, Quadtree based, and chessboard segmentation) were executed for the hierarchical network,
and at a finer image object level, the detailed land-cover classes
were identified within parcels (Li, Myint, et al., 2014).
Twelve land-cover classes were distinguished, of which five
are common to SFR parcels: building, bare soil/rock, tree/shrub,
turf grass (henceforth, grass), and swimming pool. The impervious surfaces of driveways and sidewalks could not be consistently
distinguished from bare soil/rock. Land covers adjacent to but not
part of the SFRs, such as roads and playgrounds, were not considered in this analysis. We suspect minimal effects of these adjacent
covers on parcel LST. Heat island research indicates a well-defined
land-cover impact on LST when near-surface wind speeds are light,
below 5 mph or 2.2 m s−1 (Fast, Torcolini, & Redman, 2005).
A total 12,265 SFRs in 35 neighborhoods were used to calculate
the land architecture metrics for five land-cover types in each residential parcel. The parcels were delineated from the 2010 Maricopa
County Tax Assessor roles, and ranged from 38 to 2643 per neighborhood. All SFR parcels do not contain a swimming pool and the
variability of pool sizes is small. For this reason, the presence of

110

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

Fig. 2. Research design.

a pool was included as a factor variable in the statistical analysis.
Examples of the land-cover results are provided in Fig. 3.

2.2.2. Land surface temperature data
The MODIS/ASTER Airborne Simulator (MASTER) was used to
account for the fine-grain heterogeneous of the urban land covers.
It is an imaging spectrometer that acquires high spectral and spatial resolution imagery with fifty channels ranging from 0.4–14 ␮m
with an approximate 6.8 m pixel size (Hook, Myers, Thome, Fitzgerald, & Kahle, 2001; Chang, Han, Fan, Chen, & Chang, 2002). The
MASTER data were acquired by flights during the daytime (11:55
a.m.–2:00 p.m., July 12, 2011; 11:15 a.m.–1:05 p.m., July 13, 2011)
and nighttime (1:40 a.m.–4:10 a.m., July 15, 2011; 1:20 a.m.–3:10
a.m., July 16, 2011).
The MASTER data flight lines were segmented into smaller subsets in order to improve geo-referencing by minimizing distortions
and removing erroneous data (for flight lines see: Jenerette et al.,
2016). The LST was extracted and converted to ◦ C using ENVI software and its atmospheric correction module: Quick Atmospheric
Correction (Bernstein et al., 2005). The emissivity normalization
function for thermal IR bands was applied on each of the subsets.
The LST data were geo-referenced with NAIP orthophotography
(Section 2.2.1) from Ground Control Points with ENVI software.
These data were further extracted for each parcel polygon (obtained
from cadastral GIS vector layer) to represent the parcel mean LST
at daytime and nighttime. LSTs calculated this way are accurate for
light wind conditions as noted above. For the dates in question the
wind speeds for 52 weather stations across the metro-area were
within the aforementioned limits. We assume, therefore, that nearsurface winds did not affect our measures of LST, although each
neighborhood examined did not have a weather station. Examples
of the LST results are depicted in Fig. 3.

2.2.3. Socio-economic data
The Central Arizona-Phoenix Long-term Ecological Research
project has identified 40 neighborhoods that capture the socialenvironmental diversity (e.g., income, ethnicity, landscaping
preferences) within the metropolitan area. Socio-economic data
from 35 neighborhoods were used as exploratory variables to
address the possibility of neighborhood effects. In this case, the
group or relational effects were the role of neighborhood parcels on
an individual parcel. Socio-economic data for these neighborhoods
were drawn from the 2010 U.S. Census and included median household income and age of residents and ethnicity. Each neighborhood
was assigned to one of three locational categories: within the (1)
central core or (2) not of their respective cities and (3) adjacent to
the metropolitan fringe. Also, the ratio of living area divided by the
total area of each parcel (P LIVING; county assessor data) was calculated. This explanatory variable was included because preliminary
analyses indicated that the percentage of the parcel area covered
by buildings (P BUILDING; see 2.3.1) was not strongly correlated
with LST, a result that appeared to be counterintuitive.

2.3. Methods
2.3.1. Land architecture metrics
FRAGSTATS (or landscape metrics; Table 1) were developed for
ecosystem and landscape research but have been employed in various research dealing with land configuration (McGarigal et al.,
2012). To maintain consistency with this research, FRAGSTATS
were used as independent, or predictor variables in our statistical models. Both disaggregate (land-cover specific) and aggregate
(parcel level) metrics were used in the development of predictive models for LST. In order to maximally reduce the redundancy
among the many FRAGSTATS metrics as well as to include as many
important measurements as possible, five metrics (PLAND, PD, ED,

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

111

Fig. 3. Examples of the land cover of neighborhoods (1 m) and their land surface temperatures (LST) (6.8 m). (a) Low-level (xeric) and (b) high level (mesic) vegetated
neighborhoods; (c) daytime LST of xeric and (d) nighttime LST of mesic neighborhoods.

LSI, FRAC) were chosen to quantify the character of each land-cover
type. These five metrics were grouped in two categories: land composition (PLAND) and spatial configuration (the other four). PLAND
measures the fraction of patches of each land-cover type within a
unit, while the spatial configuration metrics characterize the shape
complexity and spatial distribution of patches in the unit (Table 1).
PLAND and six configuration metrics (PD, ED, LSI, FRAC, CONTAG,
and SHDI) were selected to measure the aggregated architecture of
land patches per parcel.
The metrics were calculated using the FRAGSTATS 4.2 software
package, and obeyed the “8-cell rule” (consider all 8 adjacent cells,
including the 4 orthogonal and 4 diagonal neighbors) in defining

patch neighborhoods (McGarigal et al., 2012). Rather than using
a fixed pixel size as the sample unit common to other studies
(e.g., Connors et al., 2013), SFR parcels ranging from 372 m2 to
6039 m2 were addressed, permitting a more precise spatial fit
among parcels, neighborhoods, and LST.

2.3.2. Statistical analysis
Linear mixed-effects models were employed to quantify the
land architecture-LST relationship for summer nighttime and daytime, while accounting for the effects of time, land composition,
socio-economic and locational predictors. LMEs can be viewed
as generalizations of neighborhood-specific regression models,

112

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

Table 1
Land architecture metrics.
Metric/Abbreviation

Description

Class Specific Metrics

Aggregated Parcel Metrics

Percent cover of a
land-cover
class/PLAND
Patch density/PD

Proportion of the land-cover type
(patch type) on the unit landscape plot
(0 < PLAND ≤ 100)
Number of patches/ha (>0, determined
by grain or pixel size)
Total length of edges for all patches/ha
(≥0, where 0 refers to a landscape
composed of one patch)
Total length of all patch edges divided
by the minimum possible length of the
area of the unit landscape plot (≥1,
where the greater the LSI above 1 the
more the shape deviates from a
compact shape, i.e., a square)
A measure of shape complexity by
calculating the departure of the patch
from its Euclidean geometry
(1 ≤ FRAC ≤ 2, where 1 corresponds to
very simple shapes and 2 to extremely
complex shapes)
A measure of adjacency of patches
(0 < CONTAG ≤ 100, where patches are
maximally disaggregated and
dispersed when the values are small;
100 is the reverse)
A measure of the diversity of patches
(0 is a landscape with 1 patch)

PLAND of Building, Soil, Soil,
Tree/Shrub, Grass

N/A

PD of Building, Soil, Soil, Tree/Shrub,
Grass
ED of Building, Soil, Soil, Tree/Shrub,
Grass

PD

LSI of Building, Soil, Soil, Tree/Shrub,
Grass

LSI

FRAC of Building, Soil, Soil, Tree/Shrub,
Grass

FRAC

N/A

CONTAG

N/A

SHDI

Edge Density/ED

Landscape Shape
index/LSI

Fractal dimension/
FRAC

Contagion/CONTAG

Shannon’s diversity
index/SHDI

which summarize statistical associations when measurements are
collected from a group of neighborhoods. In contrast with regression, the adopted model class provides valid inferences with
regard to the significance of the predictors by accounting for
nested, neighborhood-specific data structures and spatial correlation effects within each neighborhood. Parcels are sampled in
clusters. Those from the same neighborhood are expected to be
correlated. Ignoring the clustered data structure may lead to inappropriate estimates of standard errors of model parameters and
consequently to errors in statistical inference due to inaccurate pvalues. In addition to the underestimation of standard errors, by
ignoring the multilevel structure of the data, important relationships involving each level of the data (e.g., important predictors
that characterize neighborhoods; Finch, Bolin, & Kelley, 2014) may
be missed.
The estimated models are formulated as:
Y i = X i × ˇ + Z i × bi + εi
bi ∼N (0, D)
εi ∼N (0, ˙ i )
b1 , . . ., bN , ε1 , . . ., εN independent

⎫
⎪
⎪
⎪
⎪
⎬
⎪
⎪
⎪
⎪
⎭

.

(1)

In (1), Yi denotes the vector of observed LST in neighborhood i with
i = 1, . . ., N; its length depends on the number of sampled parcels in
i. Xi represents the matrix of standardized aggregate and disaggregate architecture and socioeconomic predictors, with ˇ indicating
the vector of coefficients for the fixed-effects. Neighborhoodspecific random effects are captured by the Zi × bi term; vectors
bi are assumed to follow a multivariate normal distribution with
variance-covariance matrix D which is symmetric and positivedefinite, while Zi is a matrix that typically contains a subset of the
columns of Xi .
Fixed-effects represent ‘average’ model parameters whereas
random-effects denote neighborhood-specific deviations from
‘average’ dynamics. We report estimated coefficients on standardized variables. These coefficients represent the effect (on LST) of a
standard deviation change for a predictor, keeping the other predictors fixed, and allow us to evaluate the relative significance of

ED

the explanatory variables. The random effects bi and the withinneighborhood errors εi are assumed to be independent for different
neighborhoods. The neighborhood-specific covariance matrix of
the error terms ˙ i is specified to account for spatially correlated
parcels within each neighborhood. Specifically, spatial correlation
is parameterized by





cor εi , εj =





1

h εi , εj , c 0 , 



if i = j
else

(2)

with parameters c0 ,  denoting the nugget and the range, respectively. Three alternative forms were evaluated for h: Exponential,
Gaussian and Spherical (for details on models for spatially correlated residuals see Pinheiro & Bates, 2009; Chapter 5).
Different models were developed for daytime and nighttime
temperatures, based on weighted restricted maximum likelihood
estimation (REML) that accounted for total lot size. Two model
building procedures were implemented. First, the entire set of predictors and interactions of socio-economic and locational factors
with architecture metrics was used as input in a regression model.
A penalized estimation method with near oracle asymptotic properties, namely adaptive LASSO (Zou, 2006), was used to reduce
the initial set (>30) of explanatory variables. Adaptive Lasso performs simultaneous predictor selection and coefficient estimation;
it was applied here since stepwise LME building is computationally
infeasible when the set of explanatory variables is large. A BoxCox procedure was applied to identify functional transformations
of LST (e.g., logarithmic) that ‘linearize’ statistical associations; the
results of 5-fold cross-validation indicated negligible improvement
in predictive performance, hence it was decided to analyze LST in
its original scale.
In addition to the explanatory variables presented above, information on date, time and geographical coordinates was used in
developing LMEs. Given the predictors identified from adaptive
LASSO, mixed effects models were developed following the protocol described in Zuur, Ieno, Walker, Savaliev, and Smith (2009) and
the diagnostics presented in Pinheiro and Bates (2009). Specifically,
a stepwise implementation of REML was performed to identify
significant random effects: predictors with varying effects across

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

neighborhoods. Another stepwise procedure based on conditional
F-tests was implemented to derive the structure of the fixed-effects
part, Xi × ˇ. Finally, the three alternative spatial correlation structures for the residuals were evaluated against the null model, which
assumes independent residuals using Akaike’s Information Criterion. The models derived based on the abovementioned procedure
are referred to as Model 1D and Model 1N , for daytime and nighttime respectively.
The second model building procedure differed from the first
in the initial step: instead of adaptive LASSO, a classic stepwise
method, namely best subset regression (Kutner, Nachtsheim, &
Neter, 2004; James, Witten, Hastie, & Tibshirani, 2013) was implemented using the original set of predictors. Interaction terms were
not included since this iterative method is extremely demanding in
terms of computation time when the number of predictors is large
(>40). The corresponding models for daytime and nighttime LST are
dubbed Model 2D and Model 2N, respectively.
Mixed-effects models can be used to predict LST at neighborhoods not included in the original sample, based solely on
fixed-effects. Therefore, models derived by the two alternative
approaches were compared by leave-one-neighborhood-out crossvalidation, based on weighted accuracy metrics. Prediction errors
within neighborhoods were weighted for parcel size and overall performance was based on weighted averages that accounted
for the number of parcels per neighborhood. Three metrics are
reported: (1) weighted mean error (WME), a measure of bias; (2)
weighted mean absolute error (WMAE); and (3) weighted mean
absolute percentage error (WMAPE). To emphasize the significance
of predictors related to spatial configuration, the model building
procedures were repeated excluding those predictors. The generalization capabilities of these reduced LMEs were compared to the
full models using leave-one-neighborhood-out cross-validation.
3. Results
Beyond the descriptive statistics, attention is given to those
explanatory variables with strong significance (p < 0.01) and higher
standardized coefficients.

113

best in a leave-one-neighborhood-out cross-validation, included
predictors related to both variable types that comprise land architecture. Furthermore, the LMEs derived by ignoring predictors
based on configuration, displayed inferior generalization capability (Table 3, parenthetical values). The best performing model for
nighttime LST was derived from the best subset regression (Model
2N ) and achieved an accuracy close to 1 ◦ C in terms of WMAE (Fig. 5),
a satisfactory result. Daytime temperatures display more variability, influencing the performance of the LMEs with a WMAE slightly
below 3 ◦ C (Fig. 5). Accounting for spatially correlated residuals
resulted in improved specifications for nighttime, but not for daytime, temperatures. The models also account for temporal increases
and decreases LST by the timing of the remote sensing data. The day
and night signs of the effects were in accordance with prior expectations, with positive effects on LST as daytime progresses from 11
a.m. to 2 p.m. (Table 5) and negative effects as nighttime advances
from about 1:30 a.m. to 4 a.m. (Table 4).
3.2.1. Effects of land composition
The percentage of grass and the presence of a pool proved to be
especially significant for nighttime LST, with negative and positive
effects on LST, respectively (Model 2N , Table 4). The effects of grass
vary significantly across neighborhoods; this variability could not
be explained by land configuration, socio-economic, or locational
factors. In addition to these predictors, the suboptimal (in terms of
generalization capability relative to the model based on best subsets regression) LME based on adaptive LASSO (Table 3) included
the composition of built land and the ratio of living area relative to
the total area of the parcel.
The percentage of grass-covered surface in a parcel and the presence of a pool, display negative associations with daytime LST as
expected, with high standardized coefficients (Model 1D and Model
2D , Table 5). These coefficients reveal a stronger cooling effect of
PLAND of trees on daytime temperatures relative to the effect of
grass and pool. Again, the daytime effects of grass displayed significant variability across neighborhoods; this variability could not
be explained by land configuration, socioeconomic, or geographic
factors.

3.1. Descriptive statistics
Fig. 4 and Table 2 provide the descriptive statistics of the SFR
parcels examined. Their mean daytime and nighttime LSTs are
49.6 ◦ C and 23.7 ◦ C, respectively. The range among the parcels was
37.60 ◦ C to 62.41 ◦ C (s.d. = 3.8 ◦ C) for the daytime and 19.11 ◦ C to
30.43 ◦ C (s.d. = 1.4 ◦ C) for the nighttime. Recall that these are land
surface, not air, temperatures. Such extreme highs and parcel to
parcel variations have been recorded in other studies of Phoenix
(Jenerette et al., 2016; Middel, Brazel, Kaplan, & Myint, 2012).
On average, building (SFR structure) occupies 32% of the parcels,
with the remainder composed of the other land covers (Table 2).
Of those other covers, the PLAND of soil has greatest variation
(s.d. = 13.76), reflecting mesic to xeric landscaped parcels, whereas
the PLAND for grass had the smallest variation (s.d. = 8.09). The
average values for aggregated architecture (parcel level) metrics are
3.57 (LSI), 25,649 (PD), 1.23 (FRAC), and 46.26 (CONTAG), respectively, whereas different ranges of variation existed for individual
land-cover types (see Table 2 for metric meanings). As noted above,
statistical models were based on standardized predictors; hence
the corresponding coefficients can be interpreted despite the different ranges of variation of the original variables.
3.2. Mixed-effects models
Land composition and configuration are both significant in
regard to their effects on LST (Tables 3–5). The LMEs that performed

3.2.2. Effects of land configuration
Controlling for the effects of land composition, socio-economic,
and locational factors, the LSI for grass is significant for nighttime
LST (Model 2N , Table 4), with a positive standardized coefficient.
Larger values of LSI GRASS (i.e., less compact shape) are associated with higher temperatures; the estimated effect is substantially
weaker relative to the significant composition variables. The less
parsimonious (and less effective in terms of predictive power)
LME for nighttime LST (Model 1N , Table 4) includes additional
class-specific metrics. Interestingly, increasing the edge density
(i.e., numerous small patches) of grass is positively associated with
nighttime LST. For the aggregated architecture configuration metrics (Model 1N ), the diversity of patches within each parcel (SHDI)
holds a positive relationship with LST.
Only aggregated configuration metrics, namely PD and LSI,
appear in the set of significant predictors for daytime LST, with standardized coefficients about 50% reduced compared to the effects of
PLAND of grass. PD is positively related to LST, whereas the LSI
is negatively associated. Based on the estimated coefficients, an
increase on the levels of PD and LSI equal to the observed standard deviation of PD and LSI is expected to lead to an increase and
decrease, respectively, in daytime LST of about 0.28 ◦ C. PD is not
strongly correlated with LSI (Spearman’s correlation coefficient is
<0.5), hence their effects do not ‘cancel out’. The effects of both
metrics varied significantly across neighborhoods; this variability
could not be explained by socio-economic and locational factors.

114

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

3.2.3. Effects of neighborhood socio-economic characteristics
None of the socio-economic predictors, not even the geographical coordinates, were included in the final steps of the model
building procedures for LMEs. It was expected that such explanatory variables would account for some of the variability of the
random effects that relate with land architecture predictors, as indicated in research using coarser resolution LST data (e.g., Li et al.,
2016) and based on recent work elsewhere (Huang & Cadennaso,
2016). On the contrary, it appears that given information on land
composition and configuration, the additional variables provide
redundant information.

4. Discussion
The two factors of land architecture—composition and configuration of land cover—taken together provide the best modeling
results. Considered individually, composition proves to be the most
important for the LST of SFR parcels. This result is consistent with
those from larger grain assessments of the Phoenix metro area
and elsewhere, as well as with a recent fine-grain assessment for
Phoenix (Jenerette et al., 2016). It differs, however, from a fine-grain
Phoenix study, employing a Normal Moment of Inertia measure of
configuration, that found configuration to be more important than
composition in affecting LST (Li et al., 2016).

Nighttime LST decreases as the percent of grass cover increases,
apparently due to evapotranspiration (Weng, Lu, & Schubring,
2004, Li et al., 2012). Congruent with the findings of Connors et al.
(2013), the strength of this relationship varies significantly across
neighborhoods (Table 4), perhaps affected by different levels and
timing of irrigation by parcel and neighborhood. Pool water absorbs
and stores energy as do trees in and under their canopies, releasing
it during the evening with positive links to nighttime LST (Myint
et al., 2013). Our models for nighttime LSTs identified a strong positive effect related to the presence of pools; on the other hand, the
positive effect of trees/shrubs on LST could not be identified, despite
occupying about 16% of the parcel areas examined.
A large percentage of building coverage is expected to increase
nighttime LST due to reradiated energy built-up during the day and,
perhaps, to feedbacks from near surface air warming from the waste
heat from air conditioning and refrigeration systems (Salamanca,
Georgescu, Mahalov, Moustaoui, & Wang, 2014; Zhou et al., 2011).
Given information on grass and pools, the percentage of building,
trees/shrubs, and soil cover did not add significant predictive power
to our models for nighttime LST. Consistent with previous analyses
of LSTs in Phoenix (Connors et al., 2013; Zheng et al., 2014), this
result is not surprising since the percentage of grass is negatively
correlated with the percentage of buildings (Spearman correlation
coefficients equals −0.29).

Fig. 4. Distributions of daytime and nighttime land surface temperature (LST) by neighborhood and parcels.

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

115

Fig. 4. (Continued)

Table 2
Statistical summary of explanatory and predictor variables.
Variables

Mean

Std Dev

Lower 95% Mean

Std Err Mean

Upper 95% Mean

Aggregated Architecture Metrics

PD
ED
LSI
FRAC
CONTAG
SHDI

30078.12
3290.29
3.58
1.23
44.59
1.22

11254.32
880.40
0.80
0.05
9.30
0.18

29882.33
3274.98
3.56
1.23
44.43
1.22

99.89
7.81
0.01
0.00
0.08
0.00

30273.92
3305.61
3.59
1.23
44.75
1.22

Class specific Metrics

PLAND BUILDING
PD BUILDING
ED BUILDING
LSI BUILDING
FRAC BUILDING
PLAND SOIL
PD SOIL
ED SOIL
LSI SOIL
FRAC SOIL
PLAND TREE/SHURB
PDTREE/SHURB
EDTREE/SHURB
LSITREE/SHURB
FRACTREE/SHURB
GRASS
PDGRASS
EDGRASS
LSIGRASS
FRACGRASS

31.58
2171.28
1035.92
1.48
1.13
40.28
7781.37
2530.83
3.96
1.26
15.91
9355.13
1706.08
3.64
1.24
8.88
8206.45
1022.57
3.05
1.21

11.29
1463.83
318.19
0.39
0.09
13.76
5416.51
709.21
0.98
0.09
11.05
5445.73
940.56
1.23
0.09
8.09
4116.30
586.39
1.10
0.11

31.38
2145.81
1030.38
1.47
1.13
40.04
7687.14
2518.49
3.95
1.26
15.72
9260.39
1689.72
3.62
1.24
8.74
8134.84
1012.37
3.03
1.21

0.10
12.99
2.82
0.00
0.00
0.12
48.07
6.29
0.01
0.00
0.10
48.33
8.35
0.01
0.00
0.07
36.53
5.20
0.01
0.00

31.77
2196.74
1041.45
1.49
1.14
40.52
7875.60
2543.16
3.98
1.27
16.10
9449.87
1722.44
3.67
1.25
9.02
8278.06
1032.77
3.07
1.21

Land Surface Temperature (◦ C)

Parcel daytime
Parcel nighttime

49.66
23.76

3.81
1.48

49.59
23.74

0.03
0.01

49.73
23.79

116

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

Table 3
Akaike (AIC) and Bayesian (BIC) information criteria, and weighted cross-validation performance criteria for the daytime- and nighttime-specific linear mixed effect models.
Weighted cross-validation performance criteria in parentheses correspond to values obtained when model building was applied without predictors related to land configuration (not reported for Model 2D which does not include such predictors). h is the functional form of the spatial correlation structure for the residuals: nS ,rS denote the
nugget and the range of a spherical variogram, ne ,re represent the nugget and the range of an exponential variogram, respectively. Generalization performance of alternative
models is evaluated using leave-one neighborhood-out cross-validation. Three metrics are reported: Weighted Mean Error (WME), Weighted Mean Absolute Error (WMAE)
and Weighted Mean Absolute Percentage Error (WMAPE).
h

AIC
BIC
WME
WMAE
WMAPE

Model 1N

Model 2N

nS = 0.68, rS = 789

ne = 0.61, re = 813

27714
27885
0.06 (−0.12)
1.32 (1.80)
0.06 (0.08)

28556
28667
−0.06 (0.02)
1.06 (1.35)
0.05 (0.06)

Model 1D

Model 2D

58594
58690
0.16 (0.4)
2.73 (3.43)
0.05 (0.07)

59723
59797
0.14
3.16
0.07

Table 4
Estimated mixed-effects models for nighttime land surface temperature. Fixed effects coefficients are significant with p < 0.01. Standard deviations of significant random
effects are shown in parentheses, next to the corresponding fixed-effects. Model 1 and 2 correspond to alternative model building procedures.
Variable

INTERCEPT
PLAND GRASS
POOL
P LIVING
PLAND BUILDING
LSI BUILDING
ED SOIL
LSI TREE/SHURB
PD GRASS
ED GRASS
LSI GRASS
CONTAG
SHDI
DATE
TIME

Model 1N

Model 2N

Coefficient

Std Error

Coefficient

Std Error

23.89 (1.04)
−0.40 (0.11)
0.23
0.06
−0.18 (0.11)
0.07
−0.13
0.08
−0.06
0.15

0.18
0.03
0.02
0.01
0.02
0.01
0.01
0.01
0.01
0.02

23.64 (0.91)
−0.18 (0.10)
0.30

0.30
0.02
0.02

0.10

0.01

0.04
0.17 (0.12)
0.16
−0.19

0.01
0.02
0.04
0.02

0.17
−0.22

0.04
0.02

Table 5
Estimated mixed-effects models for daytime land surface temperature. Fixed effects coefficients are significant with p < 0.01. Standard deviations of significant random effects
are shown in parentheses, next to the corresponding fixed-effects coefficients. Model 1 and 2 correspond to alternative model building procedures.
Variable/Metrics

INTERCEPT
PLAND TREE/SHRUB
PLAND GRASS
P LIVING
POOL
PD
LSI
DATE
TIME

Model 1D

Model 2D

Coefficient

Std Error

Coefficient

Std Error

49.37 (2.39)
−0.80
−0.67 (0.17)
−0.42
−0.68
0.28 (0.34)
−0.28 (0.44)
0.92
1.67

0.38
0.04
0.05
0.03
0.06
0.08
0.09
0.13
0.07

49.39 (2.57)

0.44

−0.68
−0.42
−0.61

0.03
0.03
0.06

0.90
1.73

0.13
0.07

Increasing percentages of building and soil cover are expected
to increase LST during the day owing to rapid re-radiation, while
grass and tree cover lower LST due to evapotranspiration (Zhou
et al., 2011). Our statistical models identified the negative effects of
grass and tree/shrub cover. Standardized coefficients suggest that
increasing the percentage of trees is more effective than grass in
lowering LSTs in daytime, consistent with the findings by Myint
et al. (2013) and Jenerette et al. (2016). Contrary to expectations,
however, holding all other variables constant, the percentage of living area (P LIVING) was negatively associated with daytime LSTs in
our model. P LIVING is positively associated (Pearson’s correlation
metric equals 0.2) with the percentage of the parcel covered by a
pool, for which our models did not control. Thus the lurking effect
of pool size offers a plausible explanation for the significance of
P LIVING.
Land configuration is significant but proves less important than
composition, for both daytime and nighttime LST (Tables 4 and 5).

This result is based on standardized coefficients and more importantly, on cross-validation experiments where models that did not
include each variable type were evaluated in terms of their predictive performance. Grass cover decreases nighttime LST more
effectively, however, when patches are clustered; the same association has been reported in analyses of daytime LSTs in Phoenix
(Myint et al., 2013). The magnitude of the influences of spatial
pattern on daytime LST varied significantly across neighborhoods,
consistent with the results in Zheng et al. (2014). The degree of
patchiness of all cover types affects daytime LST, either positively or
negatively, depending on the land cover. This result complements
findings for early nighttime LSTs in Phoenix (Connors et al., 2013).
Interestingly, shape complexity plays an important role as well.
Increasing PD maintains warmer daytime LSTs, whereas increases
in LSI relate to cooler results, as do pools.
It was expected that high income parcels, those near their city
cores, and communities with older median age of population would

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

117

Fig. 5. Distributions of weighted (based on parcel size) mean absolute error (WMAE, top), weighted mean absolute percentage error (WMAPE, middle) and weighted mean
error (WME, bottom) across neighborhoods, for daytime- and nighttime-specific linear mixed effects models.

have cooler LST, both day and night. These expectations followed
from the association of more vegetation cover, especially grass
and tress/shrubs, with older, more expensive neighborhoods based
“desert oasis” or mesic landscaping, and more swimming pools,
on average, linked to the capacity to afford them. In contrast, low
income parcels and parcels near their city fringe were expected to
have a positive impact on nighttime and daytime LST. The former
result follows from the documented decrease in vegetation cover
among low-income parcels (Declet-Barreto, Brazel, Martin, Chow,
& Harlan, 2013; Harlan, Brazel, et al., 2006) and the latter, perhaps,
due the prevalence of “xeric-scape” parcels among the newer residential developments on the outer edges of the metroplex. Such
explanatory variables were significant only at the first stage of our
analyses, which did not account for nested, neighborhood-specific
data structures and spatially correlated parcels within the same
neighborhood. Our statistical models suggest that at the parcel

level, given information on land composition and configuration,
socioeconomic variables are not significant in explaining the variability of daytime or nighttime LST. This finding does not contradict
other work indicating that lower income areas tend to correspond
to higher LST or UHI hotspots, such as those generated by the
high proportion of impervious surfaces in low income areas of
Baltimore, MD (Huang, Zhou, & Cadenasso, 2011) or the low proportion of vegetation cover common to Hispanic neighborhoods in
Phoenix (e.g., Harlan, Budruk, Gustafson, Larson, & Ruddell, 2006).
Rather, landscaping is a more a proximate factor influencing LST
than are socioeconomic variables (Li et al., 2016), and our level
of detail concerning land-cover mutes the predictive performance
of socioeconomic variables. Indeed, a recent article that focused
on the Baltimore area (census block level) found that land cover
had a stronger association with LST than other classes of variables,

118

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

but that social variables apparently affect land cover, and thus LST
(Huang & Cadenasso, 2016).
Our best performing model for nighttime LST did not identify
a significant effect for the percentage of residential building cover
(PLAND BUILDING), given information on the presence of a pool
and the percentage and the shape index of grass. The less parsimonious model, however, indicates a negative association between
residential building cover and LST. In contrast, Myint et al. (2013)
found a positive correlation between them. The different results
may reflect the different spatial scales and areas of assessment used
to determine LST. Myint et al. (2013) employed 90 m ASTER-derived
LST data to address segments of the urban area, not individual
parcels. In addition, they combined all water sources (i.e., pools and
ponds) and vegetation in a city segment regardless of their presence on individual parcels. In contrast, our study employs 6.8 m
MASTER-derived LST data, which can capture roof areas from the
remainder of the land covers of parcels, and did not include common water bodies and vegetation, which is not part of analyzed
parcels. Our disaggregated model result is consistent with that from
Atlanta, GA (USA) in which smaller buildings (e.g., residential units)
were found to be beneficial for reducing nighttime temperatures
because roofs release thermal energy during this time (Chun &
Guhathakurta, 2016).

4.1. Urban management and planning implications
It is increasingly recognized that the relationship between LST
and land architecture has important implications for mitigating the
SUHI effects (Weng et al., 2004; Hamada and Ohta, 2010; Zhou et al.,
2011; Jenerette et al., 2016; Li et al., 2012; Li et al., 2016), with
obvious implications for water withdrawals (Wentz et al., 2016),
especially in hot, arid environments. Our study suggests that the
effectiveness of land cover in this mitigation involves a full array of
urban land covers, even at the fine-grain (i.e., parcel) level. While
land-cover composition, such as the amount of green space, is the
most important factor in this mitigation, it can be enhanced by
patch clustering and shape compactness, depending on the land
cover. An initial lesson appears to be that compactly shaped and
concentrated patches of vegetation, foremost grass, is especially
important in decreasing nighttime LST, whereas parcels maintaining larger land-cover units of irregular shape decrease daytime
LST. Balancing the spatial distribution of land-cover types and optimizing their spatial patterns and shape appears to be a means to
reducing LST among SFR parcels, or at least some of its consequences, and if replicated sufficiently, at the neighborhood level
and beyond.
An open question is the consistency of the configuration outcomes across different spatial scales. Are residential level outcomes
the same for ascending units of urban space? This question
demands more attention as does the optimal design of land covers in regard to the tradeoffs between daytime and nighttime LST.
In addition, land architecture has large impacts on water use and
runoff, aquifer recharge, biological corridors, and urban aesthetics, among other environmental services, and urban sustainability
involves tradeoffs among these and other services, not just attention to the UHI effect. In the Phoenix case, the strong mitigation
role of grass holds tradeoffs with large water requirements in a
water-sparse environment. Regardless, research on the relationship between land architecture and LST indicates that land-cover
configuration plays an important role in the functioning of these
services, and the search for design principles for new residences
and reconfiguration of older parcels to mitigate extreme heat or
other disservices warrant attention. Ultimately, of course, these
considerations need to address above ground temperature.

4.2. Limitations
This research has enlarged the number and diversity of neighborhoods and parcels examined compared to other studies, and our
assessment included both daytime and nighttime LST. Yet, a more
complete aerial coverage and a larger number of days on which
fine-grain LSTs could be generated may provide more revealing
results. While the method employed to determine emissivity and,
hence, “corrected” LST is common to studies of this kind, anisotropy
errors may be created by the 3D characteristics of the land architecture not considered here, such as varying canopy levels and housing
density by neighborhoods (Krayenhoff & Voogt 2016). Despite the
fine-grain of our LST (6.8 m) and land-cover data (1 m), the spatial
matching involves modest aggregation and averaging of LST per
pixel. As well, access to parcel specific socio-economic data may
have altered our results. Finally, we only examined the relationship
of LST and land architecture for SFR parcels. Other parcel types, such
as multiple family residential, commercial, and recreational parcels
or their bundling were not addressed. Future research will extend
study sites beyond SFR.
5. Concluding remarks
Understanding the relationship between land architecture and
SUHI effects is important for mitigating extreme heat and its consequences, especially in desert cities but increasingly for cities
residing in the tropics to mid-latitudes where global warming is
expected to increase extreme heat events. Our research, using
fine resolution, parcel level data and linear mixed effects models, corroborates emerging research findings elsewhere that both
the composition and configuration of land systems affects LST,
contributing to SUHI. It elaborates those findings, however, by
demonstrating that (1) controlling for land composition, land configuration at the parcel level proves to be a significant factor
affecting LST, and (2) socioeconomic neighborhood characteristics
are not significant in explaining daytime and nighttime LSTs for
SFRs, given information on land composition and configuration.
Such controls also indicate that the interplay among explanatory
variables results in their different directional influences on LST at
different times of day, as well as by different land architecture metrics. These results point to characteristics of individual land covers
and their parcel aggregates that ameliorate LST, foremost during
the night. Taken together, they indicate the need for more research
on the fine-grain land architecture of the “city-scape”, its impacts of
the SUHI, and its possible use to reshape individual parcels, neighborhoods, and urban areas to achieve a more temperature friendly
outcome for desert cities.
Acknowledgements
This research was supported by the National Science Foundation
(NSF) under Grant No. SES-0951366, NSF DNS Grant No. 1419593
and USDA NIFA Grant No. 2015-67003-23508, and by the Central Arizona-Phoenix Long-Term Ecological Research project (NSF
BCS-1026865). The research was undertaken in the Environmental
Remote Sensing and Geoinformatics Lab with additional support
from the Gilbert F. White Fellowship, Arizona State University.
References
Bernstein, L. S., Adler-Golden, S. M., Sundberg, R. L., Levine, R. Y., Perkins, T. C., Berk,
A., et al. (2005). Validation of the quick atmospheric correction (QUAC)
algorithm for VNIR-SWIR multi- and hyperspectral imagery. SPIE proceedings,
algorithms and technologies for multispectral, hyperspectral, and ultraspectral
imagery proceedings of SPIE, 5806, 668–678.
Brown, R. D., & Corry, R. C. (2011). Evidence-based landscape architecture: The
maturing profession. Landscape and Urban Planning, 100, 327–329.

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120
Buyantuyev, A., & Wu, J. (2009). Urbanization alters spatiotemporal patterns of
ecosystem primary production: A case of the Phoenix metropolitan area, USA.
Journal of Arid Environments, 73.
Chang, Y., Han, C., Fan, K., Chen, S. K., & Chang, J. (2002). A modular eigen subspace
scheme for high-dimensional data classification with NASA MODIS/ASTER
(MASTER) airborne simulator data sets of Pacrim II project. Imaging
spectrometry VIII proceedings of SPIE, 4816, 426–436.
Chow, W. T. L., Chuang, W., & Gober, P. (2012). Vulnerability to extreme heat in
metropolitan phoenix: Spatial, temporal, and demographic dimensions. The
Professional Geographer, 64, 286–302.
Chun, B., & Guhathakurta, S. (2016). The impacts of three-dimensional surface
characteristics on urban heat islands over the diurnal cycle. Professional
Geographer, http://dx.doi.org/10.1080/00330124.2016.1208102
City of Phoenix. (2010). Tree and shade master plan. http://www.phoenix.gov/
parkssite/Documents/0719557.pdf
Clark, W. C. (2007). Sustainability science: A room of its own. Proceedings of the
National Academy of Sciences, 104, 1737–1738.
Connors, J. P., Galletti, C. S., & Chow, W. T. (2013). Landscape configuration and
urban heat island effects: Assessing the relationship between landscape
characteristics and land surface temperature in Phoenix, Arizona. Landscape
Ecology, 28, 271–283.
Coseo, P., & Larsen, L. (2014). How factors of land use/land cover, building
configuration, and adjacent heat sources and sinks explain Urban Heat Islands
in Chicago. Landscape and Urban Planning, 125, 117–129.
Cox, K. R. (1969). The voting decision in a spatial context. Progress in Geography, 1,
81–117.
Declet-Barreto, J., Brazel, A. J., Martin, C. A., Chow, W. T., & Harlan, W. T. (2013).
Creating the park cool island in an inner-city neighborhood: Heat mitigation
strategy for Phoenix, AZ. Urban Ecosystems, 16, 617–635.
Donovan, G. H., & Butry, D. T. (2009). The value of shade: Estimating the effect of
urban trees on summertime electricity use. Energy and Buildings, 41, 662–686.
Erell, E., Pearlmutter, D., & Williamson, T. (2012). Urban microclimate: Designing the
spaces between buildings. New York: Routledge.
Fan, C., Myint, S. W., & Zheng, B. (2015). Measuring the spatial arrangement of
urban vegetation and its impacts on seasonal surface temperatures. Progress in
Physical Geography, 39, 119–121.
Fast, J. D., Torcolini, J. C., & Redman, R. (2005). Pseudo vertical temperature profiles
and the urban heat island measured by a temperature datalogger network in
Phoenix, Arizona. Journal of Applied Meteorology, 44, 3–13.
Finch, W. H., Bolin, J. E., & Kelley, K. (2014). Multilevel modeling using R. Boca Raton,
FL: CRC Press.
Forman, R. T. T. (1990). Ecologically sustainable landscapes: the role of spatial
configuration. In I. S. Zonnelfeld, & R. T. T. Forman (Eds.), Changing landscapes:
An ecological perspective (pp. 261–278). New York: Springer.
Forman, R. T. T. (2014). Urban ecology: Science of cities. Cambridge: Cambridge
Cniversity Press.
Gago, E. J., Roldán, J., Pacheco-Torres, R., & Ordoñez, J. (2013). The city and urban
heat islands: A review of strategies to mitigate adverse effects. Renewable and
Sustainable Energy Reviews, 25, 749–758.
Giridharan, R., Lau, S. S. Y., Ganesan, S., & Givoni, B. (2007). Urban design factors
influencing heat island intensity in high-rise high-density environments of
Hong Kong. Building and Environment, 42, 3669–3684.
Grossman-Clarke, S., Zehnder, J. A., Loridan, T., & Grimmond, C. S. B. (2010).
Contribution of land use changes to near-surface air temperatures during
recent summer extreme heat events in the Phoenix metropolitan area. Journal
of Applied Meteorology and Climatology, 49, 1649–1664.
Grossman-Clarke, S., Schubert, S., Clarke, T. A., & Harlan, S. L. (2014). Extreme
summer heat in Phoenix, Arizona (USA) under global climate change
(2041–2070). Die Erde–Journal of the Geographical Society of Berlin, 145, 49–61.
Hamada, S., & Ohta, T. (2010). Seasonal variations in the cooling effect of urban
green areas on the surrounding urban areas. Urban Forestry and Urban
Greening, 9, 15–24.
Harlan, S. L., Brazel, A. J., Prashad, L., Stefanov, W. L., & Larsen, L. (2006).
Neighborhood microclimates and vulnerability to heat stress. Social Science &
Medicine, 63, 2847–2863.
Harlan, S. L., Budruk, M., Gustafson, A., Larson, K., Ruddell, D., et al. (2006).
Highlights phoenix area social survey: Community and environment in a desert
city. Tempe, AZ: Global Institute of Sustainability, Arizona State University.
Contribution no. 4. CAP-LTER.
Hart, M. A., & Sailor, D. J. (2009). Quantifying the influence of land-use and surface
characteristics on spatial variability in the urban heat island. Theoretical and
Applied Climatology, 95, 397–406.
Hawkins, T. W., Brazel, A. J., Stefanov, W. L., Bigler, W., & Saffell, E. M. (2004). The
role of rural variability in urban heat island determination for Phoenix,
Arizona. Journal of Applied Meteorology, 43, 476–486.
Hook, J. S., Myers, J. J., Thome, J. K., Fitzgerald, M., & Kahle, B. A. (2001). The
MODIS/ASTER airborne simulator (MASTER) – A new instrument for earth
science studies. Remote Sensing of Environment, 76, 93–102.
Huang, G., & Cadenasso, M. L. (2016). People, landscape and urban heat islands:
Dynamics among neighborhood social conditions, land cover, and surface
temperatures. Landscape Ecology, 31, 2507-25-15.
Huang, L., Li, J., Zhao, D., & Zhu, J. (2008). A fieldwork study on the diurnal changes
of urban microclimate in four types of ground cover and urban heat island of
Nanjing, China. Building and Environment, 43, 7–17.
Huang, G., Zhou, W., & Cadenasso, M. L. (2011). Is everyone hot in the city? Spatial
pattern of land surface temperatures, land cover and neighborhood

119

socioeconomic characteristics in Baltimore, MD. Journal of Environmental
Management, 92, 1753–1759.
James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical
learning. Berlin: Springer.
Jenerette, G. D., Harlan, S. L., Stefanov, W., & Martin, C. (2011). Ecosystem services
and urban heat riskscape moderation: Water, green spaces, and social
inequality in Phoenix, USA. Ecological Applications, 21, 2637–2651.
Jenerette, G. D., Harlan, S. L., Buyantuev, A., Stefanov, W., Declet-Barreto, J., Ruddell,
B. L., et al. (2016). Micro-scale urban surface temperatures are related to
land-cover features and residential heat related health impacts in Phoenix, AZ,
USA. Landscape Ecology, 31, 745–760.
Johnston, R., Propper, C., Sarker, R., Jones, K., Bolster, A., et al. (2005).
Neighbourhood socialcapital and neighbourhood effects. Environment and
Planning A, 37, 1443–1459.
Keys, E., Wentz, E. A., & Redman, C. L. (2007). The spatial structure of land use from
1970 to 2000 in the Phoenix, Arizona, metropolitan area. Professional
Geographer, 59, 131-147.
Kleerekoper, L., van Esch, M., & Baldiri Salcedo, T. (2012). How to make a city
climate-proof, addressing the urban heat island effect. Resources, Conservation
and Recycling, 64, 30–38.
Krayenhoff, E. S., & Voogt, J. A. (2016). Daytime thermal anisotropy of urban
neighborhoods: morphological causation. Remote Sensing, 8, 108.
Kutner, M., Nachtsheim, C., & Neter, J. (2004). Applied linear regression models (4th
ed.). McGraw-Hill.
Li, J. J., Wang, X. R., Wang, X. J., Ma, W. C., & Zhang, H. (2009). Remote sensing
evaluation of urban heat island and its spatial pattern of the Shanghai
metropolitan area, China. Ecological Complexity, 6, 413–420.
Li, X., Zhou, W., Ouyang, Z., Xu, W., & Zheng, H. (2012). Spatial pattern of
greenspace affects land surface temperature: Evidence from the heavily
urbanized Beijing metropolitan area China. Landscape Ecology, 27, 887–898.
Li, X., Li, W., Middel, A., Harlan, S. L., Brazel, A., & Turner, B. L., II. (2016). Remote
sensing of the surface urban heat island and land architecture in Phoenix,
Arizona: Combined effects of land composition and configuration and
cadastral-demographic-economic factors. Remote Sensing of Environment, 174,
233–243.
Li, D., Bou-Zeid, E., & Oppenheimer, M. (2014). The effectiveness of cool and green
roofs as urban heat island mitigation strategies. Environmental Research Letters,
9, 055002.
Li, X., Myint, S. W., Zhang, Y., Galletti, C., Zhang, X., & Turner, B. L., II. (2014).
Object-based land-cover classification for metropolitan phoenix, Arizona,
using aerial photography. International Journal of Applied Earth Observations
and GeoinSformation, 33, 321–330.
McCarthy, M. P., Best, M. J., & Betts, R. A. (2010). Climate change in cities due to
global warming and urban effects. Geophysical Research Letters, 37, 9.
McGarigal, K., & Marks, B. J. (1995). FRAGSTATS: Spatial pattern analysis program for
quantifying landscape structure. Portland (OR): USDA Forest Service, Pacific
Northwest Research Station. General Technical Report PNW-GTR-351.
McGarigal, K., Cushman, S. A., & Ene, E. (2012). FRAGSTATS v4: Spatial pattern
analysis program for categorical and continuous maps. Amherst, MA: University
of Massachusettes. URL http://www.umass.edu/landeco/research/fragstats/
fragstats.html
Middel, A., Brazel, A. J., Kaplan, S., & Myint, S. W. (2012). Daytime cooling efficiency
and diurnal energy balance in Phoenix, Arizona, USA. Climate Research, 54,
21–24.
Middel, A., Häb, K., Brazel, A., Martin, C., & Guhathakurta, S. (2014). The impact of
urban form and design on mid-afternoon microclimate in Phoenix local
climate zones. Landscape and Urban Planning, 122, 16–28.
Mitchell, M. G. E., Bennett, E. M., & Gonzalez, A. (2013). Linking landscape
connectivity and ecosystem service provision: Current knowledge and
research gaps. Ecosystems, 16, 894–908.
Myint, S. W., Wentz, E. A., Brazel, A. J., & Qualttrochi, D. A. (2013). The impact of
distinctive anthropogenic and vegetation features on urban warming.
Landscape Ecology, 28, 959–978.
Myint, S. W., Zheng, B., Talen, E., Fan, C., Kaplan, S., Middel, A., et al. (2015). Does
the spatial arrangement of urban landscape matter? Examples of urban
warming and cooling in Phoenix and Las Vegas. Ecosystem Health and
Sustainability, 1, 1–15.
Myint, S. W. (2012). Effects of the spatial pattern of vegetation cover on urban
warming in a desert city. In X. Yang, & J. Li (Eds.), Advances in mapping from
aerospace imagery: Techniques and applications (pp. 261–278). Boca Raton, FL:
CRC Press.
Oke, T. R. (1981). Canyon geometry and the nocturnal urban heat island:
Comparison of scale model and field observations. Journal of Climatology, 1,
237–254.
Pinheiro, J. C., & Bates, D. M. (2009). Mixed effects models in S and S-plus. Berlin:
Springer.
Saha, M. V., Davis, R. E., & Hondula, D. M. (2014). Mortality displacement as a
function of heat event strength in 7 US cities. American Journal of Epidemiology,
179, 467–474.
Salamanca, F., Georgescu, M., Mahalov, A., Moustaoui, M., & Wang, M. (2014).
Anthropogenic heating of the urban environment due to air conditioning.
Journal of Geophysical Research, 119, 5949–5965.
Santamouris, M. (2013). Using cool pavements as a mitigation strategy to fight
urban heat island—A review of the actual developments. Renewable and
Sustainable Energy Reviews, 26, 224–240.

120

X. Li et al. / Landscape and Urban Planning 163 (2017) 107–120

Seto, K., & Reenberg, A. (Eds.). (2014). Rethinking global land use in an urban era.
Cambridge, MA: MIT Press.
Shashua-Bar, L., Pearlmutter, D., & Erell, E. (2009). The cooling efficiency of urban
landscape strategies in a hot dry climate. Landscape and Urban Planning, 92,
179–186.
Shashua-Bar, L., Pearlmutter, D., & Erell, E. (2011). The influence of trees and grass
on outdoor thermal comfort in a hot-arid environment. International Journal of
Climatology, 31, 1498–1506.
Silva, H. R., Phelan, P. E., & Golden, J. S. (2010). Modeling effects of urban heat
island mitigation strategies on heat-related morbidity: A case study for
Phoenix, Arizona, USA. International Journal of Biometeorology, 54, 13–22.
Solecki, W. D., Rosenzweig, C., Parshall, L., Pope, G., Clark, M., Cox, J., et al. (2005).
Mitigation of the heat island effect in urban New Jersey. Environmental
Hazards, 6, 39–49.
Stewart, I. D., & Oke, T. R. (2012). Local climate zones for urban temperature
studies. Bulletin of the American Meteorological Society, 93, 1879–1900.
Su, W., Gu, C., & Yang, G. (2010). Assessing the impact of land use/land cover on
urban heat island pattern in Nanjing City, China. Journal of Urban Planning and
Development, 136, 365–372.
Taha, H. (1997). Urban climates and heat islands: Albedo, evapotranspiration, and
anthropogenic heat. Energy and Buildings, 25, 99–103.
Turner, B. L., II, Lambin, E. F., & Reenberg, A. (2007). The emergence of land change
science for global environmental change and sustainability. Proceedings of the
National Academy of Sciences, 104, 20666–20671.
Turner, B. L., II, Janetos, A. C., Verbug, P. H., & Murray, A. T. (2013). Land system
architecture: Using land systems to adapt and mitigate global environmental
change. Global Environmental Change, 23, 395–397.

Turner, B. L., II. (2017). Land system architecture for urban sustainability: New
directions for land system science illustrated by application to the urban heat
island problem. Journal of Land Use Science, 12 [forthcoming].
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, 467–483.
Wentz, E. A., Rode, S., Li, X., Tellman, E. M., & Turner, B. L., II. (2016). Impact of
homeowner association (HOA) landscaping guidelines on residential water
use. Water Resources Research, 52, 3373–3386.
Wu, J., Jenerette, G. D., Buyantuyev, A., & Redman, C. L. (2011). Quantifying
spatiotemporal patterns of urbanization: The case of the two fastest growing
metropolitan regions in the United States. Ecological Complexity, 8, 1–8.
Wu, J. (2013). Landscape sustainability science: Ecosystem services and human
well-being changing landscapes. Landscape Ecology, 28, 999–1023.
Yu, K. J., Li, D. H., & Ji, Q. P. (2001). Ecological design for landscape and city:
Concepts and principles. Chinese Landscape Architecture, 6, 3–10.
Zhang, H., Qi, Z. F., Ye, X. Y., Cai, Y. B., Ma, W. C., & Chen, M. N. (2013). Analysis of
land use/land cover change, population shift, and their effects on
spatiotemporal patterns of urban heat islands in metropolitan Shanghai, China.
Applied Geography, 44, 121–133.
Zheng, B., Myint, S. W., & Fan, C. (2014). Spatial configuration of anthropogenic land
cover impacts on urban warming. Landscape and Urban Planning, 130, 104–111.
Zhou, W., Huang, G., & Cadenasso, L. M. (2011). Does spatial configuration matter?
Understanding the effects of land cover pattern on land surface temperature in
urban landscapes. Landscape and Urban Planning, 102, 54–63.
Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American
Statistical Association, 101, 1418–1429.
Zuur, A., Ieno, E. N., Walker, N., Savaliev, A. A., & Smith, G. M. (2009). Mixed effects
models and extensions in ecology with R. Berlin: Springer.