CLIMATE RESEARCH
Clim Res

Vol. 33: 171–182, 2007

Published February 22

Determinants of changes in the regional urban
heat island in metropolitan Phoenix (Arizona, USA)
between 1990 and 2004
Anthony Brazel1,*, Patricia Gober1, 2, Seung-Jae Lee2, Susanne Grossman-Clarke3,
Joseph Zehnder1, 3, 4, Brent Hedquist1, Erin Comparri2
1

School of Geographical Sciences, Arizona State University, Tempe, Arizona 85287-0104, USA
Decision Center for a Desert City, Arizona State University, Tempe, Arizona 85287-8209, USA
3
Global Institute of Sustainability, Arizona State University, Tempe, Arizona 85287-3211, USA
4
Department of Mathematics, Arizona State University, Tempe, Arizona 85287-0104, USA

2

ABSTRACT: We investigated the spatial and temporal variation in June mean minimum temperatures for weather stations in and around metropolitan Phoenix, USA, for the period 1990 to 2004.
Temperature was related to synoptic conditions, location in urban development zones (DZs), and the
pace of housing construction in a 1 km buffer around fixed-point temperature stations. June is typically clear and calm, and dominated by a dry, tropical air mass with little change in minimum temperature from day to day. However, a dry, moderate weather type accounted for a large portion of the
inter-annual variability in mean monthly minimum temperature. Significant temperature variation
was explained by surface effects captured by the type of urban DZ, which ranged from urban core
and infill sites, to desert and agricultural fringe locations, to exurban. An overall spatial urban effect,
derived from the June monthly mean minimum temperature, is in the order of 2 to 4 K. The cumulative housing build-up around weather sites in the region was significant and resulted in average
increases of 1.4 K per 1000 home completions, with a standard error of 0.4 K. Overall, minimum temperatures were spatially and temporally accounted for by variations in weather type, type of urban
DZ (higher in core and infill), and the number of home completions over the period. Results compare
favorably with the magnitude of heating by residential development cited by researchers using
differing methodologies in other urban areas.
KEY WORDS: Urban heat island · Phoenix · Land use change · Urban fringe · Housing development ·
Statistical approach
Resale or republication not permitted without written consent of the publisher

Metropolitan Phoenix’s urban heat island (UHI), the
phenomenon of warmer temperatures at the core of
the built-up urban area compared with the surrounding rural countryside, has been well documented in the
scientific literature (Balling & Brazel 1987, Brazel et al.
2000, Baker et al. 2002, Hawkins et al. 2004, Fast et
al. 2005). Large-scale, rapid urbanization (from fewer
than 1 million people in 1970 to almost 4 million today)
has typically increased the warm season minimum

temperature by 5 K and the daily temperature by more
than 3 K in a desert environment that is naturally hot
and was unfit for large-scale human habitation before
the popularization of air conditioning during the 1950s
(Cooper 1998, Baker et al. 2002). There is an urgent
need among planners and city officials for a more practical understanding of the effects of new construction
on the local climate because warmer temperatures
reduce human comfort, increase energy and water use,
and compromise the region’s capacity to market itself
as a year-round tourist destination.

*Email: abrazel@asu.edu

© Inter-Research 2007 · www.int-res.com

1. INTRODUCTION

172

Clim Res 33: 171–182, 2007

Oke (2006) represented the urban-rural continuum
with 7 ‘urban climate zones’ and theorized their effects
on climate via variation in roughness, aspect ratio, and
percent built area. Following this logic, but adapting
the idea to the built environment and settlement
history of Phoenix, we divided the area into 5 development zones (DZs): core, infill, agricultural fringe,
desert fringe, and exurban (hereafter labeled CORE,
INFILL, AGFRINGE, DESFRINGE, and EXURBAN)
and then empirically investigated the effects of residential construction and synoptic conditions on June
mean minimum temperatures between 1990 and 2004.
Using a data set that combines information from 4 historical weather networks and annual information on
new housing construction from the Maricopa Association of Governments (MAG, the local council of governments), we evaluated whether DZ affects temperature, the changes in temperature over the study period,
and the way new home construction and synoptic conditions affect temperature. Resulting coefficient estimates will enable local planners to better understand
the climatic effects of new home development in different zones of the city.
Our research was undertaken under the auspices of
a National Science Foundation Project, the Decision
Center for a Desert City (DCDC), that is charged with
producing scientific results and decision tools that support better water management decisions in the face of
growing climatic uncertainty. Although DCDC’s mission is primarily concerned with long-term climate
change and annual climate variability, potential interactions between the UHI and regional and global climatic processes and the implications of these processes for water demand and supply also represent
important avenues of research. Spatio-temporal analyses of the heat island and its changes over time are
vital inputs to other studies of the effects of rising
nighttime temperatures on water use, including
impacts on lake and pool evaporation.

2. LITERATURE
The UHI is the most intensely studied climate feature
of cities (Arnfield 2003, Souch & Grimmond 2006).
Generally, researchers use fixed weather stations or
mobile transects to study the spatial and temporal air
temperature structure of cities. Oke (2006) called for
a standardization of methods and communication in
urban climatology, and argued that any new placement or use of ongoing fixed stations or mobile surveys
in an urban environment for analysis of the UHI should
consider (1) the general urban zone within which a site
is located, (2) the setting of the immediate site and its
surroundings, and (3) the layer of atmosphere being

observed, e.g. canopy, roughness sublayer, surface, or
outer layer. Grimmond (2006) noted that there are a
multitude of data consistency issues — including the
exposure of sites, data quality, and station representativeness — in a historical weather network. As a result,
there is a continued need to evaluate temperature
fields relative to the role of surface/building geometry,
land use, vegetation, and patterns of anthropogenic
heat release.
Recent research summarized by Souch & Grimmond
(2006) shows that the UHI is primarily a nighttime phenomenon, can occur throughout the year, is dependent
on weather conditions (wind, cloud cover, etc.), and
generally consists of higher temperatures in the urban
core and commercial locations, with lower temperatures in residential and rural sites. Previous research
on residential areas and changes in air temperature
suggest that single-family housing developments raise
nighttime temperatures by 2 to 5 K above rural background values (e.g. Kuttler et al. 1996, Svensson 2004,
Unger 2004, Grossman-Clarke et al. 2005, Alcoforado
& Andrade 2006, Hedquist & Brazel 2006, Hartz et al.
2006). Our approach was to evaluate spatio-temporal
patterns of temperature, by developing relationships
among factors that are simultaneously operating at
both local (within the city and in rural areas) and
regional (climate setting and terrain) scales to gain further insights into the spatial distribution of temperature trends across an urban environment that is both
infilling and spreading outward.
The rapidly urbanizing region of central Arizona is
an excellent site for evaluating spatio-temporal trends
in temperature because of the magnitude of recent
growth and urban development, and because of the
variety of environments in which this new development occurs (Gammage 1999, Gober 2006). Growth
occurs from infilling of the already built-up area and
from development at the urban fringe. Moreover,
despite the popular image of Phoenix as the epitome of
a sprawling metropolis, new residential densities are
higher today than in the recent past because of rapid
growth, growing land prices, and intense competition
for land. According to our calculations based on files of
the Maricopa County Assessors’ Office, the average lot
size of a single-family home constructed in the 1970s
was 784 m2 (8434 ft2), which reduced to 772 m2 (8308 ft2)
during the 1980s, 696 m2 (7482 ft2) during the 1990s,
and 688 m2 (7400 ft2) between 2000 and 2003. This
means that the built environment does not correspond
to traditional expectations of declining densities with
distance from the urban core; there is a more even distribution of densities across the Phoenix metropolitan
area than Oke’s (2006) model would suggest.
Previous empirical and modeling studies of Phoenix’s
UHI were either short-term in nature, focused on the

173

Brazel et al.: Urban heat island in Phoenix

300 to 450 m. For the array of sites used, no significant
spatial attributes of the heat island (e.g. Brazel et al.
correlation was found between minimum temperature
1993, Hawkins et al. 2004, Fast et al. 2005, Grossmanand elevation within the station network (r2 = 0.07),
Clarke et al. 2005) or limited to a coarse urban-rural
dichotomy when specifying temporal dimensions (e.g.
and thus this was not specifically included in our subBrazel et al. 2000, Baker et al. 2002, Stabler et al. 2005).
sequent multivariate analysis. Data obtained from
No study has made use of the annual housing time
these networks have been quality controlled (by netseries database to link air temperature patterns to the
work specialists or by the state climate office), then
process of urbanization as manifested in new housing
systematically organized in a GIS. With the exception
construction. Our further contribution is in
Table 1. Weather stations used in statistical analysis. See Fig. 1 for locations.
the simultaneous consideration of space
AGFRINGE: agricultural fringe; DESFRINGE: desert fringe; AZMET: Ariand time, and in the choice of urban zones
zona Meteorological Network; MCFCD: Maricopa County Flood Control
that reflect the urban growth processes of
District; PRISMS: Phoenix Real-Time Instrumentation for Surface MeteoroPhoenix in particular, and North American
logical Studies; NOAA: National Oceanic Atmospheric Administration
cities more generally.
Development zone
(DZ)

Elevation
(m)

Cumulative homes
in 1 km buffer
1990–2004

Network

EXURBAN
Paloma
Queen Creek
Bartlett Dam
Falcon
Stewart Mountain Dam

219
430
503
417
378

0
26a
0
4a
0

AZMET
AZMET
NOAA
PRISMS
PRISMS

AGFRINGE
Litchfield
Waddell
Crossroads
Chandler heights
Laveen 3S
Litchfield Park
Collier
Rittenhouse

309
407
387
433
342
314
309
426

1480
39
9680
2030
9770
3200
5950
31

AZMET
AZMET
MCFCD
NOAA
NOAA
NOAA
PRISMS
PRISMS

DESFRINGE
Carefree Ranch
Thunderbird
Carefree
Fountain
Pima
Sun Lakes

976
436
772
482
665
364

2130
45
1010
3160
1510
8270

MCFCD
MCFCD
NOAA
PRISMS
PRISMS
PRISMS

INFILL
Phoenix Greenway
Scottsdale
ACDC
Buckeye
South Phoenix
Youngtown
Corbell
Sheely
Stapley

401
469
372
305
354
345
369
328
383

80
2310
2100
1300
74
1590
5690
1520
74

AZMET
AZMET
MCFCD
NOAA
NOAA
NOAA
PRISMS
PRISMS
PRISMS

CORE
Phoenix Encanto
Durango Complex
IBW
Alameda
Arcadia
Pera
Pringle
Sky Harbor
Tempe ASU

335
320
377
360
380
385
376
344
357

5a
6a
33a
9a
17a
1a
35a
0
15a

AZMET
MCFCD
MCFCD
PRISMS
PRISMS
PRISMS
PRISMS
NOAA
NOAA

3. DATA AND METHODS
Fixed point, surface weather station network data were used and co-located in a
Geographical Information System (GIS)
classified by DZ type, together with a spatial housing completion data set. In addition, a synoptic variable was developed
for the study period of 1990–2004. The
sub-sections below describe the data, variables, and methods used in the statistical
analyses.

3.1. Weather stations
Weather network data of air temperatures within the urban canopy layer (ca. 2 m
height) were obtained from 4 networks in
the metropolitan area: (1) Phoenix Realtime Instrumentation for Surface Meteorological Studies (PRISMS, Pon et al.
1998), (2) Arizona Meteorological Network
(AZMET, see http://ag.arizona.edu/azmet),
(3) Maricopa County Flood Control District
(MCFCD, see http://fcd.maricopa.gov), and
(4) over 20 co-op stations obtained from
the National Climatic Data Center of the
National Oceanic Atmospheric Administration (NCDC, NOAA; data available
from www.ncdc.noaa.gov and summarized
monthly at www.wrcc.dri.edu). Table 1
summarizes site name, elevation, numbers
of homes built within a diameter of 1 km of
a site, and the type of weather network.
Fig. 1 shows the distribution of stations and
their DZ code. Elevations of the stations
ranged from 219 m to over 976 m. However, most sites clustered near a range of

a

Homes completed were not within 200 m of immediate weather monitor
in the buffer zone

174

Clim Res 33: 171–182, 2007

Fig. 1. Weather network stations, coded by development zone (DZ). Filled symbols are those sites used in statistical analysis. Open symbols
are sites used additionally for kriging (Fig. 5). Inset shows general distribution of desert, agriculture, and urban lands in Phoenix. See Table 1
for site descriptions. AGFRINGE: agricultural fringe; DESFRINGE: desert fringe

of PRISMS, the network websites (above) contain
information on immediate site locales, photos, and
other metadata about the sites. Information on PRISMS
sites is available through the Office of the State Climatologist at Arizona State University, in cooperation with
the Salt River Project (an energy and power company),
which manages the monitoring sites (data are displayed at http://data.afws.org/sui/). For this study, we
restricted the analysis to a monthly time scale, and to a
time of year (June) during which Phoenix experiences
relatively hot, clear, calm conditions for most days, an
excellent time for UHI evaluation (Brazel et al. 2000).
Due to this desert area having over 200 clear days per

year and low winds, the UHI phenomenon is evident
on a year-round basis and June characteristically represents the nature of the spatial dimension of the UHI.
We encountered missing data for approximately
13% of the data points, and used spatio-temporal geostatistical tools to estimate the value of these observations by integrating general knowledge based on
background information and summary statistics, and
site-specific knowledge obtained through experience
with the specific situation (Christakos 2000, Christakos
et al. 2001). The majority of missing records occurred
near the beginning of the time period, because not all
sites extended back in time to June 1990. A Bayesian

Brazel et al.: Urban heat island in Phoenix

175

maximum entropy (BME) method was employed to
estimate missing data and involved 3 steps: (1) a structural stage leading to a prior probability density function (PDF) that uses the mean and covariance functions
to model space/time dependencies, trends, and the
basic physical structure of the variable (June minimum
temperature), (2) a specification stage in which BME
assesses data uncertainty in a specific set of sampling
points, and (3) an integration stage in which the prior
PDF is enriched by blending the 2 knowledge bases
through Bayesian conditionalization. The latter stage
produces a posterior PDF at all points of interest.
We modeled spatio-temporal variability in terms of a
covariance model with 2 exponential elements. The
first accounts for the short range of spatio-temporal
autocorrelation (i.e. 3 decimal degrees in space and
6.5 mo in time), which covers metropolitan Phoenix
and entails a seasonal time scale. The second accounts
for the long-range correlation (i.e. 46 decimal degrees
in space and 60 mo in time). The posterior PDF based
on space and time significantly increases the accuracy
of our predictions over a purely spatial analytic procedure. There may still be some uncertainty relative to
the station network comparability itself, but we consider that less than one-third of the stations (mostly the
NOAA network sites in Table 1) that we used would
have any problems in this regard.

3.2. MAG housing completion data
We used an archived data set available from the
MAG, which was based on single-family housing completions in metropolitan Phoenix for each year during
the period 1990 to 2004. A completion is recorded
when the certificate of occupancy is awarded by its
respective municipality. Municipalities in turn submit
the georeferenced information to MAG, and MAG
assembles the data region-wide and updates the
records quarterly. We recognize that the use of this
readily available data set acts as a surrogate for the full
landscape development process from place to place
(homes, streets, yards, etc.). The main advantage is
a consistent, complete annual data set for the entire
15 yr study period.
Maps of housing completions for three 5 yr periods
illustrate the areas of new housing completions during
the study period, the continuous outward movement of
the urban fringe of new development, and the relatively large geographic area in which new housing
construction is occurring (Fig. 2). The urban fringe in
Phoenix is not, as commonly conceived, a narrow band
of disturbed landscape, but rather a large area that is
gradually but continuously filled in (Gober & Burns
2002).

Fig. 2. Residential completions from 1990 to 2004 in 5 yr increments. Gray areas are new housing completions

176

Clim Res 33: 171–182, 2007

Also, using the housing data, we buffered each
weather station with a 1 km diameter buffer, and
extracted cumulative numbers of homes completed
within this buffer for each year from 1990 to 2004. We
performed this analysis in view of Oke (2006), who
suggested that influences closer to a site play a major
role in driving the variability observed at sites in urban
areas (Table 1). We acknowledge that our choice of a
1 km buffer relaxes an idealistic, more stringent rule
for urban site representativeness; however, it does
match the modelling resolution of the PSU/NCAR
mesoscale model (known as MM5) (Grossman-Clarke
et al. 2005), and allows for the fact that most network
weather sites have a core inner locale of the station
itself (ca. 25 × 25 m) before the surrounding land cover
starts a short distance away from the site, as well as
generally allowing for the capture of more of the
regional development process around weather sites.
None of the sites are in complex locales — close to
major urban canyons — where results would largely
depend on the exact placement of the sensor in that
environment.

3.3. Development zones (DZs)
Based on the housing completion maps and our general knowledge of land use patterns in the region, we
developed a DZ typology of weather stations that
included (1) stations in the inner CORE of the urbanized area where little new construction occurred, (2)
INFILL locations that were clearly part of the urbanized area in 1990 but became more densely settled by
2004, (3) AGFRINGE locations centered in areas of
new construction on former irrigated agricultural
lands, (4) DESFRINGE sites where new construction
has replaced open desert land, and (5) EXURBAN
locations outside the built-up urbanized area (Fig. 1).
The EXURBAN sites were mostly near agricultural
locations or dams and lakes owing to the ease of access
of these sites in a complex terrain considerably distant
to the city.
The typology of the weather station sites generally
relates to the recommended spatial classification of
urban forms arranged according to their ability to
impact local climate, as presented by Oke (2006). Our
EXURBAN category corresponds to Oke’s (2006) Type
7 category of scattered houses in natural land cover
area. Our AGFRINGE and DESFRINGE sites are rough
equivalents of Type 5 sites (medium, low density suburban), and our CORE and INFILL sites relate to Type 3
sites (highly developed, medium density urban) in
Oke’s (2006) classification. Our typology differs because it is based on changes in the built environment in
addition to the nature of built environment itself. We

note that little change occurred in the CORE and
EXURBAN sites, and that rapid change occurred in
INFILL locations and in the 2 urban fringe sites
AGFRINGE and DESFRINGE (Table 1).

3.4. Regional synoptic variable
We used a synoptic variable to account for inter-annual variation in temperatures experienced across the
weather network. Sheridan’s (2002) spatial synoptic
classification (SSC2) scheme was employed (available
at http://sheridan.geog.kent.edu/ssc.html, expressed
as results for the SSC2 classification). For June, 64% of
the days are typed dry tropical with light winds (< 5 m
s–1), and another 20% of days are associated with cool
air intrusions related to troughs that have developed in
the Western USA that reduce mean monthly minimum
temperature. In our analysis, the variable used was
the number of days in June that a weather type is
characterized by this second category, defined in SSC2
as Dry Moderate (later coded as DRYMOD). We conducted a correlation analysis of the frequency of days
in June classified as a particular weather type by SSC2
against June monthly mean minimum temperature.
The highest correlation (r2 = 0.75, inverse relationship)
was observed between the frequency of the DRYMOD
weather type per month and monthly mean minimum
temperature. In June, the recurrent Dry Tropical
weather type brings warm nights in general, as do
some transitional weather types and infrequent moist
weather types. Thus, during the study period, the dry
tropical frequencies per se only correlated with minimum temperature variation with an r2 of 0.09.
When using SSC2, we understand that one should
exercise caution because the SSC2 and the frequencies determined from it might conceivably contain an
‘urban effect’ as a consequence of the derivation of
weather types that enter the catalog. This could be so
because surface data are used in its derivation, often
obtained from major first-order weather sites within or
near urban areas. This would mean that there could be
discontinuities in the frequencies of various categories
of daily weather types. This could certainly be the case
if the main station used showed a very significant time
trend of weather elements attributable to local effects.
The Sky Harbor International Airport site (coded PHX
in the SSC2-derived data) may represent an example
of this influence. However, we consider that this is not
the case for Phoenix, and that our use of this factor in
this study is warranted on several grounds: (a) a
nearby station —Tucson, Arizona — shows similar temporal patterns and variability to Sky Harbor for the
June 1990 to 2004 period, and the Tucson airport locale
is entirely outside of any Tucson urban effect (Comrie

Brazel et al.: Urban heat island in Phoenix

2000), (b) independent analysis of time trends in temperature for the Sky Harbor Airport station for the
period 1990 to 2004 (and, in addition, our analysis
herein) shows no major increase or major trend in temperature (however, prior to 1990 there was a considerable trend in temperature at Sky Harbor Airport; see
e.g. Brazel et al. 2000, Baker et al. 2002), and (c) there
is no obvious time trend in the Dry Tropical or DRYMOD types revealed in the SSC2 catalog within the
period 1990 to 2004. Dry Tropical and DRYMOD days
are related to the frequency of ridging and troughing
in the Western USA during June, as determined by a
separate composite analysis that used the NCEP/NCAR
reanalysis data (www.cdc.noaa.gov/Composites/Day/).
As an illustration, we chose a mid-study period year
(1998) to demonstrate the SSC2 weather type and associated atmospheric conditions (Figs. 3 & 4). We com-

177

posited a sample run of days catalogued as DRYMOD
(June 7 to 10) and Dry Tropical (June 19 to 22) in the
SSC2, and displayed the 500 mb geopotential height
surface and 850 mb temperature pattern over the USA,
and Sky Harbor’s hourly air temperature and dew
point during these 2 periods. Generally, it holds true
that when DRYMOD is prevalent in the catalog,
troughing over the Southwestern USA occurs with
cooler upper level and surface temperatures. A more
thorough analysis of the SSC2 catalog values vs. synoptics and circulation characteristics for this time of
year is certainly warranted, but is beyond the scope of
this study. Here we simply chose to use the DRYMOD
daily weather type code (summed over a monthly time
scale) for the period 1990 to 2004 to evaluate the relative role of regional weather effects on the minimum
temperature from year to year.

Fig. 3. Examples of (a,c) dry moderate and (b,d) dry tropical weather type conditions over North America for selected periods
in June 1998. Data are 500 mb geopotential heights and 850 mb air temperature. Black box: study area

178

Clim Res 33: 171–182, 2007

The reduced-form ANCOVA allowed us to examine
whether intercepts vary significantly among the 5 DZs.
Assuming that there are significant differences, we
proceeded to estimate coefficients (αk and βl) that can
be used to predict changes in temperature for DRYMOD, HOUSING, AGFRINGE, DESFRINGE, CORE,
and INFILL. We excluded EXURBAN from this phase
of the analysis because it is known automatically
when AGFRINGE, DESFRINGE, CORE, and INFILL
are known. The constant in the main model could be
interpreted as representing estimated temperature at
EXURBAN sites not yet affected by the urbanization
process. The α coefficients for AGFRINGE, DESFRINGE, INFILL, and CORE represent the change
in temperature relative to EXURBAN sites.
Fig. 4. Hourly air temperature and dew point for periods
shown in Fig. 3, Phoenix Sky Harbor International Airport
Automated Surface Observing System (first-order weather
station)

3.5. Statistical model
Our analysis evaluated the effects of the cumulative
number of new homes constructed in year j (HOUSINGj), the number of days in a June month corresponding to the dry moderate weather conditions in year j
(DRYMODj ), and DZ (including AGFRINGE, DESFRINGE, CORE, EXURBAN, and INFILL) on mean
June minimum temperature in year j (TEMPj ). In the
first stage of the analysis, HOUSING and DRYMOD
were classified as covariates, whereas DZ was analyzed as a treatment effect (i.e. as a dummy variable).
Fitting the data to the full model allowed us to evaluate
interactions between DZ and the 2 confounding variables and test for the equality of slopes for each covariate. Our goal was to determine whether DZ affects the
way DRYMOD and HOUSING influence temperature.
After using a CONTRAST command in SAS version
9.0, we were able to conclude that there was no reason
to consider any interaction terms: the slope coefficients
of the 2 confounding variables do not differ significantly across DZs at the 0.05 level of significance (see
Table 2).
Based on this conclusion, we proceeded with the
reduced-form analysis of covariance (ANCOVA)
model expressed below in Eq. 1
4

2

k =1

l =1

TEMPij = μ + ∑ α k x kij + ∑ β l z lij + e ij

(1)

where i = treatment effects, j = number of observations,
x = 4 book-keeping variates for DZs, μ = intercept, z =
2 continuous variables DRYMOD and HOUSING, e =
random error term that is assumed to be normally distributed and αk and βl are coefficients.

4. RESULTS
4.1. UHI variations from 1990 to 2004
To investigate broad trends in the areal extent and
magnitude of the UHI, we show as an example three
5 yr mean patterns for June monthly mean minimum
temperature using ordinary kriging and an exponential statistical model to interpolate the air temperatures
in GIS (Fig. 5). The monthly mean minimum temperature distribution is characteristic of a UHI pattern with
a magnitude of 2 to 4 K, comparable to Fast et al.’s
(2005) general UHI findings for Phoenix. The series
shows a heat island in Phoenix that expanded as the
population grew from 2.2 million in 1990 to 3.7 million
in 2004. The shape of the heat island changed only
slightly between the first and second time period, but
expanded substantially in the most recent 5 yr period.
The intermediate 5 years were dominated by higher
than normal DRYMOD days, resulting in unusually
cool temperatures in Phoenix. Our results are also
comparable with those of Alcoforado & Andrade
(2006), who observed a UHI of ca. 3 K for Lisbon (population 2.7 million) in their analysis of sky view factor
(SVF) and percent built up area vs. nocturnal temperatures. Our results are also comparable with a previous,
limited analysis by Brazel et al. (2000) of selected
urban CORE sites in relation to a rural desert location
southeast of the urban area.

4.2. Results of statistical analysis
The slope coefficients of the DRYMOD and HOUSING
variables did not vary significantly among the 5 DZs
(Table 2). In other words, location within the urban area
per se did not affect the process of how the DRYMOD
and HOUSING variables impact the variability of TEMP.

Brazel et al.: Urban heat island in Phoenix

179

Table 2. Full model allowing interactions between categorical and continuDRYMOD is a broad regional process, and
ous variables, which should be reduced to an ANCOVA model that does not
we expected that its impact on all sites
allow interaction terms, because there was no statistically significant result
would be similar. The impact of HOUSING,
to reject H0 of equal slopes for covariates
through unifying processes of heat retention
owing to changing SVF and heat storage
Contrast
df Contrast sum Mean
F
p
changes near weather sites, also appeared
of square
square
to act similarly among all zones.
However, there were significant differTest slopes for HOUSING 4
7.735
1.934
0.81 0.5169
Test slopes for DRYMOD
4
6.017
1.504
0.63 0.6393
ences in the intercept coefficients between
some of the DZs (Table 3). The AGFRINGE
zone differed from the CORE and INFILL
sites at p < 0.05, and from the DESFRINGE at p < 0.10.
EXURBAN sites as the point of reference. The CORE
AGFRINGE was not significantly different from EXand INFILL temperatures were 2.2 K and 1.0 K higher
URBAN. AGFRINGE and EXURBAN have similar
than EXURBAN sites, respectively, whereas the
intercepts because EXURBAN sites used for this study
DESFRINGE was 0.5 K higher. The coefficient of
were also either near water bodies (dams or lakes)
AGFRINGE was not statistically different from that of
or near the strong influences of agricultural, wellEXURBAN. Overall, the results indicate a strong
watered landscapes. The INFILL and CORE sites difeffect of zone type on temperature. Given that housfered significantly (p < 0.05) from every other zone,
ing completion is in the model, it appears that
indicating that both are distinctive in terms of their
regional location influences the intensity of the UHI
heat retention factors and likely advective properties.
above and beyond the local effects of new housing
Coefficients derived from the ANCOVA showed the
construction. Results support the call by Oke (2006) to
effects of the independent variables on TEMP, using
report UHI findings by morphology and zones of a city

Fig. 5. Monthly mean minimum air temperature patterns (using ordinary kriging methods) from 1990 to 2004 in 5 yr increments

180

Clim Res 33: 171–182, 2007

Table 3. Results of tests of significant differences between intercept coefficients. The ANCOVA model should be tested to determine whether there
are urban development zone effects on June minimum temperature. This
table demonstrates that not all cases have statistically insignificant results,
which allowed us to keep the ANCOVA model. See Table 1 for abbreviations

5. THE HOUSING EFFECT

Based on the HOUSING coefficient in
Table 4 (noted above), the construction of
1000 new homes in the buffer surrounding
a site is estimated to increase temperature
AGFRINGE CORE DESFRINGE EXURBAN INFILL
by 1.4 K within various DZ areas. InspecAGFRINGE
< 0.0001
< 0.0560
< 0.6651
< 0.0001
tion of aerial photos of many of the indiCORE
< 0.0001
< 0.0001
< 0.0001
< 0.0001
vidual weather sites that were surrounded
DESFRINGE < 0.0560
< 0.0001
< 0.0384
< 0.0283
by development from 1990 to 2004 (e.g.
EXURBAN
< 0.6651
< 0.0001
< 0.0384
< 0.0001
Collier, Laveen, Crossroads, Sun Lakes,
INFILL
< 0.0001
< 0.0001
< 0.0283
< 0.0001
and Corbell; Fig. 1, Table 1) revealed that
it would take between 700 and 1000 homes
in order to facilitate proper intercity comparisons and
in the chosen buffer zone to build out a dense neighto unravel the broader influences of urban form on
borhood of typically sized, single-family homes. In
temperatures.
order to provide an interpretation of the theoretical
The DRYMOD and HOUSING variables were sigheat retention effect through factors such as the SVF
nificantly related to TEMP, holding DZ constant
(Svensson 2004, Unger 2004), we estimated a simple
(Table 4). The 2 variables acted as expected. High valrelation between numbers of homes per unit area
ues for the DRYMOD variable resulted in cooler temaround a site and a ground reference level SVF.
peratures, whereas high levels of new-home construcWe made this estimate for the end points of 1990 and
tion caused warming owing to the combination of
2004 by calculating typical height/width (H/W) ratios
reductions in SVF and addition of heat-retaining suramong homes, given home density and spacing in the
faces (roofs, asphalt, cement, etc.) to a given neighborbuffer, then by converting to a mean SVF for midhood. In June, the average number of SSC2 DRYMOD
distance among the homes using a simple expression
days is ~3 to 4 per month, which on average would
relating H/W to SVF (SVF = 0.2905 ln[H/W] + 0.4815, a
mean a cooling impact of some 1.1 to 1.4 K on the
relationship fitting Oke’s [1981] scale model results).
monthly mean minimum temperature. The HOUSING
We used the results of a field study by Hartz et al.
impact coefficient is + 0.0014 K per home completion,
(2006) to aid in determining the relationship between
or +1.4 K per 1000 homes (equivalent to complete build
density, spacing, and house height and size in order
up around a weather site in given a buffer size of
to estimate the SVF. Hartz et al. (2006) obtained field
1 km). There is variability around this value (SE =
estimates of SVF for a range of housing densities by
0.4 K), partly because not all new homes are alike, but
mapping the sky horizon among homes and converting
an inventory of the entire home completion data set
these data to SVFs. We consider our approximations
revealed mostly repetitive, single-family, similar style
for all weather sites to be reasonable, given the
residences. A more detailed breakdown of the HOUSapproximation we used for SVF. It should be noted that
ING impact coefficient across the rapidly developing
our calculation is primarily based on the spacing
zones of the metropolitan area by morphology of house
among homes and the impact that SVF variability
styles would require further analysis with climate
would have on temperature as a consequence of the
research sampling that is not possible with the limited
HOUSING variable alone.
number of fixed sites we analyzed in this study.
Our results of SVF calculations indicated that the
majority of locations would have experienced a reducTable 4. Results of ANCOVA with average June minimum
tion in HOUSING SVF of less than 0.20 in the weather
temperature in year j as dependent variable; α, β, and μ are
site buffer zones over the study time period. Only
explained in Eq. (1). HOUSING: new homes constructed;
about 20% of sites experienced reductions in SVF from
DRYMOD: dry moderate weather conditions; AGFRINGE:
1990 to 2004 that exceeded 0.20. Our estimates of SVF
agricultural fringe; DESFRINGE: desert fringe
for housing areas (from a mean of 0.99 in 1990 to a
mean of 0.83 in 2004) were somewhat consistent with,
Parameter
Estimate
SE
t
p
but slightly higher than, typical SVF values estimated
Intercept
μ = 21.818
0.1950 112.150 < 0.0001
by Unger (2004) and Svensson (2004) for variations
HOUSING
β1 = 0.0014
0.0004
3.07
< 0.0022
within their reported single-family neighborhoods,
DRYMOD
β2 = –0.3588 0.0238 –15.060 < 0.0001
and were very similar to a single-family residential
AGFRINGE
α1 = 0.1016
0.2347
0.43
< 0.6651
DESFRINGE
α2 = 0.5124
0.2470
2.08
< 0.0384
value (0.85) reported by Grimmond et al. (2001). Most
INFILL
α3 = 0.9740
0.2250
4.33
< 0.0001
of the residential areas around our weather sites were
CORE
α4 = 2.2105
0.2215
9.98
< 0.0001
not fully built up between 1990 and 2004, with the

Brazel et al.: Urban heat island in Phoenix

exception of some outliers of very densely built neighborhoods surrounding weather sites, where values of
mean buffer zone SVF were reduced to < 0.8, which
was closer to measured values typical of Hartz et al.’s
(2006) study of a fully-developed, compact, high density desert fringe neighborhood in the Phoenix area.
Regression equations determined by Unger (2004)
and Svensson (2004) imply typical magnitudes of residential warming compared with rural counterparts,
ranging between 0°C for an SVF of 0.99 and 1.5 to 2.5
K for SVF of 0.83. For an SVF reduction from 0.99 to
0.83 (mean change across all sites), our statistical
results would imply a temperature change attributable
to HOUSING around a site of only 0.3 K. For a reduction to an SVF much lower than 0.80, the temperature
increase would approach 1.0 to 1.5 K, which is lower
than residential warming values estimated by other
studies. To reconcile this, some points should be noted:
(1) our SVF buffer calculations would not necessarily
be explicitly representative of the weather station
locales, because most stations are not located right
between narrow spacings among homes in a neighborhood, but nearby where the SVF would be greater, and
(2) the DZ (which accounts for the type of development landscape) accounts for further heating beyond
the HOUSING effect. The HOUSING and DZ factors
together yield an UHI effect ranging from 3.6 K in the
CORE to 1.9 K at the DESFRINGE, assuming the build
up of 1000 homes. Considering both DZ and HOUSING, our results are very close to those of GrossmanClarke et al. (2005) and Hedquist & Brazel (2006) of 2
to 3 K heating in residential areas of Phoenix above
rural background levels (employing MM5 modeling
for the former, and mobile transect sampling for the
latter). In a similar multiple regression approach to
heat island evaluation of Stolberg, Germany, Kuttler et
al. (1996) found a 0.35 K or 10% so-called ‘sealed area’
rate of temperature increase across the city (streets,
city squares, buildings within 200 m2 of measuring
points from mobile sampling). This would yield ca. 3 to
4 K for fully built places and is comparable with our
heating rate due to housing change and associated DZ.

6. CONCLUSIONS AND REMARKS
The UHI spatial and temporal properties are related
to both climatic and human forces operating at a range
of spatial scales. In this study, we attempted to integrate climate and urban-growth variables to explain
the air temperatures in and around the Phoenix metropolitan area between 1990 and 2004. Results revealed
that the spatio-temporal variation in June minimum
temperature stems from large-scale synoptic processes
that bring cooler temperatures into the region on about

181

20% of June days, and the pace of new housing construction immediately surrounding weather stations
within a particular DZ. Owing to the fact that DZ did
not significantly affect the slope coefficients for DRYMOD and HOUSING, we concluded that the effects of
synoptic condition and new home construction did not
vary significantly across the urban region.
Our independent variables account for spatial variation (DZ), temporal variation (DRYMOD), and a combination of the two (HOUSING). The size of the coefficients indicate the estimated effects of moving from
one type of urban development to another. For example, the agricultural fringe added just 0.1 K to exurban
sites, which was not statistically significant. The desert
fringe added 0.5 K and infill locations 1.0 K, and the
urban core was 2.2 K warmer than the rural countryside. The additive effects of new home construction
(1.4 K per 1000 homes) and DZ produced estimates of
the UHI effect that are quite comparable with levels in
cities of comparable size.
Model coefficients provide planners and city officials with a tool to estimate the temperature effects of
new home construction. In Greater Phoenix, where
more than 60 000 building permits were issued for
residential units in 2004, urban growth has a profound
influence on localized temperature conditions. It is
commonplace for entire neighborhoods to sprout up
in a matter of months and for weather stations to
be completely ‘surrounded’ by urban development.
Potential homeowners also can make a more accurate
prediction of the long-term climate consequences of
new construction, depending on where they are
located and the pace of new construction. The overall
gained results and coefficients developed in the statistical approach of this study are certainly specific to
the Phoenix climate and setting. However, the role
of housing on warming is comparable with findings
elsewhere.

Acknowledgements. This study is based upon research supported by the National Science Foundation under Grant No.
SES-0345945 Decision Center for a Desert City (DCDC). Any
opinions, findings and conclusions or recommendation
expressed in this study are those of the authors and do not
necessarily reflect the views of the National Science Foundation (NSF). We thank the Maricopa Association of Governments for access to housing completion data; Nancy Selover
of the Office of Climatology, Arizona State University, for
access to PRISMS and NOAA data; Nancy Jones, Theresa Ip,
and Michelle Schwartz of DCDC for calculating housing data
for buffer zones around weather sites and editing the manuscript; Jon Skindlov, research scientist at the Salt River
Project, for reading the manuscript and providing historical
aerial photos of some of the PRISMS site locales; Donna Hartz
and Barbara Trapido-Lurie, Department of Geography, for
providing SVF data for several desert fringe neighborhoods
from a field study, and for cartography.

Clim Res 33: 171–182, 2007

182

LITERATURE CITED
Alcoforado MJ, Andrade H (2006) Nocturnal urban heat
island in Lisbon (Portugal): main features and modeling
attempts. Theor Appl Climatol 84:151–159
Arnfield AJ (2003) Two decades of urban climate research: a
review of turbulence, exchanges of energy and water, and
the urban heat island. Int J Climatol 23:1–26
Baker LC, Brazel AJ, Selover N, Martin C, McIntyre N,
Steiner FR, Nelson A, Mussacchio L (2002) Urbanization
and warming of Phoenix (Arizona, USA): impacts, feedbacks, and mitigation. Urban Ecosyst 6:183–203
Balling RC, Brazel SW (1987) Time and space characteristics
of the Phoenix urban heat island. J Ariz-Nev Acad Sci
21:75–81
Brazel AJ, Verville HJ, Lougeay R (1993) Spatial-temporal
controls on cooling degree hours: an energy demand parameter. Theor Appl Climatol 47:81–92
Brazel AJ, Selover N, Vose R, Heisler G (2000) Tale of two climates — Baltimore and Phoenix urban LTER sites. Clim
Res 15:123–135
Christakos G (2000) Modern spatiotemporal geostatistics.
Oxford University Press, New York
Christakos G, Serre ML, Kovitz JL (2001) BME representation
of particulate matter distribution in the state of California
on the basis of uncertain measurements. J Geophys Res
106:9717–9731
Comrie AC (2000) Mapping a wind-modified urban heat
island in Tucson, Arizona (with comments on integrating
research and undergraduate learning). Bull Am Meteorol
Soc 81:2417–2431
Cooper G (1998) Air-conditioning America: engineers and
the controlled environment, 1900–1960. Johns Hopkins
University Press, Baltimore, MD
Fast JD, Torcolini JC, Redman R (2005) Pseudo-vertical
temperature profiles and the urban heat island measured
by a temperature data logger network in Phoenix. J Appl
Meteorol 44:3–13
Gammage G Jr (1999) Phoenix in perspective: reflections on
developing the desert. Herberger Center for Design,
Tempe, AZ
Gober PA (2006) Metropolitan Phoenix place making and
community building in the Desert. University of Pennsylvania Press, Philadelphia, PA
Gober PA, Burns EK (2002) The size and shape of Phoenix’s
urban fringe. J Plan Educ Res 21:379–390

Grimmond CSB (2006) Progress in measuring and observing
the urban atmosphere. Theor Appl Climatol 84:3–22
Grimmond CSB, Potter SK, Zutter HN, Souch C (2001) Rapid
methods to estimate sky-view factors applied to urban
areas. Int J Climatol 21:903–913
Grossman-Clarke S, Zehnder JA, Stefanov WL, Liu Y, Zoldak
MA (2005) Urban modifications in a mesoscale meteorological model and the effects on near-surface variables
in an arid metropolitan region. J Appl Meteorol 44:
1281–1297
Hartz DA, Prashad L, Hedquist BC, Golden J, Brazel AJ
(2006) Linking satellite images and hand-held infrared
thermography to observed neighborhood climate conditions. Remote Sens Environ 104:190–200
Hawkins TW, Brazel AJ, Stefanov WL, Bigler W, Saffell EM
(2004) The role of rural variability in urban heat island
determination for Phoenix, Arizona. J Appl Meteorol 43:
476–486
Hedquist BC, Brazel AJ (2006) Urban, residential, and rural
climate comparisons from mobile transects and fixed
stations: Phoenix, Arizona. J Ariz-Nev Acad Sci 38:77–87
Kuttler W, Barlag AB, Rossmann F (1996) Study of the thermal
structure of a town in a narrow valley. Atmos Environ
30:365–378
Oke TR (1981) Canyon geometry and the nocturnal heat
island: comparison of scale model and field observations.
J Climatol 1:237–254
Oke TR (2006) Towards better scientific communication in
urban climate. Theor Appl Climatol 84:179–190
Pon B, Stamper-Kurn DM, Smith CK, Akbari H (1998) Existing
climate data sources and their use in heat island research.
Available at: http://eetd.lbl.gov/EA/Reports/41973/
Sheridan SC (2002) The redevelopment of a weather-type
classification scheme for North America. Int J Climatol
22:51–68
Souch C, Grimmond S (2006) Applied climatology: urban
climate. Prog Phys Geogr 30:270–279
Stabler LB, Martin CA, Brazel AJ (2005) Microclimates in a
desert city were related to land use and vegetation index.
Urban Forest Urban Green 3:137–147
Svensson MK (2004) Sky view factor analysis — implications
for urban air temperature differences. Meteorol Appl 11:
201–211
Unger J (2004) Intra-urban relationship between surface
geometry and urban heat island: review and new
approach. Clim Res 27:253–264

Editorial responsibility: Helmut Mayer,
Freiburg, Germany

Submitted: July 11, 2006; Accepted: October 19, 2006
Proofs received from author(s): January 10, 2007