Data from a dense urban meteorological network (UMN) are analyzed, revealing the spatial heterogeneity and temporal variability of the Twin Cities (Minneapolis–St. Paul, Minnesota) canopy-layer urban heat island (UHI). Data from individual sensors represent surface air temperature (SAT) across a variety of local climate zones within a 5000-km2 area and span the 3-yr period from 1 August 2011 to 1 August 2014. Irregularly spaced data are interpolated to a uniform 1 km × 1 km grid using two statistical methods: 1) kriging and 2) cokriging with impervious surface area data. The cokriged SAT field exhibits lower bias and lower RMSE than does the kriged SAT field when evaluated against an independent set of observations. Maps, time series, and statistics that are based on the cokriged field are presented to describe the spatial structure and magnitude of the Twin Cities metropolitan area (TCMA) UHI on hourly, daily, and seasonal time scales. The average diurnal variation of the TCMA UHI exhibits distinct seasonal modulation wherein the daily maximum occurs by night during summer and by day during winter. Daily variations in the UHI magnitude are linked to changes in weather patterns. Seasonal variations in the UHI magnitude are discussed in terms of land–atmosphere interactions. To the extent that they more fully resolve the spatial structure of the UHI, dense UMNs are advantageous relative to limited collections of existing urban meteorological observations. Dense UMNs are thus capable of providing valuable information for UHI monitoring and for implementing and evaluating UHI mitigation efforts.
The systematic manipulation and replacement of natural environments with built environments in and around cities affect surface climate such that urban areas are warmer than their rural environs. This phenomenon, known as the urban heat island (UHI), appears at the surface and throughout the lower atmosphere, from the canopy layer to the boundary layer (Rao 1972; Oke 1976). Evidence suggesting the existence of UHIs in surface air temperature (SAT) was first documented for London, United Kingdom, by Howard (1833), and later confirmed by Chandler (1965). The basic physics that accounts for the existence of UHIs was deduced by Howard (1833) in the early nineteenth century and was formalized in the late twentieth century (Nunez and Oke 1977; Goward 1981; Oke 1982); grounded in an energetic basis, the underlying physics implies that UHIs are present anywhere that humans have urbanized the land surface in a manner that perturbs the surface energy budget from its natural background. Given the dramatic land-use changes that are associated with the construction of modern cities, it is unsurprising that UHIs are pervasive. UHIs have been documented in cities around the globe (Arnfield 2003; Peng et al. 2012; Hertel 2014).
For all the gains in descriptive and theoretical knowledge of UHIs, challenges and uncertainties remain. A paucity of urban temperature observations limits the quantification of canopy-layer UHI magnitude to simple frameworks such as the difference between averages of two or more stations separated into urban and rural categories (Hage 1972; Ackerman 1985; Basara et al. 2008). Urban–rural classification schemes have dominated the UHI literature of the past century, providing the best available observational estimates of canopy-layer UHI magnitude. Stewart’s (2011) systematic review of UHI studies has suggested that there are often serious deficiencies in station metadata, leading to ambiguous urban–rural classifications and potentially inaccurate canopy-layer UHI estimates. Furthermore, urban–rural classification schemes often reduce the rich heterogeneity of the urban environment into a single number dominated by the thermal source areas around a small number of stations that may or may not adequately represent the urban form (Oke 2006). Mobile sensing platforms have, in some cases, temporarily resolved this issue by providing skeletal snapshot views of the UHI along roadways throughout cities and their surrounding rural areas (Oke 1995). In this sense, mobile traverses tend to trade increased spatial information for decreased temporal information (Oke 1973). Cooperative observing networks coordinated by government agencies may be capable of resolving the UHI (Hausfather et al. 2013) but are susceptible to data inhomogeneity (Wu et al. 2005). Dedicated and highly controlled urban meteorological networks (UMN) composed of fixed meteorological sensors are a relatively recent development and are capable of providing improved spatiotemporal observations of canopy-layer UHIs (Grimmond 2006; Muller et al. 2013a).
This study describes results that are based on data from a dense UMN in the Twin Cities metropolitan area (TCMA), encompassing the cities of Minneapolis and St. Paul, Minnesota. It complements and advances previous observational studies of UHIs by overcoming most of the limitations inherent in the sparse networks that are used to characterize most urban areas (World Meteorological Organization 2008; Muller et al. 2013a). In this paper, we 1) describe the essential features of the network; 2) present improved estimates of the TCMA canopy-layer UHI, including its diurnal, seasonal, and nonseasonal variability; 3) relate its spatial and temporal variability to human and natural factors such as meteorological conditions and land use; and 4) discuss implications for network design, UHI monitoring, and mitigation that are pertinent to investigators and agencies seeking to replicate our efforts in other cities.
Accurate quantification of the canopy-layer UHI magnitude is important because of its direct impacts on human activity and the compounding effect of climate variability at larger scales. UHIs have a variety of economic, social, and environmental consequences; they are associated with increased energy consumption (Santamouris et al. 2001), increased heat stress (Kovats and Hajat 2008), decreased air quality (Stone 2005), and urban ecosystem stresses (Baker et al. 2002). UHIs are superimposed upon naturally occurring regional climate variability (Lowry 1977) as well as large-scale warming caused by increasing concentrations of greenhouse gases in Earth’s atmosphere (Grimmond 2007). It follows that distinguishing among local impacts of the UHI, regional climate variability, and global warming requires urban observations that are capable of resolving all three sources of local climate variability. Here we demonstrate how observations from a dense UMN may be used to facilitate the separation of the UHI from background regional climate variability.
Distinguishing between time-varying changes in the UHI and global warming is a larger challenge that is beyond the scope of this study. At the scale of individual cities and their surrounding metropolitan areas, natural variability partially obscures long-term trends, regardless of their cause (Wallace 1996). The data record used in this study is not long enough to warrant investigating secular trends in the local UHI. In the future, UMNs sustained over decades with thoroughly documented metadata may be capable of separating time-varying changes in the UHI from the local interplay of regional climate variability and global warming.
2. Study area, network details, and gridding procedure
a. Twin Cities metropolitan area and its UHI
The TCMA is composed of Minneapolis, St. Paul, and their surrounding communities in east-central Minnesota, just west of the Minnesota–Wisconsin border (Fig. 1a). The study area, indicated by the dashed box in Fig. 1b, is located in the Dfb Köppen–Geiger climate zone (snow and fully humid warm summer; Kottek et al. 2006) and covers 5000 km2 within the seven counties of the TCMA (Anoka, Carver, Dakota, Hennepin, Ramsey, Scott, and Washington). It is traversed by the Mississippi, Minnesota, and St. Croix River valleys, above which the elevation varies gradually around an average of 275 m above sea level. The landscape is interspersed with approximately 900 lakes, of which the largest is Lake Minnetonka on the western periphery of the TCMA. A variety of land cover and land uses exist in the TCMA: high-density urban development in the core communities of Minneapolis, St. Paul, and Bloomington, Minnesota; medium- to low-density urban development in the surrounding suburban communities; and agriculture, undeveloped wetlands, forests, and prairie in its rural environs.
The TCMA UHI has been studied over several decades. Winkler et al. (1981) was one of the earliest studies to produce a quantitative characterization of the TCMA UHI. Using a 10-yr data record composed of 20 cooperative weather stations and the local National Weather Service station, Winkler et al. (1981) estimated the spatial extent and average magnitude of the TCMA UHI for the annual mean (1.0°C) and the calendar months of January (1.1°C) and July (1.2°C). Todhunter (1996) used a slightly larger network of 26 cooperative and television weather stations to describe the effect of the TCMA UHI on a variety of climatological indices during 1989. Whereas Winkler et al. (1981) showed hand-analyzed maps of the TCMA UHI, Todhunter (1996) employed ordinary kriging to analyze local climate statistics on a regular grid. Malevich and Klink (2011) studied the relationship between snow and the TCMA UHI, which they defined as the difference between 19 “iButton” sensors around the Minneapolis central business district (CBD) and the Minneapolis–St. Paul International Airport (KMSP), located 10 km to the southeast. They reported that the UHI was larger during winter days than nights, that the presence of snow increased the UHI magnitude during daytime and nighttime, and that moderate snowfall tended to diminish the UHI magnitude. Their network did not detect any significant relationship between the intensity of urbanization and the TCMA UHI magnitude. As part of an analysis of UHIs in the world’s largest cities for the period of 2006–12, Hertel (2014) analyzed the TCMA UHI by contrasting the average of a set of hourly rural airport observations against concurrent observations at two urban airports: KMSP and St. Paul Downtown (KSTP). Hertel (2014) documented the TCMA UHI on the basis of the annual (1.3°C), wintertime (0.9°C), and summertime (1.8°C) averages.
b. Network design, installation, and administration
This study employs data from a dense UMN designed to advance knowledge of the TCMA canopy-layer UHI by addressing limitations intrinsic to the observation networks used in previous studies (e.g., sparse networks or a limited period of record). Installation began in June of 2011. By August of 2011, the nascent network had nearly 30 sensors located throughout the urban core and the suburban and rural areas, sufficient to capture the urban–rural temperature gradient across the study area. As of late 2012, the UMN reached a stable size of approximately 170 sensors distributed throughout the 5000-km2 study area, surpassing previous TCMA UHI studies by nearly a factor of 10.
The network samples individual neighborhoods in the urban core as well as the varied natural and agricultural environs of the surrounding periurban areas (Fig. 1). In the central cities of Minneapolis and St. Paul, the average distance between a sensor and its nearest three neighbors is 1.7 km. In the surrounding suburban and exurban communities, the average distance between a sensor and its nearest three neighbors is 3.8 km. The average elevation of observation sites is 276 m. The standard deviation of the observation sites about the average elevation is 18 m, and the range is 216–316 m. None of the observation sites are located in complex terrain subject to topographically modified flows. Only a very small fraction of the TCMA is located in the narrow upper Mississippi River valley, which by virtue of its topography is essentially unpopulated.
The observation sites in our network were determined through the following process: First, local citizen volunteers were solicited from contributors to two existing local observing networks: the cooperative Community Collaborative Rain, Hail, and Snow (CoCoRaHS) network and the Minnesota State Climatology Office’s MNgage precipitation network. We subsequently expanded our network of volunteers through our research project Internet page (http://islands.umn.edu) and local media outlets (television and newspaper), and by establishing partnerships with the municipal parks departments in Minneapolis and St. Paul. Drawing from this list of volunteers, we sought to balance areal extent and density, with the goal of a denser network in the urban core and an areal extent that reached into the rural areas beyond the TCMA. Once a volunteer is accepted into the network, we site an observing platform in consultation with the World Meteorological Organization initial guidance described by Oke (2006). To minimize aliasing differing microclimates into our data, we seek sites with similar ground cover immediately surrounding the sensor platform (generally low plants such as grasses). Sensors are always sited away from potential lateral heat sources such as large buildings, air conditioning units, exhaust vents, and large trees. Beyond these considerations, the siting reflects a compromise between the needs of the project (i.e., ground cover and sky exposure) and the desires of the volunteer (e.g., that their property remains minimally affected in terms of aesthetics and function).
Stewart and Oke (2012) introduced local climate zones (LCZs) as a formal manner of categorizing the environments in which local meteorological observations are taken. Table 1 lists each LCZ and presents the number of sites in each category as of 1 September 2014 (primary LCZ class is in the first column; secondary LCZ class is in the second column). The network represents a diversity of built environments in the TCMA: from the open high, mid-, and low rises of the urban core to the sparsely built and undeveloped natural areas in communities surrounding Minneapolis and St. Paul. No stations are sited in compact high-, mid-, or low-rise areas because, with the exception of the downtown CBDs, these built environments do not exist in the TCMA. The same is true of lightweight low-rise areas. Heavy industry and large low-rise commercial areas are present in the TCMA but are not represented among the primary LCZ classes, primarily because the volunteers for our network were private citizens and not private organizations with property in commercial or industrial areas. Because there are large low-rise commercial areas adjacent to a number of the sensor sites, these areas are represented among the secondary LCZ classes. Some of the sensors are sited in the vicinity of lakes, which are also represented among the secondary LCZ classes.
Data from individual sensors in the network are downloaded manually every 3–4 months. As new data are collected, they are converted from their native comma-separated-value form to network common data form (NetCDF), processed by an automated quality-assurance procedure, aggregated with existing station data, and archived.
c. Sensor type, calibration, and quality assurance
Individual dataloggers (HOBO1 Pro v2 model number U23-004) are deployed at each site with naturally aspirated radiation shields. Each logger records air temperature measurements via an external sensor connected to the waterproof case, as well as an internal temperature measurement. The preferred deployment for each datalogger is on a standard observation platform, composed of a 5-cm-diameter post, a logger, a sensor, and a radiation shield. The radiation shield is attached to the post such that the measurement is taken as close to 2 m as possible. In cases in which the standard observing platform cannot be deployed because of property restrictions, the logger, external sensor, and radiation shield are attached to a suitable object (e.g., flagpole, light post, or fence post) as close to 2 m as possible. The dataloggers are launched in the field and configured to record observations every 15 min.
The entire set of 200 dataloggers was calibrated in two separate groups of 100. Each logger was given a unique identifier (ID), launched with HOBOware software, and successively placed in four uniform environments to verify the sensor specifications: a laboratory at 20°C (room temperature), a deep freezer at −40°C, a cooler at 0°C, and the University of Minnesota St. Paul campus weather station at 32°C (field conditions). Dataloggers that departed from the manufacturer specifications at the time of calibration were withheld from deployment. See the supplemental online material for detailed sensor specifications. Dataloggers exhibiting abrupt discontinuities, persistent errors, or failure while deployed were removed from the field, tested, and replaced with a new or recalibrated datalogger. Successfully recalibrated dataloggers were only redeployed after being updated and launched with a new ID.
To ensure that our data are free from spurious sensor readings, we employ a series of three automated data quality-assurance (QA) procedures that are consistent with those described by Durre et al. (2010):
In a local all-time-record check, we flag any 15-min observations that fall outside the all-time records for Minneapolis–St. Paul (temperature T < −36.6°C or T > 42.2°C). This check identifies physically implausible temperature observations.
In a temporal consistency check, we flag 15-min observations that exhibit a change of ±4°C or greater relative to either the preceding or subsequent 15-min observation, as well as observations that exhibit a change of 2°C or greater relative to both the preceding and subsequent 15-min observations. This check identifies physically implausible temperature variability.
In a spatial consistency check, we flag any hourly observations that are more than six standard deviations (6σ) from the network mean at a particular time point. This check identifies temperature observations with pronounced, isolated microclimatic influences as well as some additional climatological outliers that are not identified by either of the first two checks.
The first two QA procedures are applied to the raw 15-min data. The spatial consistency check is applied after averaging the data to hourly intervals. Observations flagged as outliers are recorded and removed from the dataset. Random samples of flagged values were manually inspected to assess the validity of the observation and the fidelity of the QA check.
d. Spatial interpolation with kriging and cokriging
Hourly UMN data that pass all three QA procedures are interpolated to a uniform 1 km × 1 km grid using two geostatistical techniques: kriging and cokriging (covariate kriging). The grid was determined by the smallest station separation distances in the urban core. The prediction grid covers a domain that fully encompasses the TCMA: from 44.55° to 45.35°N latitude and from 93.80° to 92.75°W longitude (dashed line in Fig. 1). To anchor the interpolation schemes at the boundaries of the analysis domain, we supplement our UMN observations with rural airport observations at nine locations surrounding the TCMA. These airports are listed in Table 2 and are shown with black squares in Fig. 1. To evaluate the kriged and cokriged temperature fields, we cross validate the gridded temperature data by calculating RMSE using a set of six TCMA airport observations that are not used in either interpolation scheme. These urban airports are also listed in Table 2 and are shown with white squares in Fig. 1.
Kriging is a standard geostatistical tool that interpolates a distributed field of data using information inherent in the spatial covariances (Journel and Huijbregts 1978). Toward that end, kriging uses semivariograms to express the spatial dependence of the field around a known data point and to develop the kriging weights that form linear combinations at prediction data points. We develop a semivariogram for each set of hourly canopy-layer air temperature observations using an exponential model whose three parameters (nugget, range, and sill) are determined by nonlinear least squares regression (Kawashima and Ishida 1992). Because there are no dramatic differences in the background climatic data over the TCMA owing to elevation or land–sea contrasts, we employ so-called simple kriging, which assumes a stationary mean. The kriging procedure begins by estimating the semivariogram using all available data for a given time point. The three semivariogram parameters are estimated for each hour between 2000 central standard time (CST; UTC − 6 h) 31 July 2011 and 0400 CST 1 August 2014. In cases of a singular semivariogram model, a linear variogram fit is applied. A range cutoff of 25 km is applied when producing the kriged fields, obtained by averaging the range parameter over all time points. The range parameter represents the distance beyond which the autocorrelation is essentially zero. Thus, the range cutoff restricts the kriged temperature at a given point to nearby observations within a range of 25 km.
Cokriging leverages covariance between a sparsely sampled prediction variable and a well-sampled secondary variable to interpolate the former with greater accuracy than may be possible with kriging alone (Ishida and Kawashima 1993). In our application, we seek to predict canopy-layer air temperature on a uniform 1 km × 1 km grid from a set of irregularly distributed temperature observations and a field of gridded impervious surface area (ISA) data. As an extension of the kriging procedure, cokriging uses a cross semivariogram to represent the spatial dependence of the two variables.
Impervious surfaces such as pavements do not permit infiltration of rainwater and have pronounced thermal mass relative to unsaturated soil and naturally vegetated surfaces. When insolation is present, these differences contribute to contrasting thermal storage and radiative cooling rates, which may lead to the development of a UHI under certain conditions. For example, Zhang et al. (2011) reported a statistically significant positive correlation between imperviousness and the strength of the canopy-layer UHI magnitude in their study of Detroit, Michigan, during the summer of 2008. Schatz and Kucharik (2014) also reported a significant correlation between imperviousness and the magnitude of the UHI; they also found that statistical models that are based on imperviousness outperformed models that are based on the normalized difference vegetation index, a measure of vegetation abundance. Our study utilizes a set of 30-m-resolution ISA data from the National Land Cover Database 2011 (NLCD2011; Jin et al. 2013). To capture the imperviousness of the thermal source region surrounding prediction points, we regrid the 30-m-resolution ISA data to 500-m resolution. Depending on meteorological conditions and surface character, the thermal source area of a standard 2-m temperature observation may extend from tens of meters to greater than 1 km (Kljun et al. 2002; Stewart and Oke 2012). Our results were not sensitive to the choice of ISA averaging radius for values of less than 750 m. Following the approach to kriging, the temperature estimates produced by cokriging at each prediction point are based on temperature observations and 500-m ISA data within a range of 25 km.
A secondary concern is the time required to produce the field estimated with cokriging. As the volume of ISA data increases, so does the computation time. In the study area, many of the 500-m ISA values are zero, in particular over rural areas. To balance the sampling of the ISA field with manageable computation time, we randomly selected an equal number of sampling points from each quintile of the ISA data. This approach reduces the dimensionality of the 500-m ISA field from 34 000 grid points to 4000 grid points. We tested the sensitivity of our results to the choice of the number of 500-m ISA grid points included in the cokriging scheme and found little sensitivity, in particular when more than 2500 grid points were included. The same 4000 ISA grid points were used in each cokriging estimate.
To validate the interpolated SAT fields, we use hourly SAT observations from the six urban Twin Cities airport stations listed in Table 2 to calculate bias and RMSE relative to the closest grid point. These urban-airport observations were not included in the interpolation schemes. Thus they represent independent data that may be used to evaluate kriging and cokriging.
The distribution of daily-mean SAT observed at TCMA UMN sensor sites on 16 May 2012 is shown in Fig. 2a. The corresponding kriged daily-mean SAT field is shown in Fig. 2b, along with station observations from six urban airports withheld from interpolation. The kriged field captures the general pattern of urban warmth but lacks definition where observations are sparse, in particular at the edge of the study domain. By contrast, the field cokriged with ISA data exhibits greater fidelity to observations, both in terms of the pattern of urban warmth and in comparison with the urban-airport validation stations. Bias and RMSE are calculated at the six grid points collocated with the urban-airport validation sites and are shown in Table 3 along with the average over all six locations. Both metrics exhibit smaller values at all six validation locations for the cokriged field when compared with the kriged field. Some bias is expected between the UMN-based gridded fields and the observations from the airport automated surface observing system (ASOS) because of the differing manner in which the sensors housings are ventilated. The UMN temperature sensors are naturally aspirated by wind; the airport ASOS temperature sensors are mechanically aspirated by fans. The average bias is only slightly larger than the reported standard error for the HOBO sensors (±0.2°C), which lends confidence to the gridding approach. The validation statistics associated with the cokriged SAT field are clearly superior to the kriged SAT field; therefore we will hereinafter only refer to results that are based on the cokriged SAT field. Other approaches to validation, such as vehicle traverses, could help to verify the instantaneous structure and magnitude of the TCMA UHI, but they are limited in their ability to observe the UHI continuously over its temporal evolution and thus remain beyond the scope of this study.
In addition to the network and sensor details that have been described in this section, extensive metadata are provided in the online supplemental material, following the standardized UMN metadata protocol introduced by Muller et al. (2013b).
a. Defining the UHI
The canopy-layer UHI has typically been defined as the difference between urban and rural screen-level (2 m) temperature observations (Oke 1982). Despite broad agreement on this fundamental concept, the UHI literature is complicated by a diversity of station classification schemes and UHI magnitude definitions (Stewart 2011).
Because of a paucity of unambiguously urban observations in major municipalities, studies have often relied on as few as two stations (one urban, one rural) to estimate the UHI (e.g., Ackerman 1985; Brazel et al. 2000; Fortuniak et al. 2006). Where a greater number of urban observations exist, investigators have begun to define “urban” and “rural” on the basis of averages of two or more stations (e.g., Kim and Baik 2005; Basara et al. 2008). Hawkins et al. (2004) pointed out that the spatial variability of temperature within rural areas could influence estimates of UHI magnitude. Fortuniak et al. (2006) noted that their initial estimates of UHI magnitude were affected by the passage of mesoscale boundaries such as fronts and thunderstorm outflows. Thus, they filtered days and nights when these features were known to be present in the vicinity of the two stations used in their study.
To overcome the aforementioned challenges associated with defining the UHI and estimating its magnitude, we define the TCMA rural background SAT as an average of the nine rural airport observations used as rural boundary conditions in the interpolation schemes described in section 2. These municipal airports are located outside the urban core, well beyond the influence of the TCMA UHI. They are unambiguously rural in character, situated over low plants away from the influence of built structures. They also encircle the TCMA, which provides a spatially representative average of its rural environs. We do not specifically account for the passage of mesoscale boundaries through our study area, but our area-average rural background SAT and choice to consider time averages of six or more hours reduce the likelihood that these features unduly influence our results.
To investigate the spatial characteristics of the TCMA UHI, we define and analyze a time-varying SAT anomaly field by subtracting the rural background SAT from the cokriged SAT field described in section 2. Therefore the UHI is present wherever local anomalies are greater than zero. The dataset has hourly sampling and extends continuously over the 3-yr period from 1 August 2011 to 1 August 2014. This allows us to resolve the diurnal evolution of the canopy-layer UHI.
Rather than classifying the UMN sensor sites as urban or rural and presenting UHI magnitude estimates that are based on the difference of their means, we instead report two spatial statistics derived from the spatial UHI distribution: the area average (μ) and the 95th percentile (P95). The area average is comparable to studies that estimate the UHI with the average difference between many urban and rural stations, whereas the 95th percentile is comparable to studies that estimate the UHI by the peak difference between urban and rural stations, often defined as the UHImax. The area average and 95th percentile are calculated for both urban and suburban areas.
The study area is classified into urban, suburban, and rural areas in Fig. 1c. These definitions are based on the average ISA within local municipalities. The urban area is defined as contiguous communities with greater than 40% area-average ISA, including the Twin Cities of Minneapolis and St. Paul, as well as their older, more developed first-ring suburbs. The suburban area is defined as contiguous nonurban communities with greater than 20% area-average ISA, including second-ring suburbs and several exurban communities that continue to expand the spatial extent of the TCMA. The remainder of the study area is defined as rural, although it should be noted that a small degree of urbanization is present in towns that surround the TCMA.
b. Spatial structure of the time-average TCMA UHI
The study-mean, daily-mean TCMA UHI shown in Fig. 3a extends across much of the study area but is concentrated in the most developed portions of the region. Local maxima are apparent over the Minneapolis CBD, the St. Paul CBD, and the KMSP airport. In the suburban area, the UHI magnitude is weaker by a factor of 2–3 and appears to be more heterogeneous than in the urban core. Beyond the suburbs, the UHI pattern tapers off, except along narrow corridors that appear to follow the development along the major interstate highways that bisect the TCMA.
The PDF corresponding to the pattern in Fig. 3a that is shown in Fig. 3b is positively skewed with a maximum around zero and a secondary maximum around 0.8°C. The primary maximum represents the outer extent of the UHI, and the secondary maximum represents the beginning of the inner-core UHI. These respective features also correspond to what Runnalls and Oke (2000) referred to as the UHI “cliff” at the rural–urban edge and the start of UHI “peaks” in areas of significant urban development.
c. Diurnal and seasonal variation
On average, the UHI is present throughout the entire day, but it varies in strength according to time of day, time of year, and location (urban TCMA in Figs. 4a,c,e; suburban TCMA in Figs. 4b,d,f). Averaged over the entire calendar year, the TCMA UHI peaks at midnight, with a secondary maximum at midday and minima around sunrise and sunset (Figs. 4a,b). This behavior is consistent with our understanding of the energetic basis of the UHI; that is, it varies diurnally in accordance with the mesoscale differences in urban and rural heating and cooling rates largely attributable to differences in radiative and thermodynamic properties (Oke 1982). Microscale differences in heating and cooling rates (e.g., building shading, cold-air pooling) also explain local UHI magnitudes but occur on spatial scales that are smaller than our gridded analysis represents (<1 km) and influence too small of a portion of the TCMA to affect the area average.
The TCMA experiences seasonal variations in the length of day, which have attendant impacts on the diurnal evolution of the UHI (see Fig. S1 in the supplemental online material). Runnalls and Oke (2000) and Fortuniak et al. (2006) normalized time to parse out UHI changes independent of season. The current study aims to examine seasonal changes in the UHI, including those driven by changes in length of day; therefore, daytime is defined as the period from sunrise to sunset, and nighttime is defined as the period from sunset to sunrise.
In the TCMA urban core, the nighttime-mean UHI is stronger than the daytime-mean UHI (Figs. 5a,b), whereas in the suburban periphery the nighttime- and daytime-mean UHI patterns are of comparable magnitude. In other words, the nighttime UHI is characterized by a stronger rural–urban SAT gradient than the daytime UHI. This is also evident in the corresponding spatial PDFs shown in Fig. 5c. The nighttime UHI PDF is similar to the daily-mean PDF shown in Fig. 3b, with a mode near 0°C, a secondary maximum around 1°C, and a small set of extreme values. In contrast, the daytime UHI PDF has a mode near 0.25°C, followed by a mostly uniform distribution that abruptly drops off above 1.25°C. We examine these differences in greater detail in section 3.
When the data are averaged only over the summer months (June–September) and winter months (December–March), their corresponding diurnal evolutions suggest that the small differences between the average nighttime and daytime UHI magnitude in the annual mean are an artifact of averaging nearly equal and opposite seasonal behavior. In the warm season, the TCMA UHI is strongest at night, peaking in the hours preceding midnight (Figs. 4c,d). In the cold season, the TCMA UHI is strongest during the day, peaking around noon (Figs. 4e,f). The warm-season behavior is consistent with previous studies of UHI diurnal variability in midlatitude cities (Kłysik and Fortuniak 1999; Runnalls and Oke 2000; Kim and Baik 2005; Fortuniak et al. 2006; Basara et al. 2008), and the cold-season behavior agrees with the results of Malevich and Klink (2011), who used a smaller UMN to study the relationship between the TCMA UHI and snow.
Different TCMA UHI patterns are expected between the warm and cold seasons given the large seasonality of the mean climate in the north-central United States (Fig. 6a). During the study period, for example, daily-mean SAT varied by nearly 50°C from winter to summer. The nighttime-mean UHI time series shown in Fig. 6b varies on multiple time scales. On daily to weekly time scales, the nighttime-mean UHI is largest in the cold season, as exemplified by the daily peaks in the area-average time series in Fig. 6b. On seasonal time scales, however, the nighttime-mean UHI is largest in the warm season, as exemplified by the time series after being filtered by a low-pass Lanczos filter with 25 weights and a low-frequency cutoff of 0.0111 (90 days; Duchon 1979). The daytime-mean UHI time series shown in Fig. 6c exhibits less variability than the nighttime-mean time series, both on daily and seasonal time scales. These contrasts between the nighttime and daytime UHI magnitude become clearer when seasonal averages are compared with seasonal extremes.
The summertime and wintertime average nighttime and daytime UHI patterns are shown in Fig. 7 along with their corresponding spatial PDFs. UHI magnitudes (μ and P95) associated with these maps are listed in Table 4, disaggregated by diurnal and seasonal period. During the summer, the UHI is strongest at night over the urban core, with magnitudes approaching 2°C locally in and around the KMSP airport and the Minneapolis–St. Paul CBDs. During the day, the summertime UHI is more spatially uniform. Hotspots are still evident, but horizontal SAT gradients are much weaker during the day than at night. During the winter, the UHI is of comparable magnitude during the day and night, both in the urban core and the suburban periphery. In fact, some of the largest differences between the nighttime and daytime wintertime UHI are seen in the rural TCMA, where the UHI anomalies are largely negative at night and are near zero or slightly positive during the day. This may be a manifestation of the contrast between strong radiative cooling by snow-covered surfaces at night and modest radiative heating by low-albedo pavements, structures, and bare vegetation.
As discussed in section 3b, the nighttime and daytime UHI time series shown in Figs. 6b and 6c exhibit pronounced high-frequency daily variability apart from their low-frequency seasonal modulation. The time-average maps shown in Fig. 7 illustrate the low-frequency seasonal patterns. To examine the patterns of high-frequency daily extremes, we composited the 30 warmest nights and days in each season on the basis of the time series in Figs. 6b and 6c. Maps are shown in Fig. 8, and the associated spatial statistics are shown in Table 5.
Like the seasonal contrasts that are evident in the time-average UHI patterns (Fig. 7), the nighttime extreme UHIs (Fig. 8a) are much stronger than the daytime extreme UHIs (Fig. 8b). Unlike the seasonal contrasts that are evident in the time-average UHI patterns, the nighttime extreme UHIs are stronger and more spatially concentrated in the wintertime (Fig. 8a) than in the summertime (Fig. 8c), particularly in the urban core. Another difference between the time-average and extreme UHIs is evident in the contrast between the daytime UHI strength in the summer and winter. In the time average, there is little difference between the summertime and wintertime daytime UHI; in the extreme cases, however, the wintertime daytime UHI (Fig. 8d) is 50% larger than the summertime daytime UHI (Fig. 8b). Taken together, the pronounced difference in UHI magnitude and patterns between the time-average and extreme cases suggests that different processes modulate the daily UHI variability and the seasonal UHI variability.
e. Modulation by regional meteorological conditions
The preceding results indicate that the TCMA UHI exhibits differing patterns of variability on daily and seasonal time scales. Previous studies have demonstrated the pronounced influence of weather on the formation and strength of UHIs around the world (e.g., Ackerman 1985; Runnalls and Oke 2000; Morris and Simmonds 2000; Morris et al. 2001; Kim and Baik 2005; Yow 2007). Following these analyses, we examined the meteorological conditions concurrent with the top 30 extreme nighttime and daytime UHI events during our study period. Table 6 lists the number of events with particular categorical sky cover, wind speed, temperature, and dewpoint conditions. Regardless of season, pronounced nighttime UHI events are strongly favored under clear or scattered sky cover conditions with light or no winds. Furthermore, these nighttime events are generally associated with warm conditions during the summertime and very cold conditions during the wintertime. The former are indicative of the residual daytime heat retained in the urban environment during an extreme warm-season UHI event. The latter are consistent with enhanced anthropogenic heat flux to the urban environment during the wintertime (Sailor and Lu 2004). Intense daytime UHI events do not show a particular tendency toward sky cover or wind conditions, but they appear to be associated with warmer conditions that are indicative of greater energy flux into the urban land–atmosphere system. During the summer, there is some indication that extreme nighttime UHI events occur during dry conditions, when large vapor pressure deficits could conceivably drive enhanced radiative cooling in the rural areas. In contrast, Table 6 suggests that extreme daytime UHI events occur under moist conditions during the warm season. Increased atmospheric moisture would tend to make the lower atmosphere more emissive during the day, absorbing and reemitting outgoing longwave radiation from the surface, which would tend to keep temperatures in the urban core elevated and maintain any residual UHI from the previous night. Another possibility is that atmospheric and/or soil moisture act to influence the thermal admittance of the surface in rural areas, helping the surface to retain heat at night and keep cool during the day (Runnalls and Oke 2000).
To further understand the signals implied by the statistics in Table 6, we also examined composite difference maps that represent the change in UHI magnitude under opposing sky cover and wind speed conditions (Fig. 9). For example, on calm nights during summer, clear skies contribute to a TCMA UHI that is 0.5°–1°C more intense than when overcast skies are present (Fig. 9a). On clear nights during summer, the UHI is more than 1°C stronger when calm or light winds prevail across the TCMA than when skies are clear and winds are greater than 5 ms−1 (Fig. 9b). Local maxima are evident over the Minneapolis and St. Paul CBDs in the summertime patterns (Figs. 9a,b), as well as over Lake Minnetonka in the western TCMA (indicated with a dark blue outline). On the basis of the relative magnitudes of the summertime (Figs. 9a,b) and wintertime patterns (Figs. 9c,d), the influence of clouds and winds on the UHI is slightly greater in winter than in summer. The composite difference patterns in Fig. 9 bear a strong resemblance to the extreme UHI patterns shown in Fig. 8. This confirmatory evidence further supports the pronounced influence of weather variability on the formation and strength of the UHI on the daily time scale. The patterns also appear to be similar to the average UHI pattern shown in Fig. 3, which suggests that the same processes that drive the long-term average UHI also control the daily variability of the UHI.
f. Modulation by land–atmosphere interactions
The pronounced seasonal variability of the TCMA canopy-layer UHI that is evident in the statistics in Table 4, the diurnal curves in Fig. 4, and the patterns in Fig. 7 implies major changes in the urban energy balance throughout the course of the year. In particular, these results reveal diurnal and seasonal variations in the UHI magnitude. These differences are highlighted in Fig. 10, which shows differences between nighttime and daytime UHI magnitude in the summer (Fig. 10a) and winter (Fig. 10b) and differences between the summertime and wintertime UHI during the night (Fig. 10c) and day (Fig. 10d). The summertime pattern shown in Fig. 10a is characterized by positive anomalies throughout almost the entire domain, testifying to the dominance of the nighttime UHI. Two local maxima are evident: one centered just east of the Minneapolis CBD, and one centered over the eastern end of Lake Minnetonka. The maximum in the western TCMA is likely a result of both the cooling influence of the large lake during the day and its warming influence during the night. The anomaly is not centered directly over the lake but instead is displaced eastward, consistent with the prevailing southwesterly winds during that season. Further study is necessary to confirm the spatial extent of the lake’s influence.
The summertime pattern in Fig. 10a contrasts strongly with the wintertime pattern in Fig. 10b, wherein the frozen lake has no discernable influence on the local UHI. During winter, the daytime UHI is larger than the nighttime UHI everywhere except a small area that is centered just east of the Minneapolis CBD. Over the suburban TCMA the daytime and nighttime UHIs are roughly equivalent, and over the rural areas the daytime UHI is larger than the nighttime UHI, suggestive of the strongly emissive, often snow-covered surface. Unlike the lake, which freezes seasonally, the CBD areas remain exposed throughout the year and are a concentrated source of anthropogenic heating. Hence, the urban core exhibits a stronger nighttime UHI throughout the year. Conversely, the suburban TCMA experiences a larger daytime UHI only during the winter.
The pattern shown in Fig. 10c indicates that the nighttime UHI is larger in summer than in winter throughout the TCMA, except for the far southeastern portion of the domain. This area is unambiguously rural, which suggests the possibility that on average the UHI is advected southeastward during winter. This is consistent with the prevailing northwesterly winds during that season. Further analysis is necessary to confirm this adevection, however, possibly following the deployment of additional sensors to that sector of the TCMA.
The daytime pattern shown in Fig. 10d also supports the advection hypothesis, where wintertime UHI magnitudes are larger than summertime UHI magnitudes in the southeastern TCMA. Seasonal differences in the daytime UHI are small throughout the remainder of the TCMA domain.
4. Discussion and conclusions
The foregoing results confirm, refine, and extend the limited set of previous studies on the TCMA canopy-layer UHI performed over the past several decades (Winkler et al. 1981; Baker et al. 1985; Todhunter 1996; Malevich and Klink 2011). They represent an unprecedented level of spatiotemporal detail for the study region, facilitated by a UMN that exceeds previous TCMA network sampling by nearly an order of magnitude in time and space. For example, whereas previous studies presented monthly-mean or annual statistics (Winkler et al. 1981; Baker et al. 1985; Todhunter 1996), the results presented here are based on hourly data averaged over nighttime and daytime.
Malevich and Klink (2011) used a much smaller UMN restricted to a 6-km radius around the Minneapolis CBD in their study of the relationship between the TCMA UHI and the presence of snow; they found that the wintertime UHI is larger during daytime than at night. Our results support this finding and demonstrate that the daytime peak is also present in the suburban periphery surrounding the urban core. Similar to Malevich and Klink (2011), we hypothesize that the pattern and intensity of the daytime, wintertime UHI and wintertime differences between daytime and nighttime UHIs are due to the influence of snow, which limits the thermal mass of suburban regions by covering dark, impervious surfaces with its high-albedo, low-conductivity, highly emissive surface (Warren 1982; Baker et al. 1991). Because of these properties, snow reflects a large amount of insolation during the day and provides an efficient surface for radiative cooling at night. Municipal snow removal is ubiquitous in the developed portions of the TCMA, often reexposing dark, impervious surfaces that may be warmed more than the surrounding snow-covered surfaces (Yuan and Bauer 2007). Furthermore, urban activity tends to darken snow over the course of several days following accumulation (Ho and Valeo 2005). The average (Fig. 7e) and extreme (Fig. 8e) daytime UHI patterns are very uniform across suburban and urban areas, which may be associated with the absorption of insolation by low-albedo impervious surfaces and dirty snow in the developed portions of the TCMA. Outside the TCMA, small positive UHI anomalies are present in the seasonal average and seasonal extremes, which also may be related to absorption of insolation, in this case by non-snow-covered surfaces or structures above the snow surface. Notwithstanding these new insights, additional years of data and further analysis are necessary to test the snow hypothesis rigorously.
In addition to contributing to advances with regard to the TCMA UHI, the dense UMN employed for the purposes of this study also offers generalized insights for the UHI literature. By representing thermal source regions across a variety of LCZs (Table 1), the dense network captures much more of the heterogeneous urban–suburban environment than would a sparser network.
To briefly illustrate the impact of increased resolution offered by a dense UMN, we compare the UMN-based statistics shown in Tables 4 and 5 with identical statistics that were calculated using only the grid cells corresponding to the locations of two TCMA urban airports: KMSP in Bloomington and Flying Cloud Airport (KFCM) in Eden Prairie, Minnesota. Bias statistics based on seasonal averages are shown in Table 7; corresponding bias statistics that are based on seasonal extremes are shown in Table 8. The bias of airport-based seasonal average UHI magnitude tends to be positive for the area average, close to zero for the urban spatial 95th percentile, and negative for the suburban spatial 95th percentile. This suggests that TCMA airports will tend to overestimate the area-average UHI, that KMSP is a good proxy for the spatially extreme UHI magnitude, and that KFCM underestimates the suburban spatial extreme UHI magnitude. The same behavior is also observed with respect to the bias of seasonal extremes (Table 8).
Depending on where urban observations are located in relation to the large-scale UHI pattern, individual stations may or may not be representative of the area-average UHI or spatially extreme UHI magnitudes. Our use of a UMN underscores the notion that no one urban location is likely to be comprehensively representative of the entire metropolitan area; that is, the mesoscale features of urban climate must be characterized by multiple stations. In cities without UMNs, the canopy-layer UHI is generally monitored using existing, limited urban observations. Vehicle-based traverses offer instantaneous snapshots but may not be suitable for sustained observations. Our results indicate that this small network approach comes at the risk of biased estimates of the local UHI, particularly if the distribution of station LCZ classifications do not correspond to the distribution of LCZs represented within the city. Although the results presented here demonstrate the efficacy of dense UMNs for UHI monitoring, we did not investigate the optimal network size for describing the bulk features of the TCMA UHI, let alone local UHIs in general.
Pigeon et al. (2006) designed a UMN optimized for UHI monitoring by strategically distributing 20 sensors to sample the leading EOF of the summertime surface air potential temperature and specific humidity fields output from a 250-m-resolution mesoscale atmospheric model simulation centered over the city of Marseille in France. They showed that the leading EOF of the nighttime temperature field corresponds to the UHI and that the leading EOF of the daytime temperature field corresponds to the sea-breeze circulation. They validated their network design by demonstrating that the same EOF structures could be obtained using the entire field, only the 20 grid points nearest to the sensor locations, and the sensor observations themselves. Thus, Pigeon et al. (2006) established that a mesoscale numerical model could capture the large-scale features of the Marseille UHI, which suggests that well-calibrated numerical models may be viewed as an alternative approach to quantifying the large-scale features and area-average magnitude of UHIs. Similar to our study, however, Pigeon et al. (2006) did not address how many UMN sensors are necessary to adequately describe the UHI. Future work could establish a relation among city size, land surface heterogeneity, and optimal network density.
UMNs, such as the one described here, are ideal for deployment in other cities while emerging observation platforms mature because they are immediately available, are straightforward to control following published guidelines, and offer flexibility in spatial and temporal sampling. Technologies like cellphone-battery-derived temperature (Overeem et al. 2013) and remotely sensed land skin temperature (Jin 2012) are undoubtedly promising but require substantial assumptions to derive estimates of canopy-layer air temperature from their raw observations. The spatial and temporal resolution afforded by UMNs surpasses that of existing networks. This increase in information content facilitates the identification of spatial structures at finescales relevant in urban areas. For example, UMNs could inform mitigation and adaptation efforts within cities by singling out particular locations for investment such as hot spots and hot corridors. UMNs can also provide critical data for evaluating the efficacy of mitigation strategies. Last, UMNs that leverage volunteers are capable of strengthening links among citizens, businesses, municipal governments, and researchers, broadening the exposure and impact of urban climatological research.
This research was funded by the University of Minnesota Institute on the Environment under Discovery Grant DG-0015-11 and the U.S. National Science Foundation under Grant 1231325. We also acknowledge the support and cooperation of the many citizens, institutions, and municipalities who have provided ongoing access to their property for the temperature observations that are associated with this project.
Supplemental information related to this paper is available at the Journals Online website: http://dx.doi.org/10.1175/JAMC-D-14-0239.s1.
Current affiliation: The Climate Corporation, Seattle, Washington.
HOBO and HOBOware are registered trademarks of Onset Computer Corporation in Bourne, Massachusetts.