1. Introduction
Cities are the most distinctly anthropogenic and possibly the most consequential ecosystems on Earth. Their form, function, and management act directly on their inhabitants and indirectly on other ecosystems through consumption, waste, and pollution. As of 2010, city dwellers constituted 50% of the global population and 80% of the North American population, and these percentages are projected to reach 70% globally and 87% in North America by 2030 (Cohen 2006; WHO 2010), making the management of cities one of the central challenges of the twenty-first century. This is particularly true of small and midsized cities. Half of urban dwellers live in midsized cities with populations of 100 000–500 000, fewer than 10% live in megacities with populations over 10 million, and these proportions are expected to persist into the future (Cohen 2006). As such, small and midsized cities have a great deal of growth and expansion ahead of them, making it crucial to understand the consequences of various modes of development, including local climate effects.
Urban landscapes have distinct climates compared to their rural environs (Arnfield 2003). The most widely studied element of the urban climate is temperature, which is typically described in terms of the urban heat island (UHI) effect. UHIs are classically defined as the difference between urban and rural background temperatures, which can be represented in a variety of ways (Stewart and Oke 2012). UHIs have been observed from the Arctic Circle (Hinkel et al. 2003) to the tropics (Chow and Roth 2006) in cities ranging in population from a few hundred to tens of millions (Oke 1973). On a local scale, UHIs and other aspects of the urban climate directly affect the daily lives of over 3.5 billion city dwellers, including their health, quality of life, and resource consumption. On a global scale, ongoing climate change heightens the frequency and severity of heat waves and other climate related risks, particularly in cities (Patz et al. 2005; Luber and McGeehin 2008). Urban growth could magnify global temperature trends in the very places where most people live, exacerbating climate related impacts (Stone et al. 2010; Fischer et al. 2012). This creates a particular need to understand the climate of midsized cities where so much future growth and development is expected to occur.
Although past urban climatological research gives us a general expectation that urban areas are warmer than rural, applying that understanding toward specific urban issues—such as heat risk (Johnson et al. 2013), energy consumption (Kolokotroni et al. 2012), violent crime (Hsiang and Burke 2014), and air quality (Filleul et al. 2006)—requires detailed representations of local climates in different regions and urban landscapes. To date, nearly all urban climate sensor arrays have been deployed in large cities with populations in the millions, several of which were not developed or utilized for research purposes (Table 1; Muller et al. 2013). Of the arrays in Table 1, only Turku, Finland (population 225 000), our study area of Madison, Wisconsin (population 407 000), and Barrow, Alaska (population 5000) have urban agglomeration populations under 500 000 (Demographia 2013). Aside from the Arctic city of Barrow, ours is the first such North American city to be studied with a long-term sensor array, highlighting how little research has been done in small to midsized cities. Our study will address this by describing the spatial and temporal properties of the UHI in Madison using a dense, long-term sensor array.
Summary of urban climate sensor arrays with at least 10 sensors operating for at least one year. Population and area of urban agglomeration are taken from Demographia (2013). The number of urban sensors is estimated from maps when not explicitly reported. The primary purpose of each sensor array is denoted monitoring (mon) education (edu), and/or research (res). The data for the study area are in italics.
Our array is one of the most spatially dense and extensive ever deployed (Table 1), which could provide unique insights into the urban climate and will allow us to describe the UHI in richer detail than a study with less replication. In the UHI literature, urban climate effects are often described categorically in terms of UHI intensity, which is the temperature difference between a city and its rural surroundings. One of the most common measures of UHI intensity is UHImax (see the appendix for the definitions of abbreviations to measures used in the text), which is the single largest observed temperature difference between any urban and any rural location. However, UHImax represents a very thin slice of space and time that may not represent predominant local climatic conditions. It is fundamentally defined by temporal and spatial outliers, making it highly sensitive to exceptional sites and environmental conditions. Measures of UHI intensity based on average temperatures at several “urban” and “rural” sites, or based on average temperatures across broader slices of time, are likely to be more representative of the local climate and more robust to outlying sites or conditions. However, even these “average” UHI intensity metrics take the entire city as the unit of analysis, reducing substantial temperature variation across the landscape (Hawkins et al. 2004; Mohsin and Gough 2012) to a single number. It also relies on arbitrary definitions of urban and rural land cover, which range from overly simplistic to highly complex (Stewart and Oke 2012), making cross-study comparison or even interpretation of individual studies challenging. In short, UHI intensity is often reported as a categorical factor defined in stark terms of urban or rural, but urbanness is not a discrete or discontinuous property of the landscape. Roads, buildings, and other features of the built environment permeate the landscape in varying densities (Fig. 1), as will the magnitude of their climate effects. A categorical factor cannot capture this variation, but a continuous factor can. If the density of the built environment (%Built) has a known relationship with air temperature (Tair), then urban climate effects can be characterized using that relationship. For example, if the relationship between %Built and Tair has a coefficient of 0.05°C %−1, then a 100% increase in urban density (i.e., going from perfectly rural to perfectly urban) would yield a 5°C increase in average Tair. From that single continuous factor we can describe not just the urban and rural sites where we happen to have sensors, but all sites having %Built values. Our extensive sensor array allows us to develop such robust statistical relationships between land cover and Tair and will allow us to describe spatial and temporal variation in the UHI in highly descriptive detail.
Map of study area in Madison, including rural land-cover classes [Wisconsin Department of Natural Resources WISCLAND Land Cover (WLCGW930)], NLCD 2006% impervious surface coverage (Fry et al. 2012), and HOBO U23 Pro v2 temperature/RH sensor locations.
Citation: Journal of Applied Meteorology and Climatology 53, 10; 10.1175/JAMC-D-14-0107.1
In short, our study will build our knowledge of the urban climate in midsized cities using a dense sensor array and analytical tools that utilize the spatial and temporal richness of our dataset. Our overall objective is to characterize surface air temperature patterns and processes in a midsized temperate city (Madison). We will focus in particular on seasonal variation in urban climate effects. The analysis will proceed in three parts: 1) spatial patterns and processes over time, 2) daily variation in UHI intensity, and 3) seasonal patterns and drivers of UHI intensity.
2. Methods
a. Study area
Madison is a city of 233 000 in the north-central United States (43°N, 89°W) with an estimated 2012 urban agglomeration population of 407 000 (Demographia 2013). It has a humid-continental climate (Köppen classification: Dfa), 1981–2010 mean annual precipitation of 876 mm, and mean temperatures of −7°C in January and 22°C in July [obtained from the National Oceanic and Atmospheric Administration (NOAA) National Climatic Data Center data tools: 1981–2010 normals; available online at http://www.ncdc.noaa.gov/cdo-web/datatools/normals]. The city is surrounded by lakes and a rural landscape of agriculture, forests, wetlands, and grasslands (Fig. 1).
b. Data
In March 2012, 135 HOBO U23 Pro v2 temperature/RH dataloggers in solar radiation shields from Onset Computing were installed on streetlight and utility poles across the Madison area (Fig. 1) in cooperation with local energy utilities, governments, and institutions. Temperature sensor accuracy is ±0.21°C from 0° to 50°C with a resolution of 0.02°C at 25°C and a sensor drift of <0.1°C yr−1 (Onset Computing). Our analysis uses data from 20 March 2012 until 1 October 2013, representing approximately the first 18 months of data collection. The loggers were installed at a height of 3.5 m to minimize the risk of damage and disturbance. Loggers were positioned on the north side of poles except in six cases where they were turned east or west to avoid the public road right of way. Six more loggers were added in October 2012 and 10 more in August 2013 for a total of 151. Instantaneous measurements are recorded every 15 min and are planned to continue for several years.
c. Spatial analysis
To develop continuous relationships between land cover and Tair, we fit daily ordinary least squares regression models for average Tair at each sensor with land-cover traits as predictors. Separate models were fit for average daytime and nighttime Tair on each of 554 days for a total of 1108 models. Average Tair across the daytime (from sunrise to sunset) or nighttime (from sunset to sunrise) hours, rather than Tair at any single point in time, was used to better represent predominant climatic conditions over time.
1) Model fitting
Models generally were very well behaved in terms of normality, constant variance, and independence of predictors, but significant residual spatial autocorrelation occurred in 32% of daytime models and 16% of nighttime models. Our data are inherently spatial, and residual spatial autocorrelation can bias results substantially (Anselin 2002), so we used spatial regression for all days with residual spatial autocorrelation. In our spatial models, an inverse distance weighted neighbor matrix with a maximum neighbor distance of 10 km was used in either a spatial error or spatial lag model, which was selected using Lagrange multiplier tests. Spatial lag models were selected over spatial error models in nearly all cases. The appropriate spatial models reduced the percentage of models with residual autocorrelation from 32% to 15% during the day and from 16% to 5% at night. The remaining models with residual autocorrelation were a concern, but coefficients were quite stable between the spatial and ordinary least squares models, so we considered them acceptable for the purposes of this study, which is focused on overall patterns rather than fitting optimal models for each specific day.
2) Covariate selection
Our objective was to compare model coefficients across time, which required a constant set of covariates over the entire study period. As such, rather than creating a best fit model for each individual day, we selected covariates that provided the best average model fits over the entire study period. To select the covariates, we ran the full set of 1108 models with different combinations of factors and compared fits using Akaike information criterion (AIC). The model with the lowest average daily AIC rank was selected as the best fitting overall model. Covariates tested for inclusion in the monthly models were 1) built environment density, 2) vegetation abundance, 3) lake effect, and 4) topographic relief, which have all been found to be important spatial predictors of urban temperature patterns (e.g., Alcoforado and Andrade 2006; Eliasson and Svensson 2003; Suomi et al. 2012).
Built environment density was represented by percent impervious surface coverage (IMP). IMP was taken directly from the 30-m-resolution U.S. National Land Cover Database (NLCD) 2006 product (Fry et al. 2012), which was the most recent available edition. Average IMP within circular buffers ranging in radius from 100 to 1000 m (in increments of 100 m) were tested to see which produced the best model fits across the study period.
Vegetation abundance was characterized using normalized difference vegetation index (NDVI) computed from 10-m-resolution SPOT 5 data from two cloud-free summer dates (1 July 2010 and 25 July 2011) that together covered our study region. We were primarily interested in the relationship between NDVI and temperature during the summer months when vegetation was at its peak, though we also included it as a candidate covariate in the winter models as another measure of the local abundance of natural versus modified surfaces. Images from the two dates were combined using relative radiometric correction, by which spectral data from the second date was normalized to the first using linear regression on pseudoinvariant features (Schott et al. 1988). We sampled 100 pseudoinvariant features in overlapping areas of the two images to build regressions for the red and infrared bands. In the corrected images, a random sample of 100 locations in overlapping regions had a correlation coefficient R = 0.85 and a slope = 0.96 for the red band and R = 0.82 and slope = 0.97 for the near-infrared band, indicating good agreement. As with IMP, average NDVI within circular buffers ranging in radius from 100 to 1000 m were tested to see which produced the best model fits across the study period. NDVI and IMP were strongly negatively correlated (R = −0.91 at 500-m radius around sensors) and introduced unacceptable collinearity into the models, so only the best fitting of the two was included in the final models, which was IMP in nearly all cases.
Lake effect was a measure of proximity to the five largest lakes in the Madison region (Fig. 1). Various functions were tested for the relationship between lake proximity and monthly average Tair, including linear and exponential distance decay, to see which produced the best fit across the study period.
Topographic relief (TOPO) was the difference between local elevation and average elevation within a 0.8-km radius on a 3-m-resolution digital elevation model. We considered that TOPO is unlikely to have simple linear effects on Tair such that, for instance, nighttime temperatures grow ever colder in deeper and deeper valleys, or ever warmer on higher and higher hills. A logistic relationship seems more likely, but our study area is relatively flat with TOPO values ranging from −22 to +22 m, and more complex relationships gave no substantial improvement over raw TOPO values.
d. Spatial interpolation
Monthly average nighttime and daytime temperatures from April 2012 to March 2013 were visualized on a 400 m × 400 m resolution grid using regression kriging. Regression kriging is well known to improve interpolation skill compared to simpler methods such as inverse distance weighting and ordinary kriging (Szymanowski and Kryza 2009), which do not utilize information about the land surface. In brief, regression kriging fits a regression model to measured point data then predicts the variation explained by that model across the landscape using the same covariates. Remaining model residuals are then interpolated using ordinary kriging and added to the modeled surface.
For covariates, IMP, lake effect, and TOPO were used at night and during the day. Lake effect and TOPO were averaged within each 400-m grid cell, and IMP was equal to average percent impervious within a 500-m radius of each grid centroid (with water masked out). Spherical or exponential variogram fits were selected by AIC.
e. Temporal analysis
To analyze temporal patterns of urban temperature effects (ΔT), we first had to create a daily measure of ΔT. As will be described in results, IMP was the dominant driver of Tair in all months and times of day. IMP also is a measure of urban density and puts daily Tair patterns explicitly in terms of urban temperature effects. As such, we defined ΔT as the daily temperature effect of going from 0% to 100% IMP in the study region based on the coefficients of each day’s model. For example, if the IMP-versus-Tair relationship had a coefficient of 0.05°C %−1 on a given night, then we defined that night’s ΔT as 5°C.
1) Statistical models
Linear mixed-effect models (lme function in the R version 3.0.2 software package nlme) were used to test the relationship between ΔT and various weather and land-cover factors (see Table 3), with daytime and nighttime ΔTs treated separately. A first-order autoregressive correlation structure accounted for serial autocorrelation and improved AIC and log-likelihood values in all cases compared to models lacking an autoregressive term. Models first were run at monthly intervals to explore seasonal differences (not shown), which revealed two distinct seasons for model relationships roughly corresponding to the warmest and coldest parts of the year. We defined the cold season (n = 98 days; from 15 December 2012 to 22 March 2013) as periods with widespread snow cover. We defined the warm season (n = 294 days; from 27 April to 1 October 2012 and from 18 May to 29 September 2013) as periods with abundant green vegetation, which we further defined as days with regional average NDVI values >0.5. Regional NDVI was derived from weekly 500-m-resolution Earth Resources Observation and Science (EROS) Moderate Resolution Imaging Spectroradiometer (eMODIS) data interpolated to daily time steps. It was calculated as average NDVI across Dane County, Wisconsin (Fig. 1), excluding water bodies and areas of cloud cover. Models for the spring and fall months were intermediate between the warm- and cold-season models, so we report the warm- and cold-season results only since they offer the greatest contrast. Model fits were assessed using the rsquared.glmm algorithm in R (Martin 2013), which calculates marginal and conditional R2 values for mixed-effects models (Nakagawa and Schielzeth 2013). The predictors for these models are described below.
2) Model covariates
We considered a wide range of synoptic, seasonal, and biophysical factors to account for daily variation in ΔT (Table 2). As described in Table 2, average values from various time intervals were compared to see if there were particular critical periods for meteorological conditions to influence daily ΔT; in all models, average daytime values (sunrise to sunset) provided the best fits for ΔT both that day and the following night. Percent sun (%Sun) was used as a proxy of cloud cover and was calculated as measured insolation (300–1100 nm) as a percent of clear-sky insolation. Measured insolation was recorded on an S-LIB-M003 pyranometer (Onset Computing) every 15-min at a weather station in our study area. Clear-sky insolation was estimated as the maximum insolation recorded within one hour of each measurement in an 11-day moving window centered on that measurement in the three year period for which we had insolation data (May 2011–May 2014) . For instance, if the pyranometer recorded 500 W m−2 at 1400 central standard time (CST) on day 10, then the corresponding clear-sky insolation would be the maximum insolation recorded between 1300 and 1500 CST from day 5 to 15 of the years 2011–14 under the assumption that clear-sky conditions occurred sometime during that time period. Average %Sun between the hours of 0900 and 1500 CST consistently provided the best model fits among the various time intervals tested. Percent sun and relative humidity (RH) were strongly correlated (R = −0.81) such that cloudy days tended to have higher RH and vice versa. We wanted to test whether daytime RH affected ΔT independent of its correlation with cloud cover, so we fit a linear model of RH versus %Sun to generate the residual variation in RH (hereinafter, residual-RH) that was orthogonal to cloud cover. We then included residual-RH in the ΔT models to explore the independent effects of cloud cover and RH on ΔT.
Explanatory covariates and data sources for daily UHI intensity models.
3. Results
a. Spatial patterns
The models of land cover versus daily or nightly average Tair had adjusted-R2 values averaging 0.54 at night and 0.40 during the day, with the best fits in summer and the lowest fits in spring (Fig. 2). IMP outperformed NDVI in 91% of nighttime models and 97% of daytime models throughout the study period; IMP outperformed NDVI at even higher rates during the warm season when vegetative biomass was highest, so IMP was used in all final models. The optimal buffer radius for IMP averaged 455 ± 7 m at night (mean ± standard error) and 536 ± 10 m during the day, so we used 500-m IMP for all models, which remarkably is the same radius cited by Oke (2006) as a typical zone of influence for urban sensors. For the effect of lakes on temperature, the best relationship in nearly all months and times of day was exp(−4d), where d is kilometers from the nearest lake shore. This relationship yields a maximum lake effect of 1.0 at the shore and rapidly decreases inland, reaching 10% of its maximum value at a distance of approximately 0.6 km, which is consistent with our observations that lake effects on monthly average Tair appeared to be restricted to shoreline locations.
The 31-day moving-average adjusted-R2 values from daily regression models of Tair from sensors, with IMP, TOPO, and lake effect as predictors of daily average daytime and nighttime Tair.
Citation: Journal of Applied Meteorology and Climatology 53, 10; 10.1175/JAMC-D-14-0107.1
For the 554 nighttime models, IMP was significant (p < 0.05) on over 99% of days, lake effect on 88% of days, and TOPO on 62%. Of these, IMP was the dominant predictor, accounting for an average of 74% of explained variation compared to 19% for lake effect and 7% for TOPO. For the 554 daytime models, IMP again was significant on over 99% of days, lake effect on 59% of days, and TOPO on only 16%. Of these, IMP again was the dominant predictor, accounting for an average of 80% of explained variation compared to 16% for lake effect and 3% for TOPO.
In terms of relationships, IMP was positively related to Tair in all months and times of day. TOPO was positively related to Tair on most nights, such that lower spots on the landscape tended to be cooler. During the day, TOPO had much smaller, variable effects and often was not significant. Lake proximity nearly always warmed Tair at night (93% of nights). During the day, lake proximity generally cooled Tair in spring and summer but warmed Tair in fall, which essentially followed the seasonal lag between water and air temperatures and is comparable to results of past UHI studies (e.g., Suomi and Käyhkö 2012).
These seasonal patterns for mid-March 2012 through September 2013 are shown in Figs. 3a–c, which visualize each model coefficient as its daily effect on Tair. Figures 3a and 3b show the effect on Tair of going from 0% to 100% IMP, which is equivalent to ΔT as defined in section 2e. Both the daytime and nighttime effects peaked in August, averaging 4°C at night and 1.5°C during the day. Both effects averaged closer to 1°C during the rest of the year, with nighttime values slightly above and daytime values slightly below 1°C.
Spatial model coefficients plotted as their effects on air temperature. Daily ΔT (temperature difference between 0% and 100% impervious surface coverage) (a) during the day and (b) at night, both fitted with 31-day moving averages (red lines). Gray boxes indicate presence of snow cover. (c) The 31-day moving averages of the daily temperature effects of standing on the lake shore vs no lake effect (blue lines), and a −10-m change in TOPO (red lines).
Citation: Journal of Applied Meteorology and Climatology 53, 10; 10.1175/JAMC-D-14-0107.1
Figure 3c shows the effects of lakes and TOPO. During the day, the lakes tended to cool Tair in spring and early summer but warm Tair in the fall. The lake effect followed similar seasonal trends at night but was always positive and generally had a larger influence on Tair compared to its daytime effects. For TOPO, a 10-m drop in local topographic relief generally had effects less than 0.5°C, which was much smaller than the influence of IMP and water; −10 m is near the 95th percentile of TOPO values in the study area and is thus at the high end of typical topographic effects.
Figures 4a and 4b show how these processes influenced temperatures across the landscape on a monthly basis. For the 12 months interpolated, adjusted R2 for the trend models averaged 0.53 during the day and 0.73 at night, and root-mean-square errors averaged 0.17 ± 0.01°C (mean ± standard error) during the day and 0.31 ± 0.04°C at night. In all months and times of day, the most densely built areas (Fig. 1) were consistently the warmest (Figs. 4a,b), with higher temperatures often sharply circumscribed by the borders of the city, especially during the day. At night, wetlands and other low elevation rural areas were consistently the coldest locations, especially in May–September. During the day, rural areas tended to be relatively uniformly cool compared to the city.
The 400-m-resolution monthly average (a) nighttime and (b) daytime air temperature from April 2012 to March 2013 interpolated using regression kriging. Units are degrees Celsius difference from minimum sensor temperature. Black areas are lakes.
Citation: Journal of Applied Meteorology and Climatology 53, 10; 10.1175/JAMC-D-14-0107.1
b. Daily variation in UHI intensity
We analyzed the drivers of daytime and nighttime ΔT during the warm and cold seasons. Depending on the season and time of day, some combination of wind, cloud cover, soil moisture, residual-RH, and snow depth were the most important predictors. Several other factors in Table 2 were also significant, but we ultimately chose wind speed, cloud cover, soil moisture, residual-RH, and snow depth as the combination most useful for explaining ΔT. For example, diurnal temperature range, soil temperature, and daily maximum temperature were often significantly related to ΔT but were also highly correlated with %Sun, did not improve model fits, and did not offer information beyond what %Sun and other factors already provided.
The major results of our daily ΔT models included the following: 1) Wind speed and cloud cover were significantly (p ≪ 0.001) negatively related to ΔT during all seasons and times of day. 2) Soil moisture was significantly (p < 0.01) negatively related to ΔT during warm-season nights only; it was not significant on warm-season days or at any time during the cold season (p > 0.8). 3) Residual-RH was negatively associated with ΔT during warm-season days and nights (p ≪ 0.001) but not during the cold season. 4) Snow depth was positively associated with daily ΔT during cold-season nights (p ≪ 0.001) but only marginally significant during cold-season days (p < 0.1).
To quantify the relative effects of these factors, we compared the effects on ΔT of changing each predictor from −1 to +1 standard deviation of the mean values recorded during our study period (Table 3). One standard deviation comprises 68% of the data for each weather variable and is intended to be representative of typical changes in each. During the warm season, one standard deviation changes in wind and residual-RH had comparable effects on nighttime ΔT, with clouds and soil moisture having somewhat smaller but still significant effects (Table 3). During the cold season, ΔT was sensitive to daily variation in snow depth but effects were only marginally significant during the day (p < 0.1). The relative effects of wind were lower during the day than at night and were generally consistent between warm- and cold-season days (17% and 23%, respectively) and between warm- and cold-season nights (38% and 34%, respectively). Cloud effects, on the other hand, were lower during the warm season than the cold season but were generally consistent between cold-season nights and days (55% and 58%) and warm-season nights and days (29% and 26%). Notably, the absolute magnitude of ΔT was also lower during the cold season (Figs. 3a,b), so in terms of relative ΔT effects, cold- and warm-season cloud-cover effects were quite comparable.
Effect of changing from −1 standard deviation to +1 standard deviation (magnitude of changes is in parentheses) in mean observed wind speed, cloud cover, soil moisture, relative humidity (independent of its correlation with cloud cover), and snow depth on the average magnitude of ΔT (temperature difference between 0% and 100% impervious surface coverage) during the day or at night during the warm or cold season. Marginal and conditional R2 is calculated using the R algorithm rsquared.glmm (Martin 2013). Here, Ns signifies statistical nonsignificance at α = 0.05; all reported percentages are statistically significant except day winter snow depth (p < 0.1).
c. Seasonal trends in UHI intensity
During our 18 months of observation, both daytime and nighttime ΔT peaked in summer and declined in spring and fall (Figs. 3a,b). Daytime ΔT underwent a secondary peak during the winter, which coincided with snow cover (Fig. 3b). We do not yet have data from other, snow-free winter periods for comparison, but the correspondence between the presence of snow cover and a rise in daytime ΔT was quite distinct, and other possible confounding factors such as temperature and day length did not appear to matter. Snow cover also affected ΔT at night (Table 3), though nighttime ΔT was sensitive to snow depth (Fig. 6) whereas daytime ΔT displayed a more constant response to the presence of widespread snow cover (Figs. 5, bottom, and 6). The distinction between snow cover and snow depth is important here, because daytime UHI intensity was sensitive to the former (Fig. 5, bottom) but only marginally sensitive to the latter (Table 3).
Seasonal patterns of ΔT (temperature difference between 0% and 100% impervious surface coverage) and its potential drivers. Black lines are 31-day moving averages of daily average ΔT during the day (dotted lines, outer-left y axis) and night (solid lines, inner-left y axis). Red lines are 31-day moving averages of (top) daily average wind speed, (middle) %Sun between 0900 and 1500 CST, and (bottom) regional average NDVI.
Citation: Journal of Applied Meteorology and Climatology 53, 10; 10.1175/JAMC-D-14-0107.1
Snow depth vs ΔT (temperature difference between 0% and 100% impervious surface coverage). Lines are 11-day moving averages.
Citation: Journal of Applied Meteorology and Climatology 53, 10; 10.1175/JAMC-D-14-0107.1
Other factors also appeared to influence UHI seasonality. From Table 3, we know that wind, clouds, residual-RH, snow, and soil moisture all affected day-to-day variation in ΔT. Seasonally, we found that monthly average variation in wind and %Sun generally tracked monthly average variation in ΔT during both the day and night (Fig. 5). That is, wind speed and cloud cover were generally lowest during the summer in Madison, which favored higher average ΔTs during those months. Soil moisture also tended to be lowest during the summer months (not shown), which during the night also would favor higher ΔT (Table 3).
This accounts for some of the seasonal trends, but clear, calm summer nights still tended to have much higher ΔTs than clear, calm winter nights (cf. cold- and warm-season peak daily values in Fig. 3a). So something other than clouds and wind underlies the annual trends in Fig. 5. Many factors follow similar annual trends as ΔT (e.g., Tair, absolute humidity, and vegetative biomass), with peaks in the summer and troughs in the winter. As seen in Fig. 5, regional average NDVI demonstrated good agreement with daytime and nighttime ΔT. We will describe possible mechanisms by which vegetation could affect ΔT in the discussion.
To summarize our observations on the seasonality of the UHI, 1) annual trends in wind speed and cloud cover tracked annual trends in UHI intensity; 2) annual trends in vegetation cover and other factors also tracked seasonal changes in UHI intensity; and 3) snow cover affected both daytime and nighttime UHI intensity; nighttime ΔT was sensitive to daily variation in snow depth while daytime ΔT appeared to respond primarily to the presence of widespread snow cover rather than to depth.
4. Discussion
The UHI is well understood as a general phenomenon, so it is not surprising that the basic features of our results are consistent with past work. Urban–rural temperature differences were largest at night under calm, clear conditions, as is true in other cities (Arnfield 2003). Seasonally, UHI intensity in our study area peaked in late summer, which is consistent with most (Arnfield 2003), but not all (Figuerola and Mazzeo 1998; Ichinose et al. 1999; Kim and Baik 2005; Montávez et al. 2000) past results and also agrees with previous studies in the north-central United States (Ackerman 1985; Sanderson et al. 1973). Spatially, temperatures in Madison were positively related to the density of the built environment (and corresponding sparsity of vegetation) and were modified locally by water bodies and topography, all of which is consistent with previous studies (Bottyán and Unger 2003; Hjort et al. 2011; Suomi et al. 2012; Szymanowski and Kryza 2012). However, our data revealed insights beyond these generalities, as discussed below.
a. Temporal drivers
The mechanisms of wind and cloud effects on UHI intensity have been thoroughly described in past work (Oke 1982). Clear skies allow larger shortwave gains and longwave losses, facilitating greater urban–rural temperature divergence. Low wind speeds limit horizontal air movement and allow spatial heterogeneity in Tair to develop and persist, while stronger winds tend to mix and homogenize adjacent urban and rural air masses. Our results are consistent with these effects during both the night and daylight hours.
Higher soil moisture has been hypothesized to lower UHI intensity by increasing the thermal admittance of rural soils, thereby raising the resistance of rural temperatures to change and limiting urban–rural temperature divergence (Runnalls and Oke 2000). However, to our knowledge Runnalls and Oke (2000) is the only other field study to have tested this. Our data showed that soil moisture had significant (p < 0.01) negative effects on UHI intensity during warm-season nights (Table 3) but not during warm-season days or during the cold season (p > 0.8). We did not expect soil moisture to be important when snow covered the ground, but we were surprised that soil moisture did not affect UHI intensity during warm-season days. There was seemingly a biophysical mechanism by which additional soil moisture reduced urban–rural temperature differences at night but not during the day. Rural surfaces are probably more biophysically sensitive to soil moisture than urban surfaces, which are characterized by impervious cover (Fig. 1), so we assume that the explanation lies primarily in an effect of soil moisture on the rural landscape. In general, additional moisture in vegetated rural soils has several related effects, including 1) increased evapotranspiration, which increases the proportion of the energy budget partitioned to latent heat; 2) increased water content of soils, which increases thermal admittance and resistance to sensible temperature changes (this is the mechanism proposed by Runnalls and Oke 2000); and 3) increased near-surface humidity, which could similarly increase resistance of the air to sensible temperature changes and also could contribute to ground fog formation and surface inversions. All three effects are fundamentally interconnected. The question is: Which of them would alter temperature trends at night but not during the day? The precise combination of mechanisms is likely to be multifaceted. Testing among the possibilities would require measuring diurnal energy budgets on urban and rural land covers and determining which elements of the local rural energy budget are most sensitive to soil moisture at different times of day.
Relative humidity also is hypothesized to affect UHI intensity, but only a few field studies have tested this, each finding that RH was negatively related to UHI intensity (Hoffmann et al. 2012; Kim and Baik 2004, 2002). However, ours is the first analysis to explicitly deal with the collinearity between RH and cloud cover to test for independent effects of RH. We found that, independent of its correlation with cloud cover, RH still was negatively related to ΔT during the warm season, though not during the cold season. We hypothesize that warm-season RH effects were due to the greater resistance of moist air to temperature changes. That is, moister air will tend to cool and warm more slowly than dry air, and urban–rural temperatures will tend to diverge less. To test this, we ran a mixed-effects model during the warm season for residual-RH versus daily average cooling rates from 0 to 3 h after sunset, which is often reported as the peak period of UHI formation (Oke 1982). We found a highly significant (p ≪ 0.001) relationship, with lower cooling rates under higher humidity conditions (marginal R2 of 0.07 and conditional R2 of 0.24), which is consistent with our hypothesis. In the cold season, there simply may be too little absolute moisture content in the air for a percent change to substantially alter the energy budget.
Although we separated RH and cloud cover in our analysis, the strong correlation between the two variables could be useful for studies modeling daily UHI intensity. Relative humidity data often are easier to acquire and more accurate than cloud-cover data, and in our models RH alone provided comparable or even slightly better model fits than cloud cover alone, at least during the warm season (not shown). As such, if cloud-cover data are unavailable, RH may serve as a reasonable proxy in addition to its independent effects on UHI intensity.
b. Seasonality
Seasonality is evident in every aspect of our analysis. Figures 3 and 4 show clear seasonal trends in UHI intensity during both the night and day. Similarly, Table 3 and Fig. 3 show that the relationships of temperature with land cover, weather, and other factors also vary seasonally. So not only the patterns but the processes themselves change over time. Although many studies have reported seasonal changes in the UHI, relatively few have described this in detail or systematically assessed its causes. Proposed drivers have included seasonal changes in wind and clouds (Fortuniak et al. 2006; Kim and Baik 2005, 2002; Kłysik and Fortuniak 1999; Morris et al. 2001; Shahgedanova et al. 1997; Unger et al. 2000; Yamashita et al. 1986), prevalence of anticyclonic conditions (Jauregui et al. 1992; Tumanov 1999), soil moisture changes (Runnalls and Oke 2000), and day length (Yang et al. 2013). Some of these are consistent with our data and others are not. For instance, peak day length (and peak daily insolation) occurs in mid-June, which lagged peak UHI intensity by 1–2 months in both years we recorded (Fig. 3). Our results were generally consistent with the other hypotheses listed above, however, with wind and cloud cover generally tracking ΔT during year. Nonetheless, wind and clouds do not fully explain our seasonal trends, and we suspect that vegetation played an underlying role (Fig. 5).
Several climatic factors follow similar annual trends as vegetation, including air temperature, absolute humidity, and soil temperature. We have sampled fewer than two annual cycles, so we cannot robustly test which of these correlated factors best explain annual UHI trends. We can hypothesize, however, that higher plant biomass during the summer favors higher UHIs by maximizing differences in the urban–rural energy budgets, particularly the partitioning between latent and sensible heat as evapotranspiration increases and reaches a maximum during the summer months. The rural landscape around Madison is dominated by agricultural crops such as corn and soybean (Fig. 1), which typically are planted in late spring, reach peak biomass in mid- to late summer, and undergo senescence during September. This corresponds well with the general rise and fall of ΔT throughout the growing season (Fig. 3). The rapid drop in nighttime ΔT between September and October 2012 (Fig. 5) coincides particularly well with autumn senescence and harvest of crops and foliage, and the rise in ΔT in spring coincides closely with spring greening. As such, we propose that regional vegetation and snow-cover conditions set seasonal baselines for ΔT and factors like wind and clouds modify daily ΔTs around that baseline. The daily data points in Fig. 3 support the existence of a UHI intensity baseline determined by seasonal land-cover conditions. At night, virtually all ΔT values > 4°C occurred during the summer, with almost no values < 2°C in the summer despite being common during the rest of the year. Similarly, daytime ΔT > 1.5°C was common in summer but very rare during the rest of the snow-free season, and values < 1°C were rare in summer but common over the rest of the year. And despite being more prevalent in summer (Fig. 5), windless, cloudless conditions occurred in all seasons, so clearly some background factor such as vegetation shifted baseline UHI intensities throughout the year. We will continue to assess this hypothesis as we collect additional years of data.
Ours is one of the first studies to report an effect of snow cover on UHI intensity. Aside from a study in Minneapolis, Minnesota (Malevich and Klink 2011), we are unaware of any comparable results. Like Malevich and Klink (2011), we found that snow increased ΔT and that the effect differed between the daylight and night hours. Daytime ΔT displayed a relatively constant response to the presence of widespread snow cover but was insensitive to snow depth, whereas nighttime ΔT was sensitive to daily variation in snow depth (Fig. 6; Table 3). Both effects appeared to be on the order of 0.5°C, though it is difficult to estimate precisely without snow-free winter periods for comparison. We hypothesize that the daytime effect was related to albedo, with city streets, building walls, and other artificial surfaces being relatively free of snow cover in comparison with the more uniform high albedo snow blanketing rural areas. An albedo effect would be expected to be nearly constant as long as snow depth is sufficient to cover the ground, which is what we observed. The nighttime effect is likely due to the low thermal admittance and conductivity of snow, such that thicker layers of snow create greater insulating effects, causing nighttime ΔT to respond to snow depth (Fig. 6). Similar biophysical explanations are favored by Malevich and Klink (2011).
5. Conclusions
Spatial and temporal variation is fundamental to the UHI. Classical measures of UHI intensity, such as UHImax, are categorical in nature and take the entire city as the unit of analysis. Using 18 months of data from an array of up to 151 sensors, we defined UHI intensity using a continuous factor (ΔT) that allowed us to scale our results across the landscape as well as to link spatial and temporal variation in the UHI and its drivers. The ΔT is the estimated temperature effect of going from 0% to 100% IMP, so if ΔT were 5°C on a given night, then a residential neighborhood with 30% IMP would have an estimated ΔT0%–30% of 1.5°C. Lake proximity and topographic relief are independent of IMP, but their relationships with temperature (Fig. 3) can similarly be used to estimate temperature differences across the landscape. Figures 4a and 4b visualize these interactions, showing that rural areas were coolest, dense urban areas warmest, and low-density residential neighborhoods moderate in temperature, with lake proximity and topographic relief modifying these patterns locally. So in essence Figs. 3 and 4 contain the same information, with one visualizing it as monthly spatial patterns and the other as a time series of daily values. This illustrates the utility of ΔT for representing the fundamental linkages between spatial and temporal variation in the UHI and its drivers.
In addition, by using a continuous rather than a categorical measure of UHI intensity, our results are relatively free from site selection bias, unlike metrics such as UHImax. For example, we observed a UHImax of >10°C on 14% of all nights (i.e., once per week), but this tells us very little about predominant conditions across most of the landscape. The ΔT is a much richer source of information for describing the local climate and should aid cross-city comparisons.
The above analysis is focused on surface meteorological conditions and drivers, but the influence of larger-scale synoptic processes, such as wind direction and boundary layer depth, could be the subject of future work, as could detailed case studies of particular days and nights to understand how synoptic and surface processes unfold across the landscape over shorter time periods than those considered here. Additionally, although we strongly advise against applying our specific coefficients to other cities with different land cover, climatic, and built environment characteristics, a comparison of how the IMP-versus-Tair relationship varies from city to city would be of significant interest. The National Land Cover Database from which we took percent impervious values is nationally and freely available, so direct intercity comparison is limited only by a lack of temperature data. The proliferation of high-density urban sensor arrays (Table 1) gives us an opportunity to address this and other questions and to contribute new insights to urban climatology.
Acknowledgments
This work was supported by the Water Sustainability and Climate Program of the National Science Foundation (DEB-1038759) with further support from a UW–Madison University Fellowship. We give our thanks to our community partners for use of their streetlight and utility poles to make our sensor array possible: Madison Gas & Electric (R. J. Hess), Alliant Energy (Jeff Nelson), the City of Madison (Dan Dettmann), Madison City Parks (Kay Rutledge), Sun Prairie Utilities (Rick Wicklund and Karl Dahl), Waunakee Utilities (Dave Dresen), the UW–Madison Arboretum (Brad Herrick), the City of Fitchburg (Holly Powell), UW–Madison (Kurtis Johnson), and Dane County (Darren Marsh). We also give our thanks to Annemarie Schneider, Tracey Holloway, Katherine Curtis, and Ankur Desai for helpful discussions and support and to Ankur Desai for reading an earlier version of the manuscript.
APPENDIX
Abbreviation Key
%Sun Measured insolation as a percent of clear-sky insolation
AIC Akaike information criterion
IMP Percent impervious surface coverage
NDVI Normalized difference vegetation index
TOPO Local topographic relief
RH Relative humidity
Tair Air temperature (at 3.5-m height)
UHI Urban heat island
UHImax Greatest observed temperature difference between an urban and rural location
REFERENCES
Ackerman, B., 1985: Temporal march of the Chicago heat island. J. Climate Appl. Meteor., 24, 547–554, doi:10.1175/1520-0450(1985)024<0547:TMOTCH>2.0.CO;2.
Alcoforado, M.-J., and H. Andrade, 2006: Nocturnal urban heat island in Lisbon (Portugal): Main features and modelling attempts. Theor. Appl. Climatol., 84, 151–159, doi:10.1007/s00704-005-0152-1.
Anselin, L., 2002: Under the hood: Issues in the specification and interpretation of spatial regression models. Agric. Econ., 27, 247–267, doi:10.1111/j.1574-0862.2002.tb00120.x.
Arnfield, A. J., 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, doi:10.1002/joc.859.
Basara, J. B., and Coauthors, 2011: The Oklahoma City Micronet. Meteor. Appl.,18, 252–261, doi:10.1002/met.189.
Birmingham Urban Climate Laboratory, cited 2014: HiTemp project: High density measurements within the urban environment. [Available online at http://www.birmingham.ac.uk/schools/gees/centres/bucl/hitemp/index.aspx.]
Bottyán, Z., and J. Unger, 2003: A multiple linear statistical model for estimating the mean maximum urban heat island. Theor. Appl. Climatol., 75, 233–243, doi:10.1007/s00704-003-0735-7.
Chang, B., H.-Y. Wang, T.-Y. Peng, and Y.-S. Hsu, 2010: Development and evaluation of a city-wide wireless weather sensor network. Educ. Technol. Soc., 13, 270–280.
Chow, W. T. L., and M. Roth, 2006: Temporal dynamics of the urban heat island of Singapore. Int. J. Climatol., 26, 2243–2260, doi:10.1002/joc.1364.
Cohen, B., 2006: Urbanization in developing countries: Current trends, future projections, and key challenges for sustainability. Technol. Soc., 28, 63–80, doi:10.1016/j.techsoc.2005.10.005.
Demographia, cited 2013: Demographia world urban areas (urban agglomerations). 9th Annual Edition. [Available online at http://www.demographia.com/db-worldua.pdf.]
Eliasson, I., and M. K. Svensson, 2003: Spatial air temperature variations and urban land use—A statistical approach. Meteor. Appl., 10, 135–149, doi:10.1017/S1350482703002056.
Figuerola, P. I., and N. A. Mazzeo, 1998: Urban-rural temperature differences in Buenos Aires. Int. J. Climatol., 18, 1709–1723, doi:10.1002/(SICI)1097-0088(199812)18:15<1709::AID-JOC338>3.0.CO;2-I.
Filleul, L., and Coauthors, 2006: The relation between temperature, ozone, and mortality in nine French cities during the heat wave of 2003. Environ. Health Perspect., 114, 1344–1347, doi:10.1289/ehp.8328.
Fischer, E. M., K. W. Oleson, and D. M. Lawrence, 2012: Contrasting urban and rural heat stress responses to climate change. Geophys. Res. Lett., 39, L03705, doi:10.1029/2011GL050576.
Fortuniak, K., K. Kłysik, and J. Wibig, 2006: Urban–rural contrasts of meteorological parameters in Łódź. Theor. Appl. Climatol., 84, 91–101, doi:10.1007/s00704-005-0147-y.
Fry, J., and Coauthors, 2012: Completion of the 2006 National Land Cover Database for the conterminous United States. Photogramm. Eng. Remote Sens.,77, 858–864.
Gedzelman, S. D., S. Austin, R. Cermak, N. Stefano, S. Partridge, S. Quesenberry, and D. A. Robinson, 2003: Mesoscale aspects of the urban heat island around New York City. Theor. Appl. Climatol., 75, 29–42, doi:10.1007/s00704-002-0724-2.
Hawkins, T. W., A. J. Brazel, W. L. Stefanov, W. Bigler, and E. M. Saffell, 2004: The role of rural variability in urban heat island determination for Phoenix, Arizona. J. Appl. Meteor., 43, 476–486, doi:10.1175/1520-0450(2004)043<0476:TRORVI>2.0.CO;2.
Hinkel, K. M., and F. E. Nelson, 2007: Anthropogenic heat island at Barrow, Alaska, during winter: 2001–2005. J. Geophys. Res., 112, D06118, doi:10.1029/2006JD007837.
Hinkel, K. M., F. E. Nelson, A. E. Klene, and J. H. Bell, 2003: The urban heat island in winter at Barrow, Alaska. Int. J. Climatol., 23, 1889–1905, doi:10.1002/joc.971.
Hjort, J., J. Suomi, and J. Käyhkö, 2011: Spatial prediction of urban–rural temperatures using statistical methods. Theor. Appl. Climatol., 106, 139–152, doi:10.1007/s00704-011-0425-9.
Hoffmann, P., O. Krueger, and K. H. Schlünzen, 2012: A statistical model for the urban heat island and its application to a climate change scenario. Int. J. Climatol., 32, 1238–1248, doi:10.1002/joc.2348.
Hsiang, S. M., and M. Burke, 2014: Climate, conflict, and social stability: What does the evidence say? Climatic Change,123, 39–55, doi:10.1007/s10584-013-0868-3.
Hung, T. K., and O. C. Wo, 2012: Development of a Community Weather Information Network (Co-WIN) in Hong Kong. Weather, 67, 48–50, doi:10.1002/wea.1883.
Ichinose, T., K. Shimodozonob, and K. Hanakib, 1999: Impact of anthropogenic heat on urban climate in Tokyo. Atmos. Environ., 33, 3897–3909, doi:10.1016/S1352-2310(99)00132-6.
Jauregui, E., L. Godinez, and F. Cruz, 1992: Aspects of heat-island development in Guadalajara, Mexico. Atmos. Environ., 26B, 391–396, doi:10.1016/0957-1272(92)90014-J.
Johnson, D. P., J. J. Webber, K. U. B. Ravichandra, V. Lulla, and A. C. Stanforth, 2013: Spatiotemporal variations in heat-related health risk in three midwestern US cities between 1990 and 2010. Geocarto Int., 29, 65–84, doi:10.1080/10106049.2013.799718.
Kim, Y.-H., and J.-J. Baik, 2002: Maximum urban heat island intensity in Seoul. J. Appl. Meteor., 41, 651–659, doi:10.1175/1520-0450(2002)041<0651:MUHIII>2.0.CO;2.
Kim, Y.-H., and J.-J. Baik, 2004: Daily maximum urban heat island intensity in large cities of Korea. Theor. Appl. Climatol., 79, 151–164, doi:10.1007/s00704-004-0070-7.
Kim, Y.-H., and J.-J. Baik, 2005: Spatial and temporal structure of the urban heat island in Seoul. J. Appl. Meteor., 44, 591–605, doi:10.1175/JAM2226.1.
Kolokotroni, M., and R. Giridharan, 2008: Urban heat island intensity in London: An investigation of the impact of physical characteristics on changes in outdoor air temperature during summer. Sol. Energy, 82, 986–998, doi:10.1016/j.solener.2008.05.004.
Kolokotroni, M., X. Ren, M. Davies, and A. Mavrogianni, 2012: London’s urban heat island: Impact on current and future energy consumption in office buildings. Energy Build., 47, 302–311, doi:10.1016/j.enbuild.2011.12.019.
Kłysik, K., and K. Fortuniak, 1999: Temporal and spatial characteristics of the urban heat island of Łódź, Poland. Atmos. Environ., 33, 3885–3895, doi:10.1016/S1352-2310(99)00131-4.
Lee, S.-H., and J.-J. Baik, 2010: Statistical and dynamical characteristics of the urban heat island intensity in Seoul. Theor. Appl. Climatol., 100, 227–237, doi:10.1007/s00704-009-0247-1.
Luber, G., and M. McGeehin, 2008: Climate change and extreme heat events. Amer. J. Prev. Med., 35, 429–435, doi:10.1016/j.amepre.2008.08.021.
Malevich, S. B., and K. Klink, 2011: Relationships between snow and the wintertime Minneapolis urban heat island. J. Appl. Meteor. Climatol., 50, 1884–1894, doi:10.1175/JAMC-D-11-05.1.
Martin, P., cited 2013: R squared for mixed models. Ecology for a crowded planet. [Available online at http://ecologyforacrowdedplanet.wordpress.com/2013/02/26/r-squared-for-mixed-models/.]
Mikami, T., H. Ando, W. Morishima, T. Izumi, and T. Shioda, 2003: A new urban heat island monitoring system in Tokyo. Proc. Fifth Annual Int. Conf. on Urban Climate, Lodz, Poland, International Association for Urban Climate, 0.3.5. [Available online at http://nargeo.geo.uni.lodz.pl/~icuc5/text/O_3_5.pdf.]
Mingguang Li, R., and M. Roth, 2009: Spatial variation of the canopy level urban heat island in Singapore. Extended Abstracts, Seventh Int. Conf. on Urban Climate, Yokohama, Japan, International Association for Urban Climate, P3-36. [Available online at http://www.ide.titech.ac.jp/~icuc7/extended_abstracts/pdf/384784-1-090123163421-004.pdf.]
Mohsin, T., and W. A. Gough, 2012: Characterization and estimation of urban heat island at Toronto: Impact of the choice of rural sites. Theor. Appl. Climatol., 108, 105–117, doi:10.1007/s00704-011-0516-7.
Montávez, J. P., A. Rodríguez, and J. I. Jiménez, 2000: A study of the urban heat island of Granada. Int. J. Climatol., 20, 899–911, doi:10.1002/1097-0088(20000630)20:8<899::AID-JOC433>3.0.CO;2-I.
Morris, C. J. G., I. Simmonds, and N. Plummer, 2001: Quantification of the influences of wind and cloud on the nocturnal urban heat island of a large city. J. Appl. Meteor., 40, 169–182, doi:10.1175/1520-0450(2001)040<0169:QOTIOW>2.0.CO;2.
Muller, C. L., L. Chapman, C. S. B. Grimmond, D. T. Young, and X. Cai, 2013: Sensors and the city: A review of urban meteorological networks. Int. J. Climatol., 33, 1585–1600, doi:10.1002/joc.3678.
Nakagawa, S., and H. Schielzeth, 2013: A general and simple method for obtaining R2 from generalized linear mixed-effects models. Methods Ecol. Evol., 4, 133–142, doi:10.1111/j.2041-210x.2012.00261.x.
Oke, T. R., 1973: City size and the urban heat island. Atmos. Environ., 7, 769–779, doi:10.1016/0004-6981(73)90140-6.
Oke, T. R., 1982: The energetic basis of the urban heat island. Quart. J. Roy. Meteor. Soc., 108, 1–24.
Oke, T. R., 2006: Initial guidance to obtain representative meteorological observations at urban sites. WMO Instruments and Observing Methods Rep. 81, WMO/TD 1250, 47 pp. [Available online at https://www.wmo.int/pages/prog/www/IMOP/publications/IOM-81/IOM-81-UrbanMetObs.pdf.]
Oswald, E. M., R. B. Rood, K. Zhang, C. J. Gronlund, M. S. O’Neill, J. L. White-Newsome, S. J. Brines, and D. G. Brown, 2012: An investigation into the spatial variability of near-surface air temperatures in the Detroit, Michigan, metropolitan region. J. Appl. Meteor. Climatol., 51, 1290–1304, doi:10.1175/JAMC-D-11-0127.1.
Patz, J. A., D. Campbell-Lendrum, T. Holloway, and J. A. Foley, 2005: Impact of regional climate change on human health. Nature, 438, 310–317, doi:10.1038/nature04188.
Runnalls, K. E., and T. R. Oke, 2000: Dynamics and controls of the near-surface heat island of Vancouver, British Columbia. Phys. Geogr., 21, 283–304.
Sanderson, M., I. Kumanan, T. Tanguay, and W. Schertzer, 1973: Three aspects of the urban climate of Detroit-Windsor. J. Appl. Meteor., 12, 629–638, doi:10.1175/1520-0450(1973)012<0629:TAOTUC>2.0.CO;2.
Schott, J. R., C. Salvaggio, and W. J. Volchok, 1988: Radiometric scene normalization using pseudoinvariant features. Remote Sens. Environ., 26, 1–16, doi:10.1016/0034-4257(88)90116-2.
Shahgedanova, M., T. P. Burt, and T. D. Davies, 1997: Some aspects of the three-dimensional heat island in Moscow. Int. J. Climatol., 17, 1451–1465, doi:10.1002/(SICI)1097-0088(19971115)17:13<1451::AID-JOC201>3.0.CO;2-Z.
Stewart, I. D., and T. R. Oke, 2012: Local climate zones for urban temperature studies. Bull. Amer. Meteor. Soc., 93, 1879–1900, doi:10.1175/BAMS-D-11-00019.1.
Stone, B., J. J. Hess, and H. Frumkin, 2010: Urban form and extreme heat events: Are sprawling cities more vulnerable to climate change than compact cities? Environ. Health Perspect., 118, 1425–1428, doi:10.1289/ehp.0901879.
Suomi, J., and J. Käyhkö, 2012: The impact of environmental factors on urban temperature variability in the coastal city of Turku, SW Finland. Int. J. Climatol., 32, 451–463, doi:10.1002/joc.2277.
Suomi, J., J. Hjort, and J. Käyhkö, 2012: Effects of scale on modeling the urban heat island in Turku, SW Finland. Climate Res., 55, 105–118, doi:10.3354/cr01123.
Szymanowski, M., and M. Kryza, 2009: GIS-based techniques for urban heat island spatialization. Climate Res., 38, 171–187, doi:10.3354/cr00780.
Szymanowski, M., and M. Kryza, 2012: Local regression models for spatial interpolation of urban heat island—An example from Wrocław, SW Poland. Theor. Appl. Climatol., 108, 53–71, doi:10.1007/s00704-011-0517-6.
Tumanov, S., 1999: Influences of the city of Bucharest on weather and climate parameters. Atmos. Environ., 33, 4173–4183, doi:10.1016/S1352-2310(99)00160-0.
Unger, J., Z. Bottyán, Z. Sumeghy, and A. Gulyas, 2000: Urban heat island development affected by urban surface factors. Idojaras, 104, 253–268.
WHO, 2010: Hidden cities: Unmasking and overcoming health inequities in urban settings. World Health Organization, UN Habitat, Kobe, Japan, 126 pp.
Yamashita, S., K. Sekine, M. Shoda, K. Yamashita, and Y. Hara, 1986: On relationships between heat island and sky view factor in the cities of Tama River basin, Japan. Atmos. Environ., 20, 681–686, doi:10.1016/0004-6981(86)90182-4.
Yang, P., G. Ren, and W. Liu, 2013: Spatial and temporal characteristics of Beijing urban heat island intensity. J. Appl. Meteor. Climatol., 52, 1803–1816, doi:10.1175/JAMC-D-12-0125.1.
Yow, D. M., and G. J. Carbone, 2006: The urban heat island and local temperature variations in Orlando, Florida. Southeast. Geogr., 46, 297–321, doi:10.1353/sgo.2006.0033.