Urban heat islands (UHIs) are one of the best-recorded incidences of anthropogenic climate change. Studies from across the globe have examined this phenomenon, but relatively few have focused on cold-winter cities and the effects of snow cover and snowfall. This study uses hourly temperature measurements from 1 December 2008 through 28 February 2009 at 22 urban sites in Minneapolis, Minnesota, to investigate the effect of snow cover and snowfall on the UHI. Snow effects on temperature are estimated for calm conditions using a linear mixed-effect (LME) model. For the winter of 2008/09, the average wintertime UHI was approximately 1.0°C, with a peak near midday rather than at night. The LME model results suggest that a snow cover of 5 cm or more increased the magnitude of the UHI by about 1.0°C during the day and by about 0.5°C at night. The model also indicates that the presence of moderate snowfall decreased the magnitude of the UHI by up to 2.0°C, although this result is based on a comparatively small number of events. The high albedo of snow is posited to contribute to the enhanced UHI during the day, and snow’s insulating properties are likely contributors to the characteristics of the nighttime UHI.
Urbanization is known to cause significant changes in the climatic properties of an area. Perhaps the most thoroughly documented consequence of this “inadvertent climate modification” is the urban heat island (UHI; Oke 1988). The magnitude of the UHI (ΔTu−r) is roughly defined as the difference between the air temperature of the urban area Tu and the air temperature of the surrounding rural area Tr. Studies in cities across the globe have explored the meteorological and climatological aspects of this phenomenon. The majority of urban climate research has been conducted in cities with warm-to-moderate climates (e.g., Yow 2007). There have been fewer investigations of UHIs in cities with significant seasonal temperature variation and/or seasonal snow cover [exceptions include Minneapolis, Minnesota (Todhunter 1996; Winkler et al. 1981); Barrow, Alaska (Hinkel et al. 2003); Lódź, Poland (Offerle et al. 2006); and Hamburg, Germany (Schlünzen et al. 2010)]. The Montreal Urban Snow Experiment (Lemonsu et al. 2008) was launched more recently to help to fill this gap in our knowledge of how snow—and cold conditions in general—interact with the urban canopy layer of a cold-winter city.
It is well known that snow cover has a cooling effect on air temperature (e.g., Baker et al. 1992), and a sufficient snow cover can effectively mask the effects of surface heterogeneity, at least for bare soil and vegetated surfaces (Baker et al. 1991). This study seeks to determine whether snow cover has an effect on the magnitude and spatial pattern of the urban heat island within Minneapolis. The study is guided by three questions: 1) What is the general magnitude of the wintertime UHI in Minneapolis? 2) Do snow events and/or the presence of snow cover alter the magnitude of the wintertime UHI? 3) How may the intensity of urbanization within the city affect the magnitude of the wintertime UHI?
Minneapolis is located near the geographic center of North America (Fig. 1) and has a continental climate with cold, snowy winters. Snow cover typically is established by late November and persists through mid-March (Kuehnast et al. 1982). The area has little topographical relief and is undisturbed by mountains or large bodies of water, although small lakes and the Mississippi River can influence the local climate. A summary of select climate variables for the study’s observation period is shown in Table 1.
The UHI observations in this research extend from 1 December 2008 through 28 February 2009. From a climatological perspective, temperatures in Minneapolis are below freezing from December through February (DJF) (1971–2000 averaging period; NCDC 2009a), with a mean maximum temperature of −3.7°C and a mean minimum temperature of −12.9°C. Normal precipitation for the 3-month period is 7.2 cm, with 80.5 cm of snow. During 2008/09, the DJF mean maximum and minimum temperatures were cooler than normal (by 1.4° and 1.9°C, respectively) and total precipitation was near normal (4.6 mm below average—a 6% reduction). Total snow amount, however, was 12.7 cm (16%) above average, with 2.54 cm (1 in.) snow totals observed on 11 days (normal is 10 days).
The study was conducted in a 113-km2 area centered on the Minneapolis central business district (CBD). The observations are focused within Minneapolis rather than across the entire metropolitan area because the objective was to observe snow cover’s effect on the local-scale variations in temperature within the city and because Winkler et al. (1981) showed that the highest urban temperatures are found near the Minneapolis CBD. The small geographic domain allowed for exclusion of any UHI effects from the nearby St. Paul CBD and also helped to minimize the likelihood of having significant differences in snowfall across the area (Doesken and Judson 1997).
Urban temperature sampling sites (Fig. 2) were identified through consideration of land use/land cover and distance from the Minneapolis CBD. Landsat-derived impermeable surface data for 2002 (obtained online from the University of Minnesota at http://www.land.umn.edu), field observations, and a land-use data layer (obtained online from MetroGIS at http://www.datafinder.org) were used to classify locations within the study area as high urbanization intensity (HI), low urbanization intensity (LI), or “no go” areas (e.g., areas of steep terrain, rivers and lakes, railroad yards, or highways). The HI areas had a large proportion (>50%) of impermeable surface (e.g., parking lots, industrial warehouse areas, and multifamily housing), and LI areas had a lower proportion (<50%) of impermeable surface (e.g., inner-city parks, residential areas, and golf courses). The Minneapolis CBD was used as the center of a series of six sequential 1-km concentric buffers. Within each of these 1-km rings, sample sites were randomly generated to include four sites at LI locations and four sites at HI locations, for a total of 48 possible sites across the study area. Of these possible sites, 24 locations were selected on the basis of category (LI or HI), accessibility, and an estimate of how well the site represented the dominant land-cover type of its immediate surroundings.
a. Data collection
Hourly air temperatures at the sample sites were recorded with high-capacity (8 kByte) iButton Thermochron dataloggers (range: from −40° to 85°C, ±0.5°C), a relatively inexpensive and low-profile instrument that minimized the likelihood of tampering, theft, or damage (see Oke 2006). The dataloggers were installed roughly 2 m above ground level on the nearest mountable surface (usually trees and utility poles), and away from canyon walls, identifiable heat sources, or other possible interference. All dataloggers were placed on the north-facing side of the surface so as to avoid the effects of direct sunlight (Fig. 3).
Guidelines proposed by Oke (2006) were used to collect metadata for each site,1 including current land use (commercial, residential, or industrial), cover types (e.g., lawns, concrete sidewalks, or asphalt parking lots), building types and estimated heights (number of stories), building materials (e.g., concrete, brick, or stucco), road materials (asphalt or concrete), traffic density (light, medium, or heavy), and an estimate of the height of trees in the immediate area. At each site, photographs were taken to the north, east, south, and west of the datalogger and of the Thermochron-mounted surface itself. Sky-view factor was estimated based on observations and sketches of local elevation, topography, and obstructions (in degrees above the horizon) to the north, east, south, and west of each datalogger. The Thermochrons were deployed beginning in early November of 2008. Sample sites were visited twice during the winter, and additional photographs were taken as needed to record the basic condition of snow at each site.
The urban (Thermochron) temperature data were compared with temperatures from the Automated Surface Observing System (ASOS) station at the Minneapolis–St. Paul International Airport (KMSP). The ASOS instrumentation is located in a flat, open, grassy area with no nearby obstructions, although parking lots and taxiways are near the site (P. Boulay 2010, personal communication; Fig. 4). The KMSP site was judged to be the most appropriate for representing “urban” snowfall (and other hourly weather) observations because it is only 12 km from the Minneapolis CBD. Snow measurements from more rural sites were not used because, from climatological data, total snowfall in the city typically is lower than for areas just outside the city (NCDC 2009a). The close proximity of KMSP to the urban Thermochron sites also reduces the likelihood of time lags between the onset of snow at the “rural” (airport) location versus the urban (Minneapolis) locations.
b. Quality control
For the 24 Thermochron sites, three HI sites in the first, second, and third rings and a single LI site from the sixth ring contained incomplete observations because of instrument damage or loss/theft over the observation period. Two other sites had suspect data: a sixth-ring LI site (likely from instrument loss/theft and the resulting, skewed, partial record), and an HI site from ring 3 that appeared to be influenced by reflected insolation from a nearby building. The two sites with suspect data (shown with an X in Fig. 2) were removed from the dataset, and the final dataset includes observations from 11 HI sites (including three with incomplete records) and 11 LI sites (including one with an incomplete record).
The magnitude of the UHI depends on urban morphology (e.g., building height, building density, the types of building materials, and the amount of impervious surface) as well as on prevailing environmental conditions (e.g., cloud cover, wind speed, and day/night). To isolate the specific influence of snow, a linear mixed-effect (LME) modeling scheme was implemented using the “lme4” package (Bates and Maechler 2009) as implemented in the R software package (version 2.11.1; http://www.R-project.org).
Much like multiple linear regression models, LME models are used to describe the relationship between a response variable and a set of predictor variables. The form of the multiple linear regression model is
where y is the vector of responses, is the matrix of observations of predictors, β is a vector of coefficients, and ε is a noise vector. In LME models, additional terms are used to separate “fixed effects” from “random effects” while the error vector is expanded to accommodate these two terms:
where and , is a fixed-effect model matrix, β is a fixed-effect vector of coefficients, is a random-effect model matrix, b is a vector of random-effect coefficients that is independent of ε, and is a variance–covariance matrix (Bates 2008). The random-effect predictors typically comprise a set of one or more categorical variables, so that b depends on whether the predictor is in state 1, state 2, …, state n (where n is the number of possible states or groups; e.g., urban/suburban/rural).
Mixed-effect modeling provides a number of advantages over traditional multiple regression: 1) the LME model is able to handle correlated observations (as are common in time series data); 2) it is able to consider the influence of a random effect, considering a set of categorical variables as a random sample of all possible values rather than a complete sample; and 3) it is able to consider the influence of variables measured at multiple scales (hierarchical data). For more details on this implementation of LME models, see Bates (2009).
In the LME model used here, ΔTu−r was treated as the response to six fixed-effect variables: hourly wind speed (from KMSP), distance from the Minneapolis CBD (depicted by the ring of the sample site), the urbanization intensity (HI or LI) of the sampling site, the depth of snow cover (from KMSP), the intensity of snowfall (from KMSP), and time of day (day/night) (Table 2). The location of the Thermochron within a given ring was treated as a random effect.
The hourly wind speed at KMSP (m s−1) was included as a continuous variable in the LME model. Hourly weather observations from KMSP (NCDC 2009b) were used to create a four-level categorical variable for falling snow: no snowfall (SF0), “light” (SF1) for observations of “continuous fall of snowflakes, slight at time of observation,” “moderate” (SF2) for observations of “continuous fall of snowflakes, moderate at time of observation,” and “heavy” (SF3) for observations of “continuous fall of snowflakes, heavy at time of observation” (NCDC 2005). Calculated values of daily sunrise and sunset were used to create a binary variable for day (from sunrise to sunset) and night (from sunset to sunrise) for each hourly value of ΔTu−r. Snow cover observations from KMSP were used to create a (daily) snow cover indicator, with “no snow cover” (SC0) present when the snow depth was below 5 cm, “light” snow cover present when the depth was between 5 and 10 cm inclusive (SC1), and “moderate” snow cover present when the depth was greater than 10 cm (SC2). These thresholds were chosen on the basis of results from Baker et al. (1991), who show that a 10-cm snow depth effectively masks the albedo of the underlying surface. A time series of snow depth at KMSP is shown in Fig. 5.
The final LME model was fit using maximum likelihood with model estimates chosen to optimize log-likelihood scores. The Akaike information criterion (AIC), Bayesian information criterion (BIC), and likelihood ratio tests were used to compare possible models, to determine the significance and influence of individual variables and the interactions between variables on the magnitude of ΔTu−r, and to judge model performance. Table 3 shows results of a likelihood ratio test that compares a simple model without snow variables (model0) with models that include variables for snow cover (SC), snowfall (SF), or both. Including the snow terms results in significantly improved model performance (significance level α = 0.01) over the no-snow model, and inclusion of all six predictor variables (Table 2) further improves the model fit. Although our modeling scheme includes wind speed as a continuous variable, the analysis presented here focuses only on the LME model results for no-wind conditions because that is when we expect the UHI to be best expressed (Oke 1988).
For December 2008–February 2009, the average Tu across the 19 urban Thermochron sites (excluding data from the three incomplete sites) was −8.9°C (standard deviation of 7.0°C, computed from hourly data). Over the same time period, the average Tr at KMSP was −9.8°C (standard deviation of 7.2°C, computed from hourly data). For 2008/09, then, the magnitude of the Minneapolis wintertime UHI (ΔTu−r) was about 1.0°C (standard deviation of 1.6°C, computed from hourly data) (Fig. 6). The 1.0°C heat island is similar to the January value of 1.1°C derived by Winkler et al. (1981). On average, the wintertime UHI is largest during the day, reaching its peak around 1200 LST (Fig. 7). We speculate that anthropogenic heating is the cause of the earlier increase in temperature at the urban sites as compared with the airport: urban temperatures begin to rise between 0700 and 0800 local time, which is when home, school, and workplace heating systems would be ramping up for the day. Toward the late afternoon ΔTu−r approaches zero, and it increases during the night, although remaining smaller than its daytime value. On average, the value of ΔTu−r (as represented by the LME model) is about 0.1°C larger for HI-only sites as compared with LI-only sites (not shown). This result suggests that, for these sites at least, urbanization intensity alone has only a small impact on the difference between urban temperatures and the temperature as measured at KMSP (airport).
In contrast to the influence of urbanization intensity, the presence of a snow cover significantly changes the magnitude of ΔTu−r. Figure 8 shows the LME-modeled ΔTu−r, with 95% confidence intervals, for an HI site during the daytime (no wind and no snowfall), with and without snow cover. The modeled effects are similar under both light (SC1) and moderate (SC2) snow cover, increasing ΔTu−r by approximately 1.0°C during the daytime. The effect of snow cover is nearly the same for LI sites as it is for HI sites, but with a widening of confidence intervals by approximately 0.1°C (Fig. 9). There is little evidence of a distance-decay pattern in the magnitude of daytime values of ΔTu−r (Figs. 8 and 9). The relatively low daytime temperatures in ring 2 may be reflective of the particular urban morphology (primarily urban residential) within this ring, although this may also be the result of the random placement of these sites closer to water and highway features than had occurred in the other rings.
At night, the magnitude of ΔTu−r at both HI (Fig. 10) and LI (Fig. 11) sites decreases with distance from the Minneapolis CBD (ring 1). The highest UHI values are in the core of the city, and they approach zero in the outer (fifth and sixth) rings for both LI and HI sites. This is a typical nighttime heat-island pattern and is likely due to the change in urban morphology from (primarily) retail/commercial development in ring 1, to mixed industrial/residential development in rings 2–5, to (primarily) residential development in ring 6 (Oke et al. 1991). In contrast to the daytime results, however, and regardless of urbanization intensity, light snow cover at night has almost no impact on ΔTu−r (no wind and no snowfall; Figs. 10a and 11a) whereas moderate snow cover increases ΔTu−r by approximately 0.5°C (Figs. 10b and 11b).
The LME model also indicates that the presence of falling snow affects the magnitude of ΔTu−r. Figure 12 shows the mean modeled ΔTu−r, with 95% confidence intervals, across the sampling rings for an HI site (daytime, no wind, and no snow cover) with no snowfall (SF0) as compared with light (SF1) and moderate snowfall (SF2). There is little consistent difference between no snowfall and light snowfall (recorded for 87% and 12% of the hours observed, respectively). When there is moderate snowfall (1% of hours observed), the magnitude of ΔTu−r decreases by about 2.0°C, although with wider confidence intervals. Heavy snowfall was observed less than 0.1% of the time so that the LME results are inconclusive.
The temperature observations and the LME model results show that the Minneapolis UHI for the winter of 2008/09 had the following features: under no-wind conditions (when the UHI is best expressed), 1) urbanization intensity (as defined here) had little impact on the magnitude of the UHI; 2) the strongest heat-island effect occurred during the day, rather than at night; 3) snow cover increased the magnitude of both the daytime and nighttime UHI; and 4) the presence of moderate-intensity falling snow (“snowfall”) decreased the magnitude of the UHI. What factors can explain these observations? Although concomitant energy balance measurements are not available, some possible causes can be drawn from the extensive UHI literature.
a. Little temperature difference between HI and LI sites
The HI and LI classifications were based on Landsat-derived impermeable surface data (obtained from the University of Minnesota at http://www.land.umn.edu) and on land use data (obtained from MetroGIS at http://www.datafinder.org). Five (23%) of the sites fell into urban climate zones (UCZs) 1 and 2 (“intensely developed urban”), four (18%) would be classified as UCZs 5 and 6 (“low density suburban”), and the remaining 13 (60%) had characteristics more closely resembling UCZs 3 and 4 (“highly developed urban”) (see Oke 2006), although these classes were not neatly aligned with the HI and LI categories. The lack of any notable temperature difference between HI and LI sites suggests that the HI/LI impervious surface-based classification that was the basis of the sampling design was unable to capture (by proxy) how urban characteristics affect the local climate, at least in Minneapolis during the winter. It also may be true that, at the spatial resolution of these data, small-scale (microscale) temperature variations between HI and LI sites were lost because of turbulent mixing within the roughness sublayer (Oke 2006).
b. Larger daytime versus nighttime UHI
One possible explanation for the strong daytime UHI comes from a consideration of the high albedo of snow, especially at the rural (airport) location as compared with the more urbanized locations of the Thermochrons, where snow is removed from roads, alleys, sidewalks, and parking lots and where the remaining snow becomes darkened by urban pollution. The low conductivity and thermal admittance of snow (Oke et al. 1991) likely are another contributing factor.
The average snow depth at the Minneapolis–St. Paul airport (KMSP) for the 2008/09 DJF period was 10 cm (NCDC 2009b). Snow has a higher albedo than nearly every other land surface and consequently the net shortwave radiation for a snow-covered surface (the airport site) would be lower than for a snow-free (or less snow-covered) surface (the urban Thermochron sites). The effect of snow cover on net shortwave radiation was demonstrated by Baker et al. (1992), who show that under snow-covered conditions (≥10-cm depth), net shortwave radiation is nearly balanced by the (daytime) net longwave radiation. When there is little or no snow cover, however, the net shortwave flux is 2–4 times the (daytime) net longwave flux. If the KMSP (rural) site has a deeper and/or more spatially homogeneous snow cover throughout the winter as compared with the urban sites—a plausible assumption, given the airport site characteristics (Fig. 4)—then shortwave heating at KMSP would be smaller than at the urban locations, leading to reduced daytime heating at the airport in comparison with the city and thus a larger daytime value of ΔTu−r (Fig. 7).
The decreasing intensity of the UHI toward sunset and the development of the nighttime UHI may be related to the low thermal conductivity of snow in comparison with many urban surfaces, such as asphalt, brick, glass, and concrete (Oke 1988). Late in the day (when there is little net shortwave flux) and at night (when there is none), the low conductivity of snow effectively insulates the ground (see Baker et al. 1992), reducing surface-to-atmosphere heat transfer at KMSP. At the urban sites, where snow cover is more uneven (exposing surfaces with higher conductivity and thermal admittance), heat is more easily transferred to the atmosphere, contributing to higher temperatures than are found at the airport. Anthropogenic heat from closely spaced urban structures (as compared with KMSP) likely also contributes to the nighttime UHI.
c. Differential impact of snow depth on the daytime and nighttime UHI
In addition to helping to explain the average diurnal pattern of ΔTu−r, the findings of Baker et al. (1992) also help to explain the differential effect of increasing snow depth on the magnitude of the UHI. During the day, ΔTu−r increases by about 1.0°C in the presence of both light and moderate (≥5 cm) snow depths, as compared with the “no snow” case (<5-cm depth). The daytime impact of snow cover on the UHI appears largely to be due to snow’s albedo effect (described above), which would be expected to be nearly unchanged as snow depth increases beyond 5 cm. This finding is similar to the results of Baker et al. (1991), who show that for bare ground and sod-covered surfaces (similar to the airport site), snow depths of 5 and 7.5 cm (respectively) were sufficient to mask these underlying surfaces. In contrast, light snow cover has almost no effect on the nighttime value of ΔTu−r, whereas moderate snow cover is associated with an increase in ΔTu−r of about 0.5°C. Snow is an effective insulator, but these results suggest that the full effects of snow cover may only emerge once snow cover reaches some minimum depth (in this case, about 10 cm). This result is similar to that of Grundstein et al. (2005), who also found that air temperature had a nonlinear response to snow depth. Ge and Gong (2010) showed that the insulating effects of snow were comparable in magnitude to snow’s albedo effect; if that is the case, these results may indicate that, for Minneapolis, about one-half of the 1.0°C daytime increase in ΔTu−r may be due to the insulating properties of snow—in particular, for the winter of 2008/09 during which the average snow depth was about 10 cm (NCDC 2009b).
d. UHI effects of falling snow
The LME model showed that light snowfall (identified in the airport hourly weather records as “continuous fall of snowflakes, slight at time of observation”) had little differential effect on urban temperatures as compared with airport temperatures but that moderate snowfall (“continuous fall of snowflakes, moderate at time of observation”) contributed to a decrease in ΔTu−r of approximately 2.0°C. As the atmospheric “snow load” increases, solar radiation passing through the atmosphere is more likely to be scattered or reflected upward, reducing the amount reaching the ground. The resulting decrease in incident solar radiation also would decrease the amount of radiation absorption and/or trapping within the urban canopy (as compared with the airport site). As recorded in the airport weather data, when moderate snowfall occurred, the cloud base was low (average of 200 m) and skies were overcast or obscured. It is possible that a portion of the effect associated with moderate snowfall may be due to increased cloud cover, which is known to reduce UHI intensity (Oke 1988).
This study investigated the effects of snow on the wintertime urban heat island in Minneapolis. Hourly air temperature measurements were collected from December 2008 to February 2009 at 11 “high urbanization intensity” and 11 “low urbanization intensity” sites within a 6-km radius from the Minneapolis CBD. Hourly temperature, weather, and snow observations at the Minneapolis–St. Paul International Airport were used to represent rural (more accurate would be to say less urbanized) conditions.
The observations for winter 2008/09 showed that the magnitude of the Minneapolis UHI averaged about 1.0°C, similar to the findings of other Minneapolis UHI studies. Contrary to expectations, the UHI reached its peak near midday rather than at night. We speculate that this was a consequence of the high albedo of snow. The mean winter snow depth at KMSP was 10 cm, and snow cover at the airport site 1) is spatially more homogeneous than within the city (where snow is swiftly removed from roads, sidewalks, and parking lots to expose low-albedo surfaces) and 2) has a higher albedo than snow within the city, where snow quickly becomes darkened by urban pollution. Both factors would contribute to enhanced heating within the city—in particular, during the day—as compared with the airport. There was little difference in temperature between HI and LI sites (as defined in this paper), perhaps because the impervious surface-based urbanization-intensity classification is inappropriate for the cold, dry (low evaporation) Minneapolis winter.
On the basis of the results of a linear mixed-effects model, the presence of snow cover of ≥5-cm depth contributed to an increase in ΔTu−r by 1.0°C during the day (attributed to the high snow albedo at the airport as compared with urban locations). At night, snow depths ≥10 cm increased ΔTu−r by 0.5°C (attributed to snow’s low thermal conductivity). The LME model also showed that moderate-intensity snowfall reduced the magnitude of the daytime UHI by as much as 2.0°C, possibly as a result of the reduction in ground-level incident solar radiation and the resulting decrease in radiation absorption/trapping within the urban canopy (as compared with the airport), or as a result of increased cloud cover. Moderate snowfall was observed on less than 1% of the 2160 h of record (1 December 2008–28 February 2009), however, so the model estimates of the effects of snowfall are somewhat speculative.
The characteristics of the Minneapolis wintertime UHI discussed here are derived only from temperature measurements. It is difficult to establish the relative contributions to observed temperature patterns of short- and longwave radiative forcing; sensible, latent, and ground (conducted) heat fluxes; and anthropogenic heating because of the lack of concomitant energy balance measurements. Plausible explanations can be drawn from the extant urban climatological literature, but more definitive climatological analyses will require simultaneous information about the urban energy balance, at least for intensive observing periods (see Lemonsu et al. 2008), sustained study over longer time periods (more than one winter), and investigation of a wide range of cold-city environments.
Thanks are given to Peter Boulay and James Zandlo of the Minnesota State Climatology Office for help with instrumentation, to Dr. Aaron Rendahl of the University of Minnesota Statistical Consulting Service for help with linear mixed-effect models, and to the anonymous reviewers for helpful comments on the manuscript. This study was supported by the College of Liberal Arts at the University of Minnesota and by the University of Minnesota Undergraduate Research Opportunities Program.
Metadata are available by making a request to the corresponding author.