Introduction

Extreme heat events (EHE), characterized by stagnant warm air masses and successive nights with elevated minimum temperatures (Luber and McGeehin 2008), can have catastrophic public health consequences (Huang et al. 2010; Klinenberg 2002; Robine et al. 2008; Vandentorren et al. 2004). The frequency, intensity and duration of such EHEs—many of which may be exacerbated by high levels of humidity and air pollution—are predicted to intensify across North America and Europe throughout the twenty-first century (Beniston 2004; Koffi and Koffi 2008; Meehl and Tabaldi 2004). For inhabitants of cities, the effects of a changing climate are expected to be magnified due to the urban heat island (UHI) phenomenon, which is the propensity for urban centres to experience comparatively warmer temperatures than adjacent peri-urban and rural locations (Oke and Maxwell 1975; Voogt and Oke 2003).

A recent focus of Toronto Public Health (municipal level health promotion organization) has been to identify neighbourhoods that are vulnerable to the adverse health effects of extreme heat with the purpose of targeting response measures during EHEs (Rinner et al. 2010). As part of this initiative, a spatially explicit potential heat vulnerability index incorporating exposure and sensitivity risk factors was created at the census tract level for the general population as well as for vulnerable subpopulations (e.g. seniors). Similar studies identifying urban areas with high heat vulnerability have been conducted for Montreal (Vescovi et al. 2005) and for the continental USA (Reid et al. 2009); however, in each of these studies, mapping was conducted at a relatively coarse spatial resolution (census subdivision and metropolitan statistical area, respectively). Satellite imagery measuring land surface temperature (LST) has been the primary data source used to characterize outdoor heat exposure in heat vulnerability (Johnson and Wilson 2009; Rinner et al. 2010) and UHI mapping studies (Aniello et al. 1995; Buyantuyev and Wu 2010; Liang and Weng 2008). Thermal infrared data from the Landsat satellites (band 6), acquired at a ground resolution of 60 or 120 m and at a spectral resolution of 10.40–12.50 μm, appear most frequently in the literature (Johnson and Wilson 2009; Maloley 2009; Yuan and Bauer 2007). In this study, we emphasize the importance of differentiating between air temperature and LST when evaluating potential exposure to extreme heat. Air temperature is measured in the canopy layer (a layer of air aboveground equivalent to the height of proximate buildings and trees) using in situ data from meteorological station networks, automobile transects, or specialized sensor platforms (Voogt and Oke 2003), whilst LST is usually measured using airborne or satellite remote sensing (Streutker 2002). Although air and land surface temperatures can display similar spatial and temporal patterns, microscale site characteristics have a greater influence on surface temperature than on air temperature (Arnfield 2003; Eliasson 1996; Oke 2006). The relationship between air and surface temperature is also heavily influenced by weather conditions; in cloudy, windy weather, the effect of surface temperature on air temperature is diminished (Gartland 2008). Furthermore, air temperature as influenced by humidity and wind determines the degree of comfort for the human body and directly influences morbidity and mortality during EHEs (Smoyer-Tomic and Rainham 2001; Havenith 2005).

Whilst thermal remote sensing with data acquired from a satellite platform is useful for detecting spatial variation in LST, there are fundamental limitations to the use of these data in investigating the heat stress on humans. A spatially continuous dataset of LST at medium spatial resolution (60–120 m) is only a proxy for heat exposure by city inhabitants (Wilhelmi et al. 2004); it is the understanding of the surface energy balance (Matzarakis et al. 2007)—especially the spatial and temporal variation in air temperature—that is more important in the context of mitigating and adapting to the potential health effects of EHEs. Thus, LST measurements should be considered surrogate data when studying the canopy layer UHI. A study conducted by Maloley (2009) using in situ meteorological data from 40 stations revealed that air and surface temperature in the Greater Toronto Area (GTA) were not well correlated due to the effects of humidity and convective transfer of heat by wind. Maloley (2009) further determined that the surface temperature of vegetated land cover exhibited a stronger positive correlation with air temperature and that urban land cover correlation with air temperature was highly variable. Eliasson (1996) also argues that intra-urban surface and air temperature patterns are notably different; therefore, LST data should be used with caution to study air temperature within cities.

A second limitation of using thermal satellite imagery to study the potential health impacts of heat is low temporal resolution (16-day repeat coverage cycle for Landsat satellites; Wilhelmi et al. 2004). Furthermore, researchers using Landsat satellite imagery have been limited to data captured earlier in the day (local time), a function of this satellite’s orbiting overpass time (Maloley 2009; Yuan and Bauer 2007). As a result, data acquisition has occurred neither at a time when temperature may be at a maximum nor during the overnight hours when temperatures are likely to be lowest and therefore a good indicator of whether relief from heat is present. Furthermore, satellite thermal imagery is quasi-static—it lacks repeat coverage at a temporal resolution (e.g. hourly) that would permit meaningful investigation of temperature change and the consequent implications of heat exposure duration.

Directional temperature variation, as influenced by the geometry of many urban structures (known as urban surface anisotropy), can also contribute to erroneous surface temperature measurements when imaging with a thermal remote sensor (Voogt and Oke 1998). In heavily built-up locations with complex three-dimensional structures, morning acquisition of thermal imagery may detect cooler-shaded rooftops and ‘canyon’ bottoms (i.e. road surfaces) rather than sun-warmed building walls (Maloley 2009).

Finally, surface emissivity information is required to accurately convert thermal radiance (received by the satellite sensor) to LST (Campbell 2006). Axelsson and Lunden (1988) discuss the difficulty in determining the surface temperature of constructed surfaces (e.g. buildings, roads) in urban areas due to inadequate knowledge of the specific emissivity of highly heterogeneous targets (characteristic of a city landscape). Although many of these surfaces tend to exhibit lower emissivities than do natural surfaces (Oke 1978; Rees 1990), common emissivity values have been assigned to entire areas of heterogeneous land cover.

As an alternative to thermal imagery, Mersereau and Penney (2008) identify diurnal air temperature (acquired hourly) as useful information for informing adaptation measures during EHEs. The existence of a growing number of privately operated meteorological stations within metropolitan areas presents a new opportunity for using canopy layer data to study heat exposure potential in cities. One such study used citizen-collected data from digital automobile and stationary thermometers in and around Manchester (UK) to map the city’s atmospheric heat island (Knight et al. 2010).

Meteorological data have the distinct advantage over remotely sensed thermal imagery of allowing measurement at regular and meaningful time intervals (Streutker 2002). In situ data collected frequently (e.g. hourly) enable canopy layer UHI mapping at times of the day when relief from heat is critical for vulnerable citizens (such as during overnight hours). For instance, when using meteorological station data from airports, Loughnan et al. (2010) found elevated mortality levels among the elderly when a hot day was followed by a hot night. Data acquired from meteorological stations can also provide relative humidity and wind speed measurements in addition to temperature, both of which are important to understanding heat stress as a function of energy balance in the urban canopy layer (Oke 2006).

Atlanta’s urban canopy layer UHI was investigated by Zhou and Shepherd (2010) during heat waves using minimum and maximum daily temperature averages, but their analysis used data from only three meteorological stations and did not consider humidity. Temporal variation in temperature differences between general land cover classes (urban, suburban and rural) was explored by Basara et al. (2010) during a prolonged heat wave in Oklahoma using data from 46 meteorological stations. And though they considered both heat and humidity, these authors did not pursue the creation of apparent temperature surfaces (the combination of temperature and humidity into an index reflective of perceived temperature). Reference in the literature to the mapping of changes in the canopy layer UHI over a diurnal cycle is usually to low spatial resolution meteorological networks (Figuerola and Mazzeo 1998; Oke 2006), and to our knowledge, no one has mapped spatiotemporal variation in apparent temperature during an EHE.

Recognizing the limitations of previous heat vulnerability and UHI research that has used LST as a surrogate for air temperature, we have designed this study using temperature and humidity data acquired from 65 meteorological stations to create a time series of 24 hourly prediction maps reflective of apparent temperature for a ‘typical’ extreme heat alert day. We use humidex, an index developed by Canadian meteorologists to reflect the ‘actual feel’ of hot and humid weather, to estimate apparent temperature (Environment Canada 2010; Masterton and Richardson 1979); it combines air temperature and relative humidity into a single value (Fig. 1). We then combine these apparent temperature maps using the concept of degree hours to produce a humidex degree hour (HDH) composite surface describing locations in the GTA that experience prolonged exposure to elevated temperatures during a typical EHE. Our primary motivation in conducting this research was to develop and propose a metric for mapping exposure to heat using meteorological data that would assist public health decision makers in assessing locations of vulnerability within their communities. Furthermore, we hope that this study will stimulate conversation about the value of establishing urban networks of meteorological stations for long-term heat vulnerability monitoring.

Fig. 1
figure 1

Apparent temperature. Apparent temperature (humidex) at air temperature values with a corresponding range in relative humidity. Apparent temperature thresholds associated with various levels of bodily discomfort follow the definitions provided by Environment Canada

Full size image

Materials and methods

Study area

The study area for this project is roughly delineated by the GTA (Fig. 2). Located in southern Ontario, the GTA is Canada’s most populated (5.5 million) and fastest growing metropolitan region (Ontario Ministry of Finance 2010). With 2.6 million people in 2010 (City of Toronto 2010), the City of Toronto is the most densely populated region within the GTA. The GTA’s four suburban regions of Halton, Peel, York and Durham are expected to grow significantly faster than the provincial growth rate between 2009 and 2036 (collectively adding over 2.4 million people); Peel and Halton are projected to increase their population by 84.5% and 90.2% over 2009 levels, respectively, by 2030 (Ontario Ministry of Finance 2010). Bounding the southern edge of many GTA regions is Lake Ontario, which has been shown to have an important moderating effect on temperature in urban locations within several kilometres from its shoreline (Munn 1969).

Fig. 2
figure 2

Study site. Locations of 76 meteorological stations in the Greater Toronto Area (GTA). Study area boundary is delineated with a dashed line; solid lines within regions are municipal boundaries. Meteorological data from 65 of the 76 stations were used to interpolate apparent temperature and to develop a humidex degree hours (HDH) index

Full size image

Data

Archival meteorological data were obtained from 75 monitoring sites across the study area for 6 days in 2008 when extreme heat alerts had been issued for the City of Toronto by the Medical Officer of Health (June 9, July 7, July 8, July 16–18). The City of Toronto uses a synoptic heat health alert system where the trigger of an extreme heat alert is based upon comparing forecast meteorological data with past analyses of regional air masses that have historically led to excess heat-related mortality (≥90% probability) in Toronto (Toronto Public Health 2009). Meteorological data for this study were obtained from three primary sources: Environment Canada, Natural Resources Canada (NRCan) and Weather Underground. Research groups at Ryerson University and the University of Toronto provided data from several additional meteorological stations. Weather Underground (weatherunderground.com) is a dynamic web service that allows citizens to upload real-time data from private meteorological stations. It provided data for 28 of the meteorological stations (36%) used in our study.

Selection of meteorological stations was exhaustive; data were obtained from all available stations meeting location and recording interval criteria. Air temperature, relative humidity and wind speed were collected at each site, with the exception of NRCan stations that recorded only air temperature. The distribution of meteorological stations was not spatially uniform across the study area and tended to be concentrated in locations with greater population density. We classified sample locations using local climate zones (LCZ), defined by Stewart (2009) and Stewart and Oke (2009), which are based on land characteristics within a 200-m radius of the meteorological station. Overall, 51 stations were classified as ‘built’, whilst 22 and 2 were classified as ‘natural’ and ‘mixed’, respectively (Table 1). The most common LCZ was ‘open-set low-rise’, which is typical of residential areas with low traffic flow.

Table1 Meteorological stations
Full size table

All measurements were taken within the urban canopy layer, although individual monitoring site specifications had some modest variation. In 17 cases, NRCan monitoring sites were located on rooftops (Maloley 2009); in all other circumstances, measurements were made at approximately 1.5 m above the ground surface. Because of airflow restrictions and radiation exchange by buildings, using data acquired from roof-mounted meteorological stations must be approached with caution. However, Oke (2006) argues that, for practical reasons, it is sometimes necessary to be flexible and use non-standard surfaces at non-standard heights for meteorological station placement. To investigate this potential for variability in recorded temperature, we selected all meteorological stations positioned in locations with the LCZ classification of ‘built’ and compared roof-mounted to non-roof-mounted measurements at four time intervals that encompassed a 24-h period.

Apparent temperature was calculated for each of our monitoring sites as follows:

$$ {\text{Apparent}}\,{\text{Temperature}} = \left( {{\text{air}}\,{\text{temperature}}} \right) + h $$
(1)

Air temperature (°C) was adjusted to reflect apparent temperature in Eq. 1 based on actual temperature and the corresponding influence of relative humidity, captured as h:

$$ h = 0.{5555} \times \left( {e - {1}0.0} \right) $$
(2)

where h is an increment to temperature that is added to represent the effects of humidity on thermal sensation. Equation 2 incorporates vapour pressure in the term e and was calculated as follows:

$$ e = {6}.{112} \times { \exp }\left( {{5417}.{753}0 \times \left( {\left( {{1}/{273}.{6}} \right)-\left( {{1}/{T_{{\rm d} }}} \right)} \right)} \right) $$
(3)

where T d is the dewpoint in kelvins. The Magnus–Tetens formula was used to approximate dewpoint from air temperature and relative humidity (Barenbrug 1974). Units are not ascribed to apparent temperature; instead, we treat it as an index and discuss it in the context perceived (‘feels like’) temperature.

Because relative humidity (RH) values were not available from the NRCan meteorological stations (30 in total), we used hourly data from the remaining 46 stations to interpolate 144 RH prediction surfaces (one each hour during the six heat alert days). Interpolation of RH used the approach of ordinary kriging. Predicted RH values were extracted at each NRCan station and used with corresponding air temperatures to calculate apparent temperature following Eq. 1.

Twenty-four composite datasets of hourly apparent temperature were created by averaging measurements recorded on the six extreme heat alert days (i.e. six apparent temperature values for each hour, at each meteorological station, averaged to produce one value per hour at each station). This approach provided us with hourly apparent temperature values for a ‘typical’ heat alert day at each meteorological station.

Mapping apparent temperature

We used Geostatistical Analyst (ESRI’s ArcGIS 9.3 extension) to conduct ordinary kriging (OK) with apparent temperature data from 65 of the 76 meteorological stations (data from stations in northern York Region and beyond were not included for OK because of large separation distances from the majority of sample locations in the southern GTA). Exploratory data analyses revealed that each of the 24 composite apparent temperature datasets exhibited a negatively skewed distribution. We subsequently applied a square root transformation to each dataset to adjust its distribution to approximate normal (Bailey and Gatrell 1995). OK is the best predictor among all unbiased predictors when the random function fits a normal distribution (Johnston et al. 2001; Weber and Englund 1994; Yasrebi et al. 2009).

The presence of anisotropy (greater autocorrelation in one direction when compared with another) was investigated in our apparent temperature dataset using an interactive semi-variogram modelling tool in Geostatistical Analyst (Johnston et al. 2001). Using this method, we found evidence of an inverted U-shaped trend in our data (i.e. higher apparent temperature values clustering in the central portion of our study area). To minimize anisotropy, we elected to use a global trend removal with a second-order polynomial (Bailey and Gatrell 1995; Waller and Gotway 2004). All apparent temperature datasets were then modelled using a spherical semi-variogram specifying a maximum neighbourhood search radius of 5 and minimum of 1. For each of the 24 semi-variograms, we elected to have Geostatistical Analyst optimize the parameters for major and minor axes (Johnston et al. 2001).

The root mean squared error and root mean squared standardized error (RMSSE) were the primary cross-validation statistics used to evaluate the accuracy of the interpolated surfaces (Webster and Oliver 2007). In addition to examining the cross-validation statistics, we used a priori knowledge of the study area (e.g. known areas of impervious surface where extended periods of high temperatures would be expected) to assist with a final suitability inspection of our predicted apparent temperature surfaces.

In an attempt to improve upon our OK results, we obtained LST data from NRCan (derived from 2008 June and September Landsat 5 Thematic Mapper thermal imagery at a 120-m spatial resolution; Maloley 2009) for use as a secondary (ancillary) data source with the spatial interpolation method of ordinary co-kriging (OCK; Bailey and Gatrell 1995). We then combined these data using a raster calculator in a geographic information system to create a single composite thermal image showing average LST. The spatial extent of this thermal imagery was not perfectly congruent with our study area; a section of southeast York Region had no data.

One thousand sample points from the composite thermal satellite image were sampled for the purpose of OCK. A stratified sampling regime was designed such that the proportion of sample points was equal to the proportion of land area in each regional municipality across the study area. Within each municipality, the sample locations were further stratified proportionate to the land cover area determined from an analysis of a 2004 land cover map (DMTI Spatial Inc.). We ensured that our secondary dataset (thermal surface temperature) was heterotopic (i.e. none of the sample points coincide, or were spatially co-located, with the primary meteorological data; Bailey and Gatrell 1995; Goovaerts 2000).

Humidex degree hours

A cumulative surface depicting prolonged exposure to heat was generated by calculating HDH from hourly apparent temperature prediction maps (24 in total). HDH is a measure of how much, and for how long, apparent temperature values are higher than a predefined threshold. We developed the concept of HDH based on cooling degree days (CDD), where CDD is used to compute the energy demand required to cool buildings (Sailor and Pavlova 2003). For our HDH index, an apparent temperature of 30°C was established as a base value. This is the apparent temperature value at which Environment Canada (2010) indicates that ‘some discomfort’ will be felt by humans and is the low value reported by Smoyer-Tomic and Rainham (2001) as leading to excess mortality in a study discussing heat-related mortality in Ontario. Air temperature values as low as 30°C have also been linked to excess mortality in Australia (Loughnan et al. 2010) and Spain (Tobias et al. 2010).

To generate HDH for our study area, we first rasterized our data (120-m pixel resolution) and then reclassified each of our 24 apparent temperature prediction maps by assigning each raster cell a new value derived from its relation to the pre-established base value of 30. For example, an apparent temperature of 30 was assigned a value of 1, an apparent temperature of 31 was given a value of 2, and so forth. Apparent temperature deviations from 30 in a direction towards cooler temperatures were not considered. The 24 newly classified maps were then summed to produce a cumulative HDH surface map describing spatiotemporal heat exposure for a typical extreme heat alert day.

We propose two HDH thresholds—48 (caution) and 72 (danger)—that we believe reflect a combination of heat intensity and duration of exposure that pose a risk to public health. Our HDH caution threshold corresponds to apparent temperatures observed by Smoyer-Tomic and Rainham (2001) that resulted in increased heat-related mortality in Ontario, Canada, between 1980 and 1996. For example, a HDH caution threshold corresponds to 24 consecutive hours at an apparent temperature index value of 31. Similarly, our danger threshold (12 sequential hours at an apparent temperature index value of 35) is based on the temperature range (35–39.9°C) reported by Smoyer-Tomic and Rainham (2001) that produced a statistically significant increase in mortality among individuals >65 years of age.

It is difficult to specify an exact temporal duration at which point prolonged heat exposure is considered ‘dangerous’. This is because the adverse health effects of hot weather on an individual are dependent on several personal factors that include age, body composition, fitness and clothing (Havenith 2005). However, several studies have discussed elevated mortality rates following 24 h of exposure to hot temperatures (Loughnan et al. 2010; Oeschsli and Buechley 1970; Tobias et al. 2010). It is expected that precursors to heat-related mortality, such as heat exhaustion or heat stroke, would begin to develop after a much shorter duration of exposure to elevated temperatures (e.g. several hours).

Results

Over the six extreme heat alert days investigated in this study, the average hourly apparent temperature values ranged from 28.1°C to 41.1°C. When we combined our 6 days of apparent temperature to produce an hourly composite (reflective of 24 h in a typical day of extreme heat), we found a significant difference in average apparent temperature at built sites compared with natural sites between the hours of 3 pm and 7 am (p < 0.005, Fig. 3). The coolest values were recorded during the early morning hours (3 to 5 am) in the northern municipalities, whilst the hottest values were measured mid-afternoon (1 to 4 pm) in the City of Toronto, Mississauga and Brampton. The maximum apparent temperature recorded in locations classified as built suburban showed minimal differences in diurnal warming and cooling patterns when compared with urban areas, except that a faster increase in temperature was observed in the late morning (9 am to 12 pm, i.e. built suburban locations appeared to reach maximum daily apparent temperatures more rapidly). Built suburban sites were significantly warmer than natural rural sites, except between the hours of 10 and 11 am (p < 0.05), peaking at 3 pm (3.3 difference). When we consider the entire diurnal temperature cycle, the average apparent temperature values recorded at built urban and built suburban sites were classified in Environment Canada’s ‘some discomfort’ range (≥30) for 13 and 10 h, respectively.

Fig. 3
figure 3

Hourly apparent temperature stratified by land use. Apparent temperatures for a ‘typical’ extreme heat alert day in the Greater Toronto Area (GTA). Data are stratified geographically by land use (urban, suburban and rural) and plotted hourly over the diurnal temperature cycle. Error bars represent 95% confidence intervals about the mean apparent temperature

Full size image

Our difference in the mean comparisons of roof-mounted to non-roof-mounted meteorological stations in LCZs classified as built at four time intervals (morning, afternoon, evening, night) revealed no significant temperature difference (α = 0.05). Therefore, we believe it is reasonable to conclude that, in this study, temperature observations from roof-mounted and non-roof-mounted stations are comparable and therefore appropriate for inclusion in our interpolation of apparent temperature maps.

All but four hours in our 24 h of apparent temperature data showed the presence of a global trend of higher values in the central southern portion of the study area. For hours during which a global spatial trend was absent, we did not elect to undertake any trend removal prior to using OK for surface interpolation; in all other circumstances, a second-order global trend removal was applied to each dataset in advance of OK. In all cases, empirical semi-variograms generated on an hourly basis were fit with a spherical model. Table 2 presents a summary of cross-validation statistics generated using output from each hour of apparent temperature in accordance with the selected model and level of global trend removal performed prior to undertaking surface interpolation with OK. Cross-validation results were judged to be robust, falling within the parameters generally established in the literature (Diodato 2005; Forsythe et al. 2004). In general, our models show a slight tendency towards overestimation of apparent temperature (RMSSE < 1).

Table 2 Cross-validation of statistical models
Full size table

OCK did not improve the accuracy of the interpolated apparent temperature surfaces in our project. In fact, the OCK results were virtually identical to our OK findings. This finding is the result of a weak correlation between the primary variable (apparent temperature) and secondary variable (surface temperature), which on average was 0.24 (SD = 0.04)—much less than the minimum correlation value of 0.40 suggested for OCK (Asli and Marcotte 1995). We therefore elected to use OK for all final point interpolations to create apparent temperature maps.

To provide a further check of the accuracy of our apparent temperature prediction surfaces, we generated standard error maps for each hourly OK output. This step allowed us to evaluate prediction errors spatially across the study area. We present a summary in Table 3 of standard error, stratified by municipality, for the hour with the worst prediction surface results (3 pm). In general, we find the standard error values to be acceptable, but wish to note that confidence boundaries could be tightened with greater spatial density of meteorological stations.

Table 3 Modelling error statistics
Full size table

The HDH map was calculated by weighting each hourly apparent temperature surface based on the magnitude of apparent temperature index values ≥30 and adding the resulting time series of 24 raster surfaces together (Fig. 4). HDH values averaged 45 ± 0.07 (95% CI) and ranged from 0 to 99, varying substantially across the study area. In general, HDH was lower in the northern and eastern study regions. Conversely, the south central corridor of Toronto, including the heavily developed downtown core, is expected to have the highest HDH values during a typical extreme heat alert day. Areas that meet or exceed HDH thresholds of 48 (caution) and 72 (danger) are also identified, capturing locations where potential exposure to heat is being experienced at high levels for an extended period of time. South central and northwest Toronto, as well as much of Mississauga, exist within the ‘danger’ zone, where apparent temperature was ≥35 for 12 or more hours.

Fig. 4
figure 4

Humidex degree hour (HDH) map. HDH for a ‘typical’ extreme heat alert day. Isolines identified as ‘Caution’ delineate areas where a combination of exposure time and apparent temperature (≥30) lead to an HDH of 48 (e.g. 24 consecutive hours at an apparent temperature index value of 31). Similarly, isolines described as ‘Danger’ encapsulate areas where HDH ≥ 72 (e.g. 12 consecutive hours at an apparent temperature index value of 35)

Full size image

Our findings identify the City of Toronto municipalities (pre-amalgamation) of Etobicoke, East York, York and North York as all having average HDH values exceeding 48. In a future EHE, given the diurnal cycle of warming and cooling air, these hours of high apparent temperature will most likely occur in succession. Our hourly apparent temperature surfaces confirm that elevated air temperatures in portions of these municipalities extend for consecutive hours well into the night. Several municipalities located in the immediate suburbs of the City of Toronto also have HDH values >48 during a typical EHE. These include Vaughan and Markham (York Region), and Mississauga and Brampton (Peel Region).

When we compare the four GTA regions, we find significantly different mean HDH values (F 3,249110 = 10,002, p < 0.001). Using a Scheffe post hoc comparison, we determined that the City of Toronto has the highest overall mean HDH value (55.01, SD = 10.7) and that Durham Region has the lowest (35.1, SD = 3.8). Peel and York Regions, whilst quite different in their variation of HDH values, were very similar in mean HDH: 42.6 (SD = 12.4) and 45.4 (SD = 24.1), respectively. ANOVA and Scheffe post hoc comparisons were also used to determine whether significant differences in HDH existed between municipalities within each region. Significant differences in mean HDH values exist among the municipalities in Peel Region (F 2,88172 = 93,803, p < 0.001), York Region (F 6,81244 = 25,950, p < 0.001), Durham Region (F 2,35831 = 3371, p < 0.001) and the City of Toronto (F 5,43854 = 4731, p < 0.001; Fig. 5).

Fig. 5
figure 5

Humidex degree hour (HDH) variability by municipality. Variation in HDH for each municipality in the study area. The municipalities of Etobicoke and Toronto (pre City of Toronto amalgamation) have median HDH values that exceed 65; Mississauga is close with a median value of 60. Municipalities in Durham Region, the northern portion of York Region and Caledon all have median HDH values <40. A wide range of HDH values exit for all municipalities in the City of Toronto

Full size image

Discussion

Our view is that obtaining diurnal apparent temperature measurements is essential for developing a spatially explicit heat vulnerability index. The HDH metric we develop in this project integrates apparent temperature intensity and duration into one spatially explicit value, thus allowing for the identification of geographic areas where citizens may experience the adverse health effects of prolonged heat exposure. Further to this, we have proposed the development of HDH thresholds (i.e. 48 and 72) that could be used by public health decision makers when evaluating potential heat vulnerability during EHEs. Time series epidemiological studies that link HDH to daily mortality are needed in order to properly refine different levels of HDH thresholds for vulnerable subpopulations (e.g. seniors, children).

When considering our methodological approach for apparent temperature prediction mapping, a low correlation between the hourly apparent temperature data (primary variable) and LST (secondary variable, obtained from thermal imagery) negated the potential value of OCK. A similar conclusion was reached by Eliasson (1996) and Maloley (2009) who found that air and surface temperature in urban areas show only weak associations over the diurnal temperature cycle. We therefore advocate that meteorologists and public health organizations explore ways of encouraging partnerships that would support the expansion of an urban network of meteorological stations and for public health agencies to make use of these data for heat vulnerability planning. The usefulness of the HDH metric is transferable to other densely settled locations where climate is characterized by the temporal congruence of high relative humidity and high actual temperature.

A recent discussion among public health professionals at the 2010 Urban Heat Island Summit held in Toronto focused on the importance of including wind velocity as a third variable when forecasting the spatiotemporal variation in apparent temperature associated with an EHE (U. Bickis, personal communication; Forkes 2010). This is especially important in built-up areas where the presence of urban canyons affects wind velocities at a microscale level. Wind velocity was measured at only 33 of our meteorological stations during the 6 days of extreme heat alert we targeted. This limited number of sample locations precluded the creation of a robust kriging surface (Webster and Oliver 1992). Furthermore, we had limited knowledge of meteorological station locations concerning factors that may influence local air movement. However, we wish to emphasize that in addition to air temperature and humidity data, detailed wind speed and wind direction maps for the urban canopy layer are needed to arrive at a complete understanding of spatiotemporal variation in air temperature during EHEs. A constant breeze in locations with higher HDH values is apt to provide natural relief from prolonged heat and humidity.

If public health organizations wish to pursue the use of in situ climate data to monitor canopy layer heat exposure, the spatial density of meteorological sites must be increased to improve the robustness of interpolated surfaces generated through geostatistical modelling techniques. Whilst in situ climate information has a distinct advantage over remotely sensed thermal imagery, the current challenge with meteorological data is that measurement sites are rarely arranged in a geographically uniform manner (Hicks et al. 2010; Knight et al. 2010). The accuracy of our HDH surface is influenced by the spatial arrangement of meteorological sites in the study area; large distances between measurement sites increase prediction errors.

To create effective HDH prediction surfaces, meteorological stations must be roughly uniform in their distribution across a given jurisdiction and exist at a relatively high spatial resolution. Based on an inspection of the range values associated with each of our 24 hourly semi-variograms, we suggest an ideal station density of between 1 and 2 km2 for areas with highly heterogeneous land cover (e.g. downtown Toronto, where there is considerable variation in land cover types over short distances). In areas with greater homogeneity of land cover, station density could be decreased to between 4 and 5 km2. This recommendation for station placement density is similar to that proposed by Oke (2006) who suggests “one to several kilometres” as optimal positioning for meteorological stations in urban areas. Such a network could be established in Toronto (641 km2) by including existing meteorological stations, such as those operated by government organizations (e.g. Environment Canada) and by private citizens. Selection of measurement sites should represent different types of land use (e.g. residential, parkland, commercial) as associated land cover has an important influence on both air and surface temperature (Arnfield 2003; Liu and Weng 2008; Unger 1996).

Whilst climate data acquired from the public realm can raise accuracy concerns, meteorological data aggregators such as Weather Underground require instrumentation standards and provide detailed instructions to the establishment and maintenance of meteorological stations. Municipal partnerships with educational institutions, such as primary and secondary schools, also provide potential for situating and operating meteorological stations. Data arising from these stations would supply additional educational opportunities, beyond their primary heat-related research function.

Conclusion

In this study, we have developed a simple metric called humidex degree hours (HDH) that integrates apparent temperature intensity and duration. It has potential use as a tool for identifying heat-vulnerable locations within and around urban centres. Our results highlight the value to public health organizations of in situ meteorological data when evaluating potential vulnerability during extreme heat events. We recommend apparent temperature measurements as an important input layer to heat vulnerability mapping initiatives that are underway in large metropolitan areas. To do this properly, geographic expansion and densification of a network of meteorological stations is required. Such a network will permit an improved level of spatial resolution associated with HDH mapping. With an expected increase in frequency, intensity and duration of EHEs across North America during in the twenty-first century, a detailed understanding of the spatiotemporal nature of heat exposure (measured using apparent temperature) is an essential step in developing strategic hot weather response strategies