Validating Satellite-Derived Land Surface Temperature with in Situ Measurements: A Public Health Perspective

Background: Land surface temperature (LST) and percent surface imperviousness (SI), both derived from satellite imagery, have been used to characterize the urban heat island effect, a phenomenon in which urban areas are warmer than non-urban areas. Objectives: We aimed to assess the correlations between LSTs and SI images with actual temperature readings from a ground-based network of outdoor monitors. Methods: We evaluated the relationships among a) LST calculated from a 2009 summertime satellite image of the Detroit metropolitan region, Michigan; b) SI from the 2006 National Land Cover Data Set; and c) ground-based temperature measurements monitored during the same time period at 19 residences throughout the Detroit metropolitan region. Associations between these ground-based temperatures and the average LSTs and SI at different radii around the point of the ground-based temperature measurement were evaluated at different time intervals. Spearman correlation coefficients and corresponding p-values were calculated. Results: Satellite-derived LST and SI values were significantly correlated with 24-hr average and August monthly average ground temperatures at all but two of the radii examined (100 m for LST and 0 m for SI). Correlations were also significant for temperatures measured between 0400 and 0500 hours for SI, except at 0 m, but not LST. Statistically significant correlations ranging from 0.49 to 0.91 were observed between LST and SI. Conclusions: Both SI and LST could be used to better understand spatial variation in heat exposures over longer time frames but are less useful for estimating shorter-term, actual temperature exposures, which can be useful for public health preparedness during extreme heat events.


Introduction
Use of thermal remote sensing and advanced spatial modeling are emerging trends in environ mental epidemiology and public health. Geospatial technologies provide a valuable resource to assist public health practi tioners and emergency response planners in identifying areas that are most at risk and using these scientific outputs to inform policies and practices. Thermal remote sensing products such as thermal images captured by the Landsat5 Thermal Mapper (L5TM) (NASA 2013) instrument have been used to study areas of higher relative tempera tures within a city, also known as microurban heat islands (Johnson 2009). L5TM has an advantage over other sen sors, such as the Moderate Resolution Imaging Spectroradiometer (MODIS) (NASA 2011), in that L5TM provides a spatial resolution of 120 m (compared with 1,000 m for the ther mal band of MODIS); however, it provides only 16day repeatability, at best, compared with 1day repeatability for MODIS (Aniello et al. 1995). The Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) is another sensor that could be used, but imagery is not available free of charge (NASA Jet Propulsion Laboratory 2012). The data captured by satellite can be transformed into several helpful measures, including land surface temperatures (LST) and percent surface imperviousness (SI). LSTs are a primary factor in determining surface radiation and human comfort in cities (Weng 2009).
The higher spatial resolution of L5TM information is important in microurban heat island studies, so we have focused on those data here. The SI is defined as the percent of the surface of an area that is not penetrable by water, such as concrete or asphalt, and can be mapped at a 30m resolution with L5TM. This characteristic has been commonly used in studies to assess the degree of urbanization of an environment as well as explore the spatial extent of surface urban heat islands (Roy and Yuan 2009).
The relationship between LST and vege tated areas has been documented in the literature. A study by Aniello et al. (1995) compared the spatial distribution of micro urban heat islands and tree cover in Dallas, Texas, using L5TM and geographic infor mation systems (GIS). They examined the usefulness of L5TM for classifying tree cover information and using thermal band 6 to produce a thermal map of Dallas, Texas. Their methods involved processing and clas sifying L5TM images and tree cover data in GIS. Although L5TM data were useful for mapping microurban heat islands in Dallas, the authors recommended use of exact on theground temperatures for image calibra tion in future studies. Remote sensing data have been used to help model urban surface temperatures; specifically, validating LST data with actual ontheground temperature measurements, known as groundtruthing. For example, strong correlations between satellitederived air temperatures and in situ measurements were found when charac terizing urban heat island intensity in Hong Kong, using ASTER satellite imagery (Fung et al. 2009). Comparisons between ground tempera tures and estimated temperatures using imagery from MODIS, the National Oceanic and Atmospheric Administration Advanced Very High Resolution Radiometer (NOAAAVHRR), and L5TM, showed a very high correlation in both urban and rural areas (Rigo et al. 2006). Although previous studies have used Landsat to create predic tion models for surface temperature, and shown strong correlations between surface temperature and surface imperviousness (Yuan and Bauer 2007), few studies have simulta neously explored the relationship among SI, volume 121 | number 8 | August 2013 • Environmental Health Perspectives groundbased temperature measures, and LST calculated from thermal imagery.
Adaptation to the consequences of climate change, as future scenarios of heatrelated morbidity and mortality become a major public health concern, requires predicting areas of high vulnerability to heat in cities. Epidemiologic studies on heat and health have begun to use satellitederived tempera tures (instead of temperature data from the nearest airport) and land cover to potentially provide more refined heat exposure classifica tions. A study in Philadelphia used GIS and thermal imaging to investigate the relationship between the spatial distributions of vulner able populations, urban heat island intensi ties, and heatrelated deaths (Johnson and Wilson 2009). The authors recommended that more multiyear studies use spatial modeling and remote sensing methods to better help determine areas of risk throughout cities Uejio et al. (2011) used ASTER data to determine the magnitude and spatial varia tion of mean radiant surface temperature for different densities of impervious surface area (ISA), to further document the change in air trends and air quality that can result from transforming land use from rural to urban. Harlan et al. (2013) used Landsat (30m reso lution) to calculate the Normalized Difference Vegetative Index (NDVI) and estimated sur face temperatures.
There is still a need to understand how proxies for heat exposure correlate with actual heat exposure. Harlan et al. (2013) provided correlations between the mean NDVIs and surface temperatures at the census block group level. Our study goes beyond this compari son by correlating SI, land surface temperature (estimated from Landsat), and actual tempera ture measurements made by a groundbased temperature monitor network over the sum mer. Our study is novel for evaluating the cor relations between satellitederived temperatures and a groundbased temperature network with high temporal (10 min) and spatial resolution (19 monitors over a range of levels of SI within a single county) at a height relevant to human health (1.5 m above the ground). Further, it characterizes these features in metropoli tan Detroit, Michigan, where people may be poorly adapted to heat and where several epi demiologic studies have already shown impor tant and socially unequal health consequences associated with hot weather (Anderson and Bell 2009;O'Neill et al. 2003;Schwartz et al. 2004;Sheridan and Kalkstein 2010).
Indeed, integrating information from groundbased temperature monitoring net works and satellitederived images using a GIS platform can provide useful data for exposure assessments in most urban areas, specifically for the study of heatrelated death and illness. Validation of satellite data sources by ground truthing (i.e., information that is collected on location) can help characterize and identify neighborhoodlevel urban heat islands; this information could be useful for public health professionals and urban planners to prevent heatrelated mortality and other adverse health effects from high summer temperatures. Such data could also direct intervention strategies to reduce the urban heat island effect. Some previously published work did not explicitly address the practical challenges of integrating insights from groundtruthing studies with public health research and applications (Lo and Faber 1997).
The purpose of this study is to apply a public health perspective to a determination of whether spatial variation of temperatures within a network of groundbased outdoor tempera ture monitors is correlated with satellitederived LST and SI. Although we did not expect LST (which represents the tempera ture of the ground) to completely predict the temperature of the air at a height relevant to human health, we hypothesized that the air temperatures measured by a groundbased temperature monitoring network would be highly correlated with LST as well as with values of SI.

Methods
Ground-based temperature-monitoring network and surface imperviousness. Sites in our groundbased temperature monitor network in the Detroit metropolitan region (Wayne County, MI) were selected, with site SI val ues ranging from 0 to 100% imperviousness and with buffer zones around each site ranging from 0 to 800 m. We picked both urban and rural locations to assess the temperature dif ferences among areas within the same county that might have different landuse patterns. This strategy was designed partly to evaluate the existence of urban heat island structure in the Detroit metropolitan region (Oswald et al. 2012;Zhang et al. 2011). Using SI from the U.S. Geologic Survey (USGS) National Land Cover Database (NLCD) product (Multi Resolution Land Characteristics Consortium 2013), we performed halfmile smoothing of every pixel and classified SI by decile. Once the SI surrounding various prospective residential sites for placing the temperature monitors was established and a range of SI levels was ensured by the sampling strategy, home occupants were approached, told the purpose of the study, and asked if they would be willing to have a temperature monitor outside their homes. All volunteers signed an agreement letter to participate. The research was compliant with all relevant national, state, and local human subjects regulations.
HOBO Pro v2 U23002 (external temperature/relative humidity) outdoor tem perature monitoring devices (Onset HOBO Data Loggers, Pocasset, MA) were calibrated and used to record temperature and relative humidity at 10min intervals from 13 June to 30 September 2009. The calibration process involved collocating the monitors in a con trolled environment along with a temperature probe from the National Institute of Standards and Technology (Gaithersburg, MD) to ensure that the temperatures recorded were within the accuracy range reported in the operation man ual: ± 0.21 o C, within an ambient temperature range of 0-50 o C (Onset HOBO Data Loggers 2013). Monitors were positioned in residential grasscovered backyards of volunteers follow ing a strict placement protocol that required monitors to be sited a) at least 10 ft (3 m) away from buildings, homes, and trees; b) 1.5 m from the actual surface [to better assess the level of exposure that would be experienced by a person of average height and to be consistent with the instrument siting protocols used by the World Meteorological Organization, as well as NOAA's National Weather Service (World Meteorological Organization 2008)]; c) not in the direct pathway of automatic lawn sprinkling systems; d) not in a shady area or near falling objects; e) away from power lines and swampy damp ground; and f ) facing southwest.
Processing the L5-TM satellite image to derive land surface temperature. Images of Earth were taken nearly continuously from 1 March 1984 to January 2013 on a 16day cycle by the L5TM, which consistently imaged the Detroit Metropolitan Region at about 1205 hours Eastern Daylight Time (EDT) at each pass. L5TM captured images collected at a 705km altitude, 185km swath, 120m spatial resolution for thermal band data and 30m resolution for the other spec tral bands. The satellite images captured by L5TM are free and downloadable from the USGS (USGS 2013).
Satellite images were downloaded for use only if they met the following criteria: The images were captured during the study period, covered the entire study area (geo graphically), had < 14% cloud cover, and were taken under clear weather conditions. More cloud cover and unclear weather conditions can inhibit the signal strength reflected back to the satellite and cause under estimation of the ground surface tempera ture. Each image had seven spectral bands of information. Thermal infrared band 6 (10.4-12.5 μm) provides the data that can be converted from raw digital numbers to LST. To convert from a digital number to a temperature, we needed calibration formulas, atmospheric correction tools, and transforma tions ( Figure 1). Once an image was selected by our criteria, we used ERDAS Imagine 9.2 software (Leica Geosystems, Inc., Atlanta, GA) to convert the image (from a TIFF file to an IMG file) to a usable format for GIS. ArcGIS version 9.3.1 (ESRI, Redlands, CA) was used to perform the calculations outlined below. First we converted the digital num bers taken from the raw image into atsensor spectral radiance [L λ , the temperature read at the sensor, W/m 2 × sr × μm (watts per square meter per steradian per micrometer)]. The source equations and calibration constants developed specifically for L5TM images were used from the process outlined by Chander et al. (2009). L TOA was then used to calculate an actual ground surface temperature.
Because the satellite signal received by the sensor is in space, the effects of the atmosphere (i.e., air pollution, weather) and surface emis sivity (the ratio of the radiation emitted by a surface to the radiation emitted by a black body at the same temperature: i.e., how well different surfaces reflect solar energy) can have considerable influence on the accuracy of the satellitederived surface temperature. To account for these influences, we used a tool that estimates atmospheric influences in conjunction with a data layer of emissivity for the study area. Using this webbased atmo spheric correction parameter tool developed by Barsi et al. (2003), we estimated three scene specific parameters for each satellite image. The tool required the following inputs for each satellite scene: year, month, day, Greenwich Mean Time (GMT), and latitude and lon gitude coordinates. The outputs of this tool were the three parameters: atmospheric trans mission (T, unitless), upwelling radiance (L u , W/m 2 × sr × μm) and downwelling radiance (L d , W/m 2 × sr × μm), where W/m 2 × sr × μm are the units of spectral radiance.
Values of emissivity for the study area were then estimated by examining the land use/land cover designations, downloaded for the region from the 2006 NLCD (Multi Resolution Land Characteristics Consortium 2013). We considered the differences in emissivity of various impervious surface land cover types in the main equation to calcu late surface temperature. Because our land cover data do not distinguish among different types of impervious surface, we were unable to represent possible differences in emissivities among them. We created a layer of emissivity (ε; range, 0-1) for the study area based on reference emissivity values for various land cover classes used in other studies (Lillesand et al. 2007).These values of emissivity, cou pled with the atmospheric correction param eters, were used in the following equation to calculate the radiance of a blackbody target of kinetic temperature (L T ), which ultimately represented surface temperature: where L T = radiance of a blackbody target of kinetic temperature and L TOA = at sensor spectral radiance, W/m2 × sr × μm. We then transformed L T into a temperature in Kelvin (using Planck's equation), and then converted Kelvin to degrees Celsius.
Once the scene was transformed to a sur face temperature in units of degrees Celsius, temperatures outside our range of inter est (< 0°C) or areas of no data (water) were masked out of the layer (given a value of NoData). The implausible ranges were likely a result of some cloud cover over a certain point, values over a body of water, or possibly a source of error in the reflectance value that would cause noise in the analysis.
Geographical and statistical analysis. Using spatial analysis tools in ArcGIS, we averaged the LST and SI over the areas of the follow ing seven concentric circles with different radii around each outdoor moni toring unit (buf fers): at the point (0 m) and at 100, 200, 300, 400, 500, and 800 m. We assessed the values at different buffers because they can represent physical processes that can occur at different spatial scales within the urban canopy layerthe layer of the urban atmosphere extending upward from the surface to building height (Roy and Yuan 2009). These spatial scales range from the microlevel (at the home = 0 m) to more macrolevel (block, neighborhood) exposures. The physical mechanism of cor relation is that thermometers "sense" tempera ture that is transferred from the "source area" (i.e., surfaces below, around) to the sensors through turbulent transport. Thus the rela tionship between source area (i.e., LST) and thermometer depends on both atmospheric and surface states. Furthermore, LST and groundbased temperature can be influenced by a number of physical factors rele vant to the study area-surface heterogeneity, considerable variability in temperature over small areas, and physical structures-and the varying buffers allowed exploration of on what scale these fac tors might operate. From a health perspective, understanding correlations at these different spatial scales can inform tools for use at the urbanlocal scale, to predict "hot spots" where prevention of heat illness and deaths is espe cially needed. The grid cells for SI (30m reso lution) and LST (120m resolution) that were contained completely within and intersected the corresponding buffer were included in the calculation. The zonalstats operation was used to generate the average LST and SI for each of the buffers. Spearman rank correlation coefficients were calculated between the mean LSTs and SIs and the temperatures from the outdoor monitoring network averaged over five time periods: 1205 hours [to correspond to the time of the satellite image (average of 1200 and 1210 hours)]; average temperature from 0920 to 1210 hours (3hr average); aver age temperature from 0400 to 0500 hours on 19 August 2009 (nighttime tempera ture); average temperature from 1210 hours on 18 August to 1210 hours 19 August 2009 (daily); and, August average monthly tempera ture (monthly). The maximum and minimum temperatures and standard deviation for each group of measurements were calculated. These different time periods were chosen to see whether LST and/or SI taken at one point in time would give a better picture of instanta neous versus longerterm spatial variation in temperatures in the study area.
Related to this point, we explored how well the LST captured by the one usable 2009 LST image represented the LST over a longer time frame, especially in this region where population growth and economic Figure 1. The processing method for converting raw satellite images to land surface temperature. Parallelograms denote data inputs, squares denote calculations, and the shaded oval denotes the final value (temperature leaving Earth's surface).

Results
Temperature data from 19 of 24 groundbased monitors positioned throughout the county were used in the analysis. Five of the tem perature monitors were excluded due to sit ing conditions that could have jeopardized the readings, such as being too close to a build ing. A search in the Landsat 4 and L5TM archive data sets for the Detroit area (latitude 42.331427, longitude -83.0457538), yielded 21 satellite images. Of these 17 had a cloud cover percentage > 14%, and 16 did not cover the study area geographically. Consequently, one image, taken by the satellite on 19 August 2009, met our criteria (i.e., covered the geo graphic area of interest; had cloud cover of < 14%, captured during the time period of the study (13 June-30 September 2009). This image was acquired during the daytime at 1805 hours GMT (1205 hours EDT). The quality of band 6 was scored as a 9, the highest score for images. This value reflects the quality and level of errors detected in the image (see USGS 2013 for explanation). Figure 2 shows the final pro cessed band 6 image of the study area.
The highest levels of mean SI were seen in the 500m buffer zones across all of the loca tions (Table 1). Overall, the New Center loca tion had the highest SIs (range, 42-87.7%), and the New Boston location had the low est overall SIs (range, 9.0-18.5%). The New Center area location had the highest LSTs compared with other locations, from 24.4 to 25.6°C, whereas the New Boston area had the lowest overall LSTs, from 18.2 to 18.4°C. In terms of the groundbased temperature readings, the instantaneous 1205hours time point at each location showed a higher maxi mum temperature than the other recorded time points.
At least two statistically significant cor relations, for each radius distance, were seen between LST and the groundbased tem peratures for the daily and monthly tempera ture for all buffers except 100 m, as shown in Table 2. At least three statistically signifi cant correlations were also seen for each radius distance between SI and groundbased air temperature measurements for nighttime tem perature, daily temperature, and monthly tem perature at all buffers except the 0m point. For the relationship between SI and LST, sta tistically significant correlations were found with Spearman correlation coefficients ranging from 0.49 to 0.91, as shown in Table 3.
In the analysis comparing spatial varia tion in LST using five summertime satellite images from 2002, 2003, 2004, 2008, and 2009, LST temperature ranges and the areas with the highest temperatures were consis tent over the years. The 2009 LST scene was highly correlated with the 5year composite LST scene (R 2 = 0.96). We also found a high correlation between the SI scenes from two different years, 2001 and 2006 (R 2 = 0.98).

Discussion
The purpose of this study was to assess the rela tionship between LST and SI measurements and groundbased air temperature measure ments in the Detroit metropolitan region. Our results showed a statistically significant rela tionship between LST and SI at all buffers, as well as LST and SI and the groundbased air temperatures at certain buffers. These correla tions between LST and SI are consistent with findings from other published studies (Imhoff et al. 2010;Uejio et al. 2011;Yuan and Bauer 2007;Zhou and Shepherd 2009). This suggests that SI data, which require much less processing than the LST data, could be used as a proxy for LST. Consequently, public health researchers and practitioners may still be able to use a fairly straightforward method to determine city hot spots using high SI as an indicator of potential increased temperature exposure.
Our study used standard methods that facilitate comparisons with other work, and is the first analysis of this kind during the summer time in an large, urban Midwest city. Detroit has unique features, including a higher proportion of vacant lots than in other metro politan areas. Additionally, our study simul taneously explored the relationship between SI, LST calculated from thermal imagery, and groundbased temperature measures, adding the groundtruthing element that has been called for to independently validate the satellite imagery as a proxy for humanscale exposures (Voogt and Oke 2003).
The consistency of our findings with those of other studies suggests that these unique fea tures do not impair the overall utility of satel lite imagery for public health applications. In Detroit, the 2009 satellitederived LST image-corrected for atmospheric effects and spatial variations in emissivity-as well as the SI image from the 2006 NLCD might be suitable to represent air temperature variability between sites for heat exposure studies in the region or for targeting heathealth interventions. Because our landcover data do not distinguish among different types of impervious surface, we were unable to represent possible differences in emis sivities among them, and this is a type of vari ability that contributes to possible uncertainties in our analysis. The analysis we did comparing the 2009 scene with LST calculated from four previous years' summertime scenes showed that the LST estimated from the satellite images was relatively consistent over time, suggesting  HOBO monitors that changes in land use were not substantial in the Detroit metropolitan region. Further, for application in heathealth studies, LST is bet ter suited for representing physical properties that are stable over time and can affect human temperature exposure rather than as a proxy for actual ambient air temperature at a particular point in time.
Our results complement those of two sis ter studies of the Detroit metropolitan region that examined spatial variation of temperature during the entire summer of 2008. The first study used the same observational network of air temperature monitors that we used in the present study in conjunction with airport temperature monitors and monitors operated by the state of Michigan Department of Environmental Quality (Oswald et al. 2012). This study found the correlation between summer mean daily low temperature anoma lies (the daily residuals at each location minus measurement uncertainty) in 2009 and SI in 2001 (r = 0.68 at the 200m buffer, p < 0.001) to be higher than between daily temperature anomalies and other geographic characteris tics. This suggests that in relatively sprawling cities, the urban heat island most closely fol lows SI and would have a unique structure in each city based on the SI structure (Oswald et al. 2012). A second study used geospatial approaches to create a continuous, spatial layer of estimated air temperature, and found high correlations between SI measured in 2006 and observed air temperatures in 2008 (Zhang et al. 2011). Neither of those two studies examined LST, but the fact that they observed correlations between SI and air temperatures using different methodologies supports our finding that SI and, by extension, LST are moderately correlated with air temperature.
Previous studies have groundtruthed L5TM data using airborne thermal scan ner flights (Voogt and Oke 1998) or using satellite data in conjunction with ground based air temperature measurements with other remote sensing predictors to create a model for air temperature (Cristóbal et al. 2008). Other researchers have also created a   (Kestens et al. 2011). Using ambient tempera tures recorded from multiple meteorological stations, they found that the 3day average air temperature was a strong predictor of LST, as were the NDVI and land cover categories. Their results suggest that increasing the num ber of meteorological and geographical predic tors could provide more precise estimates of heat exposure in urban areas. An added benefit of using SI as a proxy for temperature exposure is that increasing vegetative cover and other changes can reduce the heattrapping potential of the urban land scape; therefore, results of studies using SI could be of direct relevance for policy changes. This finding might be helpful in the urban planning sector.
Correlations between LST and ground based temperature measurements (Table 2), were stronger at the largest radii (e.g., 500 m and 800 m), and stronger using the aver age temperature from day 1 to day 2 and the monthly temperature. Several possible reasons for this come to mind. Urban areas are hetero geneous in topography, physical structures, land use, and the like, so averaging of the LST temperature over a larger buffer zone may drown out the physical noise that can influence the air temperature at a specific point.
The stronger correlations when tempera tures are averaged over longer time spans sug gest that instantaneous temperatures are less indicative of the overall spatial pattern of the temperature tendencies. Morning tempera tures are likely less correlated due to lack of turbulent transport (the main mechanism relating source area to sensor) and influence of cold air drainage (i.e., topography). However, in terms of estimating personal exposure, lower correlation between these satellite data sources and the actual groundbased air temperature readings at 1205 hours and the 3hour average underscores the importance of identifying other tools that can better gauge actual shortterm temperature exposure near the ground surface, especially when health outcomes that can result from acute exposures are of interest. Previous epidemiologic studies of heat and daily mortality that have included Detroit have found that heat exposure on days 0-1 have been most relevant (e.g., Anderson and Bell 2009), although heat wave durations of at least 4 days may have an additional effect (e.g., Gasparrini and Armstrong 2010). The day is the common time unit of analysis for administrative databases of health outcomes (hospitalization, deaths, births), but other clinical outcomes that could be affected by heat (e.g., blood pressure, pulse rate) may be available at a finer time scale, such that hour specific temperature data would be relevant. Longer duration of warm temperatures could also be relevant to both the exposure and the health resilience of residents, relating to air conditioning use and overall energy demand in homes.
Using the LST and SI data in conjunc tion with health outcome data could provide a more general understanding of spatial heat vulnerability. For example, an epidemiologic investigation of a 1993 extreme heat event in Philadelphia used satellite imagery and geostatistical methods to determine whether vulnerability to heatrelated mortality was higher in areas with higher urban heat inten sity (Johnson and Wilson 2009). The authors found that the heat load of the environ ment detected by the Landsat satellite data was potentially a contributing factor to heat related deaths during the summer of 1993, and that the thermal data used in this study could be used to develop models of place based vulnerability.
More frequent daily observations are made by MODIS. However, this sensor records thermal emission at a spatial resolution (1 km) too coarse for microurban heat island inves tigations. Future studies should investigate the public health implications of this tradeoff between temporal frequency and spatial reso lution. Additionally, researchers have created new methods to better use satellite imagery to assess land surface temperature. In particular, physical and statistical methods for downscal ing MODIS scenes (Liu and Pu 2008) and enhanced physical methods that will reduce downscaling uncertainty, reduce the smooth effects, and block effects due to isothermal assumption (Liu and Zhu 2012) could be incorporated into health studies.
Our groundbased temperature moni tors were mounted 1.5 m above the ground, and the nonstatistically significant relation ships that we found between LST and the groundbased temperature monitoring net work-for all buffers for the instantaneous and 3hr average temperatures as well as the 0 to 700m buffers for average daily tempera ture-might be a result of mixing, advection, and convection processes within the bound ary layer that influence the air temperatures recorded by the outdoor temperature moni tor. Because we are comparing two different types of measurements-surface tempera ture and air temperature-the correlations between these measurements might not be as strong due to logistical (e.g., timing and resolution) as well as physical (e.g., advec tion, wind) considerations that could affect the derived surface temperatures.

Limitations
The availability of satellite products is a key limitation. Of 21 LST scenes examined, only one scene was usable in that it lacked signifi cant cloud cover and covered the study area geographically. The 16day cycle on which Landsat images are acquired for a specific area does not afford researchers the opportunity to compare multiple images within a useful timeframe. Additionally, L5TM data have a large spatial resolution (120 m), which might not capture the full heterogeneity of an urban environment (Johnson 2009). Further, although we were able to match the time of acquisition of the ground data to the same time as the satellite data, the 1205 hours passing time of the Landsat satellite is not optimal for temperature-health studies. First, this time generally corresponds to a time of the day when ground temperatures transi tion from being cooler than air temperature to being warmer than air temperature. This means that, within the diurnal cycle, surface temperatures are not as significant drivers of air temperatures as they are later in the day. Second, exposure studies have tended to focus on maximum and minimum tempera tures, and this time corresponds with neither of these (Basu 2009). Although shorterterm groundbased temperature timeframes did not yield strong correlations with SI or LST, com posites of older satellite images could be one input into a more comprehensive planning tool or index to help describe vulnerability.

Conclusions
Our results support the need for an increased effort, nationally, by public and private entities, to create useful remotely sensed data sources that can be applied to public health practice. A workshop report from the National Academy of Sciences (National Research Council 2007) discussed the challenges and potential applica tions of using remotely sensed data for public health. The report indicated that one of the major challenges to applying these remotely sensed data in the health arena is the limited in situ groundtruthing data accompanying remote sensing technology to verify analysis, and the high learning curve to using the tools required to analyze remotely sensed data. Our study gathered ground truthing data needed to validate satellite derived LST as well as SI. However, our results highlight that issues of spatial resolution, image availability over certain time periods, and the complex urban landscape remain challenges in the effort to integrate remotely sensed data with public health research and practice. From a public health perspective, it is important to target resources and health inter ventions for the most vulnerable populations. The availability and usefulness of remote sens ing data, integrated with social and economic demographic data, can provide a powerful tool for assessing vulnerability. A quality of life study conducted by Athens-Clarke County, Georgia, used one cloudfree image, aerial photographs, and U.S. Census data to overlay biophysical (land surface temperature, NDVI, land use, and land cover) and socioeconomic layers (population density, per capita income, median home value, education) to create a quality of life indicator (Lo and Faber 1997). The research found a strong relationship between biophysical and socioeconomic vari ables, which could be useful to assess vulner ability in the public health arena. In the field of heat epidemiology, being able to utilize a userfriendly data source such as SI as a proxy for surface temperature exposure can further our understanding of spatial vulnerability to heat. Another study has already shown that areas of the Detroit metropolitan region with high SI had statistically significant correla tions with several sociodemographic variables: being ≥ 65 years of age and living alone, being able to leave the home, education level, living below the poverty line, and being nonwhite (WhiteNewsome et al. 2009).
There are several ways that remote sensing data could be better integrated into public health practice: a) increasing the capture fre quency of remotely sensed images available for research and planning purposes; b) provid ing more highly processed data accessible to the public, at a finer resolution (10-15 m) and if possible at a higher temporal frequency that could be more useful for city and county level authorities; and c) commissioning more research to groundtruth satellitederived land surface temperatures for differentsized urban areas, and establishing a set of fairly simple, standard best practices that can be used to esti mate the influences of atmospheric and other factors on deriving a precise LST value from remote sensed imagery could be useful for planning for extreme heat events. Landsat data are the most consistent and widely available source of relatively highresolution thermal information from satellites, but can be limited due to the number of clear images available at certain days and times. A gap in availability of these data exists, but continued acquisition of these data or data of comparable resolution has the potential to provide important spatial information about differential heat exposures.
One contribution of our study is to under score the importance of the limitations of data, and emphasize that it is critical to have data that are accessible, useful, and timely for those working in public health. One of the main objectives of this study was to see whether a "non-remotesensing professional" could create a tool-using available datathat can be used to estimate heat exposure. Reporting on the challenges we faced in doing this is one way to bring the issue to the attention of the remotesensing community. As more practitioners demand these data, our research and other research that attempts to use the simplest methods should provide the impetus to fill the gaps in data to overcome these limitations.