remote sensing

Article
Evaluating the Variability of Urban Land Surface
Temperatures Using Drone Observations
Joseph Naughton and Walter McDonald *
Department of Civil, Construction & Environmental Engineering, Marquette University, P.O. Box 1881,
Milwaukee, WI 53211, USA
* Correspondence: walter.mcdonald@marquette.edu; Tel.: +1-414-288-2117




Received: 30 May 2019; Accepted: 16 July 2019; Published: 20 July 2019


Abstract: Urbanization and climate change are driving increases in urban land surface temperatures
that pose a threat to human and environmental health. To address this challenge, we must be able to
observe land surface temperatures within spatially complex urban environments. However, many
existing remote sensing studies are based upon satellite or aerial imagery that capture temperature
at coarse resolutions that fail to capture the spatial complexities of urban land surfaces that can
change at a sub-meter resolution. This study seeks to fill this gap by evaluating the spatial variability
of land surface temperatures through drone thermal imagery captured at high-resolutions (13 cm).
In this study, flights were conducted using a quadcopter drone and thermal camera at two case
study locations in Milwaukee, Wisconsin and El Paso, Texas. Results indicate that land use types
exhibit significant variability in their surface temperatures (3.9–15.8 ◦ C) and that this variability is
influenced by surface material properties, traffic, weather and urban geometry. Air temperature
and solar radiation were statistically significant predictors of land surface temperature (R2 0.37–0.84)
but the predictive power of the models was lower for land use types that were heavily impacted
by pedestrian or vehicular traffic. The findings from this study ultimately elucidate factors that
contribute to land surface temperature variability in the urban environment, which can be applied to
develop better temperature mitigation practices to protect human and environmental health.

Keywords: land surface temperature; drones; unmanned aerial vehicles; thermal remote sensing

1. Introduction
Urban areas across the world are subject to thermal stresses caused by the surface urban heat island
(SUHI) effect where urban land surfaces experience higher temperatures than their surrounding rural
areas. This is in large part due to the replacement of undeveloped vegetated land with anthropogenic
materials that absorb more solar radiation and have different heat capacity and surface radiative
properties [1]. This results in higher surface temperatures that pose a significant threat to human
health [2], as well as higher storm runoff temperatures that can harm aquatic life [3–5]. These stresses
are only expected to grow with increases in global temperatures and urban populations; therefore, it is
critical that we understand the fundamental processes that drive land surface temperature (LST) to
develop solutions that can protect human and environmental health.
To that end, thermal remote sensing is an important tool for evaluating urban land surface
temperatures. This includes satellite sensors such as ASTER, MODIS and Landsat that can capture
land surface temperatures at 30 m–1 km resolutions [6]. Data from these satellites have been used to
extensively study urban land surface temperatures and their effects [7–14]. However, while satellite
remote sensing is valuable for evaluating LST across a city scale, the spatial resolution precludes its
applications to smaller spatial scales that better reflect the spatial complexity of the urban environment.
To acquire higher resolution thermal data, studies have used aerial reconnaissance or downscaling

Remote Sens. 2019, 11, 1722; doi:10.3390/rs11141722 www.mdpi.com/journal/remotesensing

Remote Sens. 2019, 11, 1722 2 of 18

techniques [15,16]; however, these are still at resolutions (4–10 m) that cannot capture changes that
occur on a sub meter resolution. Furthermore, satellite remote sensing is temporally constrained to
intervals between 1–14 days. Aerial flights do not have the same temporal constraints; however, doing
so at on-demand temporal resolutions would not be economically practical. Therefore, these methods
are inadequate for evaluating changes in urban LST that occur throughout the day or capturing the
spatial heterogeneity of urban LST at small scales.
This challenge is important to overcome as urban land surfaces are spatially complex and
significant variations in land cover can occur on a sub meter spatial resolution [17]. While existing
research has demonstrated that the spatial configuration of land use classifications at a city scale are
important (i.e. industrial, residential, forest) [18,19], less is known about the importance of the spatial
configuration and variations in LST at smaller scales (i.e., sidewalks, grass medians, flowerbeds, etc.).
In addition, the urban environment is dynamic and land surface temperatures can be significantly
influenced by other factors besides land cover material properties [20]. Land surface temperature
may therefore vary significantly across small spatial scales; however, the factors that control this
variation are not well defined. Doing so requires direct measurements of surface temperatures across
wide spatial and temporal scales, yet little research to date has evaluated the spatial variability in
temperature among urban land use types in sub-meter resolutions. This may be due to measurement
limitations, as satellite data is too coarse and in-situ temperature probes are too expensive to densely
distribute across an urban landscape. Therefore, new and innovative approaches to measuring land
surface temperatures at small spatial and temporal scales are needed to assess thermal variability
across land use types in the urban environment.
Unmanned Aerial Vehicles (UAVs) or drones, are a technology that can meet this challenge.
Recent advances in UAVs and radiometric thermal cameras have made it possible to capture land
surface temperatures on-demand and at sub-meter spatial resolutions that accurately reflect the spatial
complexity and detail of land surface temperatures in the urban environment [21]. UAVs also have
advantages in that they can be flown on demand to capture LST at temporal resolutions unmatched by
satellite or aerial imagery. While the limited battery life of around 30 minutes for quad-copter UAVs
constrains the area that can be captured in a single flight, their spatial and temporal resolutions offer
significant advantages for evaluating the variability of LST in the urban environment at fine spatial
and temporal scales.
We therefore present a study to evaluate the variability of temperatures across urban land surfaces
using a UAV. In this study, we apply a UAV and radiometric thermal camera to capture land surface
temperatures at high-resolutions (13 cm) in two case study locations: Milwaukee, Wisconsin and
El Paso, Texas. Using data collected throughout a calendar year, we evaluate the variability in land
surface temperatures, develop models to predict mean land surface temperature based upon weather
parameters and evaluate the diurnal trends in urban land surface temperature. To do so, we (1)
quantify land surface temperature variability across different surface types, (2) evaluate variance in
temperature across different surface types based upon meteorological and/or other derived parameters
(e.g., albedo, normalized difference vegetation index, apparent thermal inertia, etc.), (3) predict land
surface temperature based upon meteorological parameters and (4) assess diurnal variability in land
surface temperature magnitude and uncertainty. Ultimately, this study helps to elucidate factors that
contribute to land surface temperature variability in the urban environment at small spatial scales,
which can then be applied to develop better temperature mitigation strategies.

2. Materials and Methods

2.1. Case Study Locations

Two case study locations were chosen for this project: (1) a portion of Marquette University’s
campus in Milwaukee, WI and (2) a portion University of Texas El Paso’s (UTEP) campus in El Paso,
Texas (Figure 1). The Marquette and UTEP case study areas were roughly 21,300 m2 and 27,300 m2 ,
 Remote Sens. 2019, 11, 1722 3 of 18

Remote Sens. 2019, 11, x FOR PEER REVIEW 3 of 19

respectively and included a balance of both natural landscape and impervious gray surfaces. Surface
types within each case study location were manually delineated using ESRI’s ArcMap software.
nine surfaces types identified at Marquette and UTEP and their respective surface areas are listed in
The nine surfaces types identified at Marquette and UTEP and their respective surface areas are listed in
Table 1. The specific locations on each campus were chosen for their variety of surface types,
Table 1. The specific locations on each campus were chosen for their variety of surface types, similarities
similarities in land use between the two locations and suitability for drone takeoff/landing and flying.
in land use between the two locations and suitability for drone takeoff/landing and flying. In addition,
In these
addition, theseprovide
locations locations provide in
a contrast a contrast in geography,
geography, climate andclimate
weatherand weather
that that in
are helpful aretesting
helpful
thein
testing the generalizability of our findings. For example, Milwaukee’s climate is classified by
generalizability of our findings. For example, Milwaukee’s climate is classified by Koppen and GeigerKoppen
andas Geiger as DfaContinental
Dfa (Humid (Humid Continental
Hot Summers Hot With
Summers With Year
Year Around Around Precipitation)
Precipitation) and receivesand870receives
mm of
870 mm of precipitation annually, while El Paso is classified as BWk (Cold Desert Climate)
precipitation annually, while El Paso is classified as BWk (Cold Desert Climate) and receives 221 mm and
receives 221 mm of
of precipitation precipitation
annually [22,23].annually [22,23].

(a) (b)

Figure
Figure 1. 1. Visual
Visual imagery
imagery ofof thecase
the casestudy
studylocations:
locations:Marquette
MarquetteUniversity
University (a)
(a) and
and University
University of
of Texas
Texas
El Paso (UTEP) (b). Visual imagery of Marquette was captured from a drone on 11 AugustVisual
El Paso (UTEP) (b). Visual imagery of Marquette was captured from a drone on 11 August 2018. 2018.
imagery
Visual of UTEP
imagery was pulled
of UTEP from Google
was pulled Maps on
from Google 13 March
Maps 2019. 2019.
on 13 March
Table 1. Surface types and surface areas within each case study location.
Table 1. Surface types and surface areas within each case study location.
MARQUETTE UTEP
MARQUETTE 2 UTEP
Surface Type Surface Area (m ) Surface Type Surface Area (m2 )
Surface Type Surface Area (m2) Surface Type Surface Area (m2)
Grass 2738 Rooftop (rammed earth) 503
Grass
Sidewalk 2,738
904 Rooftop (rammed earth)
Desert Shrub 503
173
Sidewalk
Rooftop (composite) 904
336 Desert Shrub
Rooftop (composite) 173
2047
Road (asphalt)
Rooftop (composite) 3299
336 Parking(composite)
Rooftop Lot (asphalt) 1350
2,047
Parking Lot (concrete)
Road (asphalt) 908
3,299 Sidewalk
Parking (concrete)
Lot (asphalt) 9808
1,350
Rooftop (rubber) 6,057 Road (asphalt) 4253
Parking Lot (concrete) 908 Sidewalk (concrete) 9,808
Canopy Cover 4758 Parking Lot (concrete) 4081
Rooftop (rubber)
Shrub/mulch 6,057
2272 Road
Grass(asphalt) 4,253
1270
Canopy Cover
Solar 4,758
65 Parking
CanopyLot (concrete)
Cover 4,081
3782
Shrub/mulch 2,272 Grass 1,270
2.2.Solar
Equipment 65 Canopy Cover 3,782

Remote sensing data was collected using a DJI Matrice 100 (M100) quadcopter UAV. The M100
2.2.
wasEquipment
deployed at our case study locations with three types of camera payloads—visual, multispectral
and infrared.
Remote Thesedata
sensing cameras
was include the
collected DJI Zenmuse
using X3 visual
a DJI Matrice (12 MP),
100 (M100) Zenmuse X3
quadcopter multispectral
UAV. The M100
(Blue-Green-NIR 680–800 nm at 12 MP) and DJI Zenmuse XTR radiometric thermal (13 mm, 30 Hz and
was deployed at our case study locations with three types of camera payloads—visual, multispectral
spectral
and bandwidth
infrared. of 7–13 include
These cameras µm). Additionally, groundX3temperatures
the DJI Zenmuse wereZenmuse
visual (12 MP), validatedX3
using a Nubee
multispectral
NUB8380 Digital
(Blue-Green-NIR Infrared Thermometer.
680-800nm at 12 MP) and DJI Zenmuse XTR radiometric thermal (13 mm, 30 hz and
spectral bandwidth of 7–13 µm). Additionally, ground temperatures were validated using a Nubee
NUB8380 Digital Infrared Thermometer.
Remote Sens. 2019, 11, 1722 4 of 18

2.3. Data Collection Methods

Two datasets were collected during the 2018 calendar year: (1) surface temperature measured at
12:00 PM across the entire year and (2) surface temperature measured on a diurnal cycle. To evaluate
surface temperature across the entire year, fourteen flights in Milwaukee and one in El Paso were
recorded between 26 February and 13 September 2018 (Table 2). To evaluate the diurnal cycle of
temperature, three flights in Milwaukee and one in El Paso measured temperature throughout the
day at 9:00 AM, 12:00 PM, 3:00 PM and 5:00 PM (Table 3). Weather data was collected at Marquette
from a station on top of Engineering Hall and weather data at UTEP was collected from a weather
station 10.5 km away at El Paso International Airport. Each station recorded air temperature, relative
humidity, wind speed, wind direction, relative humidity, solar radiation and atmospheric pressure.
Drone imagery was captured autonomously using a third-party photogrammetry software called
Pix4Dcapture. Using this software, autonomous flight paths were programmed to the drone prior
to each mission. Programmed flight path information included drone speed, altitude and image
overlap. Drone speed was set at 54 km/h for visual and multispectral flights but set at a lower threshold
of 30.6 km/h for thermal flights due to the difference in image capture speed between the two camera
technologies. The flight altitude for each mission was set to the United States Federal Aviation
Administration (FAA) maximum allowable limit of 120 m, which resulted in thermal imagery at a 13 cm
pixel size. Finally, the image overlap was set to 85%, which provided reliable overlap for stitching an
orthomosaic during data processing.

Table 2. Flight log and summary of meteorological variables recorded for Marquette and UTEP during
fifteen noon flights.

Flight Flight Air Temp Relative Wind Speed Wind Dir Solar Rad Pressure
Flight Date
Number Time (◦ C) Humidity (%) (m/s) (Degrees) (kW/m2 ) (kPa)
MU 1 26 February2018 12:00 PM −1.7 54 4.0 225 0.00 102.2
MU 2 12 April 2018 12:00 PM 12.4 65.8 6.8 285 0.41 100.2
MU 3 8 May 2018 12:00 PM 26.4 22.1 4.2 218 0.81 101.7
UTEP 1 20 May 2018 12:00 PM 27.8 26 5.8 120 0.96 101.7
MU 4 13 June 2018 12:00 PM 25.3 33.3 3.4 321 0.89 101.2
MU 5 29 June 2018 12:00 PM 31.5 54.7 5.4 193 0.80 101.0
MU 6 11 July 2018 12:00 PM 25.9 44.1 2.0 91 0.78 101.9
MU 7 12 July 2018 12:00 PM 27.4 43.2 5.9 204 0.60 101.8
MU 8 17 July 2018 12:00 PM 25 38.9 3.0 38 0.74 101.6
MU 9 18 July 2018 12:00 PM 22.5 56.3 3.1 101 0.83 101.8
MU 10 25 July 2018 12:00 PM 28.6 31.9 2.5 271 0.83 101.4
MU 11 10 August 2018 12:00 PM 25.8 58.8 2.3 84 0.77 101.3
MU 12 31 August 2018 12:00 PM 25.8 49.9 4.0 158 0.09 101.6
MU 13 12 September 2018 12:00 PM 26.1 55.3 3.1 168 0.56 101.9
MU 14 13 September 2018 12:00 PM 22.8 64.9 4.2 127 0.68 102.0

Table 3. Flight log and summary of meteorological variables recorded for Marquette and UTEP during
four diurnal flights.

Flight Flight Air Temp Relative Wind Speed Wind Dir Solar Rad Pressure
Flight Date
Number Time (◦ C) Humidity (%) (m/s) (Degrees) (kW/m2 ) (kPa)
MU1 13 June 2018 9:00 AM 22.5 42.8 4.4 320 0.73 101.1
MU1 13 June 2018 12:00 PM 25.3 33.3 3.4 321 0.89 101.2
MU1 13 June 2018 3:00 PM 27.5 20.1 3.0 328 0.80 101.2
MU1 13 June 2018 5:00 PM 27.9 19.8 2.0 285 0.51 101.2
MU2 17 July 2018 9:00 AM 23.8 37.7 3.1 8 0.52 101.6
MU2 17 July 2018 12:00 PM 25 38.9 3.0 38 0.74 101.6
MU2 17 July 2018 3:00 PM 25 41.2 3.0 38 0.76 101.7
MU2 17 July 2018 5:00 PM 22.8 57.9 3.4 34 0.50 101.7
MU3 10 August 2018 9:00 AM 27.4 46.1 2.4 33 0.70 101.3
MU3 10 August 2018 12:00 PM 25.7 58.8 2.3 84 0.77 101.3
MU3 10 August 2018 3:00 PM 27.4 46.1 2.4 33 0.70 101.3
MU3 10 August 2018 5:00 PM 27.4 33.4 2.4 37 0.43 101.2
UTEP1 20 May 2018 9:00 AM 25 32 5.8 90 0.66 101.8
UTEP1 20 May 2018 12:00 PM 27.8 26 5.8 120 0.96 101.7
UTEP1 20 May 2018 3:00 PM 31.1 17 4.0 120 0.83 101.4
UTEP1 20 May 2018 5:00 PM 31.1 21 4.9 90 0.50 101.3
Remote Sens. 2019, 11, 1722 5 of 18

2.4. Thermal Data Processing

After data collection in the field, a series of post-processing steps were performed using Pix4D
and ESRI’s ArcMap to stitch the drone thermal imagery into orthomosaics, correct temperature values
for emissivity and extract surface temperature data for analysis. First, Pix4D was used to stitch the
captured thermal images into orthomosaics, export the orthomosaics as a 32-bit TIFF and georeference
them for application within ArcMap.
Once in ArcMap, an emissivity correction was applied to each thermal orthomosaic. Emissivity is
a measure of how well a material can emit energy as thermal radiation and different materials have
different values of emissivity depending on their surface properties [24]. Land use classifications that
were previously delineated for each case study area were used to apply emissivity values to the target
surfaces. The emissivity values for each land use classification used in this study are listed in Table 4
and are based upon a review of emissivity studies. These emissivity values were then applied in the
following emissivity correction equation derived from Stefan-Boltzmann Law:
s
4
4
Tsensor − (1 − ε) ∗ Tbackground
4
Ttarget = (1)
ε

where Ttarget is the actual temperature of the target surface [K], Tsensor is the temperature measured by
the infrared camera [K], Tbackground is the recorded air temperature [K] and ε is the emissivity value
of the target surface [25]. This equation was used to correct each surface type for their respective
emissivity before performing spatial data analysis.

Table 4. Emissivity values for each surface type.

Land Use Type Emissivity Value Reference

Grass 0.979 [26]
Shrub/mulch 0.928 [27]
Road (asphalt) 0.95 [28,29]
Parking Lot (concrete) 0.91 [29–31]
Sidewalk (concrete) 0.91 [24,29–31]
Rooftop (tar and stone) 0.973 [24]
Rooftop (black rubber) 0.859 [24]
Solar Panel 0.85 [32]
Canopy Cover 0.977 [33]

Once the thermal data were corrected for emissivity, spatial data analysis was performed in
ArcMap. First, a land use feature map was created that categorized the surface types in each case
study location. Then inconsistencies within these areas, such as a parked car within a parking lot,
human traffic on a sidewalk or construction materials on the street, were clipped and removed for each
flight. Once these inconsistencies were removed, zonal statistics was applied to compute summary
statistics of each surface type such as mean and standard deviation of the land surface temperature.
A complete flow-chart of the process from flight programming to developing summary statistics is
shown in Figure 2. In total this process took about 3 h to complete for each flight.

Figure 2. Flow chart of data collection and processing.

Remote Sens. 2019, 11, 1722 6 of 18

2.5. Surface Parameters

In addition to surface temperature, three other material properties were derived from visual and
multispectral imagery, converted into spatial distribution rasters and averaged for each surface type.
These include albedo (S), normalized difference vegetation index (NDVI) and apparent thermal inertia
(ATI). Albedo, a measure of solar reflectance of a material, was derived from blue, green, red and
near-IR image bands as shown in the following equation:

S = cb bk + c g gk + cr rk + ci ik (2)

where cb = 0.17, cg = −0.13, cr = 0.33 and ci = 0.54 are derived constants and bk , gk , rk and ik are the band
reflectance’s for—blue, bk (420–492 nm); green, gk (533–587 nm); red, rk (604–664 nm); and near-IR, ik
(833–920 nm) [34].
Visual and multispectral imagery were also used to derive NDVI, which is a measure of the
degree of live vegetation and is commonly used to evaluate soil moisture dynamics, erosion potential
and plant and crop health. As shown in Equation (3), NDVI is a function of near-IR and red band
reflectance and is estimated on a scale of −1 to +1, with higher values indicating higher vegetative
cover and greater plant health [35].

(NIR − Red)
NDVI = (3)
(NIR + Red)

Finally, ATI was derived for each surface type from albedo (S), solar correction (SCR) and the
diurnal temperature amplitude (DTA) (Equation (4)). ATI is an estimation of thermal inertia from
remotely sensed observations and can be estimated from diurnal changes in temperature. Specifically,
ATI is derived from solar correction (SCR), albedo (S) and the diurnal temperature amplitude (DTA),
where DTA is the difference between the maximum and minimum surface temperature recorded at the
time the remote images were captured and SCR is the solar correction factor (Equation (5)), which is
dependent on geographic location, the local latitude (θ) and the solar declination (ϕ) [36].

SCR(1 − S)
ATI = (4)
DTA
 
SCR = sinθ sinϕ(1 − tanθ tanϕ)2 + cosθ cosϕ arccos(−tanθ tanϕ) (5)

2.6. Model Development

Drone observations were applied to develop empirical models of land surface temperature.
These include (1) a regression model to predict spatially averaged surface temperatures at 12:00 PM
based upon meteorological variables and (2) a model to assess diurnal variability and predict surface
temperatures throughout a given day.
Multi-variable regression models were developed to predict spatially averaged surface temperature
of the fourteen Milwaukee and single El Paso 12:00 PM flights using MATLAB and the statistical software
package JMP 13 [37]. Response screening was performed for each of the respective datasets to identify
the strength of relationship between surface temperature (response) and meteorological parameters
(predictors). Between the two case study locations, six surface types that were common to both locations
were used as response variables: grass, canopy cover, concrete parking lot, concrete sidewalk, composite
rooftop and road surface. Meteorological predictor variables included air temperature, relative humidity,
preceding 24 h rainfall, wind speed, wind direction, barometric pressure and solar radiation. After
response screening, stepwise linear regression was then performed to predict land surface temperature
based upon meteorological parameters as represented in following equation:

y = β0 + β1 x1 + β2 x2 + · · · + βk xk (6)
Remote Sens. 2019, 11, 1722 7 of 18

where y is the response variable, β0 , β1 , . . . , βk are the regression coefficients and x1 , x2 , . . . , xk are
the predictor variables for k predictors [38]. These models were developed using data from both the
Milwaukee and El Paso flights; therefore, to evaluate the influence and leverage of the El Paso dataset
we computed Cook’s D influence and hat matrix leverage statistics [38].
Finally, we explored the variation in surface temperatures as they change throughout the day (9
AM, 12 PM, 3 PM and 5 PM) and evaluated if this variation could be explained by any meteorological
parameters. In addition to exploring diurnal changes in variability, we applied the data to develop a
model to predict land surface temperatures throughout the day for the six land use types common
to each location. To do so, we applied the drone data collected on the four diurnal flight missions to
estimate land surface temperatures based upon the solar radiation and the difference between the air
and land surface temperatures, which have been found to be statistically significant predictors for
diurnal estimates of pavement temperatures [39].
First, we computed a parameter (g) based upon the drone-derived mean land surface temperature
and measured air temperature and solar radiation:
 
g = Ts − Ta ∗ S (7)

where Ts is the mean surface temperature of the land use, Ta is the measured air temperature and S
is the measured solar radiation (kW). Next, g at a given hour i was estimated using a Gaussian peak
model given by the following:
i−b )2
gi = a ∗ e−0.5∗( c (8)

where gi , is the parameter g at hour i, a is the peak value, b is the critical point and c is the growth
rate [40]. Using this model, the mean land surface temperature can be predicted based upon air
temperature and solar radiation for any time of day using the following:

Ts,i = Ta,i + ( gi /Si ) (9)

where Ts,i is the estimated surface temperature at hour i and Ta,i and Si are the air temperature and solar
radiation at hour i. Taken as a whole, these models test both the suitability of predicting drone-derived
mean land surface temperatures based upon meteorological variables, as well as the generalizability of
our findings by including data from sites in two different geomorphologic and climatic regions.

3. Results

3.1. Surface Temperature Variability

We evaluated the land surface temperature variability of each flight across common land use
types and generally found that green surfaces had a greater degree of variability than gray surfaces,
with the exception being the rubber rooftop. As an example, the distribution of surface temperature
data (1,986,543 total data points) is shown in Figure 3 for a flight recorded on July 11, 2018. The six gray
surfaces recorded a smaller distribution of temperature on average but had more extreme values than
green surfaces (Figure 3a). Gray surfaces retain more heat from the sun because of their high emissivity
and ATI and therefore typically have higher surface temperatures. Additionally, non-normal behavior
was identified for both canopy cover and rubber rooftop (Figure 3b). Canopy cover exhibits a left skew
while the rubber rooftop exhibits a right skew. The canopy cover had a variation of tree types and
therefore a variation of leaf area indices (LAI), which may be a reason for the skew in the temperature
data. The rubber rooftop also exhibited a strong right skew, which may be due to small materials on
the roof surface, such as ventilation pipes and drainage grates, that were difficult to detect and may
not have been removed from the dataset. Therefore, this caused a distribution of lower temperatures
to be recorded.
exhibits a left skew while the rubber rooftop exhibits a right skew. The canopy cover had a variation
of tree types and therefore a variation of leaf area indices (LAI), which may be a reason for the skew
in the temperature data. The rubber rooftop also exhibited a strong right skew, which may be due to
small materials on the roof surface, such as ventilation pipes and drainage grates, that were difficult
to detect and may not have been removed from the dataset. Therefore, this caused a distribution of
Remote Sens. 2019, 11, 1722 8 of 18
lower temperatures to be recorded.

(a) (b)

Figure 3. Boxplot distribution of surface temperature (a); and histogram of surface temperature (b).
Figure 3. Boxplot
Data from distribution
flight recorded on of surface
July temperature
11, 2018. = and
Note GRS(a); SM = shrub/mulch;
histogram
grass; CPY = canopy;
of surface temperature (b). PL =
Data fromlot;
parking flight = sidewalk;
SWrecorded on July = 2018.
RTC11, Note GRS
composite = grass;
rooftop; = rubber
RTRSM = shrub/mulch; RD =
rooftop;CPY = canopy; PL== solar.
road; SLR
parking lot; SW = sidewalk; RTC = composite rooftop; RTR = rubber rooftop; RD = road; SLR = solar.
We then summarized the average temperature, standard deviation and coefficient of variation of
eachWe then type
surface summarized the average
for the fourteen temperature,
recorded flights standard deviation
in Milwaukee, WI and
and coefficient
the single of variation
flight in El Paso,
of
TXeach surface
(Table type for the
5). Generally, grayfourteen
surfaces recorded flights
exhibited in Milwaukee,
higher temperatures WIthroughout
and the single flightthan
the year in Elgreen
Paso,
surfaces. In El Paso, the asphalt parking lot exhibited the highest average temperature (51.7than
TX (Table 5). Generally, gray surfaces exhibited higher temperatures throughout the year ◦ C) and
green surfaces. In El Paso, the asphalt
◦ parking lot exhibited the highest average temperature
grass exhibited the lowest (41.6 C), while in Milwaukee the black rubber rooftop exhibited the highest (51.7 °C)
and grasstemperature
average exhibited the lowest
(57.4 (41.6canopy
◦ C) and °C), while in exhibited
cover Milwaukee thethe black(30.4
lowest rubber rooftop
◦ C). exhibited
In terms the the
of variation,
highest average temperature (57.4 °C) and canopy cover exhibited the lowest (30.4 °C). In terms of
lowest degrees of variation typically occurred in the parking lots and grass. However, there is a noted
variation, the lowest degrees of variation typically occurred in the parking lots and grass. However,
difference in the variation between the two locations; the road in Milwaukee had the highest coefficient
of variation of 0.32, while the road in El Paso had the lowest at 0.04. This may be due to a difference in
traffic on the days that flights were conducted. The location in Milwaukee is located near the city center
and is subject to heavy and constant vehicular traffic, while the location in El Paso is in a restricted
traffic area and experienced very low vehicle activity on the weekend that the flight was conducted.

Table 5. Average temperature, standard deviation and coefficient of variation of nine surface types From
14 recorded flights on Marquette University campus and from one flight recorded on UTEP’s campus.

Location Surface Type Temp (◦ C) Standard Dev (◦ C) Coeff. of Variation

Grass 34.7 7.9 0.15
Shrub/mulch 40.7 6.2 0.12
Canopy 30.4 7.2 0.16
Parking Lot 38.7 4.2 0.08
MU Sidewalk 36.3 11.2 0.21
Rooftop—Composite 47.6 6.2 0.1
Rooftop—Rubber 57.4 15.8 0.22
Road 32.5 15.4 0.32
Solar Panels 47.0 7.2 0.13
Grass 41.6 6.1 0.1
Canopy 46.6 6.3 10
Desert Shrub 46.2 6.7 0.11
Parking Lot
UTEP (asphalt) 51.7 4.9 0.07
Parking (concrete) 45.3 7.4 0.12
Sidewalk 43.1 8.2 0.13
Rooftop—Composite 47.3 7.2 0.11
Rooftop—Dzong 46.2 6.3 0.1
Road 48.4 2.8 0.04
Remote Sens. 2019, 11, 1722 9 of 18

The distribution of the average temperature, standard deviation and coefficient of variation for the
fourteen Milwaukee, WI flights is further illustrated in Figure 4. The shrub/mulch, composite rooftop
and solar panels have the most consistent variability among the land use types as shown in the boxplot
distribution of their coefficient of variation, while the greatest spread in variation occurred in the road
and sidewalk. This may indicate that areas that are not subject to human traffic (e.g., shrub/mulch
flower beds, rooftops and solar panels) have more consistent variability in their temperatures, while
other areas that are subject to intermittent human traffic (e.g., roads and sidewalks) have inconsistent
Remote Sens. 2019, 11, x FOR PEER REVIEW 10 of 19
temperature variabilities. We also evaluated if the variability in land surface temperature correlated
with any meteorological parameters but found no statistically significant predictors.

Average Temperature

60

40

20

Grass Shrub/mulch Canopy Parking Lot Sidewalk RT Composite RT Rubber Road Solar

Standard Deviation

30

20

10

0
Grass Shrub/mulch Canopy Parking Lot Sidewalk RT Composite RT Rubber Road Solar

Coefficient of Variation
0.8

0.6

0.4

0.2

Grass Shrub/mulch Canopy Parking Lot Sidewalk RT Composite RT Rubber Road Solar

Figure 4. Boxplot distribution of average temperature, standard deviation and coefficient of variation
from 14 recorded flights in Milwaukee, WI.
Figure 4. Boxplot distribution of average temperature, standard deviation and coefficient of variation
from 14 recorded
We evaluated flights in
the variation in Milwaukee, WI.
surface temperatures throughout the day and found that the highest
degree of variation occurred at noon. This is demonstrated in the Figure 5, which shows box plots of
We evaluated the variation in surface temperatures throughout the day and found that the
the standard deviation for six land use types: grass, canopy, parking lot, sidewalk, composite roof and
highest degree of variation occurred at noon. This is demonstrated in the Figure 5, which shows box
road for data from both MU and UTEP. As illustrated, all land use types have the greatest standard
plots of the standard deviation for six land use types: grass, canopy, parking lot, sidewalk, composite
deviation
roofin
andtemperatures
road for dataduring
from both12:00
MUPM,
andwith
UTEP.lower levels of all
As illustrated, deviation
land usein the have
types morning and late
the greatest
afternoon. This trend suggests that as surfaces heat up, they do so at different rates, which contributes
standard deviation in temperatures during 12:00 PM, with lower levels of deviation in the morning
to moreand
variability during
late afternoon. mid-day.
This trend suggests that as surfaces heat up, they do so at different rates, which
contributes to more variability during mid-day.
3.2. Impact of the Built Environment
We also evaluated
Grass
the spatial
Canopy
distribution of surface
Parking Lot
temperature
Sidewalk
to locate and
Rooftop
identify factors
Road
of the built20environment that contribute to temperature variability. Figure 5 illustrates the spatial
distribution of surface temperatures for a flight on July 8th , 2018. One factor of variability is the
reflectance and shaded cover from nearby buildings. For example, sidewalks in close proximity to
Standard Deviation (°C)

Engineering 15 Hall exhibited higher temperatures, most likely due to the sun’s reflectance off its glass

paneling. Two similarly sized sidewalk areas were compared and results show the average temperature
was 4.7 ◦ C hotter for the location closer to the building than the one farther away. In comparison
10
to the sidewalk, parking lot land uses had more consistent variability, perhaps because there were
fewer nearby buildings or large trees to exacerbate (glass reflectance) or reduce (shaded cover) their
temperature.5 This indicates proximity to nearby buildings or other structures can be a significant factor
of uncertainty in predicting surface temperatures.

9 12 15 17 9 12 15 17 9 12 15 17 9 12 15 17 9 12 15 17 9 12 15 17
Time (hr)
 highest degree of variation occurred at noon. This is demonstrated in the Figure 5, which shows box
plots of the standard deviation for six land use types: grass, canopy, parking lot, sidewalk, composite
roof and road for data from both MU and UTEP. As illustrated, all land use types have the greatest
standard deviation in temperatures during 12:00 PM, with lower levels of deviation in the morning
and late afternoon. This trend suggests that as surfaces heat up, they do so at different rates, which
Remote Sens. 2019, 11, 1722 10 of 18
contributes to more variability during mid-day.

Grass Canopy Parking Lot Sidewalk Rooftop Road

20

Standard Deviation (°C)

15

Remote Sens. 2019, 11, x FOR PEER REVIEW 11 of 19

10
Figure 5. Standard deviation distributions for six land use types at hour 9, 12, 15 and 17 at both the
MU and UTEP locations.

5
3.2 Impact of the Built Environment
We also evaluated the spatial distribution of surface temperature to locate and identify factors
of the built environment that contribute to temperature variability. Figure 5 illustrates the spatial
0
distribution of surface temperatures for a flight on July 8th, 2018. One factor of variability is the
reflectance9 and shaded
12 15 17 9cover
12 from
15 17nearby
9 12 buildings.
15 17 For 9 12 example,
15 17 sidewalks
9 12 15in 17 close9 proximity
12 15 17to
Engineering Hall exhibited higher temperatures, most likely due to the sun’s reflectance off its glass
Time (hr)
paneling. Two similarly sized sidewalk areas were compared and results show the average
Figure 5. Standard
temperature was deviation distributions
4.7 °C hotter for six
for the location land
closer to use types atthan
the building hourthe 9, 12,
one 15 and 17
farther at both
away. In the
MUcomparison
and UTEP to locations.
the sidewalk, parking lot land uses had more consistent variability, perhaps because
there were fewer nearby buildings or large trees to exacerbate (glass reflectance) or reduce (shaded
Other sources
cover) of land surface
their temperature. temperature
This indicates proximityuncertainty are traffic
to nearby buildings andstructures
or other parked cars.can beTraffic
a flow
along a roadway
significant intermittently blocks
factor of uncertainty the sunssurface
in predicting radiation, thereby impacting the surface temperatures
temperatures.
of the roadway Otherpavement
sources of land surface
below. temperature
This creates auncertainty
concentrated are traffic
pocket andofparked
cooler cars. Traffic flow
surface temperatures
along a roadway intermittently blocks the suns radiation, thereby impacting the surface temperatures
called a heat shadow, which results in variations in surface temperatures across the pavement. This is
of the roadway pavement below. This creates a concentrated pocket of cooler surface temperatures
especially pronounced
called in pavement
a heat shadow, lots
which results in with parked
variations cars as
in surface illustratedacross
temperatures in Figure 6b, which
the pavement. Thisshows the
distribution
is especially pronounced in pavement lots with parked cars as illustrated in Figure 6b, which shows pavement
of surface temperatures within a parking lot. In this figure a parked car rooftop,
the distribution
surface and heat shadow of surface temperatures
recorded withinof
temperatures 69.6 ◦ C,lot.
a parking In ◦this
47.8 C andfigure
49.0 ◦ C, respectively,
a parked car rooftop, all within
pavement
a space of ~50 m . 2 surface and heat shadow recorded temperatures of 69.6 °C, 47.8 °C and 49.0 °C,
respectively, all within a space of ~ 50 m2.

(a)

(b)

High: 70 °C

Low: 2.8 °C

High: 89 °C

Low: -2.8 °C

Spatial
Figure 6.Figure distribution
6. Spatial of temperature
distribution froma aflight
of temperature from flight recorded
recorded on 11on 112018
July July
(a)2018 (a) andinzoomed in
and zoomed
spatial distribution of temperature
spatial distribution of temperature for
forthe
the concrete parking
concrete parking lot from
lot from a flight
a flight recorded
recorded on 11 Julyon 11 July 2018
2018
(b). The (b). The hotter
hotter surfaces
surfaces (red) (red) in the
in the rightimage
right image are
areparked
parked carscars
and and
the cooler surfacessurfaces
the cooler (blue) are(blue)
heat are heat
shadows visible after parked cars leave.
shadows visible after parked cars leave.
Remote Sens. 2019, 11, 1722 11 of 18
Remote Sens. 2019, 11, x FOR PEER REVIEW 12 of 19

3.3
3.3.Surface Properties
Surface and Land
Properties Surface
and Land Temperature
Surface Temperature
Drone data was applied to derive surface properties including albedo, NDVI and ATI, of the
Drone data was applied to derive surface properties including albedo, NDVI and ATI, of the
surface types in the case study (Table 6). The light concrete parking lot exhibited the highest albedo
surface types in the case study (Table 6). The light concrete parking lot exhibited the highest albedo
(0.673) while grass exhibited the lowest (0.317). The spatial distribution of temperature, albedo, NDVI
(0.673)
and ATI while grass exhibited
at the Milwaukee, thestudy
WI case lowest (0.317).
location The spatial
is shown distribution
in Figure of temperature,
7. As illustrated, albedo, NDVI
these surface
and ATIproperties
material at the Milwaukee,
have a largeWI caseofstudy
degree location
variation across is
theshown in Figure
case study area. 7. As illustrated, these surface
material properties have a large degree of variation across the case study area.
Table 6. Average albedo, normalized difference vegetation index (NDVI) and
apparent
Table 6. Average thermal
albedo, inertia (ATI)
normalized values forvegetation
difference each surface type.(NDVI) and apparent thermal inertia
index
(ATI) valuesSurface
for eachType
surface type. Albedo NDVI ATI
Grass 0.317 0.369 0.198
Surface Type
Shrub/mulch Albedo
0.502 0.402 NDVI0.183 ATI
Canopy
Grass 0.378
0.317 0.490 0.369 0.209 0.198
Shrub/mulch
Parking Lot 0.673
0.502 0.091 0.402 0.121 0.183
Canopy
Sidewalk 0.378
0.472 0.144 0.490 0.195 0.209
Parking Lot
Rooftop – Composite 0.673
0.580 0.101 0.091 0.156 0.121
Sidewalk
Rooftop – Rubber 0.472
0.406 0.096 0.144 0.219 0.195
Rooftop–Composite
Road 0.580
0.518 0.117 0.101 0.179 0.156
Rooftop–Rubber
Solar 0.406
0.333 0.143 0.096 0.217 0.219
Road 0.518 0.117 0.179
Solar 0.333 0.143 0.217

(a) (b) (c) (d)

High: 89 °C
High: 0.91 High: 1 High: 0.22
Low: -2.8°C
Low: 0 Low: -1 Low: 0.02

Figure
Figure 7. Spatial
7. Spatial distribution
distribution of tempearture
of tempearture (a),(b),
(a), albedo albedo
NDVI(b), NDVI
(c) and ATI(c)
(d)and
for aATI (d)recorded
flight for a flight recorded
on 11 August
on 11 August 2018.2018.

Tofurther
To further explore
explore thisthis variability
variability and assess
and assess its on
its impact impact
surfaceontemperatures,
surface temperatures, we plotted these
we plotted these
surface
surfaceproperties
properties against
againstland surface
land temperature.
surface temperature.FigureFigure
8 illustrates temperature
8 illustrates plotted against
temperature plotted against its
its respectivealbedo
respective albedo for
forthe
the 611,460
611,460totaltotal
datadata
pointspoints
captured by the drone
captured imagery
by the droneand results show
imagery and results show
clusters that form for different surface types. Some of these clusters exhibit either a (1) low range in
clusters that form for different surface types. Some of these clusters exhibit either a (1) low range
albedo and high range in temperature or (2) high range in albedo and low range in temperature. For
in albedo and high range in temperature or (2) high range in albedo and low range in temperature.
example, the road exhibits a low range in albedo and high range in temperature, implying the
For example,
variability the road
in roadway exhibits aare
temperatures low range
more in albedo
dependent and high range
on meteorological (e.g.inexposure
temperature,
to solarimplying the
variability in roadway temperatures are more dependent on meteorological
radiation) and human (e.g. traffic) variables than physical properties (e.g. albedo). On the other hand, (e.g., exposure to solar
radiation)
the parking and human
lot has a higher(e.g.,
buttraffic)
similar variables than physical
range in albedo, yet it hasproperties
a much lower (e.g.,variability
albedo). inOn the other
temperature. This could
hand, the parking lotbe
has due to the fact
a higher butthat the parking
similar rangelotin has a range
albedo, yetofitmaterials
has a much from lower
asphaltvariability in
to concrete coupled
temperature. Thiswith
coulda much
be duelower levelfact
to the of traffic as compared
that the parking lot to the
hasroadway,
a range of which is morefrom asphalt
materials
to concrete coupled with a much lower level of traffic as compared to the roadway, which is more
homogenous and experiences constant vehicular traffic that intercepts land surface exposure to solar
radiation. Therefore, this graphic may support the previous statement that there are anthropogenic
variables, such as intermittent human foot or vehicular traffic, that are significant to land surface
 Remote Sens. 2019, 11, x FOR PEER REVIEW 13 of 19

homogenous and experiences constant vehicular traffic that intercepts land surface exposure to solar
Remote Sens. 2019, 11,Therefore,
radiation. 1722 12 of 18
this graphic may support the previous statement that there are anthropogenic
variables, such as intermittent human foot or vehicular traffic, that are significant to land surface
temperature processes. Overall these results suggest that patterns in the physical properties of urban
temperature processes. Overall these results suggest that patterns in the physical properties of urban
materials may provide insight into surface temperature variability.
materials may provide insight into surface temperature variability.

0.9
Canopy Cover
Shrub/mulch
0.8
Parking Lot
Sidewalk
0.7 Road
RT Composite
RT Rubber
0.6
Grass
Albedo

0.5

0.4

0.3

0.2

0.1
30 40 50 60 70 80
Temperature (°C)

Figure 8. Surface temperature data plotted against albedo from a flight recorded on 11 August 2018.
Figure 8. Surface temperature data plotted against albedo from a flight recorded on 11 August 2018.
3.4. Temperature Prediction Models
3.4. Temperature Prediction Models
Drone observations were applied to develop empirical models of land surface temperature. These
Drone observations were applied to develop empirical models of land surface temperature.
include (1) a regression model to predict spatially averaged surface temperatures at 12:00 PM based
These include (1) a regression model to predict spatially averaged surface temperatures at 12:00 PM
upon environmental variables and
based upon environmental (2) a diurnal
variables and (2)model to predict
a diurnal modelsurface temperatures
to predict throughout a
surface temperatures
given day.
throughout a given day.

3.4.1. Spatially Averaged

3.4.1 Spatially Surface
Averaged Temperature
Surface Regression
Temperature RegressionModel
Model
Multi-variable linearlinear
Multi-variable regression
regressionmodels
modelswere
weredeveloped
developed to predict
predictspatially
spatially averaged
averaged surface
surface
temperature
temperature and found
and it was it was that
found
air that air temperature
temperature and solar andradiation
solar radiation are significant
are significant predictors
predictors (Figure 9).
Standard(Figure
least9). Standard
squares least squares
regression wasregression
applied towas appliedmodels
develop to develop
thatmodels
predictthat
thepredict
surfacethe surface
temperature
temperature of six land use types: grass, canopy cover, parking lot (concrete), sidewalk, rooftop
of six land use types: grass, canopy cover, parking lot2 (concrete), sidewalk, rooftop (composite) and
(composite) and road. The models had an average R of 0.71 with the parking lot having the greatest
road. The models had an average R2 of 0.71 with the parking lot having the greatest of (0.89) and
of (0.89) and the road the lowest (0.37). The parked cars and heat shadows were clipped out as
the road the lowest (0.37).
inconsistencies before The parked
analysis cars and
occurred and heat shadows
therefore were clipped
the parking out as
lot surface inconsistencies
had the most
before analysis
homogenous distribution of temperatures. The grass model had the second greatest R (0.84)distribution
occurred and therefore the parking lot surface had the most homogenous
2 and had
of temperatures. The grass model
a similarly homogenous had theContrarily,
distribution. second greatest R2 (0.84)
the roadway surface and
hadhad a similarly
a much homogenous
less homogenous
distribution. Contrarily,
distribution the roadway
of temperatures andsurface
thus the had a much
road modelless
hadhomogenous distribution
a low predictive power and of temperatures
statistical
and thussignificance. This may
the road model hadbe due in
a low large partpower
predictive due to and
the difficulty
statisticalofsignificance.
clipping out inconsistencies
This may be due related
in large
to nonstationary objects (e.g. moving cars) combined with their impact on pavement
part due to the difficulty of clipping out inconsistencies related to nonstationary objects (e.g., moving temperatures.
cars) combined with their impact on pavement temperatures.
The data collected in El Paso, TX was evaluated for influence and leverage and it was found that it
did not have high influence or leverage in any of the six models. To evaluate influence we used Cook’s
D and found that the El Paso data points all fell below the threshold of 2.4 (max 0.19) to be considered
high-influence points [38]. In addition, we used the hat matrix to evaluate leverage and found that no
El Paso data points exhibited high leverage in the model. The agreeability of the data across the two
case study areas indicates that the findings in this study may have generalizability beyond the case
study locations.
 Cook’s D and found that the El Paso data points all fell below the threshold of 2.4 (max 0.19) to be
considered high-influence points [38]. In addition, we used the hat matrix to evaluate leverage and
found that no El Paso data points exhibited high leverage in the model. The agreeability of the data
across the two case study areas indicates that the findings in this study may have generalizability
beyond
Remote the
Sens. case
2019, 11,study
1722 locations. 13 of 18

(a) (b)
17.73+0.64*AT+8.29*SR

Recorded Temperature (°C)

Recorded Temperature (°C)

11.71+0.40*AT+15.85*SR

R2=0.84
R2=0.68

p-value<0.0001
p-value=0.0011

(c)
(d)
Recorded Temperature (°C)

Recorded Temperature (°C)

8.81+1.17*AT+4.31*SR 14.87+0.61*AT+11.53*SR

R2=0.89 R2=0.70

p-value<0.0001 p-value=0.0007

(e) (f)
Recorded Temperature (°C)

Recorded Temperature (°C)

20.6+1.13*AT+0.67*SR 1.85+0.62*AT+8.05*SR

R2=0.78 R2=0.37

p-value=0.0001 p-value=0.06

Figure 9. Temperature prediction models of six surface types: grass (a), canopy cover (b), parking lot
(c), sidewalk
Figure (d), composite
9. Temperature rooftop
prediction (e) and
models road
of six (f). UTEP
surface types: datapoint is fitted
grass (a), canopy in green.
cover Note the
(b), parking lot 95%
confidence
(c), sidewalk intervals are in blue.
(d), composite rooftop (e) and road (f). UTEP datapoint is fitted in green. Note the 95%
confidence intervals are in blue.
3.4.2. Diurnal Prediction Model
3.4.2Finally,
Diurnalmodels
Prediction
were Model
developed to predict land surface temperature throughout the day based
uponFinally,
the air models
temperature and solar to
were developed radiation (Equations
predict land (7)–(9)). Thethroughout
surface temperature diurnal data
the was fit with a
day based
Gaussian peak
upon the air distribution
temperature andand it radiation
solar was found that the parking
(Equations 7-9). The lot and composite
diurnal data was fitrooftop had the best
with a Gaussian
model fit with anand
peak distribution R2 it
ofwas
0.83found
and 0.78, respectively,
that the parking lotwhile all other rooftop
and composite modelshadhadthe R2 value
anbest model offit 0.53
or below
with an R(Figure
2 of 0.8310).
andWhile this approach
0.78, respectively, is constrained
while by a had
all other models limited
an Rnumber of 0.53
2 value of dataorpoints
belowfrom
four flights and only four numerical x-axis variables, there are a few insights we can gain from these
results. The first is that these models confirm what was found in the previous regression models: it
is much easier to predict the land surface temperature of homogenous materials, such as pavements
and rooftops, than it is to predict land surfaces that have a greater distribution in texture and material,
such as canopy. The second is that anthropogenic variables, such as pedestrians and vehicular traffic
that are difficult to quantify, may influence the ability to predict surface temperatures based upon
meteorological variables. This was shown by the lower model fit in the high-traffic roadways and
sidewalks as compared to the low-traffic parking lot.
than it is to predict land surfaces that have a greater distribution in texture and material, such as
canopy. The second is that anthropogenic variables, such as pedestrians and vehicular traffic that are
difficult to quantify, may influence the ability to predict surface temperatures based upon
meteorological variables. This was shown by the lower model fit in the high-traffic roadways and
sidewalks as compared to the low-traffic parking lot.
Remote Sens. 2019, 11, 1722 14 of 18

Grass Canopy Cover

15 20 ( . )
( . ) .
10.20𝑒 .
. 7.69𝑒 .

R2 = 0.33
R2 = 0.50 15
10
(GRS-AT)*SR

(CPY-AT)*SR
10
5
5

0
0
8 10 12 14 16 18 8 10 12 14 16 18
Time (hr) Time (hr)

Parking Lot Sidewalk

20 15
( . ) ( . )
. .
16.26𝑒 . 13.18𝑒 .

R2 = 0.83 R2 = 0.53
15

(SW-AT)*SR
10
(PL-AT)*SR

10
5

5
0
8 10 12 14 16 18 8 10 12 14 16 18
Time (hr) Time (hr)

RT Composite Road
25 20
( . ) ( . )
. .
20.61𝑒 . 17.12𝑒 .

20 R2 = 0.78
10 R2 = 0.41
(RTC-AT)*SR

(RD-AT)*SR

15
0
10
-10
5

0 -20
8 10 12 14 16 18 8 10 12 14 16 18
Time (hr) Time (hr)

Figure 10. Gaussian peak models of six surface types: grass (a), canopy cover (b), parking lot (c),
sidewalk
Figure 10. (d), composite
Gaussian peakrooftop
models (e)
of and road (f).
six surface Notegrass
types: that GRS = grass;cover
(a), canopy = canopy;
CPY (b), parkingPLlot=(c),
parking
lot; SW = sidewalk; RTC = composite rooftop; AT = air temperature; SR = solar radiation; t
sidewalk (d), composite rooftop (e) and road (f). Note that GRS = grass; CPY = canopy; PL = parking= time.
lot; SW = sidewalk; RTC = composite rooftop; AT = air temperature; SR = solar radiation; t = time.
4. Discussion
4. Discussion
We have presented a case study that applied high resolution drone measurements (13 cm) to
evaluate urban surface temperatures and results indicate that there is a wide variability in surface
temperature behavior across urban land use types. Some of the uncertainty in land surface temperature
variability may be attributable to human movement patterns, land surface properties or urban geometry.
Results indicate that mean land surface temperatures can be predicted based upon solar radiation and
air temperature. By elucidating some of the factors that influence land surface temperature variability,
we hope to contribute to the growing body of knowledge centered around land surface temperature in
the urban environment.
To this end, our findings suggest that when parameterizing models, it is important to understand
the unique relationship between surface material properties, urban geometry, weather and human
movement. For example, the results indicate that pedestrian or vehicular traffic may have an impact
on land surface temperature variability across sidewalks, parking lots and streets. Depending on the
volume of cars, either parked or moving, this can greatly impact the temperature profile of paved
Remote Sens. 2019, 11, 1722 15 of 18

surfaces. Parked cars can create heat shadows which cool the surface below and our study demonstrates
that when a car moves it can reveal temperatures as low as 8.3 ◦ C cooler than the exposed surface.
In addition, results have identified several factors of urban geometry that affect land surface
temperatures. Urban factors such as building reflectivity and surface altitude can impact solar radiation,
which then influences surface temperatures in locations impacted by these effects. For example,
sidewalks often lie near buildings and depending on a buildings reflectance or shadows this can
make sidewalk temperatures more vulnerable to temperature fluctuations. In this study, sidewalk
temperatures impacted by glass reflectance were on average 4.7 ◦ C hotter that sidewalks not impacted
by reflectance. Therefore, knowledge of the spatial distribution of urban geometry is important for
predicting and evaluating land surface temperatures in the built environment. While addressing urban
geometry or pedestrian and vehicular traffic within our prediction models is outside of the scope of
this project, future work should evaluate how to incorporate these important parameters into land
surface temperature predictions.
Results indicate air temperature and solar radiation are significant predictors of mean land surface
temperature in both of our models and it was found this relationship holds true in both Milwaukee,
WI and El Paso, TX. Because the model holds true across two different climatic regions, the models
developed in this project may be generalizable beyond their case study regions. In addition, these
models can also be easily applied as air temperature and solar radiation are commonly measured
across the world. The generalizability of these findings also has important implications for engineering
applications that use predictions of land surface temperatures. Urban land surface temperatures are
often used by public health officials to mitigate the impact of the urban heat island effect on human
health [2], in developing binders and mixers of pavement in roadway designs [41] or to estimate the
impact of land surface temperatures on receiving stream temperatures [42–44].
This study also demonstrates several advantages and disadvantages of using drones as compared
to satellite or in-situ imagery. The case studies we evaluated were restricted to the size of a city block
around 46,000 m2 and even though battery life would have allowed us to collect an area ten times this
size, we were restricted by United States Federal Aviation Administration UAV pilot rules that restrict
the flight of UAVs to within line of sight of the pilot. In an urban environment with tall buildings the
line of sight may be the primary constraint on coverage area. Therefore, a disadvantage of UAVs is
that flight time and legal restrictions may constrain the flight areas to small portions of a city. However,
this could be overcome with fixed-wing drones that are able cover a greater area, in addition to relaxed
regulations that allow flights beyond the line of sight [45]. Despite the restriction on the spatial extent
of the study area, advantages of UAVs over satellites or in-situ methods are their ability to collect
distributed temperature data at spatial resolutions (13 cm) that reflect small scale changes in the urban
environment. In addition, satellite data is restricted to daily to weekly observations while drones can
be flown on-demand, which allows them to capture temperature changes throughout the day.
Overall, this study highlights the utility of using drone observations to capture the variability of
urban land surface temperatures at small spatial scales. Urban environments are spatially complex,
making it difficult to capture the spatial distribution of observable phenomena outside of high-resolution
remote sensing techniques. Our findings suggest that drones could also be good tools for evaluating
the variability of other parameters of the urban environment that are important for environmental
studies such as soil moisture, leaf area index or impervious cover. Therefore, it is important for studies
such as this one that evaluate the spatial complexities of the urban environment in order to improve
the methods that we use to model and understand urban systems.

5. Conclusions
The main objectives of this work were to apply drone imagery to capture land surface temperature
variability and develop models to predict mean land surface temperatures. This was done through the
application of high-resolution thermal imagery as a parameterizing tool for model development. The
results revealed that land surface temperature variability is extensive and influenced by numerous
Remote Sens. 2019, 11, 1722 16 of 18

variables related to urban environments and that air temperature and solar radiation are significant
predictors of mean land surface temperature. Conclusions from this study hold true in both Milwaukee,
WI and El Paso, TX, indicating they could be generalizable to regions beyond these two case
study locations.
The key findings from this study were:

• Land surface variability was significant and ranged between (3.9–15.8 ◦ C) for common land
use types.
• Areas that experienced pedestrian or vehicular traffic exhibited higher variabilities than comparable
surfaces that did not. In Milwaukee, the high-traffic road had a coefficient of variation of 0.32 as
compared to 0.08 for the low-traffic parking lot. This indicates that human traffic may impact
land surface temperatures due to the heat-shadow effect.
• Urban geometry has an influence on land surface temperatures; shadows and reflectance from
buildings showed a significant influence on the temperatures of nearby land surfaces throughout
the day. Sidewalk temperatures impacted by glass reflectance were on average 4.7 ◦ C hotter that
sidewalks not impacted by reflectance.
• Land surface temperature variability is low in the morning, peaks at noon and goes back down
in the evening. This may indicate that as surfaces heat up, they do so at different rates, which
contributes to more variability during mid-day.
• Air temperature and solar radiation were significant predictors of spatially averaged surface
temperature in both of our models.
• Data were consistent in the models between Milwaukee, WI and El Paso, TX, suggesting that the
findings in this study may be generalizable beyond the case study locations.

Overall, our findings suggest that land surface temperature variability in the urban environment
can come from several sources including surface material properties, urban geometry, weather and
pedestrian and vehicular traffic. This has direct implications for land surface temperature models that
are used for urban environmental studies. As climate change and urbanization continue to exacerbate
the SUHI, studies such as this are important for gaining a better understanding of the complexities of
land surface temperatures. Ultimately this improved understanding will help to develop better methods
and procedures to mitigate the impact of land surface temperatures on human and environmental health.

Author Contributions: J.N. provided investigation, data collection, data analysis, and writing of the original
draft. W.M. contributed conceptualization, supervision, data collection and draft editing.
Funding: This project was funded by the Marquette OPUS College of Engineering Earl B. and Charlotte Nelson
Award.
Acknowledgments: The authors would like to acknowledge and sincerely thank Saurav Kumar and Wissam
Atwah at the University of Texas El Paso for their help in collecting data in El Paso, Tx.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Foley, J.A.; DeFries, R.; Asner, G.P.; Barford, C.; Bonan, G.; Carpenter, S.R.; Chapin, F.S.; Coe, M.T.; Daily, G.C.;
Gibbs, H.K.; et al. Global consequences of land use. Science 2005, 309, 570. [CrossRef] [PubMed]
2. Shahmohamadi, P.; Che-Ani, A.I.; Etessam, I.; Maulud, K.N.A.; Tawil, N.M. Healthy Environment: The Need
to Mitigate Urban Heat Island Effects on Human Health. Procedia Eng. 2011, 20, 61–70. [CrossRef]
3. U.S. Environmental Protection Agency. The Assessment, Total Maximum Daily Load (TMDL) Tracking and
Implementation System (ATTAINS); U.S. Environmental Protection Agency: Washington, DC, USA, 2019.
4. White, P.A.; Kalff, J.; Rasmussen, J.B.; Gasol, J.M. The effect of temperature and algal biomass on bacterial
production and specific growth rate in freshwater and marine habitats. Microb. Ecol. 1991, 21, 99–118.
[CrossRef] [PubMed]
5. Beitinger, T.L.; Bennett, W.A.; McCauley, R.W. Temperature Tolerances of North American Freshwater Fishes
Exposed to Dynamic Changes in Temperature. Environ. Biol. Fishes 2000, 58, 237–275. [CrossRef]
Remote Sens. 2019, 11, 1722 17 of 18

6. Lazzarini, M.; Marpu, P.R.; Ghedira, H. Temperature-land cover interactions: The inversion of urban heat
island phenomenon in desert city areas. Remote Sens. Environ. 2013, 130, 136–152. [CrossRef]
7. Jiang, Y.; Fu, P.; Weng, Q. Assessing the impacts of urbanization-associated land use/cover change on land
surface temperature and surface moisture: A case study in the midwestern united states. Remote Sens. 2015,
7, 4880–4898. [CrossRef]
8. Liu, L.; Zhang, Y. Urban heat island analysis using the landsat TM data and ASTER Data: A case study in
Hong Kong. Remote Sens. 2011, 3, 1535–1552. [CrossRef]
9. Sekertekin, A.; Kutoglu, S.H.; Kaya, S. Evaluation of spatio-temporal variability in Land Surface Temperature:
A case study of Zonguldak, Turkey. Environ. Monit. Assess. 2016, 188, 1–15. [CrossRef]
10. Sobrino, J.A.; Oltra-carrió, R.; Sòria, G.; Jiménez-muñoz, J.C.; Franch, B.; Hidalgo, V.; Mattar, C.; Julien, Y.;
Cuenca, J.; Romaguera, M.; et al. Evaluation of the surface urban heat island effect in the city of Madrid. Int.
J. Remote Sens. 2013, 34, 3177–3192. [CrossRef]
11. Mallick, J.; Kant, Y.; Bharath, B.D. Estimation of land surface temperature over Delhi using Landsat-7 ETM+.
J. Ind. Geophys. Union. 2008, 12, 131–140.
12. Zhou, D.; Xiao, J.; Bonafoni, S.; Berger, C.; Deilami, K.; Zhou, Y.; Frolking, S.; Yao, R.; Qiao, Z.; Sobrino, J.A.
Satellite Remote Sensing of Surface Urban Heat Islands: Progress, Challenges, and Perspectives. Remote Sens.
2018, 11, 48. [CrossRef]
13. Granero-Belinchon, C.; Michel, A.; Lagouarde, J.-P.; Sobrino, J.A.; Briottet, X. Multi-Resolution Study of
Thermal Unmixing Techniques over Madrid Urban Area: Case Study of TRISHNA Mission. Remote Sens.
2019, 11, 1251. [CrossRef]
14. Nguyen, K.A.; Liou, Y.A. Global mapping of eco-environmental vulnerability from human and nature
disturbances. Sci. Total Environ. 2019, 664, 995–1004. [CrossRef] [PubMed]
15. Deng, C.; Wu, C. Estimating very high resolution urban surface temperature using a spectral unmixing and
thermal mixing approach. Int. J. Appl. Earth Obs. Geoinf. 2013, 23, 155–164. [CrossRef]
16. Nichol, J.E.; Wong, M.S. High resolution remote sensing of densely urbanised regions: A case study of Hong
Kong. Sensors 2009, 9, 4695–4708. [CrossRef] [PubMed]
17. Cooper, B.E.; Dymond, R.L.; Shao, Y. Impervious Comparison of NLCD versus a Detailed Dataset Over Time.
Photogramm. Eng. Remote Sens. 2017, 83, 429–437. [CrossRef]
18. Li, J.; Song, C.; Cao, L.; Zhu, F.; Meng, X.; Wu, J. Impacts of landscape structure on surface urban heat islands:
A case study of Shanghai, China. Remote Sens. Environ. 2011, 115, 3249–3263. [CrossRef]
19. Zhou, W.; Huang, G.; Cadenasso, M.L. Does spatial configuration matter? Understanding the effects of
land cover pattern on land surface temperature in urban landscapes. Landsc. Urban Plan. 2011, 102, 54–63.
[CrossRef]
20. Chudnovsky, A.; Ben-Dor, E.; Saaroni, H. Diurnal thermal behavior of selected urban objects using remote
sensing measurements. Energy Build. 2004, 36, 1063–1074. [CrossRef]
21. Gaitani, N.; Burud, I.; Thiis, T.; Santamouris, M. High-resolution spectral mapping of urban thermal properties
with Unmanned Aerial Vehicles. Build. Environ. 2017, 121, 215–224. [CrossRef]
22. Kottek, M.; Grieser, J.; Beck, C.; Rudolf, B.; Rubel, F. World map of the Köppen-Geiger climate classification
updated. Meteorol. Z. 2006, 15, 259–263. [CrossRef]
23. National Weather Service Forcast Office. Available online: https://www.weather.gov/mkx/ (accessed on
17 July 2019).
24. Marshall, S.J. We Need To Know More About Infrared Emissivity. Proc. SPIE 1982, 313, 119–128.
25. Blonquist, J.M.; Norman, J.M.; Bugbee, B. Automated measurement of canopy stomatal conductance based
on infrared temperature. Agric. For. Meteorol. 2009, 149, 1931–1945. [CrossRef]
26. Humes, K.S.; Kustas, W.P.; Moran, M.S.; Nichols, W.D.; Weltz, M.A. Variability of emissivity and surface
temperature over a sparsely vegetated surface. Water Resour. Res. 1994, 30, 1299–1310. [CrossRef]
27. Wittich, K.P. Some simple relationships between land-surface emissivity, greenness and the plant cover
fraction for use in satellite remote sensing. Int. J. Biometeorol. 1997, 41, 58–64. [CrossRef]
28. Ramamurthy, P.; Bou-Zeid, E. Contribution of impervious surfaces to urban evaporation. Water Resour. Res.
2014, 50, 2889–2902. [CrossRef]
29. Chen, H.; Ooka, R.; Huang, H.; Tsuchiya, T. Study on mitigation measures for outdoor thermal environment
on present urban blocks in Tokyo using coupled simulation. Build. Environ. 2009, 44, 2290–2299. [CrossRef]
Remote Sens. 2019, 11, 1722 18 of 18

30. Larsson, O.; Thelandersson, S. Estimating extreme values of thermal gradients in concrete structures.
Mater. Struct. 2011, 44, 1491–1500. [CrossRef]
31. Salamanca, F.; Krpo, A.; Martilli, A.; Clappier, A. A new building energy model coupled with an urban
canopy parameterization for urban climate simulations—Part I. formulation, verification, and sensitivity
analysis of the model. Theor. Appl. Climatol. 2009, 99, 331. [CrossRef]
32. Herb, W.R.; Janke, B.; Mohseni, O.; Stefan, H.G. Thermal pollution of streams by runoff from paved surfaces.
Hydrol. Process. 2008, 22, 987–999. [CrossRef]
33. Spectrolab, Inc. Spectrolab Photovotaic Products Data Sheet; Spectrolab, Inc.: Sylmar, CA, USA, 2012.
34. Ban-Weiss, G.A.; Woods, J.; Levinson, R. Using remote sensing to quantify albedo of roofs in seven California
cities, Part 1: Methods. Sol. Energy 2015, 115, 777–790. [CrossRef]
35. Carlson, T.N.; Ripley, D.A. On the relation between NDVI, fractional vegetation cover, and leaf area index.
Remote Sens. Environ. 1997, 62, 241–252. [CrossRef]
36. Sohrabinia, M.; Rack, W.; Zawar-Reza, P. Soil moisture derived using two apparent thermal inertia functions
over Canterbury, New Zealand. J. Appl. Remote Sens. 2014, 8, 1–16.
37. SAS Institute. JMP®8 User Guide, 2nd ed.; SAS Institute Inc.: Cary, NC, USA, 2009.
38. Helsel, D.R.; Hirsch, R.M. Statistical Methods in Water Resources; Elsevier: Amsterdam, The Netherlands, 2002;
Volume 36.
39. Thompson, A.M.; Wilson, T.; Norman, J.M.; Gemechu, A.L.; Roa-Espinosa, A. Modeling the effect of summertime
heating on urban runoff temperature. J. Am. Water Resour. Assoc. 2008, 44, 1548–1563. [CrossRef]
40. Guo, H. A Simple Algorithm for Fitting a Gaussian Function [DSP Tips and Tricks]. IEEE Signal Process. Mag.
2011, 28, 134–137. [CrossRef]
41. Solaimanian, M.; Kennedy, T.W. Predicting Maximum Pavement Surface Temperature Using Maximum Air
Temperature and Hourly Solar Radiation; Transportation Research Board: Washington, DC, USA, 1993.
42. Herb, W.R.; Janke, B.; Mohseni, O.; Stefan, H.G. MINUHET (Minnesota Urban Heat Export Tool): A Software
Tool for the Analysis of Stream Thermal Loading by Stormwater Runoff ; Project Report No. 526; St. Anthony Falls
Laboratory, The University of Minnesota: Minneapolis, MN, USA, 2009.
43. Roa-Espinosa, A.; Wilson, T.B.; Norman, J.M.; Johnson, K. Predicting the Impact of Urban Development on
Stream Temperature Using a Thermal Urban Runoff Model. Methods 2003, 369–389.
44. Hailegeorgis, T.T.; Alfredsen, K. High spatial–temporal resolution and integrated surface and subsurface
precipitation-runoff modelling for a small stormwater catchment. J. Hydrol. 2018, 557, 613–630. [CrossRef]
45. Ravich, T. A Comparative Global Analysis of Drone Laws: Best Practices and Policies. In The Future of
Drone Use: Opportunities and Threats from Ethical and Legal Perspectives; Custers, B., Ed.; T.M.C. Asser Press:
The Hague, The Netherlands, 2016; pp. 301–322. ISBN 978-94-6265-132-6.

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).