**2.1 Sling psychrometer measurements**

Sling psychrometers are traditional meteorological measurement tools to determine air temperature (dry-bulb temperature) and relative humidity (wetbulb temperature). Sling psychrometers have been used in UHI studies going back several decades [32, 35, 36, 39]. They have an educational advantage over automated measurements as they require direct student involvement in the data gathering and documentation process. The instrument used in this study was a Bacharach model 0012-7012 using two red spirit filled glass thermometers with Fahrenheit scales. The instrument is made of a robust hard plastic shell, with its outer part acting as handle when extended, while the inner part bears the two identical glass thermometers. The thermometer scales allow readings as precise as 0.5°F, and are accurate to at least 1.0°F based on intercomparisons with other sling psychrometers and a research grade meteorological sensor, comparisons that were made part of the student project in this UHI study.

Here, we discuss only the dry bulb, aka air temperature data. The atmospheric UHI effect was calculated as the difference between the measured air temperature and the corresponding temperature at 2 m above ground level (agl) at the Texas A&M weather station (**Figure 1**). The weather station's combined T/RH sensor, a

#### **Figure 1.**

*Roadmap based view of the Bryan/College Station metro area in East Texas (30.6°N, 96.32 W). The major highway traversing the area, Texas-6, is labeled alongside the two stationary measurement locations in red (T = 'trails', Q = 'quad') and the weather station location (W) in blue.*

model 085 from MetOne Instruments Inc., has a high precision and accuracy multielement thermistor (±0.15°C). Corrections for elevation differences between the measurement sites and the weather station reference location were made using the dry adiabatic lapse rate.

Multiple locations were monitored within the BCS metro area during the spring semester of 2015, but only two urban locations, which were monitored throughout 2015, are discussed here. The first location, subsequently identified as the "Trails" apartment complex, was a parking lot area in front of apartment buildings (**Figure 1**) in College Station. This part of town is slightly elevated from the surrounding urban area, and located between a narrow, wooded green corridor, and a major freeway (State Highway 6). While the location was next to pervious lawns, the area was almost tree-less and dominated by buildings, roads, and parking lots. The tallest buildings are the three story apartment buildings; the larger area within half a kilometer, however, includes a midsize mall and associated parking lots to the north, commercial buildings, including hotels, along the freeway to the east, and both taller and less tall apartment buildings toward the south and west.

The second location, subsequently referred to as "Quad", was a small parking lot on the Texas A&M campus in College Station (**Figure 1**). This location is also dominated by impervious surface areas, such as more extensive parking lots to the southeast, large parking garages to the east, and the onsite multi-story dormitories and Dining Hall. It has, however, numerous trees lining the nearby streets. The surroundings within half a kilometer consist of numerous, multi-story university buildings to the north and west, an open park, a field, and a wooded, one-story residential neighborhood to the southwest and south, and a golf course further to the northeast.

Both locations can be characterized as local climate zones (LCZ) 56, open mid-rise to low-rise urban areas with 30–50% impervious area [8]. However, Trails is located closer to the east end of town, with rural areas (LCZ B/C) as close as 1.5 km to the east, while Quad is located more central to the metro area, with open, sparsely build-up LCZ 9 areas more than 2 km distant, and rural areas more than 4 km distant (**Figure 1**).

UHI intensity was determined for each location by comparing observed local temperatures to the rural temperature measurement (10-min average) at Texas A&M's weather station 10 km outside the urban area toward the southwest. In addition, we assembled weather station data (pressure, winds, solar radiation, and precipitation) for the days measurements were taken for data analysis.

#### **2.2 Mobile measurements**

Air temperature sensors and associated logging equipment have become both miniaturized and highly affordable, thus allowing both high spatial density distributed and low-key mobile data collection of the urban heat island [40]. Mobile measurements (traverses) using automobiles, but also bicycle-mounted or personally carried sensors, have been used numerous times in the past to study the UHI effect, both in long- and in short-term campaigns [32, 33, 37, 41–65].

In summer and fall 2015, we carried out an undergraduate student research project [38] using a standard HOBO U12 data logger with 12-bit temperature sensor (model TMC6-HD) from Onset Computer Inc. to evaluate the atmospheric urban heat island of the BCS metro area. The sensor has a typical accuracy of ±0.25°C and is recorded at a better than 0.2°C resolution. It was placed into a passive radiation shield (model RS3, Onset Comp. Inc.), which was mounted to an angular steel bracket attached to a metal sleeve with a flat rubber sheet that slides over a passenger car side window (**Figure 2**). With the window moved nearly all the way up, the

**29**

**Figure 2.**

*Mobile measurement setup.*

*Low-Key Stationary and Mobile Tools for Probing the Atmospheric UHI Effect*

sensor was measuring air temperature at approximately 2 m agl off the passenger side of the car, with its logger safely placed inside the air-conditioned vehicle. This avoided most temperature bias effects from the vehicle itself, or from other vehicles on the road, unless measurements occurred either at very low speeds with air moving across the front hood of the car toward the sensor, such as may have occurred when stopping at traffic lights; or when hot exhaust plumes from vehicles on the road were encountered. Such possible temperature measurement biases, though

Typical car moving speeds were 20–30 mph in town and up 50 mph outside

Raw temperature data were normalized first for the presumed linear temperature change, if present, during the typically one-hour drive for data collection, then adjusted for elevation differences along the route using the dry adiabatic lapse rate. Both corrections were always significantly smaller than the encountered tempera-

The routes driven were selected to avoid selective coverage of the metro area, and such that measurements along a route could be completed within approximately 1 h, including nearby rural areas. All routes were driven once in the

A smart phone app called RAAH (https://www.raah.co) was used to record the vehicle's location. Data logger and smart-phone times were aligned before each drive. During each traverse data were recorded every 10 s, providing for a typical horizontal resolution of 100+ m. However, considering the sensor's response time, and the accuracy and timing of the location determination, both the absolute bias and uncertainty of a recorded temperature's location were likely larger than 100 m. Such aspects should be considered in all mobile studies when moving speeds are

) during all mobile measurements.

occasionally encountered, were not removed from the data set.

ture differences between the rural and (sub-)urban areas traversed.

the urban areas (approximately 10–25 m s<sup>−</sup><sup>1</sup>

comparable or faster than sensor response times.

*DOI: http://dx.doi.org/10.5772/intechopen.89514*

*Low-Key Stationary and Mobile Tools for Probing the Atmospheric UHI Effect DOI: http://dx.doi.org/10.5772/intechopen.89514*

**Figure 2.** *Mobile measurement setup.*

*Spatial Variability in Environmental Science - Patterns, Processes, and Analyses*

both taller and less tall apartment buildings toward the south and west.

The second location, subsequently referred to as "Quad", was a small parking lot on the Texas A&M campus in College Station (**Figure 1**). This location is also dominated by impervious surface areas, such as more extensive parking lots to the southeast, large parking garages to the east, and the onsite multi-story dormitories and Dining Hall. It has, however, numerous trees lining the nearby streets. The surroundings within half a kilometer consist of numerous, multi-story university buildings to the north and west, an open park, a field, and a wooded, one-story residential neighborhood to the southwest and south, and a golf course further to

Both locations can be characterized as local climate zones (LCZ) 56, open mid-rise to low-rise urban areas with 30–50% impervious area [8]. However, Trails is located closer to the east end of town, with rural areas (LCZ B/C) as close as 1.5 km to the east, while Quad is located more central to the metro area, with open, sparsely build-up LCZ 9 areas more than 2 km distant, and rural areas more than

UHI intensity was determined for each location by comparing observed local temperatures to the rural temperature measurement (10-min average) at Texas A&M's weather station 10 km outside the urban area toward the southwest. In addition, we assembled weather station data (pressure, winds, solar radiation, and

Air temperature sensors and associated logging equipment have become both miniaturized and highly affordable, thus allowing both high spatial density distributed and low-key mobile data collection of the urban heat island [40]. Mobile measurements (traverses) using automobiles, but also bicycle-mounted or personally carried sensors, have been used numerous times in the past to study the UHI

In summer and fall 2015, we carried out an undergraduate student research project [38] using a standard HOBO U12 data logger with 12-bit temperature sensor (model TMC6-HD) from Onset Computer Inc. to evaluate the atmospheric urban heat island of the BCS metro area. The sensor has a typical accuracy of ±0.25°C and is recorded at a better than 0.2°C resolution. It was placed into a passive radiation shield (model RS3, Onset Comp. Inc.), which was mounted to an angular steel bracket attached to a metal sleeve with a flat rubber sheet that slides over a passenger car side window (**Figure 2**). With the window moved nearly all the way up, the

precipitation) for the days measurements were taken for data analysis.

effect, both in long- and in short-term campaigns [32, 33, 37, 41–65].

dry adiabatic lapse rate.

the northeast.

4 km distant (**Figure 1**).

**2.2 Mobile measurements**

model 085 from MetOne Instruments Inc., has a high precision and accuracy multielement thermistor (±0.15°C). Corrections for elevation differences between the measurement sites and the weather station reference location were made using the

Multiple locations were monitored within the BCS metro area during the spring semester of 2015, but only two urban locations, which were monitored throughout 2015, are discussed here. The first location, subsequently identified as the "Trails" apartment complex, was a parking lot area in front of apartment buildings (**Figure 1**) in College Station. This part of town is slightly elevated from the surrounding urban area, and located between a narrow, wooded green corridor, and a major freeway (State Highway 6). While the location was next to pervious lawns, the area was almost tree-less and dominated by buildings, roads, and parking lots. The tallest buildings are the three story apartment buildings; the larger area within half a kilometer, however, includes a midsize mall and associated parking lots to the north, commercial buildings, including hotels, along the freeway to the east, and

**28**

sensor was measuring air temperature at approximately 2 m agl off the passenger side of the car, with its logger safely placed inside the air-conditioned vehicle. This avoided most temperature bias effects from the vehicle itself, or from other vehicles on the road, unless measurements occurred either at very low speeds with air moving across the front hood of the car toward the sensor, such as may have occurred when stopping at traffic lights; or when hot exhaust plumes from vehicles on the road were encountered. Such possible temperature measurement biases, though occasionally encountered, were not removed from the data set.

Typical car moving speeds were 20–30 mph in town and up 50 mph outside the urban areas (approximately 10–25 m s<sup>−</sup><sup>1</sup> ) during all mobile measurements. A smart phone app called RAAH (https://www.raah.co) was used to record the vehicle's location. Data logger and smart-phone times were aligned before each drive. During each traverse data were recorded every 10 s, providing for a typical horizontal resolution of 100+ m. However, considering the sensor's response time, and the accuracy and timing of the location determination, both the absolute bias and uncertainty of a recorded temperature's location were likely larger than 100 m. Such aspects should be considered in all mobile studies when moving speeds are comparable or faster than sensor response times.

Raw temperature data were normalized first for the presumed linear temperature change, if present, during the typically one-hour drive for data collection, then adjusted for elevation differences along the route using the dry adiabatic lapse rate. Both corrections were always significantly smaller than the encountered temperature differences between the rural and (sub-)urban areas traversed.

The routes driven were selected to avoid selective coverage of the metro area, and such that measurements along a route could be completed within approximately 1 h, including nearby rural areas. All routes were driven once in the morning, and then at least two times during the afternoon and around midnight. Based on student availability during the project, completing the set of selected routes lasted from late August to early November 2015.

One route, which was driven twice during the project, brought the vehicle close to the Texas A&M weather station in the rural area approximately 10 km outside the urban area to the southwest (**Figure 1**). Good agreement between the vehicle based sensor and the weather station sensor was observed (<0.5°C differences), and minimum temperatures during each drive were strongly correlated with weather station temperature during the drive hour (slope = 0.98, r2 > 0.99). Consequently, we used calculated temperature anomalies during each traverse and determined UHI intensity from each drive's observed maximum minus minimum temperature. The remaining meteorological parameters were assembled in similar fashion as described above for the stationary measurements.

#### **2.3 Urban area characterization using remote sensing**

To aid in the interpretation of the observed atmospheric UHI effect, we estimated the spatial extent and configuration of various land-cover/land-use (LCLU) areas in Bryan/College Station (BCS), Texas. To this end, we jointly analyzed airborne digital orthophotographs and a LiDAR-derived digital surface model (DSM) of the study area. We thus characterized urban LCLU based on a remote-sensing approach. In particular, we employed a geospatial object-based image analysis (GEOBIA) method [66, 67], which effectively exploits contextual/ spatial information. GEOBIA-based image-processing algorithms and data fusion are needed to more fully exploit image information, particularly in the case of high-spatial-resolution data. For our GEOBIA analysis, we employed eCognition® Developer software, which allows objects/segments to be delineated and used to classify geospatial datasets. We divided the processing steps into two phases: an initial remote-sensing data segmentation and classification phase, and a subsequent post-segmentation/classification editing phase, which we conducted to improve classification accuracy. We performed quantitative classification accuracy assessment on the revised, post-classification edited result.

#### *2.3.1 Segmentation and classification phase*

*Data sets and data pre-processing.* Data processing for this remote-sensing segmentation/classification analysis began with the acquisition of multiple 50-cm natural-color (NC)/color-infrared (CIR) Digital Orthophoto Quarter Quad (DOQQ ) images, acquired from the Texas Natural Resource Information System (TNRIS). These images were collected between October 2014 and August 2015 as part of the 2015 Texas Orthoimagery Program. We mosaicked the DOQQs for the BCS, Texas study area, and then spatially resampled the mosaic to a 5-m grid cell size in order to facilitate subsequent completion of GEOBIA computational processing. Additionally, we utilized a high-point-density LiDAR point cloud, acquired from Texas A&M University, which was collected in conjunction with the State of Texas over the February 9–10, 2015 period. We filtered and processed the LiDAR point cloud using Esri ArcGIS LAS tools to produce a digital surface model (DSM) of the BCS area. After initial processing of the image and LiDAR data, we spatially subset the respective data sets using a polygon to fit the same areal extents.

*Information classes and training set delineation.* The LCLU information classes employed in this analysis are: roads, roofs, other impervious, trees, lawns, water, bare soil, and pasture/cropland. The pasture/cropland class is actually a combined pasture/grassland/cropland class, which contains grasslands that are not (residential)

**31**

*Low-Key Stationary and Mobile Tools for Probing the Atmospheric UHI Effect*

1 Roads 62 Roofs 281 Other Impervious 43 4 Trees 43 Lawns 153 Water 219 Bare Soil 39 Pasture/Cropland 46

**Class ID number Class name Number of training polygons**

lawns. We manually delineated the training areas based on manual/visual interpretation of the aerial photography and DSM. We initially generated training areas for a set of 15 classes, then later merged some of these classes, yielding the aforementioned final set of eight information classes. The distribution of training-area polygons

*Number of training polygons per class, used for GEOBIA-based LCLU classification of Bryan/College Station* 

*GEOBIA parameter values, input variables, and ruleset development*. We created a ruleset within eCognition® Developer using the imagery, DSM, and training areas. The process involved multiple steps, including: segmentation, class assignment using the training areas, assigning the classified segments as samples, configuring the nearest-neighbor classifier, and applying/conducting the classification. We determined GEOBIA parameter values via iterative, trial-and-error experimentation, with the objective of maximizing LCLU map classification accuracy, while also accounting for computational constraints of the GEOBIA system/computing environment. For the classification step, we used a nearest-neighbor classification algorithm. The classification input features included: mean blue band, mean green band, mean red band, mean near-infrared (NIR) band, mean DSM, standard deviation blue band, standard deviation green band, standard deviation red band, standard deviation NIR band, standard deviation DSM, normalized difference vegetation index (NDVI) [68], and normalized difference water index (NDWI) [69]. We applied the classifier at the image-object level [66], and exported classification results in both vector and raster formats, where we subjected the latter product (at

The resultant LCLU classification entailed various areas of clear misclassifications, including some roads, parking lots, and roofs/buildings. Therefore, we performed some post-/segmentation/classification editing to increase classification accuracy of the class information. In particular, publicly available geographic information system (GIS) data from the City of College Station and the Brazos County Appraisal District aided editing of the LCLU classification map. For roads, we used a vector GIS layer containing road center lines, where we buffered the center lines, 5 m on each side, followed by a dissolve operation. We assigned the road classification value to the buffered area, converted to the data to the raster data model, and then overwrote the pixels in the original LCLU map in these areas with these road classification values. For the roofs class, we followed a similar procedure, except that no buffers were used. Since we only had access to building

*DOI: http://dx.doi.org/10.5772/intechopen.89514*

across these classes is given in **Table 1**.

**Table 1.**

*(BCS), Texas study area.*

5-m cell size) to subsequent processing and analysis.

*2.3.2 Post-segmentation/classification editing phase*

*Low-Key Stationary and Mobile Tools for Probing the Atmospheric UHI Effect DOI: http://dx.doi.org/10.5772/intechopen.89514*


#### **Table 1.**

*Spatial Variability in Environmental Science - Patterns, Processes, and Analyses*

routes lasted from late August to early November 2015.

station temperature during the drive hour (slope = 0.98, r2

described above for the stationary measurements.

**2.3 Urban area characterization using remote sensing**

ment on the revised, post-classification edited result.

*2.3.1 Segmentation and classification phase*

morning, and then at least two times during the afternoon and around midnight. Based on student availability during the project, completing the set of selected

we used calculated temperature anomalies during each traverse and determined UHI intensity from each drive's observed maximum minus minimum temperature. The remaining meteorological parameters were assembled in similar fashion as

To aid in the interpretation of the observed atmospheric UHI effect, we estimated the spatial extent and configuration of various land-cover/land-use (LCLU) areas in Bryan/College Station (BCS), Texas. To this end, we jointly analyzed airborne digital orthophotographs and a LiDAR-derived digital surface model (DSM) of the study area. We thus characterized urban LCLU based on a remote-sensing approach. In particular, we employed a geospatial object-based image analysis (GEOBIA) method [66, 67], which effectively exploits contextual/ spatial information. GEOBIA-based image-processing algorithms and data fusion are needed to more fully exploit image information, particularly in the case of high-spatial-resolution data. For our GEOBIA analysis, we employed eCognition® Developer software, which allows objects/segments to be delineated and used to classify geospatial datasets. We divided the processing steps into two phases: an initial remote-sensing data segmentation and classification phase, and a subsequent post-segmentation/classification editing phase, which we conducted to improve classification accuracy. We performed quantitative classification accuracy assess-

*Data sets and data pre-processing.* Data processing for this remote-sensing segmentation/classification analysis began with the acquisition of multiple 50-cm natural-color (NC)/color-infrared (CIR) Digital Orthophoto Quarter Quad (DOQQ ) images, acquired from the Texas Natural Resource Information System (TNRIS). These images were collected between October 2014 and August 2015 as part of the 2015 Texas Orthoimagery Program. We mosaicked the DOQQs for the BCS, Texas study area, and then spatially resampled the mosaic to a 5-m grid cell size in order to facilitate subsequent completion of GEOBIA computational processing. Additionally, we utilized a high-point-density LiDAR point cloud, acquired from Texas A&M University, which was collected in conjunction with the State of Texas over the February 9–10, 2015 period. We filtered and processed the LiDAR point cloud using Esri ArcGIS LAS tools to produce a digital surface model (DSM) of the BCS area. After initial processing of the image and LiDAR data, we spatially

subset the respective data sets using a polygon to fit the same areal extents. *Information classes and training set delineation.* The LCLU information classes employed in this analysis are: roads, roofs, other impervious, trees, lawns, water, bare soil, and pasture/cropland. The pasture/cropland class is actually a combined pasture/grassland/cropland class, which contains grasslands that are not (residential)

One route, which was driven twice during the project, brought the vehicle close to the Texas A&M weather station in the rural area approximately 10 km outside the urban area to the southwest (**Figure 1**). Good agreement between the vehicle based sensor and the weather station sensor was observed (<0.5°C differences), and minimum temperatures during each drive were strongly correlated with weather

> 0.99). Consequently,

**30**

*Number of training polygons per class, used for GEOBIA-based LCLU classification of Bryan/College Station (BCS), Texas study area.*

lawns. We manually delineated the training areas based on manual/visual interpretation of the aerial photography and DSM. We initially generated training areas for a set of 15 classes, then later merged some of these classes, yielding the aforementioned final set of eight information classes. The distribution of training-area polygons across these classes is given in **Table 1**.

*GEOBIA parameter values, input variables, and ruleset development*. We created a ruleset within eCognition® Developer using the imagery, DSM, and training areas. The process involved multiple steps, including: segmentation, class assignment using the training areas, assigning the classified segments as samples, configuring the nearest-neighbor classifier, and applying/conducting the classification. We determined GEOBIA parameter values via iterative, trial-and-error experimentation, with the objective of maximizing LCLU map classification accuracy, while also accounting for computational constraints of the GEOBIA system/computing environment. For the classification step, we used a nearest-neighbor classification algorithm. The classification input features included: mean blue band, mean green band, mean red band, mean near-infrared (NIR) band, mean DSM, standard deviation blue band, standard deviation green band, standard deviation red band, standard deviation NIR band, standard deviation DSM, normalized difference vegetation index (NDVI) [68], and normalized difference water index (NDWI) [69]. We applied the classifier at the image-object level [66], and exported classification results in both vector and raster formats, where we subjected the latter product (at 5-m cell size) to subsequent processing and analysis.

#### *2.3.2 Post-segmentation/classification editing phase*

The resultant LCLU classification entailed various areas of clear misclassifications, including some roads, parking lots, and roofs/buildings. Therefore, we performed some post-/segmentation/classification editing to increase classification accuracy of the class information. In particular, publicly available geographic information system (GIS) data from the City of College Station and the Brazos County Appraisal District aided editing of the LCLU classification map. For roads, we used a vector GIS layer containing road center lines, where we buffered the center lines, 5 m on each side, followed by a dissolve operation. We assigned the road classification value to the buffered area, converted to the data to the raster data model, and then overwrote the pixels in the original LCLU map in these areas with these road classification values. For the roofs class, we followed a similar procedure, except that no buffers were used. Since we only had access to building

footprint data from the City of College Station, we only applied this operation to areas within College Station, and not the City of Bryan; however, a large portion of the study area lies within the City of College Station boundary. For the other impervious class, based on visual interpretation of the aerial photography, we digitized impervious areas—especially parking lots—where misclassifications were evident. After rasterization of the digitized polygons, we then overwrote misclassified pixels in the original LCLU raster using these newly digitized data. Based on visual inspection, these post-classification GIS operations generally markedly improved classification accuracy. One caveat though is that although the DOQQs and LiDAR point clouds were collected close in time, the data in the vector GIS layers used in this post-segmentation/classification editing phase were acquired some time prior to the DOQQ and LiDAR acquisitions, and this temporal disjunction may have resulted in some errors in the edited LCLU map, particularly given the relatively high rate of urbanization in the BCS area in recent years. Furthermore, future analysis may entail using variable buffer distances for the road centerlines, depending upon road type.
