*Article* **"Cool" Roofs as a Heat-Mitigation Measure in Urban Heat Islands: A Comparative Analysis Using Sentinel 2 and Landsat Data**

**Terence Mushore 1,2, John Odindi 1,\* and Onisimo Mutanga <sup>1</sup>**


**Abstract:** Urban growth, characterized by expansion of impervious at the cost of the natural landscape, causes warming and heat-related distress. Specifically, an increase in the number of buildings within an urban landscape causes intensification of heat islands, necessitating promotion of cool roofs to mitigate Urban Heat Islands (UHI) and associated impacts. In this study, we used the freely available Sentinel 2 and Landsat 8 data to determine the study area's Land Use Land Covers (LULCs), roof colours and Land Surface Temperature (LST) at a 10-m spatial resolution. Support Vector Machines (SVM) classification algorithm was adopted to derive the study area's roof colours and proximal LULCs, and the Transformed Divergence Separability Index (TDSI) based on Jeffries Mathussitta distance analysis was used to determine the variability in LULCs and roof colours. To effectively relate the Landsat 8 thermal characteristics to the LULCs and roof colours, the Gram–Schmidt technique was used to pan-sharpen the 30-m Landsat 8 image data to 10 m. Results show that Sentinel 2 mapped LULCs with over 75% accuracy. Pan-sharpening the 30-m-resolution thermal data to 10 m improved the spatial resolution and quality of the Land Surface map and the correlation between LST and Normalized Difference Vegetation Index (NDVI) used as proxy for LULC. Green-colour roofs were the warmest, followed by red roofs, while blue roofs were the coolest. Generally, black roofs in the study area were cool. The study recommends the need to incorporate other roofing properties, such as shape, and further split the colours into different shades. Furthermore, the study recommends the use of very high spatial resolution data to determine roof colour and their respective properties; these include data derived from sensors mounted on aerial platforms such as drones and aircraft. The study concludes that with appropriate analytical techniques, freely available image data can be integrated to determine the implication of roof colouring on urban thermal characteristics, useful for mitigating the effects of Urban Heat Islands and climate change.

**Keywords:** cool roofs; urban heat islands; land surface temperatures; roof colour; mitigation; urban growth

#### **1. Introduction**

Urbanization, and the associated urban land use and land cover (LULC) spatial structure transformations influence the urban thermal characteristics [1–3]. This process is typified by transformation from natural to impervious surfaces such as buildings and other urban fabrics that alter surface and near-surface temperatures [4,5]. The increase in temperatures attributable to urban growth are associated with a range of challenges that include adverse effects on human health, increased water and energy demand and air pollution [6–8]. As such, urbanization and consequent thermal elevation has been known to exacerbate in- and out-door ambient thermal discomfort that diminish the quality of urban

**Citation:** Mushore, T.; Odindi, J.; Mutanga, O. "Cool" Roofs as a Heat-Mitigation Measure in Urban Heat Islands: A Comparative Analysis Using Sentinel 2 and Landsat Data. *Remote Sens.* **2022**, *14*, 4247. https://doi.org/10.3390/ rs14174247

Academic Editor: Anthony Brazel

Received: 5 July 2022 Accepted: 25 August 2022 Published: 28 August 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 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 (https:// creativecommons.org/licenses/by/ 4.0/).

life [9,10]. Hence, it is increasingly becoming desirable to adopt climate-smart approaches that could enhance sustainable urban living.

Remotely sensed data offer an opportunity to determine urban spatio-temporal variations and their respective thermal characteristics [11,12]. Additionally, remotely sensed data allows for analysis at a range of time-scales that include sub-seasonal patterns. Over the years, technological advancement has facilitated the acquisition of both optical and thermal data on the same sensor platforms (e.g., Landsat, ASTER and MODIS), valuable for urban landscape transformation and thermal analysis [13–16]. Hence, these moderate resolution sensors have been widely used to determine the influence of LULCs on the thermal environment e.g., [17–22]. However, such moderate resolution datasets suffer from the mixed pixel problem, especially in urban areas characterized by landscape heterogeneity, which compromises their value for detailed surface analysis such as the detection of individual houses and their thermal properties [23].

Fortunately, recent sensor developments and advancements in computational power offer an opportunity for improved land surface analysis. For instance, whereas the Landsat series has over the years improved in spectral and radiometric properties, new generation sensors such as Sentinel 2 offer data with improved spatial resolution [24–26]. The sensors' 10-m spatial resolution for instance, ref. [27] allows for analysis of complex environments such as urban areas with reduced mixed pixel effect and high mapping accuracy. Whereas the Sentinel 2s platform lacks a thermal sensor, its integration with high quality data such as Landsat has potential to improve our knowledge of the relationship between urban LULCs and surface temperatures. Recently, Mushore et al. [28] showed that pan-sharpening of Landsat thermal data improves its Land Surface Temperature (LST) mapping accuracy, while Kaplan and Avdan [26] used Sentinel 2 s pan-sharpened 10-m to improve 20-m resolution bands. However, whereas the Sentinel's 10-m spatial resolution optical data can be used to derive detailed urban surface features, Landsat thermal data need to be at a similar spatial resolution for optimal analysis and mapping accuracy.

Several studies have demonstrated that built-up areas absorb and store large amounts of heat when compared to other LULC types, e.g., [22,27–30]. The thermal effect is enhanced by increased building densities that result in large surface areas for heat absorption. Furthermore, dense high-rise buildings increase heat storage capacity as walls present even larger surface areas for heat absorption. Buildings also concentrate heat in an area by retarding its removal by winds [31]. To date, a significant number of studies have dwelt on the effect of buildings on temperature. For instance, the effect of building density and height have been widely demonstrated in both the developed and the developing world, e.g., [28,32–35]. Besides density and height, building materials and other properties such as roof characteristics influence a built environment's thermal properties. For example, Mackey et al. [36] demonstrated that cool roofs surpassed green roofs, street trees and green spaces in cooling effects in Chicago. However, the adoption of remotely sensed data to understand the influence of roofing properties on temperature remains limited. Emphasis has been largely placed on understanding the influence at a broad scale and general LULC classes on the thermal environment. Focus on localized phenomena that include the effect of individual houses and their characteristics such as roof properties using freely available remotely sensed data has remained a grey area.

Studies on the effect of roofs on buildings thermal characteristics have mainly focused on rooftops with vegetation (i.e., 'green roofs') and commonly use data derived from installed meteorological instruments and analytical models [37–40]. Other studies have investigated roof characteristics such as roof angle; for instance, Tian et al. [41] compared the thermal characteristics of curved and flat roofs. Studies on roof colour have established that white roofs have more cooling effect than grey, red and black roofs [42–44], while coating coloured roofs with highly reflective materials can increase thermal performance and energy efficiency of buildings [45]. For instance, Libbra et al. [45] found that the use of cool roofs can reduce air conditioning energy consumption by 70%. For the same roof type, variations such as colour and age may also influence their interactions with heat [43,46,47]. However, due to limitations of the new generation sensors' spatial resolution, literature on the influence of roof colour on thermal performance of buildings remains scarce.

Zhao et al. [48] examined daytime and nighttime effects of roof footprints and configurations using high resolution airborne LIDAR and Quickbird satellite data (2.4-m resolution) and MODIS/ASTER simulated airborne 7-m-resolution surface temperature data. They observed that rooftop spectral attributes, slope, aspect and surrounding trees affected roof surface temperature. Although they accurately delineated roof configurations, they did not segment the roofs by colour. Furthermore, while sensors on aerial platforms such as drones and airplanes can provide data for detailed analysis of effects of roofs on thermal characteristics, such data remain expensive and not viable for studies over large spatial extents. Hence, there is a need to test the value of freely available moderate resolution optical and thermal datasets to enhance our understanding on the influence of building roof colour on thermal characteristics, especially in growing cities of developing countries. Such efforts are necessary to determine the potential adoption of roof type and colour to mitigate heat islands.

According to Alchapar and Correa [46], roof coating is the most influential morphological determinant of roof thermal behavior, while Libbra et al. [45] notes that roof colour controls the absorption of heat during the day and its emission at night. As such, it is necessary to consider "cool" roofs for UHI mitigation. Hence, in relation to adjacent LUCLs, this study sought to determine the value of Sentinel 2 10-m resolution and pan-sharpened Landsat image data in differentiating the influence of roof colour on surface thermal values.

#### **2. Methodology**

#### *2.1. Description of the Study Area*

The study was carried out in a low-density residential area close to the Central Business District (CBD) of the capital city of Zimbabwe, Harare (Figure 1). Since the study sought to determine the influence of roof colour on urban thermal characteristics, it was restricted to a small spatial extent to limit excessive heterogeneity that typifies urban landscapes. Also, a large area could have introduced additional variables (e.g., elevation and slope) that influence thermal characteristics. The area is in a low-to-medium-density residential type, however, some of the houses, especially towards the CBD, have been turned into offices. Low-to-medium-density residential areas in Zimbabwe are characterized by spacious housing units, high land value and higher vegetation density when compared to highdensity residential areas, which are predominantly occupied by the low-income strata. Since house units in the low-to-medium-density residential areas are generally large, they are potentially discriminable using 10-m or higher spatial resolution image data. Hence, based on the 10-m spatial resolution image data, the area was chosen to minimize the mixed pixel problem that characterizes the high-density residential areas. Furthermore, the area is dominated by houses with tiles, thus eliminating the effect of other roof types such as concrete, zinc, aluminium or thatch on the area's thermal characteristics. This enabled the study to determine the variability in temperature based on roof tiling of different colours.

**Figure 1.** Location of the study area in Southern Africa, Zimbabwe and Harare (**a**), Harare and the study area (**b**) and 10-m resolution natural colour composite showing variations of LULC in the study area (**c**).

#### *2.2. Field and Remotely Sensed Datasets*

A field survey was conducted to identify the LULCs and roof colours in the area. The survey revealed that the major LULC classes are grasslands, buildings with heterogeneous tiled roof colouring, bare soil and roads. The building class was further split into roof colours in line with the main objective of the study, and a stratified random sampling approach used to collect the ground control points (GCPs). Non-tiled roofs were categorized into "Other LULCs". For each identified category, coordinates of representative covers were collected using a handheld Global Positioning System. To maximize the spectral and thermal variability, the hot dry season (mid-September to mid-November) was chosen for the collection of the well-distributed LULCs' GCPs as it presents a period of maximum solar energy with no rainfall cooling effect. The LULC types were verified using a GoogleEarth image, which was also used to verify the roof colours and to shift the GCPs to the roof center for classification and validation purposes. The data were split: 70% to be used for classification and 30% for validation.

Landsat and Sentinel 2 data were downloaded from the United States Geological Survey's earth explorer portal at no cost. To minimize variation between field data and image scenes, cloud-free imagery was collected on dates close to field data collection. Two Landsat images (scene capture dates: 16 September 2021 and 3 November 2021) and Sentinel images (scene capture dates: 18 September and 2 November 2021) were used in this study. The dates were chosen as they correspond to the period of maximum radiation in the hot season and the proximity of the two sensor dates. The wind was calm and cloudless, presenting similar weather conditions when data from the same sensor were acquired. Given that the period is dry, vegetation conditions were assumed to be uniform and largely maintained by irrigation/watering throughout the periods. Landsat data acquired on the 16th of September were matched with Sentinel data of the 18th of September, a short enough

period to assume that LULCs did not change. Similarly, Landsat data acquired on the 3rd of November were matched with Sentinel data of the 2nd of November. This gave the best compromise to enable relating and blending multi-sensor data with different spatial and temporal resolutions. The two Landsat images were used to compute the average LST to minimize randomness associated with a single-date image, while the two Sentinel 2 image datasets were used for LULC classification. Multi-spectral optical 10-m resolution Sentinel 2 and 3-m resolution Landsat 8 data for each acquisition date were merged into a multilayer files using the 'Layer stacking' tool in ENVI software. This was done separately for Landsat 8 and Sentinel 2 data. In order to eliminate the effect of aerosols on reflectance values, atmospheric correction was done using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) module in the ENVI software. Due to the proximity to the CBD and location in an urban area, the urban aerosol mode was used in FLAASH, which produced multi-layer reflectance files. Multilayer 10-m resolution Sentinel 2 reflectance data were required to provide multi-spectral information to enhance separation of features in supervised image classification. On the other hand, spectral reflectivity bands in the near-infrared and red range from Landsat 8 were needed for the computation of normalized difference vegetation index (NDVI), useful for emissivity correction in LST retrieval.

#### *2.3. Separability Analysis*

Sentinel 2 10-m resolution bands and the 70% of the field-collected GPS points for the LULCs and roof colour categories were overlaid in an ENVI version 4.7 environment. Surface separability was done using the Transformed Divergence Separability Index (TDSI) based on Jeffries Mathussitta distance analysis. For each paired classes, TDSI ranges between 0 and 2, with values greater than 1 indicating that two classes are distinguishable and values close to 2 implying very high separability. Values below 1 and close to 0 suggests that the classes should be merged. The TDS analysis was necessary to test whether different roof colours and LULCs could be separated before classification.

#### *2.4. Land Use/Cover Mapping and Retrievals of Roof Colours*

The LULCs were derived from Sentinel 2 s 10-m bands based on the 70% GCPs using the Support Vector Machines (SVM) algorithm in ENVI version 4.7 software. Default settings of 0.083 and 100 were used for Gamma in kennel function and penalty parameter, respectively. The SVM uses two classes of training samples within a multidimensional feature space to fit an optimum dividing hyperplane. It aims to maximize the variability between the most proximal training samples (support vectors) and the hyperplane [49,50]. To achieve our objective, we chose a Gaussian radial-basis kennel function as it is ideal for working in an infinite-dimensional space and has a single parameter [49–51]. We classified the images into eight classes, namely, Roads and Bare, Trees, Grassland, Red roof, Blue roofs, Green colour roofs, Black roofs and Grey roofs. To display the roof colours from other LULCS, the Roads and Bare, Trees and Grassland classes were amalgamated into "Other LULC". Thereafter, a confusion matrix was generated. A confusion matrix compares the assigned class labels on the classified map with the location's actual LULC class observed in the field (ground truth). The confusion matrix was used to derive the most widely used accuracy indicators, namely, Overall Accuracy (OA) and Kappa (k) [52].

#### *2.5. Land Surface Temperature Retrieval from Landsat 8 Data*

Band 10 of Landsat 8 was used to retrieve LST from thermal infrared data using Planck's radiation law-based equation for single-channel Landsat thermal data [53]. Initially, thermal infrared digital numbers were converted to surface-leaving radiance using Equation (1);

$$L\_I = M\_I \ Q\_{\mathbb{C}AL} + A\_L \tag{1}$$

where, *Ll* is spectral radiance at Top of the Atmosphere measured in Watts/m2/srad/μm, *Ml* is Band-specific multiplicative rescaling factor, *QCAL* represents pixel values (Digital Numbers) and AL is the Band-specific additive rescaling factor. *Ml*, *AL* and *QCAL* are

obtained from the metadata downloaded together with the Landsat 8 data. As described by U.S. Geological Survey [54], the coefficients for converting digital numbers to thermal radiances were obtained from the metadata file accompanying Landsat 8 data download.

Mumtaz et al. [55] provides an in-depth description of steps for land surface temperature retrieval. The procedures include conversion of thermal radiances to blackbody/brightness temperature followed by emissivity correction to obtain surface temperatures. As such, derived radiances were used in Equation (2) to determine brightness/blackbody temperature.

$$T\_B = \frac{K\_2}{\ln\left(\frac{K\_1}{L\_{II}} + 1\right)}\tag{2}$$

where, *TB* is the brightness temperature (in degrees Kelvins), *K*<sup>2</sup> and *K*<sup>1</sup> area conversion constants for the thermal band (in this case Band 10), also obtained from the metadata file. Since brightness temperature over surfaces is calculated by assuming emissivity to be equal to 1, further analysis must consider actual emissivity which varies with LULC type. This was achieved through emissivity correction, which converted brightness temperatures to actual surface temperatures using Equation (3) [53,55].

$$T\_S = \left(\frac{T\_B}{1 + \left(\frac{\lambda \times T\_B}{a}\right) \ln \varepsilon}\right) - 273.16\tag{3}$$

where *TS* is the LST in Degree Celsius, *λ* is the central wavelength of emitted radiance (10.9 <sup>μ</sup>m for band 10 of Landsat 8), *<sup>ε</sup>* is the emissivity and <sup>α</sup> is a constant (1.438 × <sup>10</sup>−<sup>2</sup> mK). Due to its simplicity, Equation (4) was used to estimate emissivity from Normalized Difference Vegetation Index (NDVI) using [55–57];

$$
\varepsilon = a + b \ln(\text{NDVI}) \tag{4}
$$

where *a* = 1.0094 and *b* = 0.047. Developed in Botswana, which is close to the study area, the equation was chosen due to ease of computation, parsimony and proven applicability in a tropical environment [55]. The *NDVI* was retrieved using reflectance in the Near Infrared (Band 5) and Red (Band 4) of Landsat 8 in Equation (5) [53,58];

$$NDVI = \frac{(NIR - RED)}{(NIR + RED)} \tag{5}$$

where *NIR* and *RED* are reflectance in the near-infrared and red ranges [59] derived from Band 5 and Band 4 of Landsat 8, respectively. The steps above obtained LST at a resampled resolution of 30 m, requiring further enhancement for analysis of roofs thermal properties at a local scale.

#### *2.6. Gram-Schmidt Pan-Sharpening Based Method for LST Image Data Pan-Sharpening*

Improvement of LST data from 30-m to 10-m spatial resolution was achieved using the Gram–Schmidt pan-sharpening technique. The Gram–Schmidt method uses weighted addition of multi-spectral bands to produce a replicated pan-sharpened low-resolution image. Gram–Schmidt orthogonalization is then used to make all bands of the multispectral low-resolution data orthogonal and scalar products are computed and turned into covariances [60]. For each band of the low-resolution multispectral data, angles between the band and the simulated low-resolution panchromatic are computed. Gain and bias of the high-resolution panchromatic band is used to simulate each low-resolution panchromatic band. The process is reversed using the same transformation coefficients, and high resolution pan-sharpened bands are produced [60,61]. Using Gram–Schmidt transformation, the colours of the composite RGB pan-sharpened bands are near similar to the respective original images, thus there is minimal distortion of spatial patterns. The method was chosen because all transform coefficients are computed in the low MS

resolution, hence are more robust to spatial misalignment of the bands than most other pan-sharpening methods [60]. In this study, the Sentinel 2 s 10-m resolution Band 2 was used to improve the Landsat data. The purpose was mainly aimed at producing thermal data for retrieval of LST at 10-m resolution to match with the products from supervised image classification.

#### *2.7. Intensity Analysis for In-Depth Characterization of Local Climate Zones Changes*

The LST spatial configurations before and after pan-sharpening were compared to assess the effect of improving spatial resolution on image quality. The root mean-square error was also used to check the difference after resampling the LSTs to 30-m resolution to assess the effect on values per pixel. A 30-m resolution Landsat scene was used to derive NDVI and its correlation with 30-m LST (after resampling using a bicubic convolution) was obtained using the "Zonal Statistics as a Table" tool in ArcGIS version 10.2, ESRI, Redlands, California, USA. Similarly, NDVI was calculated using 10-m resolution near infrared and red Sentinel 2 and correlated with pan-sharpened 10-m resolution LST. The LST correlations with NDVI before and after pan-sharpening were then compared.

#### *2.8. Linking LULC Types and Roof Colours with LST*

Qualitatively, the spatial structure of LULC and roof colours was compared with that of LST using visual inspections of maps produced from the combination of Sentinel 2 10-m resolution and Landsat 8 thermal data. For quantitative assessment, field-collected points corresponding to each LULC and roof colour category were used to extract LST values using the "Extract values to points" spatial overlay function in ArcGIS version 10.2. The field-collected points were used instead of overlaying the LST with the retrieved LULC map to eliminate the effect of classification accuracy on extracted temperatures for the different categories. Box plots were used to depict the variations of LST between and within LULC and roof colour categories in the study area. The mean LSTs for the different LULC and roof colour categories were also used to compare their thermal performances. This was done to assess the effect of improving resolution on the relationship between LULC and LST using NDVI as a proxy for LULC spatial patterns.

Figure 2 summarizes the procedures from data collection to linking of roof colours to LST spatial structures in the study area.

**Figure 2.** Summary of steps followed in the study.

#### **3. Results**

#### *3.1. Separability of LULC and Roofs by Colour*

Table 1 indicates that the TDSI values ranged between 1.74 and 2.00, implying that the LULC and roof colour categories were distinguishable using spectral signatures from 10-m resolution Sentinel 2 bands. Tarred roads and Trees were the most discriminable classes while the Green and Grey roofs were least discriminable, as indicated by TDSI values of 2.000 and 1.708, respectively. However, although Green and Grey roofs were the least separable, the TDSI value was significantly above the separability threshold of 1, hence guaranteed that the two classes were distinguishable. Among the roof colour categories, blue and red roofs were the most separable with a TDSI value of 1.995. The trees LULC category was the most separable from other cover types, with TDSI values ranging between 1.997 and 2.000. Overall, TDSI values greater than 1.7 indicate that the LULC and roof colour classes in the study area were easily distinguishable.

**Table 1.** Discriminability of LULC types in the study area using 10-m resolution Sentinel 2 data.


#### *3.2. Land Use/Cover and Roof Colour Mapping Using 10 M Resolution Sentinel 2 Data*

The LULCs presented the houses surrounded by abundant vegetation, a characteristic of low-to-medium-density residential areas in Zimbabwe (Figure 3a). The study area has large grasslands, especially in the northeastern regions. The grasslands in the northeast are mainly sporting grounds. The other open grasslands within built-up areas are school grounds while the fragmented grasslands are mainly lawns around houses as well as unused land. The other abundant vegetation was in built-up areas. Figure 3b shows that due to narrow widths in relation to the 10-m resolution of the data, most roads, especially in the black roofs category were not visible.

**Figure 3.** 10-m resolution (**a**) LULC map and (**b**) roof colour map.

#### *3.3. Accuracy of LULC and Roof Colour Retrievals from Sentinel 2 Data*

LULC and roof colour categories were mapped with Overall Accuracy (OA) of 84.5% and Kappa of 0.81. Producer Accuracies (PA) were greater than 75% except for the grey roofs and tarred roads (Table 2). User Accuracies (UA) were less than 75% for the black roofs, trees and grey roofs, while greater than 77% for the other categories. The red roofs were mapped with the highest accuracy of all the other categories (PA and OA greater than 93%).


#### *3.4. Comparison of 30 M Resolution with Sharpened 10 M Resolution LST Retrievals*

Although the study area was small, variations in temperature were observed as some places were more than 15 ◦C cooler than others. Hotspots were noticed, especially on the southern half of the area where LSTs close to 49 ◦C were observed. The northern half was generally cooler, with the dominance of LSTs close to 41 ◦C. There was a general southeastward warming in the area. Comparison of Figure 4a,b shows that sharpening of LSTs to 10-m resolution by blending Landsat-derived LSTs with 10-m resolution Sentinel 2 did not compromise the spatial structure of LST and their ranges in the area. The 30-m resolution LST map was more pixelated than the 10-m resolution, indicating the latter's improved quality. When compared to the 30-m resolution, 10-m LST were retrieved with high accuracy (RMSE = 0.5 ◦C). Correlation between LULC and NDVI was −0.516 and −0.999 before and after pan-sharpening, respectively.

**Figure 4.** Spatial structure of (**a**) LST derived from Landsat thermal data at 30-m resolution (**b**) LST sharpened to 10-m resolution.

#### *3.5. Variations of LST with LULC and Roof Colours*

Although there were overlaps in temperature between different LULCs and roof colour categories, their mean LSTs were clearly distinct (Figure 5). The mean LST was lowest in the trees LULC category followed by blue roof. Highest LSTs were recorded in green-colour roofs and tarred roads areas. The grasslands LCZ showed greatest variability in LST, followed by green and red roofs. The order of roof colours from coolest to warmest based on average LST was blue (36.2 ◦C), black (35.8 ◦C), grey (36.9 ◦C), red (37.4 ◦C) and green (37.7 ◦C).

**Figure 5.** Observed variations of LST with LULC and roof colour classes in the study area.

#### **4. Discussion**

Separability of all classes was high, as indicated by a TDSI greater than 1.7. This was attributed to the strength of the spectral information at the 10-m resolution bands of Sentinel 2 to distinguish between different LULC and roof colours. As aforementioned, separability values close to 2 indicate that the classes are sufficiently separable using the remotely sensed image guided by GCPs [49]. The LULCs and roof colours in the study area were mapped with 75% and 0.73 accuracy and kappa, respectively. To facilitate effective separability of the heterogeneous study area, the study focused on a small spatial extent. Hence, the mapping accuracy was reasonable and comparable with other studies in urban environments such as Sithole and Odindi [62]. However, in this study, roads, (producer accuracy < 50%) were not effectively mapped. The low roads-mapping accuracy can be attributed to their narrow width, that they are largely below 10-m and along the road, and tree and tree shading, hence camouflage and/or mixed pixel with adjacent features.

Based on validation points, the roof colours were retrieved with reasonable accuracies. Producer accuracies (PA) ranged between 58 and 95%, while user accuracies (UA) were between 55 and 91%. The PA and UA values between 55 and 65% could be attributed to intra-class variabilities, which caused some similarities between different roof colours. For instance, some fading shades of black were near similar to dark shades of grey. Similarly, some shades of blue were closer to grey and black. Although not investigated, we speculate that roof ages and fading influence the similarities in roof colours. This is consistent with Alchapar and Correa [46] who noted that for a given roof colour, thermal properties can change due to age. Mapping accuracy could also be influenced by other effects such as roof shapes, reflectivity [63] and ventilation. For instance, Triano-Juárez et al. [64] observed variations in thermal properties for the same roof colour depending on reflectivity and presence of coating materials. On the other hand, Boji´c et al. [65] observed differences between slanted and flat roofs. However, despite the above-named factors that could influence thermal variability based on roof colouring, our study shows that roof colours could be mapped with acceptable accuracy. We however suggest that for applications that require very high mapping accuracy (>90%), the Sentinel 2 s 10-m resolution data may be insufficient. In this regard, the use of Unmanned Aerial Vehicles derived high spatial resolution data offers great potential for fine-scale mapping.

Similar LST spatial structure was observed before and after sharpening, while accuracy of retrieved 10-m resolution LST relative to the original 30-m resolution was high (RMSE of about 0.5 ◦C). Similar to a recent study by Mushore et al. [28], pan-sharpening also improved correlation between LST and NDVI. In this study, the LST maps effectively showed thermal variations. Spatial comparison of the LULC and LST maps showed that vegetation covers such as large grasslands and trees as well as built-up areas with abundant vegetation (which characterize most of the study area) had comparatively low temperature, an indication that even vegetation within built-up areas has heat mitigation value [62]. Zhang et al. [66] also highlighted that vegetation patches and spatial structure combine in contributing to the reduction in surface temperature of the area they occupy. This explains the surface-temperature-reduction effect of vegetation even within built-up areas. Besides latent heat transfer, the shading effect of vegetation, especially trees, lowers surface temperatures in areas they cover. As such, Zhao et al. [48] noticed the cooling effect of shadows of surrounding trees on roof-top surface temperatures during daytime.

The grey and red roofs were warmer than the black roofs, but cooler than green-colour roofs, which were the warmest (Figure 4). Contrary to expectation, black roofs were not the warmest. This could be attributed to variations in thermal characteristics in relation to, among others, roof and colour shading. For example, due to age, black roofs colouring ranged between dark black and grey. Red and green also had higher thermal values. This finding is consistent with Farhan et al. [44], who found that red roofing had higher thermal values than white roofing. Our findings show that green-colour roofs were the warmest, with average LST values close to tarred roads. On the other hand, blue roofs were the coolest, a finding consistent with Libbra et al. [45], who note that roof colour influences

surface temperatures and hence could be used to mitigate heat islands. Although not investigated in this study, pigments on roof materials could have influenced their thermal behaviour. For example, it was reported that for the same roof colour, cool pigments have the potential to increase albedo by at least 20% [67]. This may have caused dark roofs to absorb less than or comparable heat to light-coloured roofs.

#### **5. Conclusions**

The 10-m resolution Sentinel 2 data mapped LULC and roofs by colour with reasonable accuracy. However, findings show that Sentinel 2 s 10-m spatial resolution is still limited by the mixed pixel problem. Other roof characteristics such as age, shape and coating need to be investigated for potential improvement in mapping accuracy. Sharpening of LSTs derived from Landsat to Sentinel's 10-m resolution improved the LST spatial structure. It also increased the correlation between LST and NDVI, implying an improved relationship with LULC. Different roof colour showed variations in mean LST, which highlighted the contribution of roof colours in mitigating or intensifying the heat island effect. Due to variations in shades attributed to changes in age, black roofs were not the warmest. Blue roofs were found to be the coolest while green-colour roofs were the warmest, followed by red roofs. Grey roofs had a moderate effect, with the cooling effect increasing with lightness of the grey colour. Overall, the study showed that colour, in combination with other roof properties, determines a building unit's thermal characteristics. However, the study observed that even after pan-sharpening, Sentinel 2 s 10-m spatial resolution was still coarse for urban roof mapping.

The study observed that even after pan-sharpening, Sentinel 2s 10-m spatial resolution was still coarse for urban roof mapping. This implies the need to test other higher spatial resolution datasets, for example those derived from UAVs and aircraft platforms. Future studies should also consider separating different shades of the same colour, especially in view of colour changes associated with roof aging. Additionally, the combined effects of various physical factors, which include roof coating, thickness, ventilation, and shape, should be included for in-depth analysis of the effect of roofs on the area's thermal environment. Among the factors to be included simultaneously is the presence and effect of any pigment that may affect albedo and heat absorption capacities, even for rooftops of the same colour. Given the inadequacy of freely available moderate-resolution Landsat 8 and Sentinel datasets in mapping thermal properties of rooftops, there is a need to test other higher spatial resolution datasets, for example those derived from UAVs and aircraft platforms.

**Author Contributions:** Conceptualization, T.M., O.M. and J.O.; methodology, T.M.; software, T.M.; validation, T.M., J.O. and O.M.; formal analysis, T.M.; investigation, T.M., O.M. and J.O.; resources, T.M., O.M. and J.O.; data curation, T.M., O.M. and J.O.; writing—original draft preparation, T.M.; writing—review and editing, T.M., O.M. and J.O.; visualization, T.M.; supervision, O.M. and J.O.; project administration, O.M. and J.O.; funding acquisition, T.M., O.M. and J.O. All authors have read and agreed to the published version of the manuscript.

**Funding:** The research of this article was supported by DAAD within the framework of the climapAfrica program of the Federal Ministry of Education and Research. The publisher is fully responsible for the content. The work and article processing charge was also funded by the National Research Foundation of South Africa (NRF) Research Chair in Land Use Planning and Management (Grant Number: 84157).

**Data Availability Statement:** Remotely sensed data used in this study can be freely downloaded from Earth Explorer website (www.earthexplorer.usgs.gov) courtesy of United States Geological Survey (USGS). Accessed on 10 January 2022.

**Acknowledgments:** We acknowledge the Climate Modeling Group of the climapAfrica fellowship for the support and the Discipline of Geography at the University of KwaZulu-Natal and the Department of Space Science and Applied Physics at the University of Zimbabwe for providing conducive working environment.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


### *Article* **Cooling Effects of Urban Vegetation: The Role of Golf Courses**

**Thu Thi Nguyen 1,2, Harry Eslick 3, Paul Barber 2,3, Richard Harper <sup>2</sup> and Bernard Dell 4,\***

<sup>1</sup> Faculty of Forest Resources and Environmental Management,


**Abstract:** Increased heat in urban environments, from the combined effects of climate change and land use/land cover change, is one of the most severe problems confronting cities and urban residents worldwide, and requires urgent resolution. While large urban green spaces such as parks and nature reserves are widely recognized for their benefits in mitigating urban heat islands (UHIs), the benefit of urban golf courses is less established. This is the first study to combine remote sensing of golf courses with Morphological Spatial Pattern Analysis (MSPA) of vegetation cover. Using ArborCamTM multispectral, high-resolution airborne imagery (0.3 × 0.3 m), this study develops an approach that assesses the role of golf courses in reducing urban land surface temperature (LST) relative to other urban land-uses in Perth, Australia, and identifies factors that influence cooling. The study revealed that urban golf courses had the second lowest LST (around 31 ◦C) after conservation land (30 ◦C), compared to industrial, residential, and main road land uses, which ranged from 35 to 37 ◦C. They thus have a strong capacity for summer urban heat mitigation. Within the golf courses, distance to water bodies and vegetation structure are important factors contributing to cooling effects. Green spaces comprising tall trees (>10 m) and large vegetation patches have strong effects in reducing LST. This suggests that increasing the proportion of large trees, and increasing vegetation connectivity within golf courses and with other local green spaces, can decrease urban LST, thus providing benefits for urban residents. Moreover, as golf courses are useful for biodiversity conservation, planning for new golf course development should embrace the retention of native vegetation and linkages to conservation corridors.

**Keywords:** ArborCam; high-resolution airborne imagery; morphological spatial pattern analysis; land surface temperature; golf courses; vegetation structure

#### **1. Introduction**

Urban development has transformed the land cover of cities causing profound changes in the biological and physical characteristics of the transformed surfaces [1,2]. These changes often result in environmental degradation leading to negative impacts on the quality of life for city dwellers [3]. One of the consequences of urbanization is the relatively higher temperature in urban compared to surrounding peri-urban/rural areas, producing "urban heat islands" (UHIs) [4]. This is due to differences in land use/land cover resulting from human activities. The combined effect of global warming and UHIs is called urban heat [5]. Over recent decades, extreme summer heat has become more frequent across many cities in the world, making urban heat an increasingly important topic in environmental research [6,7]. It is projected that this problem will increase in many regions of the world under the influence of climate change [8] and increased urbanization.

Extreme temperatures have serious impacts on human health, such as heat rash, sunburn, fainting, and heat exhaustion [9,10], which lower the life quality of city dwellers [11].

**Citation:** Nguyen, T.T.; Eslick, H.; Barber, P.; Harper, R.; Dell, B. Cooling Effects of Urban Vegetation: The Role of Golf Courses. *Remote Sens.* **2022**, *14*, 4351. https://doi.org/10.3390/ rs14174351

Academic Editors: Elhadi Adam, John Odindi, Elfatih Abdel-Rahman and Yuyu Zhou

Received: 12 July 2022 Accepted: 27 August 2022 Published: 1 September 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 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 (https:// creativecommons.org/licenses/by/ 4.0/).

A large number of deaths related to heat occurred during heat waves in Chicago in 1995, and in 16 European countries in 2003 [9]. Moreover, rising temperatures in urban areas create an uncomfortable environment for residents that results in increasing demand for energy for cooling systems in homes during extreme heat events [12]. Therefore, understanding the spatial distribution of temperature and underlying drivers associated with cooling effects in urban landscapes is a key concern for urban planners.

In order to deal with urban heat issues, studies have been undertaken to identify the dynamics of warming in urban areas [13,14]. In general, an increase in urban green space results in a cooling effect, while impervious land cover leads to a warming effect [15,16]. Impervious surfaces absorb and retain solar energy, with heat slowly released heat back into the atmosphere [17]. In contrast, grass, trees, and other vegetation have a natural heating and cooling cycle that is disrupted by urban structures. Vegetation cools the surrounding area by providing shade and through evapotranspiration [17–19]. Shaded surfaces may be 10–25 ◦C cooler than unshaded surfaces [20]. Evapotranspiration can help reduce peak summer temperatures by 1–5 ◦C [21]. Therefore, the amount and quality of vegetation in a city can influence the rate of atmospheric CO2 sequestration and the amount of heat that a city retains [22–25]. Whether urban vegetation occurs as large nature reserves or as more fragmented and less functionally healthy green spaces for purposes other than conservation (such as public parks, golf courses, cemeteries, military bases, hospitals, university campuses, or streetscapes), they are critically important in cooling cities and making them more livable [26]. Therefore, livable city planning should require a flexible approach that takes advantage of all opportunities to retain green spaces, combining efforts both in formal parks and other recreational spaces.

Golf courses are a type of recreational green space established for commercial and public purposes. They are often a controversial land-use due to their heavy use of water, chemical herbicides, and exotic ornamental vegetation, and this has led to criticism from ecologists [27]. However, other studies have emphasized the ecological values of golf courses for biodiversity conservation [28,29] and for enhancing the connectivity of vegetation networks in urban landscapes [30]. Although the rough (out of play) vegetated areas and irrigated lawns in golf courses are expected to play a role in cooling cities, the ecological value of golf courses in reducing urban heat has been largely ignored by ecologists [31].

Remote sensing-based studies have allowed researchers to assess the spatial distribution of Land Surface Temperature (LST) in urban areas and to establish correlations between vegetation and urban LST models [32–35]. However, most studies have used low and moderate-resolution satellite imageries, such as MODIS or Landsat, to calculate LST as a proxy of urban heat [36,37]. These approaches do not provide information about how vegetation characteristics such as fragmentation, vertical structure, and crown health impact on the local cooling effect. The moderate resolution (30 m) satellite imagery (the Landsat Thematic Mapper (TM)) sensor limits the capacity to detect vegetation of different height classes and their associated LST variability. In contrast, airborne high spatial resolution imagery (0.3 m) has a much greater capacity to detect more detailed vegetation characteristics such as vegetation height classes [38]. Furthermore, studies that have investigated variation in LST among urban formal parks, open spaces, and residential gardens [39–41] have not included golf courses as a separate urban land use. Not surprisingly then, urban planners often lack information for planning urban development that can help to reduce heat exposure.

Native vegetation is undervalued and is often lost during urban development unless it is protected in biodiversity reserves. Remnant vegetation in golf courses is often under pressure from players who want trees removed that are close to fairways. There is a need to quantify some of the benefits of this vegetation to the broader community in golf course management plans in the future. Hence, we explore the value of golf courses and their vegetation in reducing LST. This study examines the hypotheses that golf courses in urban landscapes play a role in reducing urban heat, and that vegetation structure (height class and spatial configuration) influences variations in LST in urban landscapes. Using high-resolution airborne ArborCam imagery, we compared the land surface temperature within golf courses with those of other urban land-use categories, and determined the influence of vegetation traits and geographic location on the function of golf courses for urban heat reduction. The study was conducted in the suburbs of Perth, Australia, where many golf courses retain some native vegetation and provide green connectivity in the urban landscape [30]. The research examines the potential of using golf courses as a green space out-side of protected area networks, and will thus inform the planning of vegetation configuration and vegetation management to optimize cooling at the local and city scales.

#### **2. Materials and Methods**

#### *2.1. Study Area*

#### 2.1.1. General Description

Perth city is located at latitude −31.953512 and longitude 115.857048 (Figure 1). The Perth Metropolitan area covers 6418 km2, with a population density of 317.7 people per square kilometer. The area is in Australia's southwest corner, a global biodiversity "hotspot" with outstanding natural environments having the highest concentration of rare and endangered species on the entire continent [42,43].

**Figure 1.** Map showing land use categories and the location of golf courses in the two study regions of the Perth Metropolitan Region.

Perth has a Mediterranean-type climate with hot dry summers, lasting from December to late March, and cool wet winters [22]. Extreme heat events (substantial rises in temperature, duration, and frequency of very hot days) have increased in Perth over the past two decades, and are projected to increase in coming years [44]. These events pose health risks for urban citizens especially the elderly, young, sick, and people from lower socio-economic areas [45].

Perth has experienced extensive urban expansion since the 1960s and this has caused sustainability concerns due to the large-scale conversion of natural land to impervious surfaces [46], which can contribute to an increasingly warming urban environment. It is projected that by 2030 the annually averaged warming of this region will be about 0.5 to 1.2 ◦C above 1986–2005 levels [47]. Therefore, the Western Australian government issued a long-term strategic guide for the development of Perth by 2050, which identified reducing urban heat as one of sixteen aspirations under the strategy for the Planning Commission, and State and Local Government by expanding the tree canopy in high urban heat risk areas [45,48].

#### 2.1.2. Spatial Subdivision

This study focuses on the western suburbs (WESROC suburbs in the south) and the Joondalup suburbs (in the north), covering 16,205 ha (Figure 1). The WESROC suburbs are a group of old suburbs established prior to the first urban development planning of Perth (i.e., pre-1950s), and are located west of the city's central business district and north of the Swan River. These suburbs are characterized by low to moderate-density residential areas, recreation areas, nature reserves, and wetlands. Joondalup is a younger urban area that was developed as a result of northerly urban expansion following extensive urban development in the 1990s, and is characterized by dense commercial and residential areas. The suburbs of the two subdivisions, established through different times in the history of Perth's planning with different urban designing styles, are representative of residential suburbs across Perth's sprawling urban landscape. There are six golf courses in the WESROC suburbs and one in the Joondalup region (Figure 1). The golf courses vary in ownership (public, private, and semi-private), size (small < 40 ha, moderate 40–70 ha, and large > 70 ha), number of holes, and the linkage of golf courses to other vegetation (Table 1). The Lake Claremont Golf Course was converted to parkland in recent years. Images of vegetation and environment are readily available online for the respective golf courses used in this study. Figure 2 shows a typical scene. In general, the tees, fairways, and greens are reticulated and irrigated during the dry season (November to April) from underground aquifers. With declining groundwater supplies and a warming climate, the Golf Course Superintendents Association of Western Australia is collaborating with the Department of Water to assist golf courses to become more water efficient.

**Table 1.** Key characteristics of the seven golf courses.


**Figure 2.** Typical scene in a golf course in the study region in summer. The irrigated fairway is bordered by corridors of vegetation containing tall trees (*Eucalyptus*), mid-story trees (*Acacia*, *Agonis*), small shrubs, and grasstrees (*Xanthorrhoea*) (photo by Paul Barber).

#### *2.2. Data Sources and Geospatial Analysis*

To provide accurate information about the daytime LST in relation to urban land use, land cover, and vegetation characteristics, the study used multispectral, high-resolution airborne imagery, acquired at ~11:00 to 13:00 h on two typical hot late summer days (10 and 11 March 2020) with a daily maximum temperature at Perth Metro station (number: 009225) of 35.1 ◦C for both days [49] and calm conditions.

High-resolution RGB, seven-band multispectral, and long wave thermal radiation were acquired concurrently using the custom ArborCamTM vegetation monitoring system (ArborCarbon Pty Ltd., Perth, Australia). Imagery was acquired on dedicated flights using a customized Piper PA-28 aircraft with specifications as described in Table 2.


**Table 2.** Acquisition parameters and resulting image Ground Sample Distance (GSD) for each of the imaging sensors for the two study areas.

The ArborCam sensor captures seven distinct narrow multispectral bands strategically located between 450 and 780 nm of the electromagnetic spectrum [50]. Long-wave thermal Infra-red radiation (Thermal IR 7500–14,000 nm) was converted to LST in degrees Celsius by applying a standard emissivity correction across the scene of 0.95 to produce a single-band 32-bit raster, with each pixel representing land surface temperature.

All imagery was orthorectified and radiometrically corrected using a series of propriety image processing workflows. A Digital Surface Model (DSM) was generated using a Structure from Motion processing technique during orthorectification. This DSM was further classified to identify ground surface pixels, which were then interpolated to produce a Digital Terrain Model (DTM). The difference between the DSM and DTM was calculated to determine the Feature Height Model. Final imagery was converted to units of surface reflectance using radiometric targets placed throughout the scene. Finally, vegetation within the scene was classified using a segmentation and supervised classification approach. The Arbor-Cam thermal imaging sensor uses a microbolometer with a spectral range of 7.5–14 μm, and a resolution of 640 × 480 with a 15◦ × 11◦ field of view. Thermal radiance is corrected and converted to LST on board the camera using a standard emissivity correction of 0.95, and relative humidity of 50% at 20 ◦C. Linear temperature data are recorded in 16 bits, with a sensitivity of 0.05 K and a stated accuracy of ±2 ◦C or ±2%. This is a standard approach for studies of urban land surface temperature. More precise methods of emissivity correction for individual surface materials require the classification of surface materials, which is beyond the scope of the current study. The current study is concerned primarily with the relative differences in LST; therefore, the validation of the reading vs. actual LST is of lesser value.

#### 2.2.1. NDVI Calculation

Normalized Difference Vegetation Index (NDVI) maps were developed by calculating the ratio between the red (R) and near-infrared (NIR) using Equation (1) [51]:

$$\text{NIDVI} = (\text{NIR} - \text{red}) / (\text{NIR} + \text{red}) \tag{1}$$

#### 2.2.2. Morphological Spatial Pattern Analysis (MSPA)

Morphological Spatial Pattern Analysis (MSPA) was employed in this research for analysis of spatial configuration of vegetation cover (turf and all other vegetation types) as described previously [52–54]. In order to undertake the MSPA analysis, the input data (foreground class) were defined. The binary maps (vegetation and non-vegetation) obtained from the classification of PlanetScope 3B images were used as input data with the vegetation being defined as the foreground pixels (green landscape) in the MSPA approach using the MSPA-Toolbox for ArcGIS. The output of the MSPA analysis includes the seven structural categories belonging to two groups: (1) urban vegetation patches (Cores, Edge, Perforation, and Islets) and (2) urban vegetation paths (Bridges, Branches, and Loops) [52–54]. Each of these categories was described at the pixel level [52–54] and described in ecological meaning terms based on the concept of "habitat availability" and "graphic theory" [52–54], and this can be briefly described as:


#### 2.2.3. Land-Use Data

We selected eight land-use categories representing the main components in the urban landscape, as follows: conservation land (1); golf course (2); green space (3); commercial (4); industrial (5); residential (6); main road (7); and other land-use (8). This selection of land-use categories is comparable to those used in ecological research in other urban landscapes [55,56] and is described in Table 3.


**Table 3.** Description of land-use categories.

#### *2.3. Statistical Analysis*

To address the first objective (variation in LST among land-use categories and among golf courses), LST mean values were derived for each of the eight land-use categories using vector data analysis (zonal statistics) in ArcGIS 10.3 and descriptive statistics in R 3.6.1. The land-use layer (Figure 1) was used to define zonal boundaries.

To address the second objective (factors influencing the cooling effects of golf courses), the variation and correlation of LST with each driving factor were derived. We randomly generated more than 500 random points within the seven golf courses and the study area. Values for each independent variable were assigned to each point using Extract Multi Values to Points tool in ArcGIS 10.3. All geographical analyses were conducted using ArcGis version 10.3 and statistical analyses were performed in R 3.6.1 [58]. Based on the initial description of the relationship between LST and the variables, a multiple linear regression model was built with the F-statistic in R 3.6.1 to describe the effects of vegetation characteristics and geographic location that drive LST.

The explanatory variables examined are listed in Table 4. These variables were subdivided into four groups: vegetation height class, MSPA class, NDVI, and distance to water

resources (Table 4). Previous studies have explored the distance to the coast as a factor impacting urban temperature [59]. However, the study area has a network of water bodies, the ocean, the estuary of the Swan River, and groundwater-derived lakes. These water bodies are likely to influence the LST, and thus the distance to the water resources (ocean, lakes, and river) was measured using the near tool in ArcGIS 10.3. The values were added to random points as an independent variable in the regression analysis. The multiple linear regression model was performed in R 3.6.1 [58] to determine the relationship between the dependent variable (LST) and its driving factors (Table 4).

**Table 4.** Independent variables considered in the multivariate model of LST within golf courses and other land use categories. The MSPA descriptors are the same as published by the original workers [52–54].


In terms of vegetation variables, previous studies have focused on vegetation cover [59] and vegetation type (grass, shrubs, and trees) [22]. In this study, more details of vertical vegetation structure were explored where vegetation was classified into height classes (Table 4), and spatial vegetation configuration where vegetation was classified into habitat type (MSPA classes) based on the patch size and their connectivity to other vegetation areas and they were added to random points for the regression analysis.

#### **3. Results**

#### *3.1. Variation in LST among Land-Use Categories and among Golf Courses*

Overall, the conservation land was the coolest land-use category (mean LST of approximately 30 ◦C). Golf courses had the second lowest mean LST (around 31 ◦C) in the study area (Figure 3A), and thus golf courses in high-density urban areas play a role as cool islands (Figure 3B). Joondalup Resort Golf Club is shown as an example (Figure 4). The average LST for industrial, residential, and main road land uses were high, ranging from approximately 35 to 37 ◦C. The land use types in the order of highest to lowest temperatures were main roads, residential, industrial, other, commercial, green space, golf course, and conservation (Figure 3A).

**Figure 3.** Variation in the mean values of LST among (**A**) land-use categories and (**B**) the seven golf courses.

**Figure 4.** Joondalup Resort Golf Club and surrounds: (**A**) True color orthomosaic; (**B**) Land-use categories; (**C**) Vegetation height-strata; (**D**) Day time LST; (**E**) MSPA classes; (**F**) NDVI map. Data were collected in late summer (maximum temperature 35.1 ◦C) on 10 and 11 March 2020.

Notably, the LST differed markedly between golf courses. The highest mean LST occurred within Joondalup Resort Golf Club at about 32 ◦C (Figure 3B). Nedlands Golf Club had the lowest mean LST among studied golf courses (around 29 ◦C). Golf courses located nearby the coast (Cottesloe Golf Club, Sea View Golf Club, Lake Claremont Golf Course, and Wembley Golf Course) had similar mean LSTs at about 30 ◦C. Mosman Park Golf Course had the second highest mean LST (around 31◦C) (Figure 3B). These results indicate that the golf courses have different capacities to cool their local environments, and that there may be underlying drivers leading to this variation.

#### *3.2. Factors Influencing Cooling Effects of Golf Courses*

There was a positive relationship between LST and distance to water resources (Figure 5B), which indicates the availability of water bodies can be beneficial on hot summer days. Moreover, vegetation characteristics can impact cooling. Figure 5C shows the LST in non-vegetated areas was much higher than any type of vegetation (LST median around 35 ◦C) indicating the role of vegetation cover in providing a cooling effect. Within areas of vegetation cover, NDVI values reflect vegetation health, and these showed a strong negative relationship with LST (R = 0.77). This means that green, healthy vegetation has a good capacity to cool urban areas (Figure 5A) with the example of Joondalup Resort Golf Club and surrounds illustrated (Figure 4D,F).

**Figure 5.** Factors affecting Land Surface Temperature: (**A**) Relationship between LST and NDVI; (**B**) Relationship between LST and distance to the water resources; (**C**) Variation in LST among vegetation height classes; (**D**) Variation in LST among MSPA classes.

However, not all types of vegetation have the same cooling effect. The vertical structure and horizontal configuration of vegetation further determine the capacity of vegetation for cooling. In general, the taller the vegetation the cooler the LST (Figure 5C and Table 5). Vegetation of >10 m height had a median LST of around 27 ◦C, 0–3 m high vegetation had a mean LST of around 29 ◦C, and for turf, the LST was around 33 ◦C (Figure 5C).


**Table 5.** Estimates for each independent variable from the multivariate model for predicting the LST.

\*\*\* *p* < 0.001; \*\* *p* < 0.01; \* *p* < 0.05.

The size of vegetation patches and the linkages among them also influence LST (Figure 5D). A large vegetation patch, comprised of the outer Edge and Core categories, had a low LST of 27–28 ◦C. Moreover, the Bridge class that connects two Cores also had a similar low LST (Figure 5D). The Islet representing small, isolated patches, and the Perforation representing the inner Edge had the highest LST (Figure 5D).

The multiple linear regression model for predicting LST (Thermal = Vegetation Strata + MSPA classes + Distance to water resource + NDVI) had an R-squared value of 0.72 and an adjusted R-squared value of 0.7 with *p* < 0.0001 indicating that LST was closely related to the vegetation variables and distance to water resources. However, in the subset of vegetation height class variables, only the coefficient of vegetation 10–15 m and >15 m had *p*-values < 0.05 (Table 5).

Similarly, among MSPA variables, only the classes representing large patches (Core, Edge) had *p*-values of <0.05 (Table 4). This means that taller vegetation (10–15 m and >15 m) and large patches of vegetation that combine the outer Edge and Core areas had significant effects on LST. Moreover, the multi-regression model further determined that the important factors influencing LST were the health of vegetation indicated by the NDVI value and the distance to water (coefficient *p*-values < 0.01) (Table 5). This suggests that these six independent variables are statistically significant predictors of the LST.

#### **4. Discussion**

#### *4.1. Golf Courses as Cooling Islands in Urban Environments*

The study revealed that urban golf courses had lower day-time land surface temperatures than other urban land-use categories, except for conservation land. This means that in a warming climate, golf courses, with most of their surface area covered with a combination of vegetation (shrubs and trees), water bodies, and turf, will be cool-islands and natural havens in cities where most of the surface is dominated by building structures and heat-absorbing hard surfaces. Green spaces in golf courses include irrigated fairways and out-of-play areas, but the cooling effects of golf courses are strongest for woodlands with complex multiple-tiered biomass structures. A similar finding was made for urban green space in Hong Kong [31,60].

In industrial and commercial land uses, the buildings often have flat concrete or metallic roofs, as seen in the aerial imagery. Concrete surfaces have low albedo from 0.1 to 0.35 [61]. In contrast, the vegetation acts as a buffer between the ground and solar radiation, and this helps to reduce the LST [62]. The similarity in LST of golf courses and conservation land can be explained by similar surface characteristics related to their vegetation cover.

Previous studies from Sydney (Australia) and Toronto (Canada) showed that mean temperatures are significantly lower for parks and recreational land uses than for highly intensive urban land-use such as industrial and commercial areas [41,63]. Thus, future changes in land-use from forest and grasslands to new urban developments (industrial, commercial, and residential) are projected to further enhance temperature increases caused by climate change [63]. The UHI problem is serious in Australia's hotter cities such as Perth, Adelaide, and Alice Springs [64]. This is due to the large proportion of impervious surfaces as a result of urbanization. Other Mediterranean-climate cities are predicted to have increases in average minimum temperatures compared to other rural areas [65]. Therefore, in hot dry climates, urban planning should pay more attention to designing cool islands to mitigate the UHI effect and its impact on city residents. With the cooling effects of golf courses identified in our study, we propose that urban golf courses should be considered as a type of cooling island in urban planning within urban heat mitigation strategies.

#### *4.2. Vegetation Characteristics Influence Cooling Effects of Urban Green Spaces*

Previous studies have confirmed the role of vegetation cover in mitigating urban heat [22,47,59], which can help to explain why urban areas without vegetation heat up easily and retain heat [59,66]. The cooling effects of vegetation within urban golf courses have not been well investigated. For example, the study of microclimate at a sub-tropical golf course in Hong Kong only investigated the differential cooling abilities of woodland, a rough grass area, and a bare-concrete rooftop control site within the golf course [31,60]. However, the role of vertical vegetation structure (vegetation height classes) and the spatial arrangement of vegetation patches in cooling urban environments is largely unexplored [67].

By using the high-resolution (0.3 × 0.3 m) airborne imagery, our study provides new insights into the cooling capacities of vegetation of different high classes within golf courses and other green spaces. This study suggests that tall urban forests (>10 m tall) will have the strongest effects in reducing urban heat islands, while shorter vegetation, including turf grass, will be less effective. This provides a new understanding of the relationship between vegetation and urban heat and indicates that urban heat mitigation strategies using green spaces should not solely focus on the extent of vegetation coverage but should also consider the height and vertical structure of the vegetation. Due to the limited space for vegetation in urban areas, it is necessary to maximize the effects of green space by maintaining and increasing the number of big trees to regulate temperature and improve the urban microclimate.

Moreover, our study also explored how vegetation complexity in terms of spatial configuration and arrangement might facilitate the management of urban heat. The results showed stronger cooling effects of large vegetation patches (Core area and their outer Edge) as well as the vegetation paths that link Cores (Bridge) as being more effective vegetation structures than small, isolated patches (Islet). This finding supports previous studies [68–71] where large patch sizes reduced LST due to greater shading of the periphery. Furthermore, larger vegetation patches have more interior areas, which are less affected by the ambient environment, whereas smaller, isolated patches (Islets) tend to have a greater proportion of edge areas, and thus are vulnerable to disturbance from the peripheral region [67].

In addition, our study revealed that vegetation connectivity and patch size are important when designing urban green space. The connectivity of vegetation cover can contribute to decreasing surface temperature in urban areas [72]. Therefore, increasing urban vegetation, maintaining large patches of vegetation, and reducing insolation can help to decrease urban LST. Hence, urban planning should consider the size and configuration of green spaces to operate as cooling islands without becoming masked by surrounding buildings.

Vegetation health is an important factor influencing its effectiveness in cooling urban areas. Healthy vegetation patches with NDVI values from 0.6 to 0.8 had the strongest cooling effect (Figure 5A). Therefore, together with maintaining water bodies in combination

with protecting and restoring big trees in large patches, caring for vegetation health is vital to ensure the cooling effects of green spaces are optimized. Several abiotic and biotic disorders pose a threat to urban tree health with some of these, such as Phytophthora, thriving in irrigated urban parklands [73].

The correlation of LST with impact factors may explain why LST varied among landuse categories. For example, the proportion of vegetation (without turf) was highest on conservation land, and most of the vegetation existing as Core or Bridge areas in this land-use category may explain why the LST was lowest. High vegetation connectivity appears to be an important factor for conservation land having the strongest capacity to reduce urban LST. Vegetation within golf courses that is healthy with a high proportion of tall trees also contributes to the low values of LST in this land-use category. Furthermore, the golf courses were close to or contained water bodies and this helped to mitigate LST in summer in our study.

#### *4.3. Implication for Vegetation Management and Urban Planning*

This study suggests that urban vegetation management approaches are required to mitigate urban heat. Golf courses can contribute significantly to the mitigation of LST in urban landscapes. As large trees play an important role in reducing LST, golf course managers and designers should pay attention to the conservation of these trees. It is recommended that golf course managers should not only increase the natural tree canopy by planting more trees, but also actively protect tall trees and large vegetation patches to improve the cooling capacity of golf courses. It is important though to always consider the conflict between turf health and trees when designing or re-designing golf courses. Large trees provide large amounts of shade with potential negative impacts on turf health when insufficient light is received.

Urban expansion on undeveloped land containing large patches of native vegetation that involves tree clearing should embrace tree planting for future cooling effects. This study suggests that increasing the urban tree canopy should benefit heat mitigation. However, it will be difficult to reach targets without promoting planting on private land. Nowhere is this more pressing than in urban environments where there is a scarcity of available land with native vegetation. The regression equation in this study also provides an indication of how temperature can be reduced in other urban land-use categories (e.g., residential, commercial areas) by tree planting and vegetation patch maintenance. Because increasing vegetation coverage is difficult in some dense urban landscapes, measures to improve the quality of existing vegetation patches, such as tall tree conservation and irrigation, are important for mitigating temperature for improving the well-being of city dwellers.

Novel approaches for heat reduction and livable neighborhood policies should embrace the importance of developing incentives that promote multipurpose use of land and that stimulate cooperation among people and different societal sectors to support urban green space maintenance. Golf courses are an example of commercial land that can contribute to urban heat reduction that should be integrated into livable neighborhood policies to improve the life quality of urban citizens.

#### **5. Conclusions**

This study develops and demonstrates a robust and objective approach to quantify and compare variation in LST among urban land-use categories. The research used highresolution multispectral airborne imagery to classify vegetation height classes and this helped to fill gaps in the current literature that compares the LST of different vegetation types. From our study, it is clear that vertical vegetation structure and horizontal vegetation configuration and arrangement are important in urban heat reduction. It is also evident that the vegetation of golf courses can play a beneficial role in helping to reduce urban heating during hot summer days. Effective management of vegetation for urban heat reduction and livable neighborhoods should consider the maintenance of big trees and large patches of vegetation across the urban landscape. Our study is significant because it provides

insight into the ecological benefits of recreational green spaces such as golf courses in urban landscapes where such ecological roles are often valued in conservation land. The findings from this study suggest that planning for further urbanization of peri-urban land should consider opportunities for the co-planning of golf course development in conjunction with the retention of functional vegetation corridors.

**Author Contributions:** T.T.N.: Conceptualization, Methodology, Formal analysis, Investigation, Writing—Original draft, Writing—Review and editing. H.E.: Methodology, Resources, Review and editing. P.B.: Supervision, Conceptualization, Methodology, Resources, Review and editing, Funding acquisition. R.H.: Supervision, Conceptualization, Review and editing. B.D.: Supervision, Conceptualization, Methodology, Review and editing, Funding acquisition. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was part of a PhD study by Thi Thu Nguyen made possible by a joint PhD scholarship between Vietnam International Education Development (VIED) and Murdoch University. The costs of the ArborCam Imagery were borne by ArborCarbon Pty Ltd., Perth, Australia.

**Data Availability Statement:** Not applicable.

**Acknowledgments:** Naviin Hardy from ArborCarbon Pty Ltd. provided technical support with the high-quality datasets of the airborne imagery. Thanh Trang Pham from the Vietnam National University of Forestry gave advice on data analysis.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


MDPI St. Alban-Anlage 66 4052 Basel Switzerland Tel. +41 61 683 77 34 Fax +41 61 302 89 18 www.mdpi.com

*Remote Sensing* Editorial Office E-mail: remotesensing@mdpi.com www.mdpi.com/journal/remotesensing

MDPI St. Alban-Anlage 66 4052 Basel Switzerland Tel: +41 61 683 77 34

www.mdpi.com ISBN 978-3-0365-6276-6