1. Introduction
Accelerated urbanization has led to drastic changes in the types of land cover, as well as the physical and chemical properties of soil and water. Consequently, the processes of water vapor and energy exchanges between soil and the atmosphere have shifted, thereby affecting the urban thermal environment system [
1]. In turn, urban thermal environment changes have affected the regional climate, urban hydrology, air quality, and many urban ecological processes, such as material metabolism and the energy cycle. All the above have caused a series of safety problems regarding the ecological environment, thereby seriously affecting the normal production and life of urban residents; they constitute the bottleneck restricting the sustainable development of cities. Urban green space (UGS) can change the air flow and heat exchange through the transpiration of vegetation leaves and solar radiation blockage; they form urban “cold island” centers, thereby improving the urban thermal environment [
2,
3]. Therefore, as living urban infrastructures, UGS is among the most effective, lasting, and economical strategies for mitigating the effects of urban heat islands (UHIs) [
4,
5].
Many scholars have conducted in-depth research on the mitigation of the effects of UHIs through UGSs. Research on the urban thermal environment mainly focuses on the atmospheric temperature (AT) and land surface temperature (LST) [
6]. These parameters, albeit numerically different, correlate well with each other [
7] and can both express the spatial distribution of the urban thermal environment well. AT data are mainly obtained through in situ monitoring, which can have high temporal resolution; however, spatial coverage can be poor due to the usually small number of monitoring stations. Conversely, research on LST is mainly based on LST inversion, based on a thermal infrared band of remotely sensed imagery. This procedure is more time efficient and labor saving than AT data acquisition, and can effectively and comprehensively detect LST characteristics. Surface urban heat islands (SUHIs) based on remote sensing have also become the main focus of current UHI research [
8]. Previous research mainly focused on understanding the relationships among land surface characteristics, vegetation abundance, and the LST [
9,
10]. In recent years, due to the advantages of quantitative remote sensing with regard to landscape patterns, remote sensing studies have paid special attention to the influence of spatial UGS patterns on the LST [
11,
12,
13].
According to the landscape ecology theory, diverse UGS landscape features can be classified as composition and configuration characteristics [
14]. The former refers to the abundance and variety of land cover types, including UGS area, type, vegetation quality, and the latter refers to the spatial layout of land cover types, where the UGS shape and distribution determine the UGS configuration [
1,
15,
16]. Through remote sensing, the relationship of each metric with the LST can be revealed. Previous comprehensive research on the impact of UGSs on the LST preferred regular grids or concentric buffers as analysis units [
14,
17,
18,
19], while performing simple linear regressions and curve fitting on the landscape metrics and LST of each analysis unit and dividing to obtain a global correlation coefficient [
20]. The results they produced were based on the overall correlation of the overall study area, yielding only an overall correlation coefficient. However, urban landscapes are heterogeneous, and the different composition and configuration of the landscape in each area of the city affects ecological processes. For example, it has been demonstrated that urban heterogeneity affects urban ecological resilience and sustainability [
21,
22]. Urban heterogeneity may lead to different UGS cooling effects in different spatial units [
18,
23]. Furthermore, some scholars have previously concluded that the cooling effect of UGS shapes in different cities has the opposite effect. For example, Zhou et al. (2017) found a better cooling effect for complex shapes in Sacramento, but they also found a better cooling effect for simple UGS shapes in Baltimore; they attributed this contradictory finding to the different background climatic conditions [
24]. Their study likewise calculated a global correlation coefficient for the entire city. Therefore, we analyzed the cooling effect of UGS in different areas of a city from the perspective of urban heterogeneity. We obtained positive and negative results for different areas of a city by calculating the local correlation between UGS and the LST in different areas, which helps us to enhance our understanding regarding the effects of UGS spatial configuration on UHIs.
In addition, related studies have also shown that spatial autocorrelation may affect the relationships of landscape metrics with the LST [
24,
25]. Geographically weighted regression (GWR) is a new method for modeling spatial heterogeneity [
26]; based on the GWR model, analyzing the relationship between UGSs and the LST in different regions becomes feasible. Although the application of the GWR model is common, we applied the GWR model to analyze the correlation between the LST and UGS from the perspective of spatial heterogeneity based on Landsat data. We combined the cold- and hot-spot analyses to visualize the cooling effect of spatial inhomogeneity of UGS and to obtain a good explanation. This series of methods constitutes a new approach for studying the correlation between UGS and LST.
In this study, we use Landsat 8 imagery for UGS extraction from within the sixth ring road of Beijing and retrieve the LST based on its thermal infrared band. On the macroscopic scale, from the perspective of spatial autocorrelation and heterogeneity, we applied the GWR method to comprehensively consider the UGS cold source intensity and the different UGS cooling effects under different ecological and artificial environments in order to better guide UGS planning and construction.
2. Methodology
2.1. Study Area
Beijing (39.4–41.6° N, 115.7–117.4° E) is a mega city and one of China’s national central cities, i.e., a national political, economic, and cultural center. It was classified as an Alpha+ city (i.e., a highly integrated city fulfilling advanced service needs) by the globalization and world cities research network (GaWC) in 2020. By 2020, the built-up area of Beijing had reached 1469 km2 with a permanent population of more than 21 million. The GDP was 3537.1 billion yuan, the annual household electricity consumption for urban and rural residential was 20,494.47 million kWh, the total gas supply of natural gas was 19,243.47 million m3, the total gas supply of liquefied petroleum gas was 434,286 tons, and the amount of anthropogenic heat sources was immense. Therefore, the effect of UHIs has become exceptional. In recent years, in order to become an international first-class harmonious and livable city, Beijing has paid considerable attention to ecological and greening constructions, with the greening rate of urban areas reaching 48.5% in 2020. Our study area is the urban area within the sixth ring road of Beijing, which reflects the rapid urbanization of Beijing. This area is also the most densely populated area of the city, and with its greening construction reaching certain benchmarks, it is an ideal area for analyzing the role of UGSs on UHI mitigation.
2.2. Data Sources
From the USGS website (
https://earthexplorer.usgs.gov), accessed on 20 March 2022, we downloaded a Landsat 8 OLI/TIRS image of Beijing (path: 123; row: 32; cloudiness: 0.2%; resolution: 30 m). It was acquired on 20 September 2020 and was used to extract UGS patches, as well as to perform LST inversion using its band 10. Due to the low resolution of Landsat 8 images, the UGS extraction is coarse and prohibits the precise determination of land cover types, such as forests, shrubs, grasslands, and croplands. In this study, such land cover types were included in the UGSs when extracting them. For this reason, we also utilized the 10-m land cover data for 2020 published by the ESA (
https://viewer.esa-worldcover.org, accessed on 20 March 2022), which include a subdivision of the land cover into 11 types, such as forest, shrubland, grassland, cropland, built-up, bare vegetation, water bodies, etc., with an official accuracy of 74.4% [
27]. To ensure higher classification accuracy, we used visual interpretation for correction, and used its more detailed land-cover classification (i.e., forest, shrubland, grassland, and cropland) for proving the rationale behind categorizing all forest, shrubland, grassland, and cropland patches as UGS to establish a GWR model.
Based on the convenience of obtaining Landsat data, we mainly study the extraction of UGS using Landsat data, while the 10-m land cover data for 2020 published by the ESA is mainly used to verify and analyze the general pattern of the different cooling effects of forests, shrubs, and grasses in UGSs. Although it has finer classification accuracy, it only has data for 1 year, 2020. Therefore using Landsat data can extend our approach into other years and can ensure data homogeneity based on Landsat data for LST retrieval and extraction of UGS patches.
2.3. Extraction of UGS Patches
After preprocessing, i.e., applying spatial and atmospheric corrections, since the original Landsat 8 image was of a single band, it was necessary to combine multiple single-band images into a multi-band color image to improve the accuracy of land classification; however, there was considerable redundant information between bands. To reduce mutual interference, bands with large amounts of information were selected for information extraction, while the correlation between them was the smallest. In this study, we used a principal component analysis, a band correlation analysis, and a normalized difference vegetation index (NDVI) calculation based on ENVI5.3. Subsequently, we determined the optimal band combination according to the correlation analysis of various band data, including each band, each principal component, and the NDVI to increase image interpretation accuracy.
Table 1 shows that the principal component 1 (PC1) in the image had the largest amount of information; therefore, 7 bands of the image, as well as the generated bands NDVI and PC1, were selected for correlation analysis and we determined the optimal band combination. The optimum index factor (OIF) proposed by Chavez in the United States to determine the most informative 3-band combination was used as basis for selecting the band combination [
28,
29,
30]. The formula is as follows:
where
Si is the standard deviation of band
i, and
Rij is the correlation coefficient between band
i and band
j (see
Table 2). The larger the OIF, the greater the amount of information contained in the 3-band combination, the higher the independence between bands, and the smaller the information redundancy.
Finally, bands 4, 5, and PC1 were selected as the optimal band combination. In order to greatly improve interpretation accuracy, the images were fused with the Landsat 8 panchromatic band after band combination (resolution: 15 m). The fused image is shown in
Figure 1; the image clarity is greatly improved, fully meeting the classification requirements.
The supervised classification tool based on ENVI5.3 was used for classifying the processed images. As already mentioned, because of the low Landsat 8 resolution, we divided land cover into 3 types (impervious surfaces, water bodies, and UGS) which included forests, shrubs, grasslands, and croplands. Based on the macroscopic scale analysis, the 3 land cover types met the research requirements (see
Section 3.3.1 for the rationality of this classification). The classification results are shown in
Figure 2a. In this study, urban forests, shrubs, grasslands, and croplands were all classified as UGS, and the kappa coefficient was 0.96, thereby meeting the research requirements.
2.4. LST Retrieval
The LST is the radiative skin temperature of the land derived from infrared radiation [
31]. Bands 10 and 11 of Landsat 8 are thermal infrared bands and can be used for LST retrieval. The method has been fully verified and used by scholars and experts internationally. In this paper, we selected band 10. First, the digital number (DN) of band 10 was converted to spectral radiation; second, the spectral radiation was converted to the brightness temperature of the sensor using Plank’s law; and third, the surface brightness temperature was converted into the actual surface temperature by estimating the surface radiance. We refer to the literature for details on the specific inversion method [
20,
32]. The LST map is shown in
Figure 2b.
2.5. Calculation of Landscape Metrics Based on the Moving Window Method
According to previous studies, the cooling effect of UGS has a scale effect [
33] and is influenced by landscape metrics, such as UGS area, shape, and distribution. Here, we calculated the UGS landscape metrics based on the moving window method of FRAGSTATS v4.2. As the cooling effect of UGS has a scale effect, the choice of moving window size is crucial. If the window is too small, large UGS patches will be cut, thereby affecting the estimation of the cooling effect of UGS; if the window is too large, the cooling effect of UGS patches will also be greatly weakened, resulting in insufficient samples [
34].
We therefore selected the percentage of UGS (PLAND) index in FRAGSTATS v4.2 as the basis for window size selection. As the image resolution was 30 m, a multiple of 30 was selected for window size according to the software characteristics, thereby ensuring calculation accuracy. According to UGS extraction, the average area of UGS patches was 62,032 m
2 (i.e., approximately 249 × 249 m), while 99.8% of UGS patches were smaller than the 1800 × 1800 m window size. Therefore, we selected the window sizes of 300 × 300, 600 × 600, 900 × 900, 1200 × 1200, 1500 × 1500, and 1800 × 1800 m, and the final window sizes were determined based on the correlation analysis results of their respective PLANDs and LSTs.
Figure 3 shows that the 1500 × 1500 m window size exhibited a better fit between PLAND and LST, which better reflected the cooling effect of UGS at this scale, while the sample was sufficient. The number of samples in the 1800 × 1800 m window was slightly smaller.
2.6. Statistical Analysis Using Moran’s I and Getis-Ord Gi*
The purpose of using Moran’s I and Getis-Ord Gi* was to find the spatial location of cold and hot spots in the study area and to evaluate the spatial LST distribution pattern. Both indices were realized by the geographic statistical analysis extension module of ArcGIS [
35].
Moran’s
I is the most commonly used index for analyzing spatial autocorrelation. The formula is as follows:
where
I is Moran’s
I,
n is the number of 1500 × 1500 m windows, and
xi and
xj are the average temperatures of the
i and
j windows, respectively.
is the average temperature of all windows, and
wij is the weight of window
i relative to window
j. The inverse distance weighted method is used here to generate.
Getis-Ord Gi* is a local spatial autocorrelation index based on the total distance matrix for detecting the locations of high or low values of each element clustering in space and has strict statistical test standards. The formula is as follows:
where
wij is the spatial weight matrix between window
i and
j,
xj is the average temperature of window
j, and
n is the total number of windows. Using the statistical results of Getis-ord Gi*, the region can be divided into 3 different temperature zones: hot spot, cold spot, and insignificant regions.
2.7. GWR Analysis Based on LST and UGS
Landscape composition and configuration play important roles regarding the spatial LST distribution in an area; however, due to different spatial planning contexts and ecological backgrounds in cities, the influence of the UGS landscape composition and configuration on the LST cooling effect can vary in different areas [
20,
36]. Therefore, we used the GWR method for investigating the degree of influence of each UGS landscape metric on the cooling effect in different locations within the study area.
GWR is a local regression technique that considers spatial heterogeneity. The model has the form:
where Y
i denotes the LST,
Xji represents the landscape metrics, k is the total number of spatial units involved in the analysis, ε
i denotes the random error term, (
ui, vi) indicates the spatial location of sample
i,
β0 (
ui,
vi) is the intercept at the location
i, and
βj (
ui,
vi) is the local estimated coefficient of the independent variable
Xji. The Gaussian function was used to determine the weight, and the Akaike information criterion (AIC) method was used to determine the optimal bandwidth [
37].
Based on FRAGSTATS v4.2, we used the 1500 × 1500 m windows and screened them out by the above moving window analysis to analyze the UGS landscape pattern within the sixth ring road of Beijing. Each window is an analysis unit represented in the above formula.
To better and more accurately explore the comprehensive influences of LST and various UGS landscape metrics, we selected as many commonly used landscape metrics from the class level as possible to characterize the UGS landscape pattern. Landscape metrics are selected as shown in
Figure 4a. These landscape metrics covered all aspects of green landscape pattern, and the LST was used as the dependent variable for establishing a GWR model. Before the regression analysis, each variable was normalized to eliminate the influence of dimension on the regression analysis [
20]. However, as shown in
Figure 4a, each landscape metric was usually correlated with the others, and the inclusion of various landscape metrics in the model led to multicollinearity. This is also the drawback of selecting landscape metrics in past studies, where there is no standard method, but the landscape metrics are selected empirically for correlation analysis separately, which leads to high collinearity among the selected indices, resulting in some meaningless conclusions while obtaining consistent results. Therefore, we used the RStudio regsubsets function and selected the best subsets of calculated regressions to filter the landscape metrics. As shown in
Figure 4b, when landscape metrics were 8, the adjusted R
2 of the subset regression equation would maximize. Nevertheless, in order to minimize the number of selected landscape metrics, we assumed that 4 landscape metrics were sufficient for the adjusted R
2 to approximate the maximum, as choosing more landscape metrics would be of little significance. The final selected landscape metrics were PLAND, LSI, SHAPE_MN, and DIVITION. PLAND represented landscape composition, and the other metrics represented landscape configuration.
After screening the metrics, the ordinary least squares regression (OLS) model was used for further verifying the collinearity and significance of the selected landscape metrics based on ArcGIS. OLS is the most widely used regression model and a suitable starting point for all spatial regression models. After analysis, DIVITION, i.e., a metric with high collinearity, was excluded, while PLAND, SHAPE_MN, and LSI were selected. PLAND represents the UGS proportion in the 1500 × 1500 m window and can characterize the UGS landscape composition in the study area. SHAPE_MN represents the shape complexity of regional UGS patches. LSI is similar to SHAPE_MN, albeit representing the complexity of landscape shape constructed by all UGS patches within the window, and it represents the patch shape irregularity of the overall landscape. In Fragstasts4.2, it is a patch aggregation metric, reflecting the degree of aggregation of patches. These 3 metrics were able to represent the spatial UGS pattern characteristics. In addition, to better explain the interpretation ability of UGS landscape metrics on LST and to avoid the influence of the water-cooling effect, the PLAND of water bodies was also added as an independent variable. All the selected independent variables are shown in
Table 3. Finally, a GWR analysis was conducted based on ArcGIS, thereby obtaining the spatial distribution of local regression coefficients for each landscape metric (see
Section 3.4). As shown in
Table 4, by comparison with OLS models based on global parameter estimations, the GWR model can effectively deal with spatial non-stationary phenomena in regression analysis and can better quantify the relationships between landscape patterns and LST, having higher explanatory power and reliability [
38,
39].
5. Conclusions
In this paper, we proposed a new approach to investigate the spatially heterogeneous cooling effects of UGSs. Based on the analysis, we determined a 1500 × 1500 m grid as the analysis cell and optimized the selection of landscape metrics to establish the GWR model.
We firstly analyzed the impact of the patches’ own characteristics on the cooling effect, and then analyzed it at the regional scale. The conclusions are as follows: (1) the areas with strong influence of each landscape metrics on LST are basically consistent with the hot LST spots; (2) the cooling effect of various UGSs is according to the following pattern: forest > shrub > cropland >grass; (3) PLAND has the largest contribution to the cooling effect of LST, followed by LSI, and SHAPE_MN shows a positive correlation; (4) the internal LST would not change significantly if the UGS patch area was larger than a certain range; and (5) the cooling effect of regular UGS patches is better than that of complex shapes, and the denser the patches are, the better the cooling effect is.
These conclusions would help to explain the distribution characteristics of regression coefficients generated by the GWR model. The distribution of regression coefficients could visualize the cooling effect of spatial heterogeneity of UGS. The approach can clearly indicate the degree of influence of the spatial pattern of UGS on LST in each part of the study area, which can enhance the understanding on the effects of spatial configuration of UGSs on UHIs and better guide the future planning and construction of UGSs.