How Do Two- and Three-Dimensional Urban Structures Impact Seasonal Land Surface Temperatures at Various Spatial Scales? A Case Study for the Northern Part of Brooklyn, New York, USA

Abstract

Identifying the driving factors of urban land surface temperatures (U-LSTs) is critical in improving urban thermal environments and in supporting the sustainable development of cities. Previous studies have demonstrated that two- and three-dimensional (2D and 3D) urban structure parameters (USPs) largely influence seasonal U-LSTs. However, the effects of 2D and 3D USPs on seasonal U-LSTs at different spatial scales still await a general explanation. In this study, we used very-high-resolution remotely sensed data to investigate how 2D and 3D USPs impact seasonal U-LSTs at different spatial scales (including pixel and city block scales). In addition, the influences of various functional zones on U-LSTs were analyzed. The results show that, (1) generally, the links between USPs and U-LSTs at the city block scale were more obvious than those at the pixel scale, e.g., the Pearson correlation coefficient (r) between U-LST and the mean building height at the city block scale (summer: r = −0.156) was higher than that at the pixel scale (summer: r = −0.081). Tree percentage yielded a considerable cooling effect on summer U-LSTs on both the pixel (r = −0.199) and city block (r = −0.369) scales, and the effect was more obvious in regions with tall trees. (2) The independently total explained variances (R²) of 3D USPs on seasonal U-LSTs were considerably higher than those of 2D USPs in most urban functional zones (UFZs), suggesting the distinctive roles of 3D USPs in U-LST regulation at the local scale. Three-dimensional USPs (R² value = 0.66) yielded more decisive influences on summer U-LSTs than 2D USPs did (R² value = 0.48). (3) Manufacturing zones yielded the highest U-LST, followed by residential and commercial zones. Notably, it is found that the explained variances of the total study area for seasonal U-LSTs were significantly lower than those of each UFZ, suggesting the different roles of 2D and 3D USPs played in various UFZs and that it is critical to explain U-LST variations by using UFZs.

1. Introduction

Urbanization creates a well-known phenomenon called urban heat island (UHI, acronyms used throughout the manuscript are listed in Appendix A,

Table A1. List of abbreviations.

Abbreviations	Definition
U-LSTs	Urban land surface temperatures
2D	two-dimensional
3D	three-dimensional
USPs	Urban structure parameters
UHI	Urban heat island
LiDAR	Light Detection and Ranging
LCZ	Local climate zone
UFZ	Urban functional zone
DSM	Digital surface model
NYCDTT	New York City Department of Information Technology and Telecommunications
USGS	United States Geological Survey
OLI	Operational Land Imager
TIRS	Thermal Infrared Sensor
OSM	Open Street Map
RF	Random Forest
BC	Building coverage
ISC_G	Impervious surface coverage at ground-level
TP	Tree percentage
GP	Grass percentage
MBH	Mean building height
MBV	Mean building volume
FAI	Frontal area index
CAR	Canyon aspect ratio
FAR	Floor area ratio
SVF	Sky view factor
MTH	Mean tree height
DI	Distribution Index
MSE	Mean Square Error
MBI	Morphological building index
MSI	Morphological shadow index
UCI	Urban complexity index
GLCM	Gray-level Co-occurrence Matrix
HOM	Homogeneity
CON	Contrast
ENT	Entropy
NDVI	Normalized difference vegetation index
NDWI	Normalized difference water index
R	Red Band
G	Green Band
B	Blue Band
NIR	Near-Infrared Band
Man.	Manufacturing zone
Com.	Commercial zone
Res.	Residential zone
M&R	mixed industrial and residential zones
Is	the importance of the variable for summer U-LST
Iw	the importance of the variable for winter U-LST

). This phenomenon refers to urban regions experiencing higher temperatures than their surrounding rural areas [1]. Generally, UHI effects negatively impact eco-environments and people’s health, such as air pollution, energy consumption, and greenhouse gas emissions. These impacts can be exacerbated during summer heatwaves [2,3,4]. The United Nations points out that, by 2050, the urban population will reach 66% in the world [5], which means more people will be exposed to heatwaves. Therefore, mitigating UHIs becomes a significant challenge to achieve global sustainable development.

Both atmosphere temperature and land surface temperature (LST) are widely used to measure the intensity/magnitude of the UHI [6]. Atmospheric temperature is often measured by meteorological stations [6], whereas LST can be defined by satellite-derived radiative temperature [7,8,9,10]. The former is limited by the few meteorological stations and thus cannot reveal the spatial variations in urban temperature; in contrast, the latter features good spatial coverage. Moreover, the urban LST (U-LST) can regulate the lower-layer air temperature of the underlying surfaces and can be regarded as a critical indicator of surface energy exchanges, local climates, and human comforts [11]. Two-dimensional (2D) and three-dimensional (3D) urban structure parameters (USPs) can significantly influence the urban surface energy/water balance and can thus determine the variations in U-LSTs [12]. Currently, Studies that reveal how 2D and 3D USPs affect U-LSTs have been widely reported. For example, Lu et al. [6] analyzed the influence of impervious surfaces and vegetated areas on LSTs in Xian, China, based on polygonal grids at different spatial scales. They found that the links between normalized difference impervious surface index and U-LST increased with the rising spatial scales. Recently, given that the use of Light Detection and Ranging (LiDAR) point clouds data brings the possibility to obtain high-resolution 2D and 3D USPs [13], Yu et al. [14] explored the relationships between 3D landscape patterns and LSTs over an area of approximately 665.637 km² in central Shanghai, China, using the extreme Gradient Boosting tree regression method. They found that 3D USPs yield higher explanatory capabilities for U-LSTs than 2D USPs. Additionally, Imhoff et al. [7] investigated surface UHIs in 38 of the most populated cities in the continental US. They found that urban impervious surface area and vegetation can largely explain the surface UHI variations. These studies undoubtedly provide critical insight into U-LST mitigation and a better understanding of the influences of 2D and 3D USPs on urban thermal environments. However, how 2D and 3D USPs impact U-LSTs at different scales still awaits a holistic investigation.

To further reveal the influences of 2D and 3D USPs on seasonal U-LSTs, Bechter et al. [15,16] provided a new urban classification scheme—Local Climate Zone (LCZ)—to reveal how U-LSTs vary with different local climate zones [17]. Nevertheless, the classification of cities is involved in the underlying surface heterogeneity and is closely related to human activities (e.g., settlement, production, and recreation) [18]. Moreover, those underlying surfaces and human activities can significantly impact regional energy/water balance. With this in mind, Urban Functional Zone (UFZ) may explain the spatiotemporal patterns of U-LSTs. Urban Functional Zone incorporates different urban ecological environments with significant differences in urban functions, human activities, and energy consumption [19] and can be regarded as one of the basic units of urban planning and management [20]. Therefore, the impact of various UFZs on seasonal U-LSTs must be further explored.

Based on the above discussion, we selected the northern part of Brooklyn, New York, USA, as the study area and investigated the influences of 2D and 3D USPs on seasonal U-LSTs using a multi-scale analysis (including pixel and block scales). The questions to be addressed include the following:

(1)

How do 2D and 3D USPs impact seasonal U-LSTs at different spatial scales (including both city block and pixel)?

(2)

Which USPs (2D or 3D) yield a relatively strong influence on seasonal U-LSTs?

(3)

How do UFZs affect the variations in seasonal U-LSTs?

2. Study Area and Data

2.1. Study Area

The study area lies in the northern part of Brooklyn, New York, USA (Figure 1). According to the Köppen climate division, the area experiences a humid subtropical climate (Cfa) [21] and features plentiful precipitation all year, with approximately 1300 mm rainfall per year [22]. It covers about 6.15 km², with nearly 8000 buildings, 7098 parcels, and 496 blocks (Figure 1a). The area’s Digital Surface Model (DSM) changes from −4.85 to 133.70 m (Figure 1b). According to the New York City Department of Information Technology and Telecommunications [23], UFZs in the area include manufacturing, commercial, residential, and park zones (Figure 1c). The mean building heights in the study area are generally lower than 90 m (Figure 1d).

2.2. Data

2.2.1. LiDAR Point Clouds

LiDAR point cloud data were acquired from the New York City Department of Information Technology and Telecommunications (NYCDTT) [23] (

Table 1. Data used in the study.

Data	Type	Resolution	Time	Usage
Remote sensing data	LiDAR point cloud	1.5 m (5.0 ft)	5/2017	Land cover mapping and surface roughness retrieval
	High-resolution orthophoto	0.3 m (1.0 ft)	2017	Land cover mapping
	Landsat 8 remotely sensed data	Multispectral data: 30 m; TIRS data: 100 m	03/19/2015 10:39:26; 04/29/2015 10:32:56; 05/22/2015 10:38:55; 06/07/2015 10:39:05; 07/25/2015 10:39:28; 08/26/2015 10:39:39; 10/06/2015 10:33:42; 11/23/2015 10:33:50; 01/23/2015 10:33:38	LST retrieval
Geographical information data	Land-lot data	Vector	2019	Image segmentation and label optimization
	Road data	Vector	2017	Block data extraction

). Two Leica systems were mounted on a Cessna 402C or Cessna Caravan 208B aircraft to produce LiDAR point clouds with a point density of higher or equal to 8.0 points/m² in May 2017. The data’s horizontal and vertical position accuracies were lower than 42 cm and about 12.1 cm, respectively [24]. Specifically, we used “StatisticalOutlierRemoval” filter operation of Point Cloud Library 1.6 to remove the anomaly and outlier points. Additionally, a voxel grid filter was adopted to reduce the redundant points. Furthermore, we used an interpolation algorithm, i.e., binning approach, to generate accurate DSM, which was used for surface roughness retrieval and land cover mapping (Figure 1e).

2.2.2. Landsat Data

Landsat 8 remotely sensed images were acquired from the United States Geological Survey (USGS) website (http://earthexplorer.usgs.gov/, accessed on 8 November 2020). Landsat 8 satellite carries a land imager (Operational Land Imager, OLI) and a thermal infrared sensor (Thermal Infrared Sensor, TIRS), used for U-LST retrieval. Given that the UHI intensity is more intense during the daytime [25], we retrieved seasonal daytime U-LSTs covering the study area in spring (March to May), summer (June to August), fall (October and November), and winter (January to February). Details of the data acquisition time concerning Landsat 8 images can be found in Table 1.

2.2.3. Supplementary Data

The road data, acquired from Open Street Map (OSM), were used to extract city blocks in 2017. The land-lot data containing UFZ attributes obtained from the New York City Department of Information Technology and Telecommunications (NYCDITT) in 2017 were used to extract the UFZs (Table 1). Additionally, the high-resolution orthophotos, released by the NYCDITT, provide rich spectral information for the land cover classification.

3. Methods

3.1. U-LST Retrieval

The U-LST was retrieved using the radiative transfer method (i.e., high correction accuracy in the atmospheric correction method) based on the Landsat 8 TIRS data [26,27]. The expression of the thermal infrared radiance value L received by the satellite sensor can be defined as follows:

$$L_{\lambda }=\left[\epsilon B \left(T_S\right) \left(1-\epsilon \right) L\downarrow \right] \tau +L\uparrow$$

(1)

where ε is the land surface emissivity, T_S is the actual surface temperature, B(T_S) is the thermal radiance of the blackbody at T_S obtained from the Planck formula, and τ is the atmospheric transmittance of the corresponding thermal infrared band. The radiance B(T_S) of a blackbody in the thermal infrared band at temperature T is as follows:

$$B \left(T_S\right)=\frac{L_{\lambda }-L\uparrow -\left(1-\epsilon \right)L\downarrow }{\tau \epsilon }$$

(2)

T_S can be defined as a function of the Planck formula:

$$T_S=K_2 / \ln \left(K_1/B \left(T_S\right)+1\right)$$

(3)

where K₁ = 774.89 W·m⁻²·μm⁻¹·sr⁻¹ and K₂ = 1321.08 K. In this way, we retrieved U-LSTs covering the study area in different seasons.

3.2. Land Cover and UFZ Mapping

We extracted different features, including morphological features, urban complexity indices, geometric features, and spectral features, to perform the land cover classification [18,28,29,30,31,32,33,34,35,36,37] (details can be found in Appendix A,

Table A2. The multiple features considered in this study.

Category	Features	Principle	Reference
	Morphological Building Index (MBI)	The MBI was used to extract building by stablishing the relationship between implicit features and morphological operators of buildings.	[32]
Morphology features	Morphological Shadow Index (MSI)	The MSI is defined as the dual function of the MBI, i.e., the black top-hat morphological profiles, to highlight the shadow structures.	[33]
	Urban Complexity Index (UCI)	The UCI is constructed on the basis of 3D-WT, where the spatial variation of natural features is relatively smaller than the spectral variation but, in urban areas, shows more variation in the spatial domain.	[34]
Geometric features	Flatness	Flatness is obtained from DSM and refers to the flatness of non-ground points. Generally, the surface of buildings is flatter than vegetation.	[37]
	Vnd	The normal vectors of vegetation are more scattered and irregular, but are fixed in several directions of buildings.	[29]
	nDSM	DSM contains the land surface information.	[28]
Textural feature	Gray-level Co-occurrence Matrix (GLCM)	This is generated from LiDAR. In the height image, the vegetation is more textured than the buildings.	[31]
	Homogeneity	This was generated from high-resolution orthophotos in 2017. The texture features of the study area were acquired based on GLCM.	[18]
	Contrast	This was generated from high-resolution orthophotos in 2017. The texture features of the study area were acquired based on GLCM.	[18]
	Entropy	This was generated from high-resolution orthophotos in 2017. The texture features of the study area were acquired based on GLCM.	[18]
Spectral feature	Wave Spectrum of High-Resolution Orthophotos	Red Band (R), Green Band (G), Blue Band (B), and Near-Infrared Band (NIR)	[30]
	Normalized Difference Vegetation Index (NDVI)	NDVI = (NIR - R) / (NIR + R)	[35]
	Normalized Difference Water Index(NDWI)	NDWI = (G - NIR) / (G + NIR)	[36]

). Based on the spectral response of features on high-resolution orthophotos, we identified six land covers, i.e., buildings, impervious ground surfaces, trees, grasses, water bodies, and bare soils.

We selected the random forest (RF) classifier to perform the land cover classification. The RF classifier is essentially an improvement of the decision tree algorithm by combining several decision trees, where the creation of each tree depends on an independently drawn sample. Moreover, each tree in the forest has the same distribution and the classification error depends on the classification power of each tree and the correlation between them. The results of the multi-classification model are obtained by combining the majority voting principle of integrated learning with higher accuracy and generalization performance. Thus, the random forest model produces better classification results than a single tree by randomly generating many decision trees and effectively reduces the risk of overfitting associated with decision tree algorithms.

In particular, the multi-resolution segmentation method was first utilized to produce segmented objects. Then, we calculated the features of the objects and used the RF classifier for classification to label land covers. For the validity and reliability of the classification results, we randomly selected 146,074 samples and utilized the 10-fold cross-validation method to train and validate the model. The overall accuracy of the land use classification results was 87% (for details of the evaluation regarding land cover classification, see Appendix A,

Table A3. Evaluation of land cover classification using the RF algorithm.

Type	Precision	Recall	Fl-Score	Support
Building	0.93	0.91	0.92	9143
Tree	0.82	0.82	0.82	3801
Grass	0.80	0.84	0.82	2723
Bare soil	0.82	0.87	0.85	2830
Impervious ground surface	0.84	0.83	0.83	8780
Water body	0.96	0.99	0.98	1938
Accuracy			0.87	29215
Macro avg	0.86	0.88	0.87	29215
Weighted avg	0.87	0.87	0.87	29215

). Therefore, the classification results were reliable in supporting the subsequent 2D and 3D USP extraction. The boundaries of the city block were delineated using road data. The land-lot data was used to label the blocks’ UFZ attributes, and we merged those UFZs with the same attributes to obtain the final UFZ data (Figure 1c).

3.3. Extraction of 2D and 3D USPS

Based on the land use classification results, we extracted 11 USPs to reveal their impacts on U-LSTs.

Table 2. Two- and three-dimensional USPs considered in the study.

Category	Parameter	Abbreviation	Meaning	Reference
3D metric	Mean building height	MBH	Average of building height in a spatial unit (30 m × 30 m)	[39]
	Mean building volume	MBV	Average of building volume in a spatial unit (30 m × 30 m)	[39]
	Frontal area index	FAI	Frontal area divided by the area of the spatial unit (30 m × 30 m)	[40]
	Canyon aspect ratio	CAR	Urban canyon height divided by the road width (30 m × 30 m)	[41]
	Floor area ratio	FAR	Ratio of building’s total floor area to the area in a spatial unit (30 m × 30 m)	[40]
	Sky view factor	SVF	Sky view factor influenced by buildings (30 m × 30 m)	[38]
	Mean tree height	MTH	Average of tree height in a spatial unit (30 m × 30 m)	[39]
2D metric	Building coverage	BC	Percentage of build area in a spatial unit (30 m × 30 m)	[42]
	Impervious surface coverage at ground-level	ISC_G	Percentage of impervious surface area at ground-level in a spatial unit (30 m × 30 m)	[42]
	Tree percentage	TP	Percentage of tree area in a spatial unit (30 m × 30 m)	[42]
	Grass percentage	GP	Percentage of grass area in a spatial unit (30 m × 30 m)	[42]

shows the definitions of USPs selected in the study. Those parameters can be divided into two categories, i.e., 2D and 3D USPs (see Appendix A, Figure A1). The former includes building coverage (BC), impervious surface coverage at ground level (ISC_G), tree percentage (TP), and grass percentage (GP). The latter includes mean building height (MBH), mean building volume (MBV), frontal area index (FAI), canyon aspect ratio (CAR), floor area ratio (FAR), sky view factor (SVF), and mean tree height (MTH). Following Reference [38], the SVF was calculated based on DSM by setting 32 directions and a search radius of 80.

3.4. Analyses of the Influences of UFZs and USPS on U-LSTs

3.4.1. Influences of UFZs on U-LSTs

We used the distribution index (DI) to quantify the impacts of different UFZs on the urban thermal environments [43]:

$$\mathrm{DI} =\frac{S_{hi}}{S_i}/\frac{S_h}{S}$$

(4)

where S_hi and S_i are the areas of high-temperature regions and the functional zone i, respectively. S_h and S are the areas of the high-temperature regions and the study area, respectively. If the DI value is higher than 1, the proportion of high-temperature area in the functional zone i is higher than that of the study area, indicating that functional zone i significantly influence urban thermal environments [12].

3.4.2. Statistical Analysis

The correlations between USPs and U-LSTs were quantified using Pearson correlation analysis at pixel and city block scales [18]. To explore the relative importance of 2D and 3D USPs to seasonal U-LSTs, we introduced RF variable importance measure method to estimate the significance of each parameter to seasonal U-LSTs. RF regression analysis uses the impurity, expressed as Mean Square Error (MSE) in the regression, or out-of-bag (OOB) error to quantify the importance of variance to the model. Compared with the OOB error, impurity saves time by generating noise-free data when evaluating each feature. To better explain the final model, we used the MSE impurity to quantify the importance of 2D and 3D USPs for seasonal U-LSTs. The MSE can be defined as follows:

$$\underset{A,s}{MSE=\underbrace{min}}\left[\underset{e_1}{\underbrace{min}}{\displaystyle \sum }_{x_i\in D_1\left(A,s\right)}{(y_i-e_1)}^2+\underset{e_2}{\underbrace{min}}{\displaystyle \sum }_{x_i\in D_2\left(A,s\right)}{(y_i-e_2)}^2\right]$$

(5)

where $e_1$ and $e_2$ are the sample output means of the D₁ and D₂ data sets, respectively. A lower MSE means that the feature more excellent discrimination on the current training data. For the importance of feature j at node m, i.e., the amount of change in MSE after branching of node m is as follows:

$$VI{M}_{jm}^{\left(mse\right)}=MS{E}_m-MS{E}_l-MS{E}_r$$

(6)

where $MS{E}_m$ is the MSE before branching; $MS{E}_l$ and $MS{E}_r$ represent the MSE of the two new nodes before and after branching, respectively. If feature j appears in the decision tree i at the set of node M, the importance of feature j in the ith tree is ${{\displaystyle \sum }}_{m\in M}VI{M}_{jm}^{\left(mse\right)}$, assuming that there are n decision trees in RF. Thus, the total MSE variation of feature j is as follows:

$$VI{M}_j^{\left(mse\right)}={{\displaystyle \sum }}_{i=1}^n{{\displaystyle \sum }}_{m\in M}VI{M}_{jm}^{\left(mse\right)}$$

(7)

The value of feature j after normalization of its importance is noted as the importance of feature j, which is calculated as follows:

$$I=\frac{VI{M}_j}{{{\displaystyle \sum }}_{i=1}^{c}VI{M}_i}$$

(8)

where I represents the importance of feature j after normalization; note that we used Is and Iw to describe the variable importance for summer and winter U-LSTs, respectively. $VI{M}_j$ is the MSE of feature j, and ${{\displaystyle \sum }}_{i=1}^{c}VI{M}_i$ is the sum of MSE differences of all feature. The model inherits the structural visualization characteristics of the tree model, which is convenient for feature understanding and interpretation and reduces the risk of overfitting brought by the decision tree algorithm. Thus, the model can provide a better understanding of feature importance for seasonal U-LSTs. In addition, we quantified the interpretation of U-LST by 2D and 3D USPs and expressed it in terms of the total explained variance [44]. In particular, three experiments were designed to reveal the explained variance for U-LST changes: (I) single 2D USPs, (II) single 3D USPs, and (III) the combined effect of 2D and 3D USPs.

4. Results

4.1. Results of LST Retrieval across Different UFZs

We retrieved seasonal U-LSTs in the northern part of Brooklyn in 2015 (Figure 2a–d). Figure 2a–d show that, despite the variations in seasonal U-LSTs, consistent influences of UFZs on U-LSTs were observed for different seasons. Manufacturing zones yielded the highest U-LST, followed by residential and commercial zones. In contrast, parks featured the lowest U-LST. Those findings suggest that manufacturing zones are a critical factor in generating UHIs, whereas parks can help the urban cooling. In addition, the seasonal U-LSTs ranking from high to low was summer (26–50 °C) > fall (16–35 °C) > spring (−3–12 °C) > winter (−8–9 °C). Given that the highest U-LST often appears during summer daytime [25], we thus investigated U-LSTs in different UFZs in summer. Specifically, the Natural Break method was utilized to divide the study area into three categories, namely, high-temperature regions (higher than 38 °C), middle-temperature regions (35–38 °C), and low-temperature regions (lower than 35 °C) [45]. Figure 2e shows that the highest DI value (2.87) was observed in manufacturing zones; in contrast, the DI values in the other zones were found lower than 1, suggesting that manufacturing zones yield significant contributions to the high-temperature regions. It is consistent with an investigation of U-LSTs variations in various UFZs by Reference [18]. Figure 2f gives the percentages of different temperature categories in various UFZs. As shown, manufacturing zones yielded the highest and lowest rates of high-temperature regions (40%) and low-temperature areas (13%). The parks, on the contrary, produced the immense contribution of low-temperature regions (47%). It indicates that manufacturing zones have a significant warming effect on U-LST, whereas the parks yield a significant cooling effect.

4.2. Correlation between USPS and U-LSTs

Figure 3 shows the relationships between seasonal U-LSTs and USPs. As shown in Figure 3, at the pixel scale, significant cooling effects of TP were observed in summer (r = −0.199, p < 0.01) and fall (r = −0.143, p < 0.01); however, the links became weak in winter (r = −0.051, p < 0.01) and spring (r = −0.034, p < 0.01). It is found that the correlation coefficients between TP and U-LST at the city block scale were −0.369 (−0.309) in summer (fall), higher than those at the pixel scale (summer: r = −0.199, fall: r = −0.143), indicating that the cooling effects of TP at the city block scale were more evident than those at the pixel scale. Additionally, MTH was observed negatively associated with U-LST in various spatial scales. Notably, the links between MTH and U-LST were observed considerably lower than those between TP and U-LST, suggesting that the cooling effect of tree percentage was higher than that of tree roughness. Complex influences of grass on seasonal U-LST were observed. As shown, at the pixel scale, a negative relationship between GP and U-LST was observed in summer (r = −0.072, p < 0.01) and fall (r = −0.037, p < 0.01); in contrast, a positive relationship was found in winter (r = 0.075, p < 0.01). Few influences of GP on U-LST were noted in spring (p > 0.05). Generally, the influence of GP on U-LST at the city block scale (e.g., summer: r = 0.10) was higher than that at the pixel scale (e.g., summer: r = −0.072). It is noteworthy that GP has a much lower cooling potential than tree coverage, which is in line with the previous findings [18]. Those findings are valuable since they suggest that the urban surface cooling related to trees and grasses should consider the spatial scale effects. At the pixel scale, the warming effects of ISC_G in summer (r = 0.097, p < 0.01), autumn (r = 0.095, p < 0.01), and winter (r = 0.092, p < 0.01) were about two times higher than those in spring (r = 0.045, p < 0.01). However, at the block scale, few influences of ISC_G on U-LST were observed in seasons except for winter. In addition, the warming effects of ISC_G were higher than those at the pixel scale. BC was negatively related to winter U-LST, yet a positive relationship between BC and summer U-LST was observed at pixel and city block scales. Note that a strong positive link between BC and U-LST was observed at the city block scale (r = 0.273, p < 0.01).

A negative correlation between MBH and seasonal U-LSTs was found, indicating that the shadow effect of high-rise buildings influenced the U-LST in this area based on the Landsat 8 remotely sensed daytime data. However, the correlation changed with various spatial scales, i.e., the correlation coefficient at the city block scale was considerably higher than that at the pixel scale. Additionally, MBV yielded a positive correlation (r = 0.031) with summer U-LST at the pixel scale, yet a negative correlation between MBV and U-LSTs in other seasons was observed. Notably, the highest correlation coefficient of MBV was observed in winter (r = 0.206, p < 0.01). At the city block scale, the influence of MBV in winter was found more obvious than that in summer. Moreover, in winter, the cooling effect of MBV was more apparent at the city block scale (r = −0.273, p < 0.01) than that at the pixel scale (r = −0.206, p < 0.01).

Consistent influences of FAI on seasonal U-LSTs were observed at different scales. In addition, the cooling effects of FAI at the city block scale were significantly higher than those at the pixel scale in spring, fall, and winter; however few influences of FAI on summer U-LSTs were noted at both the pixel and city block scales (p > 0.05). FAR was negatively associated with seasonal U-LSTs, and the link at the city block scale (r = −0.183, p < 0.01) was considerably higher than that at the pixel scale (r = −0.050, p < 0.01). At the pixel scale, the cooling effects of CAR on U-LSTs in autumn (r = −0.032, p < 0.01) and winter (r = −0.050, p < 0.01) was higher than those in spring (r = −0.029, p < 0.05) and summer (r = −0.026, p < 0.05). In addition, the cooling effects of CAR at the city block scale were more obvious than those at the pixel scale. A significant positive correlation between SVF and seasonal U-LST was found, and the links at the city block scale (e.g., summer: r = 0.192) were stronger than those at the pixel scale (e.g., summer: r = 0.114). Notably, the correlation coefficient of SVF was remarkably higher than that of ISC_G and BC.

4.3. Relative Importance of USPS for Seasonal LSTs

The RF regression method was used to identify the relatively important factors for seasonal U-LST variations at the pixel scale (

Table 3. Variable importance of USPs revealed by random forest across different UFZs. R² represents the total explained variance of seasonal U-LSTs, and Is and Iw represent the variable importance of USPs on summer and winter U-LSTs. Man., Com., Res., and M&R in panel b represent manufacturing, commercial, residential, and mixed industrial and residential zones, respectively.

		Summer Variable Importance (Is)					Winter Variable Importance (Iw)
		Man.	Com.	Res.	Park	Total	Man.	Com.	Res.	Park	Total
3D USP	MBH	0.179	0.160	0.176	0.108	0.228	0.275	0.217	0.204	0.338	0.228
	MBV	0.104	0.064	0.068	0.074	0.085	0.036	0.038	0.074	0.033	0.104
	FAI	0.099	0.094	0.084	0.024	0.068	0.052	0.036	0.061	0.043	0.068
	CAR	0.017	0.102	0.005	0.099	0.034	0.048	0.072	0.004	0.098	0.055
	FAR	0.097	0.095	0.092	0.033	0.092	0.095	0.126	0.085	0.022	0.096
	SVF	0.195	0.181	0.163	0.150	0.185	0.294	0.184	0.168	0.196	0.224
	MTH	0.121	0.123	0.099	0.120	0.099	0.052	0.094	0.100	0.066	0.073
2D USP	BC	0.135	0.050	0.195	0.152	0.116	0.122	0.119	0.193	0.091	0.108
	ISC	0.049	0.097	0.078	0.145	0.047	0.022	0.073	0.071	0.033	0.029
	TP	0.002	0.034	0.035	0.029	0.039	0.001	0.036	0.025	0.015	0.005
	GP	0.001	0.000	0.006	0.067	0.008	0.003	0.005	0.015	0.065	0.007
	R2	0.515	0.761	0.501	0.925	0.378	0.554	0.712	0.464	0.882	0.366
	I	Low									High

). In particular, we investigated the distinctive roles of both 2D and 3D USPs in summer and winter U-LSTs across four UFZs, including industrial, commercial, residential, and park. As shown in Table 3, in the model involving the total study area, summer and winter U-LSTs were largely impacted by MBH (Is = 0.228, Iw = 0.228). MBH was also found to exert large influences on U-LSTs in the models involving four UFZs. It suggests the cooling effects of the building roughness. Additionally, SVF (Is = 0.185, Iw = 0.224) and BC (Is = 0.116, Iw = 0.108) were highly related to seasonal U-LSTs. U-LSTs increased with the rising SVF and BC in summer whereas U-LSTs increased with the decreasing BC in winter. In manufacturing zones, SVF (Is = 0.195, Iw = 0.294), MBH (Is = 0.179, Iw = 0.275), and BC (Is = 0.135, Iw = 0.122) were observed to be important for the variations in summer and winter U-LSTs. Moreover, the influences of MBH and SVF were more obvious on winter U-LSTs than those on summer U-LSTs; in contrast, an opposite trend was observed for BC. In commercial zones, SVF (Is = 0.181), MBH (Is = 0.160), and MTH (Is = 0.123) yielded relatively higher influences on summer U-LSTs, whereas relatively higher influences of MBH (Iw = 0.217), SVF (Iw = 0.184), and FAR (Iw = 0.126) on winter U-LSTs were observed. In residential zones, BC (Is = 0.195) was the most important factor for summer U-LST variation, followed by MBH (Is = 0.176) and SVF (Is = 0.163); yet in winter, the variable importance ranking from high to low was MBH (Iw = 0.264) > BC (Iw = 0.193) > SVF (Iw = 0.168). In park zones, summer U-LSTs were dominated by BC (Is = 0.152), SVF (Is = 0.150) and ISC (Is = 0.145); however, MBH (Iw = 0.338) and SVF (Iw = 0.196) yielded relatively higher influences on winter U-LSTs.

Figure 4 shows the explained variances of 2D and 3D USPs on summer and winter U-LSTs in different UFZs. Different explained variances were observed across different UFZs and seasons. Generally, the independently explained variances of 3D USPs on seasonal U-LSTs were considerably higher than those of 2D USPs in most UFZs, suggesting the distinctive roles of 3D USPs in U-LST regulation at local scales. For example, the influences of 3D USPs on seasonal U-LSTs were more decisive in manufacturing and residential zones. In commercial zones, 3D USPs (R² value = 0.66) yielded more decisive influences on summer U-LSTs than 2D USPs did (R² value = 0.48), whereas the opposite trend was found in winter. In parks, the effects of 2D USPs (R² value = 0.76) on summer U-LSTs were greater than those of 3D USPs (R² value = 0.67), yet 3D USPs (R² value = 0.83) have observed more decisive impacts in winter. Notably, it is found that the explained variances of the total study area for seasonal U-LSTs were significantly lower than those of each UFZ, suggesting that different roles of 2D and 3D USPs played in various UFZs, and it is critical to explain U-LST variations by using UFZs.

5. Discussion

5.1. Differences in the Effects of Green Infrastructure Parameters on Seasonal U-LSTs between City Block and Pixel Scales

As expected, TP exerted a significant cooling effect on seasonal U-LSTs (Figure 3 and Table 3), consistent with the results of References [46,47,48]. Trees can reduce U-LST primarily via evapotranspiration, can create a casting shade to decrease direct solar radiation received by surfaces, and can thus favor partition solar radiation into latent rather than sensible heat [11,48,49]. In particular, we first found that the cooling effects of TP were more evident at the city block scale than at the pixel scale, suggesting that spatial scales should be taken into consideration for urban greening regulation. Grasses can mitigate U-LSTs in summer and autumn, but the cooling effects of grasses were lower than those of trees. Grasses, featuring high water contents, can cool urban surfaces via transpiration and shading effects. However, given that grasses feature simple hierarchical structures and no vertical structures, the cooling abilities of grasses are deficient compared with that of trees [11,48,49]. In addition, GP yielded a warming effect on winter U-LSTs. The possible reason is that grasses feature a higher specific heat capacity and thus decrease the cooling rate compared with impervious surfaces [50,51,52]. In addition, at the block scale, few influences of GP were found on seasonal U-LSTs. The possible reason is that a small grass percentage was observed relative to other urban elements, e.g., impervious surfaces, at the city block scale, resulting in its limited cooling effects at the block scale. We found that MTH was an essential cooling factor for seasonal U-LSTs and that the impact at the city block scale was more pronounced than that at the pixel scale. In particular, the cooling effects of MTH on U-LSTs were lower than those of TP, suggesting that the mitigation effects of evapotranspiration caused by tree proportion are higher than those of land–air convection caused by tree roughness [53]. Nevertheless, shades may significantly impact U-LSTs more than evapotranspiration because the shading effect can help mitigate LST around trees [18]. Therefore, it is critical to design the proper heights of urban trees in a way that combines the effects of tree shading and evapotranspiration, resulting in an optimal cooling effect.

5.2. Differences in the Effects of Built-Up Infrastructure Parameters on Seasonal U-LSTs between City Block and Pixel Scales

In this study, the effects of eight built-up infrastructure parameters (i.e., ISC_G, BC, SVF, MBH, MBV, FAI, CAR, and FAR) on seasonal U-LSTs at different scales were considered. It was found that ISC_G can increase U-LSTs at the pixel scale. The possible reason is that impervious surfaces (e.g., concrete, cement, and asphalt) often exhibit lower emissivity and higher heat capacity than nature surfaces [54,55]. In addition, impervious and dry surfaces can reduce the evapotranspiration efficiency compared with natural surfaces [56] and can thus produce more sensible heat instead of latent heat [11]. However, the links between ISC_G and U-LST were not evident in seasons except winter (r = 0.111, p < 0.05) at the city block scale, suggesting that regulating impervious surface coverage for U-LST mitigation should consider the effects of spatial scales. BC conducted warming effects on seasonal U-LSTs (except for winter U-LSTs), particularly on summer U-LSTs. The possible reason is that connected buildings impede ventilation in summer, leading to the trapping of heat. In addition, air conditioning systems may be heavily used during a period with high temperatures and poor natural air circulation, thus releasing further heat [56]. In winter, building materials with lower specific heat capacities cool more rapidly than natural surfaces, leading to lower U-LSTs. We found a strong positive correlation between BC and U-LSTs at the city block scale (r = 0.273, p < 0.01). Those suggest that proper design and planning of building coverage for mitigating U-LSTs should consider the effects of spatial scales.

It is found that MBH yielded a significant cooling effect on seasonal U-LSTs at local scales, which is well in line with previous observations [57]. High-rise buildings can cast more shadows and can improve surface roughness, reducing the direct solar radiations received by surfaces and increasing the efficiency of land–air convection. Thus, high-rise buildings can significantly mitigate U-LSTs [58,59]. In addition, we found that the cooling effect of MBH in winter was more evident than that in other seasons. The possible reason is that vegetation transpiration becomes weak in winter and that building roughness becomes the dominant factor for U-LST mitigation. It is noteworthy that MBH did not produce a warming effect in summer, implying that the warming effect caused by building energy consumption may be offset by the cooling effect caused by land–air convection.

Moreover, in summer, the sun-facing side of a tall building, featuring a high albedo, shades the low-albedo natural surfaces behind it [59,60]. MBV was positively correlated with summer LST at the pixel scale (Figure 3), implying that a high building energy consumption can increase the anthropogenic heat release, contributing to the accumulation of U-LSTs [14,61,62]. SVF was found to be positively correlated with U-LST, contributing to high seasonal U-LSTs at various spatial scales, and the warming effects at the city block scale were higher than those at the pixel scale. The possible reason is that a high SVF can lead to more direct solar radiation entering the street canyon and thus increasing U-LSTs [63]. Additionally, multi-scale analyses indicate that the deep street canyon at the city block scale can help the U-LST mitigation. Furthermore, we also investigated the effects of FAI, FAR, and CAR on U-LST variations. At different scales, all three USPs were negatively correlated with seasonal U-LSTs. The correlations are particularly prominent in summer and autumn as well as at the city block scale.

5.3. The Effects of UFZs on Seasonal U-LSTs

The explained variance of seasonal U-LSTs in the model involving the total study area was significantly lower than that of each UFZ, suggesting that 2D and 3D USPs play different roles in each UFZ. It is crucial to explain the variation of U-LST using UFZs. For example, BC was most important for the U-LST increases in residential and park zones. In addition, the independently explained variances of 3D USPs on seasonal U-LSTs were considerably higher than those of 2D USPs in most UFZs (Figure 4), suggesting the critical roles of 3D USPs in U-LST regulation at local scales. For example, it is found that SVF has a considerable impact on seasonal U-LSTs for commercial and industrial areas. To better understand the different implications of UFZs on U-LSTs, we analyzed the composition of 2D and 3D USPs in different UFZs (Figure 5).

Manufacturing zones yielded the highest U-LST, followed by residential and commercial zones. In contrast, parks featured the lowest U-LST (Figure 2). This could be largely explained by their different composition of 2D and 3D USPs. For example, we found that manufacturing zones featured a relatively low value of cooling factors, such as low values of MBH (Figure 5a), MBV (Figure 5b), FAI (Figure 5c), TP (Figure 5j), and GP (Figure 5k). In contrast, a relatively high value of warming factors was observed, such as high values of CAR (Figure 5d), SVF (Figure 5f), MTH (Figure 5g), and BC (Figure 5h). Additionally, a relatively high value of typical cooling factors was found in park zones, such as high values of TP (Figure 5j) and GP (Figure 5k). However, it is found that park zones had a relatively low value of warming factors, such as low values of BC (Figure 5h) and MBV (Figure 5b).

5.4. Limitations and Future Research Directions

We noted several limitations. First, the Landsat 8 data of 2015 was utilized to retrieve seasonal U-LSTs using the radiative transfer method. Due to the influence of cloudiness and weather conditions, the number of high-quality images was limited, and thus, high-quality remote sensing images cannot be obtained every month. Therefore, there was some uncertainty in the estimation of U-LSTs [64]. Second, since Landsat satellite can only observe urban surfaces in the daytime, we did not analyze nighttime U-LSTs and how 2D and 3D USPs impact diurnal changes in U-LSTs. It is noteworthy that differences in energy balance and stability between urban and rural areas lead to different land surface warming/cooling rates, generating large daily changes in air temperatures. Moreover, those differences at a given time determine the urban heat island intensity [9,65]. In addition, some scholars have shown that urban structures are critical for nighttime U-LST variations, especially in dense urban areas [28]. In future studies, multiple daytime and nighttime thermal data, such as high-resolution Landsat 8, medium-resolution Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), and their fused data, can be used to explore the daytime and nighttime relationships between USPs and U-LSTs. Lastly, the overall value of our land cover classification was 87%, and thus, there may be some uncertainty in the extraction of 2D and 3D USPs. This undouble would influence subsequent detections of USP–LST relationships. Therefore, it is vital to develop new methods to improve land use classification accuracy.

6. Conclusions

This study investigates the influence of 2D and 3D USPs on seasonal U-LSTs based on a multi-scale analysis (including pixel and city block scales) by using high-resolution remote sensing imageries. To perform a quantitative analysis, we used a Pearson correlation analysis to quantify the relationship between USPs and U-LSTs. Moreover, we analyzed the relative importance of 2D and 3D USPs on seasonal U-LSTs in various UFZs using the RF algorithm. The main conclusions can be drawn as follows.

Manufacturing zones were a critical factor in generating UHIs, whereas parks can help the urban cooling. In addition, the seasonal U-LSTs ranking from high to low was summer (26–50 °C) > fall (16–35 °C) > spring (−3–12 °C) > winter (−8–9 °C). MTH was an important cooling factor for seasonal U−LSTs, and the effect at the city block scale was more pronounced than that at the pixel scale. In particular, the cooling effects of MTH on U-LSTs were lower than those of TP. Generally, the links between USPs and U-LSTs at the city block scale were more obvious than those at the pixel scale, e.g., the Pearson correlation coefficient with LST for the mean building height (MBH) at the city block scale (summer: r = −0.156) is higher than that at the pixel scale (summer: r = −0.081). In addition, the independently explained variances of 3D USPs on seasonal U-LSTs were considerably higher than those of 2D USPs in most UFZs, suggesting the distinctive roles of 3D USPs in U-LST regulating at local scales.

The main contributions of the study can be summarized as follows. First, we analyzed the influence of 2D and 3D USPs on seasonal U-LSTs at different scales (including pixel and city block scales), which has been rarely observed in previous pieces of literature. Second, we first employed the different roles of 2D and 3D USPs across different UFZs. We have shown that SVF yielded an enormous impact on seasonal U-LSTs for commercial and industrial areas. This finding is valuable since it provides new insight into heat mitigation measures using UFZs. In summary, this study can provide useful references for urban management and planning.