A Quantitative Analysis of the Complex Response Relationship between Urban Green Infrastructure (UGI) Structure/Spatial Pattern and Urban Thermal Environment in Shanghai

Abstract

The urban heat island (UHI) effect has evolved into one of the key environmental problems affecting the urban ecological environment and sustainable development. Based on 52 Urban Thermal Heat spots (UTHSs) with significant differences between land use structure and urban green infrastructure (UGI) spatial layout within the influence range of UHI in Shanghai, Landsat-8/9 satellite images were used to construct a high-dimensional dataset reflecting the impact of built environment components on urban thermal environment. Descriptive statistical analysis was used to analyze the spatial difference qualitatively. Using the stepwise regression (SWR) model and partial least square regression (PLSR) model, the complex response relationship between UGI’s structure/spatial pattern differentiation and urban thermal environment in three spatial stratification ranges of UTHSs was quantitatively analyzed. Overall, the statistical explanatory power of the PLSR model is much better than the stepwise regression model. The PLSR model points out that moderately increasing the average building height, class area (CA), percentage of landscape (PLAND), landscape shape index (LSI), and largest patch index (LPI) play a positive role in inhibiting the growth of land surface temperature (LST), and the cooling effect of index weights decreases in order. However, the interaction effects of the box-cox transformed indices with underlines, e.g., CA × Cohesion × AI × LPI and PLAND × CA × Cohesion × AI × LPI, exert relatively small weight on the cooling effect. According to the results, suggestions such as optimization of the UGI structure and urban construction layout were proposed, which can effectively mitigate the UHI effect.

1. Introduction

Since the modern industrial revolution, global urbanization has accelerated the formation of urban areas. Compared with rural areas, urban areas exhibit significantly higher temperatures, creating the urban heat island effect (UHI effect) [1,2,3]. Urbanized areas are typically densely populated and face intense competition for land development. In these areas, natural and semi-natural surfaces are often replaced by artificial structures with hard paving materials, such as buildings, roads, and bridges [4]. This severely unbalanced land use structure significantly alters the thermal energy storage and re-emission processes, sensible/latent heat distribution, and subsurface heat fluxes in urban areas [5], leading to a continuous increase in heat storage within the city [4,6]. Regarding the spatial extent of urban built-up areas, the formation of the UHI effect is closely related to human activities and land use/cover conditions. However, at the macro-scale, the UHI effect is closely related to background climatic conditions, and the combined impacts of climate change-induced large-scale heat waves and local-scale heat island effect further exacerbate the vulnerability of urban ecosystems [7,8]. Cities are exposed to the risk of frequently high temperatures, posing a huge threat to the urban ecological environment, social economy, and human health [1,9,10,11,12,13]. The UHI effect has evolved from a simple climate phenomenon to one of the key environmental issues affecting the urban ecological environment and sustainable development [14,15,16].

Urban Green Infrastructure (UGI) is a common economic and environmental regulation strategy based on ecosystem services to alleviate the UHI effect [17,18,19]. These infrastructures cover natural, semi-natural, and artificial levels, including green infrastructure, water systems, and other aspects that interact with other components of the city [20]. A large number of studies have shown that through sound design and management, UGI can provide a wide range of ecosystem services, such as regulating climate, purifying air, conserving water and soil, and, to a certain extent, inhibiting urban warming [21,22,23]. In densely populated areas, UGI research and management respond directly to the actual needs of the project, adapting the construction of UGI to specific problems [24,25,26,27]. So far, most studies have focused on two-dimensional indicators such as plant species selection, green space coverage, and patch size. They mainly examine the impact of green space area on the thermal environment while neglecting the analysis of green space type, shape, connectivity, and other aspects [28,29,30,31,32,33]. In contrast, there is a lack of studies on the complexity of the three-dimensional spatial configuration pattern of vegetation and buildings in the built environment and how they quantitatively impact the urban thermal environment [29]. Few studies have effectively handled complex/high-dimensional datasets and quantitatively revealed the spatial differentiation and response mechanisms of urban thermal environment patterns at local/micro scales. This paper aims to address this issue.

As a fast-growing megacity with tremendous growth in local population and impervious surfaces, Shanghai has been prone to the summertime UHI effect over the past decades [34,35,36,37]. Shanghai has implemented corresponding energy-saving and emission-reduction plans. However, it is inadequate to take these measures to achieve the goal of UHI mitigation at local/micro scales, especially in densely populated areas that lack well-planned UGI [38,39]. Therefore, this study takes Shanghai as an example due to its importance in China’s socioeconomic perspective and vulnerability to climate change. Through remote sensing and statistical methods, we construct and validate a feasible approach to effectively deal with complex/high-dimensional datasets, quantitatively reveal the spatial differentiation and response mechanism of the urban thermal environment pattern and determine the relative weights of each variable in the high-dimensional dataset. It will provide a reference basis and measures for improving the urban thermal environment and enhancing urban resilience [40,41,42].

2. Materials and Methods

2.1. Study Area

Shanghai is located in China’s eastern coastal economic zone. Its permanent population reached 24.89 million by the end of 2022 [43]. This region is controlled by the East Asian subtropical monsoon climate. It has four distinct seasons: warm spring, hot summer, cool autumn, and chilly winter. The climate is mild and humid; the annual average rainfall is 1097.3 mm, and the annual average temperature is 15.2~15.7 °C [44].

As a typically populous city with 24.89 million permanent residents [45], the developed lands such as commercial districts, residential areas, transportation lines, and industrial parks predominance the share (approximately 71.20%) of land uses in our study area. In contrast, the existing UGI consists of green space, and water accounts for approximately 28.80% of the land use structure. Consequently, such an unbalanced land use structure should be responsible for the severe summertime UHI effect attributed to the absence of UGI, which shapes the contrasting UHI patterns across the highly, mediumly, and slightly urbanized areas in Shanghai [15,35,37,46,47,48,49]. Thus, based on our prior knowledge from the references mentioned above, 52 representative urban thermal hotspots (UTHSs) located in the area of interest addressed by the references mentioned above, which mainly depict remarkable differences in land use structure, building form, UGI configuration, and UHI intensity across the highly, mediumly, and slightly urbanized areas, were selected for this study (Figure 1). The extent of each UTHS was spatially divided into three layers: core area, buffering area, and whole area. The relatively concentrated spatial extent of UGIs was taken as the core area. According to the accessibility requirements of residents to green space, as stipulated in the urban green space system plan of Shanghai [50], the maximum distance to public green space should not exceed 500 m. The surrounding 500-m area was designated as the buffer area. The combination of the core and buffer areas represents the whole area.

2.2. Data Sources

In this study, high-resolution land use maps, Sentinel-2 L1C multispectral data images, Landsat-8/9 OLI/TIRS images, and building outline data covering the spatial scope of 52 UTHSs were used as the main datasets. In addition, together with the digital city thematic products of downtown Shanghai, two online high-resolution satellite imagery platforms (91Weitu Map and Tianditu map) and in-situ field survey data are used as auxiliary datasets.

Table 1. Data sources and description.

Data	Description
Landsat-8/9 OLI/TIRS images	Among the available high-quality cloud-free images collected in the summer of 2013–2022, considering the time span and interval of the whole study period, five phases of images were selected in this study: 29 August 2013, 3 August 2015, 24 August 2017, 16 August 2020, and 14 August 2010. These satellite images were downloaded via www.gscloud.cn (accessed on 1 June 2023).
Sentinel-1/2 images	Sentinel is a series of Earth observation satellites launched by the Copernicus Program of the European Space Agency (ESA). Three images dated 24 February 2020 16 August 2020, 23 February 2020, and 16 August 2020 were downloaded from the Open port provided by the European Space Agency (https://dataspace.copernicus.eu/browser/?zoom=3&lat=26&lng=0&visualizationUrl=https%3A%2F%2Fsh.dataspace.copernicus.eu%2Fogc%2Fwms%2Fa91f72b5-f393-4320-bc0f-990129bd9e63&datasetId=S2_L2A_CDAS&demSource3D=%22MAPZEN%22&cloudCoverage=30, accessed on 1 June 2023).
Land use map	This map of land use cover in 2013 was originally generated using an object-oriented classification method based on orthophose-corrected high-resolution Quickbird satellite imagery. Based on the field investigation data, the classified products were further manually corrected and verified, and resampled to the TIF grid (1 m resolution), with an overall correction accuracy of 91.1% [51].
Building profile data	The building outline is a high-resolution Quickbird satellite image using orthographic correction, and outside the range is manually drawn using the 91 Weitu Map.
Digital city thematic products	Commercial thematic layers contain specific land use covers, such as buildings, warehouses, industrial parks, transportation lines, vegetated areas, and bodies of water. (Beijing Digital Space Technology Co., Ltd., Beijing, China)
Baidu map	Baidu Maps Baidu web products, including high-resolution satellite images (still/no historical review), thematic features (such as buildings, roads, traffic lines, etc.), and street views with retrospective photos.
91Weitu Map	The online high-resolution satellite image and city digital thematic service layer products operated by Beijing Qianfan World View Company (https://www.91weitu.com, accessed on 20 June 2023).
Tianditu map	operated by the National Platform for Common Geospatial Information Services (https://vgimap.tianditu.gov.cn/, accessed on 20 June 2023)
Ground truth data	Collected in 8 annual field surveys conducted between 2013 and 2020, with intervals of 3–6 months, focusing on the land use type and development pattern of each typical sample area, building height was measured on-site using the Edkors™ model AS1000H handheld height finder (Changzhou Edkors Instrument Co., LTD, Changzhou, China) .

lists the descriptive information of these datasets.

2.3. Methods

2.3.1. Data Preprocessing

In this study, all the satellite images and auxiliary datasets were converted to the same geographic coordinate system (WGS 84 system) and reprojected to the Universal Transverse Mercator system (UTM_Zone_51N). Then, atmospheric radiometric calibration for Sentinel-2 Level 2A images was performed using the Sen2Cor v2.9 [52]. Finally, the vector data of the Shanghai administrative division is used as the boundary for cutting.

2.3.2. LST Retrieval and Generation of Thermally Sharpened Products

LST has been widely used to evaluate the urban thermal environment and the UHI effect [53,54]. In order to obtain LST data, the radiative transfer equation (RTE) method was used in this study for surface temperature inversion.

In order to generate Shanghai land use and classification products, first, bands 11, 8a, and 5 of three Sentinel-2 images were fused to generate false color images. Second, referring to the land use map and field survey data, the classification scheme, including four typical land use types, such as buildings, hardtop pavement, vegetation, and water body, was used to describe the land use structure in general. Third, the maximum likelihood algorithm is used for the supervised classification of images. In each image scene, 80–100 training points for each land use type are randomly selected to ensure that all spectral classes it covers are adequately represented in the training statistics. Then, the vectorized classification results were overlaid with the auxiliary dataset and the author’s field survey data for manual correction and verification. Finally, three high-precision Shanghai LULC products with a resolution of 20 m were generated, and the overall classification accuracy was verified to be in the range of 90.33–93.91%.

Based on this high-precision product, the classification-based emissivity method (CBEM) was used to estimate the surface emissivity (LSE), while the unsharped-thermal infrared image resolution of Landsat-8/9 OLI/TIRS was 30 m. Therefore, a dataset of 5 million to 3 million random points with a distance of 20 m was constructed in the study area, and the atmospheric radiation (TOAR) value of Landsat-8/9 band 10 data was extracted and then used for ordinary kriging spatial interpolation analysis. The reconstructed data output is resampled to 30 m, overlapping with the original 30 m TOAR data layer. The minimum root-mean-square error (RMSE) of 100,000 randomly generated data point pairs was used to determine the optimal sampling point value, and the fitting data with a resolution of 20 m was obtained.

Then, according to the known surface-specific emissivity and atmospheric profile parameters such as atmospheric upwelling, downwelling radiation, and transmission measured by satellite, a general radiative transfer equation [55] is used to deduce the radiation luminance of blackbody in the thermal infrared band. Based on Planck’s law, the actual surface temperature is estimated using the below equation:

$$B_{\lambda }\left(T_s\right)={\displaystyle \frac{L_{sensor,\lambda }-L_{atm,\lambda }^{\uparrow }-{\tau }_{\lambda }\left(1-{\epsilon }_{\lambda }\right)L_{atm,\lambda }^{\downarrow }}{{\tau }_{\lambda }{\epsilon }_{\lambda }}}$$

(1)

$$T_s={\displaystyle \frac{K_2}{\ln \left(1+\frac{K_1}{B_{\lambda }\left(T_s\right)}\right)}}$$

(2)

where $L_{sensor,\lambda }$ in the band 10 thermal infrared band (in W · m⁻² · sr⁻¹ · μm⁻¹); ${\epsilon }_{\lambda }$ is the surface emissivity; $B_{\lambda }\left(T_s\right)$ is the radiant brightness (in W · m⁻² · sr⁻¹ · μm⁻¹) of the Planck blackbody when the temperature is equal to $T_s;L_{atm,\lambda }^{\downarrow }$ is the radiant brightness of the atmospheric downwelling in W · m⁻² · sr⁻¹ · μm⁻¹; ${\tau }_{\lambda }$ is the whole-layer atmospheric transmittance of the sensor path (dimensionless); $L_{atm,\lambda }^{\uparrow }$ is the upwelling atmospheric radiant brightness (in W · m⁻² · sr⁻¹ · μm⁻¹). K₁ and K₂ are the conversion constants for the band 10 thermal infrared band in the Landsat-8/9 OLI/TIRS imagery metadata file, respectively, with K₁ = 774.89 W/(m²·ster·μm) and K₂ = 1321.08 K.

2.3.3. Extraction of Main Variables/Indicators of Built Environment Affecting LST

Urban built environment consists of many factors such as land use mode, human behavior, and building form [56,57]. In this study, we constructed a high-dimensional dataset reflecting the complex influencing factors of the built environment. Due to the complexity of the built environment, two typical building indexes, the proportion of impervious surface area and the average building height, are selected. For the landscape pattern index on the UGI scale, combined with the research purpose and the accessibility, simplicity, representativeness, and comparability of the measurement indicators, we selected 9 landscape pattern indicators for this study, all of which are shown in

Table 2. The main influencing variables of built environment impact LST in the study area.

Dimension	Indicator Name	Formula	Meaning
Building index	Proportion of impervious surface area	$A_{\nu }={\displaystyle \frac{\sum _{i=1}^n{A}_i}{A}}\times 100\%$	The proportion of surfaces in a given area that are artificially constructed or artificially enclosed by buildings, roads, sidewalks, etc.
	Building height (BH)	/	The vertical height of a building usually indicates the distance from the outdoor floor to the roof of the building.
UGI index	Class area (CA)	$CA={\displaystyle \sum _{j=1}^n}{a}_{ij}\left({\displaystyle \frac{1}{10000}}\right)$	It can directly reflect the size of different landscape element types.
	percentage of landscape (PLAND)	$PLAND={\displaystyle \frac{{\sum }_{j=1}^{n}a\mathrm{i}}{A}}=100$	The relative percentage of a certain patch type in the total landscape area can be used to judge landscape dominance.
	largest patch index (LPI)	$LPI={\displaystyle \frac{a_{max}}{A}}\times 100\left(0<LPI\le 100\right)$	The maximum continuous patch area as a percentage of the entire landscape area.
	patch density (PD)	$PD={\displaystyle \frac{N\mathrm{i}}{A_i}}$	It reflects the degree of fragmentation and spatial heterogeneity of landscape segmentation.
	CLUMPY	$Given{G}_i=\left({\displaystyle \frac{g_{ii}}{\sum _{\kappa =1}^m{g}_{2,k}}}\right)$ $CLUMPY=\left[\begin{array}{c}{\displaystyle \frac{G_i-P_i}{1-P_i}}for{G}_i\ge P_i\\ g\\ {\displaystyle \frac{G_i-P_i}{1-P_i}}for{G}_i<P_i;P_i\ge 5\\ {\displaystyle \frac{P_i-G_i}{-P_i}}for{G}_i<{P;P_i<5}_i\end{array}\right]$	It reflects the aggregation and dispersion of patches in the landscape, and the value is between −1 and 1.
	COHESION	$COHESION={\displaystyle \frac{1-\frac{{\sum }_{j=i}^n{p}_{ij}}{{\sum }_{\mathrm{j}=1}^n{p}_{ij}\sqrt{{\alpha }_{ij}}}}{\left(1-\frac{1}{\sqrt{A}}\right)}}\times 100$	Represents the distance and arrangement pattern of patches in the landscape, reflecting the continuity.
	Aggregation Index (Al)	$AI=\left[{\displaystyle \frac{g_{\bar{u}}}{\mathrm{m}\mathrm{a}\mathrm{x}\to g_{\bar{u}}}}\right]\left(100\right)$	AI ∈ (0,100). AI examined the connectivity between patches of each landscape type.
	Splitting Index (SPLIT)	$I_{\mathrm{S}\mathrm{P}\mathrm{L}\mathrm{I}\mathrm{T}}={\displaystyle \frac{A^2}{\sum _{i=1}^m \sum _{j=1}^n a_{ij}^2}}$	SPLIT is the sum of the square of the total landscape area divided by the square of the patch area.
	Landscape Shape Index (LSI)	$LSI={\displaystyle \frac{0.25P}{\sqrt{A}}}$	Reflects the complexity of landscape structure; that is, the larger the value, the more complex the shape.

Notes: The class-level metrics of UGI are calculated using Fragstats 4.2. The definition and calculation of adoption indicators are detailed in the User Guide [58].

. The research framework diagram of this paper is shown in Figure 2.

2.3.4. Statistical Analysis

First, in order to explore the complex relationship between predictor and response variables, exploratory data analysis (EDA) was used for predictor and response variables, including normality test, skew data transformation, and correlation. Exploratory data analysis shows that the summer LST data of each year are consistent with the effective sampling of an approximately normal distribution population, and the main variables of the built environment may not be consistent with the normal distribution population sampling. The Box–Cox transformation was carried out on the data with skewed distribution in UGI indices (CA, IS, PLAND, PD, LPI, CLUMPY, COHESION, AI) and Building height (BH) to generate almost every transformation variable that followed the normal distribution, such as Box-IS and Box-PLAND.

Second, the three spatial levels of the core area, the buffer area, and the whole area are divided into five classes by systematic cluster analysis (named C1–C5, B1–B5, W1–W5). Classes I to V are, respectively, highly urbanized areas, mediumly to highly urbanized areas, mediumly urbanized areas, slightly to mediumly urbanized areas, and slightly urbanized areas. The Tukey postmortem test (HSD) was then applied to examine the UGI distributions and LST differences between the five categories under the three spatial stratifications.

Third, considering a large number of predictors and complex interactions among them in this study, multiple-collinearity, and overfitting problems need to be well dealt with when selecting the model. To reduce biases from the multiple-collinearity and overfitting problems, there are several traditional methods, such as the stepwise regression (SWR) and partial least squares regression (PLSR) models and machine-learning regression (MLR) models using random forest, supporting vector machine, and adaptive boost. These methods were proven effective in dealing with high-dimensional datasets. However, both can yield explicit regression equations, whereas the MLR models yield complex and confusing equation formations. Therefore, we prefer to adopt SWR and PLSR models to quantitatively reveal the quantitative response relationship between LST and UGI under the conditions of the complex building environment. The SWR and PLSR have similar equation forms but vary in their computational methods. The former is essentially a simplified ordinary least square regression (OLSR) method, which establishes a group of linear regression models and examines to what extent adding or removing a specific independent variable would cause the interpretation power of the regression models. The latter is essentially an improved OLSR using the non-linear iterative algorithm, which generates a smaller number of uncorrelated components via dimension reduction of the independent variables and performs the OLSR on these components to yield the results with the highest predictive ability [59,60]. Based on the statistical regression model, the complex relationship between UGI multiple indicators, building indicators, and LST was modeled, the weight of each indicator in reducing/inhibiting LST growth was quantitatively evaluated, and the response mechanism between LST and UGI was discussed under the complex characteristics of the built environment. In this study, all statistical processes were analyzed using the basic functions of R version 4.2.2 (R Core Team, 2022). PLSR analysis uses the “pls” library [61].

3. Results

3.1. The Relationship between UGI and LST under Different Spatial Stratification

Table 3. Table of the area proportion of five types of UTHS regions under three spatial layers.

Type	Whole Area = Core Area + Buffer Zone			Core Area			Buffer Area
	Impervious Surface Area (%)	Building Area (%)	UGI Area (%)	Impervious Surface Area (%)	Building Area (%)	UGI Area (%)	Impervious Surface Area (%)	Building Area (%)	UGI Area (%)
C1	98.95 ± 0.91	34.62 ± 10.97	1.05 ± 0.91	98.28 ± 1.18	21.2 ± 5.39	1.72 ± 1.18	98.98 ± 0.89	35.05 ± 11.03	1.02 ± 0.89
C2	93.20 ± 4.58	24.86 ± 9.60	6.80 ± 4.58	72.03 ± 13.52	12.8 ± 8.38	25.82 ± 15.16	94.62 ± 4.91	27.72 ± 7.43	5.38 ± 4.91
C3	89.83 ± 8.29	23.22 ± 8.22	10.17 ± 8.29	85.73 ± 7.58	16.1 ± 4.78	14.27 ± 7.58	89.63 ± 10.48	23.38 ± 9.15	10.37 ± 10.48
C4	93.75 ± 3.29	21.30 ± 3.23	6.25 ± 3.29	91.84 ± 3.59	16.6 ± 4.55	8.16 ± 3.59	93.98 ± 3.26	21.73 ± 3.18	6.02 ± 3.26
C5	83.78 ± 15.02	19.27 ± 8.15	15.54 ± 14.93	53.88 ± 14.60	6.22 ± 6.48	52.27 ± 13.85	88.27 ± 15.01	21.28 ± 7.75	11.20 ± 14.93
Entirety	89.14 ± 11.73	22.61 ± 9.20	10.56 ± 11.54	69.51 ± 20.05	11.3 ± 8.20	32.64 ± 8.20	91.51 ± 11.30	24.36 ± 8.56	8.26 ± 11.19

Notes: UGI area ratio is equivalent to the PLAND index. The proportion of building construction area is included in the proportion of impervious surface area.

shows the overall trend of impervious area ratio, building construction area ratio, and UGI area ratio of the core area, the buffer area, and the whole area. Combined with the analysis and comparison results of the HSD.test, the differences in LST distribution are reflected in Figure 2. Table 3 and Figure 3 qualitatively show that the spatial distribution pattern of UGI under different regional types and spatial stratification has statistically significant differences, and this has a significant impact on LST.

First, the order of impervious area ratio and building construction area ratio is buffer > whole area > core area, while the order of UGI area ratio is opposite. The average LST in the core, buffer, and whole areas increased successively. The UGI pattern of the core and buffer areas presents obvious imbalance and spatial discontinuity. The LST of the core area is significantly lower than that of the buffer area. In contrast, the UGI distribution pattern and LST of the whole area are closer to that of the buffer area, indicating that the UGI and LST of the whole area mainly depend on the UGI spatial layout of the buffer area. Second, from the perspective of different types of areas, the proportion of impervious surface (IS) area and building construction area of the core area and buffer zone from class I to class V generally shows a gradually decreasing trend, and the proportion of UGI area generally shows a gradually increasing trend. At the three spatial levels, the LST of the five classes decreased successively, which was consistent with the spatial distribution pattern of UGI. The error in individual years is within acceptable limits. The results show that the urban development intensity of class V is low, the ecological infrastructure is good, the UGI area ratio is high, and the IS and the building area ratio is low, resulting in a low LST value.

3.2. Response of LST to UGI Pattern in UTHS Range

Figure 4 quantitatively shows the correlation of LST among the three spatial layers and the results of unitary non-linear regression analysis. Among the three spatial layers, the correlation between the buffer area and the whole area is the most significant, followed by the correlation between the core area and the whole area, and finally, the correlation between the buffer area and the core area. This finding further verifies the influence of UGI spatial pattern on LST, and the spatial pattern difference between the core area and the buffer area is significant. In contrast, UGI in the core area does not have a good cooling effect on the adjacent areas. In general, the three spatial layers have a high degree of consistency. However, as can be seen from the scatter distribution in Figure 4, the data in some regions deviate from the 95% confidence interval. In Figure 5, satellite image maps of 10 typical UTHSs are selected to reflect the land use in the core area and the surrounding 500 m buffer zone of the 10 typical areas. Combined with the LST distribution map in Figure 5, different land development modes and their impacts on the urban thermal environment are shown. In low-temperature areas (e.g., Senlan Lake and Gucun Park), vegetation and water bodies dominate the core areas, and UGI has high continuity in the region. Meanwhile, in high-temperature areas (Poly Champagne Garden and Riverside Garden) are in Figure 4. Non-linear fitting diagram of LST correlation among three spatial stratification areas dominated by dense buildings with low vegetation coverage, UGI is only concentrated in small blocks, and the spatial connectivity is poor. From the perspective of spatial stratification, the vegetation coverage rate of the core area of Longhua Martyrs Cemetery is relatively high, and the surrounding green space has a certain continuity with it, so LST has high consistency across the three spatial layers.

3.3. Quantitative Model Analysis of Pattern Response Relationship between LST and UGI

Table 4. Analysis results of LST and predictor SWR model.

	2022					2020
	Coef	S − Coef	T	p	VIF	Coef	S − Coef	T	p	VIF
Constant	49.145	0.443	110.953	0.000		45.634	1.364	33.465	0.000
CA	−	−	−	−	−	−5.452	0.989	−5.515	0.000	3.152
IS	0.000	0.000	12.289	0.000	1.348	0.000	0.000	9.747	0.000	8.184
PD	−	−	−	−	−	0.407	0.165	2.463	0.015	1.616
LPI	−	−	−	−	−	−0.642	0.217	−2.961	0.004	16.625
Cohesion	−	−	−	−	−	0.000	0.000	5.176	0.000	20.986
Height	−0.903	0.114	−7.949	0.000	1.300	−0.647	0.102	−6.356	0.000	1.583
SPLIT	0.000	0.000	2.067	0.040	1.042	−	−	−	−	−
S	1.48198					1.20508
R-sq	53.94%					74.14%
R-sq(adj)	53.02%					73.08%
R-sq(pred)	51.53%					70.40%
	2017					2015
	Coef	S − Coef	T	p	VIF	Coef	S − Coef	T	p	VIF
Constant	50.947	1.232	41.353	0.000		51.06	1.57	32.61	0.000
CA	−5.052	1.128	−4.480	0.000	3.264	−6.53	1.14	−5.75	0.000	3.15
IS	0.000	0.000	8.856	0.000	7.916	0.000	0.000	9.39	0.000	8.18
PD	−	−	−	−	−	0.387	0.190	2.04	0.043	1.62
LPI	−0.428	0.235	−1.822	0.070	15.508	−0.396	0.249	−1.59	0.114	16.63
Cohesion	0.000	0.000	4.584	0.000	15.968	0.000	0.000	4.15	0.000	20.99
Height	−0.855	0.113	−7.575	0.000	1.545	−0.920	0.117	−7.87	0.000	1.58
SPLIT	0.000	0.000	1.774	0.078	1.175	−	−	−	−	−
S	1.35071					1.38388
R-sq	69.24%					70.63%
R-sq(adj)	67.98%					69.43%
R-sq(pred)	63.08%					66.30%
	2013
	Coef		S − Coef		T		p		Coef
Constant	49.731		0.788		63.15		0.000
CA	−1.923		0.674		−2.85		0.005		2.04
IS	0.000		0.000		12.36		0.000		5.27
PD	−		−		−		−		−
LPI	−		−		−		−		−
Cohesion	0.000		0.000		3.53		0.001		7.04
Height	−0.4381		0.079		−5.54		0.000		1.32
SPLIT	−		−		−		−		−
S	1.023
R-sq	72.84%
R-sq(adj)	72.11%
R-sq(pred)	70.89%

Notes: In this table, Coef and S − Coef are abbreviations for stepwise regression model coefficient and standardization coefficient, respectively, and the underlined variable is transformed by Box–Cox. A value of 0.000 in the table indicates that the value is very small, close to 0.

shows the analysis results of the stepwise regression (SWR) model as the basic model, and

Table 5. Summary of PLSR results of LST and predictors for the whole area.

Effect Term		LST2022		LST2020		LST2017		LST2015		LST2013
		Coef	S − Coef	Coef	S − Coef	Coef	S − Coef	Coef	S − Coef	Coef	S − Coef
	Constant	51.77		49.61		53.71		55.09		53.42
Main effect	CA	−1.18	−0.13	−1.28	−0.12	−1.45	−0.13	−1.74	−0.14	−0.93	−0.10
	IS		0.17		0.16		0.17		0.18		0.13
	PD	0.11	0.05	0.12	0.04	0.13	0.05	0.16	0.05	0.09	0.04
	PLAND	−0.54	−0.10	−0.66	−0.11	−0.69	−0.11	−0.85	−0.12	−0.54	−0.10
	LPI	−0.05	−0.04	−0.09	−0.06	−0.08	−0.05	−0.10	−0.06	−0.09	−0.07
	Cohesion		0.02		−0.02						−0.04
	AI		0.07		0.03		0.06		0.05		−0.01
	Height	−0.57	−0.47	−0.49	−0.35	−0.65	−0.43	−0.74	−0.44	−0.25	−0.20
	LSI	−0.07	−0.25	−0.07	−0.20	−0.08	−0.23	−0.09	−0.24	−0.04	−0.14
	SPLIT		0.04		0.04		0.04		0.05		0.04
Interaction	PD × SPLIT × LSI		0.04		0.05		0.04		0.05		0.05
effect	CA × Cohesion × AI × LPI		−0.11		−0.11		−0.11		−0.12		−0.10
	IS × Height		−0.05		−0.01		−0.04		−0.03		0.03
	IS × Height × PD × SPLIT × LSI		0.03		0.04		0.04		0.04		0.04
	PLAND × PD × SPLIT × LSI		0.05		0.05		0.05		0.06		0.05
	IS × Height × PLAND		−0.03		0.01		−0.02		−0.01		0.03
	PLAND × CA × Cohesion × AI × LPI		−0.11		−0.11		−0.11		−0.12		−0.10
	F	15.77		19.33		16.67		23.81		20.72
	R2	0.673		0.652		0.670		0.699		0.749

Notes: In this table, Coef and S − Coef are abbreviations for PLSR model coefficients and Standardized coefficients, respectively. Underlined variables have been processed by Box–Cox transformations.

,

Table 6. Summary of PLSR results of LST and predictors for the core area.

Effect Term		LST2022		LST2020		LST2017		LST2015		LST2013
		Coef	S − Coef	Coef	S − Coef	Coef	S − Coef	Coef	S − Coef	Coef	S − Coef
	Constant	52.01		50.88		55.05		56.52		54.84
	CA	−0.93	−0.10	−0.72	−0.07	−0.49	−0.04	−0.95	−0.07	−0.12	−0.01
Main effect	IS		0.26		0.36		0.42		0.39		0.39
	PD	−0.05	−0.02	−0.27	−0.10	−0.44	−0.15	−0.34	−0.11	−0.40	−0.17
	PLAND	−0.70	−0.13	−1.09	−0.18	−1.26	−0.19	−1.39	−0.19	−1.05	−0.19
	LPI	−0.06	−0.05	−0.12	−0.08	−0.12	−0.07	−0.14	−0.08	−0.13	−0.10
	Cohesion		0.04		0.04		0.08		0.06		0.04
	AI		0.09		0.09		0.14		0.12		0.09
	Height	−0.64	−0.53	−0.76	−0.54	−0.98	−0.65	−1.06	−0.63	−0.56	−0.45
	LSI	−0.07	−0.24	−0.08	−0.23	−0.09	−0.26	−0.11	−0.27	−0.05	−0.18
	SPLIT		0.02		0.04		0.03		0.03		0.04
	PD × SPLIT × LSI		0.02		0.04		0.04		0.04		0.05
Interaction	CA × Cohesion × AI × LPI		−0.09		−0.08		−0.06		−0.08		−0.05
effect	IS × Height		−0.02		0.03		0.02		0.02		0.07
	IS × Height × PD × SPLIT × LSI		0.01		0.03		0.03		0.03		0.04
	PLAND × PD × SPLIT × LSI		0.03		0.05		0.05		0.05		0.06
	IS × Height × PLAND		0.02		0.09		0.10		0.08		0.14
	PLAND × CA × Cohesion × AI × LPI		−0.09		−0.09		−0.07		−0.09		−0.06
	F	19.69		39.29		33.05		43.47		48.57
	R2	0.692		0.634		0.651		0.672		0.722

Notes: In this table, Coef and S − Coef are abbreviations for PLSR model coefficients and Standardized coefficients, respectively. Underlined variables have been processed by Box–Cox transformations.

and

Table 7. Summary of PLSR results of LST and predictors in the buffer area.

Effect Term		LST2022		LST2020		LST2017		LST2015		LST2013
		Coef	S − Coef	Coef	S − Coef	Coef	S − Coef	Coef	S − Coef	Coef	S − Coef
	Constant	52.54		47.90		52.56		51.97		51.66
Main effect	CA	−0.51	−0.04	−1.70	−0.13	−1.52	−0.11	−2.26	−0.16	−0.50	−0.05
	IS		0.47		0.73		0.71		0.75		0.63
	PD	−0.23	−0.08	0.03	0.01	−0.18	−0.06	0.15	0.05	−0.09	−0.04
	PLAND	−0.77	−0.14	−0.73	−0.12	−0.82	−0.14	−0.72	−0.11	−0.62	−0.13
	LPI	−0.09	−0.07			−0.01	−0.01	0.03	0.03	−0.02	−0.02
	Cohesion		0.19		0.36		0.36		0.39		0.28
	AI		0.01		−0.01		0.02		0.01		0.02
	Height	−0.94	−0.52	−0.94	−0.49	−1.08	−0.55	−1.15	−0.55	−0.61	−0.38
	LSI	−0.03	−0.08	−0.03	−0.07	−0.05	−0.11	−0.05	−0.11	−0.02	−0.05
	SPLIT		0.10		0.11		0.13		0.11		0.06
Interaction	PD × SPLIT × LSI		0.03				0.02		0.01		0.01
effect	CA × Cohesion × AI × LPI		−0.11		−0.10		−0.09		−0.11		−0.10
	IS × Height		−0.03		−0.03		−0.04		−0.05		0.03
	IS × Height × PD × SPLIT × LSI		−0.01		−0.06		−0.04		−0.04		−0.02
	PLAND × PD × SPLIT × LSI		0.03		0.02		0.04		0.04		0.04
	IS × Height × PLAND		0.13		0.30		0.26		0.30		0.30
	PLAND × CA × Cohesion × AI × LPI		−0.09		−0.05		−0.04		−0.06		−0.06
	F	25.95		54.73		42.29		46.17		53.60
	R2	0.576		0.753		0.727		0.718		0.751

Notes: In this table, Coef and S − Coef are abbreviations for PLSR model coefficients and Standardized coefficients, respectively. Underlined variables have been processed by Box–Cox transformations.

show the analysis results of the PLSR model of 52 UTHSs within the whole area, core area, and buffer area, respectively. The stepwise regression model has a general fitting effect for different years and can explain 53.94–74.14% of the variance of LST. There is a gap between R-sq (pred) and R², indicating that there may be overfitting of the model. In addition, independent variables such as CA, PD, LPI, and Cohesion are not identified in the data for 2022. Other years also have the problem that the number of indicators introduced fluctuates greatly and is unstable. High VIF (VIF greater than 5) is common among several variables in the SWR models, indicating there exists serious multiple-collinearity. Therefore, the statistical validity of the model is doubtful.

Compared with the SWR model, the PLSR model explained 56.6–75.3% of the variance variation of LST, showing a higher statistical effect. It is worth noting that the PLSR model takes all independent variables and their interaction effects into account, and the corresponding results between each variable and LST are more reasonable than that of the stepwise regression model. This model has better robustness and prediction ability.

From the perspective of the three spatial levels, the main independent variables that play a positive role in promoting the increase of LST are, in turn, IS, AI, and SPLIT. Among the main predictors and interaction effects of the negative normalization coefficient, the results of the three spatial layers are also roughly the same. The cooling effects among them are ranked as Height > CA > PLAND > LSI > LPI > CA × Cohesion × AI × LPI > PLAND × CA × Cohesion × AI × LPI according to the index weights. Their mean standardized partial regression coefficients are −0.62, −0.59, −0.48, −0.12, −0.06, −0.05, and −0.04. When CA, Height, and PLAND are high, it indicates that from the perspective of land use structure, the area of UGI is large, and the corresponding water bodies and vegetation are conducive to exerting the cooling effect of ecological infrastructure. From the perspective of building morphology, the higher the building, the larger the area of the building shadow and shelter the ventilation corridor from direct solar radiation. The varying interactive effects of floor height, UGI dispersion and aggregation metrics, as well as the area shares of UGI and impervious surfaces, have diverse influences on the LST. The higher the floor, the larger the proportion of UGI and the reasonable distribution, and the more conducive to the formation of low-temperature areas and the alleviation of the UHI effect.

4. Discussion

4.1. Improving UGI Spatial Pattern to Mitigate UHI Effect

The structure and spatial structure difference of UGI poses an impact on the urban thermal environment. It has been suggested that vegetation and water play an important role in cooling and mitigating UHI [62,63,64]. Others argue that UGI should be used as a mainstream adaptation measure to climate change [64,65,66,67]. Different views hold that the influence of building form on surface temperature is much greater than that of UGI [68,69,70,71].

In this study, the 20 m thermally sharpened LST products were generated by spatial interpolation analysis, which well suppresses the heterodyne isospectral and its bias phenomenon caused by mixed image elements. Unlike previous studies that focused only on the entire study area or a specific typical area, this paper also focuses on the radiation effects of UGI facilities in the core area on the surrounding areas. This study verifies that UGI plays a significant role in improving the urban thermal environment, while the modeling analysis reveals that UGI’s cooling effect is also affected by the distribution of its spatial pattern. Over the past decades, Shanghai has experienced unprecedented urban growth at the loss of huge, predeveloped landscapes such as croplands and inland water bodies. The recently established UGI patches in some parts of downtown Shanghai and big city parks at the urban periphery help suppress the expansion of impervious surfaces with higher LSTs. Regrettably, the new UGI patches are located far apart and cannot connect the heat sources and sinks along the urban core-periphery gradient. It is essential to address the imbalance of the UGI pattern between the core and buffer areas along the urban core-periphery gradient and enhance the spatial continuity between these areas as much as possible [72,73]. According to the weights of the influence factors, it is recommended to prioritize the built-up areas with a lower UGI ratio and implement renovation projects and structural optimization designs that focus on increasing the UGI ratio. Specific referable methods mainly include:

(1)

Optimizing three-dimensional greening design. Emphasis is placed on strengthening the optimization of the vertical structure of vegetation and the design of shade so as to effectively increase the vegetation coverage rate in the limited urban greening space. Different types of small green spaces, such as lawns and plants, can be set up on the roofs and external walls of buildings with suitable conditions. On the plane, the greenway system is introduced into urban planning, green belts, tree canopies, and other structures are integrated into the urban landscape, and importance is attached to the role of UGI components in urban management;

(2)

Optimizing the layout and management of urban water bodies [74,75,76]. Adapting to local conditions and appropriately increasing small water bodies, mainly with water features such as artificial wetlands, can make good use of the evaporation and heat dissipation effects of water. The rate of rainwater runoff is slowed down by permeable pavements and green lawns, forming large-scale vegetation–water mosaic landscapes, enhancing regional UGI connectivity, lowering urban surface temperatures, and forming a relatively cool microclimate environment.

4.2. Improving Ecological Building Design to Enhance Urban Sustainability

Unlike previous studies that analyzed a single factor, this study evaluates the weight of each factor on surface temperature from the perspective of multiple influencing factors. It is demonstrated that improving the balanced configuration of green spaces, buildings, and impervious surfaces is a way to minimize urban surface temperature in the future [77,78]. From a building perspective, building height and density have a significant effect on the thermal environment. The positive correlation between building height and the UHI effect may be related to the heat absorption and exothermic properties of taller buildings, which influence air flow and radiant heat distribution. Taller buildings may lead to more significant shading effects, with higher floors resulting in larger building shadows that can be created by the floors, as well as the formation of ventilation corridors that block direct sunlight, affecting local heat distribution [76,79,80]. Building density, on the other hand, has a complex interrelationship with the urban microclimate and thermal environment. In view of this, it is recommended to optimize the building’s spatial structure and layout in urban planning to avoid the centralized distribution of excessively tall buildings. Moderate building density and height mixing can help to form appropriate airflow and mitigate the heat island effect.

In addition, this study reveals that there is a synergistic effect of built environment factors (e.g., UGI’s spatial pattern, building height, area shares of UGI and buildings, and their interactive effects) that contribute to the UHI effect. Therefore, more attention should be paid to optimizing the three-dimensional morphology and overall spatial layout of the city to maximize the role of factors that inhibit/slow down the growth of LST [81]. On the basis of ensuring reasonable land development intensity and enhancing the demand for ecosystem services, rational urban high-rise building planning and design and neighborhood layout can create conditions for the construction of UGI in building-adjacent spatial greens and small water bodies, as well as moderately increase the building’s floor area ratio and increase building spacing. This can further achieve the purpose of regulating the temperature of the urban thermal environment and enhancing urban sustainability [82,83].

Finally, more environmentally friendly building designs and technologies can be used to realize the redevelopment of existing buildings and can also be used for new building construction. For example, the use of heat-insulating materials and intelligent cooling systems can help reduce heat absorption in buildings and improve their energy efficiency [73]. Regarding the relationship between building materials and their thermal performance that influences the urban thermal environment, a rational selection of building materials is needed, which includes the use of highly reflective building materials, road materials, and increasing green cover to improve reflectivity. Reducing the building’s absorption of solar radiation by increasing the reflection of urban surfaces helps to control the building’s surface temperature, which in turn reduces the surrounding air temperature.

5. Conclusions

By dividing 52 UTHSs into three spatial levels, this study quantified the complex response relationship between UGI structure/spatial pattern and the urban thermal environment in Shanghai. We found that the regional spatial differentiation of each spatial level is significant, and the UGI pattern has a strong negative correlation with LST. Secondly, the PLSR model was used for quantitative evaluation of each indicator, and it was found that the independent variables and their interaction effects were roughly the same on the three spatial levels. Weight indicators Height, CA, PLAND, LSI, LPI, CA × Cohesion × AI × LPI, PLAND × CA × Cohesion × AI × LPI cooling weights are ranked from high to low. The results show that when the UGI area is relatively large, and the building floor is high, it is beneficial to realize the cooling effect. However, when the patch density is small, and the degree of spatial fragmentation is high, the cooling effect of UGI will be affected. It is necessary to make good use of the synergistic effect between these parameters to maximize the cooling effect of each landscape pattern element. In view of this, it is proposed to make good use of UGI spatial pattern and improve ecological architectural design. Specifically, the UGI area ratio can be improved by strengthening the construction of three-dimensional greening and optimizing the layout and management of water bodies. In addition, rational planning and design of the layout of urban high-rise buildings and neighborhoods is encouraged, which can improve urban air circulation and radiant heat distribution and create more green and blue spaces, thus addressing the UGI imbalance.

However, this study still has some limitations. First, the spatial discrepancy due to the optical and thermal infrared images with different spatial resolutions used in this study still makes it difficult to completely suppress the bias caused by the mixed pixel effect. Second, in terms of time scale, the paper lacks a comprehensive grasp of the long-term trend of LST with high temporal resolution, like the widely used air temperature data. Besides, regarding the artificial factors and their influences on the UHI effect, the role of human activities is not taken into account, which may cause slight deviations in the data in this study. Therefore, in future studies, it is necessary to adopt longer time scale data, conduct quality control of datasets through reasonable cross-validation approaches, continuously expand the scope of factor analysis, develop more reasonable quantitative evaluation models, and apply the research results in actual urban planning and governance management.