Cooling Intensity of Urban Blue–Green Spaces and Its Driving Mechanisms in the Yangtze River Delta Urban Agglomeration

Abstract

As global warming intensifies, urban heat stress is increasing. Urban blue–green spaces (UBGSs) reduce heat exposure, and refined planning can enhance cooling in space-limited cities. Yet drivers and mechanisms at the urban-agglomeration scale remain unclear. This study quantified the cooling intensity (CI) of 162 UBGS across 27 cities in the Yangtze River Delta using summer 2020 land surface temperature data (June–August). CI is defined as the average temperature difference (ΔLST) between the UBGS and its surrounding buffer. Extreme gradient boosting (XGBoost) and Shapley additive explanations (SHAP) methods were applied to analyze the driving factors. Key findings include (1) the top five factors by SHAP value: Area (0.38), buffer NDBI (0.16), lake/river landscape shape index (0.07), buffer tree height (0.06), and buffer population (0.06). (2) CI rises with Area only up to 50 hm² and then plateaus at ~6 °C. Buffer NDBI strengthens CI, with faster gains when buffer NDBI < −0.2. Water-shape effects are stepwise, and a lake/river shape index of 5–8 delivers near-optimal cooling. Interactions indicate stronger cooling in large parks with high water coverage and in sites combining complex water edges with dominant grassland patches. These findings support structure-focused UBGS design and targeted mitigation in dense urban areas.

1. Introduction

Under the dual pressures of rapid global urbanization and climate change, the urban heat island (UHI) effect has become a severe challenge faced by cities worldwide [1]. UHI exerts substantial adverse impacts on both natural ecosystems and socioeconomic systems, leading to increased energy consumption, more frequent extreme heat events, deteriorating air quality, and public health crises [2,3,4]. It also elevates health risks for urban residents by increasing morbidity and mortality rates and adversely affecting mental health [5,6]. With rising urban greenhouse gas emissions and the growing frequency of extreme climatic events, there is an urgent need to integrate climate change mitigation and adaptation strategies into urban planning frameworks [7].

Compared with engineering-based approaches to heat mitigation, nature-based solutions (NbS) regulate the urban thermal environment through natural processes such as vegetation evapotranspiration, shading, and water evaporation, and are therefore regarded as more sustainable and long-term cooling strategies. Among various NbS, Urban blue–green spaces (UBGSs) refer to an integrated network of green land and water infrastructure in urbanized areas, including parks, wetlands, rivers, and vegetation. These spaces are key natural assets in urban areas [8,9]. UBGS provide essential ecosystem services such as urban heat mitigation, climate change adaptation, air quality improvement, and recreational areas, and are widely regarded as effective nature-based solutions for adapting to environmental warming and mitigating extreme heat. They balance economic progress with environmental protection [10], playing a critical role in local climate regulation and urban heat environment management [11,12,13]. Urban heat environment refers to the spatial and temporal temperature variations within urban areas, influenced by factors such as land use, building density, and the presence of green spaces, leading to the urban heat island effect (UHI) [14]. Consequently, how to improve the urban thermal environment through refined planning and design of UBGS within limited urban space has become an important research topic in sustainable development and urban governance.

However, the cooling effects of UBGSs are not spatially homogeneous and vary considerably across cities, spatial scales, and landscape configurations. To systematically characterize the capacity of UBGSs to mitigate UHI effects, research has gradually shifted from qualitative descriptions to quantitative assessments, leading to the development of multiple indicators for measuring cooling performance. For example, Guo et al. [15] proposed a set of metrics—including park cooling intensity (PCI, temperature difference between the park and its surroundings), park maximum cooling intensity (PMCI, maximum temperature difference generated by the park), park cooling area (PCA, spatial extent of cooling), park cooling efficiency (PCE, efficiency of cooling relative to area and vegetation), and park cumulative cooling intensity (PCCI, total cooling effect over time)—to investigate the spatial patterns of park cooling effects from multiple perspectives. Among these indicators, cooling intensity (CI) is defined as the difference in LST between the average temperature of the impervious surface within the buffer and the average temperature of UBGSs [16]. It is widely used to assess the cooling effect of urban blue–green spaces and is one of the most frequently utilized indicators to characterize “cooling islands” [17,18,19]. Compared to other indicators, CI accounts for the spatial continuity of the cooling effect and is typically used in large-scale, cross-regional studies [20]. Therefore, identifying and quantifying the key drivers of CI is of great practical significance for optimizing the configuration of UBGS and scientifically mitigating the UHI effect.

The study of UBGS cooling effects and their influencing factors has been extensively conducted. Higher Cooling Intensity (CI) indicates a larger cooling range and stronger cooling effect [21,22,23]. CI is influenced by factors such as the size of UBGS [15], landscape composition [24], shape characteristics [25], climatic conditions, and surrounding environmental factors [26]. Specifically, CI is related to the shape characteristics of UBGS and its internal landscape patterns (such as the percentage of land use and land cover, and spatial configuration of land use/land cover (LULC). CI generally increases logarithmically with the area size [11,23], while the relationship between shape and CI varies across different studies [17,27]. Vegetation and water bodies within UBGS typically enhance CI, while impervious surfaces such as roads and buildings usually reduce it [13,28]. Kong et al. [29] found that as the average forest patch area increases, CI decreases with the density of forest patches in parks. These findings reveal the driving factors behind CI and provide valuable knowledge for better urban planning and management [30,31,32]. However, the conclusions of these studies often differ due to variations in study area, sample selection, and analytical scale, resulting in a lack of consensus regarding the underlying mechanisms of each factor. For instance, some studies indicate that more compact and simplified green spaces show stronger cooling abilities [33,34], while others argue that shape has a negligible effect [35,36].

In addition to the geometric shape and landscape pattern within UBGS, the surrounding 2D/3D built environment, such as building density, height, and sky openness, has been shown to significantly affect CI in UBGS [37,38]. However, external indicators of UBGS do not directly affect CI, but rather indirectly influence it by affecting the temperatures inside and outside UBGS, with internal and external temperatures interacting with each other [39]. High-temperature buffer zones may impose heat stress on park vegetation, thereby reducing its cooling effect [21]. Previous studies have mostly used multiple linear regression methods, with CI as the dependent variable and driving factors as the independent variables [21,40,41,42]. However, single-equation regression, estimated with Ordinary Least Squares (OLS), may ignore potential endogeneity, leading to biased and inconsistent results. For example, Geng et al. [41] reported a positive correlation between CI and the green cover percentage of the buffer zone, while Shah et al. [43] and Qiu and Jia [42] found a negative correlation between CI and buffer greening degree (i.e., normalized difference vegetation index). These conflicting results highlight the importance of using robust statistical methods to obtain reliable results. An increasing number of researchers are now using machine learning models to gain deeper insights into UBGS cooling effects. For instance, Yang et al. used random forests (RF) to investigate the impact of various indicators on UBGS cooling ability [24]. Deng et al. used random forest regression and Shapley Additive Explanations (SHAP) [44] to uncover the driving factors of urban wetland park cooling effects [45]. Luo et al. applied XGBoost [46] and SHAP techniques to explore the spatial effects, relative importance, and interactive impacts of 2D/3D built environments on LST [47].

In recent years, academic attention has increasingly shifted from isolated city-level UHI studies to investigations of regional thermal patterns at the urban agglomeration scale [20,48]. Urban agglomerations are characterized by high spatial connectivity, economic integration, and dense development [49], causing surface UHI effects to transcend individual city boundaries and evolve into regional thermal environmental issues [50,51]. This trend is particularly evident in the Yangtze River Delta (YRD), one of the three major urban agglomerations in the Yangtze River Basin. At the urban agglomeration scale, exploring the internal and external driving factors affecting the cooling intensity of UBGS and systematically quantifying CI and its driving mechanisms is a research topic worthy of further exploration.

This study focuses on the Yangtze River Delta urban agglomeration, using land surface temperature data from June to August 2020. The study employs methods such as extreme gradient boosting (XGBoost) and Shapley additive explanations (SHAP) to quantify the summer cooling intensity of urban blue–green spaces and analyze its main driving factors. A total of 162 urban blue–green spaces across 27 cities in the YRD were compiled for analysis. The specific objectives of this study are to: (1) quantify the summer CI of urban green spaces in the YRD region based on land surface temperature data; (2) identify and analyze the key drivers of CI from 2D landscape characteristics, 3D built-environment features, and socio-economic dimensions; (3) reveal the nonlinear relationships and interactions between influencing factors and CI, as well as among the factors themselves, to explore their synergistic effects.

Given the highly urbanized and heatwave-prone context of the Yangtze River Delta, understanding how UBGSs regulate surface cooling at the urban-agglomeration scale is particularly necessary. Therefore, this study examines the joint effects of multi-dimensional drivers on CI, with an emphasis on their nonlinear relationships and interactions. The results are expected to provide region-specific evidence to support UHI mitigation and UBGS-oriented planning in the YRD and other rapidly urbanizing urban agglomerations.

2. Materials and Methods

2.1. Study Area

The Yangtze River Delta (YRD) region, located along China’s eastern coastline, is one of the nation’s most economically developed and highly urbanized areas. Geographically, it spans from 115.5° E to 122.5° E longitude and from 27.5° N to 34.5° N latitude. The region encompasses Shanghai, southern Jiangsu, northern Zhejiang, and eastern Anhui, covering a total area of approximately 211,700 km² (Figure 1). According to the Köppen–Geiger climate classification, the study area falls within the subtropical monsoon climate zone (Cwa) [52]. With an urbanization rate exceeding 70%—substantially higher than the national average—the YRD exhibits distinct developmental characteristics while facing significant environmental challenges, including uneven green-space distribution and insufficient heat island mitigation services.

Figure 1. (a) Location of provinces and municipalities in the Yangtze River Delta region within China; (b) 27 cities and UBGS within the Yangtze River Delta region.

Major cities within the region, such as Shanghai, Nanjing, and Hangzhou, have undergone rapid urbanization, resulting in extensive expansion of built-up areas and intensifying ecological pressures. Notably, Hangzhou, Nanjing, and Hefei were identified as “Top Ten Furnace Cities” by the National Climate Center of the China Meteorological Administration based on analyses of meteorological records from 1981 to 2010 (https://www.cma.gov.cn/2011xwzx/2011xqxxw/2011xqxyw/202110/t20211029_4039203.html, accessed on 5 October 2025). The YRD frequently experiences prolonged summer heatwaves. For example, during the summer of 2025, Shanghai recorded an exceptionally persistent heatwave, with 24 consecutive high-temperature days—tying the historical record since observations began in 1873 (https://www.thepaper.cn/newsDetail_forward_31470705, accessed on 5 October 2025).

Despite increases in green coverage to over 40% in most urban built-up areas and improvements in the service coverage of parks and green spaces, the region continues to face challenges such as imbalanced green-space allocation and limited cooling-island effectiveness. As urban boundaries increasingly merge and interconnect, the urban heat island (UHI) effect has transitioned from a localized issue within individual cities to a broader regional thermal environmental concern. The National Climate Change Adaptation Strategy 2035 [53] emphasizes enhancing the climate resilience of major urban agglomerations—including the YRD—where effective mitigation of UHI is a critical component of developing climate-adaptive cities.

Given that the Yangtze River Delta (YRD) urban cluster serves as a development model for emerging urban agglomerations across China—such as the Beijing–Tianjin–Hebei region, the Guangdong–Hong Kong–Macao Greater Bay Area, and the Chengdu–Chongqing Economic Circle—the findings of this study are not only of local relevance but also offer valuable insights for other rapidly urbanizing regions.

This study considers the influence of large bodies of water, such as the East China Sea and Yellow Sea, as well as mountain ranges, on local microclimates and their effects on the cooling performance of urban blue–green spaces. As a result, the study focused on urban blue–green spaces greater than 1 hectare and located away from coastal areas [54]. A total of 27 cities within the Yangtze River Delta region were included in the analysis. High-resolution Google Earth satellite imagery and LULC data [55] were utilized for visual identification, ensuring that the selected spaces met the criteria for urban blue–green spaces. In total, 162 urban blue–green spaces were selected for analysis. The boundaries of these spaces were delineated using ArcGIS Pro 3.1 [56], as shown in Figure 2c.

Figure 2. Land cover map of Yiwu City, Zhejiang Province (a), Land surface temperature (b), and cooling effect of green space taking the Futian Wetland Park as an example (c,d). (c) Enlarged green space in Futian Wetland Park, with Google Image as the background; (d) Enlarged cooling effect of green space in Futian Wetland Park, with land surface temperature as the background. The wight line represents the border of the green space, and the blue line represents the cooling distance of the green space.

2.2. Data and Preprocessing

The data used in this study were primarily sourced from and processed using the Google Earth Engine platform. Detailed information on the data sources, descriptions, and preprocessing steps can be found in Table 1. Specifically, Landsat satellite imagery was processed through radiometric calibration, cloud masking, index calculation, and temporal synthesis to generate long-term time series of LST and NDBI. All data underwent standard preprocessing steps, including spatial clipping and resampling. Landscape pattern indices were calculated using tools such as FRAGSTATS [57].

Table 1. Data collection and preprocessing.

Data	Data Resource	Data Description	Pretreatment
Administrative division data	Downloaded from the RESDC cloud platform (http://www.resdc.cn/, accessed on 1 April 2025)	Provides study area’s boundaries.	-
Land use and land cover (LULC)	GLC_FCS10 [55]	Global land use raster data for 2023 at 10 m resolution, generated based on Sentinel-1 and Sentinel-2 imagery.	Used directly as input for calculating landscape pattern indices; Figure 2a shows the land classification map of Yiwu City, Zhejiang Province.
Land surface temperature (LST)	Landsat 8 Collection 2 Level 2 https://www.usgs.gov/landsat-missions/landsat-8, accessed on 1 April 2025	Average summer (Jun–Aug) land surface temperature for the Yangtze River Delta, 2020, 30 m resolution.	Processed in Google Earth Engine (GEE): Cloud filtering, radiation calibration and brightness temperature conversion, Kelvin to Celsius conversion, mean compositing and clipping
Normalized difference built-up index (NDBI)	Landsat 5/7/8 Collection 2 Level 2 https://www.usgs.gov/landsat-missions, accessed on 1 April 2025	Long-term series of built-up indices (1986–2023) derived from Landsat series. Summer (Jun–Aug, 2020) average images for the Yangtze River Delta.	Multi-temporal image merging and filtering, radiation calibration and cloud masking, batch index calculation, compositing and clipping
Albedo	MODIS MCD43A3 Version 6.1 https://lpdaac.usgs.gov/products/mcd43a3v061/, accessed on 1 April 2025	Average summer (Jun–Aug, 2020) shortwave albedo for the Yangtze River Delta, 500 m resolution.	Temporal filtering (summer), median compositing and scaling (×0.001)
Precipitation (PRE)	ECMWF ERA5-Land Monthly Dataset	Total summer (Jun-Aug) precipitation for the Yangtze River Delta area, 2020, ~9 km (0.1°) resolution.	Calculation of summer total (sum), unit conversion (m to mm), clipping to study area
Digital elevation model (DEM)	SRTMGL1 Version 3 (USGS/SRTMGL1_003) https://lpdaac.usgs.gov/products/srtmgl1v003/, accessed on 1 April 2025	NASA SRTM Global 1 arc-second (approx. 30 m resolution). Global elevation data collected during the Shuttle Radar Topography Mission in 2000.	-
Tree height (TH)	Meta Sub-meter Global Canopy Height	Forest canopy height for the Yangtze River Delta area (~2020), ~1–10 m resolution.	-
Building height (BH)	European Commission JRC GHSL P2023A GHS-BUILT-H https://ghsl.jrc.ec.europa.eu/, accessed on 1 April 2025	Global built-up height data based on multi-source remote sensing and census data, 30 m resolution.	-
NTL	https://payneinstitute.mines.edu/eog/, accessed on 1 April 2025	2020 year, 500 m	-
POP	https://www.worldpop.org, accessed on 1 April 2025	2020 year, 100 m	-

2.3. Quantification of Cooling Intensity

As shown in Figure 2, distinct heat accumulation zones are observed in densely populated commercial districts, residential areas, and large public spaces, based on the land surface temperature spatial pattern. In contrast, cooler areas are associated with ecological infrastructures, such as urban parks, rivers, and lakes. These landscapes exhibit lower surface temperatures, functioning as cooling islands within the urban environment. For example, Futian Wetland Park in Yiwu City, Zhejiang Province, contains a prominent cool island (indicated in blue) that helps mitigate the urban heat island effect. This study employs a novel area-adaptive method to determine buffer zones [16]. Assuming circular green spaces, the cooling distance is defined by adjusting the buffer area to be approximately equivalent to the urban blue–green space area. The area-adaptive method avoids the drawback of using a fixed threshold for different sizes of UBGS [40] and the inability to determine the cooling distance of small UBGS using the turning point method [58], making the calculation more efficient and convenient. The cooling zone defined in this study is illustrated in Figure 2c. Given the urban blue–green space area $S$ and buffer distance $l$, the calculation formula is:

$$l=(\sqrt{2}-1)\times \sqrt{{\displaystyle \frac{S}{\pi }}}$$

(1)

This method has been shown to rapidly and effectively quantify CI [16,19]. Subsequently, the CI of urban blue–green spaces is defined as the difference between the LST of these spaces and their surrounding areas [59], where $LS{T}_B$ represents the average LST of the buffer zone and $LS{T}_{UGBS}$ denotes the average LST of the urban blue–green space. The formula for calculating CI is:

$$CI=\Delta LST={LST}_B-{LST}_{UBGS}$$

(2)

2.4. Potential Landscape Factors and Description

Numerous factors mitigate the urban heat island effect. Two-dimensional (2D) elements of the urban landscape, such as variations in the size, shape, and clustering of landscape features, influence the thermal environment and energy flux, thereby enhancing spatial heterogeneity [60,61]. Three-dimensional (3D) factors, including building height, volume, and floor area ratio, also play a significant role in shaping the urban climate [61,62]. Population density and nighttime light data, which reflect human activity and socioeconomic conditions, have been shown to correlate with urban thermal environments [12,63].

Building upon these three dimensions, this study incorporates a buffer-to-urban-green-space difference metric to ensure a comprehensive examination of variables. Ultimately, 15 indicators were summarized for both the interior and surrounding buffer zones of urban blue–green spaces (Table 2). Multicollinearity among variables was assessed using variance inflation factors (VIF), with severe multicollinearity typically indicated by VIF values exceeding 10 [64]. The results show no significant multicollinearity among the indices.

Table 2. Summary of CI potential driving factors.

Category	Influencing Factors (Abbreviation)	Formula and Range	Definition
2D factors	Area (Area)	$A_P\ge 1$	The area of each urban blue–green space (ha)
	Normalized difference built-up index in the buffer (NDBI_b)	$NDBI={\displaystyle \frac{SWIR-NIR}{SWIR+NIR}}$, $-1\le NDBI\le 1$	Measures of the building distribution of buffer
	Albedo (Albedo)	$0<Albedo<1$	The ratio of reflected solar radiation flux to the incident solar radiation flux
	Precipitation (PRE)	$PRE\ge 0$	The precipitation of each urban blue–green space
	Digital elevation model (DEM)	-	The digital elevation model of each urban blue–green space
	Landscape shape index of lake/river flat (LSI_LR)	$LSI={\displaystyle \frac{P}{2\times \sqrt{\pi A}}}$, $LSI\ge 1$	Landscape shape index of lake/river flat within the urban blue–green space (LSI_LR) and urban impervious surfaces in the buffer (LSI_UIS_b)
	Landscape shape index of urban impervious surfaces in the buffer (LSI_UIS_b)
	Landscape shape complexity of water bodies (ΔLSI_W)	$\Delta LSI={LSI}_b-{LSI}_{UBGS}$	The difference in the landscape shape complexity of water bodies between the buffer and the interior of the blue–green space
	Largest patch index of forests (LPI_F)	$LPI={\displaystyle \frac{a_{max}}{A}}$, $0<LPI\le 100$	Largest patch index of forests (LPI_F) and grassland (LPI_G) within the urban blue–green space
	Largest patch index of grassland (LPI_G)
	Proportion of water bodies (ΔPLAND_W)	$PLAND={\displaystyle \frac{100}{A}}\times {\sum }_{i=1}^n{W}_i,$ $\begin{array}{cc}& \Delta PLAND\hfill \\ & ={PLAND}_b\hfill \\ & -{PLAND}_{UBGS}\hfill \end{array}$	The difference in the proportion of water body area between the buffer and the interior of the urban blue–green space (%)
3D factors	Tree height in the buffer (TH_b)	$TH\ge 0$	The average tree height in the buffer (m)
	Building density in the buffer (BD_b)	$BD={\displaystyle \frac{{Area}_{bba}}{{Area}_{pc}}}$; $BD\ge 0$	The ratio of the total area of building base in buffer to the buffer area
Socio-economic factors	Nighttime light in the buffer (NTL_b)	$NTL\ge 0$	The mean nighttime light in the buffer of the green space (×10−9 Wcm−2sr−1)
	Population in the buffer (POP_b)	$POP\ge 0$	The mean population density in the buffer of the green space (person/m2)

2.5. Methods

The main workflow of this study is illustrated in Figure 3. Data sources and preprocessing steps have been described earlier. This section primarily focuses on the XGBoost machine learning model and the SHAP explanation method.

Figure 3. Flowchart of this study.

The Pearson correlation coefficient, a standard method for analyzing correlations, effectively assesses the degree of linear association between two variables. For example, it has been used to reveal correlations between urban green landscape pattern characteristics and UHI or LST [65,66]. In this study, Pearson coefficients between CI and all potential influencing factors listed in Table 1 were sequentially calculated using SPSS 26. More detailed theories and formulas regarding Pearson coefficients can be found in the literature [67].

Given the complexity and nonlinearity of urban thermal processes, machine learning methods were employed to complement traditional correlation analysis. XGBoost is a gradient-boosting-based ensemble learning technique that offers high computational efficiency and predictive accuracy. The algorithm incorporates regularization strategies and an efficient gradient descent optimization mechanism. Specifically, XGBoost applies both L1 (Lasso) and L2 (Ridge) regularization to leaf weights, which penalizes model complexity and mitigates overfitting. This approach not only enhances model precision but also effectively prevents overfitting [46]. It has been widely applied in recent studies to identify key factors influencing thermal intensity [47,68]. The XGBoost regression model [46] was implemented using the Python 3.13 package xgboost, with CI as the dependent variable and potential influencing factors as independent variables. The dataset was randomly divided, with 70% used for model training and 30% for validation, and explained variance was employed to evaluate model performance. During model training, several key hyperparameters were optimized, including the number of trees (n_estimators), maximum tree depth (max_depth), learning rate (learning_rate), subsampling ratio (subsample), feature sampling ratio (colsample_bytree), and the minimum loss reduction required for node splitting (gamma). The optimal hyperparameter configuration was determined using 10-fold cross-validation to enhance model robustness and generalization ability. Subsequently, the SHAP method was applied to interpret the contributions of individual influencing factors to CI. As a local interpretation technique, SHAP computes Shapley values to quantify the relative importance of each explanatory variable, enabling an assessment of both feature importance in model predictions and the contribution of individual samples [44]. PDP is a model-agnostic interpretation method that illustrates the average marginal effect of one or two features on the predicted outcome while averaging out the influence of all other variables [44]. The revised text clarifies that PDP is used to visualize the marginal effect of a feature on the target prediction, which helps assess whether the feature–target relationship appears approximately linear, monotonic, or nonlinear [69].

3. Results

3.1. Descriptive Statistics

Table 3 presents the descriptive statistics for all influencing factors. The average CI value for urban blue–green spaces across 160 cities was 4.85 °C, with a standard deviation of 1.44 and a range from 1.85 °C to 11.11 °C. Significant variations exist in the performance of blue–green spaces across different cities in terms of mitigating local heat island effects. The average area of the urban blue–green spaces selected for this study was 43.65 hectares, with a range from 1.32 to 504.39 hectares. The average NDBI_b (Neutral Dry Bulb Index) of the buffer zone was −0.17, ranging from −0.33 to −0.01, indicating relatively low surface hardening in the buffer zones. The mean albedo value was 0.13. The average precipitation (Pre) value was 29.98 mm, suggesting favorable rainfall conditions that promote vegetation growth and indirectly enhance cooling effects. The mean Digital Elevation Model (DEM) elevation was 27.02 m, ranging from −2.70 to 241.59 m. Regarding landscape morphology, the average landscape shape index for internal lakes and rivers (LSI_LR) was 4.12 (with a standard deviation of 3.14), indicating relatively regular water body shapes. The average urban impervious surface shape index (LSI_UIS_b) in the buffer zone was 5.20, reflecting high complexity and fragmentation in the surrounding built environment, which may negatively impact the cooling performance of green spaces. The difference in landscape shape complexity between internal and external water bodies (ΔLSI_W) reveals heterogeneity in landscape configuration between green spaces and their surroundings, with external water bodies exhibiting greater fragmentation. The average water proportion difference (ΔPLAND_W) was −39.55, indicating a significantly higher proportion of water bodies within urban blue–green spaces compared to external areas. Extensive water bodies play a positive role in regulating local microclimates.

Table 3. Descriptive statistics of the thermal and explanatory variables.

Variable	Mean	Range	Standard Deviation
CI	4.85	1.85–11.11	1.44
2D factors
Area	43.65	1.32–504.39	73.64
NDBI_b	−0.17	−0.33–−0.01	0.06
Albedo	0.13	0.00–0.18	0.02
Pre	28.98	0.00–41.06	6.30
DEM	27.02	−2.70–241.59	36.37
LSI_LR	4.12	0.00–17.54	3.14
LSI_UIS_b	5.20	0.00–21.12	3.61
ΔLSI_W	1.16	−3.94–7.24	2.21
LPI_F	7.97	0.00–87.37	19.59
LPI_G	0.51	0.00–10.81	1.07
ΔPLAND_W	−39.55	−85.96–5.51	29.95
3D factors
TH_b	0.65	0.00–5.48	0.74
BD_b	86.63	0.01–1766.83	161.46
Socio-economic factors
NTL_b	20.71	0.44–72.10	14.32
POP_b	25.95	0.00–146.53	20.12

Among the 3D factors, the mean TH_b value is 0.65 m, with a range from 0 to 5.48 m, indicating generally low values. This suggests sparse external vegetation coverage and limited shading capacity. The mean building density (BD_b) is 86.63, with a maximum value of 1766.83, reflecting relatively high values, with some green spaces surrounded by high-density urban development. High-density and high-rise buildings may obstruct ventilation and increase anthropogenic heat emissions, thereby counteracting the cooling effects of green spaces.

The socioeconomic indicators, Nighttime Light Intensity (NTL_b, standard deviation 14.32) and Population Density (POP_b, standard deviation 20.12), exhibit large standard deviations, indicating significant variations in human activity intensity and socioeconomic development levels surrounding different green spaces. These factors typically show a negative correlation with CI and may amplify the “background pressure” on the cooling functions of green spaces. Specifically, the more intensely developed the surrounding area, the more susceptible the cooling effect of green spaces becomes to suppression by the external thermal environment.

3.2. Correlation and Feature Contribution Analysis

Pearson correlation coefficients were used to analyze the linear relationships between various landscape indices and CI (Figure 4). The cooling effect of urban blue–green spaces is influenced not only by their internal ecological attributes but also by their urban environmental context. The results indicate that Area, NDBI_b, Pre, LSI_UIS_b, ΔLSI_W, and LSI_LR are positively correlated with CI, suggesting that larger green space areas, higher precipitation, greater patch aggregation and connectivity, and lower fragmentation all enhance cooling effects.

Figure 4. Pearson Correlation Matrix Between Predictors and CI.

It is widely recognized that large-scale green spaces exhibit superior cooling effects due to their larger canopy and leaf areas, which more effectively intercept solar radiation and provide shade. Higher precipitation values help sustain healthy plant growth in green spaces and enhance transpiration cooling efficiency, thereby intensifying cooling effects. The positive correlation between ΔLSI_W and CI reflects the difference in the shape indices of external versus internal water bodies, indicating that more regular internal landscape structures favor the formation of stable cooling zones. When water bodies exhibit highly regular and simple shapes—such as near-circular or square lakes—they better maintain their lower temperatures, creating a stable and potent “cold core.” This pattern reduces thermal radiation dispersion and enhances the localized regulatory capacity of water bodies and vegetation.

The impervious surface index (NDBI_b) and impervious landscape shape index (LSI_UIS_b) of buffer zones show significant positive correlations with CI. This finding suggests that the cooling effect of green spaces depends not only on their internal ecological quality but also on their thermal contrast with the surrounding environment. When buffer zones exhibit higher levels of surface hardening or more complex built-up patterns, their surface temperatures are relatively higher, increasing the temperature difference with the interior of green spaces and thereby enhancing the “cool island” effect.

Albedo, DEM, ΔPLAND_W, BD_b, TH_b, NTL_b, and POP_b were negatively correlated with CI. Due to the large proportion of water bodies within green spaces in the study area, vegetation coverage was relatively low, weakening the positive contribution of Albedo to cooling effects. A smaller difference in the ratio of internal to external water bodies resulted in better cooling performance. External building density (BD_b) negatively correlates with CI, further validating that high-density urban development suppresses green space cooling functions. Dense building clusters not only obstruct ventilation but also increase anthropogenic heat emissions, limiting the effective propagation of green space cooling effects. Both NTL_b and POP_b exhibit negative correlations with CI. High-intensity human activities significantly increase external thermal loads, diminishing the cooling effects of green spaces.

Urban planning must holistically consider both the external ecological quality of blue–green spaces and the urban environment to maximize the cooling benefits of green spaces.

This study employs XGBoost to identify nonlinear relationships between the physical environment within and surrounding urban parks and their cooling intensity. Subsequently, SHAP models are used to interpret the predictions generated by XGBoost. The training and testing sets achieved R² values of 0.924 and 0.607, respectively (Figure 5). The testing R² of 0.607 reflects moderate but meaningful predictive power, which is considered acceptable for modeling complex environmental processes influenced by multiple interacting factors.

Figure 5. Comparison of CI Prediction Results Between XGBoost Training and Test Sets.

Figure 6a illustrates the relative importance of variables for CI prediction. Based on the average SHAP values across all factors, park area (Area: 0.38) is significantly more important than the other factors. The second most important factor is buffer NDBI (NDBI_b: 0.16), followed by the LSI of internal lakes/rivers (LSI_LR: 0.07), tree height in the buffer (TH_b: 0.06), population in the buffer (POP_b: 0.06), and landscape shape index of urban impervious surfaces in the buffer (LSI_UIS_b: 0.05). In contrast, the importance of factors such as the difference in water body LSI (ΔLSI_W: 0.01), difference in water body area proportion (ΔPLAN_W: 0.02), and surface reflectance (Albedo: 0.03) is relatively low.

Figure 6. Mean SHAP (a) and SHAP value (b) of influencing factors.

In terms of importance distribution, at the urban agglomeration scale, the size of UBGS is the primary factor in explaining CI variation. Additionally, the intensity of surrounding development constitutes a key external background condition. In contrast, indicators such as ΔLSI_W, Albedo, and ΔPLAN_W rank at the lower end of importance, suggesting that these factors have a diminished average contribution in this study’s sample and scale. Their effects are more likely to be reflected through interactions with other variables, which should be further validated through subsequent PDP analysis.

Figure 6b displays a SHAP summary plot with variables sorted in descending order of relative importance. Each point represents a data sample, with the x-axis position indicating the SHAP value for that point. Points on the right side indicate a positive correlation between the feature and the prediction, while points on the left side indicate a negative correlation. Point color reflects the magnitude of the variable’s feature value: red indicates high values, and blue indicates low values. For example, for LSI_LR, low values (blue) are predominantly distributed on the left, while high values (red) are mainly on the right. Thus, increasing LSI_LR can enhance its CI. LSI_LR exhibits a long “right tail” and a short “left tail,” indicating that higher LSI_LR values have a relatively minor impact on CI, whereas lower LSI_LR values can significantly reduce CI. The Area variable is predominantly blue on the left side and red on the right side, with the largest horizontal span, indicating a positive contribution of increasing area to CI, which is much higher than other variables. Larger areas typically correspond to more pronounced cooling cumulative effects. LSI_LR shows a similar red-blue distribution pattern to the park area, but with a smaller span, suggesting its contribution is less significant compared to area. The red tail on the right side is longer than the blue tail, indicating that when LSI_LR is at higher levels, its positive contribution to CI becomes significantly amplified. For LSI_UIS_b, the red points are clearly clustered on the right, while the blue tail on the left is longer. When LSI_UIS_b is higher, its contribution to CI is more likely to be positive and concentrated.

3.3. Marginal Effects of Predictor Variables

Partial Dependency Plots (PDPs) were used to analyze the marginal effects of each potential factor on CI (Figure 7). While Pearson correlation analysis had already demonstrated correlations between the 15 variables and CI, marginal effect analysis further revealed the nonlinear nature of these relationships. Additionally, the nonlinear response patterns among different indicators exhibited significant differences.

Figure 7. Nonlinear relationships between influencing factors and CI.

Among the top three ranked variables, area showed a positive correlation with CI when ranging from 0 to 50 hm². However, when the area exceeded 50 hm², CI plateaued at its maximum value of 6 °C and ceased to change. Thus, larger park areas do not necessarily result in greater cooling effects. NDBI_b exhibits a positive correlation with CI. When NDBI_b is below −0.2, CI increases significantly faster than when NDBI_b exceeds −0.2. Higher hardscaping levels in buffer zones intensify thermal contrast with internal green spaces, amplifying CI. Green spaces deliver particularly pronounced cooling benefits in low-development-intensity areas. However, in highly urbanized external environments, while cooling effects persist, their growth rate slows. LSI_LR shows a positive correlation with CI, but its influence on CI does not increase continuously; instead, it exhibits stepwise growth. In urban blue–green space design, more complex water body edges are not necessarily better; LSI_LR within the 5–8 range achieves optimal cooling benefits. Beyond a certain edge complexity threshold, its contribution to cooling stabilizes. Among other variables, ΔPLAND_W shows a clear negative correlation with CI: larger water areas within urban blue–green spaces yield greater cooling effects. When Pre falls within the 22–30 mm range, CI exhibits a negative correlation with Pre. This indicates that during periods of low precipitation, vegetation transpiration in green spaces is constrained, and the evaporative cooling effect of water bodies is also insignificant, resulting in an overall weaker cooling capacity. However, when Pre increases to the 30–32 mm range, CI rises rapidly with increasing precipitation and then stabilizes. NTL_b and BD_b show a clear negative correlation with CI, indicating that the urban built environment significantly suppresses the cooling function of green spaces.

3.4. Interaction Effects Between Influencing Factors

Through 2D and 3D Partial Dependence Plot (PDP) analysis, the synergistic effects of the following factors on CI were examined (Figure 8). When large UBGS (Area > 50 hectares) are combined with high internal water coverage (ΔPLAND_W < −60), the cooling effect of UBGS are significantly enhanced through vegetation transpiration and large-scale water evaporation. When both high river/lake boundary complexity (LSI_LR > 7.5) and high grassland patch dominance (LPI_G > 1) are present within UBGS, CI increases significantly, indicating a clear synergistic effect between these two factors in enhancing cooling. Under low precipitation (Pre) conditions, even with complex water body boundaries, evaporation is limited. However, under high precipitation conditions, complex boundaries can more effectively utilize water for evaporative cooling. As NDBI_b increases, the cooling capacity of the green space is greatly enhanced. High NDBI_b indicates large surface heat capacity, while high LSI_UIS_b signifies disordered building layouts and poor ventilation. When both rise, heat accumulates easily and dissipates poorly through natural convection, leading to sustained local temperature increases. Maximizing the cooling effect of urban blue–green spaces requires an integrated consideration of internal structural optimization and external environmental control.

Figure 8. Interaction Effects Between influencing factors and CI.

4. Discussion

4.1. Key Factors Influencing the Cooling Effect of Urban Blue–Green Spaces

This study systematically evaluated the summer cooling intensity of 162 UBGS in the Yangtze River Delta region, with an average CI of 4.85 °C. The cooling intensity is the result of multiple contributing factors.

This study confirms a significant positive correlation between urban blue–green space (UBGS) area and CI, with larger areas generating stronger CI—consistent with findings in the existing literature [11,70]. Larger areas typically contain more trees and greater biomass, making them more likely to form cooling islands due to enhanced shading and evapotranspiration effects [21,43]. Furthermore, larger areas exhibit more stable microclimates and greater resilience against adverse thermal stress from surrounding built environments [13]. However, CI does not increase linearly with area but follows a logarithmic pattern. Consequently, cooling effects increase at diminishing rates until spatial scale reaches a threshold. This logarithmic relationship between CI and area has been documented in numerous studies [23,58]. However, different studies have proposed varying area thresholds: for example, some studies indicate that cooling intensity reaches its peak when the area exceeds 10 hectares [23], while others suggest that CI stabilizes after exceeding 20 hectares [71]. In contrast, this study estimates that the threshold for cooling efficiency in the Yangtze River Delta region is approximately 50 hectares. These differences may arise from the interaction of regional climate conditions, vegetation structure, urban morphology, and other factors, as the same threshold may not apply to all urban systems. Future research should further quantify threshold ranges in different regions and, by integrating remote sensing and field observation data, reveal the biophysical mechanisms behind the nonlinear relationship between area and cooling effects, such as the saturation characteristics of evapotranspiration and shading effects [27,72].

Research indicates that the percentage of water coverage within urban blue–green spaces (UBGS) positively correlates with CI, a mechanism attributed to the potent cooling effect provided by water through evapotranspiration, consistent with prior findings [13,73]. Concentrated patches experience less thermal stress from surrounding built environments, maintaining better health and higher evapotranspiration rates. More densely concentrated green spaces typically remain cooler than fragmented green spaces of equivalent total area [74]. However, high relative humidity can also intensify thermal sensation, making the ambient temperature feel higher, often referred to as “mugginess.” Future studies should consider the interactions between cooling intensity, humidity levels, and thermal comfort to provide a more comprehensive understanding of the thermal environment and guide the design of urban blue–green spaces to optimize comfort.

CI is also influenced by the buffer zone environment [75]. Studies reveal a significant negative correlation between building density (BD) and CI, as buildings alter energy balance patterns, directly affecting land surface temperature (LST) [37]. Additionally, surrounding structures impede the diffusion of cool air within parks [76]. Previous research indicates a quadratic polynomial relationship between BD and CI [77]. As BD increases, cooling intensity gradually rises, but decreases when BD exceeds a tipping point. This tipping point is determined by the amount of anthropogenic heat released in the neighborhood and the blocking effect of building form. However, since high-temperature buffer zones may impose heat stress on park vegetation, thereby reducing its cooling effect [21], this study found that the relationship between buffer zone building density (BD_b) and blue–green space CI exhibits a quadratic polynomial (V-shaped curve), which contradicts previous findings. Specifically, as building density increases, CI initially decreases gradually, but after reaching a turning point at approximately 80 (BD_b = 80), CI begins to increase with further increases in building density.

Studies also found that the shape index of impervious surfaces within the buffer significantly influences CI, consistent with previous research [37,42]. The normalized difference built-up index (NDBI) is considered an indispensable factor in China’s urban temperature rise; higher densities of buildings or impervious surfaces capture and store more heat [78]. However, this study found that NDBI’s impact on the Yangtze River Delta region is not particularly significant. This may be attributed to the separation between built-up areas and natural land, resulting in insufficient spatial variation in the built index to significantly influence CI [78]. Similar findings have been reported by previous scholars [79].

Socioeconomic factors are widely recognized as the primary drivers behind urban heat islands [80,81]. Therefore, this study selected multiple socio-related factors, such as population density (POP) and nighttime light intensity (NTL), to analyze the factors influencing CI. The findings reveal that NTL exerts a significant negative effect on CI. This may stem from NTL’s strong correlations with human activity, population density, socioeconomic development, and regional urbanization. Higher NTL levels indicate greater anthropogenic heat emissions, which directly intensify the heat island effect [78]. Furthermore, increased heat storage in buildings and roads during the day leads to greater heat release at night [78,79], further diminishing CI.

4.2. Implications for Sustainable Urban Planning

As China’s economic strategy unfolds, urban clusters continue to expand, accompanied by various environmental challenges and the impacts of climate change [82]. Therefore, exploring heat mitigation strategies at the regional level may yield insights that transcend single-city research [18,41]. It is crucial to strictly protect existing large-scale blue–green spaces, as they serve as irreplaceable regional cooling assets. When designing urban blue–green spaces, priority should be given to concentrated, large-scale water bodies and vegetation patches with relatively regular shapes to maximize evapotranspiration cooling effects. Additionally, efforts should be made to actively increase canopy coverage and water body coverage in both new and existing parks. To mitigate the thermal impact on surrounding environments, impervious surfaces in buffer zones should be reduced, and green infrastructure (e.g., green roofs, street trees) should be integrated into the urban fabric. This approach lowers background land surface temperature (LST) and extends the reach of urban blue–green spaces as cooling islands.

4.3. Limitations and Future Work

In this study, we quantified CI and identified the effects of factors within the urban blue–green space (UBGS) area and its buffer zone. This research has several limitations. First, this study focuses on near-field cooling intensity and does not examine the distance-decay effect beyond the buffer zone or include comparisons with non-UBGS control areas. Extending the analysis to these spatial dimensions represents an important direction for future research. Meteorological conditions such as wind speed, air temperature, and relative humidity are known to influence CI, yet this study did not account for these factors. Future research should incorporate climate variability [75]. Furthermore, to validate the accuracy of our model, future work could involve installing in situ temperature sensors to collect ground-truth data, as demonstrated in Boqué-Ciurana et al. [83], who used such sensors to reschedule human tower exhibitions based on thermal comfort. This approach would provide empirical validation and enhance the robustness of the findings. Since our dataset only covers land surface temperature data under clear-sky conditions from June to August 2020, and does not include data from other seasons, the interpretation of results for other seasons may be limited and should be treated with caution. It is recommended to integrate year-round temperature, weather, and land cover data to generate results for each season, as well as annual averages. Future studies could further expand to multi-year datasets to analyze CI variations across different years, thereby capturing long-term trends and the impact of climate change on the cooling effect of UBGS. Additionally, future research should consider other potential drivers, such as traffic density, industrial activities, and urban design characteristics, to gain a more comprehensive understanding of the formation mechanisms of the urban heat island effect.

5. Conclusions

This study provides a scientific basis for effectively planning urban mitigation of extreme heat effects through a systematic analysis of the cooling intensity and its drivers across 162 urban blue–green spaces in the Yangtze River Delta region. The key findings are as follows.

Using Pearson correlation analysis, the XGBoost machine learning algorithm, and the SHAP explanation method, we identified the complex relationships between 15 potential influencing factors and CI. The top five factors in terms of relative importance are: Area (0.38), NDBI in the buffer (0.16), Landscape shape index of lake/river flat (0.07), Tree height in the buffer (0.06), and Population in the buffer (0.06). Among the 2D landscape factors, Area shows a significant positive effect on CI, followed by NDBI in the buffer and LSI of rivers and lakes within UBGS. Higher water body coverage and a lower water body shape index within UBGS promote the formation of stable cooling zones. Regarding 3D landscape factors, Building density shows a significant negative correlation with CI, as dense building clusters not only obstruct ventilation but also increase anthropogenic heat emissions. Socioeconomic indicators also significantly influence CI, with NTL_b and POP_b showing strong negative correlations. This reflects how urbanization intensity suppresses the cooling effects of blue–green spaces.

Therefore, in practical green space planning and management, prioritizing the scale and internal landscape structure of UBGS is essential. Increasing the edge length and area of water bodies maximizes cooling effects, making core area expansion the most valuable strategy for enhancing cooling performance. Additionally, reducing hard surface ratios, optimizing three-dimensional spatial configurations, and appropriately controlling surrounding building layouts and heights can significantly mitigate the urban heat island effect.