An XGBoost–SHAP Explainable Machine Learning Approach Linking Urban Park Cooling Effects in a Megacity Core to Landscape Features Inside and Outside the Parks

Abstract

Urban parks, as essential components of green infrastructure, play a critical role in mitigating the urban heat island effect and improving thermal comfort. This study evaluated the cooling performance of 170 parks in the urban center region of Beijing using seven cooling effect indicators. An integrated framework combining XGBoost, SHAP, and Ordinary Least Squares was proposed to examine the nonlinear relationships and threshold effects between cooling effect indicators and landscape features. Moreover, the Elbow method and K-means clustering were applied to classify parks based on their cooling performance. The results showed that both internal and external landscape features are closely associated with park cooling effects. Eight features were identified as key landscape features, possessing higher relative importance. Pronounced nonlinear relationships and threshold effects were identified. For example, the optimal park area is approximately 25 hm², while more regular shapes enhance cooling performance, with an optimal ratio of perimeter and area of about 0.05 m⁻¹. Furthermore, four distinct park types with significantly different cooling performance were identified, reflecting different capacities for regulating urban thermal environments. These findings improved the understanding of park cooling effects and provided guidance for optimizing park design in megacity cores.

1. Introduction

With the rapid urbanization, approximately 45% of the global population resides in urban areas [1], and this proportion is expected to continue increasing in the future. In particular, megacities, commonly defined as the metropolitan areas with total populations exceeding 10 million [2], have emerged as a representative urban form worldwide characterized by dense populations, intensive land development, and complex built environments. However, urban expansion has substantially altered land cover and landscape patterns [3], which in turn modifies the thermal properties of urban surfaces and the structure of the near-surface atmosphere [4]. These changes disrupt the energy balance at the land–atmosphere interface, leading to the urban heat island effect, characterized by higher temperatures in urban areas compared to their surrounding rural regions. Under the influence of climate change, the impacts of the UHI effect have become increasingly pronounced [5,6], especially in densely built megacity cores [7]. The frequency and intensity of extreme heat events have risen [8,9], reducing thermal comfort and increasing heat exposure risks, with consequent threats [10] to both public health and socioeconomic stability [11,12]. As an important component of urban green infrastructure [13], urban parks integrate several types of blue–green spaces, including forests, grasslands, wetlands, and water bodies. Through mechanisms such as shading, evapotranspiration, increased surface albedo, and enhanced ventilation [14,15,16], parks can effectively reduce temperatures and increase humidity in both park interiors and surrounding urban areas, thereby mitigating the UHI effect. Consequently, the cooling effect of urban parks has become as a hotspot in studies of the urban thermal environment.

Currently, many studies have examined the cooling effect of urban parks and developed quantitative indicator systems to characterize their cooling performance [17,18,19]. Based on the buffer analysis method, a series of metrics were proposed to capture cooling effects from multiple dimensions [20], including park cooling intensity (PCI), cooling area (PCA), cooling gradient (PCG), cooling efficiency (PCE), maximum cooling distance (PD), and cooling service (PCS). The indicator systems provide a comprehensive framework for evaluating park cooling performance. Based on such assessments, many studies further explored the variables (influencing factors) associated with park cooling effects and their potential driving mechanisms. Internal landscape features of urban parks were widely confirmed to be closely associated with cooling effects, including size attributes (e.g., park area and perimeter [21,22,23]), shape attributes (e.g., ratio of perimeter to area and shape index [24,25,26]), and landscape composition attributes (e.g., area proportions of green space, water bodies, and impervious surface [22,27,28,29]). External surrounding features were also identified as an important correlated factor, encompassing building density, impervious surface scale, ventilation conditions, and the area proportions of surrounding blue–green spaces [18,30,31,32]. In addition, recent studies extended the analysis from 2D to 3D variables [33], incorporating indicators such as sky view factor and building height, thereby enriching the analytical framework. In terms of time scale, with the growing availability of multitemporal remote sensing data, some studies expanded from single daytime explorations in summer [34] to multiple periods within a day and across different seasons [20], providing a more comprehensive understanding of park cooling performance.

Despite these advances, studies on the role of spatial scale in urban park cooling effects remain limited. Existing studies generally did not distinguish park scale [35,36]. A recent study across different scales suggested that the relative importance of landscape features on cooling effects may vary with park scale, indicating the potential scale-dependence mechanism [37]. Although individual studies began to focus on small scale parks, such as pocket parks, or conducted cross-scale comparisons [38,39], systematic investigations of urban park cooling effects at the medium and large scale, particularly community and comprehensive parks closely related to daily life, remain scarce. This gap constrained a more comprehensive understanding of urban park cooling effects at the medium and large scale and its implications for urban planning.

In terms of methodology, previous studies commonly employed correlation analysis and linear regression [15,40,41] to examine the relationships between landscape features and park cooling effect indicators. These approaches can effectively identify the direction of associations and exploring potential threshold values [42], and thus provide strong interpretability and practical relevance. For example, Feng et al. (2023) [43] found that for PCA, the thresholds of park area and perimeter were approximately 30 hm² and 4000 m, respectively, while a shape index of 2.8 was associated with optimal PCE. However, park cooling effects are typically regulated by multiple interacting variables [36], and their responses often exhibit nonlinear patterns. Under such multivariate conditions, linear regression models may be limited in capturing complex relationships and nonlinear responses among variables. In recent years, machine learning methods, such as Random Forest and eXtreme Gradient Boosting (XGBoost), have been increasingly applied in urban thermal environment studies due to their strong predictive performance and flexibility in handling nonlinear relationships [37,44]. Nevertheless, conventional machine learning models are frequently criticized as “black-box” approaches because their internal decision-making processes and prediction results are difficult to interpret, which has raised concerns regarding their application in urban planning and environmental management [45]. To address this issue, explainable machine learning methods have emerged as an important analytical framework that can not only improve prediction accuracy but also reveal the underlying mechanisms behind model outputs. Among these approaches, the SHapley Additive exPlanations (SHAP) algorithm has been widely adopted as a post hoc interpretation technique for quantifying the contribution and interaction effects of individual features, which enables researchers to better understand how different environmental variables jointly influence urban thermal processes and to identify relative contributions and nonlinear thresholds that may be difficult to detect using conventional approaches [46]. For instance, Liu et al. (2023) [47] used XGBoost and SHAP models to analyze 814 parks in Zhejiang Province and found that park area was the most important predictor of PCI, with 4 hm² of a minimum design threshold and 20 hm² of an optimal size. Despite these advantages, explainable machine learning methods mainly describe statistical associations rather than strict causal relationships. Therefore, integrating explainable machine learning approaches with traditional statistical methods and domain knowledge can provide a more comprehensive understanding of the relationships between variables and park cooling effects, particularly in identifying nonlinear response patterns and threshold effects, and can offer more reliable support for park planning and thermal regulation.

Beijing, as a representative megacity, exhibits a highly concentrated and uneven population distribution, intensive land development, and complex urban spatial structures, characterized by a “dense center and sparse periphery” pattern. The urban center region of Beijing accounts for only approximately 8% of the total area but accommodates 10.88 million residents, nearly half of the total population. Similar to many famous megacities worldwide, the high-density built environment and extensive impervious surfaces in Beijing’s core area have intensified urban thermal environmental problems and increased population exposure to heat risks. Zhao et al. (2024) found that around 3/4 of Beijing’s population were exposed to the urban heat island, with high-risk areas primarily concentrated in the urban center region [7]. Against this backdrop, a systematic assessment of the cooling effects of urban parks in the urban center region is essential for advancing nature-based solutions to mitigate urban thermal stress. Accordingly, taking the urban center region of Beijing as the study area, this study selected 170 community parks and comprehensive parks at the medium (1–10 hm²) and large (more than 10 hm²) scale as the study objects. LST was derived from remote sensing data on the Google Earth Engine platform, and a multi-dimensional indicator system was developed to quantify park cooling effects, including PCI, PCA, PCG, PCE, PD, PCS, and the mean land-surface temperature of park (PLST). Methodologically, an integrated framework combining the XGBoost, SHAP, and linear regression models was employed to examine and interpret park cooling performance. By analyzing relative importance and marginal effects, the nonlinear relationships and threshold effects between 17 landscape features and cooling effect indicators were systematically explored. Furthermore, the Elbow method and K-means clustering were applied to classify parks, enabling a comparative analysis of their roles and advantages in mitigating urban heat islands. Specifically, this study aims to: (1) identify the key and important landscape features associated with park cooling effects; (2) investigate the nonlinear relationships and threshold effects between landscape features and cooling effect indicators; (3) classify urban parks based on their cooling performance and summarize their functional features and optimization strategies. These findings can help provide scientific support for optimizing the park configuration and enhancing urban thermal regulation.

2. Materials and Methods

2.1. Study Area

Beijing, the capital city of China, is located in the northern part of the North China Plain, bordering Tianjin to the east and Hebei Province to the west, south, and north. The topography gradually declines from northwest to southeast, with the Yanshan and Taihang Mountains surrounding the western and northern areas, while the southeastern region forms a plain sloping toward the Bohai Sea. Beijing experiences a temperate continental monsoon climate with obvious seasonal differences, cold and dry winters, hot and humid summers, an annual average temperature of approximately 12 °C and an extreme temperature of about 42 °C in summer. In recent years, rapid urban expansion in Beijing has led to high population concentration in the urban center region, forming a “dense center and sparse periphery” spatial pattern. The increasing building density and proportion of impervious surfaces have intensified the urban heat island effect leading lots of urban problems [48], particularly in the core areas. Therefore, this study focused on the urban center region of Beijing, including six districts (Dongcheng, Xicheng, Chaoyang, Haidian, Shijingshan, and Fengtai) and covering a total area of approximately 1384 km². In recent years, the region has accommodated a permanent resident population of approximately 10.988 million, with the population density exceeding 7900 people/hm². Among them, Dongcheng and Xicheng districts represent the core functional zone for capital, whereas the remaining four districts serve as urban expansion zones. To improve residents’ living environment and mitigate the urban heat island effect, the government actively promoted “Garden City” construction over recent years, implementing strategies such as pocket parks, green corridors, and ecological networks to improve the quantity, quality and equal distribution of urban blue–green spaces. According to the park classification standards of the Beijing Municipal Bureau of Landscaping and Forestry, this study selected 170 community and comprehensive parks at the medium or large scale in the urban center region as study objects and systematically analyzed their cooling effects and associated landscape features, which aimed to provide scientific guidance for urban park planning and urban thermal environment management. (Figure 1).

Figure 1. Location of the study area. (a) Map of China; (b) map of Beijing and study area, and (c) distribution of parks.

2.2. Research Framework

The research framework of this study is illustrated in Figure 2. Firstly, referring to previous study, seven indicators were utilized to quantify and characterize the cooling effect of urban parks based on a multi-ring buffer approach (Table 1). In addition, 21 landscape features potentially associated with park cooling effects were initially identified from previous studies [21,22,23,31,32]. Through correlation analysis, 17 landscape features were retained for further analysis (Table 2), which were classified into two categories: internal park features and surrounding environmental features. Next, this study adopted an explainable machine learning method to investigate the nonlinear relationships between landscape features and park cooling effects. Specifically, the XGBoost algorithm was employed because of its strong predictive capability and flexibility in modeling complex nonlinear relationships. To overcome the limited interpretability of conventional machine learning models, the SHAP method was further integrated to interpret model outputs by quantifying the relative importance of individual landscape features for the seven cooling effect indicators. Subsequently, Partial Dependence Plots were used to interpret the marginal effects of key landscape features, and Ordinary Least Squares regression was then applied to explore the threshold effects between internal park landscape features and the cooling effect. Finally, the Elbow method and K-means clustering were used to classify parks based on their cooling effect features. The landscape characteristics and cooling strengths of each park type were summarized. We expect that this comprehensive analytical framework can provide a robust scientific basis for the planning, construction, and improvement of urban parks and urban thermal environment.

Figure 2. Consecutive research steps undertaken by the study.

Table 1. The full names, meanings and formulas of cooling indicators of urban parks.

Indicators	Full Name	Meaning	Formula
PCI	Cooling intensity	The scale of drop in temperature	$\mathrm{P}\mathrm{C}\mathrm{I}=\mathrm{B}\_\mathrm{L}\mathrm{S}\mathrm{T}-\mathrm{P}\mathrm{L}\mathrm{S}\mathrm{T}$
PCA	Cooling area	Cooling coverage area	-
PCS	Cooling service	The total cooling extent within the cold island effect range	$\mathrm{B}\mathrm{G}\mathrm{C}\mathrm{S}=\mathrm{P}\mathrm{C}\mathrm{A}\times \mathrm{P}\mathrm{C}\mathrm{I}$
PCE	Cooling efficiency	The area of park that generates the cooling island effect per unit area	$\mathrm{B}\mathrm{G}\mathrm{C}\mathrm{E}={\displaystyle \frac{\mathrm{P}\mathrm{C}\mathrm{A}}{\mathrm{P}\mathrm{A}}}$
PCG	Cooling gradient	The degree of temperature drop per unit distance	The degree of temperature drop per unit distance
PLST	Park land-surface temperature	Average surface temperature of parks	-
PD	Cooling distance	The farthest cooling distance	-

Table 2. Description of landscape features related to cooling effects in parks.

Categories	Variables	Description	Formula	Unit
Internal landscape features	Park area (PA)	Area of the park	Area > 1	hm2
	Park perimeter (CPARK)	Perimeter of the park	Perimeter >0	m
	Ratio of perimeter and area (RCA)	The ratio of perimeter to the area of park	$\mathrm{R}\mathrm{C}\mathrm{A}={\displaystyle \frac{\mathrm{C}\mathrm{P}\mathrm{A}\mathrm{R}\mathrm{K}}{\mathrm{P}\mathrm{A}}}$	m−1
	Green area (GA)	Green space area of park	GA > 0	m2
	Water area (WA)	Blue space area of park	WA > 0	m2
	Impervious surface area (WTA)	Impervious surface area of park	WTA > 0	m2
	Percentage of green space (GP)	Ratio of green spaces in park	$\mathrm{G}\mathrm{P}={\displaystyle \frac{\mathrm{G}\mathrm{A}}{\mathrm{P}\mathrm{A}}}\times 100\mathrm{\%}$	%
	Percentage of water (WP)	Ratio of blue spaces in park	$\mathrm{W}\mathrm{P}={\displaystyle \frac{\mathrm{W}\mathrm{A}}{\mathrm{P}\mathrm{A}}}\times 100\mathrm{\%}$	%
	Percentage of impervious surface (WTP)	Ratio of impervious surface in park	$\mathrm{W}\mathrm{T}\mathrm{P}={\displaystyle \frac{\mathrm{W}\mathrm{T}\mathrm{A}}{\mathrm{P}\mathrm{A}}}\times 100\mathrm{\%}$	%
External landscape features	Buffer green area (B_GA)	Green space area of buffer	B_GA > 0	m2
	Buffer impervious surface area(B_WTA)	Impervious surface area of buffer	B_WTA > 0	m2
	Buffer_GP (B_GP)	Proportion of the green space in the buffer	$\mathrm{B}\_\mathrm{G}\mathrm{P}={\displaystyle \frac{\mathrm{B}\_\mathrm{G}\mathrm{A}}{\mathrm{P}\mathrm{C}\mathrm{A}}}\times 100\mathrm{\%}$	%
	Buffer_WP (B_WP)	Proportion of the blue space in the buffer	$\mathrm{B}\_\mathrm{W}\mathrm{P}={\displaystyle \frac{\mathrm{B}\_\mathrm{W}\mathrm{A}}{\mathrm{P}\mathrm{C}\mathrm{A}}}\times 100\mathrm{\%}$	%
	Buffer_WTP (B_WTP)	Proportion of the impervious surface in the buffer	$\mathrm{B}\_\mathrm{W}\mathrm{T}\mathrm{P}={\displaystyle \frac{\mathrm{B}\_\mathrm{W}\mathrm{T}\mathrm{A}}{\mathrm{P}\mathrm{C}\mathrm{A}}}\times 100\mathrm{\%}$	%
	Buffer normalized difference vegetation index (B_NDVI)	Mean NDVI for vegetated areas per buffer	$\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}}}$	-
	Gross floor area (GFA)	Total gross floor area	GFA > 0	m2
	Building area (BUA)	Base area of construction	BUA > 0	m2

2.3. Data Collection and Pretreatment

This study utilized multiple data types, including remote sensing imagery, land-use data, Areas of Interest data, and building vector data. The remote sensing imagery was acquired from Landsat 8 OLI/TIRS C2 L2 products with a spatial resolution of 30 m, covering June to September for the year of 2025. The images were under favorable atmospheric conditions with cloud cover less than 5%, ensuring the accuracy requirement of land-surface temperature (LST) retrieval. Cloud removal was performed on the Google Earth Engine platform (https://earthengine.google.com/, accessed on 28 December 2025), and the relevant bands were used to derive the LST and normalized difference vegetation index.

Vector data included the necessary information on buildings and parks. Areas of Interest data for community and comprehensive parks were obtained from Baidu Maps. Integrating the List of Parks in Beijing (2025) published by the Beijing Municipal of Forestry and Parks Bureau (https://yllhj.beijing.gov.cn/, accessed on 6 February 2026), 170 parks were identified. Meanwhile, according to the Measures for Classification and Grading Management of Parks in Beijing, the suitable size for community parks is above 1 hm², with a minimum size greater than 0.5 hm²; for comprehensive parks, the suitable size is above 10 hm². Therefore, parks with an area of 1–10 hm² were defined as medium-sized parks, and those above 10 hm² were classified as large-sized parks. In contrast, parks with an area smaller than 1 hm² were defined as small-sized parks. Building vector data were sourced from the Bigemap platform, including building outlines and heights, and total building area was calculated in ArcGIS 10.8.1.

Land-use data were obtained from Zhang et al. (2025) [49] with a spatial resolution of 10 m for the year 2023. The dataset integrates Sentinel-1 and Sentinel-2 time-series imagery and was constructed via Google Earth Engine, achieving an overall accuracy of 83.16%, superior to many other publicly available land-use datasets. The dataset contains 29 land-use categories, of which 27 types are included within the study area. On the ArcGIS 10.8.1 platform, farmland, forests, shrubs, grasslands, and other vegetation types were classified as green space, while wetlands and rivers were classified as blue space. Due to the similar albedo and summer LST performance [50,51], impervious surfaces and bare land were combined into the new impervious surface category.

2.4. Study Methods

2.4.1. Land-Surface Temperature Retrieval Using the Single-Window Algorithm

The Single-Window Algorithm is a widely used method for retrieving LST from single thermal infrared (TIR) band, particularly suitable for medium-resolution remote sensing data. By accounting for surface emissivity, sensor-observed brightness temperature, and atmospheric transmittance, it can efficiently estimate LST without requiring multi-band TIR data (Jiménez-Muñoz et al. 2009) [52]. Compared to other approaches, it is straightforward to implement and is particularly applicable for analyzing large urban areas.

In this study, Landsat-8 OLI/TIRS C2 L2 data were processed using the Google Earth Engine platform, selecting images acquired between June and September 2025 under favorable atmospheric conditions and cloud cover below 5% to capture representative summer urban thermal conditions. The LST was retrieved using the following equation:

$$\mathrm{L}\mathrm{S}\mathrm{T}={\displaystyle \frac{\mathrm{B}\mathrm{T}}{1+(\mathsf{\lambda }\cdot \mathrm{B}\mathrm{T}/\mathsf{\rho })\ln \mathsf{\epsilon }}}$$

where BT is the brightness temperature of the thermal infrared band (K), λ is the central wavelength of the thermal infrared band, $\mathsf{\rho }=\mathrm{h}\mathrm{c}/\mathsf{\sigma }\approx 1.438\times 10^{-2}\mathrm{m}·\mathrm{K}$, and ε represents the surface emissivity.

2.4.2. Selection and Analysis of Cooling Effect Indicators for Urban Parks

During model training, the dataset was randomly divided into 60% for training set, 20% for validation set, and 20% for testing set, respectively, to ensure robustness and predictive ability. Maximum tree depth and minimum split samples were constrained to capture nonlinear relationships while preserving generalization. Model performance was evaluated using the coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE), ensuring reliable and interpretable predictions. All operations were conducted in Python 2.7.

Referring to previous studies, a multi-ring buffer approach was employed to quantify the cooling effect of urban comprehensive and community parks. Firstly, concentric buffers were established for each park with an interval of 30 m and a total length of 600 m. The average LST differences (ΔT) among adjacent buffers were calculated. If a park exhibits the cooling effect within 30 × n meters, both ΔT_n and ΔT_n+1 − ΔT_n will be more than 0. The distance from the park boundary to the first inflection point was defined as the longest cooling distance (PD) (Figure 3). The difference between the park land-surface temperature (PLST) and the buffer surface temperature was defined as cooling intensity (PCI). Based on PD, PCI, and PLST, three additional cooling indicators were defined: cooling service (PCS), cooling gradient (PCG), and cooling efficiency (PCE), as detailed in Table 1. Finally, the above seven cooling effect indicators of 170 parks were quantified on ArcGIS platform.

Figure 3. Diagram of cooling effect of urban park. (a) Baiziwan Park and its buffers, and (b) variation curve of LST change in the park cooling process.

Furthermore, according to previous studies, 21 landscape features were initially selected, which probably be correlated with cooling effect indicators. To avoid potential model overfitting, we used Pearson correlation analysis to filter and remove irrelevant indicators. Eventually, 17 landscape features associated with cooling effect indicators were identified (Table 2), and these features were further classified into two types: internal park landscape features and external park landscape features.

2.4.3. XGBoost–SHAP Explainable Machine Learning Approach

Conventional machine learning methods can automatically identify complex nonlinear relationships and potential interactions from datasets, but they are often criticized as “black-box” models due to their limited interpretability. In this study, an explainable machine learning approach was constructed by combining XGBoost with the SHAP interpretation model. XGBoost (eXtreme Gradient Boosting) is a tree-based ensemble learning algorithm capable of handling high-dimensional data, nonlinear variables, and multivariate interactions, while maintaining fast training speed and strong resistance to overfitting [53]. Compared to traditional random forest (RF) and linear regression models, XGBoost demonstrated superior prediction accuracy, model stability, and remarkable flexibility [54].

During model training, the dataset was randomly divided into 60% for training set, 20% for validation set, and 20% for testing set, respectively, to ensure robustness and predictive ability. Maximum tree depth and minimum split samples were constrained to capture nonlinear relationships while preserving generalization. Model performance was evaluated using R², RMSE, and MAE, ensuring reliable and interpretable predictions. All operations were conducted in Python 2.7.

SHapley Additive exPlanations (SHAP) is a game-theory-based post hoc interpretation technique widely used to explain machine learning models, providing both global and local interpretability. SHAP values quantify the contribution of each feature to the prediction by computing weighted averages of all possible feature subsets. Larger absolute SHAP values indicate higher contribution to the target variable. In this study, global SHAP values (Beeswarm summary plots) were used to assess overall relative importance of all landscape features, providing consistent and reasonable XGBoost model explanations.

Additionally, Partial Dependence Plots were used to illustrate the marginal effects of landscape features on urban park cooling effect indicators. These plots showed the average impact of a single feature on the dependent variable while keeping all other features at their observed distribution. The marginal effect curves were derived from the feature importance outputs of the XGBoost–SHAP model, providing an intuitive visualization of the nonlinear responses of cooling effect indicators to different feature values. This analysis captured the contribution patterns of both internal park and buffer zone features, providing scientific guidance for urban park design and optimization.

2.4.4. Ordinary Least Squares Method

Ordinary Least Squares (OLS) is a classic linear regression method used to explore linear relationships between independent and dependent variables [20,55]. Compared with explainable machine learning models, OLS provides explicit coefficient interpretations, allowing researchers to understand the linear contribution and direction of landscape features on urban park cooling effects.

In this study, OLS complemented the nonlinear insights provided by XGBoost and SHAP. Using the linear regression method, potential threshold effects between internal park landscape features and seven cooling effects were identified [42,56], which can help understand the optimal scale, landscape composition, and shape of urban park.

2.4.5. Classification and Identification of Urban Parks Based on Cooling Effect Indicators

K-means clustering was applied to classify urban parks according to their cooling features. The seven cooling effect indicators were first standardized, and parks were grouped based on similarity in high or low value combinations. The optimal number of clusters (K) was determined using the Elbow method on Python 2.7 platform, reducing parameter search space while maintaining accuracy [57]. The results indicated that when K = 5, the sum of squared errors achieved the minimal value. Therefore, we determined that the number of the park categories should be five.

The 170 parks were initially divided into five clusters. However, there was one cluster containing only a single park (Jiangfu Park), indicating extremely limited sample representativeness and weak generalizability for comparative analysis. Therefore, this cluster was excluded from further classification analysis. The remaining four clusters all contained sufficient numbers of parks (29–65 parks per cluster) and exhibited relatively distinct cooling effect characteristics, indicating reasonable representativeness and classification effectiveness. Finally, four park clusters were obtained, and the cooling advantages and landscape features of each cluster were further analyzed.

3. Results

3.1. Cooling Effect of Urban Parks

The results indicated that 170 urban parks exhibited a cooling effect (Figure 4). The PLST values ranged from 35.85 °C to 43.86 °C, with an average of 39.64 °C, reflecting the high temperatures in Beijing during June to September. The PCA values ranged from 1.56 hm² to 144.31 hm², with an average cooling area of 16.66 hm². Jiangfu Park exhibited the largest PCA, whereas Ande Urban Forest Park possessed the smallest. The PCE values ranged from 0.04 to 11.67, with an average cooling efficiency of 2.68. Jianguomen Green Square demonstrated the highest cooling efficiency, while the Beijing Olympic Forest Park showed the lowest. The PCG reached a maximum cooling rate of approximately 0.07 °C/m at Beixiaohe Park and a minimal value at Beierhuan Urban Park, with an average of 0.02 °C/m. The PCI values range from 0.009 °C to 5.41 °C, with an average of 1.66 °C. Higher PCI values indicate a more pronounced cooling effect provided by the park. The PCS ranged from 0.04 °C to 468.91 °C, with an average of 40.27 °C. Jiangfu Park recorded the maximum PCS, whereas Beierhuan Urban Park showed the minimum. The PD values ranged from 30 m to 210 m, with an average of 85.24 m. The maximum and minimum PD values corresponded with those of PCS, suggesting that larger cooling distances are associated with greater overall cooling capacity.

Figure 4. Box plots of cooling indicators. (a) PLST and PCA; (b) PCS and PCI, and (c) PCE and PCG. (Note: “×” represents the mean value, the box length represents the interquartile range, and the horizontal line represents the median).

3.2. Impact of Influencing Factors on Cooling Effect of Urban Parks

3.2.1. Accuracy Validation of XGBoost Model

The accuracy performance of the XGBoost model was evaluated using R², RMSE, and MAE. As shown in Table 3, both RMSE and MAE values of the testing dataset were relatively low across all cooling effect indicators. Specifically, the errors of PCI, PCA, PCE, PCG, and PLST were all below 4, while those of PD and PCS remained below 20, indicating that the overall model errors fell within acceptable ranges. In addition, the R² of seven cooling effect indicators were 0.738, 0.814, 0.91, 0.818, 0.301, 0.391, and 0.834, respectively, with most values ranging from 0.7 to 0.85 (Figure 5). Overall, the XGBoost model showed good fitting performance, particularly for PCS, PCA and PD. These results demonstrated that the selected feature variables contributed substantially to the model performance, and the XGBoost model was capable of accurately capturing the variation in park cooling effect indicators. Therefore, the model met the accuracy requirements for subsequent analysis.

Table 3. Accuracy performance of XGBoost model.

Cooling Effect Indicators	RMSE	MAE	R2
PCI	0.538	0.410	0.738
PCA	4.399	1.977	0.814
PCS	10.715	7.132	0.910
PCE	0.651	0.471	0.818
PCG	0.011	0.007	0.301
PLST	1.153	0.903	0.391
PD	16.821	11.421	0.834

Figure 5. Regression models between real cooling effect indicators and those of simulated by XGBoost model. (a) PLST; (b) PCE; (c) PCA; (d) PCG; (e) PD; (f) PCS, and (g) PCI. (Note: The bar chart and curve represent the sample distribution).

3.2.2. Relative Importance Among Influencing Factors

The XGBoost model combined with SHAP was applied to evaluate the relative importance of landscape features for the seven cooling effect indicators. The left part of Figure 6 shows the mean absolute SHAP values of the 17 variables, as well as the top 3 and top 4–5 variables were identified as key landscape features and important landscape features, respectively. The right part, beeswarm summary plots, illustrates the distribution of feature contributions, where higher values on the right indicate positive effects, while those on the left indicate negative effects.

Figure 6. The relative importance of different landscape features on cooling effect indicators.

Overall, the results suggested that the relative importance of influencing factors varied across cooling effect indicators. Both park and buffer features showed notable contributions to the cooling effect. Specifically, the key landscape features included ratio of perimeter and area (RCA), buffer normalized difference vegetation index (B_NDVI), proportion of the impervious surface in the buffer (B_WTP), buffer impervious surface area (B_WTA), gross floor area (GFA), buffer green area (B_GA), park perimeter (CPARK), and park area (PA), while the important features comprised green area (GA), proportion of the green space in the buffer (B_GP), CPARK, impervious surface area (WTP), water area (WA), RCA, building area (BUA), GFA, PA, B_GA, and B_WTP.

In terms of key landscape features, B_WTA was identified as the most variable for PCI, PCA, PCS, PCE, and PD, with contribution rates of 27.2%, 48.0%, 41.8%, 28.0%, and 41.9%, respectively, showing consistent positive relationships with these indicators. This finding aligned with some previous studies [33,58], suggesting that larger extents of surrounding impervious surfaces tend to improve external LST, thereby increasing the thermal gradient between parks and their surroundings. Such gradients can enhance heat exchange processes and amplify cooling effect indicators, although they may also contribute to elevated internal park LST. For PCI, GFA and RCA also exhibited notable contributions (12.7% and 9.9%, respectively), with GFA showing a positive relationship and RCA a negative one. In the case of PCA, B_GA and GFA were identified as key variables. B_GA was positively associated with PCA, PCS, and PD, contributing 25.4%, 9.6%, and 13.2% to the overall explanation, respectively. This suggests a potential synergistic interaction between buffer green spaces and parks, enhancing the intensity, spatial extent, and service capacity of the cooling effect. Conversely, CPARK was negatively associated with PD, contributing approximately 13%. For PCS, GFA also demonstrated a relatively strong contribution (11.9%). The PCG indicator was mainly influenced by RCA, GFA, and B_GA, contributing 17.1%, 11.1%, and 10.3%, respectively. Among these, GFA showed a positive relationship, whereas RCA and B_GA exhibited negative relationships. Additionally, PA and impervious surface area (WTA) were negatively correlated with PCE, contributing 19.5% and 10.0%, respectively. This implied that while larger parks may provide broader and stronger cooling capacity, their cooling efficiency per unit area may decline [56,59]. For PLST, RCA, B_NDVI, and B_WTP were identified as dominant variables, contributing 17.5%, 17.3%, and 16.4%, respectively. Notably, WA made important contributions to both PCG and PCS (8.9% and 6.2%, respectively), with positive effects. This can be attributed to the relatively high heat capacity and thermal inertia of water bodies, along with their lower emissivity [60,61]. In addition, GA was an important contributor to PLST and PCE (8.7% and 9.0%, respectively), exhibiting negative relationships with both indicators.

3.2.3. Nonlinear Relationships Between Dominant Influencing Factors and Park Cooling Effect Indicators

As shown in Partial Dependence Plots (Figure 7), significant nonlinear relationships existed between different influencing factors and park cooling effect indicators. Some factors exhibited a type of nonlinear positive correlation. Specifically, as factor values increase, the corresponding cooling effect indicators also increase nonlinearly. However, this growth trend gradually slows once specific thresholds are reached. For instance, when B_WTA exceeds approximately 0–25 hm² (Figure 7d), 22–55 hm² (Figure 7j), 10–42 hm² (Figure 7g), 0–23 hm² (Figure 7p), and 0–30 hm² (Figure 7m), respectively, the variations in PCI, PCA, PCS, PCE, and PD become less pronounced. Similarly, when GFA exceeds approximately 0–11 hm² (Figure 7e), 12–25 hm² (Figure 7h), and 0–12 hm² (Figure 7t), respectively, the increases in PCI, PCS, and PCG are no longer significant. Similarly, B_WTP, B_GA, and RCA also showed threshold effects, with critical values around 60–88% (Figure 7c), 0–10 hm² (Figure 7n), and 0.017–0.04 m⁻¹ (Figure 7a), respectively. Above results suggested that within these specific ranges, changes in influencing factors may induce substantial variations in cooling effect indicators. Additionally, B_GA also showed a positive correlation with PCA and PCS, but the change trends were different with these mentioned above. Initially, PCA and PCS increase relatively smoothly with the growth of B_GA (Figure 7i,k). When B_GA reaches approximately 20–33 hm², there is a rapid increase in the indicators, followed by a smooth change trend again.

Figure 7. Partial dependence relationships of key landscape features in urban parks.

Conversely, there were some factors exhibiting nonlinear negative correlations with cooling effect indicators. For example, PCE is negatively associated with PA and WTA, PLST with B_NDVI, PD with CPARK, and PCG with RCA and B_GA. These cooling effect indicators generally decline rapidly as the influencing factors increase, but the changes become negligible once PA, WTA, B_NDVI, B_GA, CPARK, and RCA reach approximately 0–22 hm² (Figure 7q), 0–8 hm² (Figure 7r), 0.27–0.34 (Figure 7b), 0–11 hm² (Figure 7u), 300–2200 m (Figure 7o), and 0.01–0.025 m⁻¹ (Figure 7s), respectively. Overall, these results highlighted clear nonlinear relationships and threshold effects between park landscape features and their cooling performance, which are crucial for optimizing park design and improving cooling efficiency.

3.3. Threshold Effects Between Urban Parks and Internal Landscape Features

Previous studies indicated that internal park features, such as PA, CPARK, and RCA, showed notable relevance to cooling effect indicators [62,63]. To better support park design and optimize park cooling performances, this study selected five internal park features, including PA, CPARK, RCA, GA, and WA, and employed the OLS model to examine their threshold effects on cooling effect indicators.

In terms of size variables, PA showed evident threshold effects with PCI, PCE, PCG, PLST, and PCS, except for PCA. PA is positively correlated with PCI, PCG, and PCS, indicating that the increases in park size are linked to stronger cooling intensity, faster cooling gradients, and enhanced cooling service. When PA increases from 0 to 20–25 hm², PCI, PCG, and PCS rose rapidly, but further increases in PA resulted in limited changes (Figure 8a,b,d). Conversely, PA was negatively correlated with PCE and PLST, suggesting that larger parks may reduce their internal LST and cooling efficiency [64]. When PA increases within the range of 0–20 hm², both PCE and PLST decline markedly, while changes become less pronounced beyond this range (Figure 8c,e). These results suggested potential trade-offs among different cooling effect indicators, particularly in dense urban center regions where they are facing the huge challenge of balancing cooling performance and land-use efficiency. CPARK showed a clear threshold effect only for PLST, with a negative correlation (Figure 8f). when CPARK exceeded about 1500 m, the changes in PLST become limited.

Figure 8. The linear relationship between the internal landscape features of parks and cooling indicators. (Note: *** indicates statistical significance at the p < 0.001 level).

For shape variables, the results showed that RCA is negatively correlated with PCI, PCA, PCG, and PCS, which indicated the more complex park shape is generally associated with weaker cooling intensity, area, gradient, and service. The specific threshold ranges varied across indicators; when RCA increased within the respective ranges of 0–0.05 m⁻¹ (Figure 8g), 0–0.015 m⁻¹ (Figure 8l), 0–0.03 m⁻¹ (Figure 8j), and 0–0.015 m⁻¹ (Figure 8h), PCI, PCA, PCG, and PCS decreased rapidly. In contrast, RCA is positively correlated with PLST and PCE (Figure 8), suggesting that more irregularly shaped parks tend to possess higher LST and cooling efficiency, with threshold values of approximately 0–0.017 m⁻¹ (Figure 8i) and 0–0.04 m⁻¹ (Figure 8k), respectively.

Regarding landscape composition variables, GA exhibited threshold effects similar to PA, showing positive correlations with PCI, PCG, and PCS, and negative correlations with PCE and PLST (Figure 8m–q), with corresponding thresholds of approximately 7 hm², 9 hm², 7 hm², 5 hm², and 8 hm². WA only showed a significant threshold effect on PCI (Figure 8r), rapidly increasing as WA grew from 0 to 2.5 hm², after which its effect plateaued.

3.4. Classification and Identification of Parks Based on Cooling Effect Indicators

Clustering analysis was conducted on urban parks using five major cooling effect indicators (Figure 9b). Based on the distinctive performance of each cluster across the indicators, parks were classified into four types: high cooling intensity parks (Cluster 1), high cooling gradient parks (Cluster 2), high cooling efficiency parks (Cluster 3), and internal comfort-oriented parks (Cluster 4). Among them, Clusters 1 and 2 belonged to the large-scale category (more than 10 hm²); however, the parks in Clusters 3 and 4 were relatively smaller, with sizes approaching those of medium-sized parks (1 hm²–10 hm²). Normalized values of the five cooling effect indicators and landscape features for the four clusters were displayed using box plots (Figure 9b,c), while Figure 9a showed their spatial distribution along with representative park examples.

Figure 9. Classification of parks based on cooling effect indicators. (a) The spatial distribution and park examples of the identified cooling clusters, (b) box plots of urban park’s cooling effect indicators for five cooling clusters, and (c) normalized average landscape features.

High cooling intensity parks (Cluster 1) consisted of 33 parks (average 11.57 hm²), primarily dominated by cooling intensity (PCI) (Figure 9b), with Fengtai Park serving as a representative example (Figure 9a). Their landscape features included the largest buffer green area, small ratio of perimeter and area, regular shape, and relatively large internal water and green spaces (Figure 9c). Regarding cooling effects, these parks exhibited the highest cooling capacity, service, and coverage, with other indicators also relatively high, providing significant and sustained cooling benefits, particularly during high temperature periods, substantially improving the surrounding thermal environment.

High cooling gradient parks (Cluster 2) comprised 29 parks (average 51.63 hm²), with Haidian Park as an example (Figure 9a), dominated by cooling rate (PCG) (Figure 9b). These parks had the lowest average temperature, largest area, highest vegetation cover and water body area, smallest ratio of perimeter and area, and most regular shape (Figure 9c). Their cooling effects were characterized by the highest cooling rate and strong cooling capacity, rapidly reducing local temperatures to create comfortable thermal conditions. However, due to their large size, cooling efficiency was the lowest.

High cooling efficiency parks (Cluster 3) include 42 parks, with Qianmen Park as an example (Figure 9a), dominated by cooling efficiency (PCE) (Figure 9b). These parks were the smallest in size (average 5.2 hm²) with complex boundary shapes and limited internal vegetation and water bodies, but they did maintain some green space in the buffer (Figure 9c). They exhibited the highest cooling efficiency, quickly lowering local temperatures over small areas, offering high cost-effectiveness.

Internal comfort-oriented parks (Cluster 4) were the most numerous, with 65 parks, mainly linear or small community parks (Figure 9b), distributed around the central districts (Figure 9a). Most cooling effect indicators were relatively low, except for cooling efficiency, but surface temperatures were relatively low, land use was efficient, and distribution was widespread. Landscape characteristics included the largest ratio of perimeter and area, most complex boundary shapes, slightly larger park area (average 5.69 hm²), and internal water and vegetation slightly larger than Cluster 3, while buffer green area was minimal (Figure 9c). Although the cooling effect on surrounding areas was limited, these parks create internally comfortable microclimates, were land-efficient, and provided recreational functions, making them suitable for areas with lower cooling demand.

4. Discussion

4.1. Importance, Nonlinear Relationships, and Threshold Effects of Cooling Effects in Community and Comprehensive Parks

This study explored the cooling effects of 170 urban parks in the urban center region of Beijing, focusing on community and comprehensive parks at the medium (area ranges from 1 hm² to 10 hm²) and large (area larger than 10 hm²) scale. This study achieved the three research aims proposed in the Introduction by identifying the important and key landscape features associated with park cooling effects, revealing the nonlinear relationships and threshold effects between landscape features and cooling indicators, and classifying urban parks according to their cooling performance and functional characteristics. By integrating a multi-dimensional indicator system with an XGBoost–SHAP explainable machine learning method, the relative importance and marginal effects (nonlinear relationships) of multiple landscape features on park cooling effects were quantified, while OLS regression model was applied to identify threshold effects of key internal park landscape features. Finally, parks were classified using the Elbow and K-means clustering methods based on their cooling performance, enabling a comparative analysis of their functional advantages and implications for urban planning.

This study demonstrated the effectiveness of the explainable machine learning method for park cooling effect analysis. The XGBoost model exhibited good predictive performance for cooling effect indicators, with relatively high R² and low RMSE and MAE, indicating that the model was capable of accurately capturing variations in park cooling effects. The combination of XGBoost and SHAP methods in this study improved the interpretability of the model by effectively revealing the relative contributions of individual landscape features and further helped identify their marginal responses and nonlinear trends, thereby partially overcoming the “black-box” limitation commonly associated with conventional machine learning approaches. It also provided important advantages for identifying the hidden effects of individual features on cooling performance that are often difficult to detect using conventional linear models. In addition, many previous studies on park cooling effects did not explicitly consider park scale. Studies that distinguished scale remained limited, primarily focusing on pocket parks [38,39] or conducting comparisons cross different scale [59]. Therefore, this study provided a targeted analysis of comprehensive and community parks at the medium and large scale, contributing to a specific understanding of park cooling performance and offering insights with greater relevance for urban planning.

The relative importance analysis indicated that, for several cooling effect indicators such as cooling intensity, area, service, efficiency, and distance, external landscape features exhibited higher importance than internal features. This finding was consistent with Zhang et al. (2026) [59], whose cross-scale analysis showed that the relative importance of landscape variables varied with park scale. Internal landscape features tend to dominate in parks at the small and medium scale, whereas external landscape features become more influential in larger scale parks. However, some results of previous studies emphasized the stronger role of internal park variables, such as park area, green space area, and water bodies area. We believed that the differences may be related to the scale of selected parks and the essence of cooling effect indicators. For small parks, park cooling performance is relatively limited and tends to increase with area, making the internal features appear more influential. In contrast, as park scale increases, cooling performance does not grow linearly but instead approaches a plateau, reducing the relative importance of size-related variables. For larger parks, the cooling effect is more closely associated with the surrounding urban environment. In particular, cooling intensity reflects the temperature difference between parks and adjacent urban surfaces unaffected by park. When surrounding impervious surfaces expand, higher background temperatures may amplify this difference, resulting in stronger associations between cooling effect indicators and external features, such as impervious surface area of buffer, proportion of the blue space in the buffer, buffer normalized difference vegetation index, and buffer gross floor area. In addition, it should be noted that the higher importance of external park features does not imply a negligible role of internal features. In this study, park area, perimeter, and ratio of perimeter and area ranked among the top three importance ranking for several indicators, while park green space area and water bodies area were frequently within the top five, identified as key and important landscape features, respectively. These features remained closely associated with park cooling effect indicators and contributed substantially to thermal environment regulation.

The results indicated that both internal and external landscape features of urban parks exhibited pronounced nonlinear relationships and threshold effects with cooling performance. From the perspective of internal park features, park area showed significant positive relationships with cooling intensity, service, and gradient, while being negatively correlated with cooling efficiency and park temperature, indicating a clear trade-off between total cooling capacity and efficiency. A distinct threshold effect was identified, with the park area threshold in Beijing being approximately 25 hm². However, this threshold varies across cities, such as approximately 60 hm² in Nanjing [65] and 56 hm² in Fuzhou [22], suggesting strong spatial heterogeneity influenced by climatic and urban morphological conditions. The influence of ratio of perimeter and area remained debated in the literature. While some studies reported that higher ratio enhanced cooling effects [66], others suggested the opposite [24]. In this study, the ratio of perimeter and area was positively correlated with park temperature and cooling efficiency, but negatively associated with cooling intensity, service, gradient, and area. These results indicated that increasing ratio of perimeter and area raises internal surface temperature while improving cooling efficiency but weakened overall cooling intensity and spatial extent [65]. A possible explanation was that higher ratio of perimeter and area leads to more complex boundaries and increases edge interfaces, which intensifies heat exchange with the surrounding environment and reduced the stability of internal cooling effects. Therefore, to achieve stronger cooling capacity and lower internal surface temperatures, parks should adopt more regular shapes, with an optimal ratio of perimeter and area of approximately 0.05 m⁻¹. Moreover, water bodies area showed a strong positive relationship with cooling intensity, while green space area was positively associated with cooling intensity, gradient, and service and negatively correlated with other indicators. Notably, the water bodies area had a stronger cooling effect than the green space area. Our study showed that under the same area, water bodies can reduce temperature by approximately 3.5 °C, compared to about 1.7 °C for vegetation. This difference can be attributed to the higher heat capacity and evaporation rate of water, as well as its lower surface roughness and radiative properties, which facilitated more efficient heat dissipation [67]. In addition, reducing the park impervious surface area can enhance cooling gradient, consistent with previous findings [22]. From the perspective of external landscape features, previous studies suggested that cooling distance was generally around 300 m, with the most effective cooling occurring within 200 m [17]. In this study, the cooling distance was approximately 210 m and parks may interact synergistically with other parks and environmental factors within this range. The results further showed that as impervious surface area of buffer increases, several key cooling effect indicators also increased. This may be because larger impervious surface area of buffer was often associated with larger cooling area, which was jointly determined by cooling distance and park area. Parks with greater cooling distances and larger sizes tend to exhibit stronger overall cooling effects. Meanwhile, increasing the buffer normalized difference vegetation index and reducing proportion of the impervious surface in the buffer can significantly reduce park temperature, likely due to reduced temperature difference between parks and their surroundings, which weakened heat exchange. Interestingly, gross floor area was positively associated with cooling intensity, services and area, which contrasted with previous studies [18]. This discrepancy may be attributed to the high-density urban context of the study area. Larger parks tended to have stronger cooling effects and broader influence ranges, thereby encompassing more built-up areas within their cooling extent, leading to a positive relationship between gross floor area and cooling effect indicators.

In recent years, urban greening strategies such as “green space insertion,” ecological corridors, ventilation corridors, and vertical greening have been widely promoted, particularly in Beijing, to improve land-use efficiency and alleviate urban thermal stress. Based on the findings of this study, park planning and green infrastructure design in megacity core areas should place greater emphasis on optimizing both internal park composition and surrounding environmental conditions to improve thermal comfort under limited land resources. Specifically, increasing the proportion of water bodies within existing parks may provide a practical and space-efficient strategy for enhancing cooling intensity and reducing surface temperatures, particularly in densely built urban environments where opportunities for large-scale park expansion are constrained. In addition, parks with relatively regular shapes and appropriate spatial scales may contribute to more stable and effective thermal regulation. From the perspective of the surrounding environments, increasing green space area and vegetation coverage within park buffer zones may help strengthen the synergistic cooling interactions among parks and other urban green spaces, thereby enhancing cooling distance, cooling coverage, and the overall thermal regulation performance of surrounding urban areas.

4.2. Trade-Offs Among Cooling Effect Indicators and Implications for Park Planning

The results of this study revealed a clear trade-off between cooling intensity and cooling efficiency: while larger parks generally exhibit stronger cooling capacity, their cooling efficiency tends to decrease. Under constraints in urban land availability and fiscal resources, simply expanding park size is no longer a feasible strategy in megacity core areas. Instead, urban park planning should emphasize thermal comfort improvement, land-use efficiency, and the differentiated cooling functions of various park types to support context-specific thermal regulation strategies.

For areas with high cooling demand, such as industrial zones and densely built commercial districts, high-cooling-capacity parks (Cluster 1) are most appropriate. These areas are typically characterized by high surface temperatures and significant heat accumulation, requiring strong thermal regulation. These parks exhibited strong cooling capacity, large cooling coverage, and relatively stable cooling performance due to their larger sizes, regular shapes, and abundant internal vegetation and water bodies. Therefore, they can function as major cooling sources for mitigating regional heat accumulation and improving surrounding thermal environments.

In areas where land resources are relatively abundant and rapid cooling is required, high-cooling-rate parks (Cluster 2) are recommended, which means establishing large-scale parks (area larger than 10 hm²). Their large sizes, high vegetation and water coverage, and regular shapes contribute to faster cooling rates, stronger cooling intensity, and lower land-surface temperatures. Although their cooling efficiency is relatively low, these parks can provide significant thermal regulation service and recreational benefits in areas with relatively sufficient land resources [68].

In contrast, high-cooling-efficiency parks (Cluster 3) may be more applicable in high-density urban areas with limited land resources, such as transportation hubs and city centers. Although relatively small in size, these parks exhibited high cooling efficiency and rapid local cooling capacity, indicating that optimizing cooling benefits per unit area may represent a more feasible strategy than large-scale park expansion in highly urbanized environments [69]. Moreover, their distributed spatial layout can improve residents’ accessibility to cooling services within walking distances [18].

Internal comfort-oriented parks (Cluster 4) are more suitable for residential communities, schools, and daily activity spaces where human-scale thermal comfort is prioritized. Although their external cooling effects were relatively limited, these parks can effectively improve internal microclimatic comfort and livability through shading and localized thermal regulation. Their widespread distribution also enables synergistic interactions with street greenery, pocket parks, and ecological corridors, thereby contributing to neighborhood-scale thermal adaptation.

Overall, the results suggested that urban park systems in megacities should not rely on a single cooling-oriented strategy. Instead, different park types should be integrated according to local thermal regulation demands, urban density, land availability, and population activity characteristics to improve urban heat resilience and outdoor thermal comfort in highly urbanized environments.

4.3. Limitations and Future Research Directions

This study has several limitations that should be acknowledged. First, although the XGBoost–SHAP explainable machine learning method can effectively quantify the relative contributions and nonlinear relationships of individual variables, the interpretation results mainly reflect statistical associations derived from the model rather than strict causal mechanisms. Therefore, the identified feature importance and threshold values should be cautiously interpreted with domain knowledge. Second, the accuracy of land-use and land-surface temperature data used in this study is subject to certain limitations, partly due to the availability of high-quality remote sensing imagery. These data constraints may introduce uncertainties in quantifying park cooling effects and thus represent an inherent challenge in urban climate studies. Third, although this study examines the effects of multiple landscape factors on cooling performance, it does not explicitly address the interactions and synergistic effects among these variables. In reality, the cooling process is governed by complex, coupled mechanisms, and neglecting such interactions may limit the explanatory power of the results. In addition, urban parks are fundamentally designed to serve human needs; however, this study does not incorporate factors related to population distribution, human activity patterns, or accessibility to green spaces. These aspects are crucial for evaluating the actual effectiveness of parks in improving thermal comfort at the human scale. Meanwhile, parks, as a cluster, are inherently complex, and this study only explores their functional role in climate regulation, without considering the cultural, symbolic, and social functions that parks may also possess, which represents an aspect that could be improved in future research.

Future research should therefore adopt a more integrative framework by incorporating multi-dimensional indicators, including environmental, cultural, spatial, and socio-economic variables. Such an approach would enable a more comprehensive and accurate assessment of urban park cooling effects and provide stronger support for climate-responsive urban planning.

5. Conclusions

This study quantified the cooling effects of 170 urban parks using seven indicators and identified 17 potential landscape features from both internal and surrounding environments. By integrating the XGBoost model with SHAP approach, the relative importance and nonlinear relationships between feature variables and cooling effect indicators were systematically examined. In addition, an OLS model was employed to explore threshold effects of five key internal landscape features. Furthermore, parks were classified using the Elbow method and K-means clustering based on their cooling performance, enabling a comparative analysis of their functional advantages and implications for urban planning. The main conclusions are as follows:

(1)

Both internal and external landscape features were closely associated with the park cooling effects. Among them, RCA, B_NDVI, B_WTP, B_WTA, GFA, B_GA, CPARK, and PA were identified as key features, ranking among the top three in relative importance, while GA, B_GP, CPARK, WTP, WA, RCA, BUA, GFA, PA, B_GA, and B_WTP ranked among the top four to five. Specifically, B_WTA, GFA, B_WTP, and WA showed positive correlations with most cooling effect indicators. B_GA was positively associated with PCA, PCS, and PD, but negatively correlated with PCG. RCA exhibited significant negative relationships with PCI and PCG, while being positively associated with PLST. In contrast, CPARK, WTA, B_NDVI, and GA were generally negatively correlated with cooling effect indicators.

(2)

Pronounced nonlinear relationships and threshold effects were observed between landscape features and park cooling effect indicators. From the perspective of the thermal environment regulation, the results indicated that when PA reaches approximately 25 hm², cooling intensity, cooling service, and cooling gradient were maximized, whereas cooling efficiency was minimized. More regular park shapes were associated with lower land-surface temperatures and higher cooling rates; therefore, overly complex geometries should be avoided, with an optimal RCA approximately 0.05 m⁻¹. In addition, water bodies provided stronger cooling effects than vegetation, suggesting that increasing WA—up to an optimal size of around 2.5 hm²—can enhance cooling performance. Moreover, increasing CPARK reduced the effective cooling distance, thereby constraining the spatial extent of cooling, with a threshold of approximately 3500 m. With increasing B_WTA, PCI, PCA, PCS, PCE, and PD all showed upward trends, with threshold values of approximately, 25 hm², 75 hm², 78 hm², 40 hm², and 50 hm², respectively.

(3)

Four typical park types were classified through clustering analysis: high cooling intensity type, high cooling rate type, high cooling efficiency type, and internal comfort-oriented type. These types differed substantially in both landscape configuration and cooling performance. Therefore, urban park planning and optimization should adopt differentiated design strategies tailored to the specific cooling demands of each site, thereby enhancing cost-effectiveness and thermal regulation benefits to a certain extent.

Overall, this study provides robust theoretical support for urban park planning and design, particularly in enhancing cooling effects. The findings also offer valuable insights for mitigating the urban heat island effect and addressing increasing urban heat stress. Future research should further investigate seasonal variations in park cooling effects, regional heterogeneity, as well as issues related to park accessibility and environmental equity. In addition, it should explore the interactions and synergistic effects between the variables studied and the human factors involved in the processes under investigation, in order to provide practical guidance for urban policy and strategic urban planning.