Multidimensional characteristics of urban green space and its impact in mitigating urban heat Island effects: a case study of Guangzhou

Abstract

Urban Heat Island (UHI) effect is a significant concern in cities, increasing environmental and living conditions challenges. Urban green spaces are essential in mitigating UHI effects by regulating temperature. However, most studies do not adequately consider the multidimensional characteristics of green spaces in their cooling effects on the UHI. To address this gap, this study systematically investigates the cooling effects of green space by examining global and local mitigation mechanisms, with Guangzhou as a case study. Land surface temperature (LST) data were retrieved using the Mono-Window algorithm to analyze spatial variations in UHI intensity. Green space was further evaluated in terms of distribution, quality, and morphology, providing a robust evaluation. A Random Forest (RF) regression model, combined with SHapley Additive exPlanations (SHAP), was used to quantify the global contributions of UGS indicators to cooling effects. A Multiscale Geographically Weighted Regression (MGWR) model was employed to examine spatial heterogeneity. Results show an increase in high-temperature zones in Guangzhou from 2019 to 2024, particularly in Panyu and Nansha, indicating intensified UHI effects. Northern forested areas exhibited strong cooling effects, with Cooling Effect Index (CEI) values above 0.45. The distance to green space (DGS) was found to be the primary factor influencing cooling, accounting for 75.3% of the variance in CEI in 2024. The MGWR model confirmed that DGS has a negative correlation with LST, indicating that closer proximity to green spaces is associated with lower temperatures. This study provides evidence for optimizing green space planning and thermal environment management in Guangzhou and other megacities.

Introduction

In the context of global climate change, urban heat environments have become increasingly severe, particularly in megacities of tropical and subtropical regions^1,2. The combination of intense solar radiation, high temperatures, high humidity, and urban impervious surfaces exacerbates localized warming effects, representing a significant manifestation of urban ecological degradation³. As cities continue to expand, changes in land cover, such as the reduction of vegetation and the increase in artificial surfaces, disrupt the surface energy balance, raising heat capacity and thermal inertia⁴. Consequently, this amplifies the intensity of the Urban Heat Island (UHI) effect. The UHI effect not only drives up urban energy consumption and worsens air pollution, but also poses ecological risks, such as altered phenology, which further disrupts urban ecosystem functions^5,6. More critically, the combined effects of UHI and extreme heat events significantly elevate local heat environment risks, thereby undermining urban ecosystem stability^7,8. Consequently, mitigating UHI and implementing sustainable urban cooling strategies have become central to improving urban living conditions.

In recent years, engineering cooling technologies such as cool roofs, high-reflectivity coatings, and cool pavements have been widely applied to alleviate UHI⁹. While these methods have proven effective in localized areas, their limited applicability and high construction and maintenance costs present significant challenges for long-term sustainability¹⁰. In contrast, Urban Green Spaces (UGSs) provide more durable and comprehensive benefits for temperature regulation, ecosystem services, and social well-being^7,11,12. Studies have demonstrated that various types of UGS, including forests, grasslands, green roofs, and vertical green walls, effectively mitigate UHI effects, with localized temperature reductions ranging from 2 °C to 17°C¹³. However, most existing studies primarily focus on the cooling effects of individual UGS types or investigate these effects at small scales, often not fully integrating multidimensional characteristics, such as spatial patterns, quality, and structural features, in regulating the UHI effect⁷. This highlights the importance of better understanding how the spatial complexity and configuration of UGS influence urban thermal environments.

An important aspect in assessing urban heat environments is obtaining high-precision Land Surface Temperature (LST) data, which is essential for effective evaluation¹⁴. Compared to traditional ground-based monitoring, LST retrieval from remote sensing imagery allows for large-scale temperature data collection at a single point in time, becoming the mainstream method for monitoring urban thermal conditions¹⁵. The first application of satellite observations to study the UHI effect dates back to 1972 by Rao (1972), marking the beginning of remote sensing’s role in urban thermal research¹⁶. While early studies used Radiative Transfer Equations (RTE) for LST retrieval, their reliance on atmospheric profiles and aerosol data posed challenges due to data uncertainty at local scales¹⁷. To address these limitations, methods like the Split-Window Algorithm (SWA)¹⁸ and Mono-Window algorithm (MWA)¹⁹ were developed. MWA, in particular, has proven more stable, especially in the presence of fluctuating image quality or incomplete meteorological data, making it more suitable for UHI studies at urban scales.

In addition to LST data acquisition, evaluating the regulatory mechanisms of UGS characteristics on urban thermal environments has become increasingly crucial²⁰. Early research into UGS’s role in mitigating UHI primarily focused on green space area, quantity, and the use of Normalized Difference Vegetation Index (NDVI) to assess quality, correlating these with LST^21,22. Some studies identified the diffusion effect of UGS, exploring the distance over which cooling effects extend^23,24. However, as research evolved, it became evident that the cooling effects of UGS are not only influenced by their scale but are also tightly linked to spatial configuration²⁵. Highly fragmented UGS, lacking connectivity, may reduce overall cooling effects^21,26. As a result, more recent studies have incorporated landscape pattern analysis to assess spatial complexity, connectivity, and aggregation, revealing how green space morphology influences UHI mitigation^22,27,28. Despite this progress, gaps remain in the integration of multidimensional UGS characteristics. Therefore, there is a pressing need to develop a more systematic framework that integrates current research methods and lays the foundation for green space optimization strategies.

A common approach to studying the relationship between UGS and LST has been to use Pearson and Spearman correlation analyses, alongside geographic detectors to assess spatial heterogeneity^29,30. These methods, while valuable, mainly address linear or monotonic relationships and fail to capture complex nonlinear interactions. To address this, nonlinear machine learning models, such as Gradient Boosting Decision Tree (GBDT)^31,32, Extreme Gradient Boosting (XGBoost)³³, Light Gradient Boosting Machine (LightGBM)³⁴ and Random Forests (RF)³⁵, have been increasingly applied to explore the drivers of UGS cooling capacity. RF as an ensemble learning method, provides high stability and fault tolerance when handling high-dimensional variables, particularly in the presence of missing or noisy data. They also provide interpretability through SHapley Additive exPlanations (SHAP), providing insights into the mechanisms of driving cooling effects^36,37. Despite these advances, localized correlation analysis remains insufficient. In response, Geographically Weighted Regression (GWR) and its multiscale version (MGWR) have been widely applied to characterize the spatial heterogeneity of UGS and temperature relationships across regions^38,39. Combining global nonlinear variable analysis with localized spatial modeling provides a promising approach to fully uncover the impact mechanisms of UGS characteristics on urban thermal environments and provides deeper insights into local ecological regulation mechanisms.

This study focuses on Guangzhou as a case study and develops a comprehensive analytical framework that integrates LST retrieval, green space spatial characterization, and multiscale spatial statistical modeling. The objective is to systematically uncover the mechanisms through which UGS structure and ecological functions influence the mitigation of UHI. Specific goals include: (1) assessing the spatial pattern evolution of the urban thermal environment before and after the implementation of current green space system planning using the MWA; (2) systematically characterizing the multidimensional spatial features of urban green spaces by integrating spatial distribution, ecological quality, and structural characteristics; (3) analyzing the mechanisms and spatial heterogeneity of different green space features in influencing local cooling effects through a combination of random forests and MGWR models. This research aims to clarify the coupling relationship between green space spatial patterns and urban thermal environments at a mechanistic level, thus advancing the interdisciplinary field of landscape ecology and climate regulation. Ultimately, the study will provide a scientific basis for optimizing UGS patterns in mitigating UHI.

Study area and data source

Study area

Guangzhou, located in the central part of the Pearl River Delta (N 22°26′–23°56′, E 112°57′–114°03′), covers an area of approximately 7,434.4 km² (Fig. 1). The city experiences a subtropical monsoon climate, characterized by long, hot, and humid summers, typically spanning from March to November. Rapid urbanization has led to a pronounced UHI effect, primarily driven by the expansion of impervious surfaces and high-density urban development. In response, the Guangzhou Green Space System Plan (2021–2035), hereafter referred to as the Green Space Plan, aims to increase the greening coverage rate in built-up areas by approximately 2% compared to 2018 levels by 2025, with the objective of enhancing the regulatory function of UGSs⁴⁰.

Fig. 1

Location of the study area.

Data source and processing

The administrative boundary data used in this study was obtained from the National Platform for Common Geospatial Information Services. Land-use data were derived from the Dynamic World dataset, which is generated from Sentinel-2 imagery using a deep learning model to produce near real-time land cover classifications with a spatial resolution of 10 m. This dataset offers high temporal resolution and continuous updates, categorizing land surfaces into nine types: trees, shrubland, grassland, cropland, built-up areas, bare ground, wetlands, water, and snow/ice⁴¹. Surface temperature data were obtained from Landsat 8 imagery acquired in 2019 and 2024. The 2019 scenes represent the baseline before the implementation of the Green Space Plan. In contrast, the 2024 scenes correspond to the first full execution year, following an initial preparatory phase and subsequent policy interventions, and also serve as the final observation point before the Plan’s 2025 near-term targets are reached. To ensure data quality, images with cloud cover below 20% were prioritized, and cloud masking procedures, as described in the methodology section, were applied. Previous studies have demonstrated that in subtropical cities such as Guangzhou, seasonal variations exert only a limited influence on long-term UHI trends, and remotely sensed LST can provide a reliable proxy for annual urban thermal patterns^42,43. Therefore, to minimize data gaps and atmospheric interference caused by frequent cloud cover and precipitation during the hottest months, four high-quality Landsat scenes were selected: October 29, 2019, November 14, 2019, October 10, 2024, and November 27, 2024. All spatial datasets were preprocessed using ArcGIS Pro 3.0, including projection transformation, geometric correction, spatial registration, and clipping to the study area boundary to ensure spatial consistency. Prior to analysis, all datasets were standardized to the WGS_1984_UTM_Zone_49N coordinate reference system. Detailed information on the data sources is provided in Table 1.

Table 1 Data sources.

Methods

This study systematically investigates the cooling effects of UGS from the perspectives of spatial distribution, ecological quality, and landscape morphology by integrating both global and local spatial statistical methods (Fig. 2). First, LST was retrieved from Landsat 8 imagery and standardized using a deviation-based approach to quantify the spatial distribution of UHI intensity, as well as to analyze the spatial variation in the urban thermal environment. Second, a classification system for UGS was developed to characterize the spatial patterns and ecological quality of UGSs within the urban area. Finally, the RF regression model, combined with SHAP, was employed to quantify the global contributions and relative importance of various UGS indicators in mitigating urban heat. Additionally, MGWR model was used to capture the spatial heterogeneity of UGS effects on urban cooling, revealing localized relationships between UGS characteristics and the urban thermal environment across different regions.

Fig. 2

Framework of this study.

Urban thermal environment assessment

Land surface temperature analysis

To assess the urban thermal environment, the LST was retrieved from thermal infrared data obtained from Landsat 8, using the Mono-Window algorithm proposed by Qin et al. (2001)¹⁹. This algorithm incorporates parameters such as brightness temperature, mean atmospheric temperature, atmospheric transmittance, and surface emissivity to estimate LST with improved accuracy. The LST is calculated as follows:

$$\:{T}_{\text{s}}=\left[\alpha\:\left(1-C-D\right)+[\beta\:\left(1-C-D\right)+C+D]{T}_{\text{s}\text{e}}-{DT}_{\text{a}}\right]/C$$

(1)

where $\:{T}_{\text{s}}$ is the land surface temperature, $\:{T}_{\text{s}\text{e}}$ is the sensor brightness temperature, and $\:{T}_{\text{a}}$ is the mean atmospheric temperature. The coefficients $\:\alpha\:$ and $\:\beta\:$ are empirically defined as − 67.355351 and 0.458606, respectively⁴⁴. The intermediate variables $\:C$ and $\:D$ are calculated as:

$$\:C=\epsilon\:\tau\:$$

(2)

$$\:D=\left(1-\tau\:\right)\left[1+\left(1-\epsilon\:\right)\times\:\tau\:\right]$$

(3)

where $\:\epsilon\:$ is surface emissive, and $\:\tau\:$ is the atmospheric transmittance. $\:{T}_{\text{s}\text{e}}$ is derived using Planck’s radiation law:

$$\:{T}_{se}={K}_{2}/ln\left(1+\frac{{K}_{1}}{{L}_{\lambda\:}}\right)$$

(4)

where $\:{L}_{\lambda\:}$ is the spectral radiance obtained through radiometric calibration. For Landsat 8, the calibration constants are $\:{K}_{1}$ = 774.89 and $\:{K}_{2}$ = 1321.08⁴⁴. $\:{T}_{\text{a}}$ is estimated using a regionally adapted parametric model for subtropical zones:

$$\:{T}_{\text{a}}=17.9769+0.91715{T}_{\text{l}\text{a}}$$

(5)

where $\:{T}_{\text{l}\text{a}}$ is the near-surface air temperature on the image acquisition date.

To support LST retrieval, several biophysical parameters were derived from optical bands, including the Normalized Difference Vegetation Index (NDVI), Fractional Vegetation Cover (FVC), and surface emissivity ($\:\epsilon\:$)⁴⁵. NDVI is calculated as:

$$\:NDVI=\left(NIR-IR\right)/\left(NIR+IR\right)$$

(6)

where $\:NIR$ and $\:IR$ denote reflectance values in the near-infrared and red bands, respectively. Higher NDVI values indicate greater vegetation density. Based on NDVI, FVC is computed using the pixel dichotomy model:

$$\:FVC=\left(NDVI-{NDVI}_{soil}\right)/\left({NDVI}_{veg}-{NDVI}_{soil}\right)$$

(7)

where $\:{NDVI}_{soil}$ and $\:{NDVI}_{veg}$ represent the NDVI values of bare soil and fully vegetated surfaces, respectively, and are set to the 5th and 95th percentile values across the study area following Sun et al. (2024)⁴⁶.

Surface emissivity ($\:\epsilon\:$), a critical input for the MW algorithm, varies with land cover type and is estimated based on FVC. Land surfaces are classified into three categories: water bodies, natural surfaces, and built-up or bare surfaces. The emissivity for each category is estimated using the following empirical relationships⁴⁴:

$$\:{\epsilon\:}_{w}=0.995$$

(8)

$$\:{\epsilon\:}_{ns}=0.9625+0.0614\times\:FVC-0.0461{\times\:FVC}^{2}$$

(9)

$$\:{\epsilon\:}_{b}=0.9589+0.086\times\:FVC-0.0671{\times\:FVC}^{2}$$

(10)

where $\:{\epsilon\:}_{w}$ is used for water bodies, $\:{\epsilon\:}_{ns}$ for natural surfaces, and $\:{\epsilon\:}_{b}$ for built-up or bare surfaces. The overall surface emissivity for each pixel is then derived through weighted averaging based on the proportion of each surface category and incorporated into the LST retrieval process. All preprocessing steps for LST retrieval, including radiometric calibration, thermal band correction, emissivity estimation, and final LST generation, were conducted in ENVI 5.6.3.

Urban heat Island analysis

To assess the spatial distribution of UHI intensity, it is essential to account for temporal fluctuations in surface and atmospheric conditions. Direct comparisons of LST values across different time periods may yield biased results due to meteorological variability. To mitigate this issue and ensure robust intra-scene comparability, this study employed a standardization-based approach that transforms each pixel’s LST into a dimensionless UHI index, calculated based on its statistical deviation from the regional mean⁴⁷. Furthermore, to address the substantial spatiotemporal variability of UHI, we averaged the two LST scenes for each year, thereby deriving representative annual values for subsequent analysis. The UHI index is computed as:

$$\:UHI=\left(T-\mu\:\right)/std$$

(11)

where $\:T$ represents the LST of a given pixel, $\:\mu\:$ is the mean LST of the entire study area, and $\:std$ denotes the standard deviation. This formulation captures the degree to which each pixel’s temperature deviates from the regional average, providing a reliable basis for identifying thermal anomalies.

To further analyze the spatial variation of UHI intensity, the standardized LST values were classified into five thermal categories using a mean–standard deviation method⁴⁸. Each category reflects a specific level of thermal deviation relative to the scene-wide mean temperature. The classification criteria are presented in Table 2.

Table 2 Classification of land surface temperature levels based on standardized LST.

Green space spatial pattern assessment

Green space distribution and quality assessment

To systematically characterize the spatial distribution and ecological quality of UGS, this study defines UGS as encompassing four land cover types: trees, shrubland, grassland, and wetlands²⁸, all derived from the Dynamic World dataset. Three representative distributional metrics were selected: Green Space Density (GSD), Distance to Green Space (DGS), and NDVI. Specifically, GSD quantifies the proportion of UGS within each analysis unit, indicating the overall level of spatial coverage²¹. DGS, measured in meters, represents the Euclidean distance from each pixel to the nearest UGS boundary, capturing spatial proximity²³. NDVI reflects vegetation greenness and vitality, serving as a key ecological indicator of UGS quality²².

Green space structural characteristics assessment

To comprehensively characterize the structural attributes of UGS, this study selected five representative landscape metrics that reflect distinct dimensions: patch coverage, fragmentation, connectivity, shape complexity, and internal compactness (Table 3)^22,27. Considering the pronounced spatial heterogeneity of landscape patterns in urban environments, a moving window approach was adopted to compute and analyze the spatial dynamics of these metrics. Based on previous studies, a 600 m × 600 m grid size was selected as the spatial analysis unit, as this scale has been widely recognized as appropriate for capturing urban-scale landscape variations²⁸. All metrics were calculated at the landscape level using Fragstats 4.2.

Table 3 Landscape metrics describe green space structure.

Impact of green space on urban thermal environment

Quantifying the cooling effect of green space

To ensure consistency with the spatial units used in regression modeling, this study adopts a grid-based approach to quantify the cooling effect of UGS. Specifically, the Cooling Effect Index (CEI) is calculated by measuring the difference in mean LST between green-covered and non-green areas within each analysis grid²³. This method enables a consistent assessment of the local cooling capacity of UGSs at the same spatial resolution as the structural metrics, thereby facilitating direct comparison and integration with explanatory variables in the subsequent modeling process.

Global correlation analysis between green space and cooling effects

Before constructing spatial regression models to examine the impact of UGS on the urban thermal environment, diagnostic assessments of both the dependent and explanatory variables are essential to ensure the robustness and interpretability of the modeling results²⁰. In this study, the CEI was used as the dependent variable. To assess its global spatial autocorrelation, Moran’s I statistic was calculated, formulated as:

$$\:I=\frac{n}{{\sum\:}_{i=1}^{n}{\sum\:}_{j=1}^{n}{w}_{ij}}\cdot\:\frac{{\sum\:}_{i=1}^{n}{\sum\:}_{j=1}^{n}{w}_{ij}\left({x}_{i}-\stackrel{-}{x}\right)\left({x}_{j}-\stackrel{-}{x}\right)}{{\sum\:}_{i=1}^{n}{\left({x}_{i}-\stackrel{-}{x}\right)}^{2}}$$

(12)

where $\:n$ denotes the number of spatial units, $\:{x}_{i}$ and $\:{x}_{j}$ represent the LST values of units $\:i$ and $\:j$, $\:\stackrel{-}{x}$ is the global mean LST, and $\:{w}_{ij}$ is the spatial weight. A positive Moran’s I (I > 0) indicates clustering of similar temperature values, suggesting the presence of thermal hotspots or cool zones. A negative value (I < 0) implies spatial dispersion, while values close to zero indicate a random spatial distribution of LST across the study area. In addition, local spatial autocorrelation was examined through Local Indicators of Spatial Association, which were implemented using Anselin’s Local Moran’s I with a Queen’s contiguity spatial weights matrix³⁸. Clusters of High–High, Low–Low, High–Low, and Low–High were identified at a significance level of p < 0.05, serving as a complement to the global Moran’s I⁴⁹. To mitigate multicollinearity among explanatory variables, an Ordinary Least Squares (OLS) regression was first performed to calculate the Variance Inflation Factor (VIF) for each candidate variable. Variables with VIF values exceeding 7.5 were excluded to improve model stability and interpretability⁵⁰.

To quantify the overall influence of UGS characteristics on cooling effects, the RF regression model was employed for global analysis. RF, a robust ensemble learning method, can effectively capture nonlinear relationships and complex interactions without requiring assumptions about data distribution or model specification³⁶. This makes it well-suited for high-dimensional and heterogeneous spatial datasets. To ensure a fair and rigorous model selection process, RF was systematically compared with other widely used ensemble algorithms, including GBDT, XGBoost, and LightGBM. For each candidate model, hyperparameters were optimized through a grid search combined with 5-fold cross-validation, and training–validation procedures were repeated across different partitions to ensure robust performance and avoid overfitting^51,52. Model performance was evaluated using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE), allowing an objective comparison under their optimal configurations^33,34. After model training, SHAP values were used to interpret the contribution of each variable to CEI prediction³⁵. Rooted in cooperative game theory, SHAP attributes the prediction fairly to each feature by considering both marginal effects and feature interactions^37,52. This provides an interpretable and transparent explanation of how different UGS characteristics influence cooling effects at the global level, serving as a theoretical foundation for the subsequent spatially explicit modeling.

Local correlation analysis between green space and cooling effects

To account for potential spatial heterogeneity in the effects of UGS on the urban thermal environment, this study employed the MGWR model to capture local variations in explanatory strength. Unlike traditional regression models with fixed global parameters, MGWR incorporates a spatial weighting mechanism that allows model coefficients to vary across geographic locations, making it suitable for identifying localized relationships between UGS characteristics and cooling intensity³⁸. To maintain consistency with the preceding analysis of UGS structural metrics, the same moving window scale was adopted as the spatial analysis unit for subsequent spatial modeling. The MGWR model was implemented using MGWR 2.2, and is formulated as:

$$\:{Y}_{i}={\beta\:}_{0}\left({u}_{i},{v}_{i}\right)+\sum\:_{j=1}^{n}{\beta\:}_{j}\left({u}_{i},{v}_{i}\right){X}_{ij}+{\epsilon}_{i}$$

(13)

where $\:{Y}_{i}$ denotes the CEI value for grid unit $\:i$, $\:{X}_{ij}$ is the j-th explanatory variable in unit $\:i$, $\:{\beta\:}_{j}\left({u}_{i},{v}_{i}\right)$ represents the spatially varying coefficient of variable $\:j$ at location $\:\left({u}_{i},{v}_{i}\right)$. $\:{\beta\:}_{0}\left({u}_{i},{v}_{i}\right)$ is the local interception, and $\:{\epsilon}_{i}$ is the residual term. The model was fitted using a Gaussian distribution, with an adaptive bisquare kernel for spatial weighting and bandwidths determined by the AICc criterion. Bandwidths were adaptive and variable-specific, enabling multi-scale analysis. All coordinates were projected to WGS_1984_UTM_Zone_49N, and Euclidean distance was used as the metric. Coefficient significance was evaluated using adjusted t-values at the 95% confidence level.

Results

LST and UHI intensity

LST values were retrieved for four time points using the MW algorithm. In 2019, the LST ranged from 5.45 °C to 48.45 °C on October 29 and from 12.24 °C to 44.54 °C on November 14. In 2024, the LST ranged from 8.40 °C to 59.43 °C on October 10 and from 8.07 °C to 42.30 °C on November 27. To address missing data caused by cloud coverage, the LST values were normalized and averaged, providing mean normalized LST values for the years 2019 and 2024.

Normalized LST values were then used to derive spatial representations of UHI intensity (Fig. 3). The total area of low-temperature zones, including the Ltz and Sltz, remained nearly constant, with a slight increase from 2287.55 km² in 2019 to 2288.09 km² in 2024. In contrast, the area of the Mtz decreased from 2820.01 km² to 2720.07 km². The high-temperature zones, including the Htz and Shtz, expanded from 2176.99 km² to 2276.39 km², with approximately 100 km² shifting from the middle-temperature zone to the high-temperature zone (Table 4), indicating a continued expansion of UHI.

Spatially, the intensity and area of UHI decreased in the western and northern parts of Guangzhou, specifically in Huadu District and Conghua District, between 2019 and 2024. However, the southern regions, including Panyu District and Nansha District, saw significant increases in both UHI intensity and area. This suggests a southward shift in UHI intensity over the study period.

Fig. 3

UHI distribution in 2019 and 2024.

Table 4 UHI classification area and percentage for 2019 and 2024.

Green space Spatial patterns and quality

The spatial distribution of UGS indicators in Guangzhou revealed distinct regional patterns (Fig. 4). High values of GSD, PLAND, and NDVI were concentrated in the central and northern areas, such as Baiyun Mountain Forest Park and Maofeng Mountain Forest Park. These regions exhibited high UGS density, coverage, and vegetation health. Conversely, the western region, from Baiyun Railway Station to Baiyun Airport, and the southeastern area, including Panyu District and Nansha District, showed lower values, indicating an uneven distribution of UGS (Fig. 4a, c and d). The DGS was notably lower in the western part of the city center, suggesting poorer accessibility to UGSs in high-density urban areas compared to suburban areas (Fig. 4b). The COHESION, CONTIG_MN, and PD indicated lower connectivity of UGSs in the western city center and Panyu District. In contrast, the central and northern areas showed higher aggregation and connectivity of UGSs (Fig. 4e, f and h). The distribution of LSI and CONTIG_MN was relatively uniform, with localized complexity, reflecting subtle variations in the shape and connectivity of UGSs at the urban periphery.

Table 5 presents the magnitude and variability of changes in UGS indicators between 2019 and 2024. NDVI increased by 0.016 from 2019 to 2024, indicating improved vegetation health and reflecting the effectiveness of urban greening initiatives. Over the same period, DGS declined markedly by − 6.392, suggesting enhanced accessibility of UGSs, although localized reductions in the western city center and Nansha District warrant further spatial examination. PD rose significantly by 0.576, particularly in the west of the city center, including Panyu District, and in the corridor between Baiyun Railway Station and Baiyun Airport. In contrast, PLAND decreased by − 0.736, and GSD showed only a marginal decline of − 0.007, pointing to increasing fragmentation of UGSs in these areas, albeit without strong supporting evidence from GSD. Meanwhile, both LSI and cohesion increased by 0.079 and 0.512, respectively, indicating a deeper structural integration of UGSs within the urban fabric.

Fig. 4

Spatial patterns and quality of green space in 2019 and 2024.

Table 5 Magnitude and variability of green space feature changes in 2019 and 2024.

Relationship between green space and urban thermal environment

Cooling effect of green space

In 2019 and 2024, the spatial distribution of the CEI revealed distinct cooling patterns across Guangzhou (Fig. 5). Areas with higher CEI values aligned closely with low-temperature zones, confirming the significant role of UGSs in reducing urban surface temperatures. The highest cooling effects were concentrated in the central and northeastern regions, where continuous forests and natural hills are located. In these areas, CEI values typically exceeded 0.45, indicating a strong cooling capacity. In contrast, the urban core and southern districts consistently showed lower CEI values, generally below 0.30, indicating a weaker cooling effect.

Spatial disparities in cooling performance were evident across the city. Suburban UGSs, huge forest patches in the northern hills, provided a significantly more substantial cooling effect compared to fragmented UGSs within the high-density urban center. This pattern reflects the influence of patch size and vegetation continuity on local thermal regulation, as small, isolated UGSs embedded within dense built-up areas were less effective in mitigating urban heat.

Fig. 5

Spatial distribution of CEI in 2019 and 2024.

Global correlation between green space and cooling effects

The spatial distribution of UGS cooling effects showed significant spatial clustering across Guangzhou. Moran’s I values of the CEI were 0.511 in 2019 and 0.506 in 2024, with corresponding Z-scores of 20.201 and 19.912, and P-values of 0.000 in both years (Table 6). Local indicators of spatial association (Fig. 6) further showed stable spatial patterns. High–high clusters were concentrated in the northern forested areas, accounting for 15.53% of the total in 2019 and 15.95% in 2024, with small low–high outliers scattered in the surrounding zones at 3.20% and 3.13%, respectively. Low–low clusters were primarily located in the central and southern urban zones, expanding from 18.39% in 2019 to 19.18% in 2024. The slight decline in Moran’s I suggests an increase in spatial heterogeneity of cooling effects between 2019 and 2024.

Table 6 Global Spatial autocorrelation.

Fig. 6

Local indicators of spatial association of UGS cooling effects in 2019 and 2024.

Prior to regression modeling, multicollinearity among the explanatory variables was assessed using VIF. DGS, NDVI, LSI, COHESION, and PD were retained as final explanatory variables not only because their VIF values were all below the threshold of 7.5, but also due to their strong correlations with the CEI (Table S1).

After comparing four machine learning algorithms (Table S2), the RF regression model exhibited the best predictive performance. The optimal RF configuration in both years was max depth = 7, min samples leaf = 1, and max features = 1, while min samples split and n estimators differed: in 2019 they were min samples split = 8 and n estimators = 400, whereas in 2024 they were min samples split = 5 and n estimators = 600. Moreover, as shown in Table S3, the training and test sets in both 2019 and 2024 achieved comparable MAE, RMSE, and R² values, indicating that the RF model maintained stable generalization performance without signs of overfitting.

Feature importance analysis based on SHAP values (Fig. 7) showed that DGS was the dominant factor influencing CEI. Its relative importance increased from 51.8% in 2019 to 75.3% in 2024. SHAP values further indicated a consistent and strong negative relationship between DGS and CEI. NDVI contributed 3.9% in 2019 and increased to 8.2% in 2024. In contrast, the importance of LSI and COHESION decreased over time, with LSI contributing 29.0% in 2019 and only 5.4% in 2024, while COHESION decreased from 14.3% to 9.9%. PD consistently exhibited minimal influence on CEI, contributing between 1.0% and 1.2%.

Fig. 7

SHAP-derived importance of green space indicators for CEI in 2019 and 2024.

Local correlation between green space and cooling effects

The spatially varying effects of UGS on urban cooling were assessed using the MGWR model, applied to 20,915 spatial units across the study area. Model comparison demonstrated that MGWR outperformed the global OLS model in capturing the spatial heterogeneity of UGS effects on CEI (Table 7). In 2019, the MGWR model achieved an R² value of 0.772, which was substantially higher than the 0.482 recorded for the OLS model. Similarly, in 2024, the R² value of MGWR reached 0.738, compared to 0.491 for OLS. Improvements in adjusted R² were also observed. In both years, the Akaike Information Criterion corrected (AICc) values of MGWR were lower than those of OLS, confirming the superior model fit and enhanced explanatory power of MGWR.

As shown in Table S4, the bandwidth in 2024 is more variable, indicating that the MGWR model captures more complex spatial patterns compared to 2019. Residual spatial autocorrelation was also checked using Global Moran’s I on MGWR residuals, yielding values of − 0.109 for 2019 and − 0.014 for 2024, indicating that the model has a better fitting effect in 2024 and has effectively eliminated most of the spatial dependencies.

Table 7 Comparison of OLS and MGWR models.

Table 8 MGWR regression results.

Regression coefficients obtained from the MGWR model were spatially visualized to depict the geographic variation in predictor effects (Table 8; Fig. 8). In 2019, the average regression coefficient of DGS with CEI was 0.830, reflecting firm spatial heterogeneity. Positive effects largely influenced this positive mean in the forested regions of Conghua and northwestern Huadu, whereas adverse effects dominated the central and southern urban areas. By 2024, the relationship shifted to a consistently negative effect across the city, with only slight south–north differences and coefficients ranging from − 0.081 to − 0.074. This indicates that the association between DGS and CEI became highly homogeneous, reflecting a nearly uniform negative relationship.

NDVI showed an average regression coefficient of 0.139 with CEI in 2019. Positive effects were mainly concentrated in forested hills and suburban woodlands, where vegetation density and coverage were relatively high. In contrast, negative correlations occurred in the urban core, the Panyu District, and development corridors, such as the Baiyun Airport area. By 2024, the effect of NDVI on CEI weakened, with negative associations expanding across larger areas. The average coefficient declined to − 0.058, and spatial variation contracted considerably, suggesting a more homogeneous but less influential role of vegetation greenness in urban cooling.

PD had an average regression coefficient of 0.220 in 2019, indicating positive associations with CEI in most forested and suburban areas, such as Baiyun Mountain and Lu Tian Town. In contrast, several regions in the northwestern and southern parts of Conghua exhibited negative associations. By 2024, the average coefficient decreased to 0.104, and the spatial pattern shifted from mixed positive–negative associations to a continuous gradient, with the strength of the relationship progressively increasing from south to north.

COHESION exhibited negative associations with CEI in both years, with relatively stronger magnitudes overall. In 2019, the average coefficient was − 0.490, with negative correlations predominantly distributed in northern urban areas, particularly central Conghua and parts of Huadu. In contrast, positive relationships were observed in southern districts, such as Panyu and Nansha. By 2024, the average coefficient decreased to − 0.349, and correlations weakened markedly across both directions. Although the spatial pattern remained broadly consistent, the strength of both positive and negative associations declined, indicating a general reduction in the influence of landscape connectivity on cooling effectiveness.

LSI exerted predominantly adverse effects on CEI in both periods, with only a few small positive patches persisting in the southern areas. The mean coefficient declined slightly from − 0.880 in 2019 to − 0.920 in 2024. Although the strongest negative association weakened, spatial variability diminished, and regions of high negative values expanded outward from the northern mountainous belt. In contrast, the southern positive patches continued to contract. By 2024, nearly all areas exhibited negative correlations, most evident in the south core, where UGSs were more irregular and fragmented in form.

Fig. 8

Spatial variation of green space effects on urban cooling in 2019 and 2024.

Discussion

Crucial contribution of green spaces in mitigating UHI effects

UGSs play a pivotal role in mitigating UHI effects, yet their cooling capacity is highly heterogeneity across spatial patterns and scales. In Guangzhou, the distribution of cooling effects follows a clear gradient, with restricted impacts in the urban core, significant cooling in mountainous forests, and delayed mitigation in newly developed areas. The effectiveness of UHI mitigation depends not only on the total amount of UGS but also on the continuity and integrity of large patches as well as the spatial configuration of the UGS system.

Large forests and parks in the northern and western parts of Guangzhou, particularly Baiyun Mountain and Maofeng Mountain, serve as core areas. The continuous forest patches here overlap with low temperature zones⁵³, and reduce surface heat by enhancing latent heat flux while suppressing sensible heat flux, thereby stabilizing local energy balance⁵⁴. The canopy blocks solar radiation, while the high albedo and soil moisture of the forest floor further decrease radiative temperatures⁵⁵. Additionally, the mountainous terrain induces slope winds and convective exchange, reinforcing vertical heat diffusion and creating a macro scale cooling source⁵⁶. These findings are consistent with studies from other subtropical cities such as Hong Kong and Hangzhou^57,58. The quantified results further demonstrate that although COHESION generally exerts negative correlations with cooling, the northern forests exhibit local positive effects. This suggests that when strong connectivity is combined with complex topography and natural ventilation, the cooling function of UGSs can be amplified, highlighting the importance of synergy between ecological structure and terrain.

In the urban core, cooling effects remain weak despite an overall increase in green coverage. Much of the additional greenery consists of small parks and roadside vegetation that are highly fragmented. Our analysis shows that high PD values correspond to numerous small patches; however, the positive association between PD and cooling has weakened over time, and its spatial heterogeneity has decreased. This suggests that excessive fragmentation is unable to sustain regional cooling and may compromise ecological functionality. The low COHESION values confirm limited connectivity among patches, restricting the spatial diffusion of cooling. LSI exerts a strong negative influence, reflecting that irregular and complex patch shapes undermine efficiency. Consequently, small UGSs in the dense core, constrained by insufficient area, significant gaps, and poor continuity, cannot form effective cooling surfaces⁵⁹. Their transpiration and shading effects are confined to localized spaces²⁵. High intensity building development further impedes airflow and heat diffusion, intensifying heat retention⁶⁰. This situation resembles observations in Melbourne and Beijing^61,62, though Guangzhou’s subtropical monsoon climate with high temperatures and humidity produces a higher baseline heat environment than temperate cities. Under such humid conditions, evaporative cooling is constrained, and with urban wind channels blocked, the dependence on air dynamics exacerbates heat accumulation⁶³.

The expansion areas in the south and east, such as Panyu and Nansha, illustrate another pattern. Urban growth has outpaced the development of UGSs, which remain fragmented and low in tree coverage. PD has increased, but its maximum correlation with cooling has declined, indicating a contraction in the spatial extent of positive effects. Low COHESION and negative LSI values further demonstrate that isolated and irregularly shaped patches lack sufficient buffering capacity. Without continuous forests or structured green wedges, as found in Munich or Toronto, the new districts of Guangzhou fail to provide effective thermal regulation⁶⁴. This outcome parallels evidence from other rapidly urbanizing Asian cities including Shanghai and Mumbai^27,65.

The mitigation of UHI by UGSs involves multiple physical and ecological processes. Large and cohesive patches enhance latent heat flux, reduce sensible heat flux, and suppresses surface temperatures warming⁵⁴. They also facilitate ventilation and convective exchange, stabilizing the thermal environment at broader scales⁶⁶. Conversely, small and fragmented UGSs influence only local balances, contributing minimally to overall UHI intensity²⁵. The quantitative metrics support this distinction. PD exhibits limited and diminishing positive contributions, COHESION is predominantly negative except where connectivity interacts with favorable terrain, and LSI consistently shows strong negative impacts, particularly in dense urban and expansion zones. These results confirm the direction and magnitude of fragmentation and connectivity effects, reinforcing that optimization of spatial configuration is as critical as increasing the total area of green space.

Mechanisms of green space characteristics in cooling effects

The ability of UGSs to mitigate the UHI effect is shaped not only by their extent but also by spatial characteristics that operate across scales. Citywide SHAP analysis captures the dominant contributions of individual factors, while MGWR coefficients reveal localized heterogeneity and interannual variability. Together, these perspectives demonstrate that between 2019 and 2024 the cooling role of UGSs in Guangzhou evolved dynamically, reflecting both urban expansion and internal spatial restructuring.

DGS was the most influential predictor of CEI, with SHAP importance rising from 51.8% in 2019 to 75.3% in 2024, confirming proximity to UGSs as the dominant determinant of cooling capacity⁶⁷. Negative SHAP values in both years underscored the robustness of accessibility as a citywide driver. MGWR revealed clear temporal shifts. In 2019, coefficients displayed strong spatial heterogeneity. Negative correlations concentrated in central and southern districts where insufficient green space aggravated heat stress, while positive values appeared in forested areas of Conghua and northwestern Huadu where accessibility had already reached saturation⁶⁸. Thus, the positive mean coefficient in 2019 reflected spatial heterogeneity rather than a contradiction of the global trend, as negative associations dominated urban districts while strong positive values in forested areas skewed the average upward. By 2024, coefficients converged to a stable negative pattern with little variability. This transition reflected the combined influence of urban development and land use policy. In the southern new districts, delays in park and green space provision intensified sensitivity to distance, whereas in the northern forest areas long-standing ecological protection reduced marginal gains from additional proximity. Mechanistically, dependence on localized ventilation corridors gave way to broader spatial coherence as green space distribution became more balanced⁶⁹. Guangzhou’s infill-oriented expansion, which often left UGSs fragmented in residual land, limited the formation of effective networks. In contrast, cities adopting proactive growth boundaries and protection measures, such as Portland, maintained connectivity and enhanced the resilience of urban cooling²⁶.

NDVI illustrates the contrast between city-scale consistency and local complexity. SHAP importance increased from 3.9% to 8.2%, suggesting that vegetation vitality gained relevance, possibly reflecting policy-driven greening. Nevertheless, MGWR coefficients showed pronounced heterogeneity. In 2019, the mean coefficient was positive, with strong cooling in suburban forests and mountainous areas, yet negative values in dense urban corridors such as Panyu District and Baiyun Airport where fragmentation constrained transpiration. By 2024, the mean coefficient became slightly negative, and spatial variability diminished, with negative effects emerging even in peri-urban zones. This indicates that vitality alone cannot guarantee cooling when canopy continuity is lacking⁷⁰. In Guangzhou’s core, declines in NDVI effectiveness were compounded by soil degradation, pollution, and human disturbance⁷¹. While similar declines have been observed in Indianapolis⁷², Guangzhou demonstrates how accelerated urbanization erodes NDVI reliability more rapidly, emphasizing that structural integrity and vegetation diversity are prerequisites for sustainable cooling.

LSI further illustrates the interplay of citywide decline and local suppression. SHAP importance dropped sharply from 29.0% in 2019 to 5.4% in 2024, suggesting its diminishing relative contribution. MGWR coefficients, however, remained strongly negative, converging toward universally suppressive effects. In 2019, isolated zones such as Dafu Mountain or Shiba Luohan Shan showed localized positive coefficients, where irregular edges occasionally enhanced ventilation. By 2024, nearly all districts exhibited negative values, particularly in the southern core. This convergence indicates that UGS boundaries intensified edge heating and constrained the development of coherent cool islands⁷³. The suppressive role of LSI, amplified by Guangzhou’s dominance of small-scale parks, mirrors findings in US cities⁷⁴, but is further reinforced by local topography and uneven renewal policies.

COHESION also demonstrated divergence between SHAP and MGWR. SHAP importance decreased from 14.3% to 9.9%, reflecting a weaker citywide contribution, while MGWR coefficients remained predominantly negative. In 2019, northern districts such as Conghua and Zengcheng showed strong negative coefficients, where excessive cohesion limited air circulation, while southern districts revealed occasional positive values. By 2024, some localized positive associations appeared in Nansha, but negative effects persisted in far northern suburbs. This suggests that overly cohesive UGS patches may hinder ventilation and facilitate heat accumulation rather than dissipation²⁸. Compared with Berlin and Melbourne⁷⁵, Guangzhou displays stronger dependence on functional rather than structural connectivity, reflecting its subtropical climate and rapid urban growth.

PD maintained consistently low SHAP importance, aligning with its limited metropolitan role. MGWR coefficients, however, revealed localized effects. In 2019, the mean was positive, with cooling benefits evident in Baiyun Mountain and Lvtian where denser patches collectively enhanced evapotranspiration. By 2024, the mean coefficient declined, and spatial range contracted, reflecting that fragmentation-driven increases in patch number failed to yield cumulative cooling. Instead, isolated patches functioned as inefficient fragmented cooling sources⁷⁶. This process, consistent with findings in Karachi²¹, underscores how unchecked fragmentation under rapid urbanization weakens cooperative ecosystem functions.

Policy implications for strengthening urban green spaces to mitigate UHI effects

In recent years, Guangzhou has made significant progress in addressing the UHI effect, developing a comprehensive policy framework that spans from general to specialized planning, and from pilot projects to technical guidelines. By integrating UGS planning with UHI mitigation, Guangzhou has effectively translated this concept into action⁷⁷. The “14th Five-Year Plan for Ecological and Environmental Protection” in 2022 marked the first inclusion of UHI improvement as a key project for climate adaptation, advancing eco-district construction and research on urban cooling technologies. In 2023, the “Guangzhou Land Spatial Master Plan (2021–2035)” refined and codified the spatial layout of UGSs, including networked park systems and rooftop and vertical greening, which provided an operational basis for temperature reduction goals. These instruments advanced UHI mitigation from plan making to citywide execution during the early plan period and clarified staged intervention windows from adoption through the plan’s near-term 2025 targets. The patterns observed in 2024 are consistent with the expected cooling pathways of increased vegetative cover, improved connectivity of green spaces, and building level greening, with the strongest signals in core districts such as Yuexiu and Tianhe. Simultaneously, areas with high-intensity development, like the Sino-Singapore Guangzhou Knowledge City and Nansha Mingzhu Bay, have integrated cooling UGSs and wind corridors in the planning phase, aligning development with temperature reduction.

Despite these advancements, several challenges persist in the layout and implementation of UGSs. In old urban areas, limited UGS and land resources hinder effective UHI mitigation. Large new parks and UGSs are primarily concentrated on the city’s outskirts, such as Nansha Mingzhu Bay and the Panyu South Da Gan Line, while the urban core continues to face a relative shortage of public green space. This uneven distribution impedes comprehensive temperature reduction, with peripheral areas experiencing more significant cooling effects than central zones. In addition, new districts like Nansha and Huangpu, while meeting UGS requirements, still suffer from high surface temperatures due to insufficient tree canopy coverage. Furthermore, the fragmented funding and governance mechanisms have impaired the long-term maintenance of UGS projects, particularly rooftop greening. The implementation of rooftop greening is hindered by limited coverage and the fragmented ownership of older buildings. Some completed greening projects have been abandoned due to lack of management, even contributing to new heat sources, thereby exacerbating UHI challenges.

Based on this study’s findings, future optimization of Guangzhou’s UGSs requires targeted adjustments across several dimensions. First, in old districts with limited ground space, small pocket parks on vacant lots and underused corners, together with rooftop and vertical greenery, can provide effective cooling⁷⁸. In dense cores such as Yuexiu and Liwan, empirical studies indicate that pocket parks within a 200 to 300 m walking radius may reduce pedestrian-level air temperature by about 0.5 °C and improve thermal comfort⁷⁹. Vertical greenery has been shown to lower façade surface temperature by up to 11.5 °C and adjacent air temperature by around 3°C⁸⁰, while rooftop greening can decrease above-roof air temperature by about 1°C⁸¹. Second, the connectivity of UGSs requires improvement, especially in new and compact developments. Creating a green network through street trees and linear greenways, as demonstrated by Singapore’s Park Connector Network, can form a more integrated point–line–surface cooling system⁸². Third, policy should strengthen mandatory requirements for rooftop and vertical greening, particularly in new projects and provide financial incentives for long-term maintenance⁸³. Finally, monitoring and dynamic assessment processes need to be enhanced. Stronger interdepartmental collaboration, supported by a one-stop approval system, can improve policy execution and ensure timely implementation of cooling measures.

Limitations and future research

This study provides a comprehensive and spatially explicit assessment of the relationship between UGS and cooling effects at global and local scales to inform urban thermal mitigation planning, though several limitations remain. Firstly, given the large spatial extent of Guangzhou and the inclusion of historical time points, LST retrieval in this study relied entirely on satellite-based remote sensing data to ensure timely and city-wide coverage. The absence of ground-based temperature measurements may introduce uncertainties in LST accuracy. Future studies should integrate field monitoring to calibrate and validate satellite-derived LST across different surface types and microclimates, and incorporate additional observation years at higher frequency to identify the timing of policy interventions and capture the dynamic evolution of their effects. Secondly, while this study evaluated the overall influence of UGS structure and distribution on urban cooling, it did not distinguish the effects of different vegetation types. Considering the structural and physiological differences among trees, shrubs, and grasslands, future studies should explore the differentiated cooling performance to provide more targeted recommendations for UGS design and management. Thirdly, while the modeling framework revealed important spatial heterogeneity, methodological uncertainties remain. Future studies could enhance robustness by evaluating the sensitivity of MGWR bandwidth selection and by comparing outcomes with alternative spatial modeling approaches. Finally, although Guangzhou represents a typical subtropical megacity, the extent to which these findings are transferable to other urban contexts requires further testing across diverse climatic and morphological settings to establish broader applicability.

Conclusion

This study systematically evaluated the spatial relationship between UGS characteristics and urban cooling effects in Guangzhou, integrating multi-source remote sensing data with both global and local spatial modeling approaches. LST was retrieved for four representative periods in 2019 and 2024, corresponding to the pre- and post-implementation phases of the Guangzhou Green Space System Plan (2021–2035). UHI intensity was quantified using standardized LST, while UGS characteristics, including spatial distribution, quantity, and structural configuration, were comprehensively assessed to analyze their influence on urban thermal patterns. The results revealed clear spatial and temporal dynamics in both LST and UHI intensity. From 2019 to 2024, high-temperature zones expanded from 2,176.99 km² to 2,276.39 km², indicating continued intensification of the UHI effect, especially in the southern districts such as Panyu and Nansha. Meanwhile, UGS coverage and vegetation health showed significant improvements, as indicated by increases in GSD, PLAND, and NDVI. However, particular areas experienced increased fragmentation of UGSs, reflecting the ecological challenges posed by ongoing urban development. Analysis of cooling effects confirmed the important role of UGS in mitigating surface temperatures. Higher CEI values were concentrated in the northern forested regions, typically exceeding 0.45. Random Forest modeling identified proximity to UGSs (represented by DGS) as the dominant variable influencing local cooling effects, accounting for up to 75.3% of the explanatory power in 2024. MGWR analysis further revealed pronounced spatial heterogeneity in UGS cooling effects, with proximity to UGSs showing an increasingly consistent negative relationship with LST across the city. Overall, this study highlights the critical role of UGS in alleviating UHI intensity and underscores the need for targeted UGS planning. Future strategies should prioritize optimizing the spatial distribution, quantity, and structural connectivity of UGSs to enhance their cooling capacity and improve urban thermal comfort in subtropical megacities.