A Transferable Modeling Framework for Improving the Cooling Effect of Urban Green Space: Multi-Temporal Sampling, 3D Morphological Reconstruction and Bayesian Network

Highlights

What are the main findings?

• A transferable multi-temporal sampling and 3D modeling framework was developed to overcome limited field data and capture internal and surrounding structural effects on urban green space (UGS) cooling;
• The UGS cooling effect is mainly associated with meteorological conditions (air temperature and wind speed) and 3D configuration (area, shape, and surrounding ventilation), with “win–win” performance in large, regularly shaped, moderately ventilated patches.

What are the implications of the main findings?

• This framework enables mechanism-based, scalable assessment of UGS cooling under data-limited conditions;
• Findings inform climate-responsive green-infrastructure design and urban ventilation optimization in heat-vulnerable cities.

Abstract

Accurate assessment of the cooling effect from urban green space (UGS) is largely hindered by insufficient field samples or consideration of the internal and surrounding three-dimensional (3D) structure. This study developed a transferable modeling-optimization framework that integrated a multi-temporal sampling strategy, multimodal 3D environmental reconstruction, and Bayesian-based optimization. First, the potential influencing factors of the cooling effect were quantified from three aspects of inner 2D/3D structure, surrounding building ventilation, and background meteorology through fusing field measurements, multi-spectral UAV images, and Sentinel-2 images. Then, a generalized additive mixed-effects model was used to explore cooling-related patterns of UGS, and a Bayesian network was further applied to identify potential optimized configurations. The results suggest the following: (1) The adopted multi-temporal sampling strategy enhances the stability of detected cooling signals and minimizes spatial interference among neighboring UGS patches and water bodies. (2) Temporal changes in the cooling effect are mainly driven by average air temperature and maximum wind speed, while the spatial variation by the UGS inner characteristics of area and shape index and surrounding ventilation. (3) The “win–win” situation of cooling intensity and range occurred in UGSs with larger areas, higher shape regularity, and medium ventilation. This approach is useful for model-based planning of climate-responsive green infrastructure and city-scale ventilation systems in heat-vulnerable environments.

1. Introduction

Driven by global climate change and human activities, urban areas tend to have higher temperatures than suburban and surrounding natural landscapes, referred to as urban heat islands (UHIs) [1]. Nearly 55% of the global population lived in urban areas in 2018, and more critically, the intensifying UHI effect poses a significant threat to residents’ health and even causes death [2]. Urban green space (UGS) has been demonstrated as one of the most effective nature-based solutions that moderates ambient temperature through shading and evapotranspiration while providing multiple co-benefits, such as heat mitigation, climate-change alleviation, air-quality improvement, and recreational value [3]. Moreover, in most cities, the cooler air produced by UGS can blow into surrounding areas and generate a cooling effect on adjacent areas through air convection and diffusion, which can act as a promising way for improving thermal comfort in densely populated urban areas [4]. How to reasonably allocate the cooling effect of UGS has become a critical issue in urban renewal and sustainable planning [5].

The cooling effect of UGS is mainly impacted by three types of factors: (1) the internal structure, such as area, shape, and vegetation abundance; (2) the external environment, such as surrounding landscape pattern, ventilation, and building density; and (3) background factors, such as meteorological conditions and socioeconomic development [6,7,8]. However, controversial conclusions remain regarding the driving mechanism of the UGS cooling effect. For example, both positive and negative correlations have been suggested between the characteristics of UGS (such as size and shape) and the cooling effect [9,10]. Such discrepancies mainly arise from the incomplete consideration of multiple influencing factors (i.e., the above-listed internal, external, and background factors) and diverse sampling techniques (i.e., whether strictly selecting samples without nearby UGSs or water bodies) [11]. As ideal samples with sufficient distance from other blue–green spaces are rare in most cities, existing studies often include the area, shape, or proximity of neighboring green–blue spaces as an explanatory variable to partially control for their interactive effects [12]. Besides that, most previous studies have assessed the cooling effect at specific times or limited time points, which also contributes to inconsistent findings [6]. Therefore, a systematic and generalizable modeling framework that accounts for temporal dynamics and multi-source drivers is urgently needed to better simulate and interpret the UGS cooling effect [13,14].

Rapid urbanization processes, accompanied with substantial growth in both horizontal and vertical directions, significantly alter air movement and circulation and further impact the urban microclimate [15]. Low horizontal wind speed within a city is always associated with poor ventilation potential, induced by complex underlying surface roughness [16]. Calm and stagnant conditions make cities accumulate and retain more heat, while winds promote convective cooling and heat advection [17]. Moreover, ventilation corridors formed by urban 3D morphology can introduce clean air, diluting and discharging exhaust gas and waste heat [18]. Thus, the optimization of ventilation environments has great potential to mitigate UHIs and enhance the spatial spillover effects of the UGS cooling service [19]. However, existing studies mainly rely on isolated 3D metrics (e.g., building height, density, or sky-view factor) to explore the impacts of vertical urban growth on the thermal environment, which cannot adequately characterize urban ventilation conditions [20]. As a result, how the surrounding ventilation environment—formed by complex 3D urban morphology—regulates the UGS cooling effect remains insufficiently understood, which hinders thermal mitigation and adaptation in cities with complex 3D morphology [21].

Several methods have been developed to simulate urban ventilation, including wind tunnel models, numerical models (such as computational fluid-dynamics models, weather research, and forecasting models), and an environmental factor simulation method [16]. The former two methods are unsuitable for city-scale high-resolution simulation due to high experiment costs, the high demand for computation resources, or their applicability for low-resolution large-area mesoscale simulation [17]. Meanwhile, the latter method combines morphological parameters to quantify ventilation resistance, such as topography, land cover, building frontal index, building height, and sky-view factor, which can simulate city-scale ventilation with high computational efficiency and intuitive maps [22]. Thus, quantifying urban ventilation using an environmental factor simulation method would be a promising way to analyze the impacts of urban 3D morphology on the UGS cooling effect [19]. From a broader perspective, advancing the understanding of the UGS cooling effect requires moving beyond isolated 3D building metrics toward a mechanism-oriented characterization of the surrounding ventilation environment shaped by complex urban morphology [21]. By explicitly operationalizing urban ventilation as a composite, transferable indicator and linking it with time-series sampling and probabilistic modeling, this study provides new evidence on how surrounding 3D morphology and climate jointly regulate both the intensity and spillover of the UGS cooling effect.

To bridge these gaps, this study developed a transferable modeling-optimization framework that integrates multi-temporal sampling with a multimodal 3D environmental reconstruction to characterize and enhance the UGS cooling effect. The framework fuses UAV multi-spectral imagery, Sentinel-2 images, and field measurements to capture both internal 3D greening structures and the surrounding ventilation environment. Using a generalized additive mixed-effects model (GAMM) and a Bayesian-based scenario optimization, the framework disentangles temporal and spatial drivers of the cooling effect and identifies optimal configurations. Lanzhou, a typical valley city in Northwestern China, provides an ideal laboratory for testing the proposed framework due to its scarce green–blue spaces and significant vertical growth rather than horizontal expansion. The main objectives were to: (1) compare the seasonal differences in the UGS cooling effect using longitudinal time-series data; (2) identify critical internal, external, and background drivers from a 3D perspective at different seasons and their non-linear impacts; and (3) establish a generalizable framework for climate-responsive urban design and decision support.

2. Materials and Methods

2.1. Study Area

Lanzhou, the capital city of Gansu Province, is a semi-arid city encircled by mountains located in Northwestern China, with Gaolan Mountain in the south and Baita Mountain in the north (Figure 1). The Yellow River flows through the main built-up area of Lanzhou from west to east, and results in the typical topographical feature of a river situated between two mountains. Thus, it is a typical valley city with a linear shape, that is, with an east–west span of over 30 km and a north–south span of only 5–10 km. During 1990–2020, the built-up area, GDP, and urban residents increased by 1.14, 40.43, and 0.95 times, respectively. Lanzhou has a mid-temperate continental climate with low annual precipitation, wind speed, and atmospheric humidity, strong solar radiation, cold winters, and hot summers. The hostile climate makes Lanzhou have a fragile natural ecosystem; thus, green infrastructure in urban areas demands high maintenance costs. Therefore, the limited capacity of land, over-exploitation of natural resources, water scarcity, and traffic congestion have seriously affected the living quality of residents.

2.2. Data Collection and Processing

This study was conducted based on multi-source and time-series observations (Figure 2). First, the cooling effect was quantified based on LST data retrieved from Landsat time-series TIRS images. Second, potential influencing factors were selected and quantified from the three aspects of UGS inner characteristics, the surrounding environment, and background climate conditions. Third, a generalized additive mixed-effects model (GAMM) was utilized to explore cooling-related patterns from the two perspectives of spatial variation and temporal change. Finally, potential optimized UGS patterns were identified using a Bayesian network and multi-scenario analysis.

2.2.1. Fine-Scale Land Use Classification

Sentinel-2 imagery was used for land use and land cover classification, while Landsat imagery was adopted for land surface temperature retrieval due to the availability of thermal infrared bands. High-Resolution Network (HBRNet), a deep convolutional neural network, was utilized to classify land use/cover types from Sentinel-2 images, including impervious land, water bodies, green space, and bare land. HRNet employed four multi-branch parallel convolutions to generate maps with a high-to-low-resolution feature, which allowed information to flow smoothly across different branches and enhanced the spatio-temporal accuracy in urban feature learning [23]. Using field surveys and visual interpretation from high-resolution Google Earth images, 8357 samples were collected for model training, validation, and testing. The overall classification and kappa metrics reached 91.77% and 0.87, respectively.

2.2.2. Estimation of Land Surface Temperature (LST)

The Landsat-8 TIRS images were utilized to retrieve LST through the Radiative Transfer Equation (RTE) method [24]. During the period from 1 January 2020 to 31 December 2023, a total of 92 Landsat scenes with cloud coverage lower than 5% over the study area were available, among which 48 scenes were finally selected following a seasonal sampling strategy, with three cloud-free scenes selected for each season in each year to ensure balanced temporal representation. This method is widely used in climate studies due to its mature process, low acquisition costs, and relatively continuous spatio-temporal scales of Landsat series images. RTE incorporates factors of surface emissivity, atmospheric interactions, and sensor characteristics to calculate LST. It mainly included five steps: the conversion of digital numbers to radiance, the transformation of radiance into brightness temperature, the implementation of atmospheric correction, the consideration of surface emissivity, and the calculation of LST, as detailed below:

$$L_{\lambda }=M_{\lambda }\times Q_{cal}+A_{\lambda }$$

(1)

$$B(T_s)={\displaystyle \frac{L_{\lambda }-L_{\mu }-\tau (1-\epsilon )L_d}{\tau ·\epsilon }}$$

(2)

$$\mathit{LST}={\displaystyle \frac{K_2}{\mathit{ln}(\frac{K_1}{B\left(T_s\right)}+1)}}-273.15$$

(3)

where LST is land surface temperature; L_λ is top of atmosphere radiance; $Q_{cal}$ is the digital number value; $M_{\lambda }$ and $A_{\lambda }$ are calibration coefficients of 0.0003342 and 0.1, respectively; B(T_S) is the radiance emitted by a blackbody at a surface temperature of T_S in Kelvin; $L_{\mu }$ and L_d are upwelling and downwelling radiance, respectively; ε is surface emissivity; τ is atmospheric transmittance; M_L and A_L are radiance multiplicative and additive scaling factors from the MTL file; K₁ and K₂ are thermal conversion constants derived from the MTL file. The above steps were conducted in Google Earth Engine (https://code.earthengine.google.com, accessed on 23 April 2024). Although uncertainties in absolute LST values may exist, the physically based RTE method combined with multi-scene seasonal sampling helps ensure the reliability of the relative spatial temperature patterns used in this study.

2.2.3. Quantification of Leaf Area Index (LAI)

An upscaling process that integrated field measurement, UAV multi-spectral images, and Sentinel-2 images was developed to estimate vegetation LAI across the whole region (Figure 3). First, the relationships between field measurement data and vegetation indexes calculated from UAV multi-spectral images (

Table A1. Vegetation indicators used to estimate city-scale LAI.

Indicator	Meaning	Equation
UAV multi-spectral-based indicators
NDVI	Normalized difference vegetation index	$NDVI={\displaystyle \frac{NIR-Red}{NIR+Red}}$
GNDVI	Green normalized difference vegetation index	$GNDVI={\displaystyle \frac{NIR-Green}{NIR+Greeen}}$
OSAVI	Optimized soil adjusted vegetation index	$OASVI={\displaystyle \frac{NIR-Red}{NIR+Red+0.16}}$
NDRE	Normalized difference red-edge index	$NDRE={\displaystyle \frac{NIR-RE}{NIR+RE}}$
LCI	Leaf area chlorophyll index	$LCI={\displaystyle \frac{NIR-RE}{NIR+Red}}$
Sentinel-2 image-based indicators
CIg	Green chlorophyll index	$CIg={\displaystyle \frac{NIR}{Green}}-1$
EVI	Enhanced vegetation index	$EVI={\displaystyle \frac{2.5\ast (NIR-Red)}{NIR+6\ast Red-7.5\ast Blue+1}}$
RVI	Ratio vegetation index	$RVI=NIR/Red$
OSAVI	Optimized soil adjusted vegetation index	$OASVI={\displaystyle \frac{NIR-Red}{NIR+Red+0.16}}$
GNDVI	Green normalized difference vegetation index	$GNDVI={\displaystyle \frac{NIR-Green}{NIR+Greeen}}$
NDVI	Normalized difference vegetation index	$NDVI={\displaystyle \frac{NIR-Red}{NIR+Red}}$
MTCI	MERIS terrestrial chlorophyll index	$MTCI={\displaystyle \frac{(Rededge2-Rededge1)}{(Rededge1-Red)}}$
CI_rededge1	Red-edge chlorophyll index-1	$CI\_rededge1={\displaystyle \frac{Rededge1}{Green}}-1$
CI_rededge2	Red-edge chlorophyll index-2	$CI\_rededge2={\displaystyle \frac{Rededge2}{Green}}-1$
CI_rededge3	Red-edge chlorophyll index-3	$CI\_rededge3={\displaystyle \frac{Rededge3}{Green}}-1$
NDRE_rededge1	Normalized difference red-edge index-1	$NDRE\_reedge1={\displaystyle \frac{NIR-Rededge1}{NIR+Rededge1}}$
NDRE_rededge2	Normalized difference red-edge index-2	$NDRE\_reedge2={\displaystyle \frac{NIR-Rededge2}{NIR+Rededge2}}$
NDRE_rededge3	Normalized difference red-edge index-3	$NDRE\_reedge3={\displaystyle \frac{NIR-Rededge3}{NIR+Rededge3}}$
LCI_rededge1	Red-edge leaf area chlorophyll index-1	$LCI\_reedge1={\displaystyle \frac{NIR-Rededge1}{NIR+Red}}$
LCI_rededge2	Red-edge leaf area chlorophyll index-2	$LCI\_reedge1={\displaystyle \frac{NIR-Rededge2}{NIR+Red}}$
LCI_rededge3	Red-edge leaf area chlorophyll index-3	$LCI\_reedge1={\displaystyle \frac{NIR-Rededge3}{NIR+Red}}$

) were established through the XGBoost algorithm, which was used to calculate the LAI map in each sample region. Then, a sample library was established, consisting of the corresponding value of vegetation indicators from Sentinel-2 images (Table A1) and the simulation results of LAI in each grid (10 m × 10 m) within the sample regions. Finally, simulation models were trained through the XGBoost algorithm to predict the LAI value of UGS across the whole city. This upscaling method could effectively reduce sampling costs and improve simulation accuracy [25,26]. The establishment of a sample library also allowed for the continuous simulation of UGS LAI for the whole region. Field LAI measurements were conducted next in different seasons using an LAI-2200C plant canopy analyzer (LI-COR Inc., Lincoln, Nebraska) (first: 13 January 2022 to 15 January 2022, second: 15 April 2023 to 17 April 2023, third: 1 August 2023 to 3 August 2023, fourth: 10 November 2023 to 12 November 2023). There were a total of 2032 ground-measured LAI values in this study. In this study, vegetation LAI was estimated for each season (spring, summer, autumn, and winter) in each year from 2020 to 2023, and the corresponding seasonal LAI values were extracted for each UGS sample and used in the subsequent analyses. Compared to NDVI, LAI was utilized to represent the 3D canopy structure and for robustness assessment.

2.2.4. Simulation of Ventilation Environment

The urban ventilation environment significantly impacts airflow patterns and the formation of ventilation corridors, which directly introduce clean and fresh air into urban areas, dilute and discharge exhaust and waste heat, and further mitigate UHIs and air pollution [27]. In the main built-up area, a ventilation environment was mainly determined by the distribution of dense buildings, and ventilation potential was simulated by an indicator of CVCI (comprehensive ventilation cost index) proposed by Ying, Wang [15], which integrated multiple 3D building morphological attributes to characterize the relative ventilation resistance at the city scale. Thus, the 3D features were reflected through the combined effect of building height, volume, and façade on airflow rather than explicit wind-speed simulation. Compared to other traditional methods, CVCI could comprehensively characterize the resistance of building 3D structures to ventilation from multiple aspects and dynamically observe the difference in ventilation characteristics under different wind directions at the city scale. CVCI was calculated from the three aspects of building height (AH), building volume (3DSC), and building façade (FAI), with detailed formulas as follows:

$$AH={\sum }_{i=1}^N{H}_i/N$$

(4)

$$3DSC={\sum }_{i=1}^N\left(A_i\times H_i\right)/(A\times H_c)$$

(5)

$$FAI=A_{facet}/A_{plane}$$

(6)

$$CVCI=w_f\times FAI+w_c\times 3DSC+w_h\times AH$$

(7)

where AH is the average height of buildings in the grid; H_i is the height of building i; N is the total number of buildings in the grid; 3DSC is the 3D space congestion; A_i is the floor area of building i; Hc is the 95% value of building height across the whole region; FAI is the frontal area index; A_facet is the total area of building facets facing the wind direction; A_plane is the grid area; w_f, w_c, and w_h are calculated using the entropy method. The range of CVCI was 0–1, and the higher the CVCI value, the poorer the ventilation potential (Figure 4).

2.3. Methodology

2.3.1. Quantifying the UGS Cooling Effect

To avoid the interactions of cooling effects produced by other green spaces and water bodies, the samples were selected based on the following principles: (1) select green spaces without water bodies; (2) select green spaces more than 300 m away from other green spaces or water bodies; (3) select samples with different areas and shapes [28]. Under these strict criteria, only 16 UGS samples across the entire study area satisfied the requirements for the independent cooling assessment and remained spatially and structurally consistent during the period 2020–2023, as confirmed by high-resolution Google Earth imagery. Rather than statistically representing the city-wide UGS inventory, the analysis was restricted to UGS patches that could be independently evaluated for cooling effects.

The selected samples span a range of area sizes, shape complexities, and spatial contexts relevant to the mechanism-based analysis (Figure 1, Figure A2, and

Table A2. The characteristics of the selected urban green space samples.

Sample ID	Area (ha)	Length (km)	Type	Location
1	0.66	0.44	Park	Urban core
2	0.98	0.58	Park	Inner city
3	1.07	0.54	Park	Inner city
4	3.22	2.76	Residential green space	Urban core
5	1.45	1.16	Residential green space	Urban core
6	1.06	0.58	Roadside green space	Urban core
7	1.46	1.08	Park	Urban core
8	0.42	0.3	Residential green space	Urban core
9	7.08	5.08	Residential green space	Inner city
10	0.92	0.52	Park	Urban core
11	1.86	1.4	Institutional green space	Inner city
12	1.47	0.88	Park	Inner city
13	2.2	1.28	Roadside green space	Inner city
14	5.44	1.68	Park	Inner city
15	4.08	2.12	Park	Inner city
16	3.74	1.5	Park	Inner city

). The cooling effect of each UGS was quantified using a buffer analysis, in which concentric buffer zones were generated outward from the UGS boundary at 30 m intervals. Cooling effects were quantified based on relative LST differences along buffer gradients rather than absolute temperature values, allowing the analysis to focus on spatial temperature contrasts under consistent observational conditions. The median distance of each buffer zone was used as the x-axis, and the mean LST within each buffer zone was used as the y-axis. The cooling range was defined as the distance between the UGS edge and the first turning point of the fitted curve, while the cooling intensity was defined as the temperature difference between the UGS edge and the first turning point. According to the results of Qi, Zhao [29], phenological seasons were divided based on time-series observations of climate stations into spring, summer, autumn, and winter as March to May, June to August, September to October, and November to March, respectively.

2.3.2. Quantifying the Cooling-Related Patterns of UGS

Based on the formation mechanism of the UGS cooling effect, conclusions of related studies, and data availability, 9 potential factors (

Table 1. Potential associated factors of cooling effect.

Type	Factor	Code	Calculation Method
Inner characteristics	Patch area	AREA
	Patch shape index	SHAPE	$SHAPE={\displaystyle \frac{edge}{2\times \sqrt{\pi \times area}}}$
	Normalized difference vegetation	NDVI
	Leaf area index	LAI	Estimated following the methodology described in Section 2.2.2
Surrounding environment	Comprehensive ventilation cost index	CVCI	Estimated following the methodology described in Section 2.2.3
Background conditions	Average monthly precipitation	precip	Observation data from monitoring station
	Average monthly temperature	temper	Observation data from monitoring station
	Maximum wind speed	wind	Observation data from monitoring station
	Average monthly air pressure	pressure	Observation data from monitoring station

* edge—the edge length of UGS patch; area—UGS patch area.

) were selected from the three aspects of inner characteristics, the surrounding environment, and background conditions [4,30]. The UGS inner characteristics of area, shape, and vegetation abundance (i.e., NDVI and LAI) determined the supply level of UGS cooling service, while surrounding ventilation estimated by CVCI would significantly impact the flow and spillover of UGS cooling service [8,21]. Besides that, the generation and flow of the UGS cooling service were also impacted by background environment [30]. Specifically, only meteorological factors were selected in the aspect of background conditions due to the unavailability of socioeconomic data on a daily or monthly scale. Socioeconomic variables were not explicitly included in the seasonal analyses because they remain largely invariant within a year and do not capture short-term or seasonal dynamics of the UGS cooling effect. Including such time-invariant variables across repeated temporal observations may introduce pseudo-replication and obscure the interpretation of seasonal drivers. Moreover, their long-term influences are indirectly reflected through relatively stable spatial characteristics of land use, urban form, and vegetation structure considered in this study. Among the available factors, correlation analysis and principal component analysis were first applied to select the factors, and 9 factors were selected with significant correlations with cooling intensity or range and low multicollinearity among factors.

Given the repeated monthly observations within each UGS (92 time points per park, from 2020 to 2023), a hierarchical longitudinal analytical framework was adopted. Generalized additive mixed-effects models (GAMMs) were implemented, with UGS identity included as a random intercept to account for within-UGS temporal dependence and between-UGS heterogeneity. GAMM is a semi-parametric extension of the generalized linear model, which has been widely applied in ecological research due to its interpretability, feasibility, and regularization in dealing with non-linear relationships [9]. Compared to other methods, GAMM strikes a nice balance between interpretable-but-biased linear models and the extremely flexible but “black-box” learning algorithms [31]. Smooth terms were applied to key continuous predictors (e.g., NDVI and time), with smoothing parameters constrained (k ≤ 6) and estimated using restricted maximum likelihood (REML) to avoid overfitting.

The above steps were conducted using R packages of “mgcv” and “stats”.

2.3.3. Identifying Optimized Patterns for the UGS Cooling Effect

A Bayesian network (BN) was employed as a probabilistic configuration screening tool to identify UGS metric combinations associated with high cooling intensity and range, rather than for causal inference [32,33]. Network structure was constrained based on GAMM-identified drivers and physical understanding of the cooling processes. BNs are widely used in statistical modeling and machine learning of complex systems due to their advantages in the integration of multi-source data, transparency in the model structure, uncertainty handling in data sources, and probabilistic reasoning in decision-making. There were two main parts in BNs: a directed acyclic graph (DAG) depicting variable state and the corresponding probability distribution in each node and the relationships between each node; and the conditional probability table (CPT) representing the strength of relationships among variables. Based on the results of Section 2.3.2, the BN was first constructed using the identified critical factors as the parent nodes, and cooling intensity and range as the target nodes. Each node was classified into intervals by data frequency (

Table 2. List of variables used in the Bayesian network.

Code	Description	Unit	State Code
			L2	L1	M	H1	H2
AREA	Patch area	ha	≤0.98	(0.98, 1.45]	(1.45, 1.86]	(1.86, 3.74]	>3.74
SHAPE	Patch shape index	-	≤1.36	(1.36, 1.8]	(1.8, 2.16]	(2.16, 2.57]	>2.57
CVCI	Comprehensive ventilation cost index	-	≤0.008	(0.01, 0.02]	(0.02, 0.03]	(0.03, 0.05]	>0.05
TEMP	Average monthly temperature	°C	≤0.9	(0.9, 8.9]	(8.9, 14.1]	(14.1, 22.1]	>22.1
WIND	Maximum wind speed	m/s	≤3.7	(3.7, 4.3]	(4.3, 4.8]	(4.8, 5.4]	>5.4
CI	Cooling intensity	°C	≤0.42	(0.42, 0.91]	(0.91, 1.53]	(1.53, 2.33]	>2.33
CR	Cooling range	m	≤150	(150, 210]	(210, 270]	(270, 320]	>320

). Then BN was calibrated and validated through the k-fold cross-validation method, which randomly splits the dataset into k subsets and uses each subset to validate the model fitted on the remaining k-1 subsets. The larger the classification error, the lower the predictive accuracy of the target nodes. Finally, scenario analysis was employed to identify the optimized UGS pattern that provided the desired outcomes (i.e., the high probability of the highest class of cooling intensity and range). The marginal distribution of the parent node was randomly changed to capture all possible combinations in critical variables and further estimate the change in the probability distributions of the target nodes. In this study, 10,000 scenarios with different configurations of input nodes were generated to identify the optimized scenario. The above steps were conducted using R packages of “bnlearn” and “gRain”.

3. Results

3.1. Temporal Changes in Cooling Intensity and Their Associated Factors

The results of Spearman’s correlation analysis (Figure A3) suggest that seasonal variations in cooling intensity are mainly associated with meteorological conditions, while the cooling range did not show significant seasonal differences (Figure A4). Given the repeated temporal observations within each UGS and the limited number of independent spatial units (n = 16), a generalized additive mixed-effects model (GAMM) was employed to explore non-linear associations between cooling intensity and key meteorological variables, with UGS identity treated as a random intercept to account for within-UGS temporal dependence and between-UGS heterogeneity. The GAMM results (Figure 5) indicate that air temperature and maximum wind speed are significantly associated with temporal variations in UGS cooling intensity. Cooling intensity generally increased with mean air temperature, remaining relatively stable below approximately 5 °C, increasing rapidly between 5 and 19 °C, and showing a slower increase at higher temperatures. Notably, the negative cooling intensity values at lower temperatures suggest warming signals of UGS under cold conditions. Cooling intensity also increased with maximum wind speed, exhibiting an approximately linear relationship.

3.2. Seasonal Patterns of the Cooling Effect and Associated Factors

Based on the results of Spearman’s correlation analysis (Figure A3), season-specific GAMMs were constructed—with UGS identity included as a random intercept—to characterize spatial variation patterns under different seasonal contexts. The results of the GAMM models (Figure 6 and Figure 7) show that the identified influencing factors are strongly associated with the spatial variance of the UGS cooling effect. Specifically, the simulation accuracy of the cooling range was generally lower than that of cooling intensity, partly due to the medium-resolution LST images.

The fitted curves between each factor and the cooling effect from the GAMM suggest that most factors had non-linear relationships with cooling intensity and range among the four seasons, except for the linear relationships between UGS shape and cooling intensity in winter. As indicated in Figure 6, cooling intensity in spring and autumn was mainly associated with three factors of SHAPE, NDVI, and CVCI, displaying general positive relationships with the former two factors and a complex curved relationship with the latter. In summer, cooling intensity was mainly associated with AREA and CVCI, generally increasing with UGS AREA but showing fluctuating patterns with CVCI. Cooling intensity in winter was mainly associated with SHAPE, while higher cooling intensity occurred at the higher level of SHAPE.

As indicated in Figure 7, the cooling ranges are mainly associated with AREA, SHAPE, and CVCI in all four seasons, while their relationships vary across different seasons. The cooling range generally exhibits an inverted U-shape relationship with AREA, except in autumn, with the turning point around 3.8 ha in spring and winter and 4.5 ha in summer. For UGS SHAPE, a U-shape relationship is observed between it and the cooling range across the four seasons. Across the four seasons, CVCI is also strongly associated with cooling range, exhibiting a fluctuating relationship, which shows different threshold values across different seasons.

3.3. Optimized UGS Patterns for Improving the Cooling Effect

BN analysis was conducted to explore planning-oriented configuration patterns under observed climatic conditions, rather than to infer causal relationships. As indicated by the 10-fold cross-validation, the average classification error of the established BN model in Figure 8 is 28% and 34% for cooling intensity (CI) and range (CR), respectively. Despite the fact that prediction error is non-negligible, the BN model was mainly used to analyze probabilistic relationships between significant factors and cooling-effect indicators rather than to provide precise quantitative predictions. Generated scenario results were reorganized to sort the first 100 scenarios that provide a higher probability of the desired outcomes (i.e., the highest state of CI, the highest state of CR, and a win–win situation of CI and CR). The results (Figure 9) show that the highest level of CI was derived under the condition that combines the highest level of UGS area and the lowest or higher levels of CVCI. Specifically, medium levels of UGS area also had a higher possibility for the highest level of CI. On the other hand, the highest level of CR was more likely to be achieved under the environment with a higher level of UGS area and medium levels of CVCI and UGS shape index. As indicated in Figure 9c, the “win–win” solutions of CI and CR are more likely to occur under the condition of the highest level of UGS area, the lowest level of CVCI, and medium levels of UGS shape index. Thus, larger UGSs (i.e., >3.74 ha) with a moderate shape complexity (i.e., 1.8–2.16) and good ventilation conditions (i.e., <0.008) were more prone to generating a higher cooling intensity and range, which has great importance in maximizing the cooling benefit of UGSs and further alleviating UHIs.

4. Discussion

Existing studies suggest that the cooling effect of UGS is impacted not only by inner landscape characteristics, such as area, shape, and vegetation state, but also by external factors of the surrounding environment, such as spatial locations, land cover features, and building density [7,8]. How to separate the impacts of surrounding configuration is critical to accurately characterize the associated processes of the UGS cooling effect [34]. In this study, we utilized a unique sampling methodology to select appropriate samples for the quantification and pattern analysis of the UGS cooling effect, which can effectively avoid the interactive impacts of other UGSs and water bodies [28]. However, this also generates one important problem, that is, the small sample size at specific times, especially in cities with continuous green spaces and water bodies. In this study, we employed multi-temporal observations at selected UGS sites to enhance the stability of detected cooling signals under different climatic conditions. Although only 16 spatially independent UGS patches were available, each was repeatedly observed monthly (nearly 92 time points), yielding a longitudinal dataset of 1472 observations. Accordingly, inference is based on hierarchical repeated-measures analysis rather than cross-sectional regression.

In addition, although only 16 spatial samples were available, repeated observations were obtained from remote-sensing images and meteorological monitoring data at multiple time points. Moreover, this study further explores the impacts of 3D features on the UGS cooling effect, considering the rapid urban growth in both vertical and horizontal directions [35].

4.1. Significant Differences Exist in the UGS Cooling Effect Between Different Seasons

Despite the low temperature in the cold–arid and semi-arid climate zones, an obvious UHI effect still exists in the northwestern cities of China [20]. Meanwhile, the harsh climate and limited water–land resources increase the costs of constructing green spaces and water bodies, which enhances the necessity of selecting appropriate locations and optimizing UGS structure within cities [36]. Evidence from this study (Figure 5, Figure 6 and Figure 7) also confirms that the cooling effect of UGS cannot be solely attributed to individual factors, but is a comprehensive result of UGS composition and configuration characteristics, vegetation state, surrounding ventilation, and background meteorological conditions. Specifically, the impacts of different factors are reflected in different aspects. As indicated by the correlation test in Figure 3 and the high R² of the GAMM model in Figure 5, the temporal change in cooling effect is mainly associated with variations in meteorological factors and vegetation state. This has also been demonstrated by previous studies [30]. In contrast, the spatial variation in cooling effects among different samples is mainly associated with the inner characteristics of area, shape, and vegetation state and surrounding ventilation, as indicated in Figure 6 and Figure 7. This is partly due to the approach to data acquisition, in which the strict sampling principles of UGSs make it difficult to acquire enough samples at a specific time point, while UGSs would not largely change in a short time period, especially in well-developed urban areas [37]. In this study, multi-temporal observations were employed to enhance the stability of detected cooling signals under climatic conditions. This approach effectively leverages the accumulated extensive remote-sensing images and field monitoring data to address the limitations associated with scarce spatial samples collected at a single time point [38]. Based on the temporal variation in vegetation status, climate conditions, and the associated cooling effect within UGS samples over time, this method collected repeated observations to support an exploratory analysis of the cooling effect. However, this approach can obscure the impacts of sample characteristics, especially the area and shape of the samples, as these metrics do not largely change over a short time period [39]. This also contributes to the lower impacts of UGS area and shape compared to other studies [7]. The critical role of meteorological factors has also been neglected in previous studies [6]. This is mainly due to the difficulty in acquiring spatial maps of meteorological factors in UGS samples at high resolution, as well as the difficulty in reflecting the high heterogeneity of the urban microclimate by in situ observations.

Local meteorological conditions can greatly impact the cooling efficiency of UGS [40,41]. With the growing ambient temperature, the saturated vapor pressure in stomata of urban vegetation and the vapor pressure deficit between stomata and air would also increase, which can promote the transpiration approach and increase the UGS cooling effect [40]. However, continuous or extremely high temperatures might not be beneficial to vegetation growth and transpiration, and even cause irreversible damage to leaves after a certain limit, especially for species with poor thermal adaptation [42]. This high-temperature condition would result in nearly closed stomata of vegetation to avoid excessive water loss and a decrease in transpiration efficiency and the cooling effect [43]. Our results (Figure 3) also confront this point with the slower increase rate of cooling intensity at higher temperatures. On the other hand, the impacts of wind speed are more complicated and context-specific [30]. Theoretically speaking, the increase in wind speed can promote air convection, increase the amount of heat dissipated through sensible heat loss, and further promote vegetation evaporation, which is beneficial for enhancing the cooling effect of UGS [44]. However, large amounts of wind would directly reduce leaf surface temperature and stomatal opening, which would reduce transpiration and cooling capacity [45]. As a typical valley city surrounded by mountains, Lanzhou always has small amounts of wind throughout the four seasons [29]. This relatively stagnant environment enhances the impact of maximum wind speed on air convection—even beyond the impact of mean wind speed—and results in the linear positive relationship between UGS cooling intensity and maximum wind speed.

4.2. Urban 3D Structure Plays a Critical Role in the UGS Cooling Effect

Numerous studies have identified key drivers of the UGS cooling effect, suggesting that factors such as patch size, spatial pattern, vegetation species and structure, background climate, and socioeconomic context all play important roles [8,46]. Among these, UGS patch size has been repeatedly highlighted as a dominant factor, exhibiting a positive non-linear relationship with cooling capacity—that is, the cooling effect generally increases with larger patch sizes [7,8]. Consistent with these theoretical expectations, our results demonstrate that UGS area critically influences the cooling range across all four seasons and has a dominant effect on cooling intensity, specifically in summer (Figure 7 and Figure 8). This seasonal variation aligns with prior work focusing predominantly on summer when urban heat island (UHI) effects are strongest; in other seasons, additional factors likely dilute the impact of UGS area. Moreover, our findings further nuance the theoretical understanding by showing that UGS can contribute to localized warming in winter (Figure 4), supporting the suggestion of Deilami, Kamruzzaman [47] that vegetation can trap radiant heat and reduce air movement during colder months. These insights caution against a simple “more is better” approach to UGS expansion: although larger UGSs generally enhance cooling, their marginal efficiency declines beyond a certain size, indicating the existence of an optimal patch size [48]. Our data identify a turning point around 4 hectares across different seasons (Figure 7 and Figure 8), aligning with but also adding specificity to previous findings that optimal UGS sizes vary significantly between cities—from 0.5–1 ha in Beijing, Tianjin, and Singapore to 4.5–6.5 ha in Fuzhou and Leipzig, and up to 20 ha or more in large Chinese cities [10,49]. These inconsistencies highlight the modifying roles of local climate conditions and socioeconomic development and emphasize the strategic importance of small UGS patches in densely populated urban environments [50].

The impact of UGS shape on the cooling effect remains uncertain and controversial. For example, some studies suggest that UGSs with a regular shape can provide a higher cooling effect, although there are also studies holding the opposite view, while other studies propose that patch shape has almost no impact on cooling effect [4]. Our results suggest the critical role of shape irregularity in altering cooling intensity, with a general positive relationship, except in summer, indicating that UGSs with more complex shapes have a higher cooling intensity. This is mainly due to the increasing heat transmission between UGS and the surrounding environment, along with the increasing UGS edge [51]. However, the cooling range would first decrease at lower levels of UGS shape, indicating that a regular shape is more suitable for small UGSs to increase their cooling range. Jaganmohan, Knapp [52] also suggest that increasing shape complexity of smaller UGSs has a negative impact on their cooling effect. Thus, increasing shape regularity for small UGSs and shape complexity for large UGSs is beneficial for both enhancing their cooling intensity and increasing their cooling range.

Vegetation regulates the urban microclimate mainly in three ways, including altering the air movement and heat exchange, absorbing latent heat through evapotranspiration by foliage and moist soil in planting beds, and intercepting incoming solar radiation through shading provided by tree canopies [53]. Thus, vegetation coverage plays a critical role in altering the cooling intensity and range of UGSs with a positive impact, which is also suggested by Figure 6. However, this impact has certain limitations. Wang, Ren [54] suggest that the cooling intensity of trees at the cluster scale may not increase further with increasing tree canopy when it surpasses 40%. Calculated from vegetation-related spectral bands, NDVI and LAI can effectively reflect vegetation state, such as leaf health, canopy closure, and vegetation abundance [55]. Bartesaghi-Koc, Osmond [56] also suggest that UGS with higher NDVI can produce a higher nocturnal warming effect than expected. Our results only suggest the critical role of NDVI on cooling intensity in summer and autumn, while its impact on cooling range may have been surpassed or masked by other factors. Thus, the impact of vegetation abundance would be enhanced in the period of vegetation growth or decay (i.e., spring and autumn), but would be weakened in the period of relatively stable vegetation (i.e., summer and winter). Moreover, the overall cooling intensity of trees has been suggested as 1.35 °C higher than that of shrubs [54]. Compared to NDVI, LAI is more sensitive to trees and shrubs, while LAI of shrubs can be similar or even surpass the LAI of trees [57,58]. Thus, LAI has a lower impact on the cooling effect than NDVI, especially in arid and semi-arid zones with thicker and narrower leaves to hold more water and improve water-use efficiency [59].

Wind can largely alter the air movement and further impact the cooling intensity and range of UGS [53]. It is difficult to estimate the contribution of wind to the UGS cooling effect, as it largely depends on local conditions (e.g., street direction, building height, etc.) and needs micro-scale data of the distribution of open space and buildings [60]. In this study, we tried to tackle this problem by quantifying urban ventilation through an index of CVCI that comprehensively considers the impacts of building height, windward façade area, and 3D space congestion, which displays higher values in environments with poor ventilation and air advection [15]. Our results suggest a significant non-linear relationship between CVCI and the cooling effect (i.e., intensity and range) across the four seasons, indicating that a moderate level of CVCI is more conducive for enhancing and increasing the range of the UGS cooling effect. The vertical growth of buildings inevitably increases the roughness of underlying urban surface and further reduces air circulation efficiency, which would promote the formation of UHIs [61]. On the other hand, a high-ventilation environment is unfavorable for the retention of cold air generated by UGS, and also impractical for urban development due to the high demand for land resources [62,63]. Thus, a moderate level of ventilation in the environment is more suitable for enhancing the UGS cooling effect and mitigating UHIs, which should be paid more attention to in urban innovation and planning.

4.3. Implications for Urban Planning and UGS Management

Urban green spaces, such as parks, greenbelts, and square greenery, have been widely considered as a nature-based solution to improve thermal comfort and resident living quality. Previous studies have proposed inconsistent and controversial conclusions for the driving mechanisms of the UGS cooling effect, which makes it difficult to formulate reasonable suggestions for urban management and planning. These controversial results can be explained by the climate background of specific case areas, data resolution, and spatial patterns of case cities [8]. This study has integrated the impacts of meteorological factors and explored the determining mechanism of the UGS cooling effect from two aspects of temporal change and spatial variation.

The results suggest that the temporal change in UGS cooling intensity is mainly driven by meteorological conditions, among which average temperature and maximum wind speed have the largest impacts with positive relationships. Thus, appropriately increasing UGS in areas with higher temperature and wind speed has a higher efficiency in mitigating UHIs and improves thermal comfort for highly populated areas, especially in upwind areas. From the spatial perspective, the UGS cooling effect is comprehensively compacted by size, shape, NDVI of UGS, and surrounding ventilation, which show significant non-linear relationships and obvious differences among the four seasons. For large UGS parks, keeping their shape index at intervals of 1.8–2.16 and keeping surrounding ventilation at the lowest level of resistance is helpful to achieve the win–win between cooling intensity and range, as indicated in Figure 9. Besides that, reasonable increases in UGSs and their shape regularity can also be beneficial for mitigating thermal discomfort, especially in high-density residential districts. This study also confirmed the importance of urban ventilation for the cooling effect; keeping the ventilation of the UGS’ surrounding environment at a medium level (i.e., building density of 0.374–0.692 and sky-view factor of 0.25–0.50) can effectively improve the cooling intensity and range. Moreover, urban management should pay more attention to establishing reasonable ventilation systems, and measures such as improving UGS quality in key nodes and securing ventilation corridors should be encouraged to promote the flow of cold air and enhance the spillover effect of the UGS cooling service.

4.4. Limitations and Future Research Directions

There are some limitations of this study that need to be addressed in future research. First, urban ventilation is impacted by not only building morphology but also meteorological conditions, such as local temperature, solar radiation, and the atmospheric environment [17]. This study only considers building morphology and distribution and presents the probability of ventilation rather than actual wind speed. More research is needed to integrate morphological factors into the quantification of ventilation, and strict aerodynamic simulations to obtain actual convection coefficients.

Second, the application of satellite-derived thermal indicators introduces certain limitations in highly heterogeneous urban environments. The limited resolution of the Landsat 8 TIRS images restricts the direct application of the findings to fine-scale urban management. Moreover, the cooling effect of UGS was quantified based on land surface temperature rather than near-surface air temperature, which is more directly related to human thermal perception. While LST is well-suited for capturing surface energy balance and fine-scale spatial heterogeneity of cooling effects, acquiring near-surface air temperature data with sufficient spatial resolution and temporal continuity at the intra-urban scale remains challenging [30]. Station-based observations are spatially sparse, and available gridded or modeled air-temperature products often lack the spatial detail required to resolve localized cooling effects of individual green spaces. Accordingly, near-surface air temperature was treated as a background climate variable in this study. Future research could integrate dense in situ measurements, mobile observations, or microclimate simulations to better link surface cooling processes with human thermal comfort.

Third, the heat transfer process among UGS, the atmosphere, and the surrounding environment is also impacted by other factors, such as vegetation species or management options. In particular, irrigation practices—which can substantially enhance evapotranspiration and cooling effects—were not explicitly considered due to the lack of spatially explicit and temporally continuous data at the city scale. Nevertheless, the effects of irrigation are implicitly reflected in the observed thermal and vegetation responses, therefore contributing to model uncertainty rather than invalidating the proposed analytical framework. Future research should seek to integrate irrigation-related information or suitable proxy indicators to further refine the quantification of UGS cooling processes.

5. Conclusions

This study has quantified cooling-related patterns from three aspects of UGS inner characteristics, surrounding building ventilation, and background meteorology, and further identified the optimized UGS patterns for summer cooling effects using a Bayesian network and scenario analysis. The results suggest that the temporal change in cooling intensity is mainly associated with background meteorology, especially air temperature and maximum wind speed, displaying positive relationships. In different seasons, the spatial variation in cooling intensity is mostly associated with the UGS inner characteristics of area and shape, while vegetation state of NDVI shows a higher association in spring and autumn. Meanwhile, the cooling range is mainly associated with UGS area and shape, with similar non-linear relationships but different thresholds across different seasons, while it can also be altered by surrounding ventilation with fluctuating relationships. The win–win situation of cooling intensity and range mostly occurs in UGSs with larger areas, lower shape irregularity, and the lowest ventilation resistance. Optimization of urban ventilation systems should be paid more attention in future urban development to mitigate urban heat islands.