Exploring the Role of Urban Green Spaces in Regulating Thermal Environments: Comparative Insights from Seoul and Busan, South Korea

Abstract

Urban heat islands are intensifying under the dual pressures of global climate change and rapid urbanization, posing serious challenges to ecological sustainability and human well-being. Among the factors influencing urban thermal environments, vegetation and green spaces play a critical role in mitigating heat accumulation through canopy cover, evapotranspiration, and ecological connectivity. In this study, a comparative analysis of Seoul and Busan—two representative metropolitan areas in South Korea—was conducted using land surface temperature (LST) data derived from Landsat 8 and a set of multi-source spatial indicators. The nonlinear effects and interactions among built environment, socio-economic, and ecological variables were quantified using the Extreme Gradient Boosting (XGBoost) model in conjunction with Shapley Additive Explanations (SHAP). Results demonstrate that vegetation, as indicated by the Normalized Difference Vegetation Index (NDVI), consistently exerts significant cooling effects, with a pronounced threshold effect observed when NDVI values exceed 0.6. Furthermore, synergistic interactions between NDVI and surface water availability, measured by the Normalized Difference Water Index (NDWI), substantially enhance ecological cooling capacity. In contrast, areas with high building and population densities, particularly those at lower elevations, are associated with increased LST. These findings underscore the essential role of green infrastructure in regulating urban thermal environments and provide empirical support for ecological conservation, urban greening strategies, and climate-resilient urban planning. Strengthening vegetation cover, enhancing ecological corridors, and integrating greening policies across spatial scales are vital for mitigating urban heat and improving climate resilience in rapidly urbanizing regions.

1. Introduction

In recent years, with the acceleration of global climate change and urbanization, the urban heat island (UHI) effect has become one of the most prominent environmental challenges facing cities worldwide [1,2,3]. Numerous studies have demonstrated that UHI not only leads to significantly higher land surface temperatures (LST) in urban areas compared to surrounding suburbs [4], but also triggers a series of risks related to the ecological environment and public health, such as heat-related mortality, increased energy consumption, and deteriorating air quality [5,6]. According to the United Nations’ “World Urbanization Prospects 2022,” the global urban population has reached 57% of the total, and is projected to rise to 68% by 2050 [7]. Driven by high-density and high-intensity human activities, the scope and severity of UHI impacts continue to expand [8]. Thus, in-depth exploration of the influencing factors of urban thermal environments is of great significance for guiding urban planning, achieving sustainable urban development, and enhancing residents’ well-being [9,10].

The UHI effect refers to the phenomenon in which land surface temperature in urban areas is generally higher than that in surrounding suburban areas as urbanization accelerates [11,12]. Its fundamental causes lie in the dramatic changes in urban built environment structures, land use patterns, and modes of human activity [13]. A substantial body of research indicates that UHI not only alters the surface energy balance, but also significantly increases the frequency of extreme high-temperature events, thereby worsening urban microclimates, increasing energy consumption, and elevating health risks for residents [14,15,16].

In recent years, research on the driving factors of urban thermal environments has continued to deepen. Zhou et al. (2021) combined field observations and deep learning to reveal the compound effects of urban greenery and building proportions on the thermal environment, emphasizing the crucial role of three-dimensional urban structural parameters in regulating urban temperatures [17]. Yang et al. (2021), using numerical simulations and correlation analysis, systematically assessed the influence of multi-dimensional urban morphological indicators such as building density and sky view factor on LST, and highlighted the dominant role of three-dimensional urban form in heat regulation [18]. Zhang et al. (2021), through spatiotemporal evolution analysis, demonstrated that urban thermal patterns are driven by multiple factors, including administrative boundary adjustments, transportation network layout, industrial relocation, and ecological construction, and stressed the necessity of integrating multi-source factors for systematic identification of thermal environment evolution mechanisms [19]. In addition, Zhao et al. (2022), using geostatistical methods and multi-source data fusion, analyzed the spatiotemporal heterogeneity of LST and identified key natural and anthropogenic factors via the geographical detector model [20]. Their findings further confirmed that the synergistic effects of land use and human activities significantly shape urban thermal patterns [21]. Despite the increasing depth of related studies, most focus on single cities or regions, and systematic multi-city comparative studies remain scarce. As a result, the commonalities and heterogeneities of urban thermal environment mechanisms under the same policy framework have not been fully revealed. Therefore, conducting quantitative multi-city comparative research is of vital academic and practical significance for understanding the commonalities and individual differences in urban thermal environment evolution under similar contexts and for developing more targeted climate adaptation strategies [22].

Researchers have employed various statistical methods to identify and determine the driving factors influencing LST. Traditional regression approaches, such as ordinary least squares (OLS), multiple linear regression (MLR), and geographically weighted regression (GWR), have been widely used in LST studies [23,24,25]. However, these correlation and multiple regression methods struggle to comprehensively reveal the nonlinear effects and potential interactions among various factors, limiting the understanding of the mechanisms underlying urban thermal environment formation [26]. In recent years, machine learning—particularly ensemble learning models such as XGBoost—has gradually become an important method for identifying the driving factors of urban thermal environments, owing to its superior nonlinear modeling capabilities and feature extraction strengths [27,28]. XGBoost can efficiently handle complex relationships among high-dimensional, highly correlated variables and significantly improve model predictive accuracy [29]. However, its “black box” nature partially hinders the deep understanding of variable contributions and interaction pathways among urban management and decision-making stakeholders [30,31]. As an emerging model interpretability tool, SHAP (Shapley Additive Explanations) provides global and local interpretations of model predictions through game theory-based feature attribution, thereby overcoming the explanatory bottleneck of traditional statistical and single machine learning models regarding variable interactions and causal mechanisms [32]. The integration of XGBoost and SHAP not only breaks through previous limitations in variable interaction identification and causal mechanism explanation, but also provides robust theoretical and methodological support for multi-factor complex system analysis in environmental science [33].

Despite the increasing depth of related research, most studies have focused on individual cities or specific regions, and systematic multi-city comparative analyses remain scarce. Consequently, the commonalities and heterogeneities in urban thermal environment mechanisms under the same policy framework have not been fully revealed. To address this gap, South Korea serves as a typical representative of rapid urbanization and economic development in East Asia. Among its cities, Seoul and Busan exemplify two distinct urban types: a megacity in the capital metropolitan area and a coastal gateway city, respectively. These cities possess highly diverse climatic conditions, topography, and urban spatial structures.

Against this backdrop, we take Seoul and Busan—the two major cities of South Korea—as case studies. By integrating Landsat 8 remotely sensed LST data with multi-source spatial datasets, and applying multiple modeling approaches including MLR, random forest (RF), LightGBM, and XGBoost, we employ the SHAP method to comparatively analyze the primary driving factors of the urban thermal environment, while revealing their nonlinear effects and interaction mechanisms. This study aims to achieve the following objectives: (1) to systematically analyze the key driving factors and spatial differences of urban thermal environments in the two Korean cities; (2) to elucidate the nonlinear main effects and interaction mechanisms of different factors on urban heat environments; (3) to propose scientifically informed recommendations for thermal environment regulation and spatial optimization tailored to different types of cities.

The main innovations of this study are as follows: (1) the integration of XGBoost and SHAP methods to achieve precise identification and interpretable analysis of complex urban thermal environment driving factors; (2) the enrichment and extension of the theoretical framework of urban thermal environment driving mechanisms based on multi-city comparative analysis; (3) the provision of data support and scientific guidance for the formulation of urban thermal environment regulation policies. The findings offer robust scientific evidence for urban thermal environment management and planning in complex urban systems in South Korea and other regions with similar urban characteristics.

2. Materials and Methods

2.1. Study Area

Seoul is located in the northwest of the Korean Peninsula and serves as South Korea’s political, economic, cultural, and transportation center. It is also one of the most densely populated cities in the country. Covering a total area of approximately 605 km², Seoul has a permanent population of nearly 10 million, a high level of urbanization, and a broad expanse of built-up areas characterized mainly by high-density residential, commercial, and mixed land uses. The city is surrounded by mountains to the north and east, with the Han River running west to east through the urban core. Seoul experiences a temperate monsoon climate with distinct seasons—hot and rainy summers, and cold, dry winters. In contrast, Busan is located on the southeastern coast of South Korea and is the country’s second-largest city and largest port, with an area of about 770 km² and a population of approximately 3.4 million. The urban spatial structure of Busan is influenced by both mountainous terrain and the coastline, forming a belt-shaped distribution. The urban area is dissected by multiple mountain ranges and rivers, resulting in a diversified urban landscape. Busan has a temperate oceanic climate, with small annual temperature fluctuations, higher humidity, and a lower frequency of extreme heat events compared to Seoul due to maritime influences. The two cities differ markedly in terms of geographic location, topographical characteristics, climatic conditions, population size, and urban spatial structure. Therefore, this study selects Seoul and Busan—two highly representative cities in South Korea—as the research areas (Figure 1).

2.2. Data Sources

The land surface temperature (LST) data used in this study were derived from the Landsat 8 OLI/TIRS Collection 2 Level-2 LST product, released by the United States Geological Survey (USGS). The product’s Thermal Infrared Sensor (TIRS) original data have a spatial resolution of approximately 100 m, which are resampled to 30 m using USGS’s standard processing to ensure consistency with OLI multispectral data, thus enabling higher-precision spatial analysis. The temporal coverage spans from June to August 2023, with the study area comprising Seoul and Busan, South Korea. All images were obtained via the USGS EarthExplorer platform, and QA_PIXEL masks were applied to remove images with more than 20% cloud cover, ensuring the spatial continuity and data quality of the results.

LST retrieval follows the USGS-recommended Single Channel (SC) algorithm, based solely on TIRS Band 10 [34]. The specific processing steps include the following: First, converting the digital numbers (DN) of the images to radiance to obtain the brightness temperature (T_B10). Then, atmospheric correction and emissivity correction are performed, and the LST is retrieved according to the following formula:

$$\mathrm{L}\mathrm{S}\mathrm{T}={\displaystyle \frac{{\mathrm{T}}_{\mathrm{B}10}}{1+\frac{\lambda \cdot {\mathrm{T}}_{\mathrm{B}10}}{\rho }\mathrm{l}\mathrm{n}\epsilon }}$$

(1)

where ${\mathrm{T}}_{\mathrm{B}10}$ is the brightness temperature (K), λ is the central wavelength of the band (10.895 μm), $\rho ={\displaystyle \frac{\mathrm{h}\cdot \mathrm{c}}{\sigma }}$, with $\mathrm{h}$ being Planck’s constant, c being the speed of light, and σ sigma being the Boltzmann constant, and ε is the surface emissivity. Surface emissivity is estimated based on the normalized difference vegetation index (NDVI) partitioned threshold method to further improve retrieval accuracy [35]. All LST data are uniformly projected to the WGS 84 coordinate system to ensure consistency with other spatial data.

In addition, this study selected eight representative independent variables, encompassing the built environment (BD, BH, RD), natural environment (NDVI, NDWI, DEM), and socio-economic factors (PD, GDP) to construct a multi-dimensional indicator system. All variables were generated in raster format, standardized to a spatial resolution of 30 m, and registered to the WGS 84 coordinate system to ensure spatial data consistency. All data preprocessing, including rasterization and georeferencing, was performed using ArcGIS software (version 10.8.1). The definitions, calculation methods, and data sources for each variable are detailed in

Table 1. Summary of Independent Variables.

Variable	Definition	Calculation Method	Data Source
BD	Building density	$\mathrm{B}\mathrm{D}={\displaystyle \frac{\sum {\mathrm{A}}_{\mathrm{b}}}{{\mathrm{A}}_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}}}$	Building height of Asia in 3D-GloBFP
BH	Average building height	$\mathrm{B}\mathrm{H}={\displaystyle \frac{\sum {\mathrm{H}}_{\mathrm{b}}}{\mathrm{n}}}$	Building height of Asia in 3D-GloBFP
RD	Road density per unit area	$\mathrm{R}\mathrm{D}={\displaystyle \frac{{\mathrm{L}}_{\mathrm{r}\mathrm{o}\mathrm{a}\mathrm{d}}}{{\mathrm{A}}_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}}}$	OpenStreetMap
NDVI	Normalized difference vegetation index	$NDVI={\displaystyle \frac{NIR-Red}{NIR+Red}}$	Landsat 8 OLI
NDWI	Normalized difference water index	$NDWI={\displaystyle \frac{Green-NIR}{Green+NIR}}$	Landsat 8 OLI
DEM	Terrain relief index	$TRI=\sqrt{\sum {\left(z_0-z_i\right)}^2}$	SRTM (30 m)
PD	Population density per unit area	$PD={\displaystyle \frac{P_{grid}}{A_{grid}}}$	SGIS Statistical Geographic Information Service (https://sgis.kostat.go.kr/ (accessed on15 December 2024))
GDP	GDP per unit area	$GD{P}_{grid}=GD{P}_{district}\times {\displaystyle \frac{A_{grid}}{A_{district}}}$	Korean Statistical Information Service (KOSIS) (https://kosis.kr/ (accessed on 15 December 2024))

.

To ensure consistency and comparability among variables during model training, all input variables in this study were subjected to normalization [36]. Given the significant differences in scale and numerical range across multi-source variables, and to avoid the interference of scale effects on model training results, the Min–Max normalization method was employed to standardize all variable values to the (0, 1) interval [37]. This normalization not only improves the efficiency and convergence speed of model training but also reduces the impact of variable scale differences on the interpretation of results—an effect particularly pronounced in ensemble learning models such as XGBoost, which are sensitive to feature distributions.

2.3. Research Models

2.3.1. Multiple Linear Regression (MLR)

Multiple linear regression is used to characterize the linear relationship between multiple variables and the dependent variable [38]. It is widely used in the fields of geography and environment for quantitative analysis. The model is expressed as

$$\mathrm{Y}={\beta }_0+\sum _{\mathrm{i}=1}^{\mathrm{n}}{\beta }_{\mathrm{i}}{\mathrm{X}}_{\mathrm{i}}+ϵ$$

(2)

where $\mathrm{Y}$ is the dependent variable (LST), ${\mathrm{X}}_{\mathrm{i}}$ denotes the influencing factors, ${\beta }_{\mathrm{i}}$ are the regression coefficients, ${\beta }_0$ is the intercept, and $ϵ$ is the random error term. This parametric method enables the quantitative assessment of the direction and significance of each variable’s impact on LST but is limited in its ability to handle nonlinear relationships among multiple variables.

2.3.2. Random Forest Regression Model (RF)

Random forest is an ensemble regression approach based on the Bagging concept, in which multiple decision trees are constructed. Each tree is trained through bootstrap sampling and generated by random feature selection, with the final result being the average of all trees [39]. Its formulation is

$$\widehat{\mathrm{Y}}={\displaystyle \frac{1}{\mathrm{M}}}\sum _{\mathrm{m}=1}^{\mathrm{M}}{\mathrm{f}}_{\mathrm{m}}\left(\mathrm{X}\right)$$

(3)

where $\widehat{\mathrm{Y}}$ is the predicted value, M is the number of trees, and ${\mathrm{f}}_{\mathrm{m}}$(X) is the output of the mth tree. RF has strong capabilities for high-dimensional and nonlinear data fitting and includes feature importance, but the model itself has limited interpretability.

2.3.3. Light Gradient Boosting Machine (LightGBM)

LightGBM is an efficient gradient boosting decision tree model that constructs strong nonlinear mapping functions by iteratively optimizing loss and combining multiple decision trees [40]. The core equation is

$${\mathrm{F}}_{\mathrm{t}}\left(\mathrm{x}\right)={\mathrm{F}}_{\mathrm{t}-1}\left(\mathrm{x}\right)+\lambda {\mathrm{h}}_{\mathrm{t}}\left(\mathrm{x}\right)$$

(4)

where Ft(X) is the model after t iterations, X is the feature set, and ht(X) is the new decision tree at iteration t. LightGBM can efficiently handle large-scale data and is capable of modeling nonlinear relationships and complex interactions.

2.3.4. XGBoost Regression Model

XGBoost is an efficient implementation of gradient boosting decision tree (GBDT) ensemble methods. The core idea is to sequentially add new trees based on the previous model, each time optimizing the ensemble’s loss function by minimizing the overall loss [41]. The XGBoost loss function and regularization are

$$\mathcal{L}=\sum _{\mathrm{i}=1}^{\mathrm{n}}\mathrm{l}\left({\mathrm{y}}_{\mathrm{i}},{\widehat{\mathrm{y}}}_{\mathrm{i}}\right)+\sum _{\mathrm{k}=1}^{\mathrm{K}}\Omega \left({\mathrm{f}}_{\mathrm{k}}\right)$$

(5)

where $\mathrm{l}$ is the loss function, ${\mathrm{y}}_{\mathrm{i}}$ and ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ are the true and predicted values for the $\mathrm{i}$th sample, Ω is the regularization term, and K is the number of trees. XGBoost supports automatic calculation and selection of optimal features, and its accuracy and efficiency have made it widely used in urban and environmental remote sensing modeling.

2.3.5. SHAP Value Interpretation

To enhance model interpretability, this study adopted the SHAP method to interpret machine learning model results. SHAP is based on Shapley values from game theory and attributes the model’s output to the actual contribution of each feature [42]. For a sample xxx, the SHAP value is defined as

$${\varphi }_{\mathrm{j}}=\sum _{\mathrm{S}\subseteq \mathrm{F}\backslash \left\{\mathrm{j}\right\}}{\displaystyle \frac{\left|\mathrm{S}\right|!\left(\left|\mathrm{F}\right|-\left|\mathrm{S}\right|-1\right)!}{\left|\mathrm{F}\right|!}}\left[\begin{array}{c}{\mathrm{f}}_{\mathrm{S}\cup \left\{\mathrm{j}\right\}}\left({\mathrm{x}}_{\mathrm{S}\cup \left\{\mathrm{j}\right\}}\right)-{\mathrm{f}}_{\mathrm{S}}\left({\mathrm{x}}_{\mathrm{S}}\right)\end{array}\right]$$

(6)

where ϕj is the SHAP value for the jth feature, F is the full set of features, S is a subset excluding j, and ${\mathrm{f}}_{\mathrm{S}}$ is the model trained with feature set S. SHAP quantitatively explains the contribution size and direction of each variable to LST prediction, improving the scientific interpretability of machine learning models.

2.4. Model Evaluation Metrics

To comprehensively assess the predictive performance of different regression models, this study employed three main evaluation metrics: mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination ($R^2$). MAE measures the average absolute deviation between predicted and observed values [43]; RMSE reflects the average magnitude of the prediction errors and is sensitive to larger errors [43]; $R^2$ represents the model’s explanatory power regarding the variance in observed data, with values approaching 1 indicating better fit [44]. The metrics are calculated as follows:

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\widehat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left(\stackrel{\mathrm{‾}}{\mathrm{y}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}}$$

(7)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\widehat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right)}^2}$$

(8)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\left|{\widehat{\mathrm{y}}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}}\right|$$

(9)

where n is the sample size, ${\widehat{\mathrm{y}}}_{\mathrm{i}}$ is the predicted value for the ith sample, yi is the observed value, and $\stackrel{\mathrm{‾}}{\mathrm{y}}$ is the mean of the observed values. Higher ${\mathrm{R}}^2$ and lower RMSE and MAE indicate superior model performance and predictive capability.

2.5. Research Framework

The research framework of this study is illustrated in Figure 2. First, Seoul and Busan were selected as the study areas, and multi-source spatial data were collected for the summer of 2023, including land surface temperature (LST), built environment variables (BD, BH, RD), natural environment factors (NDVI, NDWI, DEM), and socio-economic indicators (PD, GDP). All datasets were georeferenced to the WGS84/UTM zone 52 N coordinate system and rasterized to ensure spatial consistency and comparability. In the spatial data processing stage, ArcGIS 10.8.1 software was employed to calculate various spatial indicators and perform statistical aggregation within a 500 m × 500 m grid cell, which was adopted as the basic analysis unit (Supplementary Materials). This scale selection comprehensively considered both the spatial resolution of Landsat 8 imagery and the spatial heterogeneity characteristics of the urban heat island effect within cities. Smaller grids facilitate the capture of fine-scale urban microclimate variations; however, excessively small units may introduce noise and increase computational burden. The 500 m × 500 m grids can efficiently reveal spatial variability without sacrificing accuracy or introducing unwanted noise in urban planning tasks [45]. The 500 m scale balances spatial accuracy and computational efficiency while also aligning with common spatial analysis units in urban thermal environment research, thereby enhancing the stability of model analysis and the interpretability of results.

For the modeling stage, the performances of several regression models—including MLR, RF, LightGBM, and XGBoost—were compared. The entire modeling pipeline, from data preprocessing to interpretation, was implemented in Python 3.9 using libraries such as pandas, NumPy, scikit-learn, XGBoost, Optuna, and SHAP. The XGBoost model was selected as the principal analytical tool due to its optimal predictive accuracy and feature identification capability. Hyperparameter tuning for the XGBoost model was performed using Optuna, which conducted 30 trials with a TPE sampler to minimize the validation mean squared error. The search space for optimization included n_estimators (50–500), max_depth (2–8), learning_rate (0.01–0.3), min_child_weight (1–12), subsample (0.5–1.0), and reg_lambda (0–10). The best-performing model from this process was then interpreted using the SHAP method to elucidate the global and local linear and nonlinear mechanisms, as well as interaction effects, among variables. The findings of this study provide a more refined and differentiated regulatory basis for urban planners committed to alleviating the urban heat island effect.

3. Results

3.1. Model Performance Evaluation

To verify the advantages of the XGBoost model in modeling the driving factors of urban thermal environments in different cities, this study compared it with multiple linear regression (MLR), LightGBM, and random forest (RF) models. The core performance indicators included mean absolute error (MAE), root mean square error (RMSE), and R², as shown in Figure 3.

In Seoul, the XGBoost model achieved the lowest MAE and RMSE, at 1.0069 and 1.2859, respectively, indicating the smallest error among all models. Its R² reached 0.6182, significantly higher than that of MLR (0.3750), RF (0.5793), and LightGBM (0.5652), demonstrating the strongest explanatory power for the observed data. While the RF and LightGBM models outperformed MLR in MAE and R², their overall performance still lagged behind XGBoost. In Busan, XGBoost also exhibited the best performance, with MAE and RMSE of 1.0220 and 1.3966, which were significantly lower than those of MLR (1.3964, 1.8087), RF (1.1075, 1.4942), and LightGBM (1.1487, 1.5223). Meanwhile, XGBoost achieved an R² of 0.5876, much higher than MLR (0.3082) and superior to RF (0.5279) and LightGBM (0.5100).

As a comprehensive and high-performing ensemble learning model, XGBoost not only excels in predictive accuracy but also possesses the ability to handle complex multivariate relationships and nonlinear structures [46]. Its unique feature selection and ensemble mechanisms effectively enhance model stability and generalization capability, while accommodating missing values and reducing the risk of overfitting [47]. These advantages render it more suitable for interpreting the complex relationships between urban spatial factors and land surface temperature. Therefore, this study adopts XGBoost as the core model for subsequent analysis of variable mechanisms to enhance the scientific validity and practical value of the research conclusions.

3.2. SHAP Model Interpretation and Feature Importance Analysis

To comprehensively reveal the key driving factors of urban thermal environments, this study utilized the XGBoost model in combination with the SHAP method to conduct feature attribution analysis for LST in Seoul and Busan, as illustrated in Figure 4. Figure 4(a1,a2) presents the feature importance ranking of each variable, based on the mean absolute SHAP value across all samples, which measures the overall impact of each variable on model prediction. A larger value indicates a more significant contribution to LST prediction, making it a representative quantitative indicator for interpreting model output [48]. Figure 4(b1,b2) further demonstrates the direction and magnitude of each variable’s impact on LST predictions. Through color-coded SHAP values, the positive or negative regulatory effects of different variable values on the urban thermal environment are visualized: high variable values are represented by red points and low values by blue, with the vertical axis denoting SHAP values and representing the marginal contribution of each variable to LST [49]. This visualization intuitively reveals the warming or cooling trends of each factor at different value intervals, thereby providing important support for understanding nonlinear mechanisms and variable threshold characteristics.

Figure 4 shows the distribution of SHAP values and feature importance ranking for the main variables in Seoul and Busan. The results for Seoul indicate that BD and PD are the most critical factors driving changes in the urban thermal environment, with mean absolute SHAP values significantly higher than other variables, highlighting the dominant roles of high-density built-up areas and population concentration. The SHAP values in high BD and PD regions are mainly positive, further confirming the close association between intensive urban development and the urban heat island effect [50]. DEM ranks third in feature importance; positive SHAP values in low-elevation areas indicate that such areas are more prone to heat accumulation and the formation of intense thermal environments [51]. The SHAP distribution of GDP exhibits both positive and negative contributions, reflecting the dual regulatory effect of economic activity on the thermal environment—enhancing infrastructure while increasing development intensity. RD and BH also exhibit positive influences, pointing to the compounding effects of dense road networks and high-rise buildings on surface temperatures. Among natural environment variables, SHAP values for NDVI and NDWI are generally negative, indicating that increased urban greenery and water bodies can effectively mitigate thermal stress, though their overall influence is limited compared to built environment variables. In summary, the thermal environment in Seoul is predominantly controlled by built environment factors, with insufficient regulation by natural environment components.

For Busan, DEM and GDP exhibit the highest importance in driving the urban thermal environment, as indicated by their markedly higher mean absolute SHAP values, highlighting the dominant role of topography and socio-economic factors. The SHAP distribution for DEM shows that SHAP values in low-elevation areas are significantly positive, reflecting the substantial impact of topography on heat accumulation and on the spatial distribution of urban LST. The SHAP values for GDP are also primarily positive, suggesting a strong correlation between increased economic activity and urban thermal environment risks [52]. Although less influential than DEM and GDP, PD also demonstrates a pronounced positive effect in high-density areas. Among natural environment variables, NDVI SHAP values are mainly negative, suggesting a certain mitigation function for the thermal environment, albeit with limited overall contribution. RD, BD, BH, and NDWI are all ranked lower in importance, indicating that, in a spatially complex city like Busan with concentrated economic activity, the regulatory effects of built environment and water body factors are relatively weakened. In conclusion, the formation mechanism of Busan’s thermal environment is primarily driven by topography and economic factors, while the cooling effect of ecological systems is not prominent in the overall pattern.

3.3. Nonlinear Analysis of Main Effects of Variables

To further elucidate the main effects and nonlinear relationships of influencing factors on the urban thermal environment, this study employed partial dependence plots (PDP) based on the XGBoost model for the two cities. Figure 5 illustrates the PDP curves of urban thermal environment influencing factors, reflecting the specific response characteristics of each factor.

Figure 5a presents the analysis results for Seoul, showing marked differences in the main effects and nonlinear response characteristics of each variable on LST. Both BD and PD exhibit significant positive main effects; as BD and PD increase, LST rises continuously, with a marked surge in surface temperature during the transition from low to high BD, and the warming effect further intensifies in high PD intervals. GDP has a particularly pronounced enhancing effect on the urban thermal environment in the high-value range, with the PDP curve rising sharply when GDP exceeds 75, indicating that intensified economic activity increases urban surface heat (higher LST). NDVI and DEM both show negative regulatory effects: higher NDVI significantly lowers LST, and its mitigation effect plateaus when NDVI exceeds 0.6; increasing DEM rapidly decreases and stabilizes LST. BH overall demonstrates a negative main effect on the thermal environment, with the PDP curve indicating a gradual decrease in LST as BH increases, suggesting that high-rise buildings have a certain mitigating effect [53]. NDWI is generally negatively correlated with LST, but the curve fluctuates considerably, reflecting spatial heterogeneity in its mitigation effect. RD displays a non-monotonic, U-shaped relationship: LST is high in low RD intervals, decreases initially with increasing RD, and then rises again, implying a complex regulatory mechanism of road network structure on the urban thermal environment. Overall, Seoul’s urban thermal environment is influenced by socio-economic, morphological, and natural factors, each with distinct main effects and pronounced nonlinear relationships, providing data support for subsequent regulatory strategies.

Figure 5b presents the analysis results for Busan, highlighting certain differences in main effects and nonlinear response mechanisms compared to Seoul. BD exhibits an overall negative correlation with LST; the PDP curve shows that as BD increases, LST slowly decreases, suggesting that high-density built areas in Busan, possibly due to ventilation corridors or intensive management, contribute to heat mitigation. BH displays a complex nonlinear relationship, with certain intervals (e.g., BH ~10–15) having a warming effect, while extremely high or low building heights correspond to lower temperatures, indicating interactive effects between building height distribution and the thermal environment. DEM exhibits the most significant negative main effect, with the PDP curve showing LST plummeting from over 30 °C to 27.5 °C as DEM rises from 0 to 300, underscoring the critical role of topographical elevation in lowering urban LST. GDP mainly enhances LST in the high-value range (>60), with a steep increase in the PDP curve, indicating that economic activity intensity is a principal driver of localized thermal anomalies in Busan. PD is significantly positively correlated with LST, especially after PD exceeds 0.04, where LST rises sharply and stabilizes at a high level, highlighting the pronounced thermal risk in densely populated zones. NDVI generally displays a negative correlation, with the lowest LST in areas where NDVI > 0.8, emphasizing the mitigating effect of ecological space. The PDP curve for NDWI indicates that LST rises in intervals above −0.7, reflecting limited regulatory effect of water bodies on the thermal environment, and even local positive correlations where water area is limited. RD is positively correlated with LST, with a marked increase in LST after RD exceeds 0.03, suggesting that road network density intensifies thermal effects [54]. Overall, Busan’s thermal environment is co-driven by topography, socio-economic, urban morphological, and ecological factors, each exhibiting spatial heterogeneity and threshold characteristics in their main effects and nonlinear responses, providing an important data foundation for refined urban thermal environment regulation.

3.4. Analysis of Interactions Among Variable

To further understand the synergistic and coupled effects of influencing factors in the formation of urban thermal environments, this study employed the XGBoost model with the SHAP method to analyze the interactions among major variables. The SHAP interaction value matrix visualizes the joint influence of variable pairs on the urban thermal environment, while joint SHAP scatter plots further illustrate the direction and nature of variable interactions.

Figure 6 displays the SHAP interaction effects of key variables in Seoul. The interaction between DEM and BD is the most prominent; in low-elevation areas, the warming effect of building density on the thermal environment is particularly pronounced, while this effect is markedly attenuated with increasing DEM, indicating that higher elevations can mitigate the thermal risk of high-density development [55]. The interaction between GDP and PD also exhibits a strong additive effect, with heat environment risk significantly heightened in areas of both high GDP and high PD, reflecting the amplification mechanism of economic activity and population concentration on the LST effect. Further analysis shows that the synergy between BD and PD is also critical, with simultaneous increases in both exerting the strongest positive effect on LST, identifying intensively developed and densely populated areas as high-sensitivity zones for the urban thermal environment. The synergistic effects of natural environment factors are also noteworthy; the interaction between NDVI and NDWI shows that their simultaneous enhancement results in the strongest cooling effect, reflecting the integrated regulatory role of ecological spaces on the urban heat island effect. Moreover, in high-building density areas, increased building height significantly enhances the mitigating effect on the thermal environment, indicating that vertical spatial development can help optimize the local microclimate [56]. The interaction between RD and DEM further indicates that road density contributes substantially to the thermal environment in low-elevation zones, but its influence is limited in high-elevation areas. In summary, the nonlinear evolution of Seoul’s urban thermal environment is primarily driven by high-order interactions among topography, development intensity, economic-population coupling, and ecological synergy, providing robust data and theoretical support for precise urban thermal risk regulation.

Figure 7 presents the SHAP interaction effects of key variables in Busan. The interaction between DEM and GDP is most significant, with the positive impact of economic level on the thermal environment being more pronounced in low-elevation areas, while the effect of rising GDP on the urban thermal environment diminishes in higher-elevation zones. The interactions between GDP and spatial structure variables such as BD and BH are also notable, showing that increases in GDP, coupled with high building density and building height, further strengthen the positive effect on the thermal environment; in low GDP regions, the contribution of spatial structure variables is relatively limited. Additionally, DEM not only significantly modulates the effect of GDP on the thermal environment, but also interacts with NDVI, RD, and other variables, with the direction and intensity of their impacts on the thermal environment shifting as elevation increases. The interaction between NDWI and GDP reveals that the cooling effect of water bodies is more limited in areas with high economic levels. In the interaction between PD and GDP, higher economic levels intensify the warming effect of population density on the thermal environment [57]. Overall, the evolution of Busan’s thermal environment is not only influenced by single variables but is highly dependent on the compound effects of topography, economy, population, and spatial structure, with economic development and topographic factors playing particularly dominant roles, and the regulatory effects of spatial structure and ecological elements being most pronounced in high-economic and low-elevation areas.

4. Discussion

4.1. Differences in the Driving Factors of Urban Thermal Environments Between Seoul and Busan

Based on the results of the XGBoost-SHAP model analysis, a comparison of the driving factors of urban thermal environments in Seoul and Busan reveals both commonalities and differences. In both cities, an increase in NDVI significantly reduces land surface temperature, while high-intensity urban development (e.g., BD, PD) generally leads to temperature increases—characteristics typical of the urban heat island effect [58,59]. However, the relative strengths and mechanisms of the main driving factors differ markedly between the two cities. In Seoul, socio-economic and built environment indicators (such as PD, BD, GDP) associated with the degree of urbanization contribute more prominently to thermal environments, indicating that high development density and intense economic activity have reinforced the urban heat island effect [60]. In contrast, Busan, characterized by pronounced topographic variation and proximity to the ocean, is dominated by natural geographic and economic factors (e.g., DEM, GDP), with mountainous terrain and the marine environment providing inherent cooling effects [61]. The SHAP interaction analysis for the two cities further reveals differing mechanisms. In Seoul, there is a clear interaction effect between green spaces and the built environment: when development intensity is high, the cooling benefit of increasing NDVI diminishes, indicating that the mitigating effect of urban greening is limited in high-density areas [62]. In Busan, topographical factors profoundly influence other variables—for example, on higher-elevation slopes, abundant vegetation not only directly lowers temperatures but also further mitigates heat accumulation in low-lying built-up areas by enhancing ventilation [63]. In terms of economic level, high-GDP areas in Seoul often correspond to commercial centers with dense high-rise buildings, where energy consumption and waste heat emissions exacerbate local warming. In Busan, economic activities are mainly distributed along the coastal zone, where sea breezes help buffer thermal environment pressures in these areas [64]. In summary, Seoul’s thermal environment is dominated by built environment factors, with relatively limited ecological regulation, whereas Busan is primarily driven by topography and economic factors, and the ecological cooling effect is not significant in the overall pattern. This demonstrates that differences in topographic conditions, development density, and ecological patterns between the two cities have shaped distinct mechanisms underlying their urban heat island effects.

4.2. Urban Planning and Policy Recommendations

The results of this study provide a scientific basis for strategies to mitigate urban thermal environments. Based on the findings, interventions can focus on terrain utilization, green space configuration, vertical development, and infrastructure optimization to enhance urban heat regulation capacity. In terms of terrain utilization, urban planning should adapt to and leverage natural topography to facilitate heat dissipation [65]. For example, in Seoul, ventilation corridors along rivers and valley orientations should be preserved to alleviate heat retention caused by basin topography. In Busan, coastal wind channels and mountain gorge ventilation paths should be protected, utilizing sea breezes and elevation gradients to reduce heat accumulation in urban districts. For green space configuration, increasing urban green coverage and optimizing spatial distribution are necessary [66]. Within high-density built-up areas, the planning and construction of sufficient parks and urban tree canopies should be prioritized to enhance NDVI and offset the warming effect of impervious surfaces [67]. Connectivity within green space networks should be emphasized to form continuous ecological corridors from surrounding forests to central urban areas, thereby strengthening the overall ecological cooling function [68].

With respect to vertical development, building height and density should be reasonably controlled, and the layout of building clusters optimized to improve urban ventilation and solar access [69]. Excessive concentration of super high-rise buildings should be avoided to prevent the formation of “street canyon” heat island effects; instead, planning can involve distributing building heights or reserving open spaces [70]. The promotion of vertical greening and rooftop gardens is also encouraged, enabling high-rise buildings to enhance land use efficiency while reducing their adverse impacts on local thermal environments [71]. Regarding infrastructure optimization, it is recommended to improve urban hard surfaces and related facilities to reduce heat storage [72]. For instance, the adoption of high-albedo cool roofs and exterior wall materials, permeable and cooling pavements, increased street sprinkling and misting, and the addition of urban water features (such as fountains and reflecting pools) can all enhance the albedo and evaporative cooling capacity of urban surfaces [73,74].

At the policy level, these measures should be incorporated into urban planning and climate adaptation policy frameworks. Governments can promote the implementation of comprehensive strategies—such as subsidies for green roofs, ventilation corridor protection policies, building height regulation, and infrastructure upgrades—through improved regulations and incentive mechanisms. Such integrated approaches would systematically enhance urban thermal environment regulation and climate resilience [75].

4.3. Limitations and Future Directions

Several limitations of this study warrant attention and improvement in future work. First, the coverage of variables is not exhaustive. Although this study included topography, population, economy, and ecological factors, other potential influencing factors—such as anthropogenic heat emissions, underlying surface material properties, and more refined meteorological elements (e.g., wind speed, humidity)—were not incorporated. The absence of these variables may constrain the model’s ability to fully capture the causes of urban thermal environments. Future research should consider integrating a broader range of influencing variables, or adopt a fusion analysis combining remote sensing, meteorological, and socio-economic data to improve explanatory power. Second, regarding model scalability, although the XGBoost-SHAP model effectively reveals nonlinear and interactive relationships among factors, the model was trained on data from specific cities, and the generalizability of its conclusions to wider regions or other cities remains unverified. Furthermore, spatial autocorrelation and larger-scale climate factors were not fully addressed. Future studies could apply transfer learning and multi-city comparative approaches to extend the methodology to cities of different scales and contexts, thereby exploring whether the driving factors and mechanisms of urban thermal environments display universality or variability across cities.

5. Conclusions

This study took Seoul and Busan, two representative cities in South Korea, as research objects, and, by integrating multi-source data—including built environment (BD, BH, RD), natural environment (NDVI, NDWI, DEM), and socio-economic factors (PD, GDP)—employed the XGBoost model and the interpretable SHAP machine learning method to systematically and quantitatively analyze the main driving factors, nonlinear main effects, and interactions affecting urban thermal environments. The main conclusions are as follows:

(1)

The XGBoost model demonstrated significant superiority over traditional models such as multiple linear regression (MLR), random forest (RF), and LightGBM in both predictive accuracy and the characterization of complex nonlinear relationships.

(2)

In terms of feature importance, SHAP analysis revealed notable differences in the dominant factors of urban thermal environments between the two cities. In Seoul, LST is mainly driven by built environment and socio-economic factors, such as building density (BD), population density (PD), and GDP, highlighting the significant role of built environment and economic activity in elevating urban temperatures. In contrast, in Busan, topographic factors (DEM) and GDP are the primary determinants, with topographical variation and economic activity jointly shaping the local thermal environment pattern.

(3)

Nonlinear main effect analysis further uncovered the threshold characteristics of key variables: when GDP and PD reach high levels, their warming effects on LST are significantly enhanced. Meanwhile, the cooling effect of NDVI is particularly evident in the medium–high value range (approximately 0.6–0.8).

(4)

Multi-factor interaction analysis indicated clear synergistic and amplifying mechanisms among different driving factors. In low-elevation areas, DEM strengthens the warming influence of built environment factors such as high BD. The combined effect of NDVI and NDWI enhances the mitigation of urban heat by ecological factors, while the superimposition of high population density and high GDP markedly increases urban heat load.

By systematically identifying the driving factors and mechanisms of urban thermal environments in Seoul and Busan using multi-source data and machine learning methods, this study enriches theoretical understanding of urban thermal evolution in different city contexts. The results provide a solid empirical foundation for climate-adaptive urban planning and refined management of urban thermal environments, and also open new avenues for the application of interpretable artificial intelligence methods in urban environmental studies. Future research will incorporate meteorological, anthropogenic emission, and social behavioral data with higher temporal and spatial resolution, as well as extend dynamic comparative analyses across multiple regions and scales to further deepen scientific understanding and practical guidance for the evolution and regulation of urban thermal environments.