Next Article in Journal
Multi-Resolution and Multi-Temporal Satellite Remote Sensing Analysis to Understand Human-Induced Changes in the Landscape for the Protection of Cultural Heritage: The Case Study of the MapDam Project, Syria
Previous Article in Journal
Urban Agglomerations Promote the Coordinated Development of Urbanization and Intensive Land Use
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Land
Volume 14
Issue 11
10.3390/land14112232
land-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

City-Specific Drivers of Land Surface Temperature in Three Korean Megacities: XGBoost-SHAP and GWR Highlight Building Density

by
Hogyeong Jeong
1,
Yeeun Shin
2 and
Kyungjin An
1,3,*
1
Department of Forestry and Landscape Architecture, Graduate School, Konkuk University, Seoul 05029, Republic of Korea
2
Laboratory of Spatial Design Research, Konkuk University, Seoul 05029, Republic of Korea
3
Department of Bio & Healing Convergence, Graduate School, Konkuk University, Seoul 05029, Republic of Korea
*
Author to whom correspondence should be addressed.
Land 2025, 14(11), 2232; https://doi.org/10.3390/land14112232
Submission received: 22 October 2025 / Revised: 7 November 2025 / Accepted: 10 November 2025 / Published: 11 November 2025
(This article belongs to the Special Issue Beyond Boundaries: Exploring the Nexus of People, Land, and Water Along Land Use Gradients)
Browse Figures
Versions Notes

Abstract

Urban heat island (UHI), a significant environmental issue caused by urbanization, is a pressing challenge in modern society. To mitigate it, urban thermal policies have been implemented globally. However, despite differences in topographical and environmental characteristics between cities and within the same city, these policies are largely uniform and fail to reflect contexts, creating notable drawbacks. This study analyzed three cities in Korea with high land surface temperatures (LSTs) to identify factors influencing LST by applying Extreme Gradient Boosting (XGBoost) with Shapley Additive explanations (SHAP) and Geographically Weighted Regression (GWR). Each variable was derived by calculating the average values from May to September 2020. LST was the dependent variable, and the independent variables were chosen based on previous studies: Normalized Difference Vegetation Index (NDVI), Normalized Difference Built-up Index (NDBI), ALBEDO, Population Density (POP_D), Digital Elevation Model (DEM), and SLOPE. XGBoost-SHAP was used to derive the relative importance of the variables, followed by GWR to assess spatial variation in effects. The results indicate that NDBI, reflecting building density, is the primary factor influencing the thermal environment in all three cities. However, the second most influential factor differed by city: SLOPE had a strong effect in Daegu, characterized by surrounding mountains; POP_D had greater influence in Incheon, where population distribution varies due to clustered islands; and DEM was more influential in Seoul, which contains a mix of plains, mountains, and river landscapes. Furthermore, while NDBI and ALBEDO consistently contributed to LST increases across all regions, the effects of the remaining variables were spatially heterogeneous. These findings highlight that urban areas are not homogeneous and that variations in land use, development patterns, and morphology significantly shape heat environments. Therefore, UHI mitigation strategies should prioritize improving urban form while incorporating localized planning tailored to each region’s physical and socio-environmental characteristics. The results can serve as a foundation for developing strategies and policy decisions to mitigate UHI effects.

1. Introduction

The expansion of impervious surfaces driven by urbanization leads to an increase in LST, which exacerbates the UHI phenomenon [1,2,3]. UHIs, together with climate-related stressors (e.g., heat waves, floods, and air pollution), are major drivers of increased urban energy demand. In particular, elevated temperatures increase health risks for urban populations, with vulnerable groups such as older adults [4]. The most direct impact of urban heat islands on human health is exposure to high temperatures, thereby elevating the risk of heatstroke and cardiovascular disease in summer [5]. Therefore, rising urban LST can adversely affect ecosystems as well as human health and well-being [6,7].
To address the worsening urban heat island phenomenon, numerous studies have been conducted and policies have been implemented, both domestically and internationally [8,9,10]. Among available indicators, LST is extensively used to evaluate and monitor the UHI intensity and serves as a key variable in urban climate studies [11,12]. However, many studies have focused on associations and temporal patterns between LST and urban variables [13,14,15]. Specifically, prior research has tended to investigate one-dimensional relationships such as between topography and LST [16,17], between land use/land cover (LULC) and LST [11,18,19], and between NDVI and LST [20]. Several recent studies have considered multiple factors that influence LST [21,22,23,24]. A summary of representative studies that analyzed the relationships between LST and various influencing factors is provided in Table 1 [23,24,25,26,27,28].
As shown in Table 1, multiple variables have been used in previous studies. Key indicators describing urban thermal environments, such as NDBI, NDVI, and DEM, have been frequently employed.
Such work failed to adequately reflect the heterogeneity and complexity of the various impacts on LST, making it challenging to translate these findings into policy implications [29]. In particular, policies to improve urban thermal environments tend to rely heavily on single-type measures, such as expanding green spaces [8,30]. While numerous studies have demonstrated that expanding green spaces positively impacts thermal environments [31,32,33], insufficient attention to local context and the resulting uniformity in policies across most cities remains a limitation [30]. Differentiated strategies that reflect regional social structures, topographical conditions, and population densities remain limited [9,34,35]. Therefore, it is essential to understand the drivers underlying temperature increases and prioritize these causes in each region.
Hence, this study not only compares and analyzes determinants of land surface temperature in three major cities in South Korea but also identifies regional differences and provides baseline data to support the need for future customized responses. This study aims to analyze the factors and impacts of urban thermal environments by combining the nonlinear predictive power of XGBoost, a tree-based model, with the spatial interpretability of GWR.
In summary, this study aims to identify the factors driving improved thermal environments and prioritize them across cities with different regional characteristics. To achieve this goal, the analysis combines XGBoost with the feature importance evaluation of SHAP, thereby harnessing superior forecasting capability for complex relationships, with GWR, which facilitates local spatial interpretation [36,37]. Unlike previous studies that relied on a single modeling approach, this study introduces a hybrid analytical framework that combines the nonlinear predictive strength of machine learning with the spatial diagnostic capability of GWR. This integrative approach enables both global and local interpretations of LST drivers, offering a more comprehensive understanding of spatial heterogeneity. The city-specific results derived from this framework provide practical insights for designing regionally customized urban heat mitigation strategies and advancing sustainable spatial planning.

2. Research Methodology

The causes of the UHI effect encompass multiple domains, including climate, environment, population, and physical structure. Therefore, a multivariate approach is essential to analyze urban thermal environments [38,39]. Given the multifactorial nature of urban thermal environments, selecting an appropriate analytical technique to capture complex interdependencies is crucial. Linear regression has been widely used to identify factors influencing LST [40,41,42]. While it offers clear interpretability of variable relationships, it often fails to capture the nonlinearities and complex interactions inherent in urban systems. To address these limitations, various machine learning techniques have been introduced, and tree-based models are prominent nonlinear methods [43,44].
XGBoost-SHAP, one of the most powerful tree-based models, provides robust predictive performance and enables the quantification of variable importance. However, as a nonspatial model, it cannot fully account for spatial patterns, clustering effects, or spatial autocorrelation because predictions are based solely on attribute data, not geographic coordinates [45]. Previous studies using XGBoost have primarily focused on ranking or characterizing the relative importance of factors associated with LST [23,24,27], but their ability to represent spatial heterogeneity remains limited.
While SHAP provides a global interpretation of variable importance across the entire dataset, it does not capture how these effects may vary across space. In contrast, GWR allows us to identify spatial heterogeneity in these relationships. To address these challenges, this study integrates XGBoost-SHAP, a gradient boosting-based machine learning technique, with the spatial regression model GWR to analyze both the nonlinear contributions of explanatory variables and spatially varying effects. Through this, key factors influencing the urban thermal environment in Seoul, Daegu, and Incheon were identified, and empirical evidence was provided for developing customized thermal environment management strategies. The overall research process is presented in Figure 1.
(1) Using MODIS LST data, the 20-year summer (June–August) average LST was calculated at the district level nationwide. Three metropolitan cities (Daegu, Incheon, and Seoul) were selected that are consistently ranked in the top 15 in the analysis. (2) Key variables influencing LST were chosen and compiled. (3) An XGBoost model was built using the selected variables, and the SHAP values were calculated to analyze the effects and nonlinear interactions of the variables. (4) The spatial influence distribution of the variables within each study area was evaluated using GWR analysis.

2.1. Study Areas

Although the underlying mechanisms of UHI phenomena are generally consistent across cities, variations in land-cover composition, population density, and coastal proximity can generate distinct local thermal responses. Therefore, comparative analyses across cities with contrasting spatial and environmental settings remain essential for uncovering both shared and context-specific thermal mechanisms.
To ensure data-driven city selection, the Google Earth Engine (GEE) was used to analyze nationwide average LSTs for all Korean cities and counties using MODIS LST data collected every five years from 2000 to 2020 (Table 2). This long-term temporal perspective reduces the influence of short-term climatic anomalies and ensures that city selection reflects persistent thermal patterns rather than isolated heat events. This preliminary screening identified the 15 hottest areas that consistently ranked at the top over two decades. Among them, Daegu, Seoul, and Incheon were chosen as representative study sites because they not only maintained persistently high LSTs but also exhibited distinct geographic and environmental characteristics—offering an ideal foundation for investigating both common and divergent drivers of urban heat formation.
Daegu is a large city, with a population of 2,357,023 and total area of 1499.49 km2 as of 2025, featuring a mosaic of downtown and surrounding green spaces [46]. Nevertheless, since this study applied to administrative districts as of 2020, the analysis was conducted based on an area of 885 km2 before Gunwi-gun was incorporated into Daegu in July 2023. In the LULC ratio of Daegu, tree cover accounted for the lowest proportion at approximately 75% of the total area, followed by built-up areas at approximately 18% and cropland at approximately 5.8%. This shows that while Daegu has a wide distribution of green spaces because of its mountainous terrain, impervious areas are considerably developed in the downtown area due to its dense population.
Incheon, with a population of 3,031,361 in 2025 and total area of 1067.09 km2, is adjacent to Seoul and Gyeonggi Provinces and borders the sea to the west, encompassing many islands [47]. The LULC of Incheon was fairly even overall, with tree cover accounting for 38.34%, built-up areas for 24%, croplands for 20%, and bare/sparse land for 10%.
Seoul, the capital of South Korea, has a population of 9,335,724 and a total area of 605.21 km2 as of 2025. It is notably urbanized [48]. The LULC of Seoul was chiefly composed of built-up areas, which accounted for approximately 55%, followed by tree cover (35%) and water bodies (5%). This distribution indicates intensive urban development in the city center and the presence of water bodies along the Han River.
Overall, the three metropolitan cities share certain characteristics in their land use and urban form. All exhibit a high degree of urbanization with densely built-up downtown areas surrounded by green or mountainous regions, reflecting the typical spatial structure of South Korean Cities constrained by topography. However, distinct differences are evident. Seoul is the most intensively urbanized, with built-up areas exceeding half of its total area, indicating severe land use pressure. Daegu, although mountainous and covered by extensive tree areas, still experiences significant urban heat accumulation in its central basin due to limited ventilation. Incheon, in contrast, shows a more balanced land use composition and coastal influence, with numerous islands and a relatively high proportion of cropland and bare land.
These contrasts imply that while similar urban green spatial patterns exist, the thermal environments of each city are influenced by differing geographical and developmental contexts. The climatic and geographical characteristics of the three cities are summarized in Figure 2.

2.2. Variable Selection and Model Construction

Based on previous research, salient indices that could influence urban thermal environments were selected in this study. Initially, several candidate variables were evaluated, including NDBI, NDVI, normalized difference water index (NDWI), normalized difference soil index (NDBSI), wetness component (WET), POP_D, SLOPE, DEM, and ALBEDO. However, high correlations were observed among some variables, including the NDBI, NDVI, NDWI, WET, and NDBSI indices. Therefore, the final set was selected. The dependent variable was LST, and the predictors comprised NDVI, NDBI, ALBEDO, POP_D, DEM, and SLOPE. Although NDBI and NDVI are highly correlated variables, numerous analyses have included them jointly as key indices in urban thermal environment analysis. To address this issue, a residual variable (NDVI_resid) was generated using R version 4.4.1 (R Core Team, 2024) and R Studio based on a simple linear regression model (NDVI ~ NDBI). After converting NDVI into a residual variable, the variance inflation factors (VIFs) for all variables fell below ten, ensuring model stability (Table 3).
The data were collected from June to August in the summer. However, owing to the characteristics of the Korean climate, the period was extended from 1 May to 30 September, as it was difficult to secure accurate information and data were missing during the summer monsoon season because of the cloudy weather. Despite the masking process, prolonged cloud cover may have reduced the number of valid observations, which should be considered when interpreting the seasonal averages.
Core datasets were collected and processed using QGIS and GEE. NDVI, NDBI, ALBEDO, DEM, and SLOPE were derived using Landsat 8 Operational Land Imager/Thermal Infrared Sensor (OLI/TIRS) Surface Reflectance (L2) imagery provided by the U.S. Geological Survey (USGS). DEM and SLOPE were derived from the USGS Shuttle Radar Topography Mission (SRTM v3, SRTMGL1_003) dataset with a spatial resolution of 30m. POP_D was obtained from WorldPop (2020) [49]. All indices were calculated as average values in 1 km grid units and were used for the analysis.

2.3. Research Model

2.3.1. Implementation of XGBoost-SHAP for Variable Importance Analysis

XGBoost is an algorithm built upon gradient boosting decision trees [50]. This is suitable for this study because it effectively captures complex interactions and nonlinear relationships, such as those between LST and variables, which are difficult to explain using simple linear models [24]. Furthermore, it is widely used owing to its efficiency, superior performance, and strong predictive accuracy compared with other machine learning methods [51,52,53]. The XGBoost formula is as follows:
o b j θ =   ∑ i = 1 n L ( y i , y ^ i ) + ∑ k = 1 K Ω f k
where L ( y i , y ^ i ) denotes the loss function quantifying the error between the observed value y i and predicted values y ^ i . Ω ( f k ) denotes the regularization term, which discourages complexity to prevent overfitting. Here, i is the sample index, k is the index of each tree and θ represents the model parameters.
Machine learning models, such as XGBoost, excel at producing accurate predictions. However, these are commonly viewed as opaque because they do not intuitively reveal the extent and direction of the contribution of each input variable. SHAP is an indispensable tool for interpreting tree-based machine-learning models [54]. Developed by Shapley in 1953, SHAP is a principal approach based on the cooperative game theory that can interpret the contributions of various factors [55]. It also indicates whether contributions increase or decrease the predicted value, making it a valuable tool for enhancing the transparency of tree-based machine learning models [56]. The formula for SHAP is as follows:
ϕ i =   ∑ s ⊆ N S ! N − S − 1 ! N ! · v S ∪ i − v ( S )
Each ϕ i denotes the SHAP value of feature i , which captures the effect of each variable on the model prediction. S indicates any subset of the features excluding i , N denotes the set of all features, and v ( S ) is the model output when only the features in S . This formula calculates the average marginal contribution of feature i across all possible combinations of other features.

2.3.2. Implementation of GWR for Spatial Variability Analysis

Next, the spatial heterogeneity of the variables affecting LST was examined using GWR. GWR is a spatial regression technique that was first developed in 1970 by the British university professors Fotheringham, Charlton, and Brunsdon. The GWR model incorporates spatial coordinates into traditional regression analysis, enabling the spatial visualization of regional relationships [57]. In this study, a Gaussian kernel was used as the kernel function, and the regional regression coefficients were estimated using a nearest-neighbor kernel of size nineteen at every site. Distance calculations employed straight-line metrics, and regression was performed using the actual observation point coordinates. The model settings, including bandwidth optimization, were determined to minimize AICs. The GWR model is as follows:
y i =   β 0 u i , v i +   ∑ k = 1 K β k u i , v i x i k + ∈ i
where y i denotes the response at the location i (e.g., LST); x i k is the k th independent variable at location i ; ( u i , v i ) represents the spatial coordinates of observation i ; β k ( u i , v i ) is the local regression coefficient for the k th variable at position ( u i , v i ); ∈ i is the random error term. In the GWR framework, regression coefficients are estimated individually for each spatial location, thereby revealing spatial heterogeneity and local relationships between predictors and the dependent variable.

2.3.3. Model Training and Evaluations

The assembled dataset was divided into training and testing phases to evaluate the generalization performance of the model. Eighty percent of the data were used for training to estimate the model parameters, while the remaining 20% were used for validation to limit overfitting and evaluate the predictive accuracy. This partitioning method is a common practice for consistently validating the performance of XGBoost, a machine-learning-based model, and the comparative model GWR. The model performance was evaluated using the root mean square error (RMSE) and coefficient of determination (R2), which provide interpretable measures of discrepancy.

3. Results

3.1. Spatial Patterns of LST and Socio-Environmental Variables

Analysis of the spatial distribution of the variables revealed that within the same city, variations occurred across different urban zones such as the urban core and periphery (Figure 3). This intra-urban land use gradient underscores the linkages between land use patterns, urban thermal conditions, population distribution, and building density.
The dependent variable LST exhibited higher values in the urban cores across all three cities. However, in Daegu and Incheon, distinct spatial patterns were observed, with temperature differences of approximately 25 °C between urban and non-urban areas along the land use gradient. In contrast, Seoul showed relatively smaller temperature differences between the urban core and periphery. Similarly, urbanization indicators, such as POP_D and NDBI, were spatially distinct in Incheon and Daegu, whereas in Seoul, urbanization was more evenly distributed, resulting in no distinct spatial anomalies.

3.2. Analyzing XGBoost-SHAP Results

The relationship between LST and explanatory variables was examined using the XGBoost model for Daegu, Incheon, and Seoul. The RMSE and R2 values were calculated to evaluate the predictive performance of the model; the results are presented in Table 4. When the model trained with the training dataset was applied to the validation dataset, an RMSE of 1.042 (95% CI: 0.849–1.235) and an R2 of 0.966 were obtained for Daegu. For Incheon, an RMSE of 1.375 (95% CI: 1.155–1.602) and an R2 of 0.856 were observed. For Seoul, the model showed somewhat lower explanatory power, with an RMSE of 1.683 (95% CI: 1.428–1.946) and an R2 of 0.727. These results confirmed that the predictive accuracy of the model was generally strong across all three metropolitan areas, indicating strong predictive performance.
After verifying the predictive performance, the primary predictors affecting the LST were derived using XGBoost (Table 5).
Additionally, SHAP analysis was conducted to clarify the directional effects of the explanatory variables on LST, which could not be confirmed through variable importance alone. The SHAP results are shown in Figure 4.
The interpretation of the derived XGBoost-SHAP results showed significant similarities and differences. First, NDBI was identified as a variable with substantial influence across all cities, with values of 2.649 in Daegu, 1.026 in Incheon, and 1.293 in Seoul. However, the other significant variables, excluding NDBI, differed in importance across the three target sites. In Daegu, SLOPE had an influence of 1.144; in Incheon, POP_D was 1.337; and in Seoul, DEM influenced 0.873. To further examine the relationships in detail, dependency plots for the top three variables for each target site are shown in Figure 5.
The plot shows how the variable values change in LST prediction, revealing nonlinear patterns and potential interactions between the variables. In each scatterplot, the x-axis denotes the variable value, and the y-axis represents the SHAP value of the variable. The color indicates the interaction variable with the most influential variable. In Daegu and Seoul, DEM functioned as the interaction variable with NDBI, whereas ALBEDO emerged as the primary interaction variable in Incheon. Meanwhile, NDBI was identified as the interaction variable with POP_D in Daegu and Incheon, whereas other variables exhibited site-specific effects. This is likely because of the various geographical, environmental, and structural characteristics of each region, even for the same variable.
Although the variables influencing the sites differed, the LST trends across the variables were similar. NDBI and ALBEDO generally had the SHAP values to increase as the variable values increased, whereas POP_D showed a tendency to grow to a certain level before saturation. The DEM and SLOPE exhibited a pattern of rising to a certain level and then rapidly decreasing.

3.3. Analyzing GWR Results

XGBoost-SHAP analysis was used to identify the global effects and patterns of each variable across the entire study area. Subsequently, a GWR analysis was used to characterize spatial variation in variable impacts within each study area.
Similarly to XGBoost, the RMSE and R2 were calculated in the GWR analysis to verify the predictive performance of the model, and the results are shown in Table 6. When the model trained on the training data was applied to the validation data, an RMSE of 1.143 (95% CI: 0.912–1.381) and an R2 of 0.958 resulted for Daegu, an RMSE of 1.294 (95% CI: 1.212–1.377) and an R2 of 0.889 were obtained for Incheon, and an RMSE of 1.428 (95% CI: 1.324–1.542) and an R2 of 0.813 were obtained for Seoul. These results indicate that the proposed model can achieve a strong fit without overfitting.
Unlike typical regression analyses, GWR produces unique regression coefficients for each location. This shows that the degree and direction of the influence of a specific factor on LST can vary spatially. The GWR results for Daegu are presented in Table 7. As shown in Table 7, the NDBI in Daegu generally exhibits a positive range. Conversely, the SLOPE and DEM exhibited negative ranges. This indicates that LST decreased as DEM, SLOPE, and NDVI_resid increased. ALBEDO and NDVI_resid exhibited broad ranges of maximum and minimum values.
Figure 6 shows the spatial distribution of the GWR coefficients for three top-ranked variables (NDBI, SLOPE, and POP_D) based on XGBoost-SHAP for each 1 km grid within Daegu.
The colors indicate the sign and magnitude of the coefficients, with red representing an appreciable rise in LST and blue representing a considerable decrease. The NDBI coefficient is mostly red, except in the southeastern mountainous regions. The SLOPE coefficient is red only in some southern regions, where industrial complexes cluster, with blue dominating most other regions. Finally, POP_D exhibited a red coefficient citywide, with its impact being particularly pronounced in the southern region.
Table 8 confirms that NDVI_resid, NDBI, and ALBEDO exerted a strong influence. Conversely, DEM and SLOPE showed lower LST with increasing values, suggesting an auxiliary role. The variable with the lowest value was POP_D, which had a minimal effect.
Figure 7 shows the spatial distribution of the GWR coefficients for three variables with high importance rankings (POP_D, NDBI, and ALBEDO) based on XGBoost-SHAP for each 1 km grid within Incheon.
The POP_D coefficient for Incheon showed positive values but relatively small absolute values in most grids. NDBI showed particularly strong positive coefficients in the downtown and airport areas, and ALBEDO showed a similar spatial distribution pattern. Compared to Incheon, which is a coastal basin city where built-up intensity strongly affects surface temperature, Seoul is an inland city with more complex topography. Therefore, while broadly similar coefficient patterns are expected due to Seoul’s highly urbanized characteristics, local variations may emerge owing to its inland and mountainous terrain.
The following presents the GWR results for Seoul. As shown in Table 9, the NDBI coefficient for Seoul shows positive values across the entire grid. Conversely, the DEM was negative, confirming that the LST decreased with increasing elevation. The remaining variables—SLOPE, NDVI_resid, ALBEDO, and POP_D—spanned positive and negative coefficients. Among them, POP_D had the weakest effect.
Finally, the results were mapped for Seoul. Figure 8 shows the spatial distribution of the GWR coefficients for NDBI, DEM, and NDVI_resid, three variables with high importance based on XGBoost-SHAP, at 1 km grid levels in Seoul.
The NDBI coefficient in Seoul showed predominantly positive values across all regions, whereas the DEM coefficient was largely negative, particularly in the northern mountainous regions. NDVI_resid mostly showed a positive coefficient distribution except in mountainous areas.
In summary, the LST was elevated in the city center in all three cities, but the gap between city centers and non-city centers varied depending on the development pattern of each city. The XGBoost-SHAP analysis showed that NDBI was the primary variable in all three cities, SLOPE had a salient effect in Daegu, DEM in Seoul, and POP_D in Incheon. The GWR results show that these influences vary by region. This indicates that the factors influencing LST differ depending on the structural characteristics of the city, supporting the need for customized thermal environment management strategies for each city.

4. Discussion

This study employed XGBoost-SHAP and GWR to analyze the factors influencing the LST in three major metropolitan cities in South Korea that have shown high LSTs over the past two decades. The results revealed that UHIs and the rise in LST are not solely determined by environmental variables but are the outcomes of a complex interplay among human activities, land use patterns, and topographic characteristics. Unlike previous studies that primarily focused on a single region, this study conducts a comparative analysis across three cities with distinct regional characteristics to identify and explain the spatial differences in thermal environmental drivers. Furthermore, moving beyond the limitations of studies that rely solely on a single machine learning model, this study combines XGBoost-based predictive modeling with GWR analysis. This integrative approach provides explanatory power for the influencing factors and spatial interpretability, highlighting the academic significance of integrating machine learning and spatial statistical models in urban climate research.

4.1. Interpretation of XGBoost-SHAP Results

According to the XGBoost-SHAP results, NDBI emerged as the most influential factor across all three study areas. This finding aligns with the fact that the target cities are large metropolitan regions with multi-million populations. Previous research on the relationship between urban LST and its drivers has similarly emphasized the roles of impervious surface expansion and building density in increasing LST by increasing surface heat capacity and thermal conductivity [33,51,52].
Beyond NDBI, however, the dominant secondary factors varied among the cities: SLOPE in Daegu, POP_D in Incheon, and DEM in Seoul. Daegu has basin-shaped terrain, surrounded by mountains to the north and south. While the urban core sits on flat terrain, the surrounding mountainous and hilly landscapes create a distinct spatial gradient on the slope, significantly affecting LST. This pattern can be attributed to the basin topography. In basin-shaped areas, air circulation is often restricted by surrounding mountains, causing heat to become trapped near the surface [58]. During the daytime, radiative heat accumulates, and at night, limited ventilation prevents the dissipation of stored heat, resulting in minimal cooling effects [59]. Such phenomena have been clearly demonstrated in previous studies using numerical modeling, field observations, and satellite-based analyses [53,54].
In Incheon, POP_D exerted the greatest effect on LST. This reflects the unique spatial setup of the city, featuring a highly urbanized mainland core and numerous peripheral islands. Although the central districts are dense, 128 of 168 islands of Incheon (76%) lack residents, creating a sharp contrast in settlement patterns. This spatial imbalance in population density drives concentrated impervious surface coverage, enhancing the explanatory power of POP_D for LST. However, the GWR results indicated that the local coefficients for POP_D were close to zero in most areas, suggesting that its localized influence was limited. This discrepancy underscores the different sensitivities of the models; POP_D may be a significant predictor in certain regions, but its overall impact at broader spatial scales is limited. This limited local influence may partly result from the 1km grid resolution of the dataset, which constrains the ability to fully capture fine-scale variations in population density and impervious surfaces.
In Seoul, the DEM was more relevant than the slope. This may be due to the complex terrain of the city, which includes river plains, lowlands, and mountainous zones around the Han River. Unlike Daegu, where slope differences are more pronounced, Seoul shows a more uniform distribution of elevational gradients. Therefore, the DEM has a greater influence in explaining the spatial variation in LST.

4.2. Interpretation of GWR Results

The GWR analysis indicated that NDBI and ALBEDO generally exerted a positive influence on LST across most areas. This aligns with prior studies emphasizing the role of impervious surfaces and reflective materials in heat accumulation in urban environments [60,61]. However, unlike studies that included both rural and urban spatial contexts, our analysis focused on large metropolitan cities characterized by diverse and complex land use patterns. As a result, topographic variables such as DEM and SLOPE did not demonstrate uniform cooling effects. For instance, while steep slopes in Daegu are largely associated with forested mountainous areas, steep terrains in Seoul also include densely developed urban districts. This suggests that elevation or slope alone cannot guarantee LST reduction unless accompanied by continuous vegetation or ventilation pathways.
Furthermore, the spatial discrepancies observed between the XGBoost-SHAP and GWR results indicate that certain predictors-particularly NDBI and POP_D-may behave differently depending on the analytical scale. Whereas XGBoost captures the overall predictive importance of variables at the city level, GWR reveals that their local effects may be weak or inconsistent. This pattern was particularly evident in Incheon, where some uninhabited islands displayed artificially strong positive POP_D coefficients, illustrating the limited spatial explanatory power of demographic variables in heterogeneous urban morphologies. Similar findings were reported by Chen et al. (2023) [1], who noted that while population density is often correlated with LST in global models, its spatial explanatory power diminishes when local variations in land use and development form are considered.
Overall, while NDBI remains the dominant factor influencing LST, its local expressions vary across different functional areas such as airports, historic downtowns, and industrial employment clusters. This highlights that urban heat formation is shaped not only by land cover composition but also by socio-spatial dynamics and historical development patterns [62,63]. These tailored strategies, using GIS-based spatial modeling, can enhance both policy effectiveness and urban resilience while advancing the understanding of people-land-climate linkages.

4.3. Limitations and Future Research

This study aimed to prioritize the factors influencing LST in three Korean cities characterized by high LST due to their topographical and environmental conditions. XGBoost-SHAP was employed to determine the relative importance of variables, and GWR was applied to analyze spatial variations. Notably, by integrating GWR with the XGBoost-SHAP framework to enhance spatial interpretability, this study demonstrates methodological novelty that distinguishes it from previous research.
Nevertheless, this study has several limitations. First, the resolution of the analytical dataset was limited to a 1 km grid. This limits the ability to fully reflect fine-scale variability, such as the effects of fine green spaces, small buildings, and structures within the city. Second, the inability to directly incorporate LULC details into the analysis hindered clear attribution for specific patterns across heterogeneous land-cover types. Third, the study focused on three large metropolitan cities—Seoul, Daegu, and Incheon—providing insights specific to highly urbanized settings. Therefore, the findings may not be directly applicable to smaller or rural cities, and the localized nature of the study may restrict the universality of its conclusions. Future studies should expand the comparative scope to include cities of varying scales and morphological contexts to enhance generalizability.

5. Conclusions

The UHI effect is intensifying, and urban thermal environment policies and plans to mitigate this effect are becoming increasingly important. However, current greenspace policies tend to focus on uniformly expanding green space across the entire city, failing to adequately consider the diverse spatial characteristics of cities, such as mountainous areas, coastal areas, and urban centers. This approach undermines actual heat island mitigation effects and administrative performance.
To address this issue, a comprehensive analysis of the factors influencing the LST in each region is essential. This study analyzed the key factors influencing LST in three major South Korean cities with prominent heat island effects. XGBoost-SHAP was used to derive the relative importance of variables, and GWR assessed the spatial differences within the cities. The XGBoost-SHAP results revealed that NDBI was most influential on the LST in Seoul and Daegu, whereas POP_D had the greatest influence in Incheon. In an additional GWR analysis, NDVI_resid, NDBI, and ALBEDO exerted pronounced effects primarily in urban areas, whereas SLOPE and DEM showed greater influences in non-urban areas. POP_D was of limited importance owing to its small coefficient.
Collectively, the results indicate that NDBI, a proxy for building density, is the most influential predictor of the thermal environment across metropolitan areas. Therefore, urban structural improvement strategies should be prioritized when developing plans and policies to mitigate heat-island impacts. Furthermore, as revealed by the spatial regression, the effect of these variables varies by region, requiring customized plans that reflect regional characteristics. At the intra-urban scale, customized planning that reflects the distinct characteristics of urban cores and peripheral areas can enhance both policy effectiveness and urban resilience. For example, high-NDBI areas could implement building density regulations and reflective surface standards to reduce local heat accumulation. The findings of this study provide a foundation for developing strategies and policy decisions aimed at mitigating heat-island impacts, while also shedding light on how LULC configurations influence local thermal environments. Overall, this study demonstrates the methodological value of integrating machine learning and spatial regression for urban climate analysis, providing a replicable framework for assessing and mitigating urban heat across diverse city contexts.

Author Contributions

Conceptualization, H.J. and K.A.; writing—original draft preparation, H.J.; writing—review and editing, Y.S.; supervision, K.A. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data Availability Statement

The data presented in this study are available on request from the corresponding author.

Acknowledgments

This paper was supported by Konkuk University in 2025.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Chen, Y.-C.; Chiu, H.-W.; Su, Y.-F.; Wu, Y.-C.; Cheng, K.-S. Does urbanization increase diurnal land surface temperature variation? Evidence and implications. Landsc. Urban Plan. 2017, 157, 247–258. [Google Scholar] [CrossRef]
  2. Fu, P.; Weng, Q. A time series analysis of urbanization induced land use and land cover change and its impact on land surface temperature with Landsat imagery. Remote Sens. Environ. 2016, 175, 205–214. [Google Scholar] [CrossRef]
  3. Yin, C.; Yuan, M.; Lu, Y.; Huang, Y.; Liu, Y. Effects of urban form on the urban heat island effect based on spatial regression model. Sci. Total. Environ. 2018, 634, 696–704. [Google Scholar] [CrossRef]
  4. Heaviside, C.; Macintyre, H.; Vardoulakis, S. The urban heat island: Implications for health in a changing environment. Curr. Environ. Health Rep. 2017, 4, 296–305. [Google Scholar] [CrossRef]
  5. Ho, J.Y.; Shi, Y.; Lau, K.K.; Ng, E.Y.; Ren, C.; Goggins, W.B. Urban heat island effect-related mortality under extreme heat and non-extreme heat scenarios: A 2010–2019 case study in Hong Kong. Sci. Total. Environ. 2023, 858, 159791. [Google Scholar] [CrossRef]
  6. Bao, Y.; Li, Y.; Gu, J.; Shen, C.; Zhang, Y.; Deng, X.; Han, L.; Ran, J. Urban heat island impacts on mental health in middle-aged and older adults. Environ. Int. 2025, 199, 109470. [Google Scholar] [CrossRef]
  7. Mirzaei, M.; Verrelst, J.; Arbabi, M.; Shaklabadi, Z.; Lotfizadeh, M. Urban heat island monitoring and impacts on citizen’s general health status in Isfahan metropolis: A remote sensing and field survey approach. Remote Sens. 2020, 12, 1350. [Google Scholar] [CrossRef]
  8. Tun, K.Z.; Pramanik, M.; Mallick, S.K.; Chakrabortty, R.; Halder, B.; Moharir, K.N.; Pande, C.B.; Zhran, M. Cooling the Cities: A Comprehensive Review of Urban Heat Island Mitigation Strategies in Southeast Asia. Hum. Settl. Sustain. 2025, 1, 91–102. [Google Scholar] [CrossRef]
  9. Kumar, P.; Debele, S.E.; Khalili, S.; Halios, C.H.; Sahani, J.; Aghamohammadi, N.; Andrade, M.d.F.; Athanassiadou, M.; Bhui, K.; Calvillo, N.; et al. Urban heat mitigation by green and blue infrastructure: Drivers, effectiveness, and future needs. Innovation 2024, 5, 100588. [Google Scholar] [CrossRef] [PubMed]
  10. Lee, H.-K.; Lee, M.-G.; Joo, J.-H.; Kim, H.-B. A Cost-benefit Analysis for Urban Heat Island Mitigation Policy: Focused on the Installation of Green and Cool Roofs in Seoul, Korea. J. Korea Plan. Assoc. 2020, 55, 97–108. [Google Scholar] [CrossRef]
  11. Tran, D.X.; Pla, F.; Latorre-Carmona, P.; Myint, S.W.; Caetano, M.; Kieu, H.V. Characterizing the relationship between land use land cover change and land surface temperature. ISPRS J. Photogramm. Remote Sens. 2017, 124, 119–132. [Google Scholar] [CrossRef]
  12. Grover, A.; Singh, R.B. Analysis of urban heat island (UHI) in relation to normalized difference vegetation index (NDVI): A comparative study of Delhi and Mumbai. Environments 2015, 2, 125–138. [Google Scholar] [CrossRef]
  13. Jamal, S.; Saqib, M.; Ahmad, W.S.; Ahmad, M.; Ali, A.; Ali, B. Unraveling the complexities of land transformation and its impact on urban sustainability through land surface temperature analysis. Appl. Geomat. 2023, 15, 719–741. [Google Scholar] [CrossRef]
  14. Zhu, T.; Wang, X.; Luo, Y.; Qiu, H. Systematic Exploration of the Effect of Urban Green Space on Land Surface Temperature: A Scientometric Analysis. Buildings 2025, 15, 1032. [Google Scholar] [CrossRef]
  15. Chen, J.; Du, P.; Jin, S.; Ding, H.; Chen, C.; Xu, Y.; Feng, L.; Guo, G.; Zheng, H.; Huang, M. Unravelling the multilevel and multi-dimensional impacts of building and tree on surface urban heat islands. Energy Build. 2022, 259, 111843. [Google Scholar] [CrossRef]
  16. Peng, X.; Wu, W.; Zheng, Y.; Sun, J.; Hu, T.; Wang, P. Correlation analysis of land surface temperature and topographic elements in Hangzhou, China. Sci. Rep. 2020, 10, 10451. [Google Scholar] [CrossRef] [PubMed]
  17. Moradi, M.; Salahi, B.; Masoodian, S.A. On the relationship between MODIS Land Surface Temperature and topography in Iran. Phys. Geogr. 2018, 39, 354–367. [Google Scholar] [CrossRef]
  18. Njoku, E.A.; Tenenbaum, D.E. Quantitative assessment of the relationship between land use/land cover (LULC), topographic elevation and land surface temperature (LST) in Ilorin, Nigeria. Remote Sens. Appl. Soc. Environ. 2022, 27, 100780. [Google Scholar] [CrossRef]
  19. Zhou, W.; Huang, G.; Cadenasso, M.L. Does spatial configuration matter? Understanding the effects of land cover pattern on land surface temperature in urban landscapes. Landsc. Urban Plan. 2011, 102, 54–63. [Google Scholar] [CrossRef]
  20. Yue, W.; Xu, J.; Tan, W.; Xu, L. The relationship between land surface temperature and NDVI with remote sensing: Application to Shanghai Landsat 7 ETM+ data. Int. J. Remote Sens. 2007, 28, 3205–3226. [Google Scholar] [CrossRef]
  21. Malik, M.S.; Shukla, J.P.; Mishra, S. Relationship of LST, NDBI and NDVI using landsat-8 data in Kandaihimmat watershed, Hoshangabad, India. Indian J. Geo-Mar. Sci. 2019, 48, 25–31. [Google Scholar]
  22. Alademomi, A.S.; Okolie, C.J.; Daramola, O.E.; Akinnusi, S.A.; Adediran, E.; Olanrewaju, H.O.; Alabi, A.O.; Salami, T.J.; Odumosu, J. The interrelationship between LST, NDVI, NDBI, and land cover change in a section of Lagos metropolis, Nigeria. Appl. Geomat. 2022, 14, 299–314. [Google Scholar] [CrossRef]
  23. Wei, J.; Li, Y.; Jia, L.; Liu, B.; Jiang, Y. The Impact of Spatiotemporal Effect and Relevant Factors on the Urban Thermal Environment Through the XGBoost-SHAP Model. Land 2025, 14, 394. [Google Scholar] [CrossRef]
  24. Huang, C.; Liu, K.; Ma, T.; Xue, H.; Wang, P.; Li, L. Analysis of the impact mechanisms and driving factors of urban spatial morphology on urban heat islands. Sci. Rep. 2025, 15, 18589. [Google Scholar] [CrossRef] [PubMed]
  25. Ullah, W.; Ahmad, K.; Ullah, S.; Tahir, A.A.; Javed, M.F.; Nazir, A.; Abbasi, A.M.; Aziz, M.; Mohamed, A. Analysis of the relationship among land surface temperature (LST), land use land cover (LULC), and normalized difference vegetation index (NDVI) with topographic elements in the lower Himalayan region. Heliyon 2023, 9, e13322. [Google Scholar] [CrossRef]
  26. Guha, S.; Govil, H.; Gill, N.; Dey, A. Analytical study on the relationship between land surface temperature and land use/land cover indices. Ann. GIS 2020, 26, 201–216. [Google Scholar] [CrossRef]
  27. Kim, M.; Kim, D.; Kim, G. Examining the relationship between land use/land cover (LULC) and land surface temperature (LST) using explainable artificial intelligence (XAI) models: A case study of Seoul, South Korea. Int. J. Environ. Res. Public Health 2022, 19, 15926. [Google Scholar] [CrossRef]
  28. Hoang, N.-D.; Nguyen, Q.-L. Geospatial Analysis and Machine Learning Framework for Urban Heat Island Intensity Prediction: Natural Gradient Boosting and Deep Neural Network Regressors with Multisource Remote Sensing Data. Sustainability 2025, 17, 4287. [Google Scholar] [CrossRef]
  29. Gao, Y.; Zhao, J.; Han, L. Exploring the spatial heterogeneity of urban heat island effect and its relationship to block morphology with the geographically weighted regression model. Sustain. Cities Soc. 2022, 76, 103431. [Google Scholar] [CrossRef]
  30. Drapeau, L.-M.; Beaudoin, M.; Vandycke, L.; Brunstein, M. Urban Heat Island Mitigation Strategies; Institut National de Santé Publique du Québec: Québec, QC, Canada, 2023; Available online: https://www.inspq.qc.ca/pdf/publications/1513_UrbanHeatIslandMitigationStrategies.pdf (accessed on 2 October 2025).
  31. Sun, R.; Chen, L. Effects of green space dynamics on urban heat islands: Mitigation and diversification. Ecosyst. Serv. 2017, 23, 38–46. [Google Scholar] [CrossRef]
  32. Gunawardena, K.R.; Wells, M.J.; Kershaw, T. Utilising green and bluespace to mitigate urban heat island intensity. Sci. Total Environ. 2017, 584–585, 1040–1055. [Google Scholar] [CrossRef]
  33. Mishra, B.; Sandifer, J.; Gyawali, B.R. Urban heat island in Kathmandu, Nepal: Evaluating relationship between NDVI and LST from 2000 to 2018. Int. J. Environ. 2019, 8, 17–29. [Google Scholar] [CrossRef]
  34. Ko, Y.J.; Cho, K.-H.; Kim, W.-C. Analysis of Environmental Equity of Green Space Services in Seoul—The Case of Jung-gu, Seongdong-gu and Dongdaemun-gu. J. Korean Inst. Landsc. Arch. 2019, 47, 100–116. [Google Scholar] [CrossRef]
  35. Irfeey, A.M.M.; Chau, H.-W.; Sumaiya, M.M.F.; Wai, C.Y.; Muttil, N.; Jamei, E. Sustainable Mitigation Strategies for Urban Heat Island Effects in Urban Areas. Sustainability 2023, 15, 10767. [Google Scholar] [CrossRef]
  36. Sisman, S.; Aydinoglu, A.C. A modelling approach with geographically weighted regression methods for determining geographic variation and influencing factors in housing price: A case in Istanbul. Land Use Policy 2022, 119, 106183. [Google Scholar] [CrossRef]
  37. Wang, Z.; Zhao, Y.; Zhang, F. Simulating the spatial heterogeneity of housing prices in Wuhan, China, by regionally geographically weighted regression. ISPRS Int. J. Geo-Inf. 2022, 11, 129. [Google Scholar] [CrossRef]
  38. Huang, W.T.K.; Masselot, P.; Bou-Zeid, E.; Fatichi, S.; Paschalis, A.; Sun, T.; Gasparrini, A.; Manoli, G. Economic valuation of temperature-related mortality attributed to urban heat islands in European cities. Nat. Commun. 2023, 14, 7438. [Google Scholar] [CrossRef]
  39. Liu, Q.; Zhang, H.; Peng, J.; Li, Q.; Liu, T.; Hu, F.; Huang, H.; Shen, Y.; Ye, B.; Guo, Y.; et al. Mitigation strategies for regional heat island effects: Integrated network construction of cold and heat islands. Ecol. Indic. 2025, 176, 113703. [Google Scholar] [CrossRef]
  40. Gupta, N.; Aithal, B.H. Urban land surface temperature forecasting: A data-driven approach using regression and neural network models. Geocarto Int. 2024, 39, 2299145. [Google Scholar] [CrossRef]
  41. Zhi, Y.; Shan, L.; Ke, L.; Yang, R. Analysis of land surface temperature driving factors and spatial heterogeneity research based on geographically weighted regression model. Complexity 2020, 2020, 2862917. [Google Scholar] [CrossRef]
  42. Oukawa, G.Y.; Krecl, P.; Targino, A.C. Fine-scale modeling of the urban heat island: A comparison of multiple linear regression and random forest approaches. Sci. Total. Environ. 2022, 815, 152836. [Google Scholar] [CrossRef]
  43. Kaggle, R. State of Data Science and Machine Learning 2021; Kaggle: San Francisco, CA, USA, 2020. [Google Scholar]
  44. Lundberg, S.M.; Erion, G.; Chen, H.; DeGrave, A.; Prutkin, J.M.; Nair, B.; Katz, R.; Himmelfarb, J.; Bansal, N.; Lee, S.-I. From local explanations to global understanding with explainable AI for trees. Nat. Mach. Intell. 2020, 2, 56–67. [Google Scholar] [CrossRef]
  45. Batunacun; Wieland, R.; Lakes, T.; Nendel, C. Using SHAP to interpret XGBoost predictions of grassland degradation in Xilingol, China. Geosci. Model Dev. Discuss. 2020, 2020, 1–28. [Google Scholar]
  46. City, D.M. Daegu Metropolitan City GIS Portal. 2025. Available online: http://gis.go.kr/intro.do (accessed on 24 September 2025).
  47. City, I.M. Incheon Metropolitan City GIS Portal. 2025. Available online: https://www.incheon.go.kr/index (accessed on 24 September 2025).
  48. Government, S.M. Seoul Metropolitan Government GIS Portal. 2025. Available online: https://www.seoul.go.kr/main/index.jsp (accessed on 24 September 2025).
  49. WorldPop. WorldPop Data Hub. 2025. Available online: https://hub.worldpop.org (accessed on 29 September 2025).
  50. Chen, T.; Guestrin, C. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016. [Google Scholar]
  51. Ma, X.; Yang, J.; Zhang, R.; Yu, W.; Ren, J.; Xiao, X.; Xia, J. XGBoost-based analysis of the relationship between urban 2-D/3-D morphology and seasonal gradient land surface temperature. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2024, 17, 4109–4124. [Google Scholar] [CrossRef]
  52. Song, S.; Shi, J.; Fan, D.; Cui, L.; Yang, H. Development of downscaling technology for land surface temperature: A case study of Shanghai, China. Urban Clim. 2025, 61, 102412. [Google Scholar] [CrossRef]
  53. Wang, M.; Li, Y.; Yuan, H.; Zhou, S.; Wang, Y.; Ikram, R.M.A.; Li, J. An XGBoost-SHAP approach to quantifying morphological impact on urban flooding susceptibility. Ecol. Indic. 2023, 156, 111137. [Google Scholar] [CrossRef]
  54. Li, S.; Wong, M.S.; Zhu, R.; Shi, G.; Yang, J. Impacts of land surface temperature and ambient factors on near-surface air temperature estimation: A multisource evaluation using SHAP analysis. Sustain. Cities Soc. 2025, 122, 106257. [Google Scholar] [CrossRef]
  55. Kuhn, H.W. Classics in Game Theory; Princeton University Press: Princeton, NJ, USA, 2020; 328p, Available online: https://www.torrossa.com/en/resources/an/5641636#page=87 (accessed on 10 October 2025).
  56. Zhang, J.; Ma, X.; Zhang, J.; Sun, D.; Zhou, X.; Mi, C.; Wen, H. Insights into geospatial heterogeneity of landslide susceptibility based on the SHAP-XGBoost model. J. Environ. Manag. 2023, 332, 117357. [Google Scholar] [CrossRef]
  57. Jia, H.; Yang, D.; Deng, W.; Wei, Q.; Jiang, W. Predicting land surface temperature with geographically weighed regression and deep learning. WIREs Data Min. Knowl. Discov. 2021, 11, e1396. [Google Scholar] [CrossRef]
  58. Li, G.; Zhang, X.; Mirzaei, P.A.; Zhang, J.; Zhao, Z. Urban heat island effect of a typical valley city in China: Responds to the global warming and rapid urbanization. Sustain. Cities Soc. 2018, 38, 736–745. [Google Scholar] [CrossRef]
  59. Wu, Y.; Wu, H.; Li, S.; Wu, L.; Guo, C.; Zhang, Z.; Fan, R.; Yang, X.; Zhou, G.; Yang, J.; et al. Integrating urban development and ecological sustainability: A study on land surface temperature and urban heat island effect in a Karst River Basin. Front. Built Environ. 2024, 10, 1457347. [Google Scholar] [CrossRef]
  60. Lu, Y.; Qin, Y.; Huang, C.; Pang, X. Albedo of Pervious Concrete and Its Implications for Mitigating Urban Heat Island. Sustainability 2023, 15, 8222. [Google Scholar] [CrossRef]
  61. Smith, I.A.; Fabian, M.P.; Hutyra, L.R. Urban green space and albedo impacts on surface temperature across seven United States cities. Sci. Total. Environ. 2023, 857, 159663. [Google Scholar] [CrossRef] [PubMed]
  62. Ivajnšič, D.; Kaligarič, M.; Žiberna, I. Geographically weighted regression of the urban heat island of a small city. Appl. Geogr. 2014, 53, 341–353. [Google Scholar] [CrossRef]
  63. Zhang, H.; Qi, Z.-F.; Ye, X.-Y.; Cai, Y.-B.; Ma, W.-C.; Chen, M.-N. Analysis of land use/land cover change, population shift, and their effects on spatiotemporal patterns of urban heat islands in metropolitan Shanghai, China. Appl. Geogr. 2013, 44, 121–133. [Google Scholar] [CrossRef]
Land 14 02232 g001
Figure 1. Study workflow.
Figure 1. Study workflow.
Land 14 02232 g001
Land 14 02232 g002
Figure 2. Climatic and geographical characteristics of the study areas.
Figure 2. Climatic and geographical characteristics of the study areas.
Land 14 02232 g002
Land 14 02232 g003aLand 14 02232 g003b
Figure 3. Spatial patterns of LST and socio-environmental variables along the intra-urban land use gradient.
Figure 3. Spatial patterns of LST and socio-environmental variables along the intra-urban land use gradient.
Land 14 02232 g003aLand 14 02232 g003b
Land 14 02232 g004
Figure 4. SHAP results for each study area.
Figure 4. SHAP results for each study area.
Land 14 02232 g004
Land 14 02232 g005aLand 14 02232 g005b
Figure 5. SHAP dependence plots.
Figure 5. SHAP dependence plots.
Land 14 02232 g005aLand 14 02232 g005b
Land 14 02232 g006
Figure 6. Spatial distribution of GWR coefficients by key influencing variables in Daegu. Red areas indicate regions with higher LST influence (positive coefficients), whereas blue areas indicate regions with lower LST influence (negative coefficients).
Figure 6. Spatial distribution of GWR coefficients by key influencing variables in Daegu. Red areas indicate regions with higher LST influence (positive coefficients), whereas blue areas indicate regions with lower LST influence (negative coefficients).
Land 14 02232 g006
Land 14 02232 g007
Figure 7. Spatial distribution of GWR coefficients by key influencing variables in Incheon.
Figure 7. Spatial distribution of GWR coefficients by key influencing variables in Incheon.
Land 14 02232 g007
Land 14 02232 g008
Figure 8. Spatial distribution of GWR coefficients by key influencing variables in Seoul.
Figure 8. Spatial distribution of GWR coefficients by key influencing variables in Seoul.
Land 14 02232 g008
Table 1. Summary of key variables and findings from previous studies.
Table 1. Summary of key variables and findings from previous studies.
Study (Year)—Study AreaVariablesMain Results
Hoang, N.D., and Nguyen, Q.L. (2025) [28]—Da NangElevation, slope, aspect, TPI, distance to coastlines, distance to river, distance to wetlands, land use, land coverBuilt-up density has the greatest impact
Wei Junqing et al. (2025) [23]—SeoulPOP_D, NTL, RD, BD, BH, DEM, NDVI, NDBSI, WET, PLAND, LPI, ED, CA,
CONTAG		Daytime: DEM has the greatest effect in both urban and rural areas
Nighttime: POP_D has the greatest impact in both urban and rural areas
Huang, Caiyi et al. (2025) [24]—TianjinBH, BHD, BD, SVF, NDVI, NDWI, PIS, PGS, ALBEDO, NLINDVI has the greatest effect
Ullah, Waheed, et al.
(2023) [25]—Northern Pakistan		DEM, LULC, NDVIDEM and NDVI negatively
correlate with LST
Built-up areas:
larger area → higher mean LST
Water bodies:
larger area → lower mean LST
Kim, M., Kim, D., and Kim, G. (2022) [27]—SeoulNDBI, NDWI, GNDVI, DEM, slope, LULC NDBI has the greatest effect
Guha, Subhanil, et al.
(2020) [26]—Raipur		NDVI, NDWI, NDBI, NMDINDBI positively correlates with LST
NDVI, NDWI, and NMDI negatively correlate with LST
Table 2. Summary of key variables and findings from prior work.
Table 2. Summary of key variables and findings from prior work.
YearDaeguIncheonSeoul
20007th (30.65 °C)3rd (31.22 °C)2nd (31.87 °C)
20058th (30.41 °C)5th (30.78 °C)2nd (31.28 °C)
20107th (31.04 °C)6th (31.42 °C)4th (31.68 °C)
20158th (31.67 °C)3rd (32.45 °C)7th (32.06 °C)
202011th (30.96 °C)7th (31.30 °C)5th (31.89 °C)
Table 3. VIFs by variables.
Table 3. VIFs by variables.
VariableDaeguIncheonSeoul
NDVIAfter (Before)2.624925 (13.874368)2.016321 (13.874368)3.818013 (25.984242)
NDBIAfter (Before)2.133849 (10.337122)7.401509 (10.337122)4.115698 (19.280302)
ALBEDOAfter (Before)2.422610 (1.932140)1.932140 (1.932140)3.243649 (3.243649)
POP_DAfter (Before)1.368502 (2.588510)2.588510 (2.588510)2.136446 (2.136446)
DEMAfter (Before)5.426983 (2.9848)2.9848 (2.9848)5.867107 (5.867107)
SLOPEAfter (Before)7.527535 (8.125609)8.125609 (8.125609)7.724673 (7.724673)
Table 4. Verification of XGBoost predictive performance.
Table 4. Verification of XGBoost predictive performance.
MetricDaeguIncheonSeoul
RMSE1.0421.3751.683
R20.9660.8560.727
Table 5. Ranking of variables affecting LST.
Table 5. Ranking of variables affecting LST.
Metric1st2nd3rd4th5th6th
DaeguNDBI
(2.649)		SLOPE
(1.144)		POP_D
(0.879)		DEM
(0.707)		NDVI_resid
(0.329)		ALBEDO
(0.251)
IncheonPOP_D
(1.337)		NDBI
(1.026)		ALBEDO
(0.776)		SLOPE
(0.544)		DEM
(0.478)		NDVI_resid
(0.454)
SeoulNDBI
(1.293)		DEM
(0.873)		NDVI_resid
(0.696)		ALBEDO
(0.308)		SLOPE
(0.297)		POP_D
(0.259)
Table 6. Validation of GWR predictive performance.
Table 6. Validation of GWR predictive performance.
MetricDaeguIncheonSeoul
RMSE1.1431.2941.428
R20.9580.8890.813
Table 7. Summary of GWR coefficient estimates for Daegu.
Table 7. Summary of GWR coefficient estimates for Daegu.
VariablesMin1st Qu.Median3rd Qu.Max
NDVI_resid−41.002−19.564−1.7136.93631.712
NDBI3.89818.5325.13628.70944.116
ALBEDO−53.34824.3361.79107.34203.893
POP_D−0.00009890.0001540.0002630.0004510.0023
DEM−0.0249−0.00809−0.00616−0.003340.0131
SLOPE−0.526−0.1−0.0510.0320.365
Table 8. Summary of GWR coefficient estimates for Incheon.
Table 8. Summary of GWR coefficient estimates for Incheon.
VariablesMin1st Qu.Median3rd Qu.Max
NDVI_resid−29.0079.96423.69835.55871.031
NDBI−19.56111.89819.33332.70179.567
ALBEDO−63.3498.63629.75563.709163.542
POP_D−0.005259−0.00008120.00054860.0030460.0301
DEM−0.18748−0.05811−0.02540−0.012730.0106
SLOPE−0.4935−0.1408−0.05820.24041.404
Table 9. Summary of GWR coefficient estimates for Seoul.
Table 9. Summary of GWR coefficient estimates for Seoul.
VariablesMin1st Qu.Median3rd Qu.Max
NDBI1.486419.637025.690033.345057.5625
SLOPE−0.37810.00560.12630.20910.4945
POP_D−0.0001270.0000270.0000360.0000880.0002
DEM−0.0772−0.0410−0.0300−0.01550.0187
NDVI_resid−16.99313.262038.481049.932073.2679
ALBEDO−32.42127.504053.776072.1910139.842
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Jeong, H.; Shin, Y.; An, K. City-Specific Drivers of Land Surface Temperature in Three Korean Megacities: XGBoost-SHAP and GWR Highlight Building Density. Land 2025, 14, 2232. https://doi.org/10.3390/land14112232

AMA Style

Jeong H, Shin Y, An K. City-Specific Drivers of Land Surface Temperature in Three Korean Megacities: XGBoost-SHAP and GWR Highlight Building Density. Land. 2025; 14(11):2232. https://doi.org/10.3390/land14112232

Chicago/Turabian Style

Jeong, Hogyeong, Yeeun Shin, and Kyungjin An. 2025. "City-Specific Drivers of Land Surface Temperature in Three Korean Megacities: XGBoost-SHAP and GWR Highlight Building Density" Land 14, no. 11: 2232. https://doi.org/10.3390/land14112232

APA Style

Jeong, H., Shin, Y., & An, K. (2025). City-Specific Drivers of Land Surface Temperature in Three Korean Megacities: XGBoost-SHAP and GWR Highlight Building Density. Land, 14(11), 2232. https://doi.org/10.3390/land14112232

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop