The Impact of Spatiotemporal Effect and Relevant Factors on the Urban Thermal Environment Through the XGBoost-SHAP Model

Abstract

The urban thermal environment is a critical topic in contemporary urban studies. However, the mechanisms driving the relationships between influencing factors and the urban thermal environment across different spatial scales and temporal dimensions remain unclear, particularly as most of these relationships exhibit nonlinearity. This study utilizes XGBoost and SHAP models, combined with a partial dependency plot, to analyze the influence of population activities, built environment, urban topography, ecological and climatic conditions, and urban landscape pattern on the diurnal and nocturnal land surface temperature (LST) changes across urban and rural areas of Hangzhou throughout the year. The results indicate that during the daytime, urban topography exerts a strong influence on LST changes in both urban and rural areas of Hangzhou. At nighttime, the influence of population activities becomes more pronounced. Meanwhile, urban landscape patterns show no significant impact on LST in either urban or rural areas, regardless of daytime or nighttime. Additionally, we analyzed the specific nonlinear relationships between influencing factors and LST. Finally, our findings suggest that influencing factors can interact synergistically in pairs to affect LST, with this mechanism being more prominent in urban areas. Overall, the study categorizes and examines the factors contributing to urban thermal environment changes from spatial and temporal perspectives, providing insights for developing urban planning strategies to mitigate urban heat issues in the future.

1. Introduction

In recent years, extreme heat events have become increasingly frequent in major cities around the world, leading researchers to increasingly focus on the challenges associated with urban heat [1,2,3,4,5]. In this context, to control the spread of urban heat, some researchers have focused on the trends in LST or urban heat island (UHI) changes and the factors driving these variations [6,7,8,9,10]. The aim is to identify relevant patterns that can guide future urban planning to mitigate the urban heat phenomenon. Based on this, we review studies in this field from three main perspectives. The first focuses on the factors influencing urban heat issues, the second examines the application of relevant research methods to urban heat studies, and the third addresses spatial scale considerations in the study areas.

In studies of urban heat issues, most researchers choose variables such as LST, SUHI (surface urban heat island), UHI, etc. to reflect the extent of urban heat. Among them, LST represents the temperature of surface objects (such as buildings, vegetation, roads, etc.). It is influenced by solar radiation, and exhibits significant diurnal variation, making it suitable for analyzing urban heat issues from a temporal perspective. SUHI focuses on the difference in LST between urban and rural areas (ΔLST = LST_urban area − LST_rural area). It reflects the extent of surface temperature change caused by urbanization and is a relative measure. On the other hand, UHI indicates the temperature difference between the urban and rural areas’ air temperatures. Therefore, UHI variations are also influenced by factors such as pollutants and atmospheric wind speed. This study focuses on reflecting urban heat issues through the variation in LST.

The factors influencing LST variation are multifaceted. First, changes in land use are particularly significant [11,12]. During urbanization, extensive use of impervious surfaces such as asphalt and concrete results in areas that absorb more heat but have lower heat dissipation efficiency compared to natural surfaces [13,14]. As such, these surfaces contribute to the accumulation of heat, making it difficult for it to dissipate, particularly in urban centers [15]. Second, the layout of buildings and urban landscapes also exert an influence on LST. At the building level, densely clustered high-rise buildings form urban canyons, restricting airflow and facilitating the accumulation of heat within these canyons, making it harder for heat to disperse [16,17]. At the landscape level, different landscape patterns contribute to the formation of urban cool corridors. Recent studies have shown that scholars have employed multi-layer networks combined with Morphological Spatial Pattern Analysis (MSPA) to identify cold and heat source patches in cities based on different landscape forms. This research has, to some extent, uncovered the urban cooling network and established the role of landscape morphology in mitigating urban heat [18,19]. Third, population activity and climate factors also play crucial roles in the formation of urban heat. In terms of population, high-density areas typically indicate an increase in energy consumption and waste heat emissions [20]. High population density often exacerbates air pollution and carbon dioxide emissions. These emissions affect atmospheric conditions and are related to the intensification of the urban heat [21]. As for climate factors, temperature and moisture are significant contributors to urban heat [22]. During the day, high temperatures lead to stronger solar radiation absorption by the surface, while the heat is gradually released at night, slowing down the cooling process [23]. In conclusion, current research explores the mechanisms through single sets of factors affecting urban heat issues and insufficient attention has been given to the influence of multidimensional factors.

In terms of research methodologies, scholars have traditionally relied on linear regression models to identify the factors contributing to urban heat issues [24,25]. Some researchers have also employed spatial error or spatial lag models to explore factors influencing urban heat, leveraging the spatial autocorrelation observed in Moran’s I analysis of urban heat patterns [26]. However, these models often struggle to capture the nonlinear interactions present among complex datasets. Such nonlinear mechanisms play a critical role in determining the effects among research variables, making the representation of nonlinear regression relationships particularly significant. Currently, machine learning offers substantial advantages in addressing nonlinear research problems [26,27]. It not only handles multidimensional data through its robust function-fitting capabilities but also mitigates overfitting by increasing the volume of training data [28,29,30].

On the spatial scale, existing research adopts an urban–rural dichotomy to investigate the mechanisms underlying changes in urban heat. The degree of urban heat aggregation in urban areas differs from that in rural areas, with pronounced temperature gradients often observed at the urban–rural interface [31,32]. Thus, spatial scale delineation is essential to facilitate the development of more targeted policy recommendations for urban heat mitigation in future planning and design. In recent studies, scholars have examined the urban–rural dual structure in Shanghai to explore how influencing factors impacted urban LST changes between 2015 and 2020 [33]. Other studies have taken a broader perspective, conducting comparative analyses of the urban heat outcomes across different urban and rural contexts within larger metropolitan regions [34].

Synthesizing the research gaps identified in studies on urban heat issues, we propose that this study should align with these research trends and employ machine learning to investigate the multidimensional mechanisms influencing urban heat. Specifically, the research focuses on the following aspects: (1) Identifying the mechanisms driving the spatial distribution of urban heat across five dimensions: population activity, built environment, urban topography, ecological climate, and landscape pattern. (2) Utilizing the XGBoost model and explanatory models to analyze the complex spatial relationships between influencing factors and the urban heat environment. (3) Dividing the study area into urban and rural spaces based on an urban–rural dichotomy and exploring these two distinct spatial types. (4) Considering temporal influences in addition to spatial scales by analyzing and discussing the variations in the urban heat environment during different times of the day, specifically daytime and nighttime. This comprehensive approach aims to advance our understanding of the multifaceted factors shaping urban heat dynamics.

2. Materials and Methods

2.1. Study Area

The study area for this research is Hangzhou, China. As a megacity in Asia, Hangzhou has a population of 12.52 million and covers a total area of 16,850 square kilometers. Located in northern Zhejiang Province, along the lower reaches of the Qiantang River, Hangzhou experiences a subtropical monsoon climate. Influenced by climate change and rapid urban expansion, the city’s annual average temperature has shown a gradual upward trend in recent years. In 2022, the average annual temperature reached 18.8 °C, the highest recorded since the establishment of the Hangzhou meteorological station.

To address the frequent extreme heat events in recent years [35], the Hangzhou municipal government has introduced policies such as the National Carbon Peaking (Hangzhou) Implementation Plan and the Hangzhou Measures for Climate Resource Protection and Utilization. Additionally, meteorological authorities have conducted urban heat island effect assessments and emphasized the need for coordination among various levels of urban management organizations.

In previous studies on urban heat, scholars have generally recognized a temperature difference between urban built-up areas and rural areas. This temperature disparity typically occurs at the urban–rural interface. The delineation of UHI study areas is often based on the urban–rural dichotomy. The definitions of urban areas and rural areas can be understood from multiple dimensions. First, some studies define the distinction between urban and rural areas based on population distribution and different countries and regions set different population thresholds for cities and rural areas [36]. Second, urban and rural areas can be distinguished by land use, as urban land types are more diverse compared to rural areas [37]. Third, the level of infrastructure development, particularly in terms of transportation, commerce, and cultural and recreational facilities, can also serve as a criterion [38]. Fourth, administrative boundaries can be used for classification [39]. In this study, the urban–rural dichotomy is based on administrative boundaries. Given the degree of land development, we believe this classification approach is reasonable.

Accordingly, we divided the study area into two zones—urban areas and rural areas—for separate analysis. Based on the current administrative divisions of Hangzhou, the urban area includes ten districts: Shangcheng, Gongshu, Xihu, Binjiang, Xiaoshan, Yuhang, Linping, Qiantang, Fuyang, and Lin’an. The rural area consists of Tonglu, Chun’an, and Jiande (Figure 1).

2.2. Data Source

This study investigates the mechanisms driving changes in the urban thermal environment under the influence of multidimensional factors. The dependent variable is LST. The independent variable is urban heat and research categories include data on population activities, built environment features, urban topography, ecological and climatic conditions, and urban landscape pattern indices. These datasets encompass the five key dimensions identified in existing research as influencing urban heat issues [40,41].

We obtained the daytime and nighttime LST data using the MOD11A1.061 Terra Land Surface Temperature/Emissivity Daily Global 1 km product. This product is provided by NASA and is derived from the MODIS sensor onboard the Terra satellite. It provides daytime (LST_Day) and nighttime (LST_Night) LST data corresponding to the overpass times of the Terra satellite over Hangzhou. Specifically, the daytime LST (LST_Day) is acquired at approximately 10:30 AM (local solar time), while the nighttime LST (LST_Night) is acquired at approximately 10:30 PM (local solar time). The data were obtained from the Google Earth Engine (GEE) platform (https://code.earthengine.google.com/). In this study, we used this platform to extract the annual average daytime and nighttime LST for Hangzhou in 2022 as the dependent variable for our analysis.

The measurements for the influencing factors mentioned above are as follows: LST, population density (POP_D), nighttime light (NTL), road density (RD), building density (BD), building height (BH), Digital Elevation Model (DEM), slope, aspect, Normalized Difference Vegetation Index (NDVI), Normalized Difference Build-up and Soil Index (NDBSI), wetness (Wet), Patch Landscape (PLAND), Largest Patch Index (LPI), Edge Density (ED), Class Area (CA), and Contagion Index (CONTAG). Population data were sourced from WorldPop. Nighttime light data were derived from the dataset published by Chen [42] (https://www.escience.org.cn), with masking applied to exclude water bodies. Road data were obtained from OpenStreetMap (OSM). Building density and building height data were sourced from the Tibetan Plateau Data Center (https://data.tpdc.ac.cn/zh-hans/data/60dac98d-eec4-41df-9ad5-b1563e5c532c/) and the Zenodo platform (https://doi.org/10.5281/zenodo.7827315, accessed on 14 April 2023), respectively. Urban topography data were obtained from the NASA Earth Science Data website (https://earthexplorer.usgs.gov/), specifically high-resolution terrain data captured by the ALOS satellite. From this dataset, DEM, slope, and aspect data were computed. For ecological remote sensing indices (RESIs), we selected NDVI, wet, and NDBSI as core indicators. These indices were calculated using Google Earth Engine (GEE) (https://code.earthengine.google.com), with “USGS Landsat8 Level 2, Collection 2, Tier 1” used as the dataset. Given its temporal resolution of 8 days, we calculated the average values of RESI indices over one year, using 8-day intervals as the time cycle for analysis. Landscape pattern indices were calculated using Fragstats 4.3. The required land cover data were derived from the 30 m annual land cover raster dataset for China (1990–2022) published by Professors Jie Yang and Xin Huang of Wuhan University (https://zenodo.org/record/8176941, accessed on 1 August 2023) (

Table 1. Summary of the data source.

Research Categories	Research Measurement	Data Source
Population activities	POP_D	https://hub.worldpop.org/geodata/summary?id=131
	NTL	https://www.escience.org.cn
Built environment	RD	https://www.openstreetmap.org/
	BD	https://data.tpdc.ac.cn/zh-hans/data/60dac98d-eec4-41df-9ad5-b1563e5c532c/
	BH	https://doi.org/10.5281/zenodo.7827315
Urban topography	DEM	https://earthexplorer.usgs.gov/
	Slope
	Aspect
Ecological and climatic conditions	Wet	https://code.earthengine.google.com
	NDBSI
	NDVI
Urban landscape pattern	PLAND	https://zenodo.org/record/8176941
	LPI
	ED
	CA
	CONTAG

).

To integrate the above datasets, projection and clipping were performed using the standardized coordinate system WGS_1984_UTM_Zone_50N. Since the spatial size and extent of the urban area and rural area are similar, it was unnecessary to account for heterogeneity in research granularity caused by spatial expansion. Thus, a uniform grid of 2 km × 2 km was selected as the analytical unit. Specifically, the urban area contains 2090 grids, the rural area contains 2123 grids, and the total number of grids across the global area is 4213.

2.3. Research Model

XGBoost (Extreme Gradient Boosting) is a gradient boosting framework widely applied in machine learning tasks. Its key advantage lies in its ability to handle nonlinear relationships effectively while providing a certain degree of model interpretability. In analyzing the multifactorial influences on urban LST, XGBoost demonstrates strong nonlinear fitting capabilities, enabling a more intuitive understanding of the interaction mechanisms among variables. The loss function of XGBoost consists of an empirical loss term and a regularization term, which are designed to measure the discrepancy between the model’s predictions and the true values, as well as the complexity of the model:

$$obj\left(\theta \right)=\sum _{i=1}^{n}L\left(y_i,{\widehat{y}}_i\right)+\sum _{k=1}^K\Omega \left(f_k\right)$$

(1)

In this context, ${\sum }_{i=1}^{n}L\left(y_i,{\widehat{y}}_i\right)$ represents the empirical loss term and it quantifies the loss between the predicted values and the true values of the training data; ${\sum }_{k=1}^K\Omega \left(f_k\right)$ denotes the regularization term, which represents the sum of the complexities of all t trees and is used to control model overfitting. $L\left(y_i,{\widehat{y}}_i\right)$ represents the model parameters, $\Omega \left(f_k\right)$ denotes the loss value of the i-th sample, and $\theta$ indicates the complexity of the K-th base model.

SHAP (Shapley additive explanation) is a method for interpreting the predictions of machine learning models. It is based on the Shapley value, a concept from cooperative game theory, to compute each feature’s contribution to the model’s predictions. This approach ensures fairness in analyzing all research data and quantifies each feature’s contribution to the model’s outcomes. Furthermore, SHAP adheres to the principle of consistency, meaning that if a feature’s actual impact increases, its SHAP value will not decrease, ensuring the stability and reliability of interpretations. SHAP also considers all possible feature combinations, enhancing the comprehensiveness of the analysis. Its derivation is based on a set $N$ composed of features and a specific feature $i$. The Shapley value measures the contribution of feature $i$ to the model’s output. For each feature $i$, the Shapley value is defined as:

$${\phi }_i=\sum _{S\subseteq N\setminus \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}}·\left[\nu \left(S\cup \left\{i\right\}\right)-\nu \left(S\right)\right]$$

(2)

Here, $N$ represents the set of all features, $S\subseteq N\setminus \left\{i\right\}$ denotes a subset that does not include feature $i$, and $\nu \left(S\right)$ is the value function of the feature set $S$, which is typically the model’s prediction output in machine learning. $\left|S\right|$ represents the size of the subset $S$.

When XGBoost is combined with SHAP, the efficiency of the algorithm is significantly improved, particularly in calculating baseline values, assessing the marginal contributions of features, and analyzing the interactive effects of multiple features.

Model Evaluation Parameters are crucial for validating the effectiveness of machine learning models. For this study, we selected the mean squared error ($MSE$) and coefficient of determination ($R^2$) as evaluation metrics for the regression task. The dataset was split into 80% for the training set and 20% for the testing set to ensure robust model evaluation.

$$MSE={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\left(\widehat{y_i}{-y}_i\right)}^2$$

(3)

$$R^2=1-{\displaystyle \frac{{\sum }_i{\left(\widehat{y_i}{-y}_i\right)}^2}{\sum _i{\left(\overline{y_i}{-y}_i\right)}^2}}$$

(4)

In this context, $n$ represents the number of samples. $\widehat{y_i}$ denotes the predicted value for the $i$-th sample, while $y_i$ represents its true value. A lower $MSE$ corresponds to a higher$R^2$. A higher $R^2$signifies the better performance of the urban thermal environment model.

In this research, predicting the target variable (LST) using the XGBoost regression model, the dataset is first randomly split into training and testing sets, with 80% of the data used for training and 20% for testing. Then, the grid search CV with 5-fold cross-validation is applied to find the optimal combination of hyperparameters. Finally, the best model is used to make predictions on the test set, and R² and RMSE are computed as performance metrics. In the SHAP plot, features are ranked by their importance, with the most important features displayed at the top. The direction of the SHAP values (positive or negative) indicates the impact of each feature on the model’s prediction, reflecting whether the feature contributes positively or negatively. The magnitude of the SHAP values represents the degree of influence a feature has on the prediction—larger values indicate a greater impact. The plot type is a scatter plot. It visualizes the SHAP values for each sample on a given feature.

2.4. LST Intensity

The 2022 urban surface temperature data for Hangzhou were standardized, and the surface temperature intensity was calculated separately for morning and evening throughout the year. After standardization, urban surface temperature levels were classified to analyze the spatial distribution changes in the heat island effect. The formula for surface temperature standardization is as follows:

$${LST}_{std}={\displaystyle \frac{{LST}_i-{LST}_{min}}{{LST}_{max}-{LST}_{min}}}$$

(5)

LST_std is the standardized LST. LST_min is the minimum LST across all study units. LST_max is the maximum LST across all study units and LST_i is the LST of a specific study unit. The mean standard deviation method was employed to classify surface thermal field levels. This method combines the mean LST (μ) and multiples of its standard deviation (std) as thresholds to define LST levels. Based on the surface temperature distribution in Hangzhou and practical requirements, the surface thermal field was divided into five levels using the thresholds μ ± 0.5 std and μ ± std. This approach enables systematic classification of surface temperature levels. The classification of pixels allows for a more intuitive understanding of the spatial distribution of high and low LST areas in Hangzhou.

3. Results

3.1. The Spatial Distribution of LST Intensity

During the daytime, the high land surface temperature areas (HSTAs) in Hangzhou are concentrated in the urban area, with significant hotspots in Gongshu, Yuhang, Shangcheng, Binjiang, Xiaoshan, and Linping. In the rural area, HSTA is localized to a few regions, aligned with linear infrastructure. Other areas are dominated by middle-temperature areas (MTAs). Low land surface temperature areas (LSTAs) are distributed in rural regions, forming patches, particularly along the urban fringe (Figure 2a).

At nighttime, HSTA is distributed across both the urban area and rural area. In the urban area, the spatial pattern of HSTA shifts compared to daytime, becoming more dispersed. Key hotspots include Gongshu, Shangcheng, Xihu, Xiaoshan, and Binjiang. In the rural area, HSTA shows significant clustering around the large water bodies in Chunan. Other rural areas are primarily characterized by MTA distribution. LSTA remains concentrated in the rural area, at the junctions of Tonglu, Chunan, and Jiande. Within the urban area, extensive LSTA can be found in the center and northwest periphery of Linan (Figure 2b).

3.2. The Spatial Distribution of Relevant Research Measurements

For ecological and climatic factors, the distribution of NDVI shows a pattern of lower values in the northeast and higher values in the southwest, indicating that dense vegetation cover is concentrated in rural areas. Overall, vegetation density increases progressively from the urban center to the suburban and rural areas (Figure 3a). NDBSI exhibits stronger values in the city center, the northwestern edge, and along the linear infrastructure connecting urban and rural areas (Figure 3b). Regions with higher wet values are concentrated in the city center and near large lakes in the rural area (Figure 3c).

For the built environment, the selected measurements for the built environment are BD, BH, and RD. For BD, high-density building areas are mainly located in the urban area, particularly in Shangcheng, Gongshu, Xihu, Binjiang, and Xiaoshan (Figure 3d). In the rural area, Chun’an exhibits relatively high-density built-up areas. Regarding BH (Figure 3e), the distribution is more dispersed compared to BD but remains concentrated in the urban area, especially in Shangcheng, Gongshu, Xihu, Binjiang, and Xiaoshan. For RD, high road density is observed throughout all urban areas and along the linear infrastructure connecting urban and rural regions, with sparse distribution in other areas (Figure 3f).

For urban topography, the distribution of DEM is concentrated in the northwestern and southwestern edges of Hangzhou, with some areas in the rural region (Figure 3g). Slope is distributed in rural areas, while the central urban area is largely flat (Figure 3h). The aspect shows a random distribution across the city (Figure 3i).

For landscape pattern, the selected indicators for landscape pattern analysis include PLAND, LPI, CA, SHEI, ED, and CONTAG. Results indicate that large landscape patches are primarily concentrated in the rural areas and the urban areas of Fuyang and Lin’an. PLAND, LPI, and CA exhibit high spatial similarity, while ED and CONTAG also show strong spatial similarity. According to SHEI, the distribution of landscape diversity in Hangzhou is relatively uniform, displaying an overall even and random pattern (Figure 3i–o).

For population activities, the selected measurements for socioeconomic factors are PD and NTL. PD is concentrated in the urban area, with Shangcheng and Gongshu exhibiting the highest density (Figure 3p). Similarly, NTL is mainly clustered in urban areas, though it also aligns with linear infrastructure (Figure 3q).

3.3. The Impact Mechanisms of LST and Relevant Research Measurements

3.3.1. XGBoost Model Tuning and Results Analysis

The study explores the mechanisms influencing LST from both temporal and spatial dimensions. In daytime analysis, the model results indicate that the mean square error (MSE) is 0.066, and the coefficient of determination (R²) is 0.791 for the urban area. For the rural area, the MSE is 0.057, and R² is 0.655. In nighttime analysis, the model results show an MSE of 0.060 and an R² of 0.669 for the urban area. For the rural area, the MSE is 0.074, and R² is 0.511. Overall, the relatively low MSE values demonstrate high prediction accuracy, and the R² values, all exceeding 0.5, confirm that the model provides an acceptable fit.

Using the XGBoost model, a feature importance analysis was conducted to evaluate the factors influencing urban LST variations. The results indicate that DEM is the primary factor influencing daytime LST changes in both urban and rural areas. Conversely, POP_D is the dominant factor affecting nighttime LST variations. In urban areas during the daytime, aside from DEM, the key influencing factors, in order of importance, are wet, POP_D, NDBSI, NDVI, RD, and NTL. In rural areas during the daytime, apart from DEM, the main factors are POP_D, wet, RD, NDVI, aspect, and NDBSI. At nighttime in urban areas, besides POP_D, the primary factors are DEM, RD, wet, NDBSI, NDVI, and NTL. In rural areas at night, excluding POP_D, the key factors are wet, NDBSI, NDVI, RD, DEM, and slope (Figure 4).

The prediction results of the optimized XGBoost regression model were analyzed through visualization and processed using locally weighted scatterplot smoothing (LOWESS) to reduce noise and enhance trend clarity. In this analysis, 10% of the data points were used for local smoothing with three iterations.

For daytime, in the urban area, at the landscape pattern level, PLAND, LPI, and CA showed no effect on urban LST. ED exhibited a slight suppressive effect on LST around 0.3, but its influence diminished thereafter. CONTAG displayed a promotive effect on LST at approximately 0.2, followed by a strong suppressive effect up to 0.5, after which its influence became negligible. Overall, the impact of landscape pattern changes on LST was minimal, with variations in LST being less than 0.001. From an ecological and climatic perspective, NDVI exhibited a general promotive effect on LST, which became significant around 0.7 but diminished beyond that point. Wet exerted a suppressive effect on LST after 0.2, with slight fluctuations until 0.5, after which its influence became negligible. NDBSI had a strong promotive effect on LST after 0.2, peaking at 0.4, and its impact diminished thereafter. At the population activity, NTL’s influence on LST was concentrated below 0.05; while it continued to have a positive effect afterward, it was not significant. POP_D exhibited notable fluctuations in its effect on LST below 0.1, stabilizing thereafter. At the built environment level, BD and BH exhibited some fluctuations in their effects on LST and the range of variations was not significant. RD had a strong positive effect on LST within the range of 0–0.4, with a relatively large impact on LST values. Regarding urban topography, DEM had a substantial influence on LST, showing strong suppressive effects within the range of 0–0.4, where variations in LST were significant. Slope and aspect had minimal variations in their effects on LST (Figure 5).

In the rural areas at the landscape pattern level, PLAND suppressed LST at 0.6. ED promoted LST at approximately 0.25 but suppressed it around 0.4. CONTAG showed a strong promotive effect on LST at 0.1, which persisted until 0.8, after which its influence waned. LPI and CA had no discernible impact on LST. In terms of magnitude, CONTAG exhibited the largest positive influence, with a difference exceeding 0.1. From an ecological and climatic perspective, NDVI suppressed LST, with the suppressive effect becoming pronounced beyond 0.7. Wet exhibited a strong suppressive effect on LST after 0.2, peaking at 0.4 with an impact magnitude of 0.5, after which its influence diminished. NDBSI had a strong promotive effect on LST after 0.2, peaking at 0.4, with minimal influence thereafter. At the population activity level, NTL’s influence on LST was concentrated below 0.02, with minor positive effects thereafter. POP_D exhibited a positive effect on LST beyond 0.1, stabilizing thereafter. At the built environment level, BD and BH showed some fluctuations in their effects on LST and the variations remained stable within a range of 0.02. RD had a strong positive effect on LST within the range of 0–0.15, with an impact variation of approximately 0.1. Regarding urban topography, DEM had a substantial influence on LST, exhibiting strong suppressive effects within the range of 0–0.6, where significant variations in LST were observed. Slope and aspect had some influence on LST changes, but the variations were minimal (Figure 6).

During nighttime in urban areas, the variations in LST influenced by related studies are relatively small. At the landscape pattern level, PLAND exerts a promotive effect on LST when it reaches 0.6. Conversely, ED suppresses LST within the range of 0.25–0.4. CONTAG demonstrates a strong suppressive effect on LST at 0.3. However, LPI and CA exhibit no impact on LST. In terms of the magnitude of influence on LST, the negative effect of CONTAG in the landscape pattern is the most significant, with a difference exceeding 0.2. From the ecological and climatic condition, NDVI generally promotes LST at 0.3, but exhibits a noticeable suppressive effect beyond 0.7. Wet has a strong suppressive effect on LST after reaching 0.2, persisting until 0.4, after which it shifts to a strong promotive effect. NDBSI suppresses LST significantly after 0.2, maintaining this effect until 0.4. After that, its impact diminishes substantially. At the population activity, NTL has a positive impact on LST. Similarly, POP_D positively affects LST beyond 0.1, with its influence peaking at 0.4 and stabilizing thereafter. At the built environment level, BD exhibits a strong positive impact on LST at the 0.1 and 0.5 thresholds. RD shows a strong positive influence on LST within the range of 0–0.15, with an associated variation of approximately 0.04. Regarding urban topography, DEM has a significant influence, strongly suppressing LST within the 0–0.7 range, with substantial variations in LST observed in this interval. Slope positively affects LST at 0.4, while aspect has a limited influence on LST, with a relatively small range of variation (Figure 7).

In rural areas, at the landscape pattern level, PLAND suppresses LST at 0.9. ED also suppresses LST at 0.4. CONTAG promotes LST within the range of 0.1–0.4, while LPI positively impacts LST at 0.5. CA exhibits no influence on LST. From the ecological and climatic perspective, NDVI promotes LST at 0.8. Wet strongly promotes LST after 0.5 but ceases to exert an impact beyond 0.6. NDBSI shows significant fluctuations in its impact on LST after 0.2, persisting until 0.4, after which its influence diminishes. At the population activity, NTL positively impacts LST, while POP_D exhibits significant fluctuations in its influence on LST. At the built environment level, BD positively influences LST between the 0.1 and 0.15 thresholds. RD has a notable impact on LST. Regarding urban topography, DEM had a substantial influence on LST, exhibiting strong suppressive effects within the range of 0–0.6, where significant variations in LST were observed. Slope and aspect had some influence on LST changes, but the variations were minimal. Regarding urban topography, DEM had a substantial influence on LST, exhibiting strong suppressive effects within the range of 0–0.6, where significant variations in LST were observed. Slope and aspect had some influence on LST changes, but the variations were minimal (Figure 8).

3.3.2. SHAP Model Interpretation and Feature Importance

The SHAP impact model shows the influence of study metrics on LST. The x-axis shows the SHAP value. The color of the sample points shows the sample value. Higher sample values lead to lower SHAP values, showing a negative contribution. Lower sample values lead to higher SHAP values, showing a positive contribution. We studied the metrics for the daytime period. In urban areas, DEM is the most influential factor and shows a negative influence. NTL is the second most influential factor and shows a positive influence. Wet ranks third and shows a negative influence. NDBSI ranks fourth and shows a positive influence. POP_D ranks fifth and shows a positive influence. Slope ranks sixth and shows a negative influence. RD ranks seventh and shows a complex influence. NDVI ranks eighth and shows a positive influence, but some areas show a negative influence. BD ranks ninth and shows a positive influence. Other study metrics have minor impacts and are not discussed (Figure 9a).

In rural areas, RD is the primary factor influencing LST variation, mainly exhibiting a positive impact mechanism. DEM ranks as the second most influential factor, primarily showing a negative impact mechanism. Wet is the third most significant factor, with a predominantly negative influence. NDVI ranks fourth, generally displaying a positive influence, though certain localized areas exhibit a negative impact. POP_D is the fifth most influential factor, primarily contributing positively. NDBSI ranks sixth, mainly showing a negative influence. BD ranks seventh, with a predominantly negative impact. NTL is the eighth most influential factor, primarily exhibiting a positive impact mechanism. Aspect ranks ninth, showing both positive and negative impacts. Slope ranks tenth, mainly exhibiting a negative influence, though its overall effect is limited. Other study metrics have minimal impacts on LST and are not discussed in detail (Figure 9b).

During nighttime in urban areas, POP_D is the most significant factor influencing LST variation, primarily exhibiting a positive impact mechanism, though negative effects are observed in certain localized areas. DEM is the second most influential factor, predominantly showing a negative impact mechanism. Wet ranks third, with a primarily positive influence. NTL is the fourth most significant factor, displaying a negative impact. RD ranks fifth and shows a predominantly positive influence. NDBSI is the sixth influential factor, primarily contributing negatively. BD ranks seventh and mainly exhibits a positive impact. NDVI is the eighth influential factor, showing a negative impact. Slope ranks ninth, primarily demonstrating a positive influence. ED ranks tenth and predominantly shows a negative impact. Other study metrics have minimal effects on LST and are not discussed in detail (Figure 10a).

In rural areas, POP_D is the most significant factor influencing LST variation, primarily exhibiting a negative impact mechanism. Wet ranks second and mainly shows a positive impact. DEM is the third most influential factor, predominantly showing a negative impact. NDBSI ranks fourth, contributing negatively. Slope ranks fifth, primarily demonstrating a negative influence. RD is the sixth most significant factor, mainly showing a negative impact. BD ranks seventh and primarily contributes negatively. Aspect ranks eighth, predominantly showing a negative impact mechanism, with positive effects observed in some localized areas. NDVI ranks ninth, exhibiting both positive and negative impacts. CONTAG ranks tenth, primarily showing a positive influence. NTL is the eleventh most influential factor, mainly exhibiting a negative impact mechanism. Other study metrics have minimal effects on LST and are not discussed in detail (Figure 10b).

Multivariable interaction analysis provides a deeper understanding of how multiple variables interact to influence the target variable. In this analysis, the target variable is typically represented by the SHAP value, which corresponds to LST in this study. The research process involves several steps. First, a summary plot is used to create a scatter plot, offering an overview of the extent to which multiple variables influence the target variable and identifying variable groups with significant interactive effects. Subsequently, a dependence plot is employed to conduct a detailed analysis of the selected variable groups, allowing for the observation of specific mechanisms through which multiple variables impact the target variable.

3.3.3. SHAP-Based Multivariable Interaction Analysis

Multivariable interaction analysis helps to understand how multiple variables interact to influence the target variable. The target variable is represented by the SHAP value. In this study, the SHAP value corresponds to LST. The analytical process has several steps. A summary plot generates a scatter plot. This plot shows the influence of multiple variables on the target variable. It also identifies variable groups with strong interactive effects. A dependence plot analyzes the selected variable groups in detail. It shows how multiple variables impact the target variable. We combined model recommendations with the influence rankings of paired metrics from related studies and selected the top seven pairs with the strongest impact on LST for analysis.

During the daytime in urban areas, we observed significant interactions between pairs of study metrics. These metrics include DEM, NTL, wet, NDBSI, POP_D, and slope (Figure 11h). The variable pairs with strong interaction effects are DEM and NTL, DEM and wet, DEM and NDBSI, DEM and POP_D, DEM and RD, DEM and NDVI, and wet and POP_D. We used dependence plots to analyze the interaction mechanisms of these pairs on LST. For DEM and NTL (Figure 11a), as DEM increases and NTL decreases, LST decreases. When DEM decreases and NTL increases, LST increases. For DEM and wet (Figure 11b), wet values are randomly distributed along the decreasing trend of LST. As DEM increases, LST shows a clear downward trend. This suggests that the decrease in LST is mainly caused by the increase in DEM. Wet has little effect on LST in this process. For DEM and NDBSI (Figure 11c), a similar pattern appears. NDBSI values are randomly distributed during the LST decrease. This indicates that the reduction in LST is driven by the increase in DEM. For DEM and POP_D (Figure 11d), as POP_D decreases and DEM increases, LST decreases significantly. This shows a strong synergistic effect of these variables on LST. For DEM and RD (Figure 11e), as DEM increases and RD decreases, LST decreases. For DEM and NDVI (Figure 11f), as both DEM and NDVI increase, LST decreases clearly. For wet and POP_D (Figure 11g), as wet increases, LST decreases. However, POP_D values are randomly distributed. This indicates that wet has a greater effect on LST, and the synergistic effect between wet and POP_D is not strong.

In rural areas, we observed significant interactions between pairs of study metrics. These metrics include RD, DEM, wet, NDVI, POP_D, and NDBSI (Figure 12h). The variable pairs with strong interaction effects are RD and DEM, RD and wet, RD and POP_D, RD and NTL, DEM and wet, wet and POP_D, NDVI and DEM, and NDVI and BD. We used dependence plots to analyze the interaction mechanisms of these pairs on LST.

For RD and DEM (Figure 12a), when DEM exceeds 0.075, RD and DEM both have a positive impact on LST. For wet and RD (Figure 12b), wet is mostly randomly distributed. LST is influenced mainly by the increase in RD, which leads to an increase in LST. For RD and POP_D (Figure 12c), POP_D is also mostly randomly distributed. LST is mainly influenced by the increase in RD, causing an increase in LST. For RD and NTL (Figure 12d), NTL shows a low-value distribution before 0.075 and a random distribution afterward. NTL’s contribution to the synergistic effect is not significant. For DEM and wet (Figure 12e), wet is primarily randomly distributed. LST is mainly influenced by the increase in DEM, which causes a reduction in LST. For wet and POP_D, wet is also randomly distributed. LST is again influenced mainly by the increase in DEM, resulting in a decrease in LST. For NDVI and DEM (Figure 12f), and NDVI and BD (Figure 12g), the influence of DEM is mainly randomly distributed. Before NDVI reaches 0.8, it has a positive impact on LST. After 0.8, NDVI has a negative impact on LST.

At nighttime in urban areas, we observed that the interactions between pairs of study metrics, including POP_D, DEM, wet, NTL, RD, and NDBSI, have a significant influence on LST (Figure 13h). The variable pairs with more pronounced interaction effects include POP_D and DEM, POP_D and RD, NTL and DEM, DEM and wet, DEM and RD, DEM and NDBSI, and DEM and slope. To further validate the synergistic relationships between these variable pairs, we employed dependence plots to analyze their interaction mechanisms on LST. For POP_D and DEM (Figure 13a), as DEM decreases and POP_D increases, the combined effect has a positive impact on LST. For POP_D and RD (Figure 13b), as both RD and POP_D increase, the combined effect has a positive impact on LST. For NTL and DEM (Figure 13c), as NTL increases and DEM decreases, LST tends to increase. For DEM and wet (Figure 13d), wet values are primarily randomly distributed. The decrease in LST is mainly driven by the increase in DEM. For DEM and RD (Figure 13e), as DEM increases and RD decreases, LST shows a downward trend. For DEM and NDBSI (Figure 13f), NDBSI values are mostly randomly distributed. As DEM increases, LST decreases. Thus, NDBSI does not contribute significantly to the interaction mechanism. For DEM and slope (Figure 13g), as both DEM and slope increase, LST exhibits a downward trend.

In rural areas, we observed significant effects of variable interactions on LST. The variables include POP_D, wet, DEM, NDBSI, slope, and RD (Figure 14h). The pairs with stronger interaction effects are POP_D and wet, POP_D and DEM, POP_D and NDBSI, POP_D and RD, RD and NDVI, wet and DEM, and wet and slope. For POP_D and wet (Figure 14a), wet does not strongly influence LST directly. The effect of POP_D is also limited. When wet exceeds 0.6, an increase in wet and a decrease in POP_D lead to a rise in LST. For DEM and POP_D (Figure 14b), LST decreases as DEM increases to 0.4. This decrease is caused by the decline in POP_D. Before DEM reaches 0.4, an increase in POP_D reduces the decline in LST. For NDBSI and POP_D (Figure 14c), POP_D is mostly randomly distributed. LST decreases as NDBSI increases. The impact on LST is driven mainly by NDBSI, with little contribution from its interaction with POP_D. For RD and POP_D (Figure 14d), both variables show random variations in their impact on LST. No clear interaction is observed. For RD and NDVI (Figure 14e), as both RD and NDVI increase, LST does not show a significant trend. For wet and DEM (Figure 14f), when wet exceeds 0.6, an increase in wet and a weaker DEM lead to a positive trend in LST. For wet and slope (Figure 14g), there is no clear interaction mechanism affecting LST.

4. Discussion

This study uses a nonlinear model and its interpretive framework to elucidate the mechanisms influencing the urban thermal environment. Most influencing factors exhibit nonlinear relationships with the urban thermal environment, highlighting the complexity of these mechanisms and the need to consider multiple interacting factors. Among the key influencing factors, certain mechanisms are particularly significant. The ranking of these significant features is largely consistent between the XGBoost model and SHAP analysis. This consistency demonstrates that SHAP effectively explains the prediction results of the XGBoost model and accurately reflects the contribution of each feature to the model’s predictions. Such alignment enhances our understanding of the model’s decisions, improving its interpretability and reliability.

4.1. The Influence Mechanism of a Single Research Measure on LST

4.1.1. The Impact Mechanisms of LST Variation Across Urban and Rural Spatial Scales

Existing research shows that DEM has the strongest impact on LST during the daytime at the urban area scale. The XGBoost model reveals that DEM has a significant negative effect on LST. This negative impact decreases after reaching a certain threshold [43]. Most scholars agree that DEM strongly influences LST in a negative way. Some studies report that DEM has little effect on LST due to small elevation differences in urban areas [44]. Hangzhou’s highest point is Qingliang Peak in Lin’an District, with an elevation of 1787 m. Its lowest point is Dongtiaoxi in Yuhang District, with an elevation of 3 m. This creates a large elevation difference of 1784 m. This difference explains why DEM has a significant influence on LST in Hangzhou (

Table 2. The ranking of the significance of the research measure.

	Daytime		Nighttime
Model	XGBoost	SHAP	XGBoost	SHAP
Urban area	DEM	DEM	POP_D	POP_D
	Wet	NTL	DEM	DEM
	POP_D	Wet	RD	Wet
	NDBSI	NDBSI	Wet	NTL
	NDVI	POP_D	NDBSI	RD
	RD	Slope	NDVI	NDBSI
	NTL	RD	NTL	BD
	Slope	NDVI	Slope	NDVI
	Aspect	BD	BD	Slope
	BD	Aspect	Aspect	ED
Rural area	DEM	RD	POP_D	POP_D
	POP_D	DEM	Wet	Wet
	Wet	Wet	NDBSI	DEM
	RD	NDVI	NDVI	NDBSI
	NDVI	POP_D	RD	Slope
	Aspect	NDBSI	DEM	RD
	NDBSI	BD	Slope	BD
	BD	NTL	Aspect	Aspect
	Slope	Aspect	BD	NDVI
	NTL	Slope	NTL	CONTAG

Note: The study identified the top 10 research measures under the detection of the XGBoost and SHAP models.

).

At night, POP_D has a stronger influence on LST than DEM. The XGBoost model shows that POP_D has a positive effect on LST. The relationship between POP_D and LST is often non-linear. Studies suggest that a population density of more than 14,500 inhabitants/km² can increase air temperature by more than 1 °C. These studies also show a strong positive correlation between population density and LST [45]. This matches our findings on the non-linear influence of POP_D on LST at night. Multiple thresholds exist (several smooth linear segments) and exceeding them causes a stronger positive effect on LST. Wet, RD, and NTL also have strong effects on LST. Landscape pattern metrics, however, have weaker effects on LST. This is related to the distribution of urban morphology and topographic features in Hangzhou. Landscape patterns have little influence on LST at larger spatial scales (Table 2).

At the rural area scale, DEM has a significant impact on LST during the daytime. RD also strongly influences LST during the daytime. The influence of DEM on LST in rural areas is similar to its effect in urban areas. Therefore, we focus on discussing RD. RD is part of the linear infrastructure (LI). Studies suggest that LI changes urban morphology and impacts LST. In rural areas, RD creates linear impervious surfaces. These surfaces, along with related infrastructure, form LI. Impervious surfaces absorb large amounts of heat during the daytime. This heat dissipates slowly, which causes RD to have a strong positive effect on LST.

At night, POP_D is the main factor influencing LST in rural areas. The effect of POP_D on LST is predominantly positive. Wet also strongly influences LST. The Normalized Difference NDBSI has a strong impact on LST as well. Landscape patterns have a weak influence on LST in rural areas.

4.1.2. The Influence Mechanism of LST Variations Across Temporal Dimensions

By distinguishing the feature importance of factors influencing LST between daytime and nighttime, we found that during the daytime, DEM has a significant impact on both urban and rural areas, with a consistent negative influence. At nighttime, POP_D is the most important factor influencing LST, showing a consistently positive effect. Among other influencing factors, DEM, RD, wet, NDBSI, and NTL also exhibit strong impacts on LST. Overall, the research metrics influencing urban and rural areas during the daytime are primarily related to urban topography and climate. At nighttime, the influencing metrics are mainly associated with population distribution and activities, as well as urban topography.

Using SHAP analysis, we found that POP_D has the largest difference in its influence on LST between morning and evening in urban areas. At night, there is no solar radiation. Population clusters act as concentrated heat emission sources. Therefore, population clusters have a significant positive impact on LST at night. In rural areas, RD shows the largest difference in influence on LST between morning and evening. During the day, solar radiation heats the ground. This causes rapid warming near roads. At night, the accumulated heat dissipates. Roads cannot accumulate new heat at night [46]. As such, the influence of roads on LST is much lower at night (Table 2). Upon analysis, we believe that the primary reason for this result is related to the location of the roads. As a major component of linear infrastructure, roads exhibit different thermal behaviors depending on their spatial context. In urban areas, nighttime human activities, such as vehicular traffic, remain relatively frequent. It may prevent road surface temperatures from decreasing significantly compared to daytime. In contrast, in rural areas, the level of human activity is much lower, resulting in minimal disturbances to road surface temperature variations. Therefore, the observation of lower nighttime LST in rural areas is considered reliable.

4.2. The Influence Mechanism of LST Under the Interaction of Multiple Research Measure

Through research and analysis, we found that in urban areas during the daytime, the variables with significant synergistic effects on LST are DEM and NTL, DEM and POP_D, DEM and RD, and DEM and NDVI. The cooling effect on LST is characterized by positive DEM and negative NTL, positive DEM and negative POP_D, positive DEM and negative RD, and positive DEM and positive NDVI. In this series of synergistic influence mechanisms, the direction of the combined effect of the metrics on reducing LST is consistent with their individual effects. This indicates that the results are accurate and logically valid.

At nighttime, the variables with significant synergistic effects on LST include POP_D and DEM, POP_D and RD, DEM and NTL, and DEM and RD. The cooling effect on LST is characterized by positive POP_D and negative DEM, positive POP_D and negative RD, positive NTL and negative DEM, positive DEM and negative RD, and positive DEM and positive slope. Among these synergistic metric combinations, except for DEM and slope, the combined effect on reducing LST aligns with their individual effects.

Analysis shows that DEM decreases LST when it acts as an independent variable. Higher DEM values result in lower LST. Slope increases LST when it acts as an independent variable. Higher slope values result in higher LST. The synergistic effect of DEM and slope is mainly driven by DEM. Slope has little impact on LST.

In rural areas, no variable combinations exhibit strong synergistic effects on LST during the daytime. At nighttime, wet and DEM have a significant synergistic effect on LST. However, this significant effect is observed only when wet values exceed 0.6. The direction of the combined effect is consistent with the direction of their individual effects when acting as independent variables (Figure 14f).

In summary, research metrics in urban areas are more likely to form significant multivariable synergistic effects on LST compared to rural areas. We believe this result is due to two main reasons. First, higher LST values in urban areas tend to be more spatially clustered. This clustering amplifies the synergistic cooling effect of multiple variables on LST. In rural areas, the clustering of high LST values is relatively weaker, reducing the interference of related variables on LST. Second, certain variables have extremely strong individual effects on LST. Even when synergistic variables are randomly distributed, LST is still primarily influenced by these dominant variables, exhibiting relatively consistent variation trends (

Table 3. The pairs of research measures have a significant impact on LST.

	Daytime	Nighttime
Urban area	DEM&NTL	POP_D&DEM
	DEM&POP_D	POP_D&RD
	DEM&RD	DEM&NTL
	DEM&NDVI	DEM&RD
	Wet&POP_D	DEM&Slope
Rural area		Wet&DEM

).

4.3. Implications for Urban Planning and Future Planning

Based on the existing conclusions, we found that HSTA in Hangzhou is distributed in urban areas and the southern regions of Chunan and Jiande within rural areas. The distribution patterns of HSTA differ between daytime and nighttime. There are two main differences. First, large water bodies release significant heat at night, resulting in noticeably higher LST in areas near water bodies in rural areas during nighttime. Second, during the daytime, RD and LI connecting urban and rural areas absorb substantial solar radiation, causing higher LST in these regions.

LSTA is concentrated in the northwestern regions of Hangzhou, characterized by higher DEM values. Areas most influenced by LSTA during both daytime and nighttime include large water bodies and LI, particularly water body regions located in rural areas. During the daytime, these water bodies demonstrate a significant cooling effect, forming distinct cold spots and contributing to a strong cooling island effect in the surrounding areas.

To address the characteristics of all-day LST distribution in Hangzhou, we propose a multi-faceted approach. First, our findings indicate that landscape patterns have an overall insignificant impact on Hangzhou’s LST. Therefore, in future large-scale urban spatial planning, it is unnecessary to overemphasize landscape connectivity and shape. Instead, planning should focus on three key aspects: urban topography, population distribution, and the layout of linear infrastructure. Additionally, attention should be given to urban climate regulation by increasing the quantity of blue-green infrastructure, particularly in urban areas. These infrastructures can help regulate urban humidity and mitigate LST growth to some extent.

Second, it is important to recognize the differences between daytime and nighttime urban heat dynamics. For example, large water bodies absorb substantial solar radiation during the day, creating a cooling effect and forming cold islands in the surrounding areas. However, at night, these water bodies release stored heat, resulting in elevated LST in nearby areas. Therefore, managing water body sizes and optimizing their cooling potential are critical for mitigating urban heat.

Finally, addressing urban heat challenges in complex heat networks requires integrating the influence of multiple factors. For instance, in areas of Hangzhou with high DEM values, policymakers can deprioritize heat source control as these regions are less affected by urban heat. Conversely, low DEM areas, especially those with high population density, should be prioritized for urban heat mitigation strategies. This targeted approach can better facilitate resource allocation and urban planning.

5. Conclusions and Limitations

This study primarily explores the annual trends in daytime and nighttime LST changes and their associated influencing mechanisms in Hangzhou. Using an urban–rural dichotomy as the spatial segmentation approach, Hangzhou was divided into urban and rural areas for separate analyses. The study presents three key conclusions:

• The spatial distribution of LST in Hangzhou reveals that HSTAs are mainly concentrated in urban areas, with higher HSTA densities closer to the city center. LSTAs, on the other hand, are primarily distributed in certain rural areas and the high-altitude regions along Hangzhou’s northwestern edge. For most regions, the distribution of HSTAs and LSTAs remains consistent between daytime and nighttime, except for large water bodies in rural areas. Due to the high specific heat capacity of water, these areas release more heat at night compared to daytime.
• The contributions of influencing mechanisms to the urban thermal environment vary across temporal and spatial scales. The ranking of influence categories is as follows:
  First, during the daytime, DEM has a strong impact on the thermal issues in both urban and rural areas. At nighttime, the distribution and activity of the population play a dominant role. Second, wetness significantly affects the global area of Hangzhou during both day and night. Third, the built environment, especially road density, has a notable impact on the global area of Hangzhou. Fourth, the distribution of landscape patterns, including landscape shape and density, has a relatively low impact on LST across Hangzhou.
• In multivariable synergy analyses, the collaborative effects of influencing factors on LST are significantly stronger in urban areas compared to rural areas. In urban areas, the most influential synergistic factors are ranked as follows: DEM, POP_D, RD, wet, NDVI, and slope. These factors should be jointly considered in future urban planning and development efforts in Hangzhou’s urban areas to comprehensively enhance the capacity to mitigate LST in the city’s core areas.

In summary, compared to previous studies [47], this paper introduces a nonlinear approach to examining the mechanisms influencing urban LST. By leveraging XGBoost and SHAP, the study ranks the importance of key factors affecting LST in Hangzhou and interprets their respective mechanisms of influence. Notably, it provides an explanatory analysis of the synergistic effects of related factors within the dual spatial scale of urban and rural areas. This research proposes the potential of addressing future urban heat challenges in Hangzhou through multivariable correlations, moving beyond the traditional [48]. Future studies can further refine this research by conducting more detailed investigations across spatial and temporal dimensions. For example, cities could be more finely divided, categorizing different areas into specific types (such as local climate zones or functional zones), and further exploring the key factors influencing land surface temperature in various urban areas [49]. Additionally, incorporating advanced machine learning and deep learning techniques could enhance both the precision and scope of the analysis, offering deeper insights into urban thermal environments.