Revealing Nonlinear Relationships Between Urban Morphology and Diurnal Land Surface Temperature via Spatial Heterogeneity

Abstract

Urban morphology, encompassing both horizontal landscape patterns and three-dimensional architectural structures, plays a pivotal role in modulating urban heat distribution. However, conventional models often fail to capture the intricate spatial nonstationarity and nonlinear coupling of these drivers at the block scale. Recognizing that land surface temperature (LST) exhibits distinct diurnal and nocturnal thermal cycles, this study explicitly incorporates spatial heterogeneity analysis to systematically evaluate the relative and local contributions, marginal effects, and interaction mechanisms of multidimensional urban morphology on diurnal LST variations. To achieve this objective, geographically weighted extreme gradient boosting and SHapley Additive exPlanations were employed to decipher these complex driving mechanisms from a morphological perspective. The results indicate the following: (1) Built environment variables predominate the spatial heterogeneity of LST in Xi’an, China, with their governing mechanisms shifting diurnally—characterized by a midday NDVI-induced evapotranspiration cooling effect and an atmospheric back-radiation warming effect associated with PM2.5 during the night and early morning. (2) The driving mechanisms exhibit pronounced spatial nonstationarity; while the northeastern and northern sectors are primarily influenced by the synergistic interaction between surface albedo and PM2.5, the central-western and southern regions are governed by population density and 3D architectural morphology. (3) Significant nonlinear interaction thresholds and non-monotonic response mechanisms were identified across the variables. By resolving localized thermal responses through the lens of spatial heterogeneity, this research provides a robust scientific framework for precision urban planning and the mitigation of the urban heat island effect.

1. Introduction

The share of the global population living in urban areas is projected to reach 68% by 2050 [1]. During urbanization, natural landscapes are replaced by urban landscapes, disrupting surface energy balance and thermal inertia, such as thermal radiation, heat storage, and heat transfer. Combined with increased anthropogenic heat emissions, this results in significantly higher temperatures in urban areas compared to suburban areas, contributing to the formation of an urban heat island (UHI) [2,3,4]. A UHI is a typical urban ecological phenomenon, causing significant urban environmental problems, such as increased energy demand [5] and adverse effects on the health [4] and thermal comfort of urban residents [5,6,7]. High-temperature urban environments exacerbate heat stress and pose a threat to residents’ health during heat waves by increasing the heat risk in urban areas [6]. To mitigate the adverse effects on urban sustainability, a comprehensive understanding of the driving mechanisms of UHI and the development of targeted strategies are essential.

UHI is primarily assessed using land surface temperature (LST) data acquired from remote sensing imagery, a method widely used to analyze surface UHI (SUHI). While remote sensing technology can study the coverage and resolution of urban thermal environments, most studies focus more on the impact of daytime SUHI and less on diurnal LST data. However, considering only daytime LST while ignoring nighttime LST precludes a holistic interpretation of the underlying drivers of SUHI. Furthermore, nighttime temperature increases due to the SUHI effect have been reported to have negative health effects and are associated with increased mortality [8,9]. A holistic understanding of diurnal LST variation is indispensable for SUHI research; yet, traditional studies frequently overlook high-frequency intra-diurnal thermal fluctuations, limiting the effectiveness of heat mitigation strategies. LST data on diurnal cycles obtained through ECOSTRESS (https://appeears.earthdatacloud.nasa.gov/ (accessed on 3 November 2025)) can circumvent these limitations. Recently, several studies have used ECOSTRESS LST data to study the diurnal variation in LST [10,11,12]. Therefore, ECOSTRESS LST represents a new significant proxy indicator for exploring the diurnal UHI effect [13,14,15].

The diurnal cycle of LST depends on multiple factors. Urban morphology, by regulating heat distribution, thermodynamics, and urban airflow, has become the key driver of the diurnal cycle of LST and is usually deconstructed into two core dimensions: landscape morphology and building morphology. Landscape morphology is categorized into three distinct typologies: (1) Water bodies: High-density clean water bodies can effectively absorb heat through evaporation and conduction processes, and this effect is particularly significant in high-density built-up areas [16]. (2) Vegetation cover: Vegetation cover reduces surface temperature by providing shade and increasing evapotranspiration [17]. It also improves the urban living environment by influencing the urban microclimate and improving air quality [18]. (3) Impermeable surfaces: An increase in impermeable surfaces reduces the land’s natural evaporation and water infiltration capacity, usually leading to an increase in surface temperature [19]. Architectural morphology refers to the horizontal layout and vertical structure of buildings [20]. Building height and density are the most commonly used indicators of architectural morphology. Increased building height and density weaken urban ventilation efficiency by increasing the roughness of the underlying surface, usually leading to an increase in surface temperature [21]. Traditional architectural morphology mainly includes two-dimensional (2D) urban morphological variables, including 2D urban morphology, which can be represented by the composition and configuration of the landscape. Composition usually refers to the percentage of different land cover grades. Because different land cover categories possess unique characteristics in terms of soil moisture content, radiation, and thermal properties, different landscape compositions exhibit significant responses to spatial thermal patterns [22].

One of the most prominent features of the urban environment is the complex three-dimensional (3D) landscape composed of buildings [23,24,25]. Although some studies have investigated the driving effect of diurnal and nocturnal LST by introducing 3D urban structure indicators [26,27,28], these studies only reflect the spatial composition of urban spatial morphology and do not comprehensively consider the influence of natural geography, human activities, and environmental factors on LST, thus requiring further research. Particulate matter with a diameter of $<$2.5 μm (PM2.5), as a major environmental disturbance factor, is a crucial indicator driving diurnal LST. Studies have shown that PM2.5, as a major component of haze, significantly enhances China’s UHI at night [29]. Further research, incorporating population density (Pop), PM2.5, and nighttime light data, has revealed seasonal differences in the spatial synergy between PM2.5 and UHI [30], particularly pronounced in autumn and winter [31]. However, these studies have not been integrated with urban morphology indicators and lack exploration of the diurnal variations in LST drivers. Furthermore, landscape and building morphology indicators can only capture macroscopic spatial patterns; the significant impact of spatial heterogeneity makes it difficult to accurately model the relationship between urban morphology and LST, introducing uncertainty into existing research results and highlighting the need for more refined analytical methods to capture such variations. In addition, existing research often focuses on the linear contribution of single variables, largely neglecting the complex nonlinear relationships between variables and their synergistic effects on LST. Therefore, it is necessary to construct a comprehensive indicator system encompassing natural geography, human activities, environmental factors, and urban morphology from a more holistic perspective. This system is crucial for studying the diurnal variation in LST drivers and the spatial heterogeneity of their driving mechanisms and for providing site-specific planning for mitigating diurnal UHI in rapidly urbanizing areas.

Accurately identifying the factors and mechanisms by which urban spatial morphology influences LST is a widely recognized problem. Traditional methods have limitations in revealing the complex nonlinear relationship between LST and urban spatial morphology [32]. For example, ordinary least squares and geographically weighted regression methods, due to their high reliance on prior assumptions of linear relationships between variables, have significant limitations in revealing the complex nonlinear relationships and interactions between urban morphology and LST. Nonlinear models, on the other hand, do not impose strict assumptions on data distribution but rather increase model complexity as the dataset grows, making their configuration more flexible. These models are particularly effective in handling the complex data relationships and patterns between urban morphology and LST [33,34]. For example, random forest, extreme gradient boosting (XGBoost), and support vector machine are commonly used to analyze the relative importance and marginal effects of urban morphology on LST [35,36].

Given that previous research has primarily focused on global statistical correlations or local physical simulations, mesoscale neighborhood effects have been largely neglected; simultaneously, most machine learning-based studies of urban thermal environments rely on global samples, ignoring the impact of urban morphology variables on the spatial heterogeneity of the urban thermal environment. Geographically weighted model-based methods have emerged to mitigate this limitation. Geographically weighted random forest, employing a parallel ensemble strategy, constructs multiple independent decision trees and averages them, demonstrating strong generalization ability and robustness in analyzing the nonlinear cooling effects of indicators such as green space morphology [37]. The Geographically Weighted Neural Network, utilizing a deep learning architecture, captures nonlinear features [38], but due to its low model transparency, it struggles to analyze specific collaborative thresholds between morphological factors. Geographically weighted XGBoost (GW-XGBoost), employing a sequential residual iterative correction mechanism, can more sensitively capture temperature gradients caused by subtle morphological changes, resulting in higher local analytical accuracy in capturing nonlinear interaction thresholds and non-monotonic response mechanisms of driving factors, offering significant advantages for studying the spatial heterogeneity of urban morphology and LST [39,40].

Although nonlinear models are effective, their “black box” nature often limits the intuitive analysis of surface thermodynamic mechanisms [33]. While the SHapley Additive exPlanations (SHAP) method, as a primary approach to interpretable machine learning, offers the possibility of improving model transparency and quantifying variable interactions, the integration of geographically weighted models with SHAP in diurnal urban thermal environment modeling remains largely unexplored [41]. Furthermore, although machine learning models have been widely used to capture such complex nonlinear problems, existing research primarily focuses on the influence mechanisms of univariate variables on LST, while the complex interactions within urban systems and the synergistic effects of urban spatial morphology variables on LST are mostly neglected. This makes it difficult for existing empirical studies to reveal how urban morphology, under different spatial contexts, jointly regulates the diurnal evolution of LST through these factors. This knowledge gap limits our comprehensive understanding of the driving mechanisms between urban morphology and diurnal LST, resulting in a lack of solid scientific support for refined thermal environment analysis.

In summary, although many scholars have conducted extensive research on the nonlinear relationship between LST and urban spatial morphology at the street scale, the diurnal variation in LST and the spatial heterogeneity of urban spatial morphology variable indicators still require further investigation. Moreover, the synergistic effects of interactions among urban morphology variables on LST are mostly ignored, and the significant spatial heterogeneity between urban spatial morphology and LST has not been considered simultaneously. Given these limitations, this study takes Xi’an, a city with rapid urban expansion and significant diurnal temperature variation, as an example to investigate the diurnal variation in urban LST and its driving factors. This study aims to: (1) Analyze the spatial pattern and spatiotemporal evolution of diurnal LST. (2) Quantify the nonlinear importance of urban morphological variables and their synergistic response mechanisms. (3) Reveal the spatial heterogeneity of the driving mechanisms, providing scientific guidance for urban thermal environment regulation based on the research findings.

2. Materials and Methods

2.1. Study Area

Xi’an (107°40′–109°49′ E, 33°42′–34°45′ N) is located in Northwest China. It has six national and provincial key development zones and is the capital of Shaanxi Province, a core city of the Guanzhong Plain urban agglomeration, and a national central city. Xi’an currently administers 11 districts and 2 counties. As of 2020, the city’s permanent resident population was 13.1676 million, and its GDP exceeded 1.2 trillion yuan (Xi’an Municipal Bureau of Statistics, 2023). The region has a warm temperate semi-humid continental monsoon climate with distinct seasons, hot and rainy summers, and cold and dry winters. The average annual temperature is 13.5–15.0 °C, with the hottest month, July, averaging 27.2 °C. The city averages 1646 h of sunshine per year. Meteorological statistics show that the average annual temperature in Xi’an has shown a significant upward trend from 1961 to 2024, with a warming rate of 0.28 °C/10 years. The warming rate has accelerated since the beginning of this century, making the trend of global warming increasingly pronounced (Figure 1).

Meanwhile, in the process of rapid urbanization, its built-up area expanded from 395 km² in 2010 to 810 km² in 2023 (Xi’an Municipal Bureau of Statistics, 2023). Since the implementation of the Western Development Strategy in 2000, the urban population, motor vehicle emissions, and industrial pollution emissions in Northwest China have surged. Large areas of natural underlying surfaces have been gradually replaced by impermeable surfaces such as concrete and asphalt. Xi’an is located in the central part of the Guanzhong Basin. Although it is bordered by the Qinling Mountains to the south, the internal terrain within the built-up area is relatively flat, with minimal topographic relief. Consequently, macro-scale basin effects outweigh micro-scale terrain variations. At the same time, the large-scale population concentration in the built-up area leads to increased anthropogenic heat emissions. Therefore, the funnel-shaped topography, combined with numerous anthropogenic heat sources, becomes the main driving force for regional heat accumulation, while the impact of local topographical changes is secondary in this specific urban environment.

In conclusion, Xi’an, with its diverse functions and increasingly severe urban thermal environment problems, is a typical research area for studying the differences in LST and its influencing factors among different urban morphologies. Given that the main urban area is a highly concentrated area of human activities, economic factors, and urban construction, its functional heterogeneity is stronger, the heat island effect is more significant, and relevant research data is more detailed. This study uses Xi’an City in the 2020 Global Urban Settlement Boundary product as the research scope, excluding non-urban areas and focusing on the urban built-up area, covering 9 administrative districts with a total area of 1174.90 $\mathrm{k}{\mathrm{m}}^2$ [42]. Within this specific study area, the average population density reaches 3760 persons/km², characterizing it as a region of high-intensity human activity. The land use composition is highly heterogeneous, predominantly consisting of residential areas (45.45%), industrial land (20.25%), commercial land (14.07%), and public service facilities (12.87%). Greenery land, which provides critical ecosystem services for heat mitigation, accounts for 7.36% of the study area. These diverse land use forms and their corresponding 3D structures constitute the complex urban morphology that drives the spatial variations in LST observed in this study.

2.2. Data Source

Table 1. Description of the data used in this study.

Data Type	Spatial Resolution	Data Usage	Data Source
Street Blocks	-	Research area block data	Tang et al., 2025 [43]
Building height	-	Building height data	Che et al., 2024 [44]
Landsat-8 satellite image	30 m	Calculate various factors	USGS, https://earthexplorer.usgs.gov/ (accessed on 9 November 2025)
ECOSTRESS	70 m	Invert LST	USGS, https://earthexplorer.usgs.gov/ (accessed on 3 November 2025)
Copernicus Global Land Service (CGLS)	5 km	Correct LST data	CMLS, https://land.copernicus.eu/en/products/temperature-and-reflectance/hourly-land-surface-temperature-global-v2-0-5km (accessed on 26 November 2025)
WorldPop	100 m	Population data	WorldPop, www.worldpop.org (accessed on 9 November 2025)
Canopy height (CH)	10 m	Tree canopy height data	Lang et al., 2023 [45]
PM2.5	1 km	PM2.5 data	Wei and Li, 2025 [46,47,48]
Buildings land cover	0.5 m	Calculate the landscape index and the water index	Zhang et al., 2025 [42,49]

shows the data used, including street and building height vector data, Landsat-8 satellite imagery, ECOSTRESS, Copernicus Global Land Service (CGLS), canopy height (CH), PM2.5, buildings land cover, and population raster data. To ensure timeliness, the data period primarily focuses on June–August 2020–2023. ArcGIS Pro 3.4.2 was used to uniformly project the data to the WGS 1984 UTM Zone 49 N coordinate system, uniformly crop it to the study area, and interpolate it to a 30 m spatial resolution using the nearest neighbor method.

LST data source: ECOSTRESS satellite. ECOSTRESS is a medium-resolution satellite launched by NASA in 2018, initially used for inverting day and night LST, with a spatial resolution of 70 m $\times$ 70 m. Previous studies have confirmed the high accuracy of ECOSTRESS-derived LST products. Therefore, in recent years, LST products derived from ECOSTRESS have been widely used. We focus on factors influencing the urban thermal environment. ECOSTRESS images are primarily concentrated between June and September to capture the more pronounced thermal characteristics of cities in summer. Weather conditions such as calm, clear, cloudless, or low cloudiness also need to be considered to ensure data accuracy. Due to the orbital limitations of the International Space Station, ECOSTRESS may not be able to fly over the same location again for several days or even longer. Therefore, surface temperatures of the same area at different times must be measured using ECOSTRESS images from different dates. During image screening, it is also necessary to ensure that meteorological conditions are similar across different dates to minimize the impact of weather.

2.3. Study Framework

The methodological framework is shown in Figure 2. First, building upon established landscape and architectural metrics, we integrate diverse multi-source data, including Pop and PM2.5 concentrations. Second, to account for the spatial heterogeneity and non-linear associations between urban morphology and LST, we employed the GW-XGBoost algorithm for model fitting, effectively capturing the spatial non-stationarity inherent across the study area. Furthermore, the SHAP method was incorporated to quantify the marginal effects, interaction mechanisms, and diurnal variations in these metrics. This integrated framework facilitates a comprehensive understanding of the intricate and multifaceted relationships between urban form and LST, ultimately providing crucial insights for resilient urban climate mitigation.

2.4. LST Processing

We acquired diurnal LST data for the summer of 2021–2023 (June–August). After QC testing, four ECOSTRESS images were selected that met the requirements. Four high-quality scenes were collected at different times (Beijing time, UTC+8): 2 July 2022, midnight (01:33); 14 June 2023, morning (07:58); 1 August 2021, afternoon (14:13); and 14 September 2021, nighttime (20:35). To synthesize a representative diurnal cycle from ECOSTRESS scenes acquired on different dates, a meteorological normalization procedure was essential to eliminate the influence of day-to-day weather and atmospheric variations. Specifically, hourly 5 km CGLS LST data (https://land.copernicus.eu/global/products/lst (accessed on 26 November 2025)) were employed as a regional background baseline to correct the LST data across different dates [50]:

$$LST\left(t\right)={LST}_{ECO}\left(d,t\right)+{LST}_{CGL{S}_m}\left(t\right)-{LST}_{CGLS}\left(d,t\right)$$

(1)

where LST(t) represents the adjusted ECOSTRESS surface temperature at time t; ${LST}_{ECO}\left(d,t\right)$ and ${LST}_{CGLS}\left(d,t\right)$ represent the ECOSTRESS and CGLS surface temperatures derived at date d and time t, respectively; and ${LST}_{CGLS\_m}\left(t\right)$ represents the average CGLS surface temperature at time t throughout the study period. The term $\left(LS{T}_{CGLS\_m}\left(t\right)-LS{T}_{CGLS}\left(d,t\right)\right)$ quantifies the deviation of the regional climate on a specific date from the summer average at hour $t$. By applying this correction as a scalar or low-frequency shift, the regional background is normalized without altering the high-frequency spatial gradients captured by the 70 m resolution ECOSTRESS data. This ensures that the observed diurnal variations primarily reflect the driving mechanisms of urban morphology rather than synoptic-scale weather fluctuations. Finally, all ECOSTRESS surface temperature data were cropped to the study area and uniformly resampled to 30 m using the nearest neighbor method.

2.5. Urban Morphological Metrics

Regarding variable selection, 25 potential variables were initially identified across seven categories: socioeconomic (e.g., Pop), environmental interference (e.g., PM2.5), landscape pattern, 2D and 3D urban morphology, and topographic features. To ensure model stability, a rigorous two-step screening process was employed to eliminate redundancy and multicollinearity. First, we performed a collinearity screening based on Pearson correlation coefficients, removing variables with coefficients greater than 0.8; for instance, nighttime light was excluded due to its high correlation with other socioeconomic indicators. Second, a screening based on the Variance Inflation Factor (VIF) was conducted, where variables with a VIF exceeding 10 were removed to avoid severe multicollinearity, leading to the exclusion of floor area ratio and the topographic wetness index.

Therefore, 22 variables (

Table 2. Urban morphological metrics in this article.

Categories	Variable Name	Description
Land cover Indicators	Surface albedo (SA)	Earth’s ability to reflect solar radiation.
	NDVI	Measures vegetation density and health.
	Water density (WD)	Percentage of the landscape area occupied by water bodies within a block.
2D Urban morphological Metrics	Building density (BD)	The ratio of building footprint area to total block area.
	Number of buildings (NB)	Total number of buildings within the study area.
	Mean area of projected architecture (MAPA)	Average area of a building’s vertical projection onto the ground.
3D Urban morphological Metrics	Building height (BH)	Average building height within the block.
	Building volume density (BVD)	The ratio of total building volume to land area.
	Number of high buildings (NHB)	Number of high-rise buildings in the study area.
	High building ratio (HBR)	The proportion of high-rise buildings in the total number of buildings.
Landscape pattern Metrics	Largest patch index (LPI)	Percentage of the landscape occupied by the largest patch, measuring dominance.
	Patch density (PD)	Number of patches per unit area, assessing landscape fragmentation.
	Landscape shape index (LSI)	Describe the complexity of the shape of the plaque.
	Edge density (ED)	Total edge length per unit area between heterogeneous landscape patches.
	Shannon’s diversity index (SHDI)	Measure landscape diversity and heterogeneity.
	Contagion (CONTAG)	Describe the spatial clustering or connectivity of different land cover types.
Topographical Indicators	Stream power index (SPI)	Represents the potential erosive power of overland water flow.
	Topographic position index (TPI)	Identifies landform positions relative to the local neighborhood.
Socioeconomic	Population density (Pop)	Number of permanent residents per unit area, reflecting anthropogenic heat intensity.
Environmental interference	PM2.5	$\mathrm{Atmospheric} \mathrm{aerosol} \mathrm{concentration} \mathrm{with} \mathrm{a} \mathrm{diameter} \le 2.5 \mathrm{\mu }\mathrm{m}$.

), including topographic feature indicators and biophysical parameters, were retained for further construction of a driving force analysis model for LST. These 22 variables were further aggregated into each block of the study area using various packages in Python 3.12.

2.6. Model Construction

Before building the model, we first ensured that the correlation coefficients between all independent variables were less than 0.8. To address the potential nonlinear relationship between the variables and the LST, we employed the GW-XGBoost model to capture spatial heterogeneity through local fitting. A stratified sampling method was used to divide the data into two parts: 70% for model training and 30% for validation. Furthermore, the model’s bandwidth expanded over time from 1270 (01:33) to 1470 (20:35), and the maximum tree depth at each time period was strictly controlled between 4 and 5 layers, ensuring the local accuracy and robustness of the driving mechanism analysis.

XGBoost is an improved gradient boosting tree algorithm that can capture nonlinear relationships and handle high-dimensional, multi-feature data [51]. It is widely used in LST modeling. Its objective function formula is as follows:

$$O\left(\varphi \right)=\sum _{i}l\left({\widehat{y}}_i,y_i\right)+\sum _n\Omega \left(f_n\right)$$

(2)

$$\Omega \left(f_n\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda {‖\omega ‖}^2$$

(3)

where $l({\widehat{y}}_i,y_i)$ is the loss function; $\Omega \left(f_n\right)$ is the regularization term; $\gamma$ represents the model complexity; $T$ is the number of leaf nodes; and $\omega$ is the weight vector of the leaf nodes.

However, standard XGBoost does not explicitly incorporate spatial context. To accurately characterize this spatial variation, this study introduces a GW-XGBoost model. Unlike feature-weighted [52] and neighborhood-feature-reconstruction-focused [39] models, this study employs a “locally weighted ensemble” strategy.

This strategy uses an adaptive bandwidth approach to address the uneven spatial distribution of urban monitoring points. For target point $i$, the k nearest samples are selected as the local training set ($k$ is the bandwidth parameter). Let $h_i$ be the distance to the $k$ nearest neighbor and use the Bi-square kernel function to calculate the spatial weight $w_{ij}$ of sample $j$. A local model $f_i$ is then constructed and optimized by minimizing the geographically weighted loss function:

$$w_{ij}=\left\{\begin{array}{c}{\left[1-{\left({\displaystyle \frac{{di}_j}{h_i}}\right)}^2\right]}^2,if {di}_j <h_i\\ 0, otherwise\end{array}\right.$$

(4)

$${\mathcal{L}}_{\mathcal{i}}=\sum _{j=1}^N{w}_{ij}\cdot l\left(y_j,\widehat{y_j}\right)+\mathrm{\Omega }\left(f_i\right)$$

(5)

where $l\left(y_j,\widehat{y_j}\right)$ represents the prediction error (e.g., squared error), and $\mathrm{\Omega }\left(f_i\right)$ is a regularization term that controls the complexity of the tree.

2.6.1. Model Evaluation

To evaluate the model’s performance, root mean squared error, mean absolute error, and coefficient of determination (R²) are used to assess the model’s fitting accuracy and generalization ability:

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

(6)

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

(7)

$$MAE={\displaystyle \frac{1}{n}}\sum _{i=1}^n\left|y_i-\widehat{y_i}\right|$$

(8)

where $y_i$ is the actual observed LST of the $i$-th sample; $\widehat{y_i}$ is the model’s predicted value for the $i$-th sample; $\overline{y}$ is the average LST of all observed samples; and $n$ is the total number of samples in the validation set.

2.6.2. Interpretability Analysis Based on SHAP

To explain the impact of each variable on LST in the XGBoost model, this study introduces the SHAP method based on cooperative game theory, as shown in the following formula:

$${\varphi }_i=\sum _{s\subseteq N\setminus \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(M-\left|S\right|-1\right)!}{M!}}\left[f_x\left(S\cup \left\{i\right\}\right)-f_x\left(S\right)\right]$$

(9)

where ${\varphi }_i$ represents the total SHAP value of feature $i$; $S$ is a subset that does not contain feature i; $\left|S\right|$ is the size of the subset; $M$ is the total number of features; and $f_x\left(S\right)$ is the model prediction value of subset $S$.

3. Results

3.1. Spatial Distribution of LST

To highlight the spatial heterogeneity of surface temperature, the study used a $2\mathrm{\sigma }$ (standard deviation) method as a threshold to stretch and enhance the images. The results show that Xi’an’s LST exhibits a significant “center-periphery” spatiotemporal differentiation characteristic (Figure 3). Spatially, high-temperature areas are mainly concentrated in the densely impermeable main urban area, exhibiting a typical UHI effect; while low-temperature areas are stably distributed in the suburbs with higher vegetation cover and around water bodies. Temporally, the spatial pattern of the thermal environment shows obvious dynamic migration characteristics: from early morning to morning, the high-temperature center moves from the main urban area to the northeast; from morning to noon, with the increase in solar radiation, the high-temperature area rapidly expands, and the center of gravity shifts from northeast to northwest; from afternoon to night, residual surface heat dissipates, and the high-temperature area gradually shrinks and flows back to the center of the main urban area.

In summary, the LST in the study area not only exhibits a “higher during the day and lower at night” temporal rhythm in terms of values, but the spatial morphology of its high-value areas also undergoes an evolutionary process of “dispersion-aggregation-migration-decline.” This spatiotemporal evolution is closely related to the spatiotemporal dynamics of the urban underlying surface properties and the intensity of human activities.

3.2. Deciphering the Global and Local Driving Mechanisms of Diurnal LST

The results showed that the model had the strongest explanatory power in the afternoon and evening, with R² reaching 0.6489 and 0.6876, respectively. In the morning period, R² was 0.2566.

The driving mechanisms of LST exhibit systematic diurnal transitions and pronounced spatial non-stationarity (Figure 4). On the global scale, NDVI governs daytime cooling through evapotranspiration, whereas PM2.5 dominates nocturnal warming via enhanced atmospheric back-radiation under stable boundary layers. Four distinct spatial patterns further characterize this relationship:

• Nocturnal surface albedo (SA)-PM2.5 synergy, where aerosol insulation restricts the cooling potential of high-reflectance surfaces.
• Noon NDVI-building density (BD) independence, reflecting the decoupled operation of biological and physical cooling paths.
• Pop centripetal driving, indicating the morning concentration of anthropogenic heat.
• PM2.5 patchy distribution, underscoring its dependence on microscale boundary layers.

Category-based clustering analysis reveals that environmental interference factors maintain a high contribution across all periods, with NDVI and SA serving as the primary cooling regulators during the day and night, respectively. Notably, the proportion of 3D urban morphological metrics escalates significantly at night, exhibiting a strong positive correlation with LST. Given Xi’an’s basin topography, increased building height enhances surface roughness and obstructs ventilation, resulting in significant heat entrapment within high-density architectural clusters (Figure 5).

3.3. Non-Monotonic Response and Interaction Threshold Representation

SHAP marginal analysis identifies critical inflection points for urban thermal regulation across multidimensional indicators. For environmental interference, PM2.5 exhibits distinct temporal patterns: a cooling effect emerges only when concentrations exceed 230.69 μg/m³ at midnight, while a U-shaped distribution is observed in the morning, with the cooling contribution concentrated within the 184.59–216.77 μg/m³ range. Regarding biophysical parameters, NDVI exhibits a cooling threshold at 0.17, beyond which its evapotranspiration efficiency increases exponentially. In terms of 2D morphology, BD turns into a major heat source once exceeding 0.22. In 3D contexts, building height (BH) demonstrates a “stepwise” warming effect above 10.83 m at night, consistent with the findings of previous studies regarding nocturnal heat interception [11,40].

Furthermore, Pop sensitivity peaks at low values (≈18.16), while large architectural footprints reach thermal saturation beyond 460.68 m at noon.

To further elucidate the interaction mechanisms of urban spatial morphology variables, this study selected the top four key features of variable importance in each time period for in-depth analysis. In the interaction effect analysis, locally weighted scatterplot smoothing regression was used to fit the curves, visually demonstrating the significance and uncertainty of the interaction. In Figure 6, the horizontal and vertical axes represent interacting variables, and the scatter point color reflects the change in SHAP interaction values, with lighter to darker colors representing decreasing values (Figure 7).

Significant non-monotonic interactions exist between aerosols and surface properties. Within the PM2.5 range of 220–238 μg/m³, high-volume clusters with a building volume density (BVD) ≥ 2.905 exhibit a synergistic cooling effect due to daytime solar shielding. Critically, SA only effectively counteracts PM2.5-induced warming ($>$222 μg/m³) when maintained within a specific window of 0.18–0.22. This suggests that under heavy pollution, the cooling marginality of high-reflectance surfaces is suppressed by aerosol insulation. Conversely, the extremely weak interaction between NDVI and BD (inflection at NDVI = 0.15) indicates that biological transpiration and physical shading operate through independent spatial pathways.

4. Discussion

4.1. The Influence of Urban Features on LST

Our study shows that urban spatial morphology exhibits significant spatial heterogeneity in its impact on daytime and nighttime LST. Furthermore, the results indicate that built environment variables have the greatest impact, with the overall average relative importance of built environment indicators reaching 29.5%, consistent with findings from many scholars [20,21]. Notably, this study found that the average relative importance of environmental interference and socioeconomic factors reached 18.5% and 12.25%, respectively, while traditional landscape morphology indicators had a relatively low impact on LST (22.25%). Consistent with the dominance of the BD indicator found in previous articles [40], this study further reveals that, as a basin city, Xi’an’s PM2.5 plays a more crucial role in hindering nighttime cooling through a heat-preserving effect induced by back radiation [31,53]. This finding corrects the bias in traditional studies that overemphasize NDVI while neglecting atmospheric background components and social activities [54,55]. It also emphasizes that in rapidly urbanizing areas, in addition to considering traditional urban morphology indicators, it is necessary to construct a comprehensive indicator system encompassing natural geography, human activities, and environmental factors to more fully reveal the complex driving mechanism of LST.

Furthermore, based on the analysis of diurnal variations in LST, combined with the complex nonlinear relationships between urban spatial morphology variables and their synergistic effects on LST, this study found that (1) during the early morning and nighttime periods, the driving force of LST shifted from local regulation dominated by SA in the early morning to a complex synergistic effect of PM2.5 and SA in the northeastern region at night. Combining the interaction effect of SA and PM2.5, this means that reducing SA when PM2.5 content is high and increasing SA when PM2.5 content is low helps to weaken the “heat blanket” phenomenon induced by PM2.5, enhancing atmospheric back radiation, thereby alleviating the problem of obstructed cooling of long-wave radiation on the ground surface at night [31]. (2) During the midday strong radiation period, NDVI showed a significant cooling effect, and its high contribution area was mainly distributed around the main urban area, corresponding to the significant cooling phenomenon of the ecological corridor [55,56]. At the same time, BD showed a strong heat interception effect in the western part of the study area, inducing significant warming. Meanwhile, the interaction effect between NDVI and BD was very weak, which suggests that under an extreme radiation environment, the transpiration cooling of vegetation and the shading/heat interception of buildings have relatively independent operating mechanisms in space. (3) Pop, as a socio-economic factor, showed a significant cooling phenomenon in the west and southwest at night, while it showed a warming phenomenon in the northeast. This differentiation mechanism is mainly attributed to the modulation effect of urban geometry on anthropogenic heat emissions [11,57,58]: The high-rise and high-density building clusters in the southwest significantly reduced the initial heat storage of the system through the shading mechanism of deep street valleys during the day, while the compact building layout in the northeast restricted local ventilation, making it difficult for anthropogenic heat emissions to diffuse and to generate a synergistic warming effect with aerosols [59].

4.2. Urban Planning Implications

With the deepening of global urbanization, Xi’an, a megacity located deep in the Guanzhong Basin, is facing an increasingly severe UHI challenge. The deterioration of its thermal environment not only affects residents’ thermal comfort but also threatens the city’s sustainable development. Although mitigating the heat island effect requires multi-dimensional and comprehensive actions, developing an overall intervention strategy from an urban planning perspective remains the core means of improving local microclimates.

In the northeastern region, surface temperature is mainly affected by environmental disturbances and surface parameters, which are synergistically dominant. Planning should prioritize blocking the thermal forcing between aerosols and the underlying surface. Delineating “local emission control zones” [60] to reduce PM2.5 concentration is crucial for restoring atmospheric transparency. Simultaneously, during community renovation, high-albedo surfaces (SA range of 0.18–0.22) [39,61,62] should be promoted to optimize long-wave radiation cooling efficiency and reduce nighttime heat accumulation. In the high-intensity development zones of the Midwest and South, dominated by socioeconomic factors and 3D urban morphological metrics, relevant regulations must optimize the interaction between anthropogenic heat and canopy physical structure. In densely populated areas (Pop > 21.8), the scientific layout of high-rise buildings (BH > 13.07 m) should be encouraged by increasing the BVD ($>$2.905). This utilizes “shading cooling” pathways to offset heat absorption by impermeable surfaces [63]. Furthermore, multi-layered ventilation networks utilizing waterways (e.g., areas through which rivers flow [64]) should be integrated to accelerate the turbulent diffusion of anthropogenic waste heat [65].

Planning must address the decoupling of biological cooling and physical shading pathways at midday. In areas with high NDVI values, connecting isolated patches into continuous, complex-shaped networks (e.g., landscape shape index, LSI > 3.024) can maximize evapotranspiration [66,67]. Conversely, in compact, sparsely vegetated neighborhoods (NDVI < 0.15), heat interception must be minimized by limiting BD ($<$0.196) and optimizing street orientation.

Furthermore, since the nighttime heat island profile is particularly pronounced in densely impermeable areas of the main urban area [68], low-rise buildings (BH < 10.31 m) should be guided to adopt a simple patchwork layout (LSI ≤ 3) to increase radiative heat dissipation. In sparsely developed areas, the “micro-greening” strategy of using pocket parks can effectively break the persistent heat island effect. These refined, site-specific interventions provide a scientific paradigm for climate-adaptive urban development in arid and semi-arid basins.

4.3. Limitations

Although this study systematically explored the mechanisms by which multi-source urban spatial morphology influences diurnal variations in land surface temperature, several limitations remain. While the ECOSTRESS land surface temperature data used in this study has unique advantages in monitoring land surface temperature, its spatial resolution limits the fine-grained capture of temperature changes in urban structure and microenvironment. Furthermore, the irregularity of satellite data acquisition times affects the analysis of short-term, high-frequency temperature variations. Additionally, due to the inherent limitations of thermal infrared remote sensing, this study primarily focuses on land surface temperature under clear-sky conditions, as thermal infrared remote sensing requires cloud-free scenes to accurately retrieve land surface temperature; future research should focus on urban heat island effect models under cloudy conditions, where changes in radiation balance may alter the driving mechanisms. Secondly, although this study currently relies on high-quality remote sensing products, it lacks integration of real-time ground observation data, such as Internet of Things (IoT) temperature sensors [68]. Future research incorporating these multi-source real-time datasets can significantly improve model validation and enhance the overall robustness and reliability of the proposed framework. Currently, this study uses Xi’an as an example to reveal the influence patterns of urban morphology under a specific semi-arid basin climate background. Since morphological characteristics and thermal response mechanisms vary depending on geographical environment, climate zone, and city size, future research should extend this assessment framework and model to more diverse urban samples for validation. This will not only help improve the generalizability of the research results but also provide more valuable empirical evidence for urban planning in different regions around the world.

5. Conclusions

This study quantifies the non-linear relationship between multi-source urban spatial morphology and diurnal LST in Xi’an using a GW-XGBoost framework integrated with SHAP interpretation. The primary findings are summarized as follows:

• Diurnal non-linearity of driving mechanisms: LST drivers exhibit a significant non-linear evolution across the diurnal cycle. NDVI serves as the core cooling agent during peak solar radiation (noon), while PM2.5 emerges as the dominant warming factor at night and dawn. This confirms that aerosols in the stable boundary layer hinder surface longwave radiation dissipation through an enhanced “atmospheric warming effect.” Consequently, urban planners should adopt a time-stratified approach to heat mitigation: prioritizing greening strategies for daytime cooling while focusing on “local emission control zones” and atmospheric transparency restoration to alleviate nocturnal heat trapping.
• Hierarchy and necessity of multi-source indicators: While built environment variables emerge as the primary drivers of diurnal LST (averaging 29.5% importance), the significant roles played by environmental interference (18.5%) and socioeconomic factors (12.25%) underscore the necessity of a multi-dimensional indicator system for precise thermal risk assessment. This hierarchy suggests that effective heat mitigation cannot rely solely on vegetation; planners must integrate socioeconomic density management and environmental pollution control into a holistic climate-resilient framework.
• Non-monotonic interaction thresholds: Synergistic effects between variables demonstrate distinct threshold characteristics. In the early morning, high-density clusters (BVD ≥ 2.905) exhibit a cooling effect when PM2.5 is within 220–238 μg/m³, attributable to 3D street-canyon shielding. Conversely, the cooling marginality of high-reflectance surfaces (SA) is suppressed by aerosol insulation at night, only stabilizing within a specific window of 0.18–0.22. These thresholds serve as quantitative benchmarks for refined design: in high-intensity zones, encouraging high-rise layouts (BH > 13.07 m) with BVD > 2.905 can utilize “shading cooling” pathways, while the promotion of high-albedo materials should be precisely targeted within the 0.18–0.22 range to optimize long-wave radiative dissipation under polluted conditions.
• Spatial non-stationarity and geographical differentiation: The influence of urban morphology shows significant geographical boundaries. The northeastern and northern sectors are primarily regulated by SA and PM2.5 interactions, with SA peaking in sensitivity across open areas. Meanwhile, BD induces significant warming in high-intensity western development zones. The extremely weak interaction between NDVI and BD suggests that biological transpiration and physical shading operate through independent spatial pathways. Therefore, differentiated spatial planning is required: connecting isolated green patches into complex networks (LSI > 3.024) can maximize evapotranspiration in ecological corridors, whereas compact neighborhoods with limited vegetation (NDVI < 0.15) must prioritize limiting building density (BD < 0.196) and optimizing street orientation to facilitate waste heat diffusion.

In summary, this study elucidates the complex, nonlinear, and nonstationary relationship between multi-source urban spatial morphology and diurnal surface temperature variation. By combining multi-source datasets with an interpretable machine learning framework, this research provides differentiated spatial planning strategies for different urban morphologies. Furthermore, the analytical framework developed in this paper provides a scalable and robust tool for global policymakers and urban planners. In the future, the application of this model in diverse geographical and climatic contexts will further deepen our understanding of urban thermodynamics, ultimately supporting global efforts to promote climate-resilient and sustainable urban development.