Characterizing the Thermal Effects of Urban Morphology Through Unsupervised Clustering and Explainable AI

Highlights

What are the main findings?

• A data-driven classification using the K-Means unsupervised clustering algorithm reveals that urban morphology is a primary determinant of the thermal environment. Among the classified types, ‘Compact Mid-rise’ exhibits the highest temperatures, whereas ‘Open High-rise’ is the coolest. The Normalized Difference Built-up Index (NDBI) is identified as the most significant warming factor, while the Sky View Factor (SVF) emerges as the most crucial cooling factor. However, the precise influence of these factors is highly contingent upon the specific urban morphology type.
• The influence of three-dimensional (3D) urban morphology on Land Surface Temperature (LST) is both nonlinear and dichotomous. For instance, within compact built-up areas, increasing Building Height (BH) and density presents a double-edged effect. On one hand, it can impede heat dissipation, leading to higher temperatures through a ‘heat trapping’ effect. On the other hand, it can provide a cooling benefit by blocking solar radiation via a ‘shading’ effect.

What is the implication of the main finding?

• These findings advocate for a shift in urban cooling strategies, moving away from a ‘one-size-fits-all’ approach towards precisely targeted policies tailored to different local urban morphologies. For instance, urban planning should prioritize the optimization of spatial building layouts in compact zones, whereas in open, low-density areas, the strategic deployment of green infrastructure should be the primary focus.
• Urban morphology significantly mediates the thermal effects of different land use functional zones. For functional zones with high anthropogenic heat emissions, such as industrial districts, planning interventions should favor sparse or open layouts to mitigate thermal stress on adjacent areas.

Abstract

The urban thermal environment poses a significant challenge to public health and sustainable urban development. Conventional pre-defined classification schemes, such as the Local Climate Zone (LCZ) system, often fail to capture the highly heterogeneous structure of complex urban areas, thus limiting their applicability. This study introduces a novel framework for urban thermal environment analysis, leveraging multi-source data and eXplainable Artificial Intelligence to investigate the driving mechanisms of Land Surface Temperature (LST) across various urban form types. Focusing on the area within Beijing’s 5th Ring Road, this study employs a K-Means clustering algorithm to classify urban blocks into nine distinct types based on their building morphology. Subsequently, an eXtreme Gradient Boosting (XGBoost) model, coupled with the SHapley Additive exPlanations (SHAP) method, is utilized to analyze the non-linear impacts of ten selected driving factors on LST. The findings reveal that: (1) The Compact Mid-rise type exhibits the highest annual average LST at 296.59 K, with a substantial difference of 11.29 K observed between the hottest and coldest block types. (2) SHAP analysis identifies the Normalized Difference Built-up Index (NDBI) as the most significant warming factor across all types, while the Sky View Factor (SVF) plays a crucial cooling role in high-rise areas. Conversely, road density (RD) shows a negative correlation with LST in Open Low-rise areas. (3) The influence of urban form is twofold: increased building height (BH) can induce warming by trapping heat while simultaneously providing a cooling effect through shading. (4) The impact of land use functional zones on LST is significantly modulated by urban form, with temperature differences of up to 2 K observed between different functional zones within compact block types. The analytical framework proposed herein holds significant theoretical and practical implications for achieving fine-grained thermal environment governance and fostering sustainable development in the context of global urbanization.

1. Introduction

Since the Industrial Revolution, the global urbanization process has accelerated at an unprecedented pace, characterized by the migration of populations to urban centers and the continuous expansion of urban space. This process has led to the extensive replacement of original natural pervious surfaces with artificial, non-evaporating, and impervious materials such as asphalt and concrete. This transformation profoundly alters the surface energy balance, giving rise to the increasingly severe Surface Urban Heat Island (SUHI) phenomenon, where urban areas are significantly warmer than their surrounding rural counterparts [1,2]. The SUHI effect not only poses a direct threat to resident health and comfort but is also intricately linked to a city’s overall energy consumption, air quality, ecological stability, and sustainable development goals [3]. Consequently, an in-depth, quantitative investigation into the urban thermal environment is of paramount importance for effectively mitigating SUHI effects and enhancing the quality of urban living.

A considerable body of research has investigated the driving mechanisms of the urban thermal environment, establishing urban morphology as a critical determinant of the local thermal climate [4,5]. These drivers are typically analyzed across different dimensions. From a two-dimensional (2D) perspective, the prevalence of impervious surfaces—comprising buildings and roads—is confirmed as a primary contributor to the SUHI effect. Conversely, vegetation and water bodies within the surface composition exert a significant cooling influence through evapotranspiration and high thermal inertia, respectively. As cities continue to expand vertically, three-dimensional (3D) spatial structure indicators, such as building height (BH) and Sky View Factor (SVF), have been demonstrated to have complex, often dual, effects on thermal energy accumulation and dissipation [6,7].

This study constructs an analytical framework for the urban thermal environment by leveraging multi-source data and eXplainable Artificial Intelligence (XAI). A key methodological feature is the adoption of irregular urban blocks, delineated by the organic structure of OpenStreetMap (OSM) road networks, as the fundamental analytical units. This approach better aligns with the authentic morphological features of the urban space compared to traditional grid-based methods [8]. Building on this block-level foundation, key metrics such as building density and average height are extracted and utilized in a K-Means unsupervised clustering algorithm. This process establishes a data-driven classification system that reflects region-specific patterns of building utilization intensity. To thoroughly investigate the drivers of Land Surface Temperature (LST), a comprehensive indicator system is developed, encompassing 2D planar patterns, 3D spatial structure, and land cover characteristics. Subsequently, an XGBoost model, coupled with the SHapley Additive exPlanations (SHAP) method, is employed to quantitatively analyze and explain the contribution and mechanism of each driving factor across the different block types. Unlike methodologies based on the Local Climate Zone (LCZ) system, the contribution of this research is not to propose a universal replacement, but rather to offer a complementary analytical tool that is highly adaptive to the local context.

2. Literature Review

The urban thermal environment pertains to the heat-related physical conditions within urbanized areas, directly influencing regional energy consumption, ecological balance, and human health and comfort [9]. The SUHI is a phenomenon characterized by anomalously higher LST in urban areas, as retrieved from remote sensing data. As a key parameter in the surface energy balance, LST intuitively reflects the composite effects of land cover and urban morphology on the thermal environment. Given its capacity for large-scale, high-precision monitoring via remote sensing technology, LST has been widely adopted as a core proxy indicator for characterizing the SUHI [10,11].

The spatial differentiation of urban LST is the result of the combined effects of multiple factors, which have been primarily investigated across three dimensions: 2D planar patterns, 3D spatial morphology, and land cover composition. Two-dimensional planar characteristics are fundamental in shaping the urban thermal environment and can be further distinguished into land cover composition and planar pattern metrics. Changes in LULC are one of the most critical factors influencing the SUHI [12,13]. Urban built-up areas and ISP significantly elevate LST due to their thermodynamic properties, such as low albedo and high heat capacity. Conversely, urban green spaces and water bodies typically exert a significant cooling effect; vegetation dissipates latent heat through evapotranspiration, while water bodies generate a distinct “cool island” effect owing to their high specific heat capacity. To quantitatively assess these land cover features, researchers widely employ spectral indices derived from remote sensing data. NDVI and NDBI are the most extensively used indicators to quantify the density of vegetation and built-up areas, respectively [14,15]. Concurrently, MNDWI is effective for extracting water body information. However, the relationships between these indices and LST are not simply linear; the strength and even the direction of their correlations are highly context-dependent, influenced by climate background, seasonal variations, and spatial scale. Within the built-up environment, planar pattern metrics further describe the physical layout. BCD and ISP jointly determine the surface’s capacity to absorb and store solar radiation, making them primary drivers of the SUHI effect. RD not only represents impervious road surfaces but is also associated with anthropogenic heat emissions from traffic. Additionally, BC is often used as a supplementary metric to gauge development intensity [16,17,18].

As cities continue to expand vertically, 2D analysis alone has become insufficient to fully explain the variation patterns of LST [19,20]. Consequently, 3D urban morphology indicators—such as BH, BCD, FAR, and SVF—are increasingly being integrated into analytical frameworks [21,22]. Existing research confirms that the impact of 3D morphology on LST is complex and non-linear. For instance, increasing building height and density can reduce the SVF, which traps longwave radiation and inhibits nocturnal cooling (a warming effect), while simultaneously providing greater shading during the day (a cooling effect) [23]. The magnitude and direction of a driving factor’s influence on LST exhibit significant scale dependency, making the selection of a scientific unit of analysis critically important. While early studies often employed regular grids, this approach is arbitrary and can fragment urban areas with cohesive morphological and functional features. To overcome this limitation, the LCZ classification system proposed by Stewart and Oke [24,25] provides a standardized analytical framework that has greatly advanced the comparability of urban thermal environment studies globally. By classifying urban landscapes into 17 standard types based on pre-defined criteria for building morphology and land cover, the LCZ scheme effectively reveals temperature differences among them; for instance, compact mid- and high-rise zones are typically warmest, whereas vegetated and water-covered zones are coolest [26,27,28]. However, the LCZ system has limitations when applied to highly heterogeneous cities. As a “top-down,” standardized framework, its universal thresholds may not fully adapt to the unique morphological characteristics of a specific city [29,30]. Although subsequent developments like the World Urban Database and Access Portal Tools (WUDAPT) [31] project and various city-specific calibration methods—such as GIS-based rule-driven classifications [32], supervised and spatial clustering approaches [33], and automated statistical methods [34]—have enhanced the local adaptability of LCZ, they may still involve subjective judgments and might not perfectly match the actual structural patterns of a city.

This highlights a key research gap: the lack of a “bottom-up” analytical approach that is adaptive to local urban characteristics. Existing classification systems largely rely on pre-set standards, which can struggle to truly reflect the diversity of urban structure and function. Addressing this, our study proposes a data-driven framework based on irregular urban blocks and unsupervised clustering. The objective is to derive a classification system from the city’s actual morphology, thereby creating a more locally attuned typology for accurately identifying and analyzing the drivers of the internal urban thermal environment.

3. Data and Methods

3.1. Study Area

The study area for this research encompasses the region within the Fifth Ring Road of Beijing, the capital of China, as shown in Figure 1. This area, covering approximately 710 km², constitutes Beijing’s core functional zone. It hosts the city’s highest population density and the largest proportion of its employment, economic, cultural, and recreational activities, representing a highly mature stage of urban development. The urban morphology within this zone exhibits extreme complexity and heterogeneity; it juxtaposes strictly protected low-rise historic districts with modern mid-rise and high-rise residential and commercial zones.

3.2. Data Sources and Pre-Processing

This study utilizes a suite of multi-source geospatial data, including remote sensing imagery, building vectors, road networks, and impervious surfaces (as detailed in

Table 1. Summary of Data Sources and Descriptions.

Data Category	Data Source	Spatial Resolution	Description and Purpose
Road Network Data	OpenStreetMap (OSM) [35]	Vector	Calculation of RD and delineation of irregular urban blocks
Building Vector Data	The World’s First National-Scale Multi-Attribute Building Dataset (CMAB) [36]	Vector	Calculation of morphology indicators, including BC, BCD, BH, and FAR
Remote Sensing Imagery	United States Geological Survey (USGS)	30 m	Retrieval of LST and calculation of land cover indices, namely NDVI, NDBI, and MNDWI
Impervious Surface Data	Global Artificial Impervious Area (GALA) Dataset [37]	30 m	Calculation of ISP
Urban Functional Zones	Essential Urban Land Use Categories of China (EULUC-China) [38]	30 m	Used for the classification of urban functional zones.

). To ensure temporal consistency, the baseline year for all data was standardized to 2019. All datasets were reprojected to the GCS WGS 1984, UTM Zone 50N coordinate system, with a uniform resolution of 30 m. Raster datasets, such as LST, land cover indices (NDVI, NDBI, MNDWI), and the Global Artificial Impervious Area (GAIA) data, were aggregated to the irregular urban block units using a zonal statistics approach. Specifically, the mean value of all raster pixels whose centroids fall within a given block polygon was calculated and assigned to that block. The annual average LST for 2019 was computed by calculating the arithmetic mean of all available cloud-free monthly LST images, thereby applying equal weighting to each valid monthly observation. Prior to this calculation, pixels contaminated by clouds, cloud shadows, and cirrus were masked out from each monthly image using its corresponding quality assessment band. This process ensures that the annual average is derived exclusively from high-quality, clear-sky pixels, under the assumption that each monthly image provides a representative snapshot of the thermal conditions for that month. The specific data sources and processing methods are detailed in the table below.

3.3. Methodological Framework

The technical framework of this study comprises three main steps, as illustrated in Figure 2:

(1)

Constructing a Data-Driven Classification System for Urban Blocks. This initial step involves delineating irregular urban blocks based on authentic road networks. Subsequently, a K-Means clustering algorithm is applied to classify these blocks into distinct urban structure types based on their internal BCD and average BH. This process establishes a classification system that reflects varying levels of building utilization intensity.

(2)

Calculating Multi-dimensional Drivers and Retrieving LST. A suite of 10 driving factors is computed from multi-source data, encompassing metrics of 2D planar patterns, 3D spatial structure, and land cover composition. Concurrently, block-scale LST is retrieved from Landsat 8 imagery using established radiometric correction and mono-window algorithms.

(3)

Statistical Modeling and Interpretability Analysis. A preliminary exploratory analysis is first conducted using Spearman’s rank correlation. Following this, nine distinct XGBoost regression models are constructed and trained, one for each urban structure type, to model the LST. Finally, the SHAP method is employed for an in-depth interpretation of the trained models. This allows for the quantitative dissection of the non-linear influence and context-dependent importance of each driver on LST.

Figure 2. Methodological framework of this study.

Figure 2. Methodological framework of this study.

3.4. Key Methodological Steps

3.4.1. LST Retrieval and Validation

In this study, LST was retrieved from the thermal infrared bands of monthly Landsat 8 images for the year 2019. LST retrieval algorithms can be categorized into four main types: (1) the radiative transfer equation method [39]; (2) single-channel algorithms [40]; (3) dual-channel methods (i.e., the split-window algorithm) [41]; and (4) multi-angle algorithms [42]. Landsat 8 is equipped with two thermal infrared bands (Bands 10 and 11). However, due to significant calibration uncertainties associated with Band 11, the United States Geological Survey (USGS) recommends against the use of the split-window algorithm, advising instead a single-channel algorithm based on Band 10 for LST retrieval [43]; Accordingly, this study adopted the mono-window algorithm developed by Qin et al. [44,45], which is derived from the radiative transfer equation. The algorithm is particularly suitable for Landsat 8 data as it can be applied without the need for atmospheric correction. The governing equation is as follows.

$${\mathrm{T}}_{\mathrm{s}}=\left\{\mathrm{a}\left(1-\mathrm{C}-\mathrm{D}\right)+\left(\mathrm{b}\left(1-\mathrm{C}-\mathrm{D}\right)+\mathrm{C}+\mathrm{D}\right){\mathrm{T}}_{\mathrm{s}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{o}\mathrm{r}}-\mathrm{D}{\mathrm{T}}_{\mathrm{a}}\right\}/\mathrm{C}$$

(1)

where ${\mathrm{T}}_{\mathrm{s}}$ is the brightness temperature at the sensor, T_a is the mean atmospheric temperature, and a and b are atmospheric function parameters. For a surface temperature range of 0–70 °C, a = −67.355351 and b = 0.458606. $\mathrm{C}$ and $\mathrm{D}$ are intermediate variables, calculated as follows:

$$\mathrm{C}=\mathsf{\tau }$$

(2)

$$\mathrm{D}=\left(1-\mathsf{\tau }\right)\left[1+\left(1-\mathsf{\epsilon }\right)\right]\mathsf{\tau }$$

(3)

In Equations (2) and (3), ε is the land surface emissivity and τ is the total atmospheric transmittance from the ground to the sensor.

The resulting LST maps for the 12 months were then used to compute a weighted annual average LST spatial distribution map. To validate the accuracy of our results, the retrieved LST values were compared against an open-source temperature product available on the Google Earth Engine (GEE) platform. The comparison, conducted across more than 740,000 validation points, yielded a coefficient of determination (R²) of 0.86 and a Root Mean Square Error (RMSE) of 0.90 K. These metrics indicate a strong agreement and affirm the high reliability of the retrieved LST data for subsequent analysis, as illustrated in Figure 3:

3.4.2. Classification of Urban Morphology Types

To overcome the subjectivity and limited local adaptability associated with the pre-set thresholds of traditional research methods, this study employs a “bottom-up,” data-driven approach to classify urban morphology. This method allows for the creation of a classification system that objectively reflects the intrinsic built-environment patterns of Beijing.

Initially, 3875 enclosed block polygons were generated from pre-processed OSM road network data to serve as the basic analytical units. To address the issue of redundant road types within the OSM dataset, this study filtered out specific categories, including bridleways, steps, cycleways, and unknown paths. Following a topological check of the OSM road network, the centerlines of the filtered roads were extracted. A 50 m buffer was then created on both sides of these centerlines to generate the final road data. The study area was subsequently partitioned into blocks using these street centerlines as boundaries for further data processing. Finally, the building coverage ratio was calculated for each block, and blocks with a coverage ratio of less than 0.5% were classified as non-built-up areas.

A variety of machine learning and data-driven methods have been employed to investigate the relationship between urban morphology and climate, including Random Forest regression [46], hierarchical clustering [47], and K-Means clustering [48]. These approaches are adept at handling the heterogeneity of the urban built environment and its impact on the local climate. Among the numerous clustering algorithms available, this study selected the K-Means algorithm. The algorithm was selected for its computational efficiency and intuitive principles; it operates by assigning each unit to the cluster with the nearest centroid, thereby ensuring that intra-cluster variance is minimized. Consequently, it is widely applied in the field of urban studies [49,50].

For the selection of classification variables, this study referenced the morphological research framework for major Chinese cities developed by Long et al. [51,52], selecting BCD and average BH. These two indicators represent the most fundamental dimensions of 3D urban spatial form and collectively define building volume and development intensity [53]. The K-Means algorithm was then independently applied to these two dimensions. Implemented in Python 3.12 using the scikit-learn library, the algorithm partitions the block vectors into k clusters by assigning each vector to the cluster with the nearest centroid (i.e., the vector of mean parameter values). K-Means clustering aims to minimize the squared Euclidean distance, a characteristic highly suitable for the block-level input data. While the elbow method suggested a k value between 3 and 5, this study, referencing LCZ classification system, set the number of clusters (k) to 3 for both dimensions. By constructing a clear, interpretable 3 × 3 classification matrix, the three BCD categories (‘Compact,’ ‘Sparse,’ and ‘Open’) were combined with the three BH categories (‘Low-rise,’ ‘Mid-rise,’ and ‘High-rise’), ultimately forming nine distinct types of urban blocks. This classification system is illustrated in Figure 4.

3.4.3. Selection of Driving Factors

The urban thermal environment is a complex, dynamic system influenced by a multitude of factors [54], with the calculation methods and roles of the key indicators detailed in

Table 2. Definition and Calculation of Driving Factors.

Dimension	Factor	Abbreviation	Formula	Description
2D Morphology	Building Count [58]	BC	$n$	Total number of buildings within a block.
	Building Coverage Ratio [59]	BCD	$BCR={\displaystyle \frac{{\sum }_{i=1}^n{S}_i}{S\_site}}$	The ratio of the total building footprint area to the total block area.
	Impervious Surface Percentage [23]	ISP	$ISP={\displaystyle \frac{{\sum }_{i=1}^n{SIS}_i}{S\_site}}$	The ratio of impervious surface area to the total area of a given region.
	Road Density [60]	RD	$RD={\displaystyle \frac{{\sum }_{i=1}^n{SR}_i}{S\_site}}$	The ratio of the total road network area to the total block area.
3D Morphology	Floor Area Ratio [61]	FAR	$FAR={\displaystyle \frac{{\sum }_{i=1}^n{S}_i\ast N_i}{S\_site}}$	The ratio of the total floor area of all buildings to the total block area.
	Building Height [62]	BH	$BH={\displaystyle \frac{{\sum }_{i=1}^n{F}_i\ast ∆H}{BC}}$	The average building height within a block.
	Sky View Factor [63]	SVF	$SVF={\displaystyle \frac{{\sum }_{i=1}^n{SVF}_i}{n}}$	The ratio of the radiation received from the sky to the total radiation emitted by the entire hemispheric environment.
Land Cover	Normalized Difference Vegetation Index [64]	NDVI	$NDVI={\displaystyle \frac{{\rho }_{NIR}-{\rho }_R}{{\rho }_{NIR}+{\rho }_R}}$	Quantifies vegetation by measuring the difference between near-infrared and red light.
	Modified Normalized Difference Water Index [65]	MNDWI	$MNDWI={\displaystyle \frac{{\rho }_G-{\rho }_{MIR}}{{\rho }_G+{\rho }_{MIR}}}$	Enhances the detection of open water features while suppressing noise from built-up land and vegetation.
	Normalized Difference Built-up Index [66]	NDBI	$NDBI={\displaystyle \frac{{\rho }_{SWIR}-{\rho }_{NIR}}{{\rho }_{SWIR}+{\rho }_{NIR}}}$	Used to extract and map built-up areas.

. Informed by a comprehensive review of existing literature [23,26,46,55], this study selected a suite of 10 driving factors that have been widely demonstrated to influence LST. These factors span three dimensions: 2D planar patterns, 3D spatial structure, and land cover composition [13,56,57], and were integrated to construct a comprehensive analytical indicator system. The selection process was guided by three main criteria: each factor had to be (1) widely recognized and applied within the research community, (2) proven effective in both theory and practice, and (3) non-redundant to avoid multicollinearity. The specific calculation method for each factor was based on established literature and implemented using the ArcGIS 10.8 and GEE platforms. Finally, all indicator values were aggregated at the level of each irregular block unit to form a comprehensive panel dataset for subsequent analysis.

3.4.4. Statistical Analysis and Explainable Modeling

(1)

A preliminary exploratory analysis was conducted using Spearman’s rank correlation coefficient to probe the monotonic relationships between each driving factor and LST. This non-parametric method was chosen for its robustness, as it does not require the data to conform to a normal distribution [67].

(2)

To capture the complex non-linear relationships and interaction effects among the driving factors, this study employed XGBoost algorithm [68]. XGBoost is an ensemble learning framework based on decision trees, renowned for its high prediction accuracy and strong robustness against overfitting. The algorithm builds the model sequentially by iteratively adding new decision trees to correct the prediction errors from the previous iteration. As an optimized gradient boosting tree model, XGBoost excels at handling complex non-linear relationships and has been proven to outperform other tree-based machine learning models, such as Random Forest and Support Vector Machines, in terms of both prediction accuracy and computational efficiency [69]. It finds wide application in urban studies, for tasks such as urban heat island effect analysis [70], electricity consumption forecasting [71], and carbon emission research [72]. For each of the nine distinct urban block types, we trained an independent XGBoost regression model with LST as the dependent variable and the 10 driving factors as independent variables. This tailored approach allows us to reveal the differential and context-dependent nature of the driving mechanisms across various urban environmental settings.

The optimization of the model is assessed through its objective function, expressed as follows:

$$\mathrm{O}\mathrm{b}\mathrm{j}\left(\mathrm{X}\right)=\mathrm{L}\left(\mathrm{X}\right)+\mathsf{\Omega }\left(\mathrm{X}\right)$$

(4)

where $\mathrm{O}\mathrm{b}\mathrm{j}\left(\mathrm{X}\right)$ is the objective function, which quantifies the structural risk of the model given the current parameters; $\mathrm{L}\left(\mathrm{X}\right)$ is the training loss function, which evaluates the model’s goodness of fit on the training dataset; and $\mathsf{\Omega }\left(\mathrm{X}\right)$ is the regularization term, which measures the model’s complexity to mitigate issues such as overfitting and instability.

To prevent overfitting and enhance predictive performance [73], this study utilized Optuna for hyperparameter tuning, performing 50 search iterations combined with 5-fold cross-validation to ensure model robustness. The dataset was randomly partitioned into an 80% training set and a 20% testing set for model validation. The model’s accuracy was evaluated using metrics such as RMSE, Mean Absolute Error (MAE), and R². The optimized hyperparameters include: n_estimators = 400 (number of decision trees), max_depth = 9 (maximum tree depth), learning_rate = 0.0391, subsample = 0.6759, and gamma = 0.0023. On the test set, the predictive model achieved an MAE of 0.218, an RMSE of 0.576, and an MSE of 0.332.

(3)

Although the XGBoost model possesses powerful predictive capabilities, it is often considered a “black box,” making it difficult to interpret the logic behind its predictions [74]. To address this issue, the SHAP method was introduced to provide post hoc model explanation [75]. SHAP is a game theory-based approach that quantifies the impact of each feature by calculating its marginal contribution to the prediction for each individual sample. Unlike traditional metrics that only provide a global ranking of feature importance, SHAP values can precisely reflect both the magnitude and direction of each feature’s influence [76]. In recent years, the SHAP method has been widely applied in the environmental sciences to reveal the complex internal mechanisms of machine learning models [77,78], thereby helping us to understand the driving factors of the urban thermal environment and their modes of action in a more in-depth and transparent manner.

In this study, SHAP values reflect the relative importance of each feature in influencing the average LST of the block samples, while partial dependence plots are used to quantify both the linear and non-linear effects of the features on the model. This research utilized the “shap” package in Python 3.12 to perform the relevant calculations. The mathematical expression for the Shapley value is defined as follows:

$$\mathrm{SHAP}=\sum _{\mathrm{S}\subseteq \mathrm{N}\setminus \left\{\mathrm{i}\right\}} {\displaystyle \frac{\left|\mathrm{S}\right|!\left(\mathrm{M}-\left|\mathrm{S}\right|-1\right)!}{\mathrm{M}!}}\left[{\mathrm{f}}_{\mathrm{x}}\left(\mathrm{S}\cup \left\{\mathrm{i}\right\}\right)-{\mathrm{f}}_{\mathrm{x}}\left(\mathrm{S}\right)\right]$$

(5)

where $\mathrm{SHAP}$ denotes the SHAP value for feature $\mathrm{i}$, with its magnitude indicating the feature’s impact on the model’s prediction. $\mathrm{S}$ represents the vector of feature values for a given sample, while $\mathrm{M}$ is the total number of features. The term ${\displaystyle \frac{\left|\mathrm{S}\right|!\left(\mathrm{M}-\left|\mathrm{S}\right|-1\right)!}{\mathrm{M}!}}$ is a weighting factor determined by the number of possible feature subsets that include feature $\mathrm{i}$. The term ${\mathrm{f}}_{\mathrm{x}}\left(\mathrm{S}\right)$ represents the model’s prediction for a given feature subset $\mathrm{S}$; features not included in $\mathrm{S}$ are considered marginalized (i.e., their influence is averaged out). By calculating a weighted average of a feature’s marginal contributions across all possible feature combinations, the SHAP value ensures a fair and equitable distribution of each feature’s contribution to the final prediction.

4. Results

4.1. Spatial Patterns of the Urban Thermal Environment and Morphology Types

Utilizing the K-Means clustering algorithm, this study classified the 3875 urban blocks within the study area into nine distinct types based on building utilization intensity. Compact Mid-rise was identified as the most prevalent block type, accounting for 31.26% of the total, whereas Open Low-rise was the least common, at only 1.70%. The spatial distribution of these types, illustrated in Figure 5, reveals distinct patterns. Broadly, there is an increasing trend in average building height moving from the city center to the periphery, while building density is highest in both the urban core and certain peripheral areas. The Compact Low-rise and Sparse Low-rise types are highly concentrated within the old city core inside the Second Ring Road. This area is a vital historic district where restrictions on building height and development intensity, aimed at preserving historical features such as hutongs and siheyuan courtyard houses, have maintained an urban form dominated by low-rise structures. The various Compact types collectively represent the largest proportion of blocks and form the main body of the urban built-up area between the Second and Fourth Ring Roads. This zone accommodates a high concentration of residential, scientific, educational, and administrative functions, resulting in a typical urban morphology of high density and medium building height. Finally, High-rise morphologies are predominantly located in the newly developed zones between the Third and Fifth Ring Roads. These zones are characterized by high-intensity land development and a concentration of high-rise and super-high-rise buildings.

An overlay analysis was conducted between the LST data and the urban block vectors to investigate the thermal performance of different morphology types. Given the significant disparities in the total area covered by each type, an area-weighted mean LST was calculated to ensure a robust comparison. The results, presented in Figure 6, reveal statistically significant differences in LST across the nine morphology types. A one-way Analysis of Variance (ANOVA) confirmed that the effect of block morphology on LST was highly significant (F = 30.85, p < 0.001). Post hoc comparisons using the Tukey HSD test further elucidated these thermal distinctions. For instance, Compact Mid-rise blocks were significantly warmer than nearly all other types, including the cooler Sparse High-rise and Open High-rise types. Notably, for a given building height category, compact types consistently exhibited higher median temperatures than their sparse and open counterparts. This finding is consistent with existing literature: Compact Mid-rise blocks displayed the highest area-weighted mean LST at 296.74 K, confirming that high-density, medium-height urban fabric most strongly promotes surface warming. Conversely, Open High-rise blocks recorded the lowest mean LST at 295.56 K. The LST differential between the hottest and coldest individual blocks reached as high as 11.29 K, underscoring the profound impact of urban morphology on the thermal environment.

The spatial distribution of the annual average LST (Figure 7) reveals significant spatial heterogeneity within the study area. The overall pattern indicates that the central urban core and southern districts exhibit higher temperatures, while the northern and peripheral areas are relatively cooler. When analyzed by density, the thermal disparities among functional zones within Compact type blocks are more pronounced, likely due to high building density and intense human activity. As the morphology transitions towards Sparse and Open types, these inter-zonal temperature differences progressively diminish, with Open types showing the minimal thermal variation.

In summary, the morphological characteristics of urban blocks, particularly building density and spatial layout, exert a significant regulatory effect on the local thermal environment. High-density, medium-height built environments are most prone to exacerbating surface thermal effects, whereas high-rise layouts with greater spatial openness can contribute to mitigating heat island intensity.

4.2. Analysis of LST Driving Mechanisms

4.2.1. Correlation Between Driving Factors and LST

A Spearman correlation analysis was conducted to examine the relationship between the 10 driving factors and LST across the different urban structure types. The results are visualized in the heatmap in Figure 8, which displays the correlation coefficient (ρ) for each factor within each of the nine morphology types. The y-axis lists the block categories, with the sample size (n) indicating the number of blocks included in each analysis. The statistical significance of each correlation is denoted by asterisks, which correspond to the p-value from the test. Among all indicators, NDBI exhibited the strongest positive correlation with LST, with a coefficient of 0.4383. This positive relationship was consistent across all block types, confirming that a higher proportion of built-up surfaces leads to a warming effect, a phenomenon particularly pronounced in high-rise environments. Conversely, SVF demonstrated a consistent negative correlation, which was most significant in high-rise areas. This supports the physical mechanism that open spaces facilitate the dissipation of longwave radiation, thereby providing a cooling effect. Notably, RD presented an interesting anomaly. While this indicator showed weak correlations in most block types, it displayed a strong negative correlation within Open Low-rise blocks. This finding, contrary to the intuition that roads typically absorb heat, may suggest that roads in these specific areas possess cooling characteristics, such as being shaded by extensive tree canopies. From an overall perspective, although most indicators were statistically significant, their correlation strengths were generally weak to moderate. This phenomenon implies that an aggregated analysis mixing all urban structure types would mask the internal complexity and heterogeneity.

The sample size (n) for each category is provided in parentheses. The color of each cell represents the strength and direction of the correlation, with red indicating a positive relationship (warming effect) and green indicating a negative relationship (cooling effect). Statistical significance is denoted by asterisks based on the p-value: * p < 0.05, ** p < 0.01, *** p < 0.001.

4.2.2. Non-Linearity and Context-Dependency of LST Driving Mechanisms

To unravel the complex mechanisms driving LST, independent XGBoost models were developed for each of the nine urban block types. The models demonstrate a robust goodness of fit, with coefficients of determination R² ranging from 0.62 to 0.93. This indicates that the selected drivers effectively explain the majority of LST variance across different urban contexts. The SHAP analysis provides deep insights into these relationships, revealing the non-linear characteristics and strong context-dependency of each factor’s impact on LST. Overall, land cover composition and 3D morphology emerge as the dominant factors determining LST.

The feature importance analysis across all block types identifies NDBI as the most influential variable, followed closely by SVF and ISP (Figure 9). Artificial surfaces, represented by NDBI and ISP, directly amplify surface energy fluxes due to their high thermal capacity, low albedo, and negligible evapotranspiration, forming the fundamental driver of UHI effect. Three-dimensional building morphology, represented by SVF, plays a critical regulatory role by modulating the energy exchange processes between the urban surface and the sky. Notably, while NDVI emerges as a powerful explanatory factor in Open Low-rise blocks, its importance diminishes significantly in the majority of medium and high-density areas. This finding does not imply the failure of green infrastructure but rather suggests a strategic hierarchy for thermal mitigation in dense urban environments. The primary focus should be on reducing the thermal contribution of artificial surfaces, which can then be supplemented by the optimized configuration of vegetation.

4.2.3. SHAP-Based Feature Importance Analysis Using the XGBoost Model

To provide a deeper insight into the complex driving mechanisms of LST, this study developed independent XGBoost models for each of the nine urban block types, as shown in Figure 10. The models’ performance was evaluated using a five-fold cross-validation, yielding an R² of 0.78 and RMSE of 0.59 K. Furthermore, a diagnostic test for spatial autocorrelation was performed on the model residuals. The Global Moran’s I statistic was 0.012 (p = 0.21), which is not statistically significant, indicating that the residuals are randomly distributed in space with no discernible clustering pattern. This result suggests that the models effectively captured the underlying spatial processes influencing LST. The SHAP analysis reveals significant spatial heterogeneity in the LST driving factors across the different block morphologies. This analysis is based on feature importance rankings, as illustrated in the subsequent plots. In these plots, red or purple dots represent high values of a feature, while blue dots represent low values. Positive SHAP values on the horizontal axis indicate a warming effect on the LST prediction, whereas negative values signify a cooling effect.

The analysis of the urban thermal environment’s driving mechanisms shows that built-up areas of different densities exhibit distinctly different characteristics. In high-density and medium-density zones, the driving mechanisms are remarkably consistent. NDBI consistently emerges as the most important influencing factor, followed by SVF. This indicates that the proportion of building cover is the primary determinant of LST, while the degree of sky openness, dictated by the building enclosure, is the key factor governing heat dissipation. In these building-saturated zones, LST is principally governed by a combination of surface material properties and heat dissipation efficiency. Notably, high values for 3D morphological factors representing building volume such as FAR, BH, and BCD generally exhibit a cooling effect. This suggests that in high-density areas, the daytime shading provided by buildings can outweigh their heat-trapping effect. In stark contrast, low-density blocks exhibit the greatest heterogeneity in their driving mechanisms, with the dominant factor changing significantly as a function of building height. In low-rise areas, NDVI is the dominant factor. This shifts to ISP in mid-rise areas, and further transitions to SVF in high-rise areas. This demonstrates the intricate complexity of thermal regulation mechanisms under different building height scenarios within low-density environments.

An analysis of the thermal driving mechanisms across different building height categories reveals that the influence of urban morphology on LST is strongly height-dependent. In Low-Rise Blocks. For Compact Low-rise and Sparse Low-rise blocks, SVF is the most important cooling factor, while NDBI is the primary warming factor. This reflects the direct impacts of sky openness and building density. Notably, FAR exhibits a cooling effect in these types, suggesting that the shading provided by building volume can outweigh its heat accumulation. In Open Low-rise blocks, NDVI emerges as the dominant cooling factor, and BH also shows a significant cooling influence. In Mid-Rise Blocks. NDBI and SVF continue their dominant roles in mid-rise blocks. BH demonstrates a clear cooling effect, further confirming the efficacy of increasing building height to enhance shading in dense areas. ISP is a prominent warming factor, highlighting the direct thermal contribution of surface materials. Within Open Mid-rise blocks, FAR emerges as the most critical cooling driver. In High-Rise Blocks. The warming effect of NDBI and the cooling effect of SVF are maximized in high-rise environments. A core finding is that high values of BCD produce a net cooling effect, providing strong evidence for the self-shading capacity of dense building clusters. In Open High-rise blocks, the importance of SVF surpasses all other factors, indicating that the control of sky openness on heat dissipation is paramount in such environments.

4.2.4. Impact of Urban Functional Zones on the Thermal Environment

Building upon the preceding analysis of the relationships between LST and its drivers across different morphologies, this study further investigates the LST characteristics of various functional zones within unified morphological types, as shown in Figure 11. The analysis reveals that the magnitude of temperature differences among functional zones varies significantly depending on the overarching urban morphology, as detailed in

Table 3. LST statistics by functional zone.

Functional Zone	Transportation	Industrial	Residential	Commercial	Public Service
Count	21	99	1938	461	1355
Average Temperature (K)	296.93	296.60	296.14	296.13	296.09

. Statistical results show that industrial and transportation zones have significantly higher LSTs than other functional areas. The average LST for industrial zones is 296.60 K and for transportation zones is 296.93 K. In contrast, residential, commercial, and public service zones exhibit comparable average temperatures, all falling within the narrow range of 296.09 K to 296.14 K. This phenomenon can be attributed to the substantial anthropogenic heat generated by industrial production activities, which leads to rapid daytime warming. Transportation zones similarly exhibit elevated temperatures due to a combination of factors, including the large scale of their associated buildings, high-density human activity, and concentrated heat emissions from air conditioning systems.

Further analysis reveals that urban morphology has a significant moderating effect on the temperature differences among functional zones, as illustrated in Figure 10. Compact and dense urban forms tend to exacerbate and amplify these thermal disparities, whereas sparse and open morphologies effectively mitigate and homogenize the thermal performance across different functional zones. For instance, within high-density morphologies such as Compact Low-rise and Compact Mid-rise, LST differential between industrial zones and residential or public service zones can approach 2 K. This indicates that when the background urban form is dense, factors intrinsic to the functional zone itself, such as anthropogenic heat sources, building materials, and activity intensity, become the key determinants of the local microclimate. In contrast, within more open morphologies like Sparse High-rise and Open High-rise, the temperature differences among functional zones are significantly reduced. The LST for all functional areas in these types converges within a narrow range of 295.5 K to 295.9 K. This suggests that in open areas, the influence of the overarching urban morphology in shaping the thermal environment can surpass the thermal effects generated by the functional land use, thereby neutralizing or masking the thermal signatures of the functional zones.

Based on this analysis, functional zoning can be considered a proxy variable for latent anthropogenic heat sources and the intensity of socioeconomic activities. While functional zoning is not a direct physical driver, its influence on LST is statistically significant, with this effect being particularly pronounced in high-density built-up areas.

5. Discussion

5.1. Comparative Analysis with LCZ Classification

A comparative analysis was conducted between the adaptive classification method employed in this study and the established LCZ system to investigate the discrepancy between universal standards and local characteristics within a highly heterogeneous urban environment. Following the LCZ classification standards defined by Stewart and Oke [24], the number of blocks corresponding to each of the built types (LCZ 1–9) was calculated. This result was then juxtaposed with the proportions of the nine morphology types derived from our K-Means clustering for a comparative analysis (

Table 4. Comparison of Block Proportions between K-Means Classification and the LCZ System.

K-Means Classification (This Study)	Block Proportion (%)	Corresponding LCZ Category	LCZ Proportion (%)
Compact High-rise	20.83%	LCZ 1 (Compact high-rise)	8.30%
Compact Mid-rise	31.26%	LCZ 2 (Compact mid-rise)	34.37%
Compact Low-rise	7.48%	LCZ 3 (Compact low-rise)	18.68%
Sparse/Open High-rise	11.35%	LCZ 4 (Open high-rise)	5.52%
Sparse/Open Mid-rise	23.35%	LCZ 5 (Open mid-rise)	2.68%
Sparse/Open Low-rise	5.71%	LCZ 6, 8, 9 (Sparse/open low-rise types)	16.11%
		LCZ 7 (Lightweight low-rise)	6.72%
		LCZ A-G (Natural land cover types)	7.62%

).

The comparative analysis with the LCZ framework first validates the robustness of our classification results. our “Compact High-rise,” “Compact Mid-rise,” and “Compact Low-rise” categories correspond broadly to LCZ 1, LCZ 2, and LCZ 3. Notably, both methods identify a “compact mid-rise” morphology as the dominant type in Beijing; our Compact Mid-rise accounts for 31.26% of blocks, while LCZ 2 accounts for 34.37%, indicating a high degree of consensus. More importantly, the comparison highlights the advantages of our adaptive approach. The standard LCZ classification identifies a significant proportion of sparse/open low-rise structures (LCZ 6, 8, and 9 collectively at 16.11%) and a remarkably high share of compact low-rise (LCZ 3 at 18.68%). In contrast, our method identifies a greater prevalence of Compact High-rise (20.83%) and Sparse Mid-rise (23.35%). This latter characterization more accurately reflects the on-the-ground reality of Beijing. As China’s capital and a megacity, its central areas have undergone long-term, high-intensity urbanization. This process has resulted in a dense built environment with high building density, a widespread agglomeration of mid- to high-rise buildings—typically 6 to 15 stories, especially in residential and commercial complexes built since the 1990s between the 2nd and 5th Ring Roads—and highly intensive land use [22].

The LCZ system utilizes a fixed, globally standardized set of parameters to classify urban forms, ensuring cross-city comparability by requiring that metrics like building surface fraction and height-to-width ratio fall within predefined intervals [27,28]. However, this universal approach may not fully capture the internal heterogeneity of a specific city. Our adaptive clustering method, by contrast, is data-driven and objectively delineates local morphologies based on their intrinsic statistical distributions. This approach not only provides a strong cross-validation for the dominant urban form identified by the LCZ system but also demonstrates significant efficacy and originality in adaptively identifying locally dominant morphologies, discovering key transitional zones, and enhancing the analytical depth of thermal environment drivers. It therefore serves as a powerful supplement and refinement to existing urban morphology analysis methods.

5.2. Duality and Non-Linearity of Urban Morphology’s Impact on LST

This study reveals that urban morphology exerts a dual and non-linear influence on the thermal environment. A key finding is the coexistence of a warming “enclosure effect,” where dense building arrangements reduce the SVF and impede heat dissipation, and a cooling “shading effect,” where building mass provides shade and leads to localized temperature reduction. This discovery challenges the conventional assumption of a simple, positive linear relationship between building density and LST. This result is consistent with the findings of Ding et al. [79], who also identified the non-linear impact of 3D urban factors on LST. The SHAP analysis further elucidates this dual mechanism at the sample level, revealing that in high-density urban areas, an increase in BH can, counterintuitively, produce a cooling effect. This mechanism aligns with the thermal principles of urban canyons [80], indicating that a rational 3D spatial layout can effectively mitigate the UHI effect by enhancing shading and promoting ventilation.

The analysis of land cover factors in this study also challenges traditional approaches to thermal environment governance. While NDVI is widely recognized as a key cooling factor, our results indicate its role is prominent only in Open Low-rise blocks, with its importance diminishing significantly in high-density environments. This finding contrasts with the established understanding of a universal cooling effect from vegetation [8,81] but is consistent with studies that highlight the complexity of the LST-NDVI relationship. This suggests that in land-scarce urban cores, merely increasing green space may yield minimal returns; strategies must be integrated with the optimization of 3D spatial structures to enhance heat dissipation efficiency. Furthermore, the Spearman correlation analysis identified a strong negative correlation between RD and LST within the Open Low-rise block type. This result contradicts the general consensus that roads, as impervious surfaces, typically contribute to warming. A plausible explanation is that roads in these specific areas are often accompanied by street trees with high canopy coverage, which provides a significant shading effect. This finding is in agreement with ZHANG, who also noted that road density can have a relatively more significant negative impact on LST [82].

5.3. Interaction Between Urban Morphology and Functional Zones

Based on the preliminary analysis of the relationship between LST and its drivers across different morphologies, this study further investigates the LST characteristics of various functional zones within unified morphological types. The analysis reveals that urban morphology significantly moderates the thermal impact of different functional zones. In Compact block types, the temperature differences among various functional zones are amplified [83]. This suggests that within high-density environments, the unique anthropogenic heat signatures and activity intensities associated with each function become the decisive factors in shaping the local microclimate. The combustion of fuels for transport, industrial processing, and space heating/cooling contributes to this effect. Conversely, in Open block types, these inter-zonal temperature differences are substantially reduced and tend to homogenize. This indicates that in less dense settings, the overarching thermal-shaping effect of the urban morphology itself can surpass and mask the more subtle thermal signatures of the underlying functional land use [84].

5.4. Implications for Urban Planning and Management

Based on the findings of this study, more targeted strategies for urban thermal environment optimization can be proposed. For Compact Mid-rise and High-rise built-up areas, planning and design should prioritize the optimization of building spatial layouts to enhance SVF, thereby improving conditions for heat dissipation. In medium-density zones, the cooling ‘shading effect’ of buildings should be fully leveraged, achieving cooling objectives through the rational configuration of BH. For relatively open, low-density areas such as Open Low-rise zones, expanding and optimizing the configuration of green infrastructure remains the most efficient cooling measure. Furthermore, this research reveals that compact urban forms amplify the anthropogenic heat effect from functional zones such as industrial areas. Consequently, in the context of industrial park planning or urban renewal, high-heat-emission industries should avoid Compact Low-rise and Compact Mid-rise layouts. Instead, Sparse-type or Open-type forms are more suitable. It should also be mandatory to establish sufficiently wide green or open spaces as thermal buffer zones between these industrial sites and residential districts.

5.5. Limitations of the Study

The findings of this study have significant implications for urban climate adaptation planning and refined governance. They underscore the necessity of shifting from conventional ‘one-size-fits-all’ mitigation strategies to precision-based policies tailored to the specific characteristics of the built environment. While this study makes important contributions to methodological innovation and the elucidation of underlying mechanisms, several limitations should be acknowledged. First, the analysis is based on annual average data and thus does not fully account for the influence of monthly, seasonal, or inter-annual variations on the driving mechanisms. Second, although LST serves as a robust proxy for the thermal environment, it can differ from actual human thermal comfort. Moreover, this study did not directly quantify anthropogenic heat emissions from sources such as traffic and building energy consumption. Future research should aim to incorporate anthropogenic heat source data into the models to provide a more comprehensive explanation of UHI phenomenon. Furthermore, this study is geographically focused on the megacity of Beijing; the generalizability of its findings to cities in different climate zones and of varying sizes remains to be validated. Therefore, comparative studies across multiple cities are recommended to explore both the universal principles and the regional specificities of urban thermal dynamics. Such efforts would contribute to the development of a more universally applicable theory of urban climate.

6. Conclusions

This study establishes an analytical framework for the urban thermal environment, leveraging multi-source data and explainable machine learning. Using the area within Beijing’s Fifth Ring Road as a case study, we employed irregular urban blocks as the primary analytical units. A K-Means clustering method was utilized to classify 3875 blocks into nine distinct urban morphology types: Compact Low-rise, Sparse Low-rise, Open Low-rise, Compact Mid-rise, Sparse Mid-rise, Open Mid-rise, Compact High-rise, Sparse High-rise, and Open High-rise. Subsequently, we conducted an in-depth investigation into the driving mechanisms of LST across these different morphology types. The key findings are as follows:

(1)

A significant quantitative relationship between the urban thermal environment and the urban morphology classification was identified. Compact Mid-rise blocks exhibited the highest annual mean LST at 296.59 K, with a substantial difference of 11.29 K observed between the hottest and coolest block types. The XGBoost model demonstrated robust fitting performance across all nine morphology types, with R2 ranging from 0.62 to 0.93. Through SHAP analysis, the research revealed the significant context-dependency of driver importance. NDBI was the most critical warming factor across all block types. SVF played a key cooling role in high-rise areas by facilitating radiative cooling, whereas NDVI was the dominant cooling factor only in Open Low-rise blocks. This discovery challenges the conventional notion of vegetation’s universally dominant role in urban thermal mitigation, indicating that the influence of driving factors is strongly modulated by block-scale morphological features.

(2)

The research identifies a dual mechanism through which urban morphology impacts LST: the coexistence of a “trapping” warming effect and a “shading” cooling effect. In high-density areas, an increase in BH and density reduces SVF, which traps longwave radiation and results in a warming effect. Simultaneously, however, the enhanced shading from taller and denser buildings can provide a significant cooling benefit by blocking direct solar radiation. This discovery transcends the conventional understanding of a simple linear relationship between building density and LST, offering a novel theoretical perspective on the complexity of the urban thermal environment.

(3)

The influence of functional zoning on LST exhibits a pronounced morphological modulation effect. Within Compact-type blocks, the high-temperature effects of industrial and transportation zones—often linked to anthropogenic heat emissions —are significantly amplified, with LST differences between functional areas reaching up to 2 K. In contrast, within Open-type blocks, temperatures across different functional zones tend to homogenize, with differentials narrowing to less than 0.4 K. This indicates that under certain conditions, the thermal environment-shaping role of urban morphology itself can override the influence of land use function.