Quantifying the Influence of Different Block Types on the Urban Heat Risk in High-Density Cities

Abstract

Urbanization and climate change have led to rising urban temperatures, increasing heat-related health risks. Assessing urban heat risk is crucial for understanding and mitigating these risks. Many studies often overlook the impact of block types on heat risk, which limits the development of mitigation strategies during urban planning. This study aims to investigate the influence of various spatial factors on the heat risk at the block scale. Firstly, a GIS approach was used to generate a Local Climate Zones (LCZ) map, which represents different block types. Secondly, a heat risk assessment model was developed using hazard, exposure, and vulnerability indicators. Thirdly, the risk model was demonstrated in Guangzhou, a high-density city in China, to investigate the distribution of heat risk among different block types. An XGBoost model was used to analyze the impact of various urban spatial factors on heat risk. Results revealed significant variations in heat risk susceptibility among different block types. Specifically, 33.9% of LCZ 1–4 areas were classified as being at a high-risk level, while only 23.8% of LCZ 6–9 areas fell into this level. In addition, the pervious surface fraction (PSF) had the strongest influence on heat risk level, followed by the height of roughness elements (HRE), building surface fraction (BSF), and sky view factor (SVF). SVF and PSF had a negative impact on heat risk, while HRE and BSF had a positive effect. The heat risk assessment model provides valuable insights into the spatial characteristics of heat risk influenced by different urban morphologies. This study will assist in formulating reasonable risk mitigation measures at the planning level in the future.

1. Introduction

1.1. Background

Climate change and urbanization have the potential to exacerbate the urban heat island (UHI) and extreme heat events, posing significant heat-related health risks in urban built environments [1]. Heat events can elevate the risk of various health issues, including heart, respiratory, and kidney diseases, as well as increase morbidity and mortality rates [2,3,4]. For example, the mortality rate in China has surged to 110–140% of the 1995–2014 level due to the rising heat risk [5]. It is crucial to assess the heat-related health risks in urban development [6,7]. In the summer of 2022, it was estimated that 9100 deaths in Germany were attributed to heat-related causes [8]. In India, an estimated 1116 people die from heat waves every year [9]. Many studies focus on exploring urban heat risk assessment, aiming to understand the risk pattern within communities and across cities. Urban heat risk assessment is crucial for revealing risk variation patterns and proposing heat risk mitigation strategies.

1.2. Literature Review

The interplay between urbanization, climate change, population growth, and aging can exacerbate health risks associated with high temperatures [1]. In general, heat risk is regarded as the level of health risk posed to humans by extremely high temperatures. This risk is particularly acute among socioeconomically and physically vulnerable groups, including the elderly, disabled individuals, infants, and the impoverished, as well as those who are frequently exposed to outdoor environments, such as outdoor activity participants and workers [10]. Empirical assessments of heat-related health risk have primarily relied on two methods. One approach involves identifying the spatial distribution of the risk by establishing a regression model. For example, a distributed lagged nonlinear model was used to evaluate the heat risk of 51 regions in Seoul [11]. Similarly, a weighted regression model can be used to determine areas of high heat risk in different regions of Singapore, taking into account UHI intensity and the proportion of the elderly population [12]. Another popular method for assessing heat risk is the “Crichton Risk Triangle” framework, which integrates societal, population, economic, and urban morphology aspects of a city. This framework considers the following three key dimensions: hazard, exposure, and vulnerability. Hazard encompasses elements that have the potential to trigger risks, with extremely high temperatures and heatwaves significantly increasing the likelihood of heat-related hazards [13]. Exposure primarily concerns the populations that are subjected to high temperatures, particularly the individuals engaging in outdoor activities [14]. Vulnerability serves as an indicator of an individual’s resilience to high temperatures, which is influenced by diverse factors such as encompassing income, education, and age level [15]. Many studies on heat risk assessments have been conducted at the administrative units. For example, urban heat island intensity, population density, and elderly population ratio were used to assess heat risk of Singapore [16]. Results are generally reported at the administrative unit level, limiting their ability to provide block-scale information on the spatial distribution of risk levels. Similarly, some studies have conducted risk assessments at the administrative level [11,17], but few of them have elucidated the relationship between urban morphology and heat risk at block. Block-scale assessments are crucial as they represent specific urban block planning and are the directly perceived areas by human inhabitants.

Urban morphology significantly influences the microclimate by altering local wind patterns and solar radiation, subsequently affecting temperature distributions and heat risk within various block scales [18,19]. Local Climate Zones (LCZ) is a powerful tool to represent different block types [20], which can be used to quantify the association between urban morphology and the thermal environment [21,22,23]. For example, the heat characteristics and their spatiotemporal patterns can be thoroughly analyzed based on the categorization of a LCZ [24]. Remote sensing data revealed UHI effect intensity across different LCZs, with built-up areas intensifying UHI, while land-cover areas (e.g., LCZ D and LCZ G) mitigate it. This establishes a linkage between the thermal environment and LCZs in urban areas, thereby facilitating the application of an LCZ in heat risk assessments [25]. Similarly, an LCZ map has been used to quantify heat risk level in different LCZ types. Results indicated that at least 60% of LCZs 1-5 were designated as high-risk areas, while LCZ 6 was deemed to be more suitable for implementing measures to mitigate heat hazards [26]. Additionally, an LCZ map can be integrated with urban population mortality rates to assess heat-related health risk. The findings revealed that densely built-up areas exhibit higher risk levels compared to land-cover areas [27]. In summary, the LCZ, which aggregates various urban morphology indicators, can facilitate a comprehensive understanding of the relationship between different block types and heat risk.

Revealing the correlation between urban morphology factors and heat risks is essential for developing effective mitigation strategies to reduce these risks. The complex spatial morphology significantly impacts the urban thermal environment, rendering a singular linear indicator insufficient to capture its complex relationship with heat risks. Various machine learning algorithms, such as neural networks [28] and random forests [29], provide invaluable tools for exploring the nonlinear relationships among multiple morphology factors. These algorithms have been widely used to investigate the correlation between urban areas, the urban thermal environment, and the health of residents [30,31]. For example, the XGBoost algorithm was employed to explore the correlation between the morphology factors of European cities and the urban thermal environment. The findings indicate that approximately two-thirds of temperature variations within cities can be explained by urban morphological features [32]. Additionally, the random forest algorithm was used to investigate the influence of urban morphology on Land Surface Temperature (LST). Results showed that high building density positively affects LST, while the floor area ratio exhibited a negative impact [33]. Previous studies have successfully elucidated the relationship between urban structure and its thermal environment, there remains a gap in methodology for interpreting how different urban spatial factors influence heat risk.

1.3. Research Objectives and Structure

To address the gap, this study proposes a heat risk assessment model to investigate heat risk at various block types. Firstly, urban morphology factors were obtained using multi-source data to map LCZs. The proposed model for assessing heat risk was demonstrated in Guangzhou, a high-density city in China. Furthermore, the correlation between the heat risk level and different LCZs was quantified through spatial correlation. Finally, the XGBoost model was applied to interpret the sensitivity of urban morphology on block-level heat risk. The primary novelties of this study can be summarized as follows: (1) proposing a novel heat risk assessment model by considering urban morphology factors; (2) revealing the correlation between urban morphology factors and heat risk, which facilitates the development of mitigation strategies to reduce high risk areas; and (3) quantifying the spatial autocorrelation of different heat risk levels and identifying their specific reasons these variations at a block scale. This study provides support for explaining the relationship between heat risk and urban morphology factors at block scale, thereby facilitating the development of healthy and sustainable building designs.

This paper introduces the study area and data sources in Section 2. Section 3 explains the calculation methods for risk indicators and the heat risk assessment methodology. Section 4 presents the results. Section 5 and Section 6 provide the conclusions and discussion, respectively.

2. Study Area and Data

2.1. Study Area

Guangzhou is located in the core of the Greater Bay Area in China (Figure 1), spanning between 112°57′~114°3′ E longitude and 22°26′~23°56′ N latitude. It lies within a subtropical humid climate zone, characterized by a hot summer and a warm winter, with a monthly average temperature ranging from 14 °C to 28 °C. With a population of approximately 18.81 million inhabitants and an urbanization rate of 86.48% [34], Guangzhou serves as a prototypical high-density city that has undergone rapid urbanization, resulting in the significant expansion of its built-up area. This expansion has given rise to considerable heat-related health hazards. Therefore, examining the heat risk of various block units is paramount to developing effective mitigation strategies and fostering a more sustainable urban environment.

2.2. Data Collection and Pre-Processing

2.2.1. Dataset for Urban Morphology Factors

This study integrated various data sources for Guangzhou, including urban morphology, vegetation coverage, water body coverage, and land cover information (

Table 1. Data acquisition and their application in this study.

Theme	Source	Period	Resolution	Application
Building footprint	https://www.resdc.cn/Default.aspx (accessed on 4 December 2023)	2019	-	LCZ mapping
Water cover	https://www.openstreetmap.org/ (accessed on 4 December 2023)	2020	-	LCZ mapping
Green cover	https://www.openstreetmap.org/ (accessed on 4 December 2023)	2020	-	LCZ mapping
Land use [35]	https://zenodo.org/ (accessed on 5 December 2023)	2022	1 m	LCZ mapping
Landsat-8	https://www.usgs.gov/ (accessed on 5 December 2023)	2015–2020	30 m	Hazard/Exposure calculation
Population density	https://www.worldpop.org/ (accessed on 8 December 2023)	2020	100 m	Hazard/Exposure calculation
Population density (>65)	https://www.worldpop.org/ (accessed on 8 December 2023)	2020	100 m	Hazard/Exposure calculation
Night-time Light	http://59.175.109.173:8888/app/login.html (accessed on 10 December 2023)	2019	130 m	Hazard/Exposure calculation
Anthropogenic heat flux	https://dataverse.harvard.edu/ (accessed on 7 February 2024)	2019	500 m	Hazard calculation
Mobile signaling data	China Unicom mobile phone	July 2022	-	Residents’ activity preference

). First, key urban morphology indicators such as Sky View Factor (SVF), Height of Roughness Elements (HRE), and Building Surface Fraction (BSF) were calculated using the building footprints and heights obtained from the architectural data of Guangzhou City. Then, the Impervious Surface Fraction (ISF) and Pervious Surface Fraction (PSF) were derived using vegetation coverage data and water body coverage data. Lastly, the land cover was classified using the land cover data.

2.2.2. Dataset for Heat Risk Assessment

In this study, some risk indicators were calculated using Landsat-8 data. Specifically, the TIRS10 band (thermal infrared band) of Landsat-8, spanning from July to August between 2015 and 2020, with the atmospheric correction algorithm, provided the summer average LST, which represents the hazard dimension. In addition, the Near-infrared (NIR) and Red-band of Landsat-8 were utilized for calculating the Normalized Difference Vegetation Index (NDVI). Meanwhile, the green band (Green), NIR, and Short-Wave Infrared band one (SWIR1) of Landsat-8 were used to calculate the Enhanced Water Index (EWI). Population density (PD) represents the exposure dimension. Both the density of population over 65 years old (OPD) and Nighttime Light data (NTL) were used to calculate the vulnerability dimension. The selection of these heat risk indicators were based on the previous research [26]. Due to differences in resolution, all data were resampled to achieve a 30 m resolution for calculating risk values.

3. Methodology

3.1. Framework for Heat Risk Assessment

Heat risk assessment is a comprehensive outcome that integrates urban multi-source data preprocessing, LCZs mapping, heat risk mapping, and spatial patterns analysis. Figure 2 presents the workflow of the heat risk assessment. Firstly, urban morphology factors, building data, and land cover data are obtained to generate an LCZ map, which is subsequently mapped using a GIS-based approach. Then, three heat risk assessment indicators were selected, including heat hazard, heat vulnerability, and heat exposure. Heat hazard reflects environmental temperature severity, influenced by the urban environment, which is crucial for heat risk assessments. When daily maximum temperatures exceed 35 °C, it is considered to be high heat, posing potential negative effects on human health. [36]. Heat vulnerability reflects an individual’s physical state and capacity to cope with heat risks, influencing resistance, response, and recovery [37,38,39]. Meanwhile, heat exposure is usually described as the presence of people, environmental functions, infrastructure, and cultural assets in areas and contexts susceptible to adverse impacts [40]. These three dimensions are then multiplied to generate the overall heat risk map. Finally, spatial autocorrelation and the XGBoost-SHAP methods are employed to investigate the relationship between urban spatial morphology and heat risk levels.

3.2. Improved Heat Risk Assessment Model

The proposed GIS-based risk assessment model is based on Crichton’s risk triangle [41], including hazard, vulnerability, and exposure indicators. Risk values in different blocks were calculated using Equation (1). The natural breakpoint method is used to categorize heat risks into seven levels.

$$\mathrm{H}\mathrm{R}\mathrm{K}=\mathrm{h}\mathrm{a}\mathrm{z}\mathrm{a}\mathrm{r}\mathrm{d}\times \mathrm{v}\mathrm{u}\mathrm{l}\mathrm{n}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{b}\mathrm{i}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{y}\times \mathrm{e}\mathrm{x}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}$$

(1)

The entropy weight method is used to determine the weight of each indicator. Range normalization transforms multi-source data into a unified range by using Equations (2) and (3) [42]. Information entropy is defined in (Equations (4) and (5)), and the weight is calculated in (Equation (6)). These weights are used to compile indicators within layers, as shown in (Equation (7)).

$$X=\left[\begin{array}{ccc}{x}_{11}& \cdots & x_{1n}\\ ⋮& \ddots & ⋮\\ x_{m1}& \cdots & x_{mn}\end{array}\right]$$

(2)

$$Y_{ij}=0.1+\frac{X_{ij}-min\left(X_i\right)}{\mathit{max}\left(X_i\right)-min\left(X_i\right)}\times (0.9-0.1)$$

(3)

where $X$ is the original indicator matrix composed of m research units and n indicators; $X_{ij}$ is the original value of the i-th research unit and the j-th indicator; and $Y_{ij}$ are the standardized values.

$$P_{ij}=\frac{Y_{ij}}{{\sum }_{i=1}^n{Y}_{ij}}, i=1,\dots ,n,j=1,\dots ,m$$

(4)

$$E_j=-\ln {\left(n\right)}^{-1}{\sum }_{i=1}^n{P}_{ij}\ln P_{ij}$$

(5)

where $P_{ij}$ is the variation of the indicator size; $E_j$ is information entropy; and $n$ represents $n$ research units.

$$w_i=\frac{1-E_j}{k-\sum E_j}(j=1, 2, \dots , m)$$

(6)

$$S_i={\sum }_{j=1}^m{w}_j\ast x_{ij}$$

(7)

where $w_i$ represents indicator weight and $S_i$ represents the comprehensive indicator value.

3.2.1. Heat Hazard

Land Surface Temperature (LST) was chosen as the primary indicator of heat hazards. In addition, anthropogenic heat, resulting from human activities like industrial labor, transportation, and metabolism raises environmental temperature, contributing to microclimate differences at the block scale [43]. Therefore, anthropogenic heat was considered in the hazard indicators and data from a previous publication [44]. This study calculated the summer average LST values using the radiative transfer method, known for accurate temperature estimation [45,46]. Landsat-8 imagery were available from Google Earth Engine of summer days from July to August between 2015 and 2020.

The two indicators were standardized within the range of 0.1–0.9. The standardized indicators were then multiplied by their corresponding weights to obtain the heat hazard value (Equation (8)).

$$Hazard=LST\times w_L+AHF\times w_A$$

(8)

3.2.2. Heat Vulnerability

The elderly is less resilient to high temperatures and more susceptible to related illnesses. Consequently, population density (>65 years old) (OPD) was chosen as a vulnerability indicator. Furthermore, Night-Time Light (NTL) positively correlates with economic development [47,48], which can indirectly reflect income levels [49,50] and individual resilience to heat risks. Both variables were standardized to a range of 0.1–0.9. The standardized indicators were then multiplied by their respective weights to calculate the vulnerability value (Equation (9)).

$$Vulnerability=OPD\times w_N-NTL\times w_o$$

(9)

3.2.3. Heat Exposure

Population density (PD) data were included as one of indicators of heat exposure, which can be obtained from the WorldPop website. Vegetation coverage, as reflected by the Normalized Difference Vegetation Index (NDVI), provides shade and helps reduce environmental temperatures, while water bodies absorb heat, also mitigating heat risks. These three indicators were standardized to a range of 0.1–0.9. Subsequently, the standardized indicators were multiplied by their corresponding weights to calculate the exposure value (Equation (10)).

$$Exposure=PD\times w_P-NDVI\times w_N-EWI\times w_E$$

(10)

NDVI serves as a vegetation layer indicator (Equation (11)) [51], and EWI clarifies water body boundaries using remote sensing (Equation (12)) [52]. Therefore, PD, NDVI, and EWI were chosen as heat exposure indicators.

$$NDVI=\frac{NIR-R}{NIR+R}$$

(11)

$$EWI=\frac{Green-(NIR+SWIR1)}{Green+(NIR+SWIR1)}$$

(12)

where NIR is the near-infrared band; R is the red band; Green is the green band; and SWIR1 is the short-wave infrared band one.

3.3. Block Types Classification

LCZ types were classified into 17 categories that encapsulate morphology characteristics and microclimate changes, consisting of ten built-up areas (LCZ 1~LCZ 10) and seven land cover areas (LCZ A~LCZ G). Each category possesses distinct surface characteristics that impact the urban microclimate. Previous studies have established a unified framework for evaluating urban zoning based on diverse surface characteristics [53,54]. The GIS-based classification method has proven to be effective in urban climate zone classification, accurately reflecting the 3D morphology of different block types, thereby justifying its application in generating the LCZ map [55]. The key processes for LCZ mapping are as follows: (1) determining the LCZ grid resolution; (2) assessing urban spatial morphology; and (3) utilizing fuzzy classification and majority voting methods to categorize LCZs.

3.3.1. Grid Resolution

Establishing an appropriate boundary scale significantly enhances LCZ classification accuracy. Although urban morphology can be segmented, the thermal climate remains stable within a specific area, influenced by factors such as surface roughness, architectural geometric properties, and weather conditions. In general, an LCZ’s diameter ranges from 400 to 1000 m grid resolution [56]. Based on previous research, a 240 m × 240 m grid resolution has been successfully applied for LCZ classification in Guangzhou [57,58]. Consequently, this study adopts this resolution (240 m × 240 m).

3.3.2. Assessing Urban Morphology Factors

The key parameters that define urban canyon geometry include H (mean building height on both street canyon sides), W (horizontal canyon extent), and L (canyon length). These parameters are used to calculate BSF [33], HRE, and SVF. In addition, land cover interacts with the atmosphere, altering thermal conditions within the canyon. Surface characteristics play a crucial role in influencing latent heat flux, which in turn affects temperature dynamics [59]. Land cover can be categorized into impermeable surfaces (e.g., asphalt and concrete) and permeable surfaces (e.g., soil, water, and vegetation). These can be quantitatively represented by the fraction of impermeable surface (ISF) and the fraction of permeable surface (PSF), as shown in

Table 2. Calculation method for urban morphology factors.

Property	Methods	Formulas	Description
SVF	SAGA GIS	$SVF=\frac{S_{sky}}{S_{total}}$ [60]	where Ssky indicates the visible sky area in the model space, m2; and Stotal indicates the total sky in the model space, m2.
BSF	Building footprints, ArcGIS pro	$BSF=\frac{S_b}{S_{total}}$ [60]	where Sb indicates the total building footprint area, m2; and Stotal indicates the total block area, m2.
HRE	Building height, ArcGIS	$HRE=\frac{{\sum }_i^n{S}_i\ast H_i}{S_{total}}$ [60]	where Si indicates the building footprint area, m2; Hi indicates the building height, m; and n indicates the count of typical buildings within a block.
PSF	Green cover, water cover, ArcGIS pro	$PSF=\frac{S_p}{S_{total}}$ [60]	where Sp indicates the total pervious area, m2; and Stotal indicates the total block area, m2.
ISF	ArcGIS pro	$ISF=1-\left(BSF+PSF\right)$ [61]	where Si indicates the total impervious area, m2; and Stotal indicates the total block area, m2.
TRC	Davenport classification of terrain roughness [62]	$Z_0=f_0\overline{Z_H}$	where Z0 represents the surface roughness length; $f_0$ represents the empirical coefficient; and $\overline{Z_H}$ represents the height of the surface elements [63].

.

3.3.3. LCZ Mapping and Validation

Remote sensing- and GIS-based methods are widely employed for LCZ classification, leveraging urban spatial data [64]. Furthermore, GIS-based methods offer valuable function to further investigate the relationship between heat risk and urban morphology factors [65]. This study adopts a GIS-based method to classify different LCZ types, distinguishing between built-up areas and land cover areas based on building footprints [56]. Specifically, land use data were utilized to identify land cover areas, while urban morphology data were used to map built-up areas. In addition, the fuzzy classification and majority voting methods were employed to classify built-up areas, as demonstrated in Figure 3. Figure 3a shows an example of a spatial morphology element, showcasing the application of a trapezoidal linear function to determine the fuzzy membership of each LCZ type. Figure 3b demonstrates a map of the spatial morphology element, while Figure 3c shows each neighborhood block has a dominant LCZ type. Moreover, this study utilized an area-based assessment method to determine the proportion of classified categories within the total area. This information is then leveraged to construct a confusion matrix, enabling the calculation of the overall accuracy for the LCZ classification.

3.4. Spatial Correlation Analysis

3.4.1. Pearson’s Correlation

Pearson’s correlation is used to quantify the linear association between urban morphology and heat risk factors [66]. Pearson’s correlation coefficient varies from −1 to 1 (Equation (13)), and r = 1 indicates a perfect positive correlation. r = −1 indicates a perfect negative correlation.

$$r=\frac{{\sum }_{i=1}^n\left(X_i-\overline{X}\right)(Y_i-\overline{Y})}{\sqrt{{\sum }_{i-1}^n{(X_i-\overline{X})}^2\sqrt{{\sum }_{i=1}^n{(Y_i-\overline{Y})}^2}}}$$

(13)

where $X_i$ and $Y_i$ represent the i-th observation value of the two variables, respectively; and $\overline{X}$ and $\overline{Y}$ represent the mean of the two variables, respectively.

3.4.2. Spatial Autocorrelation Method

Spatial autocorrelation is the similarity between adjacent data values resulting from spatial interaction and diffusion. This dependence weakens or disappears when the distance between the data values increases. Moran’s I is an indicator used to measure spatial autocorrelation by comparing variable values in neighboring regions [67]. In this study, the Moran’s I-based univariate local indicator of spatial autocorrelation (LISA) detects spatial clustering of heat risk, as shown in Equation (14):

$$I=\frac{N{\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}(x_i-\overline{x})(x_j-\overline{x})}{\left({\sum }_{i=1}^n{\sum }_{j=1}^n{w}_{ij}\right){\sum }_{i=1}^n{(x_i-\overline{x})}^2}$$

(14)

where $N$ represents the number of observations (points or polygons); $w_i$ and $w_j$ represent the variable values at the locations i and j, respectively; $w_{ij}$ represents the weight indicating the relationship between location i and location j; and $\overline{x}$ represents the mean of all observation values. Moran’s I value ranges from +1 to −1. An amount of +1 indicates strong positive spatial autocorrelation, 0 indicates perfect randomness, and −1 suggests dispersion.

Bivariate Moran’s I detects the spatial autocorrelation between urban morphology and heat risk, interpreting their spatial interrelationship. It consists of the following two patterns: clustering (high–high, low–low), and dispersion (high–low, low–high). The calculation method is shown in Equation (15) [68].

$$I=\frac{x_i-\overline{x}}{{\sum }_i{(x_i-\overline{x})}^2}{\sum }_j{w}_{ij}(x_j-\overline{x})$$

(15)

where $x_i$ and $x_j$ are the value of the attributes $x$ at location i and j; $\overline{x}$ is the average value of the census tract; and $w_{ij}$ is the spatial weight matrix.

3.5. Interpretable Machine Learning Model

The XGBoost machine learning model, renowned for its resistance to nonlinearity, inherent feature selection, and interpretability, was chosen for this study to analyze urban morphology indicators and heat risk levels [69]. In this study, the XGBoost model was employed, with indicators of urban morphology serving as the independent variables and urban heat risk values serving as the dependent variables. The XGBoost model was configured with a learning rate of 0.1 and a maximum tree depth of 3. To train the XGBoost model, 70% of the data were used as the training set, and 30% as the test set. The model’s performance was assessed through metrics including coefficient of determination (R²), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE).

The interpretability of the “black box model” is crucial for this study. Shapley Additive Explanations (SHAP), based on game theory, provide post hoc interpretation, elucidating the outputs of any machine learning model. The core principle of SHAP is to compute the marginal contribution of features to the model’s output, allowing for the interpretation of the “black box model” on both global and local levels. Previous research has demonstrated that SHAP’s interpretation of the XGBoost model yields spatial effect results comparable to those of the Spatial Lag Model and Multi-scale Geographically Weighted Regression [70]. Therefore, this study employs SHAP for interpretation. For each predicted sample, the SHAP model generates a “Shapley value”, which represents the sum of the values assigned to each feature.

4. Results

4.1. Classification Results of LCZ Types

In this study, the overall accuracy of LCZ mapping using the GIS-based method was 85.5%, which is fulfill the high-precision requirement for assessing heat risk [26]. The spatial pattern of LCZ results is shown in Figure 4a. Land cover areas accounted for 84.4% of all LCZs. LCZ A-B and LCZ C-D were the most common land covers, accounting for 66.6% and 20.5%, respectively (Figure 4b). They were mainly concentrated in northern hilly areas. Built-up areas accounted for 15.6% of all LCZs. LCZ 5 and LCZ 2 were the dominant type, accounting for 24.7% and 20.8%, respectively. They were concentrated in the central areas of Guangzhou city. The rest of the LCZ types and their proportion are ranked as follows (Figure 4b): LCZ 10, LCZ 9, LCZ 4, LCZ 6, LCZ 8, LCZ 1, and LCZ 3.

4.2. Analysis of Heat Risk Distribution under Different Block Type

4.2.1. Spatial Distribution of Hazard, Vulnerability, and Exposure Indicators

Three risk indicators were categorized into the following seven levels using the natural breaks method in ArcGIS Pro: very low, lower, low, medium, high, higher, and very high. The spatial distribution of the three indicators is shown in Figure 5. Figure 5a shows that hazard values range from 0.13 to 0.79. High-risk areas (i.e., high and higher levels) were mostly clustered in the west and scattered in the east. These high-risk areas share features like impermeable and exposed surfaces, receiving abundant solar radiation, resulting in rapid surface warming. Figure 5b shows that vulnerability values range from 0.10 to 0.90. High-risk areas appeared in the area around Yuexiu district, the historic center in Guangzhou. This area boasts a dense elderly population that is particularly susceptible to the adverse effects of high temperatures. Other districts exhibit clusters of medium- to low-risk areas, while the northern hills and southern riverine regions are marked by low-risk areas. Moreover, Figure 5c indicates that exposure values span from 0.10 to 0.75. High-risk levels are also found in Yuexiu, since it is also composed of numerous built-up areas with an insufficient number of shaded green spaces. Similarly, the Pearl River area exhibited similar risks. The rest of the area shows a mix of low-risk and lower-risk levels.

4.2.2. Spatial Distribution of Heat Risk Levels

The distribution of each LCZ corresponding to the heat risk level is shown in Figure 6. It indicates that heat risk levels were higher in the city center compared to the suburbs. Very high–high-risk areas and very low–low-risk areas were primarily distributed in the central and northern areas. Very high-risk areas were concentrated in central of city (Figure 6a), which can be attributed to the dense population and a significant elderly population residing in this area. Compared to the heat hazard in Figure 5a, the risk may be significantly reduced if social factors are incorporated using the proposed assessment model. The traditional heat risk assessment method may overestimate high heat risk since it only considers natural factors. However, the residents in these areas had a strong resilience to high temperatures when considering social factors.

In addition, built-up areas posed a higher heat risk than land covers, as shown in Figure 6b. For example, LCZ 4 had the most very high- and high-risk areas, accounting for 15.48% and 21.07%, respectively. Conversely, low-risk areas were the most in LCZ A and B, accounting for 93.81%, followed by LCZ F at 57.98%. This is evident in the higher heat risk values for built-up types in comparison to those for land covers, as shown in Figure 7. LCZ 4 had the widest risk distribution, with mean heat risk values of 0.21, followed by LCZ 1 and LCZ 8. In general, LCZs 1–4, which have high building and population density, are more likely to be affected by heat risks.

4.2.3. Spatial Autocorrelation between Different Heat Risk Levels

The local indicator of spatial autocorrelation (LISA) was used to understand the spatial clustering of heat risk in Guangzhou. It explains the risk distribution in different built-up areas (Figure 8). Moran’s I was 0.933, the z-score was 667.934, and the p-value was less than 0.01, indicating a significant positive correlation in the clustering of heat risk. High-risk areas (high–high clusters) were concentrated in the central area of Guangzhou. Low-risk areas (low–low clusters) were concentrated in the northern mountainous areas and southern farmland areas. Compared to the suburbs, urban built-up areas were more likely to exhibit heat risks related to high temperatures. This is true especially for the built-up areas in LCZs 1–5, which produce a clustering effect of heat risk. Parks and adjacent urban built-up areas have a variety of clusters, such as the low–high clusters in the city center. This proves that in these areas, the risk value is lower due to the abundant vegetation coverage.

4.2.4. Relationship between Heat Risk and People’s Activity Preferences

This study used the signaling data of China Unicom mobile phones in July 2022 to explore people’s activity preferences in the following three spaces: parks, shopping malls, and pedestrian streets, as shown in Figure 9. Figure 10a indicates that the hourly number of visitors in the pedestrian street (LCZs 2) are far greater than that in the park (LCZs A and B) and shopping mall (LCZs 4), with a higher heat risk level. Compared to park and shopping center visitors, the frequency of pedestrian street visitors shows greater variability and periodicity. The number of visitors to pedestrian streets decreases on weekdays and significantly increases on weekends, while parks and shopping centers are almost unaffected. Similarly, the average hourly number of elderly visitors in the street was greater than that in the park and the shopping mall (Figure 10b). Overall, open activity spaces classified as “built types”, such as pedestrian streets, attract larger crowds, including the elderly. Secondly, parks and green spaces categorized as “land cover types” provide shaded areas for cooling and recreation. In addition, a male-to-female ratio was above 1 (Figure 10c). Men favored parks, preferring low-risk areas like LCZ A and B. Conversely, women preferred shopping malls and streets despite there being a high heat risk. To safeguard women’s health, measures addressing their severe heat risk are needed. Future considerations should include strategies to mitigate heat risk in outdoor spaces, particularly in pedestrian streets [71].

4.3. Relationship between Urban Morphology and Heat Risk

4.3.1. Sensitivity Analysis of Different Urban Morphology Factors

The sensitivity of urban morphology factors to urban heat risk indicators was quantified using Pearson’s correlation heatmap, as shown in Figure 11. The map can also be used to analyze the appropriateness of exploring urban heat risk through the analysis of urban morphology. The results show that the p-value between all indicators was <0.001, and the confidence interval was 95%, indicating that urban morphology and heat risk factors are statistically significant. In terms of heat risk indicators, LST, AHF, OPD, NTL, PD, and EWI exhibited consistency. There was a negative correlation between SVF and PSF and a positive correlation between BSF, HRE, ISF, and TRC, which is consistent with the heat risk spatial clustering in Section 4.3.2. In contrast, NDVI exhibited an opposite situation. The relationship between LST and urban morphology was the highest overall, with an absolute average Pearson’s relationship coefficient of 0.64. The highest coefficient appeared between LST and the urban morphology indicator PSF, reaching −0.79. The relationship between EWI and urban morphology was the lowest overall, with an absolute average Pearson’s correlation coefficient of 0.15. The lowest coefficient appeared between EWI and morphology indicator HRE, with a coefficient of only 0.092. While urban morphology may not directly influence heat exposure and vulnerability, it can exert an indirect impact by influencing heat risk factors. For instance, higher BSF can accommodate a larger population, while higher PSF can mitigate the surrounding temperatures.

4.3.2. Spatial Relationship between Heat Risk and Urban Morphology

The spatial dependence of heat risk and lagged urban morphology is understood through bivariate LISA maps and scatter plots, which explain the spatial relationship between heat risk and urban morphology factors (Figure 12). Moran’s I of HRE (MI: 0.345), BSF (MI: 0.548), ISF (MI: 0.497), and TRC (MI: 0.329) were significantly positive, exerting a positive impact on heat risk. Conversely, SVF (MI: −0.561) and PSF (MI: −0.565) had a negative effect on heat risk, with significantly strong correlation. The corresponding cluster maps indicate the significant local correlation between the heat risk and urban morphology. Specifically, the indicators observed in the central area of Guangzhou, which has high–high clusters and dominant built-up areas, underscore that dense and compact urban morphology with fewer green spaces are more prone to heat risk in the summer. In contrast, the same indicators in the northern hilly and riverine areas, which have low–low clusters and dominant land cover areas, are less susceptible to heat risk. SVF and PSF in Guangzhou have resulted in high–low and low-–high clusters, showing how vegetation coverage and sky openness can have a negative effect on heat risk. For instance, downtown Guangzhou exhibits low risk values due to high sky openness and high vegetation coverage in the north. This can be attributed to the heat absorption capacity of the vegetation. The smaller the SVF, the more the solar radiation can be blocked with shade.

4.3.3. Effect of Urban Morphology Factors on Heat Risk

Urban morphology influences regional temperature, population, and the economy. Examining the influence of urban morphology indicators on heat risk can provide valuable insights for future urban planning and development. The XGBoost-SHAP model helps in understanding the effect of urban morphology on heat risk (Figure 13). This study identified PSF as the predominant factor influencing heat risk, followed by HRE, BSF, SVF, ISF, and TRC. It is speculated that because 84.4% of the study area comprises land covers with low heat risk values, it has increased the effect of PSF in model training. The effects of HRE and BSF were almost consistent, and their correlation with LST was closer. This result can be attributed to higher buildings and dense layouts that block solar radiation and reduce the absorption of heat by the surface.

To further clarify the impact of urban morphology indicators on heat risk, the dependence plot shows the nonlinear relationship between individual urban morphology indicators and heat risk indicators (Figure 14). Overall, SVF and PSF showed a negative correlation, and the dependence between the two was the strongest (Figure 14a,c). The negative impact of SVF became increasingly evident when the positive impact exceeded 0.75. As the SVF increases, the sky is more visible and, thus, the population density is lower, meaning that the heat risk is also lower. Conversely, a higher SVF can lead to a greater gain in solar radiation. The surface temperature then rises, increasing the heat risk. Moreover, PSF had a negative impact on heat risk after it exceeded 0.6. The impact of BSF and HRE on heat risk roughly presented an inverted U-shape (Figure 14b,e), which indicates that the impact on heat risk gradually increased as BSF and HRE increased. However, the impact declined after reaching a certain level. Nevertheless, the overall impact remained positive. It is worth noting that too large or too small a BSF can have a negative impact on heat risk. Similarly, HRE can have a significant negative impact at 15 m because the self-built houses in the suburbs are mostly 15 m tall. In this case, the regional heat risk was high. ISF had a positive impact on heat risk when the SHAP value was higher than 0.7, and vice versa (Figure 14d).

5. Discussion

This study developed a GIS-based risk assessment model to analyze the distribution of heat risk distribution across different block types. The results indicate that LCZs 1–4, characterized by high building and population density, were more prone to high heat risk. Conversely, LCZ 10 and LCZ E exhibited low heat risk levels, despite being in high hazard zones in certain regions, primarily due to their lower population density. It was found that low-risk areas are present in these areas, such as urban parks (LCZ A~LCZ E). In addition, this study observed a tendency for spatial clustering of heat risk, with the potential to transform entire districts into high-risk areas, as shown in Figure 10. Moreover, the low-high clustering of heat risk was observed at the junction of high-value areas dominated by built-up areas, and low-value areas dominated by green spaces. This indicates that green spaces have a significant role in reducing heat risk. Urban planners can effectively reduce heat risk through the cross-construction of green spaces and built-up areas based on the findings of this study.

Urban morphology correlates with surface temperature, the economy, population, etc. (Figure 12), which confirms the appropriateness of exploring the correlation between urban heat risk and urban morphology. This study found that PSF has the greatest influence on heat risk. Higher PSF means being able to absorb more solar radiation, raising LST, while population and economy are also closely related to PSF [72]. The reason may be that the proportion of green spaces in the study area is too large, increasing the influence of PSF on heat risk. Further research needs to be conducted in areas with dominant built-up types. SVF and PSF showed a decreasing impact on urban heat risk. This is because as SVF increases, it blocks solar radiation and reduces surface temperature. PSF has the same effect due to lower thermal conductivity and greater heat capacity of PSF than that of the ISF (Figure 14a,c). HRE and BSF showed an increase and then a decrease, particularly due to population density. The above results indicate that the GIS-based assessment model can explain the relationship between urban spatial morphology and heat risk.

This study has several limitations that could potentially introduce inaccuracies in the heat risk results. Firstly, the GIS-based method for mapping LCZs is limited to areas with a comprehensive database of building information. This approach may not be suitable for smaller cities where data acquisition is challenging or incomplete. Secondly, this study considered only six risk indicators due to the difficulty of obtaining precise block-scale data. However, numerous factors that influence individuals’ adaptability to heat risks, including demographic, social, and economical factors, deserve further consideration in future research. Future studies should prioritize urban built-up areas, characterized by intense human activities and a higher likelihood of heat risks. Thirdly, a broader range of socioeconomic factors should be considered, as they may serve as potential indicators of vulnerability or exposure in future urban heat risk assessments, encompassing aspects such as the accessibility of parks and hospitals, education levels, and outdoor activity patterns of residents. Due to data limitations, this study has provided limited exploration between outdoor heat risk and human activity preferences. Human activity preferences are also influenced by temporal dimensions and differences in indoor versus outdoor heat risk within the same area. Finally, the psychological impact on the perception of heat risk is an essential consideration. Residents’ perception of heat risk may be influenced by mood or cognitive levels, potentially affecting their resilience to heat risk. This could serve as a risk indicator to be included in future research.

6. Conclusions

This study proposes an urban heat risk assessment model to explore the influence of urban morphology and block types on heat risk. Firstly, urban morphology factors, building data, and land cover data were obtained to generate an LCZ map using a GIS-based approach. Then, three heat risk assessment indicators were selected, including heat hazard, heat vulnerability, and heat exposure. These risk indicators are then multiplied to generate an overall heat risk map. The proposed model was demonstrated in Guangzhou, a densely populated city in China. Finally, spatial autocorrelation and the XGBoost-SHAP methods were employed to investigate the relationship between urban spatial morphology and heat risk levels.

Results indicated that heat risk levels in built-up areas surpassed those in land covers, with LCZ 4 exhibiting the highest heat risk, boasting a hazard ratio of 55.23%. Conversely, LCZ 10 and LCZ 5 were identified as low-risk areas, accounting for 90.10% and 73.94%, respectively. In residential and commercial areas, such as LCZ 4 and LCZ 1, it is recommended to adopt mitigation strategies for heat risk. However, in LCZ 9 and LCZ 10, where population density is low, the adoption of such strategies may not be necessary. In addition, different block types not only reflect temperature differences, but also lead to spatial heterogeneity in social activities. For example, LCZ 1 and LCZ 2, which have higher population densities, exhibit higher heat risks. These results can be used for developing block-scale urban planning strategies, optimizing the overall spatial morphology of the city, and reducing the health risks due to high temperatures. Furthermore, when considering urban morphology factors, such as SVF and PSF, they had a negative effect on heat risk, while BSF, HRE, and TRC had a more positive effect. This can be attributed to the effect of solar radiation on the surface temperature and the tendency of people to congregate in areas with diverse building types. By blocking solar radiation, urban spaces absorb less heat, resulting in lower temperatures. Additionally, areas with sparser, lower buildings tend to have lower population densities, which in turn contributes to a more effective reduction in heat risk at the block level. Therefore, potential strategies for mitigating urban heat risk could involve increasing tree planting for shading and evenly distributing the population to create more conducive living environments.

Urban planning shapes the development of diverse block types, characterized by variations in building height, density, and pervious surface fraction across various urban areas. These disparities in spatial factors directly influence the absorption and reflection of solar radiation, contributing to temperature fluctuations within different built environments. Additionally, the functions and population capacities of different blocks influence residents’ living behaviors, resulting in a variability in population density. By considering the combined impacts of both natural and social environments, these factors ultimately result in spatial disparities in heat risk. The proposed heat risk assessment model can assist urban planners in evaluating and implementing mitigating strategies for heat risk at the community and neighborhood level, thereby enhancing the safety of the built environment for future residents.