Heat Risk Assessment in Arid Zones Based on Local Climate Zones: A Case of Urumqi, China

Abstract

Based on the rapid development of urbanization and the increasing severity of extreme heat disasters caused by global warming, it has become increasingly important to enhance the assessment of heat risk. In this study, in response to the urgent need for fine-grained assessment of urban heat risk in arid zones in the context of climate change, an analytical method of dividing Local Climate Zones (LCZs) into street blocks combined with the Hazard–Exposure–Vulnerability–Adaptability (HEVA) heat risk assessment framework is used in Urumqi, a representative city of China’s arid zones. In addition, Shapley Additive Explanations (SHAP) was introduced to quantitatively resolve the driving mechanisms of heat risk in different types of LCZs. The results show that the study area has the largest proportion of bare soil (LCZ F) (37.6%), which is distributed around the built-up types of LCZs, while water (LCZ G) has a very small proportion (0.39%) and only exists in the outskirts of the city. Heat risk was significantly higher in the urban core than in the peri-urban areas, but LCZ F had a very high hazard due to the unique surface characteristics of arid zones, which elevated the heat risk in the peri-urban desertification fringe; SHAP analyses demonstrated that in arid zones, land surface temperature (LST) became a determinant of heat risk for all low-density built-up types of LCZs. This study proposes targeted mitigation strategies for heat risk in arid zones based on the LCZ framework.

1. Introduction

Global urbanization continues to accelerate. By 2050, the proportion of the world’s population living in cities and towns is projected to reach 68%, with most of the growth occurring in developing countries [1]. Urban expansion has triggered the transformation of large-scale natural surfaces into artificial ones, exacerbating the urban heat island effect [2], which has led to an upward trend in the frequency, intensity, and duration of extreme heat events [3], and the deterioration of the heat environment not only threatens the health of the population (e.g., heat stress, increased risk of cardiorespiratory diseases) [4] but also has a serious socio-economic impact by reducing labor productivity and increasing energy consumption [5]. Under this double pressure, heat risk has become the focus of research worldwide [6], and the construction of a refined heat risk assessment system has become a scientific prerequisite for the development of climate resilience planning.

Current research on heat risk has been established in cities in humid climate zones (Köppen–Geiger climate classification as A-, C-, or D-type), involving a wide range of developed and developing countries in Europe [7,8,9,10,11], the Americas [12,13,14,15], and Asia [16,17,18]. The studies in the humid climate zones clearly reflect an iterative relationship, with the post-2020 studies [8,10,11,13,16,17,18] adopting more evaluation metrics, more interpretable and quantitative modeling, and more accurate risk identification than the earlier studies (pre-2020) [7,9,12,14,15]. However, research involving cities in arid climatic zones (Köppen–Geiger climate classification as B-type), which cover 30.2% of the world’s area [19], is limited, the system is not yet robust [20], and coping strategies fail to target specific cities and populations, i.e., there is a lack of granularity in the attention given to particular environmental topics [21]. In addition, such studies have focused on macro and regional scales, making it difficult to capture intra-regional heterogeneity in thermal environments. Heat risk research within the last two years has proposed a neighborhood-scale analysis paradigm, and studies in Cardiff [8], Changzhou [16], and Delhi [18] have begun to explore the correlation between heat risk and the domain of local climate. This is due to the homogeneity of buildings within the same neighborhood, which can be considered as a separate thermal zone [22]. Considering the relationship with the thermal environment on a neighborhood scale is more rational [23] and helps to link the results of risk assessments to fine-grained specific urban planning measures [16].

Local Climate Zones (LCZs) are between the urban and neighborhood scales; they can well describe and classify different urban environments [24], and can work with multi-scale urban planning [25]. One of the mapping methods is usually based on the city’s main roads or rivers, and the segmentation scale is relatively detailed, which can well address the heterogeneity of the urban surface [26], and is considered an ideal framework for studying the relationship between urban morphology and the distribution of various indicators [27]. The system deconstructs urban space into 17 types of climate-responsive units by standardizing built-up environment morphological parameters (building height/density, etc.) and surface cover characteristics (Figure 1), which establishes a physical basis for refined heat risk assessment [28]. Compared with traditional methods for studying the urban thermal environment, LCZ’s fine-grained depiction of heterogeneous structures and indicators within the city is more conducive to standardized descriptions across disciplines [29]. Current LCZ mapping methods mainly include RS-based methods and GIS-based methods, with RS-based mapping methods being the most numerous, accounting for the majority of studies (64.86%) [30]. However, this method has low resolution and weak visualization effect [31], with a verified accuracy of only 50–60% [32]. GIS-based methods can achieve higher accuracy, but need to rely on complete and detailed urban spatial data [33]. Some studies have combined the advantages of the two methods [34,35], firstly using RS-based methods to distinguish land cover types for initial classification, and subsequently combining architectural and geographic information from GIS methods to improve the accuracy of LCZ zoning. Although LCZ studies have covered more than 200 cities around the world, there is still a large gap in areas under extreme climate effects compared to economically developed cities. Therefore, increasing research on these regions will enrich the applicability of LCZs [30].

Furthermore, in order to develop response strategies for heat risk, it is necessary to go beyond the limitations of traditional metrics, such as single urban heat island (UHI) measurements or land surface temperature (LST) analyses. While these metrics are effective in quantifying temperature intensity and spatial patterns, they usually provide only partial insights [36] and do not capture the interaction of a wide range of factors such as land use, vegetation cover, and urban sprawl [37]. Therefore, the choice of a comprehensive assessment system in risk assessment is particularly important [38]. In terms of assessment frameworks, the representative Clayton risk triangle system based on the composition of hazard, exposure, and vulnerability proposed by the IPCC is widely used [39,40,41]. There is also the HVI thermal vulnerability evaluation system composed of exposure, vulnerability, and adaptability [42,43]. On this basis, Chinese scholars Wu et al. proposed the HEVA risk assessment framework based on Hazard–Exposure–Vulnerability–Adaptability to improve the accuracy of heat risk mapping [44] and form a more complete heat risk assessment framework, which has been successively used in heat risk assessment research [45,46,47].

In this study, we take the central urban area of Urumqi in the arid zone of China as an example, draw high-precision LCZ maps based on the method of combining RS and GIS, and construct the HEVA heat risk assessment system to determine the distribution pattern of heat risk of each LCZ type at the neighborhood scale. Moreover, the causes of heat risk are resolved by calculating the contribution of SHAP features. This paper aims to discuss the following:

• What is the distribution relationship between urban heat risk and LCZ types in arid zones?
• The heterogeneity of LCZ zoning in heat risk in arid zones.
• Strategy planning based on LCZ zoning in arid zones.

The subsequent part of the paper is structured as follows: Section 2 describes the steps for the division of the LCZ map and the construction of the HEVA system. Section 3 carries out the assessment of heat risk based on the system established in the previous section. Section 4 demonstrates the spatial differentiation patterns, drivers, and strategies proposed for heat risk in arid regions. Section 5 summarizes the study.

2. Materials and Methods

2.1. Study Area

Urumqi (86°48′ E–88°58′ E, 42°45′ N–45°00′ N) is located in the northern foothills of the Tianshan Mountains in the arid region of northwestern China, in the hinterland of the Asia–Europe continent. It is the capital of Xinjiang Uygur Autonomous Region and the core city of the Silk Road Economic Belt, as well as a typical representative of oasis cities in arid regions [48]. The study area has a typical temperate continental arid climate [49]; its Köppen–Geiger climate classification is Bsk-type, with cold winters and hot summers. The urban water source mainly relies on glacial snowmelt and groundwater [50], with an average annual precipitation of less than 300 mm, evaporation of up to 2000 mm or more, and a significant temperature difference between day and night. The topography is high in the north and low in the south, with an altitude of 680–920 m above sea level and a bare soil coverage rate of 37.6%. The ecological environment shows a pattern of oasis–desert interlaced distribution. The city of Urumqi consists of seven districts, namely Tianshan, Saybagh, Xinshi, Shuimogou, Toutunhe, Dabancheng, and Midong, as well as Urumqi County. Among them, Tianshan District and Saybagh are political, economic, cultural, and financial centers; Xinshi District is a high-tech industrial development zone; and Midong District is an industrial park [51]. As of 2023, the city has a resident population of 4,084,800, with an urbanization rate of 96.56%, the second highest in China (Urumqi Municipal People’s Government, 2024). The ultra-high urbanization rate responds to the concentration of population and resources in a special geographical environment. In addition, heavy industries such as chemical plants and non-ferrous metal smelting are densely distributed in Urumqi [52]. As a result, the urban expansion and industrialization process lead to significant heterogeneity of the thermal properties of the subsurface.

This study takes the main urban area within the Urumqi Bypass as the study area (Figure 2), covering the core administrative units of the Xinshi District, Saybagh District, and Tianshan District. The area is rich in natural and artificial subsurfaces such as high-density built-up areas (LCZs 1–3), industrial zones (LCZ 8 and LCZ 10), and natural bare soil (LCZ F), and has typical urban characteristics of arid zones, which makes it an ideal site for the study of urban heat risk distribution in arid zones.

2.2. Research Methodology

This study can be divided into the following 4 steps (Figure 3). Firstly, a high-precision LCZ map of Urumqi was drawn by combining RS and GIS, and then a heat risk map was drawn based on the HEVA system. Secondly, the two are coupled to form the heat risk map of Urumqi under the LCZ framework, presenting the heat risk distribution pattern at the neighborhood scale and further quantifying the characteristics of heat risk share of different types of LCZs. Finally, SHAP is introduced to analyze the key driving factors of heat risk in different LCZ types and their action mechanisms.

2.2.1. LCZ Mapping

In this study, the LCZ map is drawn using the road network as the basic unit and the scene-based classification method. Firstly, road data were obtained from the OpenStreetMap website to construct the basic road network structure. Sentinel-2 L2A data were obtained from the Google Earth Engine website, and the Sentinel-2 images were segmented according to the road network to obtain irregularly mapped parcels of appropriate size and processed with image normalization to a regular size of 320*320 m to match the image blocks of the training dataset. Subsequently, the ConvNeXt model was used to train the localized samples, and 20% of the samples were selected for testing and evaluating, optimizing, and adjusting the performance of the model. For LCZ types that are not very accurate for prediction, the LCZ classification results were optimized using average height (AH) and normalized difference built-up index data (NDBI).

2.2.2. Heat Risk System Selection

Currently, most of the research systems on heat risk are HEV [7,16,40,53,54,55] and HVI [15,56] (

Table 1. Indicator selection and weighting methodology for the global component study.

Research Cities	Indicator System	Indicator Selection	Empowerment Methodology
Karachi, Pakistan [53]	HEV	Hazard: LST Exposure: population density Vulnerability: elderly population aged 60 years and over, children aged 10 years and under, illiterate population, building density, slum population, population with access to drinking water, bathrooms, electricity, children with a family size of 9 or more, poor population, normalized difference vegetation index (NDVI)	Principal Component Analysis (PCA)
Perth, Australia [54]	HEV	Hazard: LST Exposure: population density Vulnerability: age structure, ethnicity, socio-economic disadvantage, employment, living space, need for assistance, income, availability of internet connection and mobility	Equal Weighting System
Birmingham, UK [7]	HEV	Hazard: high resolution urban heat island maps Exposure: detailed commercial data per household Vulnerability: filtered exposure layer	Equal Weighting System
Changzhou, China [16]	HEV	Hazard: LST, frequency of high temperatures Exposure: population density, frequency of high temperatures Vulnerability: MNDWI, NDVI, night light data, population density over 65 years old	Entropy Weight Method (EWM)
Yangtze River Delta City Cluster, China [55]	HEV	Hazard: surface temperature Exposure: combination of enhanced vegetation index, nighttime lighting Vulnerability: elderly population aged ≥65 years, elderly population aged 60 years or older living alone, illiteracy or semi-illiteracy rate among population aged 15 years or older, total number of beds in health facilities, number of air conditioners per 100 households, GDP per capita	PCA/EWM
Hong Kong, China [40]	HEV	Hazard: hours of heat, hours of heat at night Exposure: population density Vulnerability: population aged 65 and over, uneducated population, households living alone, low-income households, renter population, population living in non-residential buildings	PCA
Santiago, Chile [15]	HVI	Exposure: LST Sensitivity: population over 60 years old, population with young children under 5 years old, population with disabilities, family structure, level of education, unemployed population Adaptability: population with access to communication technology, population with access to water supply, housing materials, medical facilities, road network density, NDVI	PCA
Beijing-Tianjin-Hebei, China [56]	HVI	Hazard: daily surface temperature Exposure: population density Sensitivity: elderly population, child population, female population, nursing homes, kindergartens and primary schools Adaptability: water cover, vegetation cover, night lighting, medical buildings	EWM
Turin, Italy [10]	HEVA	Hazard: urban heat island distribution Exposure: the over-65 population Vulnerability: women over the age of 65, population over 85 years old, low education rate, social isolation, housing overcrowding, ethnic minorities coming from poor countries with high migration pressure, presence of ischemic heart disease, presence of cerebral vasculopathies, presence of heart failure, presence of diabetes, residential buildings in poor conditions, building density, population density, distance from watercourses Adaptability: percentage of green area, average number of floors, possibility of conversion to green roofs, proximity to social welfare facilities, proximity to cool places	Equal Weighting System
Chengdu-Chongqing City Cluster, China [57]	HEVA	Hazard: frequency of heat waves, maximum body temperature of heat waves, maximum duration of heat waves, total duration of heat waves Exposure: habitat index Vulnerability: time to hospital Adaptability: river network density, gross regional product, number of beds in healthcare facilities, general public budget expenditure, residents’ savings balance	Criteria Importance Though Intercriteria Correlation (CRITIC)
Dalian, China [45]	HEVA	Hazard: radiant heat stress intensity Exposure: population density, point of information (POI) density, road network density Vulnerability: child population, elderly population, women’s population, twitter heat Adaptability: year of construction, hot shelter sites, medical sites, GDP	CRITIC

). In this study, we chose the HEVA system, which is a new system extended from the above two traditional heat risk assessment systems. Compared with HEV, which focuses on environmental exposure, and HVI, which focuses on social vulnerability, HEVA quantifies heat risk in a more comprehensive and systematic way by combining the two and using four dimensions of risk indicators: hazard, exposure, vulnerability, and adaptability.

2.2.3. Data Sources and Standardization of Secondary Indicators

In the HEVA system, hazard reflects the degree of danger triggered by its own attributes in a natural disaster [45]; exposure reflects the activities of the population exposed to the risk and the function of the city [58]; vulnerability reflects the sensitivity of the population affected by the disaster to the harm [54]; and adaptability reflects the ability of an individual or a region to cope with extreme heat [59]. In order to accurately identify heat risk drivers and construct a spatial assessment model [60], this study determined an assessment system of 11 secondary indicators under 4 dimensions (

Table 2. Selection, meaning, and sources of indicators.

Primary Indicators	Secondary Indicators	Meaning of Indicators	Direction	Spatialization Methodology	Data Sources
Hazard	LST	Characterizing the intensity of surface thermal radiation	+	remote sensing inversion	LANDSAT8 remote sensing imagery
Exposure	Population Density	Characterizing anthropogenic heat exposure	+	Arcgis Kernel Density Tool	https://lbsyun.baidu.com. accessed on 1 August 2024.
	POI Density	Characterizing the clustering effect of urban functions	+	Arcgis Kernel Density Tool	BIGEMAP
	Road Network Density	Characterizing the level of transport facilities	+	Arcgis Kernel Density Tool	BIGEMAP
Vulnerability	Child Population	Characterizing the distribution of sensitive populations	+	Arcgis Kernel Density Tool	BIGEMAP
	Elderly Population	Characterizing the distribution of sensitive populations	+	Arcgis Kernel Density Tool	BIGEMAP
	Year of Construction	Characterizing building thermal regulation capacity	-	IDW interpolation	https://wulumuqi.anjuke.com. accessed on 5 August 2024.
Adaptability	Housing Value	Characterizing the ability to regulate living conditions	-	IDW interpolation	https://wulumuqi.anjuke.com. accessed on 5 August 2024.
	High-temperature Shelter	Characterizing public thermal adaptability resource allocation	-	Arcgis Kernel Density Tool	BIGEMAP
	Medical Facilities	Characterizing health security capacity	-	Arcgis Kernel Density Tool	BIGEMAP
	GDP	Characterizing the potential for heat adaptability inputs	-	IDW interpolation	https://www.resdc.cn. accessed on 5 August 2024.

), with reference to the mechanism of heat environment in arid cities and the data advantages of economically developed cities. The spatialization of each indicator follows the principle of matching data types and algorithms.

The sources and magnitudes of the indicators are vastly different, and in order to ensure the validity of the assessment results at the neighborhood scale, all the data were first standardized to a resolution of 30 m × 30 m, and then the data were normalized using the method of polar deviation.

For the positive indicators (hazard, exposure, and vulnerability), the formula is as follows:

$${x^{\prime }}_{ij}={\displaystyle \frac{x_{ij}-{\min (x}_j)}{{\max (x}_j)-{\min (x}_j)}}$$

(1)

For the negative indicator (adaptability), the formula is as follows:

$${x^{\prime }}_{ij}={\displaystyle \frac{{\max (x}_j)-x_{ij}}{{\max (x}_j)-{\min (x}_j)}}$$

(2)

where ${x\prime }_{ij}$ is the value of the jth indicator in the ith sample after standardization, $x_{ij}$ is the original value of the jth indicator in the ith sample, and ${\max (x}_j)$ and ${\min (x}_j)$ are the maximum and minimum values of the jth indicator, respectively.

2.2.4. Improved CRITIC Assignment Methodology

In analyzing the importance of heat risk indicators, most of the existing studies use equal weight summation or PCA (Table 1). However, analyzing the characteristics of the indicators reveals that the primary indicators of heat risk have different degrees of influence on the composite risk, and therefore, the weights of the indicators need to be quantified. The improved CRITIC method is a further improvement on the basis of the entropy weight method, the core of which is to consider the variability and conflict between indicators, calculate the comparative strength and relevance of the same indicators under different scenarios, and then calculate the amount of information contained in each indicator [57]. In view of this, this optimization method is selected as the weight determination tool in this study, and the specific implementation process is as follows:

• Calculation of contrast: The standard deviation is used in the CRITIC method as a mathematical representation of the degree of data dispersion within an indicator. According to the principle of statistics, the standard deviation, as an absolute measure of dispersion, is positively correlated with the fluctuation amplitude of the indicator data set: the standard deviation $s_j$ represents the contrast of the jth indicator, and an increase in the value of $s_j$ indicates that the heterogeneity of the distribution of the sample values of the jth indicator has increased significantly, and its information-carrying capacity and evaluation strength have also been increased. Based on the framework of information entropy theory, such indicators with high dispersion need to be given higher weights in the comprehensive evaluation process to accurately reflect their differentiated characteristics. The quantitative formula for indicator contrast can be expressed as follows:

  $$s_j=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{\left(x_{ij}^{\prime }-\overline{x_j^{\prime }}\right)}^2}{n-1}}}$$

  (3)

  where $x_{ij}^{\prime }$ denotes the data normalized to the jth indicator value for the ith sample, and $\overline{x_j^{\prime }}$ denotes the mean of the data normalized to the jth indicator value.
• Calculation of conflictivity: this study defines conflictivity as a quantitative characterization of the degree of information redundancy between indicators. When two indicators show significant positive correlation, i.e., the value of conflictivity $c_j$ is small, it indicates that the information overlap between them is high, and their independence characteristics are subsequently decayed. This phenomenon will lead to highly correlated indicator groups forming information redundancy, weakening the marginal contribution of individual indicators in the comprehensive evaluation. Therefore, it is necessary to moderately reduce the weight proportion of such indicators in the weight allocation process to optimize the signal-to-noise ratio of the evaluation system. The conflicting $c_j$ formula is as follows:

  $$c_j=\sum _{i=1}^n\left(1-r_{ij}\right)$$

  (4)

  where $r_{ij}$ indicates the correlation coefficient between indicator i and indicator j. The larger the conflictivity $c_j$, the higher the indicator independence.
• Calculation of information carrying capacity: synthesize contrast and conflict to calculate the amount of indicator information:

  $$I_j=s_j{c}_j$$

  (5)
  Determine the weight according to the proportion of information; the greater the information carrying capacity, the greater the weight. Set the weight of the first indicator as $w_j$:

  $$w_j={\displaystyle \frac{I_j}{{\sum }_{j=1}^n{I}_j}}$$

  (6)

According to the above calculation process, the final results of the weights of the indicators in this study were obtained (

Table 3. Results of weighting of indicators.

Primary Indicators	Primary Indicators Weights	Secondary Indicators	Secondary Indicators Weights
Hazard	0.367	LST	1.000
Exposure	0.195	Population Density	0.317
		POI Density	0.419
		Road Network Density	0.264
Vulnerability	0.207	Child Population	0.253
		Elderly Population	0.240
		Year of Construction	0.507
Adaptability	0.231	Housing Value	0.332
		High-temperature Shelter	0.206
		Medical Facilities	0.202
		GDP	0.260

):

2.2.5. Establishment of a Heat Risk Assessment System

In this study, the composite index method is used to calculate the composite evaluation index of each primary indicator of heat risk by cumulatively multiplying the indicator values with the weights derived in the previous section. On the basis of the standardized indicator data, the comprehensive evaluation index of primary indicators is generated by weighted superposition layer by layer with the following formula:

$$F_k=\sum _{j=1}^p{w}_j{X}_{kj}$$

(7)

where $F_k$ is the composite assessment value of the primary indicators of the kth spatial unit, $w_j$ denotes the weight of the jth secondary indicator, and $X_{kj}$ is the standardized value of the jth indicator of the kth unit. The model objectively reflects the differential impact of each indicator on heat risk through a linear combination of weights and standardized values.

Considering the interaction mechanism of the four-dimensional indicators of hazard ($H$), exposure ($E$), vulnerability ($V$), and adaptability ($A$), the formula of heat risk index ($HRI$) was constructed [46]:

$$HRI={\displaystyle \frac{H\times E\times V}{A}}$$

(8)

3. Results

3.1. Description of the LCZ Map of Urumqi

In this study, a machine learning method was used to construct a spatial distribution model of LCZs in the study area. According to the Stewart and Oke classification system [28], the study area was classified into 17 LCZ types. Among them, built-up types (LCZs 1–10) accounted for 30.35%, and land-cover types (LCZs A–G) accounted for 69.65%. The built-up areas in the study area showed a central development pattern, while the land-cover types wrapped the built-up areas in a ring shape to form a core–edge distribution structure.

Spatial statistical analysis shows that among the built-up types, LCZ 5 (open mid-rise buildings) accounts for the highest proportion of 20.1% (Figure 4b), and its spatial distribution has obvious spatial clustering with that of LCZ 2 (compact mid-rise buildings), which is mainly located in the urban core area constituted by the northwestern part of Tianshan District, northeastern part of Saybagh District, southern part of Xinshi District, and the western part of Shuimogou District, and is sporadically located in the central part of Midong District. LCZ 1 (compact high-rise buildings) accounts for only 2.2% (Figure 4b), and is distributed in the economic center of Tianshan District, such as the Democracy Road–Hongqi Road business circle. The spatial differentiation of industrial land is also significant: LCZ 8 (large low-rise buildings) and LCZ 10 (lightweight low-rise buildings) form a cluster in the matrix of LCZ F, which is mainly distributed in the southern part of the Toutunhe District and the central part of the Midong District, reflecting the distribution characteristics of the arid zone in which the industrial land is intertwined with bare land.

The land-cover types of LCZs show typical arid zone surface characteristics: LCZ F accounts for 53.9% of the total land-cover types (Figure 4c), which is also the highest among all LCZs, accounting for 37.6% of the total study area (Figure 4a) and forming a semi-annular distribution along the northeast–southeast–southwest of the built-up area; LCZ D (low plants) accounts for 32.9% of the total land-cover types and is mainly located in the agricultural cultivation area in the northwest of the city, and within it, rural settlement types such as LCZ 3 (compact low-rise buildings) and LCZ 6 (open low-rise buildings) are distributed. LCZ G (water) accounted for the lowest share of only 0.56% of the land-cover types (Figure 4c), and 0.39% of the total area of the study area (Figure 4a), which reflected the typical characteristic of the arid zone of water resource scarcity.

3.2. Heat Risk Assessment Map of Urumqi

According to the heat risk map (Figure 5), the heat risk value in Urumqi city ranges from 0.26 to 0.69, with high-risk areas distributed in high-density built-up areas in the southern part of the Xinshi District, the northeastern part of Saybagh District, the western part of Shuimogou District, and the northwestern part of Tianshan District. These areas have high building density and frequent population activities, resulting in high exposure and vulnerability and high heat risk. In addition, Midong District in northeastern Urumqi is the next highest heat risk area, with a large number of factories and heavy industrial areas, industrial heat emissions, and large areas of bare soil making this area relatively high in hazard and exposure. In contrast, areas of low heat risk are concentrated in the northwest of the city. This part of the city has a low density of buildings, low levels of human activity, and extensive vegetated spaces that effectively reduce surface heat radiation, making this area also at a lower level of risk than the rest of the city.

3.3. LCZ-Based Heat Risk Analysis

3.3.1. LCZ-Based Spatial Heterogeneity of Heat Risk

Under the LCZ framework, heat risk in Urumqi showed significant spatial differentiation characteristics (Figure 6a). The statistical results show (Figure 6b,c) that the heat risk of the built-up types LCZs is generally higher than that of the land-cover types LCZs. Among them, the median value of the heat risk index of the high-density built-up types LCZ 1 and LCZ 2 both exceed 0.5, which is the highest level in the city, and their high exposure and vulnerability are the main driving factors. Compact medium-high-rise (LCZ 1 and LCZ 2) has a higher heat risk than open medium-high-rise (LCZ 4 and LCZ 5). The relatively high heat risk of industrial types LCZ 8 and LCZ 10 may be due to the combined effect of high rates of surface hardening, anthropogenic heat emissions, and low vegetation cover in industrial zones. Low-density built-up types LCZ 6 and LCZ 9 have the lowest mean value of heat risk, mainly due to their sparse population with building shading effect. Among the land-cover types, LCZ F has a higher heat risk and accounts for the largest proportion of the land-cover types in LCZs, which indicates that desertification fringe areas in arid cities are more susceptible to the superposition of the risks of bare soil areas and built-up areas, and more attention should be paid to the prevention and control of risks in this area. The median value of the heat risk of LCZ D and LCZ G is lower than 0.4, as their ecological regulation capacity effectively mitigates the heat risk.

It is worth noting that LCZ A (dense trees) is constrained by the climatic characteristics of arid zones, where its presence is mostly due to artificial cultivation and therefore concentrated in built-up areas; therefore, exposure and vulnerability are higher due to factors such as population density. In this case, the cooling potential of LCZ A cannot be characterized.

3.3.2. Spatial Heterogeneity of LCZ-Based Primary Indicators

The distribution of hazard (Figure 7a) shows significant spatial heterogeneity, with a range of values from 0 to 1, and the proportions of each risk level are as follows: very low risk (9.66%), low risk (15.75%), medium risk (25.33%), high risk (28.10%), and very high risk (21.16%) (Figure 7b). Unlike the humid city, the high-risk areas in Urumqi are mainly located in the peripheral LCZ F rather than in the high-density built-up areas, and the proportion of “very high” risk in LCZ F is more than 40%, which is directly related to the intensity of thermal radiation and the wide distribution of bare soil. Among the built-up areas, LCZ 7, LCZ 8, and LCZ 10 have significantly higher heat hazards than LCZs 1–3 (high-density buildings) due to the high rate of hardening of the surface and low vegetation cover in the industrial areas, whereas the high-density residential areas mitigate the buildup of thermal radiation through building shade and localized greening [61]. The “very low” risk area is concentrated in the northwestern part of the city in the LCZ D cropping area, confirming the cooling effect of vegetation transpiration.

The spatial variation of exposure (Figure 7c) is highly coupled with the intensity of human activities. The overall exposure was high at the core and low at the edge, with “very high” and “high” exposures concentrated in the southern part of the Xinshi District and the northeastern part of the Saybagh District, etc. The proportion of “very high” and “high” exposures in LCZ 1 and LCZ 2 exceeds 80% (Figure 7d), and the proportion of high exposures in LCZs 3–5 exceeds 20% due to the combination of commercial and residential functions. It is worth noting that LCZ A has a particularly high percentage of high exposure among land-cover types, which is due to the fact that plantation forests in arid zones are mostly adjacent to built-up areas and have a high degree of spatial overlap with urban functions.

There is a more pronounced spatial correlation between the distribution pattern of vulnerability (Figure 7e) and exposure. The high-risk zones (3.39% of the total) are also concentrated in the LCZ 1 and LCZ 2 area. Sporadic distribution of vulnerability highs occurs in the education concentration zones in the southern part of Tianshan District (including the southern campus of Xinjiang University, primary and secondary schools, and the University for the Elderly, etc.), in the industrial parks in the south-central part of the Toutunhe District, and in the southern part of the Midong District, in the Midong Petrochemical Zone. The education zone is densely populated by sensitive people, and the industrial park and Midong Petrochemical Zone are subject to the enhancement of surface radiation by heat generation from factories and bare soil. Building year, as a negative indicator, shows a vulnerability amplification effect in LCZ 3 (mostly old neighborhoods).

Adaptability (Figure 7g) acts as a negative indicator to counteract heat risk. The “very high” adaptability (2.70%) is concentrated in the core areas of the old city, such as Saybagh District and Tianshan District, where the dense distribution of healthcare and heat evacuation sites is the main driving factor. Correspondingly, the high-GDP Xinshi District, which has been developed for a shorter period of time, has a lower level of healthcare and other infrastructure coverage than the old city, resulting in a southward shift of high adaptability values compared to high exposure and vulnerability values. Similarly, the heavy industrial zone in Midong District (LCZ 8 and LCZ 10) is densely populated but has a severe shortage of public service facilities, making it less resilient. LCZ E (bare rock or paved) is significantly more resilient than the other land-cover types due to the interspersing of functional spaces such as transport hubs and commercial centers.

4. Discussion

4.1. Heat Risk Heterogeneity in Arid Zones Under LCZ Framework

For the built-up types (

Table 4. Built-up types of LCZ with highest and lowest heat risk in different studies.

Research Cities	Köppen–Geiger Climate Classification	Risk Indicators	Highest Heat Risk Built-Up Type LCZ	Lowest Heat Risk Built-Up Type LCZ
Urumqi, China (this study)	BSk	HEVA	LCZ 2	LCZ 9
Karachi, Pakistan [53]	BWh	HEV	LCZ 3	LCZ 8
Kabul, Afghanistan [63]	BSk	LST SUHI	LCZ 8 LCZ 8	LCZ 2 LCZ 35
Phoenix, AZ, USA [20]	BWh	LST	LCZ 8	LCZ 4
Las Vegas, NV, USA [20]	BWh	LST	LCZ 8	LCZ 4
50 cities in the world [64]	/	SUHI	LCZ 2	LCZ 9
Cardiff, UK [8]	Cfb	LST SUHI	LCZ 3 LCZ 8	LCZ 9 LCZ 9
Changzhou, China [16]	Cfa	HEV	LCZ 1	LCZ 9
Harbin, China [65]	Dwb	HEV	LCZ 2	LCZ 9A

), established studies confirm the high risk of LCZ 2 and LCZ 5 [62]; the study on Karachi demonstrates that LCZ 2 and LCZ 3 exhibit the highest heat risk due to high population and building density [53]; the study on Kabul, also an arid city, shows that large areas of impervious surfaces, low-rise buildings, and concrete and steel structures make LCZ 8 the highest LST and Surface Urban Heat Island (SUHI) value zones [63]; Phoenix and Las Vegas had the same findings [20]; the study on 50 cities demonstrated that high-density types (LCZs 1–3) and industrial types (LCZ 8 and LCZ 10) generally have stronger SUHI intensities than open types (LCZs 4–6), with LCZ 9 having the lowest SUHI intensity [64]; the mean heat risk was highest in LCZ 3 and lowest in LCZ 9 for the built-up types in Cardiff [8]; in the case of Changzhou, LCZ 1 had the highest heat risk and LCZ 9 the lowest, and LCZs 1–5 with high-density populations were overall higher than the other built-up types [16]; the study in Harbin showed the highest proportions of LCZ 2 and LCZ 3 in the high risk, again validating that compact LCZs have a greater risk of causing disasters [65].

Regarding the land-cover types of LCZs (

Table 5. Land-cover types of LCZs with highest and lowest heat risk in different studies.

Research Cities	Köppen–Geiger Climate Classification	Risk Indicators	Highest Heat Risk Land-Cover Type LCZs	Lowest Heat Risk Land-Cover Type LCZs
Urumqi, China (this study)	BSk	HEVA	LCZ E	LCZ G
Karachi, Pakistan [53]	BWh	HEV	LCZ E	LCZ G
Kabul, Afghanistan [63]	BSk	LST SUHI	LCZ F LCZ F	LCZ B LCZ B
Phoenix, AZ, USA [20]	BWh	LST	LCZ E	LCZ G
Las Vegas, NV, USA [20]	BWh	LST	LCZ F	LCZ G
50 cities in the world [64]	/	SUHI	LCZ F	LCZ D
Cardiff, UK [8]	Cfb	LST SUHI	LCZ F LCZ F	LCZ A LCZ A
Changzhou, China [16]	Cfa	HEV	LCZ E	LCZ G
Harbin, China [65]	Dwb	HEV	LCZ E	LCZ G

), LCZ E was the highest heat risk type in Karachi, Phoenix, and Changzhou, while the study of Kabul, an arid zone city, claimed that LCZ F behaved as the highest risk zone among the land-cover types on different dates. This is usually due to the fact that bare soil or sand surfaces can easily absorb and re-emit solar radiation [63]; this finding is consistent with the results for Las Vegas, 50 cities, Cardiff, Changzhou, and Harbin. LCZ G was the lowest-risk land-cover type in the vast majority of the studies, with the exception of Kabul (where LCZ B had the lowest risk, followed by LCZ G). LCZ G had the lowest risk during the daytime in the 50 cities study, but LCZ D had the lowest performance in terms of daily risk due to the nighttime temperature rebound effect.

In conclusion, comparing the typical urban cases around the world, there are significant differences in the heat risk of different LCZ types due to the characteristics of built-up areas, population density, and economic activities in different cities. The present study shows that compact built-up types (LCZs 1–3) have higher heat risk than open built-up types (LCZs 4–6), which is in line with the results of studies in humid climate zones [8,66]. The pattern of performance of land-cover types LCZs, on the other hand, differs somewhat from studies in other regions. Although LCZ G exhibits a cooling effect globally [20,67,68], the extent of its influence in arid cities is diminished by the small size of the water bodies, making it difficult to create regional climate regulation. Meanwhile, the widely distributed bare soil zone (LCZ F) at the edge of arid cities has a higher heat risk, and the urban desertification fringe should be more concerned about the built-up heat environment and the dynamic enhancement of the natural climate environment. Overall, the more concentrated heat risk values for all 17 LCZ types in this study compared to the wet zone [8,16,46] suggest that different urban form factors tend to homogenize under the effect of significant climatic conditions in arid zones. This phenomenon has not been fully demonstrated in the current LCZ-based risk assessment.

4.2. Neighborhood-Scale Coping Strategies for Heat Risk in Arid Zones

4.2.1. SHAP-Based Analysis of Indicator Contributions

Based on the previous block-scale assessment of heat risk in arid zones under the LCZ framework, SHAP analysis is applied to reveal the driving mechanisms of heat risk in different LCZ types and to propose differentiated coping strategies. In order to assess the contribution characteristics of the indicator system, it is necessary to first ensure the independence between the variables. In this study, the variance inflation factor (VIF) was used to diagnose multicollinearity among independent variables. VIF is the ratio of the variance in the presence of multicollinearity among explanatory variables to the variance in the absence of multicollinearity [69] and is calculated using the following formula:

$${VIF}_i={\displaystyle \frac{1}{1-{R_i}^2}}$$

(9)

where ${R_i}^2$ is the coefficient of determination when regressing the ith independent variable on all other independent variables. The larger the VIF value, the stronger the covariance of the variable with the other variables (the usual threshold is 10.0, which corresponds to a tolerance of 0.10) [70]. Based on the results of the multiple-covariance test (

Table 6. Results of multicollinearity test.

Model	Unstandardized Coefficients		Standardized Coefficients	t	Sig.	Collinearity Statistics
	B	Std. Error	Beta			Tolerance	VIF
(Constant)	34,531,757.206	151,893.600		227.342	0.000
LST	15,783.415	72.598	0.154	217.409	0.000	0.994	1.006
Population Density	2794.473	89.165	0.038	31.341	0.000	0.335	2.983
Year of Construction	−17,188.481	75.457	−0.187	−227.793	0.000	0.739	1.353
Housing Value	39.129	0.218	0.149	179.252	0.000	0.727	1.375
High Temperature Shelter	−270.473	17.490	−0.026	−15.465	0.000	0.171	5.840
Medical Facilities	8590.258	343.350	0.047	25.019	0.000	0.140	7.138
GDP	−0.184	0.023	−0.007	−8.099	0.000	0.665	1.503

), the final seven indicators with VIF ≤ 10 were filtered to be LST, Population Density, Year of Construction, Housing Value, High-temperature Shelter, Medical Facilities, and GDP.

To explain the contribution of individual metrics to heat risk, this study invokes SHAP as an interpretable model. The SHAP value is derived from the Shapley value in the cooperative game theory, which solves the inconsistency problem of the traditional feature significance approach by equitably distributing the contribution of features to the model prediction [71]. The SHAP analysis demonstrates the key drivers of heat risk and their contribution for different LCZ types (Figure 8): LCZ 1 and LCZ 2 have the highest contribution of healthcare resource distribution to heat risk, suggesting that healthcare resource coverage plays an important role in mitigating heat risk in a high-density built-up area. Population density is the main driver of heat risk in LCZ 3, with a decrease in per capita resource availability (e.g., medical facilities) elevating risk. In contrast, the heat risk of LCZ 4 and LCZs 6–10, which have relatively low building densities, is primarily driven by LST. LCZ 9, in particular, has a significantly amplified contribution from LST due to its high degree of integration with the natural ground surface, and the enhancing effect of bare soil cover on thermal radiation is particularly prominent in this area. While LCZ 6, LCZ 7, and LCZ 10 show similar trends, LCZ 9 exhibits absolute dominance of LST due to the lowest building density.

4.2.2. Strategy Response

Combined with the results of SHAP analysis, the strategic response is proposed (Figure 9): additional smart medical emergency stations are installed in high-population-density neighborhoods (LCZ 1 and LCZ 2), such as Minzhu Road and Hongshan Business District, equipped with emergency cooling devices, and prioritized to cover neighborhoods with high heat risk. In addition, for LCZ 8 and LCZ 10, such as Midong Industrial Zone, the solution would be laying heat storage floor tiles, adding fogging cooling devices, and combining vegetation barriers to achieve surface heat blockage to reduce LST. Bare soil zone management strategies are proposed for urban desertification fringes, suggesting a combination of vegetation modification and albedo modulation to transform these neglected zones into heat buffer zones. This approach may provide a more cost-effective solution for expanding arid cities. In addition, shrub cultivation supplemented with gravel mulching could be implemented in the LCZ F zone along the Jingxin Expressway to reduce surface albedo and radiation suppression while restoring the ecology. In the LCZ E block of JD’s Asia No.1 Logistics Park, permeable pavement is laid, and grass tiles are laid in the car park to improve vegetation cover and reduce surface radiation. In the logistics park, green roofs and other means are used to reduce the heat load of buildings and achieve green warehousing. At the same time, the positive role of the blue-green space LCZ D and LCZ G should be enhanced by constructing an artificial wetland along the LCZ G of the Heping Drainage Canal and extending the cooling effect of the water body to a wider area by using the recycled water recycling system. Lay out protective forest belts in the northwestern LCZ D plots of the farming area to ensure the area of blue-green space in the arid zone. In addition, at the policy level, popularize the policy control under the guidance of the LCZ framework, and implement restrictions on different development intensities for different LCZ types. Real-time control of population, buildings, and other indicators under the LCZ block unit is constructed to achieve high-precision heat risk dynamic early warning and guide urban small area development planning.

4.3. Limitations and Prospects

Based on the LCZ classification system and the HEVA heat risk assessment framework, this study reveals the spatial differentiation patterns of heat risk in typical cities in arid zones, but there is still room for improvement. Firstly, due to the limitation of spatial resolution of socio-economic data, it is difficult to accurately characterize the numerical distribution of the 30 m rasterized data obtained by spatial interpolation, and future studies should seek more accurate numerical sources. Secondly, in the process of LCZ classification, there is still a little misjudgment of the ground cover type, and the subsequent research needs to optimize the algorithm, extract the sample features in depth, and establish the exclusive classification rules for arid zones by combining with the ground survey data.

5. Conclusions

In this study, the spatial heterogeneity of heat risk at the neighborhood scale and its driving mechanism were systematically revealed by integrating LCZ zoning and HEVA assessment framework and applying SHAP analysis in a typical arid zone city of Urumqi.

It is found that the LCZ of Urumqi shows the distribution pattern of bare soil surrounding the built-up area and the typical arid zone characteristic of water scarcity. There are both core and peripheral zones of high heat risk, with the high-density built-up area in the city center becoming the core zone of heat risk due to the double agglomeration effect of population and buildings, and the bare-soil area in the outskirts of the city forming the secondary risk zone due to the strong surface radiation in arid climate, which together lead to the overall elevated level of heat risk in the study area. This result can provide local governments with a warning of the direction of climate risk and delay land development of construction types in high-risk areas in urban planning. In addition, urban desertification fringe areas can be used as a focus of dry zone strategy to guide the role point of sustainability planning strategy. It is noteworthy that LCZ D/G shows significant thermal mitigation efficacy in localized areas, verifying its importance as an ecological safety barrier for cities in arid zones. Meanwhile, the quantitative analysis based on the SHAP model further indicates that LST is the dominant factor influencing the heat risk in low-density built-up areas. This can provide a scientific basis for the construction department to revise the controlling norms in terms of building density, as well as ideas for the macro layout of, for example, urban wind corridors and thermal buffer zones. Furthermore, the LCZ in the study demonstrated its applicability in arid regions, filling the gap of the LCZ tool in extreme climate zones around the globe, and the framework of the LCZ, together with HEVA and SHAP, expands the universality of the tool in heat risk assessment and strategy response, and also enriches the completeness of the heat risk assessment system, which provides a replicable methodology for researchers in the field of heat risk across climatic zones in the international arena.