Analysis of Spatially Varying Relationships Between Urban Environment Factors And Land Surface Temperature In Mashhad City, Iran

Land Surface Temperature (LST), in particular for the urban environment, is a key indicator to characterize urban heat changes, urban climate, global environmental change, and human-environment interactions. However, due to differences in the local spatial variations of LST and the related influence factors, few studies have discussed the spatial non-stationarity and spatial scale effects within urban areas. Moreover, in cities such as Mashhad, which are located in a hot and dry climate, have been less studied of the relationship between LST and urban influencing factors on a neighborhood scale. In the present study, the spatial distribution of the mean LST was evaluated in association with the 16 explanatory indices at the neighborhood's level in Mashhad City, Iran, as a case study. To assess the main components contributing to the LST variations, Principal Components Analysis (PCA) was employed in this study. Additionally, Ordinary Least Square (OLS) and Geographically Weighted Regression (GWR) models were used to explore the spatially varying relationships and identify the model's efficiency at the neighborhood's scale. Our findings showed the five most important components contributing to LST variances, explaining 86.2% of the variability. The most negative relationship was observed between LST and the morphological features of neighborhoods (PC3). In contrast, the landscape composition of the green patches (PC2) exhibited the lowest negative impacts on LST changes. Moreover, road and traffic density characteristics of the neighborhoods (PC4) were the only effective components to alert the average LST positively. With R 2 = 0.678, AIC c= 2125.6, and Moran's I= 0.018, the results revealed that the GWR model had better efficiency than the corresponding non-spatial OLS model in terms of the goodness of fits. It suggests that the GWR model has more ability than the OLS one to predict LST intensities and characterize spatial non-stationary. Therefore, it can be applied to adapt more effective strategies in planning and designing the urban neighborhoods for mitigation of the adverse heat effects.


Introduction
For a long period, an increasing number of studies have been conducted to reveal the relationship between LST and driving factors in various disciplines. However, concerning urban studies, LST has been considered as a fundamental parameter for controlling the physical, chemical, and biological processes of the Earth's surface (Pu, Gong, Michishita, & Sasagawa, 2006). According to the type of relations, trends, and corresponding effects, these studies can be divided into three main distinct categories. The first group includes the causality studies mainly focused on the associations between LST and urban environment components. These works particularly have a wide range of urban morphology indices, vegetation index, landscape metrics, spatial patterns of Urban Green Spaces (UGS), and climatic characteristics (Y. Wang, Chau, Ng, & Leung, 2016).
Although most of the mentioned studies refer to the various number of variables, the findings mainly include the analysis of the bivariate relationship without considering the confounding effect of other variables. Indeed, using spatial data, acquired directly from the remotely sensed cases, has been established to be non-stationary and varied in geographic characteristics (Su, Foody, & Cheng, 2012). Besides, the results of conventional statistical methods such as the Ordinary Least Squares (OLS) may not reflect all the key details of the relationships between variables if the non-stationary spatial property of data is not considered as a major limitation when geo-referenced data are applied (Brown et al., 2012;Brunsdon, Fotheringham, & Charlton, 1996;Luo & Peng, 2016).
Recently, a non-parametric technique has been developed by Brundson, Charlton, and Fotheringham to overcome this problem by extending traditional standard global regression approaches. OLS compared to Geographically Weighted Regression (GWR), considers the spatial structure of data in the modeling. GWR as a local approach is an effective model able to represent the spatial variations to explore the potential relationship between variables (Y. Huang, Yuan, & Lu, 2019;Z. Wang, Fan, Zhao, & Myint, 2020). GWR provides extra information on the spatial processes by a local model of the variables, varying over space to detect locally different parameters with spatial dependence and heterogeneity of regions (Fotheringham, Brunsdon, & Charlton, 2003;Harris, Fotheringham, Crespo, & Charlton, 2010). Particularly, this approach enables measuring the spatial instability of coefficient over the study area regarding LST and the urban environment variables, tending to be spatially varied by the type and local conditions like activities, ecological and morphological patterns.
The present study is carried out at the neighborhood scale to evaluate the effects of different indicators of the urban environment on LST in Mashhad city, Iran. This study aims to achieve the following objectives: 1) investigating the spatial non-stationarity modeling by GWR in the relationships between LST and the urban environment indicators; 2) choosing a good fitting model for comparing GWR and OLS; 3) examining the spatial distributions of LST and the impact factors; 4) exploring the relationships at the neighborhood scale to adopt proper strategies for heat mitigation and urban warming.
In this study, the characteristics of the urban environment, in particular, were identified by 16 indicators, implying various aspects of natural and physical elements, including built-up areas, streets, spatial configuration, land use, and landscape. Furthermore, the GWR model was used to highlight the details of spatial information and enhance the accuracy of relationships between the variables. Due to variation and complexity in the selected urban datasets, redundancy and multicollinearity problems may be added to the difficulty in interpreting the relationships between the influencing variables. Therefore, Principal Component Analysis (PCA) was used in combination to reduce the data dimensionality and complexity. This pre-analysis helped to organize a large number of variables into a few and more informative groups. Moreover, the current work applied the ordinary least square (OLS) to compare to the findings of the GWR concerning residual spatial autocorrelation, allowing to find a goodness-of-fit model. This has been explained in detail in the research method section.
Briefly, many scholars have applied the GWR model to explore the relationship between LST and urban factors or explanatory variables in recent years (Kalota, 2017;S. Li, Zhao, Miaomiao, & Wang, 2010;Luo & Peng, 2016;Su et al., 2012;Z. Wang et al., 2020;C. Zhao, Jensen, Weng, & Weaver, 2018). Although, in the recent literature, the related advantages of the GWR model have been well-documented over those of the other models, utilizing the idea of the GWR in combination with PCA concerning LST and the urban variables has rarely been discussed. Also, a lack of the scale's investigations concerning urban study is observed; however, this study is prominently at the neighborhood scale for better understand the distributions of LST, urban indicators, and existing relationships, as the neighborhood scale reveals more details. Therefore, the neighborhood scale can improve our knowledge concerning the driving mechanisms behind the formation of thermal environments and relative changes. Besides, the related findings can fill the gap in the corresponding scale and significantly formulate urban planning policies to alleviate LST effects. This study presents findings that are of major significance to climate experts, urban and regional planners, environmentalists, and policymakers to explore measures for mitigating uncontrolled urban expansion and thermal discomfort in Mashhad City.
With a height of 970 meters, the city has a cool-dry climate with warm-dry summers and coldhumid winters. The average temperature is 15℃, with 233.8 mm annual rainfall and 53% relative humidity. Moreover, the number of freezing days per year for this city amounts to 85 days.

Figure 1. Overview of the study area
It is the second-largest city and the most important population center in the northeast of Iran ‫و‬ Moreover Mashhad is one of the main metropolitans in Iran, subdivided into 787 neighborhoods in terms of urban management. Currently, with more than 3,000,000 people, it is the most populated city in the country. In the last three decades, Mashhad has experienced rapid population growth and corresponding physical transformation due to its excellent situation in pilgrimage, tourism, agricultural, and commercial contexts. Accordingly, the development of the urbanized areas has led to a change in climatic parameters compared to its surroundings, resulting in air pollution and excessive heat release. In this study, the characteristics of the urban environment, in particular, were identified by 16 indicators, implying various aspects of natural and physical elements, including built-up areas, streets, spatial configuration, land use, and landscape.
Furthermore, the GWR model was used to highlight the details of spatial information and enhance the accuracy of relationships between the variables. Due to variation and complexity in the selected urban datasets, redundancy and multicollinearity problems may be added to the difficulty in interpreting the relationships between the influencing variables. Therefore, principal component analysis (PCA) was used in combination to reduce the data dimensionality and complexity. This pre-analysis helped to organize a large number of variables into a few and more informative groups. Moreover, the current work applied the ordinary least square (OLS) to compare to the findings of the GWR concerning residual spatial autocorrelation, allowing to find a goodness-offit model.

2.2.Derivation and selection of variables
In the present work, two different types of datasets, namely primary and secondary data, were utilized. Primary datasets included Landsat image, Landsat Operational Land Imager (OLI), Thermal Infra Red Sensor (TIRS), land use map, population, road network, and traffic. To obtain better results in this study, LST was processed later using the maximum likelihood classification in ENVI software.

Dependent variable: LST
Landsat 8 image was launched on 14 September 2019 from the Earth Explorer website to estimate the LST. With the spatial resolution of 100m, the thermal infrared sensor (TIRS) band (band 10) was employed to enable the atmospheric correction of the thermal data and retrieve LST with minimum error. To assess LST, the following equation was used to convert the digital number (DN) values of band 10 to top of atmosphere (TOA) spectral radiance (Barsi et al., 2014). Then, TIR band data was converted from spectral radiance to brightness temperature (BT) using the following equation: Where K1 and K2 stand for the band-specific thermal conversion constants from the metadata (Avdan & Jovanovska, 2016). By adding the absolute zero, approximately equal to -273.15, the results can be obtained in Celsius (Han-qiu & Ben-qing, 2004). In the current study, the dependent variable was mean LST, calculated for each neighborhood using the zonal geometry tools in Arc GIS 10.3.

Explanatory variables
The explanatory variables examined, including 16 various influencing factors, were categorized as primary and secondary data. To clarify urban morphology effects, the urban morphology was redefined into four aspects, as shown in Table 1

landscape structure
To measure the landscape composition and configuration, landscape metrics have been broadly introduced as descriptive and quantitative spatial patterns (McGarigal & Turner, 2001;Riitters, Vogt, Soille, & Estreguil, 2009). In this study, five landscape metrics were chosen to analyze the spatial patterns based on the area, size, shape, continuity, and complexity of the urban green patches. As presented in Table 1 in detail, the selected landscape metrics include the largest patch index (LPI), class area (CA), the effective mesh size (MESH), landscape shape index (LSI), and mean fractal dimension index (Frac_MN).

Urban spatial configuration
Urban spatial configuration is the arrangement of urban land use/land cover, geometry, and layout (Krafta, 1997;Marshall, 2005). Although a few studies have addressed thermal behavior and urban spatial configuration using urban geometry, street canyons, and orientations ( required. In the current study, the space syntax approach was specifically employed to analyze the spatial configuration by evaluating the connectivity, movement intensity, and permeability of the urban network at the neighborhood level. Based on the approach, the proposed indicators were as follows. i) Integration is a measurable degree of concentration of space in comparison to other spaces in urban systems. This was employed at a global scale specified with radius n and a local scale determined with radius 3. The former scale considers the total number of steps or direction changes in urban networks and represents the overall urban structure, whereas the integration R3 can be used in investigating pedestrian movement (Hillier, 2007). ii) Connectivity is defined as the number of spaces (e.g., street) that are directly joined with a particular space, or a relative connection can exist between spaces in reality. iii) Mean depth, as a topological distance, describes the value of segregation for a specific space to all others in the urban spaces. iv) Intensity evaluates the rate of entropy change with total depth. It is an appropriate index to interpret movement efficiency by giving the distance in the network (Park, 2005). in the current work and also the possibility of the internal relationship between the variables affecting LST in the study areas, the PCA method was used in the first step. As a pre-analysis step, PCA is an exploratory technique to obtain important variables in the form of components by dimensionality-reduction method (Jolliffe & Cadima, 2016). PCA includes the following equation: where pcij is the component score of the jth principal component for cell i, Xik is the value of the kth layer for cell i, and Ekj is the element of the eigenvector matrix at row k and column j (Yue, Liu, Fan, Ye, & Wu, 2012). before PCA, the numeric datasets were standardized using the standard deviations, providing the dimensionless variables with unit variance and standard deviation.

OLS analysis:
Traditionally, OLS is a multiple regression analysis based on spatially constant features of the dependent and independent variables within the study area (S. Li et al., 2010). Generally, it suites the various datasets and could be applied equally across the whole study area (Tian, Qiu, Yang, Xiong, & Wang, 2012). According to the assumption of observation independence, OLS would be failed when applied to the spatial data referring to the geo-referenced data. OLS model can be written as (Luo & Peng, 2016): where yi=1,2,…n represents the dependent variable, xk (k= 1,2,…M) represents the explanatory variable, β0 is the intercept, βk is the regression coefficient, and ε stands for the random error. In the present study, applying the OLS regression model, the nature of relationships between the underlying factors and LST was found, directly allowing to compare the findings with GWR.

GWR analysis:
Since the observed geographical and ecological patterns had been investigated in this paper, the GWR regression model was tested to produce prediction and parameter estimates at the neighborhood scale effectively. Mathematically, the GWR model takes The coefficient of determination (R 2 ) is a value to measure the model fitness, computed to reflect the relationship of each explanatory variable. By a comparison between the predicted values at each regression point and the observed values, the higher R 2 could be used to indicate better performance and stronger predictive ability of two conducted models (Propastin, 2009).
Additionally, two main performance indicators, including the global Moran's I and the corrected Akaike Information Criterion (AICc), were tested in this study. Comparing the results, AICc is an important value to assess fit for both the GWR and OLS model's performance. In general, the best model is determined by lower AICc, indicating a closer approximation of the model to reality (Tian et al., 2012). Also, the global Moran's I is a clustering analysis statistic able to calculate the spatial autocorrelation of the residuals based on feature location and value, ranging from -1 to 1. The model will be superior if the distribution of the residuals obtained from the regression model tends to be dispersed. In other words, the lower Moran's I suggest more random spatial patterns in the residuals by the best model. When using the spatial regression models, the residuals obtained from fitting the models must be locally independent of each other; hence an autocorrelation coefficient is used to test the spatial dependence of the standard deviation of the residual (std-residual), obtained from OLS and GWR models.

LST measurement
The LST analysis of Mashhad city was done at the neighborhood scale ( Figure 2). As mentioned earlier, to compare the neighborhoods, the average temperature of each polygon was computed. Overall, the results demonstrated that the neighborhoods within the urban center had a relatively lower temperature than those in the suburbs due to an abundance of green space and gardens, in particular, Astan-e-Ghods gardens like Malekabad and Imam Reza with 42.4 °C and 41.04 °C average temperatures, respectively. Exceptionally, some neighborhoods had higher temperatures than others due to adjacent land use types like Mashhad railway station with an average temperature of 45.41 °C or being affected by the exposed solar radiation, mainly located around the holy shrine (43.51 °C).

Spatial distribution of the variables
The spatial distribution of the explanatory variables is mapped in Figure 3.

Calculation of PCA
Among the 16 mentioned variables, five principal components (PCs) were identified using the Cut-Off point of the Scree plot and the eigenvalues higher than 1. Table 2 provides the eigenvalues and percentages of the explained variance of the selected components. Also, the first five principal components exhibited the highest variance in the original data using the proportion of variance explained (PVE). Accordingly, the PCs determined 86.2% of the variability by cumulative percent variance. The first component (PC1) with the largest eigenvalue (6.58) accounted for 34.66% of the total dataset variability of the datasets. As presented in Table 3, this PC combined almost similarly sized contributions from three main related variables, including intensity, mean depth, and integration Rn. Moreover, PC2, with 20.25% of the total variance, was dominated by the contribution of MESH, LPI, and CA indices, being mainly associated with the size and distribution of green patches referred to as landscape composition. Additionally, PC3 represented the average of a group of morphological factors, such as population density, building height, and density, describing 15.37% of the variance. The next component, (PC4), was the combination of the road length, traffic, and integration R3. This component represented the characteristics of transportation and urban flows, explaining a 9.18% variability of the datasets. Finally, PC5 was the combination of average fractal and shape indices, representing only 6.74% of the data variations. However, the results revealed that connectivity and entropy had no significant contributions to the principal components among the variables.

The relationship establishment between LST and the components using OLS and GWR model.
In this section, the relationship between the PCs and LST was established using the OLS and GWR models. The OLS model was first adopted to assess the global statistical relationship between the dependent (LST) and independent variables (PCs). Furthermore, the corresponding GWR model was used to investigate the non-stationary spatial relationships between the dependent variable and the components in the OLS model. A summary of the results in both OLS and GWR models is presented in Table 4. where an increase in PC2 is associated with a decrease in LST up to 0.14 degrees.
To compare the models, the results of the GWR model are listed in Table 4. Accordingly, the strength of the relationships between the five predictors (PCs) and LST boosted significantly. Also, single-PC analysis reflected better performance than the OLS model by a relatively smaller measure of AICc and a higher value of R 2 . Similar to the OLS model, PC3 was identified as the most important factor in LST variations by the highest values of R 2 (0.84) and the lowest AICc showed the area, size, and continuity of green patches to have the least effect on reducing the average of LST in neighborhoods.

Assessment of the models' efficiency
To evaluate the efficiency of both applied models, the estimated LST maps were represented using OLS and GWR models (Figure 4). Overall, the spatial patterns of the LST generated by the models showed a similar trend and goodness of fit with the original map. The predicted map usually is characterized by LST higher values concentrated in the suburbs and lower values located mostly in the neighborhoods in the city center. Additionally, it can be seen that the minimum and maximum of predicted values are beyond the observed data. However, these indicate that the findings reflect changes in variables and highlight a much detailed spatial distribution of LST. A noticeable difference is that the generated maps indicate more variations in the neighborhoods situated in the east and northwestern of the city, where the higher values of LST happen mostly in developing neighborhoods with extensive bare lands, high scattered built areas, and insufficient UGS.
Moreover, the corresponding results of the models are statistically compared by describing data variations and addressing spatial autocorrelation. Since the model is non-stationary, the GWR model computes the predicted values well for spatial differentiation in the neighborhoods. As presented in Table 5, the coefficient of determination (R 2 ) of the OLS model is 0.343, whereas the R 2 for the GWR model is 0.678. Towards a better fitting model, the higher value of R 2 is considered to explain the relationships more efficiently. This implies that GWR is a powerful model for explaining the LST changes in response to the underlying components and would extremely enable the findings to fit the observed LST. Also, the higher R 2 suggests that the relationships between LST and the PCs may exhibit spatial nonstationarity. Moreover, the AICc value for GWR is lower than OLS, with a notable decrease of more than 377.86 points, being strong evidence for a real improvement in the results based on the local model (GWR). In addition to the comparisons mentioned above, the Moran's I for the GWR model gives a smaller value (0.018) than the OLS one (0. 272), exhibiting a more random spatial pattern of residual values in the GWR model. This means the OLS model does not fully explain the whole datasets due to a violation of normality assumption and hence cannot provide closer representations of reality than the GWR model. Therefore, these results confirm the efficiency of GWR over the OLS model, addressing the ability to predict the LST variations and spatial autocorrelation.

Discussion
This research used GWR and OLS models to explore the associations between LST and the urban environment indices in Mashhad urban neighborhoods. In the pre-step analysis, the explanatory variables were redefined by the PCA procedure and reassembled in the first five group components ( Figure 5).

Figure 5. Spatial distribution of PCs in Mashhad's neighborhoods
In general, the results reported clear connections between LST and PCs at the neighborhood level.
Furthermore, the findings confirmed the regression relationships between LST and PCs to be identified by significant spatial non-stationarity. The results provided clear evidence that the GWR model made a significantly better prediction of LST by smaller AICc, higher R 2 , and the smaller value of Moran's I compared to the OLS model. The results are discussed in the following parts:

PC1 and LST
The first component, modeled by intensity, mean depth, and integration Rn, explained 34.66% of the total variance. Configurationally, the variables contributing to PC1 are mainly dealing with urban cores, space geometry, road orientation, movements, and land use/ land cover arrangement.
As displayed in Figure 5, the high values of PC1 generally appeared in the neighborhoods extending from the urban core (in the center) to the west, whereas the low values were distributed in the suburb. Overall, the results indicated that a lower mean LST was detected in a smaller neighborhood with higher integrated spaces in comparison to larger neighborhoods and higher segregated spaces.
Our findings confirmed the LST intensities to be derived via the various configuration of the urban neighborhoods negatively. Besides PC3 and PC5, PC1 was found to be the most influential component able to mitigate the LST in two effective ways: internal spatial structure and contiguity of the neighborhoods. These findings emphasize the role of neighborhood configuration in representing the internal spatial structure with maximum integration and intensity, thereby holding the various parts of a neighborhood constant. On the contrary, neighborhoods with low configurational attributes do not have a coherent internal structure, being identified with the dispersed distribution. Moreover, spatial configuration widely organizes the arrangement of urban land use, spatial patterns of land cover, and structure of urban morphology in the small neighborhoods that tend to have more compact urban clusters, larger contiguity of integrated cores with medium and high-rise buildings, and interconnected spaces. Hence, the clustered built-up neighborhoods may have a more suitable urban canopy to reduce direct sunlight and consequently the LST effect as it is closely related to airflow and the city's ventilation conditions. Indeed, the spatial configuration can optimize high-rise building arrangement and street layout around the integrated spaces to extend shades and urban canopies. Therefore, it is comprehensible that the spatial configuration can qualitatively improve interaction with cooler airflow and cause the mitigation of LST. Although contradictory results have been reported in some cases (Coutts, Beringer, & Tapper, 2007;Schwarz, 2010;Yue et al., 2019), our findings are in agreement with previous work (Lan & Zhan, 2017;Soltanifard & Aliabadi, 2019;Yin, Yuan, Lu, Huang, & Liu, 2018). Differences in approach, statistical methods, and climatic characteristics may be contributing to these inconsistencies in the results.

PC2 and LST
The starting point to assess the landscape structure in the studied areas was to investigate the spatial patterns of the green spaces in terms of landscape composition and configuration. However, MESH, LPI, and CA were the main contributors of PC2, which were summarized as the LST predictors by the PC analysis method. As shown in Figure 5,

PC3 and LST
Based on the relationship between PCA's and LST, PCA3 has the strongest effect on the average LST among other components. Building height, building density, and population density were three main contributors to PC3, describing the morphological characteristics of neighborhoods. ways (Guo, Zhou, Wu, Xiao, & Chen, 2016;Jin, Cui, Wong, & Ignatius, 2018). However, we found the negative effect of the results to be related to building height and clusters with shadows, where increasing height and density could mitigate LST even during the strong sunlight and heating influence. Our results are corroborated by some studies (Cai, Ren, Xu, Lau, & Wang, 2018;Jin et al., 2018;Scarano & Mancini, 2017;Sun, Gao, Li, Wang, & Liu, 2019).

PC4 and LST
In the current work, PC4 was another responsible component for explaining the variance of LST at the neighborhood level. The component had a heavy loading matrix with road length, integration R3, and traffic density, indicating the transportation features of the urban network. The results showed that PC4 positively correlated with the average LST. Due to the fundamental transportation elements, such as traffic, street features, and movement patterns, indeed, the complex transportation network forms components of urban structure and its associations from the scale of settlements to the scale of the street block (Marshall, 2005). Given the main contribution of transportation elements to urban structure features, several studies confirm the role of transportation to alter the behavior of LST patterns significantly (Giridharan, Lau, Ganesan, & Givoni, 2007;Lee, Moon, Choi, & Yoon, 2018;Oke, 1988). Based on the findings, the increasing trend of LST directly is related to an effective decrease in surface albedo, in particular, the impervious surfaces, including streets and pedestrian road pavements. Moreover, land use, urban space geometry, and road construction affect traffic density in the neighborhoods, outlining the heat behavior along with both road and land use types. Accordingly, the increased input of anthropogenic heat and street surface type causes heating emission of the road to the urban environment and increases LST. These results are confirmed by previous studies (Chapman, Thornes, & Bradley, 2001;Oke, 1988;Walawender, Szymanowski, Hajto, & Bokwa, 2014).

PC5 and LST
In this study, the PC5 is characterized by the shape (LSI) and the complexity of the green patches (Frac-MN), referred to as the configuration features at the neighborhood level. Figure  however, the simultaneous impacts of the components is negative on LST. Unlike the neighborhoods in the suburbs, the results reflect that the high aggregation and density of the green patches in the neighborhoods could mitigate heat intensities by representing a higher cooling function. These suggest that less complexity and shape of the green patches may better regulate LST interestingly. The results of the present work are consistent with those of several previous studies (Connors, Galletti, & Chow, 2013;Dugord, Lauf, Schuster, & Kleinschmit, 2014;Peng et al., 2011;Xie, Wang, Chang, Fu, & Ye, 2013;Zhou, Huang, & Cadenasso, 2011).

Conclusion
In summary, the present work thoroughly highlights the following major findings: First, as regards the LST trends, urban neighborhoods increasingly showed distinct LST changes from the center to the suburb, affected by urban configuration, morphology, roads, traffic, as well as landscape composition and configuration.
Second, all explanatory variables demonstrated a significant contribution to LST intensities. The morphological features of neighborhoods (PC3) had the most negative influence on LST. In contrast, the landscape composition of the green patches(PC2) exhibited the lowest negative impacts on LST changes. Furthermore, road and traffic characteristics of the neighborhoods (PC4) were the only effective components to alert the average LST positively. Therefore, our empirical results suggested developing high-rise buildings to provide shading during days, besides increasing urban trees and vegetation to produce the cooling effect.
Third, this study confirmed that, compared to the OLS model, the GWR model had an excellent ability to detect the scale-dependence of the relationships between the LST and the components, especially in heterogeneous regions. Hence, for urban policymakers, the GWR model may be suitable assistance to determine the most reliable action plan at the neighborhood level.
Finally, although removing the hot and impervious surfaces is impossible, urban planners and designers can mitigate heat effects by reforming urban morphological context, improving planting quality, and redesigning urban configuration. In particular, for developing countries, like Iran, it is suggested to pay more attention to the arrangement of buildings, the configuration of green spaces, and the enhancement of the urban spatial configuration at the neighborhood level as effective strategies to create a comfortable outdoor environment, alleviating LST and corresponding adverse effects. This study presents findings that are of major significance to climate experts, urban and regional planners, environmentalists, and policymakers to explore measures for mitigating uncontrolled urban expansion and thermal discomfort in Mashhad City. This caters to the vision of the United Nations Sustainable Development Goals (SDGs) for sustainable cities and communities (SDG 11), and good health and wellbeing (SDG 3) and Take urgent action to combat climate change and its impacts ( SDG 13) by 2030.