Urban Climate

Volume 53, January 2024, 101791
Urban Climate

Effects of 2D/3D urban morphology on land surface temperature: Contribution, response, and interaction

https://doi.org/10.1016/j.uclim.2023.101791 Get rights and content


  • 3D urban metrics play a more critical role in shaping LST than 2D urban metrics.

  • Interactions of 3D-3D metrics pairs are stronger than 2D-3D metrics pairs.

  • Interactions between 2D-3D metrics pairs depend on the specific thresholds.

  • Interactions between 3D-3D metrics pairs generally show a synergy effect.


Urban morphology severely affects the intra-urban heat flux transport and thus directly regulates urban thermal environment. Despite previous studies suggested that urban morphology may significantly contribute to land surface temperature (LST), few studies simultaneously discuss the effects of urban morphology on LST in different urban units, particularly in the irregular block scale. Here, we explored the relationships between multi-dimensional urban morphology and LST using the eXtreme gradient boosting (XGBoost) model and the SHapley Additive exPlanations (SHAP) method. Especially, we revealed the interaction of the urban morphological metrics on LST at the block (i.e., irregular size delineated by roads) and the grid scale (i.e., a regular grid size with 200 × 200 m). The results show that the contributions of urban metrics on LST follow the rank of 3D building metrics >2D landscape metrics of impervious surface (IS) > 2D landscape metrics of urban green space (UGS). Average building height (AH) and the percentage of landscape of IS (PLAND_IS) are the most important metrics at the block and the grid scale, respectively. Moreover, some metrics have specific thresholds of the warming or cooling effects, e.g., the cooling effect when AH exceeds 19 m and warming effect when PLAND_IS >49%. Finally, we found that the interaction effects of 3D-3D metrics pairs are stronger than that of 2D-3D metrics pairs. These findings provide useful information for understanding the driving mechanism of the 2D and 3D urban morphology on LST and optimizing the urban morphology for mitigating the urban heat islands effect.

Urban heat island
Interaction effects
SHAP method
Sustainable development goals

1. Introduction

Urban Heat Island (UHI) effect caused by rapid global urbanization and industrialization has led the world's major cities to warm at twice the global average rate (IPCC, 2021). Continuing urbanization and increasing greenhouse gases emission will result in 1.6 billion exposing to high temperatures by the middle of the century (C40 Cities, 2021). Rapid urbanization has also accompanied with the formation and transformation of spatial configuration and structure changes of landscape, calling as urban morphology (Elzeni et al., 2022). Urban morphology is considered as the study of coherent neighborhood morphology (e.g., open spaces, building) and function (e.g., human activity represented as the spatial characters of land cover) (Azhdari et al., 2018; Bonsu and Bonin, 2023; Mu et al., 2022). The transformation of urban morphology changes the transmittance and evaporation of land cover and it consequently alters thermal properties of urban land surface (Grigoraș and Urițescu, 2019). Therefore, widely observed drivers of urban morphology can provide deep understanding of the relationship between urban thermal environment and urban resilience (Azhdari et al., 2018; Huang and Wang, 2019).

Changing the physical characteristics of urban morphology to mitigate the hazards of overheating in cities has been explored in previous studies (Liu et al., 2020b, Liu et al., 2020c; Zheng et al., 2023). As revealed by many urban climatology studies, for example, the traditional layouts of urban centers with high impervious surface (IS) density caused an increase of urban temperatures (Li et al., 2011; Wang et al., 2022b). In contract, the management and planning of urban green space (UGS) composition and configuration can provide pronounced cooling effect (Du et al., 2016; Wang et al., 2022a). With the continuous progress of remote sensing techniques and the widely use of remote sensing product, the variables of 2D urban morphological metrics, including land-cover/use types (Petralli et al., 2014; Tabassum et al., 2023), landscape composition and configuration patterns (Zhou et al., 2017), and street patterns, have been used to investigate the their relations between LST (Liu et al., 2018; Xiao et al., 2022). However, abovementioned research only discussed bivariate associations between 2D urban morphological metrics and LST, the relative contributions of 2D and 3D urban morphology characteristics to LST is not fully explored.

The urban thermal environment is modulated by both 2D land cover/use organization and 3D characteristics of buildings, where the tall buildings increase solar absorption in the daytime and reduce the heat loss in the nighttime (Salvati et al., 2019). Given the complex urban morphology, 2D land cover fails to represent the spatial heterogeneity of real landscape structure (Kong et al., 2022; Yuan et al., 2023). Increasing studies have investigated the influences of 3D building morphology on LST to of complex 3D structures (Li and Hu, 2022; Zeng et al., 2022; Zhou et al., 2022b). Researches have used point cloud data obtained from airborne LiDAR to calculate the metrics (e.g., height, volume, and aboveground biomass) to character vegetation geometry (Chen et al., 2022a; Kong et al., 2022; Plowright et al., 2017; Yu et al., 2020), and thereby to explore the effect of 3D vegetation on the urban thermal environment. These studies provide quantitative assessments of the impact of 3D landscape on LST. However, the effect of 2D and 3D urban morphological metrics on LST remains in disagreement with relevant studies. For example, Berger et al. (2017) concluded that 3D urban morphological metrics are consistently outperformed by some of the most widely used 2D metrics, while Chen et al., (2022b) found that summer LST variation is governed by more 2D metrics than that of 3D metrics. The inconsistent conclusions from these studies mean that the impacts of the 2D and 3D urban morphology on LST remains under debate.

The connection between urban morphology 2D/3D characteristics and LST is complex and multidimensional. For exploring the relationships between urban morphological metrics and LST, relevant studies mainly adopted the correlation coefficient (Zhou et al., 2022b), linear regression (e.g., multiple regression (Zhou et al., 2011), spatial regression (Chun and Guldmann, 2018)), geodetector (Hu et al., 2022), machine learning (e.g., elastic network regression (Zhao et al., 2020), random forest (RF) (Zhao et al., 2019), and the eXtreme gradient boosting (XGBoost) model (Zhou et al., 2022a)), etc. Recently, deep learning (i.e., convolutional neural network (CNN)) has been used to detect the effects and relative importance of urban morphology on LST (Logan et al., 2020), demonstrating the potential of CNN for geospatial data. With the advance in the analysis methods, the nonlinear characteristics in the effect of urban morphology have been gradually explored. Especially, the threshold effects of urban morphological metrics on LST (Hu et al., 2020; Yuan et al., 2021). Although the nonlinear relationships have been extensively investigated by machine learning and geodetector, however, only a few studies have considered the interaction effects (i.e., a kind of non-linearity) of 2D and 3D urban morphology characteristics on LST. Recently, the SHapley Additive exPlanations (SHAP) method proposed by Lundberg and Lee (2017) demonstrated the advantages of both demonstrating the relationship between the explainable variables and describing their impacts on the prediction. Therefore, SHAP method allows us to consider interaction of 2D and 3D urban morphology in a more continuous way and address the influential factors with complicated behavior and inter-association.

In addition, previous studies generally investigated the impacts of the urban morphology on LST using the regular grid units (Chun and Guldmann, 2014; Dai et al., 2018), which may produce biased estimates (Huang and Wang, 2019) and are difficult for offering practical, operational guidelines for urban managers (Yin et al., 2018). Besides, the landscape segmented by regular grid units may introduce uncertainties in determining the impacts of urban morphology on LST, and the segmented landscape patches cannot reflect its composition and configuration (e.g., structure-function boundaries of the landscape) well. In contrast, the irregular urban block unit can be regarded as an appropriate and operational scale for urban development projects and the implementation of urban sustainability measures (Chen et al., 2022a; Yin et al., 2018) and it has been widely taken as an analytical unit in recent LST research (Hu et al., 2020; Yao et al., 2020; Yuan et al., 2021). Since the relationship between LST and urban morphology is scale-dependent (Niemelä, 1999), we adopted the regular grid scale and the irregular block scale simultaneously to investigate the impact of urban morphology on LST to mitigate the uncertainty introduced by a single analytical scale (Plowright et al., 2017).

In summary, a sufficient explanation of intra-association between 2D and 3D urban morphology, as well as their relative importance to LST remains unclear. Considering the scale of analytical units may influence the variances of urban morphology (Plowright et al., 2017; Zhou et al., 2022a, Zhou et al., 2022b). We first characterized the 2D and 3D urban morphology at two scales (i.e., the irregular block unit and regular grid unit) using the high-resolution GaoFen-2 (4 m) remote sensing images and open sources of geographic information data. Then, we applied the XGBoost regression and the SHAP method to investigate the relationships between 2D and 3D urban morphology and LST at the block and grid scales separately. The stability and variations of these relationships aim to answer the following three questions:

  • (1)

    What are the differences of 2D and 3D urban morphology characteristics related to LST in the irregular block and regular grid units?

  • (2)

    What are the comprehensive effects of 2D and 3D urban morphology characteristics on LST?

  • (3)

    How does the interaction effects between 2D and 3D urban morphology characteristics play an important role in modulating LST variations?

2. Data and methods

2.1. Study area

Xi'an (107°24′ to 109°29′ E, 33°25′ to 34°27′ N) is in the Guanzhong Plain of China with a semi-humid warm temperate continental monsoon climate (Fig. 1). In the summer of 2018, the average air temperatures in June, July, and August reached 26.6 °C, 28.4 °C, and 29.1 °C, respectively (http://tjj.xa.gov.cn/tjnj/2019/zk/indexeh.htm). Besides, the rapid economic development led to great increases in urban built-up areas and urban population in the last decade in Xi'an (Han et al., 2022). Particularly, the urban built-up area increased by 114.6% and the urban population increased by 87.9% from 2010 to 2020 (Li et al., 2023). More seriously, the rapid urbanization and population agglomeration have exacerbated the UHI effect (Lu et al., 2020), posing a great challenge to promote resilient cities. Our study focused the scope within the expressway of Xi'an (Fig. 1b) and the total area of which is 458.8 km2.

Fig. 1
  1. Download : Download high-res image (2MB)
  2. Download : Download full-size image

Fig. 1. Study area. (a) location; (b) averaged LST in summer in 2018; (c) land cover in 2018.

2.2. Data collection

Four datasets were used in this study. The Landsat-8 OLI/TIRS images (i.e., spatial resolution of band 10 is 100 m and is resampled to 30 m by USGS, https://earthexplorer.usgs.gov/) were collected to retrieve LST via the radiative transfer equation method (Weng et al., 2004). Because we focused more on the thermal environment of the hot summer. We, therefore, screened all images from June 1, 2018, to August 31, 2018; only three images that were imaged at 11:19 a.m. (local time) on June 23, July 25, and August 10 satisfied the cloud-free condition in our study area. Besides, the weather conditions of these three days and before days were static-wind days and were no-rain days (https://rp5.ru). These three images were further averaged into one image to represent the typical LST in summer.

Gaofen-2 images (i.e., spatial resolution of spectral band is 4 m, https://www.cheosgrid.org.cn/) in 2018 were used to classify the land cover of the study area. In eCongnition 9.0 software, we applied the support vector machine (SVM) method (Bouslihim et al., 2022) to classify land cover into six types, i.e., cropland, urban tree, grassland, water body, IS, and bare land. The overall accuracy (OA) of classification results is 84.1% and the Kappa coefficient is 0.80 (Yuan et al., 2021). We merged urban trees and grassland into one type to represent the UGS.

The building footprint vector (https://map.baidu.com/) in 2018 contains the number of floors of each building, which was widely used as a source of building heights to study the urban climate in China (Li et al., 2021; Yang et al., 2019, Yang et al., 2021a). We visually checked the building floor based on Google Maps and found that only a few building footprints in the fringe area of the study area recorded the wrong building floors, which were manually updated before calculating the 3D building metrics. Note that we assumed that one floor of the building is constant as 3 m (Yang et al., 2021a; Yuan et al., 2021), and the number of floors was multiplied by 3 m to represent the building's height.

The road networks in 2019 that downloaded from www.openstreetmap.org was used to delineate the study area into fine community blocks (totaling 716 blocks). The road classes used included motorway, trunk, primary road, secondary road, tertiary road, and residential road.

2.3. Method

Our analysis framework includes four steps (Fig. 2): (1) Data processing (Fig. 2a). We obtained the basic data for calculating urban morphological metrics from satellite images and building footprint vector. (2) Definition of two analytical units (Fig. 2b). The analytical units included the block scale (i.e., an irregular size) and the grid scale (i.e., a regular size). The blocks are delineated by the road networks from OpenStreetMap. We selected six road classes (i.e., motorway, trunk, primary road, secondary road, tertiary road, and residential road) and set different buffer distances for each road class. Then, we merged all buffer regions and clipped the study area to get the final blocks. The unsuitable buffer areas were deleted manually in advance. Besides, the regular grid with the size of 200 m × 200 m was used as another analytical unit, which was used for comparing the difference of the effects of urban morphology metrics on LST against the block scale, as well as ensuring more sample data for modeling. Note that only blocks and grids that both have 2D and 3D metrics were selected for modeling to avoid the unconvincing results caused by data missing, i.e., 661 blocks and 6670 grids. The sizes of 95% of blocks are larger than a regular grid size (i.e., 200 m × 200 m). (3) Quantification of urban morphological metrics (Fig. 2c). All 2D and 3D metrics were calculated at two scales as independent variables, while the dependent variable was the mean LST at each analytic unit, representing the thermal conditions of each analytic unit. (4) Statistical modeling (Fig. 2d). We first modeled the relationship between urban morphology metrics and LST at two scales using the XGBoost regression, and then we interpreted the impact of urban morphology metrics on LST using SHAP method.

Fig. 2
  1. Download : Download high-res image (861KB)
  2. Download : Download full-size image

Fig. 2. The adopted framework for analyzing the effects of 2D and 3D urban morphology on LST. (a) collecting dataset; (b) delineating two analytical units; (c) calculating the urban morphology metrics; (d) modeling and interpreting the relationship between the urban morphology and LST.

2.3.1. Urban morphological metrics

Through comprehensively reviewed the studies on characterizing the urban morphology (Berger et al., 2017; Hu et al., 2022; Liu et al., 2017), we selected six 3D building metrics that are highly correlated with LST, namely average building height (AH), sky view factor (SVF), average building height standard deviation (AHSD), floor area ratio (FAR), mean architecture projection area (MAPA), and average volume (AV). Besides, we also considered the seven class-level landscape metrics to characterize 2D landscape composition and configuration of land cover. One metric i.e., the percentage of landscape (PLAND), represents the landscape composition, while six metrics, i.e., edge density (ED), patch density (PD), largest patch index (LPI), aggregation index (AI), mean patch size (AREA_MN), and area weighted mean shape index (SHAPE_AM), represent the landscape configuration like area, shape, and aggregation (Li et al., 2011; Zhou et al., 2022a). Briefly, this study selected 13 metrics (Table 1) to represent 2D (i.e., land cover) and 3D (i.e., buildings) urban morphology, all 2D landscape metrics of land cover were obtained in Fragstats 4.2 and the 3D building metrics were calculated in ArcGIS 10.5 and R studio.

Table 1. The summary of 2D and 3D metrics in this study.

Percentage of landscape (%)PLANDPLAND=j=1naijA100Percentage of the areas of all patches of the corresponding patch type to total landscape area (A: m2). aij is the area of each patch.
Edge density (m•ha−1)EDED=j=1neijA10000Total lengths of all edge segments involving the corresponding type divided by the total landscape area. eij is the length of jth patch of type i.
Patch density (ha−1)PDPD=NA10,000100Number of patches in the landscape area. N is the total number of patches in the landscape.
Largest patch index (%)LPILPI=MaxaianA×100Percentage of area of the patch with the largest area of the type i to total landscape area.
Aggregation index (%)AIAI=i=1ngiimaxgiiPi100The number of like adjacencies involving the corresponding class, divided by the maximum possible number. gii is the number of joins between pixels of type i. Pi is the proportion of the landscape comprised of patch type i.
Mean patch size (ha)AREA_MNAREA_MN=j=1naijnMean patch size of type i.
Area weighted mean shape indexSHAPE_AMSHAPE_AM=j=1n0.25pijaijaijj=1naijArea-weighted shape complexity of type i. pij is the perimeter of jth patch of type i.
Average building height (m)AHAH=i=1nHinThe average height of building. Hi is the height of the building i. n is the number of building.
Sky view factorSVFSVF=1i=1nsin2βiαi360°The ratio of visible sky area to the hemisphere. n is the total number of the angle elements of the obstacles in the hemisphere environment. αi and βi are the elevation and azimuth angles of the angle element i, respectively.
Average building height standard deviation (m)AHSDAHSD=i=1nHiAH2nVariation degree of the building height. n is the number of building.
Floor area ratioFARFAR=i=1nc×FAThe maximum floor area of a property that one can construct on a plot of land. c is the number of floors; F is the land area of building taken.
Mean architecture projection area (m2)MAPAMAPA=TAPAnThe mean area of architecture projected vertically to floor. TAPA is the total architecture projection area. n is the number of building.
Average volume (m3)AVAV=i=1nVinVi is the building volume of the building i. n is the number of building.

Note: we calculated these seven 2D metrics for UGS and IS, respectively. In the following contents, abbreviations like PLAND_IS means the percentage of landscape of impervious surface.

2.3.2. XGBoost regression model

XGBoost model is an optimized gradient boost tree model, which combines weak regression trees into previous models to correct the residuals in the predictions (Friedman et al., 2000), and thereby achieve higher prediction accuracy and higher computational efficiency than other tree structure-based machines learning models (e.g., RF and SVM). Recently, XGBoost has been increasingly used to detect the driving factors of LST (Sun et al., 2019; Zhou et al., 2022a) because it uses the sparsity-aware split-finding approach to train on sparse data (Sheridan et al., 2016). Due to low sampling nature of field data, the training method is valuable for processing the remote sensing data (Yu et al., 2020).We performed the XGBoost model by Python package “xgboost”.

In this study, a 10-fold cross-validation grid-search method was used to acquire the optimal form of the XGBoost model (Chen et al., 2019; Liu et al., 2020a). The analytical units were randomly split into 3/4 for training and 1/4 for validating, and the accuracy of the model was evaluated with explained variances. We referred to the previous studies (Wu et al., 2022; Yu et al., 2020) and set the model parameters as follows:

At the block scale, the number of estimators was set as from 2000 to 4000 with an interval of 100, while that was set as from 8000 to 10,000 with an interval of 100 at the grid scale. Besides, the other model parameters were same at two scales, including learning rate (0.005), maximum depth of a tree (5 and 6), minimum sum of the weight of all observation in a child tree (from 1 to 5 with an interval of 2), sampling rate for each tree (0.7), sampling rate for columns of each tree (0.6 and 0.8), and gamma (from 0 to 0.3 with an interval of 0.1). The explained variances in cross-validation of final best model are 60.2% and 59.4% at the block scale and the grid scale, respectively.

2.3.3. SHAP method

This study adopted the SHAP method (Lundberg and Lee, 2017) and the best XGBoost model to interpret the relationships between urban morphological metrics and LST. The SHAP method can solve the “black-box” problem (Lundberg et al., 2020) to determine the relative contributions of urban morphological metrics on LST variations. The SHAP method is a local interpretation method derived from the cooperative game theory, which calculates the Shapely values (Shapley, 1952; Štrumbelj and Kononenko, 2014) to represent the importance of each explained variable and examine the effects of each variable on the model. The Shapely value is calculated as follows (Lundberg et al., 2018):(1)ϕi=SM\iS!MS1!M!fSixSifSxSwhere Φi is the contribution of the variable i, the M denotes the set of all input variables, f represents the prediction function model, S represents the set of all input variables excluding i, and x represents the values' vector of a certain instance. The positive SHAP value means the warming effect of potential drivers on LST, while negative SHAP value represents the cooling effect.

The SHAP interaction values were calculated to interpret the interaction effect of the metrics pairs with strong correlation. The SHAP interaction values characterize the interaction effects as the part of combined effects of a specific explained variables pairs that cannot be explained by the sum of alone effects of the two variables, which is a prevailing way to capture and interpret the interaction effects adopted by other methods, like the generalized additive models or Friedman's statistics (Friedman and Popescu, 2008). In principle, this study replaced the external variable subset S to calculate the weighted average δi,j(S) to obtain the interaction values of feature i and feature j. δi,j(S) can be described as the combined effects of two variables minus the sum of the separate effects of two variables with external variable subset S according to eq. (2–3).(2)ϕi,j=SM\ijS!MS2!2M1!δi,jS(3)δi,jS=fSijxSijfSixSifSjxSj+fSxS

In this study, we applied the summary plot to show an overview of metrics by summing the SHAP value over all samples. The metrics were sorted by the feature importance, indicating the contribution of metrics to LST and the general relationship. The dependence plot was adopted to provide more details by visualizing the contribution of a specific metric (Wu et al., 2022). Here, we used SHAP values and SHAP interaction values to interpret the impact of metrics on LST. The SHAP interaction values capture the vertical dispersion that is caused by the interaction between the metrics pairs ranging from low (blue) to high (red) in the dependence plot. The dependence plot using the SHAP interaction values separates the interesting hidden relationships between the metrics pairs compared to the dependence plot using the SHAP values.

3. Results

3.1. Contributions of urban morphological metrics on LST

The contributions of urban morphological metrics on LST at two scales follow the rank of 3D building metrics >2D landscape metrics of IS >2D landscape metrics of UGS, which illustrate that the 3D metrics have more influence on LST than 2D metrics (Fig. 3, Fig. 4). Especially, at the block scale, the total contributions of 3D metrics, 2D landscape metrics of IS, and 2D landscape metrics of UGS are 44.6%, 32.3%, and 23.1%, respectively. The most influential metric on LST is AH (Fig. 3), which is negatively correlated with LST, indicating that LST decreases as AH increases. Followed by MAPA, PLAND_IS, AI_UGS, and AHSD. Among them, MAPA and PLAND_IS have a significantly positive correlation with LST, while AI_UGS and AHSD have significantly negative correlations with LST. Besides, despite the low contribution of SVF to LST, SVF is negatively correlated with LST.

Fig. 3
  1. Download : Download high-res image (513KB)
  2. Download : Download full-size image

Fig. 3. SHAP summary plot of urban morphological metrics at the block scale. (a) Global feature importance. (b) Local explanation. The left graph represents the global importance of metrics on LST variation, ranked from top to bottom according to their contribution to LST variation. Correspondingly, the right graph represents the local explanation of metrics on LST variation. The positive SHAP value in (b) stands for the warming effect of metrics. In contrast, a negative SHAP value stands for the cooling effect of metrics. The feature value is represented by color, which ranges from low (blue) and high (red), as shown in the colored bar on the right side. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 4
  1. Download : Download high-res image (569KB)
  2. Download : Download full-size image

Fig. 4. SHAP summary plot of urban morphological metrics at the grid scale. (a) Global feature importance. (b) Local explanation.

At the grid scale (Fig. 4), the total contributions of 3D building metrics, 2D landscape metrics of IS, and 2D landscape metrics of UGS are 59.2%, 28.3%, and 12.5%, respectively. PLAND_IS, FAR, MAPA, and SVF are positively correlated with LST variations, while AV and AH are negatively correlated. In addition, we found that AH, MAPA, and PLAND_IS among the top eight metrics at two scales have significant contributions to LST variations. The 3D metrics show more significant effects on LST variations than that of 2D metrics, and the contributions of 2D landscape metrics of IS are larger than that of 2D landscape metrics of UGS.

3.2. Responses of LST to urban morphological metrics

According to the results of the SHAP summary plot of 2D and 3D metrics at two scales, we selected the first eight metrics and PLAND_UGS to deeply examine the responses of LST to these nine metrics. The total contributions of these first eight metrics to LST variations are >70% and 81% at the block scale and the grid scale, respectively. Besides, although the contribution of PLAND_UGS is small, it is important for alleviating the UHI effect (Yao et al., 2020). The dependence plot using SHAP values of these nine metrics is shown in Fig. 5.

Fig. 5
  1. Download : Download high-res image (1MB)
  2. Download : Download full-size image

Fig. 5. The dependence plot using the SHAP value for the nine most important metrics (i.e., the value on the X-axis) at two scales. (a) and (b) are the responses of LST to six same metrics at the grid and block scales, respectively. (c) shows the response of LST to other metrics at the grid scale (c1-c3) and the block scale (c4-c6). The value on the Y-axis represents the effect of the metric (i.e., the value on the X-axis) on LST. Here, the LST is the dependent variable and the value on the X-axis is the independent variable. The color of the dots represents the value of the moderator metric, which stands for the effect of the independent variable on the dependent variable depending on the other variable (i.e., moderator metric) (Jaccard and Turrisi, 2003; Zhou et al., 2022a). The units are presented in Table 1.

The general response trend of LST to urban morphology metrics are similar at block and grid scales (Fig. 5a and b), with close and specific thresholds of the warming and cooling. Specifically, AH shows a significant cooling effect when it reaches 19 m. A significant warming effect of PLAND_IS on LST was observed when PLAND_IS >49%. Increasing FAR shows a concave warming effect. AHSD shows a cooling effect when AHSD >15 m at the block scale, while it cannot determine specific thresholds of warming or cooling effects at the grid scale. We also found that there is a significant cooling effect when the proportion of UGS is larger than 48% at the grid scale. In addition, increasing values of AV and SVF show obvious cooling and warming effects at the grid scale, respectively (Fig. 5c1 and c2). Noteworthily, AI_UGS of each unit at the block scale exceeds 76% (Fig. 5c4), indicating that more concentrated green space could provide a significant cooling effect.

3.3. Interactions between 2D/3D urban morphological metrics

The dependence plot using SHAP values shows the response of LST to the most influential metrics, while the SHAP interaction value captures the strong interactions of the metrics pairs. According to the variation in the Y-axis (SHAP value in Fig. 5) and colored dots (i.e., the value of the moderator metric), we selected one metrics pair (i.e., AI_UGS-AH) at the block scale (Fig. 6); five metrics pairs at the grid scale (Fig. 7), including PLANSD_IS-SVF, AV-FAR, AH-FAR, SVF-FAR, and PLAND_UGS-AH, to investigate the response of LST to the interaction effect of the urban morphology metrics. The SHAP interaction values illustrate that the interaction of 3D-3D metrics pairs is stronger than 2D-3D metrics pairs.

Fig. 6
  1. Download : Download high-res image (207KB)
  2. Download : Download full-size image

Fig. 6. The dependence plot using the SHAP interaction value of AH-AI_UGS at the block scale. The units are presented in Table 1.

Fig. 7
  1. Download : Download high-res image (440KB)
  2. Download : Download full-size image

Fig. 7. The dependence plot using the SHAP interaction value for the metrics pairs with strong interactions at the grid scale. The units are presented in Table 1.

At the block scale, the impact of AH on LST depends on AI_UGS (i.e., moderator metric) (Fig. 6). It is worth noting that AI_UGS is larger than 87% at all blocks in our study area, which implies that UGS patches are highly compact (i.e., pixels are highly connected) with small opportunity to exchange heat with other patches at the block scale. For the blocks with a relatively low agglomeration of UGS (i.e., AI_UGS < 91%), an increase in the AH brings a reduction of the interaction value up to 0.1 when AH is larger than 20 m, implying that the AI_UGS may promote AH's main effect (i.e., cooling), while when AI_UGS is larger than 91%, the situation is converse with a greater increase up to 0.2.

At the grid scale (Fig. 7), there is a significant interaction effect of SVF-FAR (Fig. 7a) than that of other 3D metrics. When FAR <1.5, the interaction of SVF-FAR increases, which brings the cooling effect. Noteworthily, the increase of SVF (> 0.8) brings the warming effect when FAR is 1.5 to 2.5. In terms of the AH-FAR (Fig. 7b) and AV-FAR (Fig. 7c), the strong interaction could provide a significant cooling effect. The interaction of PLAND_IS-SVF (Fig. 7d) and PLAND_UGS-AH (Fig. 7e) show a clearly different pattern. When SVF < 0.75, the interaction effect between PALND_IS and SVF decreases as PLAND_IS gets larger. The opposite is true if SVF > 0.75. For PLAND_UGS-AH, when AH > 25 m, the synergy of high-rise buildings and high UGS coverage (> 50%) provides an additional cooling effect, while the increase of PLAND_UGS can bring a warming effect when AH < 25 m.

4. Discussion

4.1. The relationship between urban morphology and LST

In this study, we found that the contributions of urban morphological metrics on LST at two scales followed the rank of 3D building metrics >2D landscape metrics of IS >2D landscape metrics of UGS (Fig. 3, Fig. 4), which illustrated that the 3D metrics have more influence on LST than 2D metrics. Previous studies in other cities also support this result (Alavipanah et al., 2018). However, our result is different from the finding by Chen et al. (2022b), which suggested that the variances of LST are more governed by 2D building metrics in summer (2009) rather than 3D building metrics. We think that the small number of tall buildings (proportion) in the year 2009 may influence the contribution of 3D morphonology characteristics related to LST. The divergent result implies that the distribution in the input data may affect the modeling result for the abovementioned effects of urban morphology. Besides, the relationship between 3D urban parameters and LST may be different due to different scan angles, geographies, and meteorological conditions. For example, the relationship between 3D urban parameters and LST is not stronger at the block scale in Germany cities (Berger et al., 2017). Nonetheless, the 3D urban parameters consistently show significant impact than some of the most widely-used 2D urban parameters (Berger et al., 2017). Hence, there is great significance for a broader and consistent understanding of the relationship between 2D/3D urban morphology and LST across diverse cities worldwide in future works.

In addition, SVF shows an opposite trend at two scales, i.e., SVF is negatively correlated with LST at the block scale (Fig. 3), while it is positively correlated at the grid scale (Fig. 4 and Fig. 5c2). Generally, there is a positive correlation between SVF and LST, because the region with high SVF will potentially receive more solar radiation (Deng et al., 2021; Kim et al., 2022). However, Yin et al. (2018) found that LST is negatively correlated with SVF (Pearson's r = −0.085, p < 0.05) at the block scale (i.e., an irregular grid), which is consistent with our result. Besides, the relationship between SVF and LST is nonlinear (Chen et al., 2012; Guo et al., 2016). Medium SVF values produce the coolest LST, whereas the maximal and minimal SVF values yield the highest LST. Rather than the well-known explanation, i.e., the increase in SVF improves the ventilation in the block (Grimmond, 2007), we observed that the reduction might be more linked to how we delineate the block. We found that the blocks with SVF approaching 0.9 are mostly the low-rise compact area, while those with SVF near 1 are mostly the open area with a small proportion of built-up land. In this case, the transition from low-rise and compact area to the open area may efficiently dissipate the heat, rather than improving ventilation.

4.2. The impact of interactions between urban morphology metrics on LST

Based on the SHAP interaction value, we found that AH shows a significant cooling effect when UGS is relatively scattered than the average condition at the block scale (Fig. 6). The increase in AH mainly increases the casted shadow area (Yu et al., 2019), if the casted area originally has a higher LST, the casted shadow may have a greater cooling effect. At the grid scale, although SVF is positively correlated with LST (Fig. 4), the high SVF and low FAR bring a significant cooling effect (Fig. 7a). One reason is the uncertainty introduced by the absence of vegetation's height during calculating SVF. The high proportion of vegetation in regions with high SVF and low FAR generally shows a cold island. Besides, the increase in SVF will increase both the sun-exposed surface in the urban canyon (i.e., heating) (Oke, 1973) and the wind speed (i.e., cooling) (Liu et al., 2020c). Our results indicate that the relative importance of the cooling effect of ventilation may increase with less volume of buildings at the grid scale. In terms of the interaction between PLAND_IS and SVF, the low SVF and increasing PLAND_IS will bring significant cooling effect (Fig. 7d), which is related to the shading effect that brings from high-rise buildings. Regarding the interaction between PLAND_UGS and AH, the results show that PLAND_UGS have a greater cooling effect when the AH is greater than the average condition (Fig. 7e). The increase in AH brings not only the additional casted shadow area but also strengthens the air circulation at the flanks of buildings (Oke et al., 2017), thereby providing an effective cooling effect. Moreover, there is usually a greater distance between residential buildings than the mid- and low-rise zones. Our results may indicate that the extra cooling effects brought by strengthened air circulation are larger than the cooling effect brought by the increased shadow area. It is worth noting that the interaction between AH and FAR significant increases with increasing AH regarding a high FAR condition, which is also observed between AV and FAR. The similar interaction patterns (i.e., AH-FAR and AV-FAR) indicate that the configuration of high-rise buildings with small volumes at a relatively high FAR condition has the potential to mitigate the UHI effect, which benefits from shadows of taller buildings, as well as less heat accumulation by small buildings (Zhang et al., 2022).

4.3. Implications for mitigating the UHI effect

This study reveals that the 3D urban morphology plays a more critical role in shaping the urban thermal environment than the 2D urban morphology, which is confirmed at the regular (200 m × 200 m) and irregular (blocks) scales. The design of the mean building height should be considered for urban development projects and the implementation of urban sustainability measures. For the blocks with a mean building height < 19 m, the uncompact urban greening may not obtain the expected cooling effect. Meanwhile, a low value of FAR is also beneficial to cooling. In terms of mid- and high-rise building regions, it is critical to design urban ventilation to facilitate heat dissipation. Although low SVF will prevent the urban surface from absorbing more solar radiation during the day, the longwave radiation will be trapped due to the compact buildings and tree canopies (i.e., low SVF) at night, causing LST to rise. Hence, providing sufficient street openness for wind penetration is beneficial to cooling the region with a compact layout. Besides, the spatial pattern of IS should also be considered. We found that the percentage of IS in a specific block should be <49% at least, otherwise, LST will rise. Meanwhile, the planning of IS should take the region's SVF into account. When the region with IS >49%, increasing the tree canopies (i.e., low SVF and high percentage of UGS) can provide an effective cooling. Hence, it is suggested that rationally arranges urban green spaces to break the agglomeration of the impervious surface (or reducing building density). From the perspective of the configuration of specific metrics pair, taller buildings with small building volumes mitigate heat accumulation, which is conducive to lowering LST. Therefore, it is advisable to increase the buildings' height and decrease the building volumes in regions with high FAR, as well as restrict to the construction of low-rise buildings, which increases the spacing and air circulation between buildings, thereby reducing surface temperature values.

4.4. Limitations and future works

There are some limitations that need more concern in the future works. First, because the current 3D characteristics of UGS are mainly calculated by airborne LiDAR data that is hardly obtained, this study only considered the 2D characteristics of UGS. Given that the 3D characteristics of UGS also significantly affect the thermal environment (Chen et al., 2022a; Kong et al., 2022), comprehensive consideration of 3D characteristics of building and vegetation in future work is needed. Second, the directions of effects of some urban morphological metrics on LST are different at the block and the grid scale, e.g., SVF and AI_UGS. The scale dependency should be considered to deepen the understanding of the impacts of these metrics on LST (Shi et al., 2021). Moreover, we found some uncertainties of interaction effects, e.g., the metrics pairs of SVF and FAR. Because SVF is correlated with urban ventilation (Hernandez et al., 2015), which can directly or indirectly contribute to the heat exchange at urban canopy layer to alleviate local UHI and improve the thermal comfort (Yang et al., 2021b). Hence, the ventilation metrics (e.g., frontal area index) should be considered to deeply explore the interaction of urban morphological metrics. Finally, meteorological factors (e.g., wind speed and relative humidity) also play important roles in regulating LST variation, more investigation needs to be conducted at a fine scale.

5. Conclusions

In this study, we explored the effects of 2D and 3D urban morphological metrics on LST using the XGBoost model and the SHAP method from the perspective of the irregular block scale and the regular grid scale. Specifically, we quantified the response of selected 2D and 3D metrics on LST and the interaction between these urban morphology metrics. This study illustrated that the 3D metrics show more significant effects on LST variations than 2D metrics, especially the mean building height and SVF should be considered in urban planning and management for mitigating the UHI effect. The general impacts and thresholds of warming or cooling effects of urban metrics on LST are similar at the irregular block scale and the regular grid scale. The increase in IS provides warming effects. Nevertheless, the rational configuration of IS and UGS is still useful for cooling regarding the regions with IS >49%. Finally, we found that the interactions of 3D-3D metrics pairs on LST are stronger than those of 2D-3D metrics pairs. The interaction between 3D-3D metrics pairs generally shows a synergy effect (e.g., AH-FAR and AV-FAR), which is beneficial to mitigate the UHI effect. Besides, the interaction of 2D-3D metrics pairs depends on the specific thresholds, which may be the synergy effect (e.g., PLAND_IS-SVF) or antagonistic effect (e.g., PLAND_UGS-AH). These results revealed that the interactions of urban morphology are complex regarding the regulation of the local thermal environment, more investigation should be conducted in the future to understand the driving mechanisms of the complex multi-dimensional urban morphology on the thermal environment, which will guide the development of more sustainable and resilient urban systems.

CRediT authorship contribution statement

Bo Yuan: Conceptualization, Methodology, Writing – original draft. Liang Zhou: Supervision, Writing – original draft, Writing – review & editing, Funding acquisition. Fengning Hu: Methodology, Writing – review & editing. Chunzhu Wei: Writing – review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.


This work has been funded by the National Key Research and Development Program of China (No. 2022YFC3800700), the National Natural Science Foundation of China (No. 42271214, 41961027), the Foundation of Key Projects of Natural Science of Gansu Province (No. 21JR7RA278, 21JR7RA281), and the Basic research top talent plan of Lanzhou Jiaotong University (2022JC01).

Data availability

Data will be made available on request.


Cited by (1)

View Abstract