Introduction

The United Nations predicts that by 2050, about 70% of the global population will inhabit metropolitan regions1. This fast urbanization, driven by population expansion, is significantly altering metropolitan climates. A notable consequence of this transition is the urban heat island (UHI) effect2, wherein urban areas exhibit markedly elevated temperatures compared to adjacent rural regions3,4, rendering it a defining feature of the urban climate. This phenomenon significantly impacts human health and the environment, leading to elevated risks of heat-related ailments and fatalities3, diminished thermal comfort4, and increased energy consumption5,6.

Land Surface Temperature (LST) serves as a crucial metric for evaluating the UHI effect7,8, as it accurately reflects spatial temperature discrepancies shaped by landscape characteristics, thus elucidating urban heat exposure patterns9,10. The urban thermal environment is influenced by various elements, with urban greenery identified as a significant method for alleviating the UHI effect11,12. Vegetation, encompassing park greenspaces and street trees, mitigates LST via shading and evapotranspiration while concurrently improving ecological quality and fostering public health13,14,15. Two principal methodologies exist for objectively assessing the quantity, distribution, and accessibility of urban green spaces16. The initial technique uses high-resolution satellite imagery for remote sensing to evaluate the cooling effects of vegetation canopies. This is accomplished by analyzing the correlation between the Normalized Difference Vegetation Index (NDVI) and LST17,18,19. Unlike simple land cover classification methods that focus only on the extent of green space, NDVI reflects vegetation density, allowing for a more nuanced and continuous assessment of the relationship between vegetation and LST, and enabling more precise analysis of urban cooling effects20. The second way employs growing street view imagery (SVI) technology, which has seen heightened popularity due to the advancement of platforms like Google, Baidu, and Tencent5,21,22,23. Compared to satellite imagery, which predominantly captures tree canopies from an aerial perspective, SVI facilitates the identification of lawns, shrubs, and green walls beneath the canopy layer, providing a hemispherical viewpoint that more accurately represents human perceptions of the environment24,25. By deriving vegetation data from SVI, researchers can acquire a human-centered comprehension of urban vegetation shape and accessibility, overcoming the constraints of conventional remote sensing in detecting fine-scale greenery. The Green View Index (GVI) is a widely utilized metric for evaluating urban green areas, precisely measuring vegetation visibility from a direct line of sight26,27. Utilizing semantic segmentation approaches, GVI may be calculated from SVI to precisely identify shrubs and green walls, thus enhancing data on street-level vegetation coverage24.

Furthermore, in addition to urban greenery, the arrangement of buildings and the design of roadways are crucial in shaping the UHI effect. Prior research has shown that two-dimensional elements, such as urban blue-green areas and population density, are essential for mitigating urban heat28. Nonetheless, due to the diversity of urban settings, the influence of urban spatial morphology, especially three-dimensional elements, must be acknowledged. Semi-enclosed areas combined with dense vegetation have demonstrated efficacy in mitigating elevated summer temperatures29. The constructed environment affects the local microclimate by altering evapotranspiration and wind flow dynamics30,31, while tall buildings can impede heat dissipation, intensify elevated temperatures, and raise energy consumption32.

In 2023, Guangzhou secured the fourth position nationally in terms of gross domestic product (GDP) and permanent population33,34. The UHI effect has inexorably impacted the metropolis as a result of population growth, urban expansion, increased building density, and less green space. Multiple research has examined the UHI effect in Guangzhou35,36,37. Notwithstanding the substantial volume of research, there remains a continuing deficiency in comparative assessments that integrate vegetation data derived from satellite imagery and street view imagery, with the objective of evaluating their differential influences on LST. The disparities in the influence of NDVI and GVI on LST are not yet well comprehended. This study focuses on Guangzhou as the research region and analyzes the varying effects of NDVI and GVI on LST. Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), and Multiscale Geographically Weighted Regression (MGWR) models are used to comprehensively assess the influence of urban morphology variables, encompassing both street-level and architectural characteristics on LST. This study systematically assesses the combined impact of these variables and NDVI on the UHI effect.

Materials and methods

Study area

Situated in southern China, Guangzhou experiences a subtropical monsoon climate characterized by persistently elevated temperatures and significant precipitation year-round. It serves as a prime example of the nation’s humid and oppressive regions. The city comprises 11 administrative districts, with a resident population of 18.827 million in 2023. Global warming and growing urbanization have led to an increased frequency of extreme heat occurrences, resulting in a notable rise in high-temperature days. In 2023, the annual average temperature approached 23.7 °C, reaching a documented peak of 39.7 °C. Since 1961, the frequency of high-temperature days has increased by 6.8 days per decade38. The UHI impact has strengthened since the late 1980s, with average UHI intensity of 1.1–1.4 °C since 2009 and peak values of up to 2.0 °C recorded in high-density areas such as Yuexiu, Liwan, western Haizhu, and southern Tianhe39. The diverse urban environment and irregular allocation of green spaces in Guangzhou establish an optimal natural laboratory for investigating the impact of vegetation on LST across various scales. This research concentrates on the center metropolitan region of Guangzhou (Fig. 1), encompassing seven districts—Liwan, Haizhu, Yuexiu, Panyu, Baiyun, Tianhe, and Huangpu—totaling approximately 1989 km2.

Fig. 1
figure 1

Visualization of the study area. Map generated using ArcGIS Pro 3.2 (Esri Inc., https://www.esri.com).

Full size image

Method

The research framework is shown in Fig. 2. Firstly, the gathered data include Landsat 8 TIRS/OLI satellite imagery, street view photographs of Guangzhou’s primary urban region, and a building dataset. Secondly, LST at a 1 km grid resolution was obtained for the research area throughout 2023 utilizing Landsat 8 satellite images. Thirdly, twelve urban morphology indicators influencing LST were computed based on the spatial vegetation index and building information. A correlation analysis was performed on 13 contributing parameters, including NDVI, to detect and address multicollinearity concerns. Fourthly, NDVI and GVI, which denote urban greenery derived from satellite and street view data, respectively, were evaluated independently to assess their statistical and spatial variations in affecting LST. Fifthly, we employed OLS regression, GWR, and MGWR models to investigate the relationships between urban morphology indices and LST. Their prediction accuracy was subsequently evaluated to select the optimal model, from which the coefficients of numerous components were investigated to ascertain their positive or negative associations with LST. The influence of urban environmental characteristics on LST during the year was assessed, resulting in policy suggestions for sustainable urban development.

Fig. 2
figure 2

Research framework.

Full size image

Calculation of LST

This study obtained LST via the Statistical Mono-Window (SMW) algorithm40, executed on the Google Earth Engine (GEE) platform, to analyze Landsat 8 images from 2023. The resulting LST represents an annual average condition, offering a temporally integrated and representative view of urban thermal environments. Created by the Climate Monitoring Satellite Application Facility, the SMW algorithm estimates LST by developing a pragmatic link between Top-of-Atmosphere (TOA) thermal radiation (\(Tb\)) and surface emissivity (\(\epsilon\)), expressed as:

$$LST= {A}_{i}\frac{Tb}{\epsilon }+ {B}_{i}\frac{1}{\epsilon }+ {C}_{i}$$
(1)

where \({A}_{i}\)​, \({B}_{i}\), and \({C}_{i}\) are coefficients obtained from radiative transfer simulations for various Total Column Water Vapor (TCWV) classes.The LST retrieval technique entailed the preparation of TOA brightness temperatures from Landsat 8’s thermal infrared bands and Surface Reflectance (SR) data to calculate NDVI, which was employed to estimate the Fraction Vegetation Cover (FVC) and surface emissivity. NDVI was computed as follows:

$$NDVI= \frac{{\rho }_{NIR}- {\rho }_{Red}}{{\rho }_{NIR}+ {\rho }_{Red}}$$
(2)

where \({\rho }_{NIR}\)​ signifies near-infrared and \({\rho }_{Red}\) denotes red band reflectance. Surface emissivity (\(\epsilon\)) was adjusted using the Vegetation-Cover method:

$$\epsilon =FVC \cdot {\epsilon }_{veg}+\left(1-FVC\right) \cdot {\epsilon }_{bare}$$
(3)

where \({\epsilon }_{veg}\)​ and \({\epsilon }_{bare}\)​ represent emissivity values for fully vegetated and bare ground surfaces, respectively. Cloud masking was implemented utilizing the QA_PIXEL band to reduce cloud contamination.

LST values were aggregated to a 1000-m grid using zonal averaging in Google Earth Engine. This was implemented through the reduceRegions function, which computed the mean LST value within each 1 km × 1 km grid cell based on original 30-m pixels.

Architecture of the semantic segmentation network

This study used semantic segmentation with the Pyramid Scene Parsing Network (PSPNet) to detect urban green spaces and analyze diverse environmental attributes in street-level imagery. The model’s hierarchical methodology facilitates precise pixel-level classification, guaranteeing correct delineation of landscape components within intricate metropolitan environments. PSPNet, an image segmentation method grounded in deep learning, proficiently identifies and categorizes diverse items inside an image41,42. PSPNet enhances the Fully Convolutional Network (FCN) architecture by incorporating a pyramid pooling module for the extraction of multi-scale context. The model comprises three primary components: (1) a ResNet-based backbone for feature extraction, (2) a pyramid pooling module for the aggregation of global and local spatial information, and (3) a concluding classification layer for pixel-wise segmentation. PSPNet employs dilated convolutions to retain intricate spatial features while ensuring computational efficiency. This structure facilitates more accurate segmentation of vegetation in varied urban areas due to the intricacy of natural roadway settings.

A systematic strategy for data gathering and preprocessing was adopted to provide complete spatial coverage. We commenced by acquiring road network data for the research area from OpenStreetMap (OSM) (https://www.openstreetmap.org/), a prevalent open-access mapping platform. Subsequently, we designated sampling points at exact intervals of 50 m. The study area was segmented into 1 km × 1 km grids, yielding 1356 distinct grid units for spatially explicit analysis. Utilizing the Baidu Maps application programming interface (https://quanjing.baidu.com/), we assembled a comprehensive dataset of 1,287,100 street view photos depicting the principal metropolitan areas of Guangzhou, with images sourced from 2017 to 2019. PSPNet was subsequently applied for semantic segmentation to classify image pixels into categories such as vegetation and other urban features. The segmentation results were employed to measure the spatial distribution of these factors throughout the study region, offering essential insights for urban environmental evaluation and planning. Although the street view data predates the satellite imagery, it remains valid for spatial comparison purposes. This is because major urban structures and large-scale vegetation patterns (e.g., tree-lined streets, green corridors) in central Guangzhou have remained relatively stable over this period. Thus, despite the temporal discrepancy, the dataset still provides a reliable basis for spatial pattern analysis.

Calculation of urban morphology indicators

This section examines the calculation of urban morphological indicators. These metrics assess the spatial attributes of urban environments and are essential for analyzing their influence on LST. Chen et al. (2023) classified urban morphology indicators into street view indicators and building form indicators43. This study utilizes this approach to choose six street view indicators (Table 1) and six building form indicators (Table 2). Street view indicators, obtained from panoramic photos, delineate the visual composition of urban streetscapes, encompassing vegetation covering (GVI), constructed edifices (WALL), transportation surfaces (ROAD, SWALK), and the openness of street canyons (SVF, BVF). Building form indicators were derived from the dataset supplied by Che et al.44. These indices quantify the density and volumetric characteristics of urban structures, encompassing spatial intensity (BD, FAR), variation in building height (BH, BH_SD), and three-dimensional morphology (BV, BV_SD). These indicators offer a thorough framework for comprehending how urban morphology influences LST, hence aiding research on urban heat reduction techniques.

Table 1 Selected street view indicators in the study.
Full size table
Table 2 Selected architectural form indicators in the study.
Full size table

Spatial autocorrelation analysis (Moran’s I)

Spatial autocorrelation evaluates the distribution patterns of geographic features both globally and locally45. Global spatial autocorrelation quantifies the extent to which similar attribute values aggregate or disperse within a geographic region, commonly assessed through Mora’s I. The index ranges from − 1 to 1: a score nearing 1 indicates a robust positive spatial correlation, whereas values close to − 1 suggest a negative spatial configuration. A value near zero signifies randomness, lacking any identifiable spatial trend. The standardized Z-score is employed to evaluate significance; a |Z| value of more than 1.96 indicates substantial spatial autocorrelation at the 95% confidence level46.

Conversely, local spatial autocorrelation examines spatial interactions at a more detailed level, detecting clusters or dispersed patterns across spatial units and their adjacent areas. Utilizing local Moran’s I value, spatial units can be classified into five distinct categories: High-High clusters (regions with elevated values adjacent to other high-value areas), High-Low clusters (high-value pockets encircled by lower-value zones), Low–Low clusters (areas with persistently low values neighboring similar regions), Low–High clusters (low-value zones adjacent to high-value areas), and regions demonstrating no significant spatial correlation47. This approach offers a more comprehensive view of spatial heterogeneity and clustering attributes within the study area. In this study, spatial units were categorized into the five cluster types based on the sign and magnitude of local Moran’s I values along with their corresponding Z-scores and pseudo p-values. Only units with statistical significance at the 95% confidence level (p < 0.05) were assigned to the High-High, High-Low, Low-Low, or Low–High categories. Units that did not meet the significance threshold were labeled as “Not significant.” This classification ensures that only statistically meaningful spatial clusters were included in the analysis.

Multiple regression models

Ordinary Least Squares (OLS) regression is the conventional analytical method for investigating the correlation between LST and urban morphology indicators28. This method presupposes that the effect of predictor factors is consistent throughout the entire geographical region being studied. The OLS model can be expressed as:

$$y=X\beta + \varepsilon$$
(4)

where \(y\) represents the dependent variable, \(X\) comprises the matrix of independent variables, \(\beta\) constitutes the regression parameter vector, and \(\varepsilon\) signifies the error component. Despite its computational simplicity, OLS neglects geographical variability, a significant constraint in urban climate research. Previous studies have shown that metropolitan areas display significant geographic variability in LST patterns, requiring more spatially adaptive regression methods48.

Unlike OLS, Geographically Weighted Regression (GWR) accommodates spatial variability by allowing regression coefficients to vary based on location49. By prioritizing proximate data points, GWR adeptly reveals localized correlations between urban landscape attributes and LST. The mathematical representation of GWR can be expressed as:

$${y}_{i}= {\beta }_{0}\left({u}_{i}, {v}_{i}\right)+ \sum_{k=1}^{m}{\beta }_{k}\left({u}_{i}, {v}_{i}\right) {x}_{ik}+ {\varepsilon }_{i}$$
(5)

where \(\left({u}_{i}, {v}_{i}\right)\) denotes the coordinates of spatial position i, with \({\beta }_{k}\left({u}_{i}, {v}_{i}\right)\) representing the spatially adaptive regression coefficients. However, a significant limitation of GWR is its assumption that all predictor variables influence LST equally over the same spatial scale, which may not effectively represent the intricate multiscale interactions in urban settings28.

The improved method, Multiscale Geographically Weighted Regression (MGWR), addresses this constraint by allowing each independent variable to operate at its optimal spatial level, hence improving model accuracy and interpretability50. MGWR generates variable-specific bandwidths through an iterative back-fitting process that minimizes the corrected Akaike Information Criterion (AICc). This approach enables each predictor to reflect the spatial heterogeneity of its influence on LST. Such adaptability allows MGWR to more effectively capture both localized and global effects than conventional GWR, resulting in enhanced model performance and more nuanced insights into the links between indicators and LST. The MGWR model is expressed as:

$${y}_{i}= {\beta }_{0}\left({u}_{i}, {v}_{i}\right)+ \sum_{k=1}^{m}{\beta }_{k}\left({u}_{i}, {v}_{i}, {bw}_{k}\right) {x}_{ik}+ {\varepsilon }_{i}$$
(6)

where \({bw}_{k}\)​ represents the optimal bandwidth for each explanatory indicator k, facilitating a more precise examination of vegetation indices and urban morphology impacts on LST.

The models’ performance was rigorously assessed using several measures, including the coefficient of determination (R2), the adjusted AICc, and the residual sum of squares (RSS). These measurements offered a comprehensive view of the model’s strengths and weaknesses51. Typically, when the AICc difference between two models is under 3, their fitting performance is regarded as comparable. A model with a lower AICc score is favored for its superior balance between goodness of fit and complexity52. A reduced RSS indicates a closer correspondence between the model and the actual data, signifying enhanced accuracy53. All calculations and model evaluations were conducted utilizing ArcGIS Pro.

Results

Spatial distribution patterns of LST

Figure 3a depicts the distribution patterns of LST over the primary urban region of Guangzhou in 2023, indicating a thermal range from 27.61 to 41.99 °C. The city displayed a distinct temperature gradient, with elevated temperatures prevailing in the central-western regions and colder conditions in the northeast and southeast. This eastward thermal shift underscores significant urban heat inequalities throughout the primary urban region. This trend illustrates the UHI effect when temperatures progressively diminish from the urban center to the outskirts. Haizhu, Liwan, and Yuexiu displayed the largest percentages of high-temperature zones, while Baiyun, Tianhe, and Huangpu registered the lowest.

Fig. 3
figure 3

(a) Distribution of land surface temperature (LST) in the research area for 2023; (b) local spatial autocorrelation of LST at a 1 km grid size. Map generated using ArcGIS Pro 3.2 (Esri Inc., https://www.esri.com).

Full size image

The spatial autocorrelation analysis yielded a Moran’s I value of 0.96, accompanied by a z-score of 122.11 and a p-value below 0.01, indicating a significant clustering tendency in LST distribution. Figure 3b illustrates that the local analysis revealed concentrated high-value clusters in the western regions and notable low-value clusters in the northeastern zone, indicating significant positive spatial dependency throughout the research area.

Comparative analysis of the effects of NDVI and GVI on LST

Before statistical analysis, both NDVI and GVI were normalized to guarantee comparability on the same dimensional scale. To assess whether the differences between NDVI and GVI were statistically significant, we conducted Mann–Whitney U tests for each district, given that the two datasets represent independent samples. Table 3 indicates that all seven districts exhibited statistically significant differences between NDVI and GVI (p < 0.1), revealing that the two indicators capture different dimensions of urban greenness. Figure 4 illustrates the changes in the statistical distributions of NDVI and GVI across the seven districts. NDVI ranges from 0.12 to 0.64, and GVI varies between 0.18 and 0.47. This signifies that the extent of NDVI variations surpasses that of GVI, with GVI demonstrating lesser variability across several districts. Moreover, in smaller districts like Liwan, Haizhu, and Yuexiu, the median NDVI is inferior to the median GVI. Inversely, in larger districts such as Panyu, Baiyun, Tianhe, and Huangpu, the median NDVI surpasses the median GVI. The median GVI of Yuexiu district (0.3701) surpassed its NDVI (0.3097) by approximately 20%. By contrast, the median GVI in Tianhe district (0.3839) was 17% lower than the median NDVI (0.4653). Tianhe District demonstrates the most variation in both NDVI and GVI among the seven districts, whilst Liwan District exhibits the least variation. Comparing NDVI and GVI within a single district illustrates the disparities in urban greening levels from two separate vantage points: satellite imagery and street view imagery.

Table 3 Statistical test results comparing NDVI and GVI in each district.
Full size table
Fig. 4
figure 4

Statistical disparities in NDVI and GVI across seven districts.

Full size image

LST, NDVI, and GVI were categorized into lower and higher ranges in relation to the entire study region. Lower ranges were delineated as LST below 36 °C, NDVI below 0.48, and GVI below 0.21, whereas higher ranges were characterized by LST above 36 °C, NDVI above 0.48, and GVI above 0.21. Significantly, these data were not standardized to preserve the original geographical distribution patterns of LST with respect to NDVI and GVI values.

Figure 5a illustrates the typological map of LST and NDVI in Guangzhou, demonstrating clear spatial trends. The western sectors of the primary urban zone displayed LST levels surpassing 36 °C, along with NDVI values beneath 0.48, signifying an inverse relationship between urban vegetation cover and LST. This pattern was especially evident in the majority of Liwan, Haizhu, and Yuexiu districts, corresponding with the overall spatial distribution of LST. Comparable patterns were also noted in extensive areas of Baiyun district and the southwestern regions of Panyu district. Conversely, regions with LST below 36 °C and NDVI exceeding 0.48—specifically the southeastern sections of Baiyun and the northeastern areas of Tianhe—further illustrated the cooling effect associated with increased vegetation coverage. Furthermore, certain regions had either low LST and NDVI or high LST and NDVI concurrently, indicating that variables beyond plant cover may exert a more significant influence on LST in specific areas.

Fig. 5
figure 5

Typology map of LST and NDVI/GVI for Guangzhou, China in 2023. Map generated using ArcGIS Pro 3.2 (Esri Inc., https://www.esri.com).

Full size image

A distinct spatial pattern arises when analyzing the correlation between LST and GVI, as depicted in Fig. 5b. Compared to NDVI, the spatial correlation between LST and GVI appears more fragmented and less consistent. In places where LST surpassed 36 °C, and NDVI fell below 0.48, GVI values exhibited significant variability; certain areas (e.g., Liwan and Yuexiu districts) recorded values exceeding 0.21, but others (e.g., Baiyun and Haizhu districts) remained beyond this threshold. This mismatch highlights the constraints of satellite imagery in depicting fine-scale urban greenery, whereas street-level imagery offers more localized and precise vegetation data. In locations characterized by lower LST and elevated NDVI, GVI values showed significant diversity, with certain places exhibiting greater values. In contrast, others remained comparatively lower, especially in the southeastern sections of the Baiyun district.

Factors contributing to LST

Multiple regression models

Prior to conducting regression models, correlation analysis was performed among all indicators (Fig. 6), revealing that most coefficients had absolute values below 0.5, indicating low multicollinearity between variables. Table 4 presents the results of the OLS, GWR, and MGWR models. The OLS regression exhibited the least efficacy among the three models, as its inflexible linear structure inadequately addressed the complex, location-specific differences in the correlation between predictor variables and LST. On the contrary, GWR demonstrated substantial enhancement in model fit relative to OLS, chiefly because it integrated spatial weighting components, which improved the model’s capacity to address local fluctuations and offered nonlinear spatial modeling capabilities. Nonetheless, GWR’s efficacy was constrained by its presumption of homogeneous geographical scales across all metrics. The MGWR model surpassed both OLS and GWR with an R2 of 0.727. Despite the slight enhancement in R2 compared to GWR, the considerable decreases in AICc and RSS values signified a superior model fit. These data demonstrate that MGWR offers a more accurate representation of the relationship between LST and urban morphology indicators, successfully identifying local spatial non-stationarity.

Fig. 6
figure 6

Correlation heatmap.

Full size image
Table 4 The results of OLS, GWR, and MGWR models.
Full size table

Figure 7a,b present the local R2 values for the GWR and MGWR model, respectively. Unlike the OLS model, both GWR and MGWR demonstrated regionally heterogeneous R2 patterns. The northern and northwestern sections of the research area had the most robust performance in elucidating LST, whereas areas with elevated local R2 values were also present in the south. Nonetheless, predicted accuracy declined in the western, southeastern, and central urban areas. The MGWR model demonstrated a broader and more spatially extensive distribution of local R2 values than GWR, including bigger regions, despite a slightly lower maximum local R2 value. The broader spatial scope highlights MGWR’s enhanced ability to consider localized differences in the interactions between LST and its influencing elements.

Fig. 7
figure 7

Distribution of local R2 values for both the (a) GWR and (b) MGWR models. Map generated using ArcGIS Pro 3.2 (Esri Inc., https://www.esri.com).

Full size image

Spatial variation of regression coefficients

The MGWR model findings demonstrated significant geographical non-stationarity in the impact of urban morphology indicators on LST, underscoring differences in their effects throughout the research area (Table 5). BD exhibited the most robust positive correlation with LST, with a mean value of 0.287, underscoring its substantial role in temperature elevation. Concurrently, BV demonstrated a favorable effect but less pronounced, with a mean value of 0.056. In contrast, markers such as FAR (− 0.130), GVI (− 0.158), SVF (− 0.173), BVF (− 0.161), and NDVI (− 0.262) had negative mean regression coefficients, indicating their cooling influence on LST. Significantly, BH (− 0.042), BH_SD (− 0.012), WALL (− 0.029), ROAD (− 0.051), and SWALK (− 0.032) demonstrated comparatively low absolute mean values (< 0.1), signifying spatial instability and diminished impacts on LST. Notably, BV_SD showed a near-zero mean regression coefficient (0.000) and a very small standard deviation (0.001), indicating that spatial variation in building volume has a negligible direct effect on LST in the study area. This result may reflect the relatively homogeneous distribution of building volumes within most 1 km × 1 km grid units, leading to limited explanatory power at the scale of analysis. The fluctuations in bandwidth indicate that various indicators function at different spatial scales, highlighting the need for a multiscale approach to more effectively understand urban morphological impacts on thermal environments.

Table 5 Regression relationship between indicators and LST together with optimal bandwidths produced by MGWR.
Full size table

Figure 8 illustrates the delineated regions within the MGWR model exhibiting substantial coefficient fluctuations of urban morphology indicators, with colored areas indicating statistically significant estimates. Among the indices of building form, BD showed consistently robust positive associations throughout the research area (Fig. 8a). BH displayed its most pronounced cooling effect in the center urban zone, with negligible variance in the eastern and southern parts (Fig. 8c). FAR had favorable relationships with LST in center areas while displaying cooling effects in outer sections (Fig. 8b). BV exhibited weak negative correlations in center locations and more robust positive correlations in adjacent parts (Fig. 8e). BH_SD had positive correlations in the center and western regions (Fig. 8d), whereas BV_SD revealed positive correlations in the southern region and negative correlations in the northern region (Fig. 8f).

Fig. 8
figure 8

Spatial distribution of coefficients for urban morphology indicators utilizing the MGWR model. Map generated using ArcGIS Pro 3.2 (Esri Inc., https://www.esri.com).

Full size image

Concerning street view indicators, both GVI and SVF consistently exhibited negative relationships with LST across the study region (Fig. 8g,k). Conversely, BVF exhibited minor regions of positive connection with LST in the eastern and southern zones (Fig. 8l). WALL, ROAD, and SWALK demonstrated a highly uniform spatial distribution of positive and negative associations (Fig. 8h,i,j). Furthermore, NDVI exhibited positive associations in major urban locales while demonstrating more pronounced negative correlations in adjacent locations (Fig. 8m).

Discussion

Comparative examination of the distinct impacts of NDVI and GVI on LST

In the western section of the study area, a clear inverse correlation between NDVI and LST is evident, displaying a high-LST, low-NDVI pattern. At the same time, central urban regions demonstrate a low-LST, high-NDVI pattern, particularly in the southeastern part of Baiyun and the northeastern part of Tianhe. The prevalent occurrence of both low-LST, low-NDVI, and high-LST, as well as high-NDVI patterns in other central regions, indicates that variables beyond vegetation cover significantly influence LST. Relative to NDVI, the spatial link between LST and GVI seems more disjointed. In regions where LST surpasses 36 °C, and NDVI is below 0.48, GVI demonstrates significant variation, creating various distribution patterns associated with LST. For example, GVI values are comparatively elevated in the Liwan and Yuexiu districts, whereas diminished values are noted in the Baiyun and Haizhu districts. A comparable differentiation pattern is also observable in regions marked by reduced LST and elevated NDVI.

While NDVI and GVI demonstrate a uniform spatial pattern in their effect on LST, the disparities between the two indices suggest that regions with elevated NDVI values do not necessarily align with correspondingly high GVI values. Prior research indicates that inconsistencies in urban settings frequently arise from complex multi-tiered transit networks, increased construction density, and the reflecting characteristics of certain building materials54. Generally, regions next to parks and green spaces exhibit elevated GVI values55, underscoring nuanced variations in urban greening relative to varying degrees of vegetation covering. Regions exhibiting elevated GVI values are typically distinguished by the prevalence of large-canopy trees, which confer significant cooling effects via transpiration and shadowing. The cooling effect of vegetation mostly arises from its shade and evapotranspiration mechanisms29,56. Conversely, NDVI’s impact on LST is predominantly limited to a two-dimensional spatial cooling effect. Researchers investigating the UHI effect encounter obstacles when employing satellite-derived LST data57, including resolution constraints that compromise accuracy and challenges in identifying vegetation within intricate urban environments58. The influence of GVI highlights the significance of three-dimensional biological environments in reducing LST, stressing the additional benefits of street-level imagery in documenting intricate environmental specifics, thus overcoming the shortcomings of satellite-based evaluations.

The influence of urban morphology on LST

The correlation between urban morphology indicators and LST was examined at various spatial scales utilizing the OLS, GWR, and MGWR models. The findings suggest that the MGWR model offers the optimal match, corroborating previous research conclusions59. The swift growth of Guangzhou’s metropolitan region has significantly influenced the spatial variability of the urban thermal environment60,61. This extension extends beyond the primary urban region, affecting nearby districts where metropolitan cores have had significant growth. Unlike the conventional OLS model, the MGWR model refines the basic GWR framework by optimizing bandwidth parameters and enhancing coefficient calculations to represent regionally variable connections between predictors and LST more accurately. This sophisticated methodology overcomes the constraints of spatial stationarity present in conventional models55. This methodology enables various explanatory variables to function at multiple spatial scales, thus enhancing model efficacy.

Multiple indicators, such as BH, BH_SD, BV_SD, WALL, ROAD, and SWALK, demonstrate spatial non-stationarity and have a comparatively minimal impact on LST. Among them, parameters linked to building height (BH and BH_SD) do not demonstrate a monotonic influence on LST. Moderate-height buildings may elevate LST, but higher edifices in urban centers often obstruct more solar radiation and offer enhanced shading, therefore diminishing LST59. Furthermore, WALL, ROAD, and SWALK constitute essential elements of the street environment. The paving materials commonly utilized in urban streets—mainly asphalt and concrete—tend to retain heat and resist water, significantly contributing to increasing metropolitan temperatures. These surfaces absorb sunlight extensively while permitting minimal drainage, so transforming streets into heat reservoirs that exacerbate the UHI effect43,62.

Both BV and BD exhibit a favorable connection with LST. Increased individual building volumes and elevated urban building density diminish ventilation effectiveness, promoting heat retention and restricting evaporation, which results in consistently high LST levels63. In the study region, the overall building density is elevated, and BD regularly exhibits a robust positive connection with LST. However, BV in the central region exhibits a little negative association. This region, marked by sophisticated urban economic growth, features numerous large-scale high-rise structures that cast shadows, thereby diminishing direct solar radiation and lowering daytime LST. The center region features extensive green infrastructure, including rooftop and vertical greening, urban parks, and efficient energy systems, which somewhat mitigate the adverse impacts of building thermal inertia. Considering the identified positive correlations and in accordance with Guangzhou’s urban development strategies, alterations in urban village redevelopment may entail the demolition of specific structures and the implementation of high-rise, narrow-width, and low-density architectural designs to mitigate the warming impacts of densely populated urban blocks. Moreover, in regions with high LSTs, the cultivation of shade-providing trees is a more efficacious cooling approach than the growth of grasslands.

FAR, SVF, and BVF exhibit an inverse correlation with LST. The cooling effect of an elevated FAR can be ascribed to multiple factors: (1) the augmented obstruction rate promotes shadow formation, consequently reducing solar radiation absorption by heat-retentive materials such as asphalt and concrete64; (2) the urban canyon effect enhances airflow over road surfaces, aiding in heat dissipation; and (3) a higher FAR encourages the creation of public spaces and ventilation corridors. The cumulative effects facilitate the cooling impact of FAR on LST, a finding corroborated by prior research65. An elevated SVF signifies a greater availability of open spaces in densely constructed urban environments, facilitating air circulation in street canyons and encouraging heat dissipation. Nonetheless, heightened exposure to direct solar radiation in open areas may lead to elevated LST values48. The overall impact of SVF on LST is determined by the equilibrium between these two conflicting effects. In contrast to other research43,66, our results indicate that the impact of air exchange is more pronounced than that of solar radiation. The impact of BVF on LST is predominantly manifested in the heightened intensity of horizontal and vertical structures adjacent to highways as BVF values grow. Increased building density results in a bigger shadowed area on the ground, leading to a localized decrease in LST.

Limitations and further research

This research has certain limitations. The collected dataset comprises Landsat 8 TIRS/OLI satellite images, from which LST and NDVI raster data at a resolution of 1 km were obtained through spatial aggregation in GEE. Although these data provide useful insights, the spatial resolution may be inadequate for comprehensive assessments of urban streets and buildings, requiring additional refinement. The temporal span of the investigation necessitates enhancement. The dependence on 2023 satellite and street view pictures may insufficiently represent long-term urban thermal dynamics, highlighting the necessity for more comprehensive time-series data. Vegetation coverage is a key factor in mitigating urban LSTs, and recent findings suggest that its spatial configuration also plays a critical role67. Guangzhou’s swift development has resulted in considerable spatial variability in the thermal environment, indicating that numerous uninvestigated elements may also affect urban heat distributions. Water bodies, characterized by their significant heat retention and poor reflection, have demonstrated considerable cooling effects68. Furthermore, socio-economic factors such as population density, economic activity, and traffic flow are crucial for a comprehensive understanding of the UHI effect69. Future studies should consider these parameters to enhance the precision and applicability of urban thermal environment evaluations.

Conclusions

This study examined the influence of urban green spaces on the UHI effect in Guangzhou by using satellite-derived NDVI and street-level GVI. The findings indicated that: (1) A significant west-to-east temperature gradient was detected, with elevated temperatures centered in the central metropolitan region, demonstrating severe spatial clustering of heat. Significant regional disparities were seen between NDVI and GVI, with NDVI predominantly indicating extensive vegetation covering, whereas GVI emphasizes localized cooling impacts of greenery. This highlights the need to examine vegetation from both vertical and horizontal viewpoints. The regression analysis indicated that the MGWR model offers improved performance over OLS and GWR in capturing the geographical variation of LST. Urban morphology indicators demonstrated geographically heterogeneous effects, with BV and BD positively linked with LST. However, the FAR, GVI, SVF, BVF, and NDVI revealed negative associations. These findings underscore the imperative to enhance urban green spaces and constructed environments to alleviate the UHI effect. Future studies should utilize higher-resolution datasets and include supplementary environmental variables, such as water bodies and socio-economic aspects, to improve urban heat regulation systems.