Revealing the Impact of Urban Morphology Evolution on the Urban Heat Island Effect in the Main Urban Area of Guangzhou: Insights from Local Climate Zones

Abstract

Local Climate Zones (LCZs) provide a critical framework for analyzing how urban morphology influences the surface urban heat island (SUHI) effect. However, the spatiotemporal heterogeneity of the driving mechanisms of urban morphology in SUHI within LCZs under urban renewal remains insufficiently understood. In this study, estimated building heights for 2018, 2021, and 2024 in the main urban area of Guangzhou were used to generate LCZ maps using GIS-based methods. Land surface temperatures (LSTs) were retrieved to quantity the SUHI effect. The Geographical-XGBoost (G-XGBoost) model was applied to evaluate the impacts of urban morphology on SUHI. The results indicated the following: (1) Building height estimation errors range from 5.92 to 7.03 m, and incorporating building height data into LCZ classification enabled sensitive detection of urban evolution dynamics. (2) Built LCZ types accounted for the majority of the study area. Between 2018 and 2024, LCZ 3 decreased markedly, by 9.57%, and land cover LCZ types declined annually to 21.35%. (3) Low-level SUHII was predominant, while the proportion of high and extremely high levels of SUHII initially rose before declining to 16.62%. LCZ 2 and LCZ 3 exhibited the highest SUHII. (4) Pervious surface fraction (PSF) is generally regarded as the most important explanatory factor across LCZ types; however, LCZ 4 represents an exception where its importance significantly decreases. This study reveals the nonlinear impacts of urban morphology evolution on SUHI under the effect of the interaction between LCZs and urban renewal, supporting efforts to optimize urban microclimates and promote sustainable development.

1. Introduction

The persistent deterioration of urban thermal environments poses a significant challenge to global urban sustainability. The surface urban heat island (SUHI) effect, one of its most prominent manifestations [1,2], profoundly impacts urban ecological and socio-economic systems [3]. SUHI negatively affects urban energy consumption and CO₂ emissions [4], public health [5], and vulnerability to natural disasters [6]. In the context of climate change, the increasing frequency of extreme heat events amplifies thermal risks by exacerbating the combined effects of SUHI and elevated temperatures [7,8]. Consequently, SUHI represents a critical issue necessitating urgent attention in urban planning and management [9]. Therefore, a comprehensive understanding of its driving mechanisms and mitigation strategies is vital for enhancing urban resilience and promoting climate-adaptive development.

The development of SUHI is largely attributed to the rapid expansion of urban areas and population growth. These processes have been accompanied by spatiotemporal changes in land use/land cover (LULC), which have altered the physical and chemical properties of the land surface and modified urban climatic characteristics [10]. The spatiotemporal dynamics of LULC have been found to influence land surface temperature (LST) through changes in surface albedo [11], variations in evapotranspiration processes [12], and modifications in heat storage capacity and urban morphology [13]. For example, urban expansion between 1990 and 2021 increased the proportion of built-up land from 18.8% to 43.2% in Bhubaneswar, thereby contributing to a rise in LST [14]. Another study, conducted in Jakarta, reported that the conversion of green spaces to built-up areas increased air temperature by approximately 0.24 °C [15]. To address the growing urban thermal risk, a range of urban management strategies have been implemented to mitigate SUHI [16].

To evaluate the impacts of management strategies and LULC dynamics, a systematic framework is needed to characterize the thermal variations induced by urban landscapes [17,18]. In this context, the Local Climate Zone (LCZ) classification system has emerged as an effective approach, accounting for differences in urban surface structure and physical properties [19]. LCZs categorize urban and land cover types into 10 built types and seven land cover types based on uniform surface cover characteristics [20], enabling the integration of urban microclimate and spatial morphology [21] and offering a standardized approach for SUHI research [22]. However, the original LCZ classification, largely based on climatic parameters, does not fully account for the morphological diversity of individual cities [23]. This limitation has led to the development of supervised classification workflows that incorporate local urban forms and expert knowledge [24]. To address this issue, the World Urban Database and Access Portal Tools (WUDAPT) project proposed a generalized workflow for LCZ classification [25]. Generating LCZ maps based on WUDAPT’s Level 0 remote sensing data products combined with machine learning methods such as random forests has become a mainstream approach in recent studies [26,27].

Despite its utility, optical remote sensing-based LCZ mapping is heavily dependent on the quality of the training sample [28], limiting its ability to represent complex urban spatial structures [29]. An accuracy assessment of global-scale LCZ mapping shows that the classification accuracy based on the WUDAPT method is at an intermediate level, with only 50–60% classification accuracy [30]. Moreover, satellite imagery does not capture accurate building height information [31,32], a key feature for distinguishing built LCZ types. In over 20 Chinese cities, LCZ maps using WUDAPT methods performed poorly due to the absence of building height data [33]. In contrast, GIS-based methods that incorporate real 3D urban morphological data provide superior classification accuracy and finer detail [34,35]. For instance, in high-density cities like Hong Kong, GIS-based LCZ maps significantly outperform WUDAPT products [36,37]. However, in most regions, building height data are not readily available and often exhibit spatial discontinuities [38,39]. Recent studies have leveraged machine learning models—including random forests [40], convolutional neural networks [41], and semantic segmentation [42] combined with high-resolution remote sensing data [43,44] to generate fine-scale building height estimates. These methods have markedly improved the accuracy and efficiency of height estimation [45,46], enabling the production of multi-temporal LCZ maps [47].

As stable and relatively independent microclimate units, LCZs offer a systematic framework for analyzing the complex interactions between urban morphology and SUHI [48,49]. Recent studies have used spatial and statistical techniques to explore spatiotemporal variations in surface urban heat island intensity (SUHII) across LCZ types [50]. Findings indicate substantial variability in SUHI patterns across cities [51,52,53], driven by the interplay between urban morphology and background climate [54], prompting cross-city [19,55] and cross-climate [56,57] comparisons. In addition, studies identifying diurnal [58,59], seasonal [22,60], and interannual [61,62] variations in SUHII indicate that the SUHI–LCZ relationship exhibits temporal dynamics [63]. Notably, intra-LCZ temperature differences may even exceed inter-LCZ differences, underscoring the critical role of urban morphology in shaping LCZ-specific thermal behaviors [64,65]. Although many studies have examined the impact of urban morphology on SUHI [66], the effect of a given morphological indicator may vary across LCZ types [67]. Moreover, urban renewal continuously alters LCZ types and reshapes the influence of key drivers over time. However, temporal analyses of intra-LCZ thermal response mechanisms remain limited.

Given the nonlinear and spatially heterogeneous nature of SUHI drivers [68], their analysis remains a major challenge. Traditional linear models [69,70] and correlation analysis [71,72] often overlook spatial heterogeneity, leading to reduced prediction accuracy. To address this, multiscale geographically weighted regression (MGWR) has been introduced to disentangle spatial scales of influence across variables [73]. However, MGWR remains confined to linear assumptions [74] and cannot fully capture the nonlinear dynamics of SUHI. The recently developed Geographical eXtreme Gradient Boosting (G-XGBoost) model enhances the XGBoost algorithm by incorporating spatial kernel functions to generate localized models [75], which are then integrated with global models. This approach offers robust capabilities for modeling nonlinear, spatially heterogeneous SUHI drivers.

To address these research gaps, this study employs the LCZ framework and G-XGBoost model to investigate the impact of urban morphology evolution on SUHI in the main urban area of Guangzhou. The specific objectives are as follows: (1) to generate building height datasets for 2018, 2021, and 2024 using deep learning techniques and create corresponding LCZ maps; (2) to analyze the spatiotemporal variation of SUHII within different LCZs; (3) to assess the spatiotemporal heterogeneity in the effects of driving factors on SUHI across LCZs. The results will offer scientific evidence for developing fine-scale urban heat mitigation strategies and optimizing spatial planning in Guangzhou, while also serving as a reference for other cities pursuing sustainable thermal environment governance.

2. Materials and Methods

2.1. Study Area

Guangzhou, as the capital of Guangdong Province and one of the core cities of the Guangdong–Hong Kong–Macao Greater Bay Area, is located at 112°57′E–114°03′E and 22°26′N–23°56′N. According to the Guangzhou City Master Plan (2017–2035), the main urban area of Guangzhou covers Liwan, Yuexiu, Tianhe, and Haizhu districts, as well as areas south of the Baiyun North Second Ring Expressway in Huangpu District, and north of the Guangming Expressway in Panyu District, as depicted in Figure 1. Since the reform and opening-up, Guangzhou’s main urban area has undergone rapid urban expansion [76]. Under the influence of a subtropical monsoon climate, the expansion has made high-temperature areas more concentrated and temperatures rise [77]. Moreover, the uneven intensity of expansion and the mixed distribution of diverse building types have resulted in a highly intricate urban texture, complicating the mechanisms through which urban morphology influences the thermal environment. Given the ongoing expansion in Guangzhou’s urban core, it is of practical significance to explore the evolution of its urban morphology and its impact on the urban thermal environment, particularly for the city’s future planning and sustainable development.

2.2. Data Sources and Preprocessing

2.2.1. OpenStreetMap Road Data

Urban blocks, defined as enclosed units formed by road networks, exhibit relatively homogeneous urban morphology and thermal environment characteristics within each unit, making them suitable as basic research units [78]. In this study, road centerlines were extracted from OpenStreetMap Road data (OSM) through morphological operations such as dilation and erosion [79]. Major roads were then used to delineate urban blocks in the main urban area of Guangzhou. Based on the research scale and characteristics of the road network, the initial block segmentation was manually refined with the aid of satellite imagery, resulting in a total of 4763 urban blocks.

2.2.2. Sentinel-1/2

To estimate building height for the years 2018, 2021, and 2024, this study employed multi-source remote sensing data from the Sentinel-1 and Sentinel-2 satellites. Sentinel-2 provides 10 m resolution multispectral imagery covering the red, green, blue, and near-infrared (NIR) bands, which offers rich spectral and textural information [80,81]. These optical images are useful for extracting building footprints and identifying shadow variations [44], thereby providing indirect cues for height estimation. However, optical imagery is often affected by weather conditions such as cloud cover, resulting in data gaps in certain regions or time periods. As a result, this study primarily used Sentinel-2 data for building footprint extraction, while the estimation of building heights relied on Sentinel-1 imagery. Sentinel-1 Ground Range Detected (GRD) Level-1 products provide VV and VH polarized radar backscatter data at 10 m resolution [82], with capability for all-weather, day-and-night imaging [83]. The backscatter characteristics are highly sensitive to surface material, geometry, and structural features, making SAR data particularly suitable for building height retrieval [84]. It is worth noting that due to the unavailability of Level-2A products in 2018, Sentinel-2 imagery for that year was obtained as Level-1C products from the European Space Agency (ESA) and atmospherically corrected. Data for the other years were acquired and preprocessed on the Google Earth Engine platform.

2.2.3. SRTM V3 Digital Elevation Data

To support the geometric and radiometric correction of SAR imagery and enhance the accuracy of building height estimation, this study incorporates the Version 3 (V3) Digital Elevation Model (DEM) from the Shuttle Radar Topography Mission (SRTM) [85]. The SRTM data were acquired during an 11-day mission in February 2000 using interferometric synthetic aperture radar (InSAR) onboard the Space Shuttle Endeavour. The V3 release improves data quality through void filling and error correction, providing near-global elevation coverage at a spatial resolution of 1 arc-second (approximately 30 m) [86]. In SAR-based applications, DEMs are essential for correcting terrain-induced geometric distortions and normalizing backscatter intensity, particularly in regions with complex topography [87]. Therefore, this study further incorporates DEM data as an auxiliary explanatory variable within the deep learning model. This integration explicitly accounts for terrain-induced variations in radar backscatter, helping to improve the accuracy and robustness of building height estimation in topographically complex areas and compensating for the limitations of conventional correction methods.

2.2.4. Reference Building Height Data

This study also collected three reference datasets to evaluate the accuracy of the estimated building heights: (1) the built-up area building height dataset produced by the Global Human Settlement Layer (GHSL) project [88], (2) the 10 m resolution building height product of China (CNBH-10m) [40], and (3) the China Multi-attribute Building dataset (CMAB) [89], which is the first national-scale dataset of individual buildings with multiple attributes in China. These datasets correspond to the years 2018, 2020, and 2024, and are used to validate the estimated building height data for 2018, 2021, and 2024, respectively. While these products provide valuable references, they differ in temporal coverage, input data sources, and production methods, which introduce systematic deviations in dense urban contexts like Guangzhou. For this reason, in this study, they are used only as reference datasets.

2.2.5. Landsat 8

Landsat 8 is an Earth observation satellite jointly launched by the National Aeronautics and Space Administration (NASA) and the United States Geological Survey (USGS). It carries two sensors, the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS), which together provide data in 11 spectral bands. Specifically, OLI captures Bands 1–7 and 9, while TIRS collects Bands 10 and 11. The visible, near-infrared, and shortwave infrared bands acquired by OLI have a spatial resolution of 30 m, while the thermal infrared bands from TIRS have a native resolution of 100 m [90]. Among them, the red band (Band 4, 0.64–0.67 μm) and near-infrared band (Band 5, 0.85–0.88 μm) are commonly used to calculate the Normalized Difference Vegetation Index (NDVI) [91], and the thermal infrared Band 10 (10.60–11.19 μm) is widely applied for LST retrieval [92]. Given the frequent cloud cover in the study area during summer, cloud-free Landsat 8 imagery was selected for LST retrieval, and a median composite approach was applied to minimize noise and data gaps caused by atmospheric interference. In addition, annual NDVI at a spatial resolution of 30 m was generated using maximum value composites from all Landsat 8 imagery throughout the year, and then aggregated to 100 m resolution by averaging.

2.2.6. CLCD Data

To support the classification of LCZs in this study, the China Land Cover Dataset (CLCD) was introduced. The CLCD is derived from time-series remote sensing imagery from the Landsat and Sentinel satellite series and is generated using machine learning classification algorithms [93]. It provides annual land cover maps of China at a spatial resolution of 30 m from 1985 to 2023. Given the recent slowdown in land use change within the study area, the 2023 CLCD data were used to represent the land cover pattern of 2024, which is considered reasonable. For the years 2018 and 2021, the corresponding CLCD data from those specific years were adopted.

2.3. Method

Figure 2 illustrates the overall methodological framework of this study, which consists of the following four steps: (1) estimating building height in the main urban area of Guangzhou for the years 2018, 2021, and 2024 using the Simultaneous building Height And Footprint extraction from Sentinel imagery (SHAFTS) model and multi-source satellite data, followed by the calculation of urban morphology indicators; (2) retrieving LST from Landsat-8 imagery; (3) generating LCZ maps based on the computed morphology indicators and CLCD land cover data; and (4) applying the G-XGBoost model to quantify the contribution of urban morphology indicators to SUHI across different LCZ types.

2.3.1. Building Height Estimation Using the SHAFTS Model

To address the issue of spatiotemporal discontinuity in building height data, this study adopts SHAFTS model [94] to estimate building heights in the main urban area of Guangzhou. SHAFTS is a deep-learning framework implemented in PyTorch (version 1.12.1), which employs a multi-task deep learning (MTDL) strategy to simultaneously predict building heights and footprints. The core of SHAFTS is a CNN with a SENet backbone, integrating residual connections and Squeeze-and-Excitation (SE) blocks to enhance hierarchical feature extraction and channel-wise attention. Two fully connected (FC) layers are used in the head for final predictions, with ReLU activation for building height and sigmoid for building footprint prediction.

Specifically, for the 100 m resolution case used in this study, the CNN uses an initial convolution layer with a 3 × 3 kernel and stride 1, followed by three SE layers, each containing one SE block with a reduction ratio of 16. The stride of the first SE block in the first SE layer is set to 1. After separate feature extraction from Sentinel imagery and SRTM elevation data, the two branches are concatenated for downstream prediction. The model is trained using the Adam optimizer with a learning rate of 0.01, weight decay of 0.0001, and batch size of 256, following a cosine annealing schedule. The source code of SHAFTS is openly available at https://github.com/LllC-mmd/3DBuildingInfoMap (accessed on 17 October 2023) [94]. Sentinel-1 SAR, Sentinel-2 optical imagery, and SRTM elevation data from the years 2018, 2021, and 2024 were used as inputs to predict building heights and footprints at a 100 m spatial resolution for the main urban area of Guangzhou.

To ensure the applicability and predictive accuracy of the SHAFTS model in the study area, it is essential to validate the estimated building height. Root Mean Square Error (RMSE), a commonly used metric for evaluating the accuracy of regression models, is employed in this study to quantify the deviation between predicted and reference building height. According to the year of data production, the GHS-BUILT-H [88], CNBH-10 [40], and CMAB [89] datasets were used to validate the predicted building heights for 2018, 2021, and 2024, respectively. The RMSE is calculated as follows:

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{\sum _{i=1}^n {\left(h_{est,i}-h_{ref,i}\right)}^2}{n}}}$$

(1)

where $h_{est,i}$ represents the estimated building height of pixel $i$, $h_{ref,i}$ is the reference building height for pixel $i$, and $n$ is the total number of pixels.

2.3.2. GIS-Based LCZ Mapping

Based on the original LCZ framework [20], Building Height (BH) (referred to as the height of roughness elements in the original framework), Sky View Factor (SVF), Pervious Surface Fraction (PSF), and Building Surface Fraction (BSF) were found to better characterize urban spatial structure and provide stronger discriminative capacity for identifying built LCZ types compared with other morphological parameters [95]. Therefore, this study introduced the four parameters. The specific definition and calculation methods are provided in

Table 1. Definitions and calculation formulas of parameters for built LCZ types.

Parameter	Definition	Method
Building Height (BH)	BH represents the area-weighted average building height within each grid, where n is the number of buildings in the given urban block, ${BS}_i$ is the ground footprint area of the i-th building, and ${BH}_i$ is its geometric height.	$\mathrm{B}\mathrm{H}={\displaystyle \frac{\sum _{i=1}^n \left({BS}_i\ast {BH}_i\right)}{\sum _{i=1}^n {BS}_i}}$
Building Surface Fraction (BSF)	BSF denotes the building surface fraction within a given urban block, where n is the number of buildings in the block, ${BS}_i$ is the ground footprint area of the i-th building, and $S_{site}$ is the total area of the block.	$\mathrm{B}\mathrm{S}\mathrm{F}={\displaystyle \frac{\sum _{i=1}^n {BS}_i}{S_{site }}}$
Sky View Factor (SVF)	SVF represents the degree of openness within an urban block. In this study, SVF was calculated using the Horizon package in R, where α is the azimuth angle of rotation, and β is the elevation angle from the block center to the surrounding building walls.	$\mathrm{S}\mathrm{V}\mathrm{F}=1-n_{i=1}^n{\mathrm{s}\mathrm{i}\mathrm{n}}^2{\beta }_i\times \left[{\displaystyle \frac{\alpha }{360}}\right]$
Pervious Surface Fraction (PSF)	PSF is defined as the proportion of pervious surfaces within a given urban block, where $A_{ps}$ is the area of pervious surfaces, identified as regions with an NDVI greater than 0.2, and $A_{total area}$ is the total area of the block.	$\mathrm{P}\mathrm{S}\mathrm{F}={\displaystyle \frac{A_{ps}}{A_{total area }}}$

. Given the complexity of the study area, particularly the presence of urban villages within the main urban region, the original classification thresholds are not entirely applicable. To improve local suitability, the classification parameters were adjusted according to the spatial morphological characteristics of Guangzhou. Representative samples were selected from high-resolution satellite imagery, and reference was made to studies focusing on Chinese cities with similar urban characteristics [21,86].

First, BSF was used to distinguish between built and land cover LCZ types. Areas with BSF exceeding 0.1 were classified as built LCZs, while those below this threshold were assigned to initial land cover LCZs [96]. Subsequently, the initial land cover LCZs were refined using CLCD data to derive the final land cover LCZ classifications. In the second step, BSF was further applied to differentiate between compact (BSF ≥ 0.4) and open (BSF < 0.4) built LCZs. BH was then used to classify vertical structures into high-rise, mid-rise, and low-rise categories, and the thresholds applied were consistent with both the original LCZ framework [20] and previous empirical studies [38,52,95]. Finally, SVF and PSF values were calculated from representative samples selected within the pre-classified LCZ units. The final thresholds generally aligned with Zheng et al. [37] and Chen et al. [96], and they were used to refine LCZ assignments and minimize potential misclassifications. It is worth noting that LCZs 7 to 10 were not observed in the study area; therefore, only LCZs 1–6 were retained for built types. The classification thresholds for all LCZ types were locally adjusted based on the actual urban conditions in Guangzhou, with detailed criteria provided in

Table 2. Values of parameters for built LCZ types. All properties are unitless except BH (m).

Built Types	BH	BSF	SVF	PSF
LCZ 1—Compact high-rise	>25	≥0.4	0.1–0.25	<0.1
LCZ 2—Compact mid-rise	10–25	≥0.4	0.25–0.6	<0.2
LCZ 3—Compact low-rise	<10	≥0.4	0.2–0.6	<0.5
LCZ 4—Open high-rise	>25	0.1–0.4	0.4–0.7	0.1–0.7
LCZ 5—Open mid-rise	10–25	0.1–0.4	0.4–0.8	0.2–0.7
LCZ 6—Open low-rise	<10	0.1–0.4	0.6–0.9	0.5–0.7

.

For land cover LCZ types, the original classification was integrated and refined based on the land use characteristics of the main urban area of Guangzhou. LCZ A (dense trees) and LCZ B (scattered trees) were merged into a mixed tree category, LCZ A-B (trees); farmland and herbaceous vegetation were grouped into LCZ D (low plants); and bare land and roads were excluded. As a result, the final land cover LCZ types used in this study included LCZ A-B, LCZ D, and LCZ G (water), aiming to enhance classification accuracy and local adaptability. To evaluate the LCZ classification, 10% of the street blocks—excluding those used as representative samples—were randomly selected as validation samples for computing the overall accuracy (OA) [96,97].

2.3.3. LST Inversion

In this study, the Statistical Mono-Window (SMW) algorithm [98] was first implemented on the Google Earth Engine (GEE) platform to retrieve LST from Landsat-8 imagery for the summers of 2018, 2021, and 2024, and the results were subsequently composited using a median approach. The SMW algorithm estimates LST by establishing an empirical relationship between Top-of-Atmosphere Brightness Temperature (TOA), brightness temperature, and LST, based on a linearization of the radiative transfer equation. Specifically, the algorithm utilizes the single thermal infrared (TIR) channel to establish the following linear regression equation:

$$\mathrm{L}\mathrm{S}\mathrm{T}=A_i{\displaystyle \frac{T_b}{\epsilon }}+B_i{\displaystyle \frac{1}{q}}+C_i$$

(2)

where $T_b$ is the TOA brightness temperature in the TIR channel, $\epsilon$ is the surface emissivity of that channel, and the algorithm coefficients $A_i$, $B_i$, and $C_i$ are derived from radiative transfer simulations. The simulation process stratifies total column water vapor (TCWV) into 10 classes (0–6 cm with a 0.6 cm interval), with TCWV values exceeding 6 cm assigned to the final class.

To correct for atmospheric effects, the algorithm integrates TCWV data from the NCEP/NCAR reanalysis [99], which is accessible on GEE and provides temporal coverage consistent with the entire operational period of the Landsat series.

2.3.4. SUHII Calculation and Classification

Due to the susceptibility of land surface temperature to weather fluctuations, the SUHII is commonly used to characterize the urban thermal environment. SUHII is typically defined as the temperature difference between urban and suburban areas [62]. In LCZ-based thermal environment analysis, SUHII is calculated by subtracting the mean LST of LCZ D from the LST of each LCZ types [100]. This approach effectively reduces the influence of background climatic noise and highlights the independent effects of urban morphology on the thermal environment. Based on the block-level LCZ classification results, the LST values were extracted for each block, and the average LST of LCZ D was computed. The SUHII of a given block was then calculated using the following equation:

$$SUHI{I}_i={\overline{LST}}_i-{\overline{LST}}_{\mathrm{LCZ} \mathrm{D}}$$

(3)

where $SUHI{I}_i$ denotes the SUHII value of block $i$, ${\overline{LST}}_i$ represents the average LST of the block, and ${\overline{LST}}_{\mathrm{LCZ} \mathrm{D}}$ is the mean LST of all LCZ D blocks.

To minimize the influence of temporal variation and improve the comparability of results across different years, the SUHII values were normalized using the mean–standard deviation classification method. This approach, based on clearly defined statistical thresholds, provides robust classification performance and has been widely adopted in previous studies [101]. The SUHII values were categorized into six levels based on combinations of the mean and standard deviation [102], with detailed classification criteria presented in

Table 3. SUHII classification based on the mean–standard deviation method.

SUHII Level	Classification Thresholds
No SUHII	$SUHI{I}_i$$*<0$
Very low	$0\le SUHI{I}_i<\mu -\sigma$
Low	$\mu -\sigma \le SUHI{I}_i<\mu -0.5\sigma$
Moderate	$\mu -0.5\sigma \le SUHI{I}_i<\mu +0.5\sigma$
High	$\mu +0.5\sigma \le SUHI{I}_i<\mu +\sigma$
Very high	$\mu +\sigma \le SUHI{I}_i$

* $SUHI{I}_i$ denotes the surface urban heat island intensity of block $i$, where $\mu$ is the mean SUHII value and $\sigma$ is the standard deviation.

.

2.3.5. Exploring the Influence of Urban Morphology on SUHI Across Different LCZs

Urban morphology is characterized as a spatiotemporal heterogeneous entity shaped by the interaction between human activities and the natural environment. This interaction is reflected in the spatial configuration and dynamic evolution of physical elements in both two and three dimensions. As a fundamental component of the urban system, buildings exert a substantial influence on urban morphology through their physical characteristics, such as height and volume [103]. This influence results in alterations to the energy balance and microclimate conditions of urban areas. This study integrates two-dimensional layout and three-dimensional characteristics to establish a comprehensive parameter system for urban morphology. In addition to the four parameters used for LCZ classification presented in

Table 4. Definitions of and calculation formulas for supplementary parameters.

Parameter	Definition	Method
Floor Area Ratio (FAR)	FAR is defined as the ratio of the total floor area within a block to its building footprint area. FAR is directly proportional to the number of floors in a building, with taller structures exhibiting higher FAR values. $S_i$ is the base area of a single building, $F_i$ is the number of floors in that building, and $S_A$ is the total plot area.	$\begin{array}{c}\mathrm{FAR}={\displaystyle \frac{\sum _{i=1}^n S_i{F}_i}{S_A}}\end{array}$
Building Roughness (BR)	BR quantifies the degree of irregularity in the micro-scale geometric morphology of building surfaces. $h_i$ is the height of building, $AVH$ is the average building height within the specified unit, and N is the total number of buildings in that unit.	$\mathrm{BR}=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N} \left(h_i-AVH\right)}$
Building Height Coefficient of Variation (BHCV)	BHCV is defined as the ratio of the standard deviation of building height within a unit to its average height, describing the degree of height variation. $HSD$ represents the standard deviation of building heights within the unit.	$\mathrm{BHCV}={\displaystyle \frac{HSD}{AVH}}$
FLUctuation (FLU)	FLU denotes the difference between the maximum and minimum building height within the unit, reflecting the degree of vertical variation among buildings.	$\mathrm{FLU}=H_{max}-H_{min}$

, four supplementary parameters were incorporated to describe the three-dimensional characteristics of buildings [48,104]: Floor Area Ratio (FAR), Building Roughness (BR), Building Height Coefficient of Variation (BHCV), and FLUctuation (FLU). The calculation formulas for these parameters are provided in Table 4.

Given the inherently nonlinear and spatially heterogeneous relationship between urban morphology and SUHI, the G-XGBoost model [75] is employed for analysis. This novel spatial machine learning model enhances the conventional XGBoost algorithm by integrating spatial weights and local regression methods. Such enhancements effectively address spatial heterogeneity and dependency issues.

First, the traditional XGBoost model is referred to as the “global XGBoost model” to distinguish it from the spatial local model. The global XGBoost model is formulated as follows:

$${\widehat{y}}_{i-gl}={\widehat{y}}_i=\sum _{k=1}^K f_k\left(x_i\right),f_k\in \mathcal{F}$$

(4)

where ${\widehat{y}}_{i-gl}$ is the predicted value for the observation $i$, $K$ is the number of trees, $f_k$ is a function within the function space $\mathcal{F}$, and $\mathcal{F}$ is the set of all possible classification and regression trees (CART).

Second, during the construction of the G-XGBoost model, spatial weights are computed using a bi-square kernel function, and local models are trained on data subsets defined by a specified bandwidth.

$$w{s}_{ij}=\left\{\begin{array}{cc}\left[1-{\left({\displaystyle \frac{d_{ij}}{b}}\right)}^2\right]& \mathrm{if} e_{i-\mathrm{l}\mathrm{o}\mathrm{c}}>e_{i-gl}\\ 0& \mathrm{otherwise} \end{array}\right.$$

(5)

where $d_{ij}$ is the Euclidean distance between the $i$-th and $j$-th points, and $b$ is the bandwidth, with its optimal value determined by minimizing cross-validation criterion.

Based on the above procedures, $n$ local models are trained for each spatial unit, with predictions presented in the formula below:

$${\widehat{y}}_{i-loc}=\sum _{k=1}^K f_k^{\left(u_i,v_i\right)}\left(w_{si},x_{ib}\right),f_k\in \mathcal{F}$$

(6)

where ${\widehat{y}}_{i-loc}$ is the prediction of the local model calibrated at location $i$, ($u_i,v_i$) are the centroid coordinates of the i-th spatial unit, and $w_{si}$ is the spatial weight matrix from center point $i$ to the spatial units $x_{ib}$ within the spatial kernel.

Finally, an ensemble model is generated by combining the predictions of the global and local models through weighted averaging:

$${\widehat{y}}_{i-\mathrm{ens}}={\alpha }_i{\widehat{y}}_{i-\mathrm{loc}}+\left(1-{\alpha }_i\right){\widehat{y}}_{i-gl}$$

(7)

where ${\widehat{y}}_{i-\mathrm{ens}}$ is the prediction of the ensemble model for location $i$, and ${\alpha }_i$ is the hyperparameter that controls spatial dependency, which is dynamically adjusted based on the residual errors of the local and global models.

In this study, eight two-dimensional and three-dimensional urban morphology indicators were used as independent variables, and the SUHII values of different built LCZ types were used as dependent variables. Separate models were established for each LCZ type in each year using G-XGBoost. Furthermore, the influence of each urban morphology indicator on SUHII was assessed by analyzing the feature importance derived from each model.

3. Results

3.1. Building Height Estimation and Accuracy Assessment

Figure 3 illustrates the spatiotemporal variation in building height across the main urban area of Guangzhou from 2018 to 2024, along with changes in the average building height for each administrative district. Overall, mid-rise buildings (10–25 m) dominate, with low-rise buildings (<10 m) primarily concentrated in outlying suburban areas, while high-rise buildings (>25 m) are concentrated in core areas such as the southwest of Tianhe District and the east of Yuexiu District. Notably, the central part of Huangpu District has gradually evolved into a new cluster of high-rise buildings. Statistical results show that the average building height in the main urban area increased from 5.49 m in 2018 to 6.74 m in 2024, exhibiting an overall upward trend. Most administrative districts experienced a steady increase followed by minor fluctuations in building height, with Huangpu District and Tianhe District showing the most significant growth, with building height increasing by 38.4% and 25.9%, respectively, over the six-year period.

Figure 4 shows the correspondence between the estimated building height and the reference data in the form of a scatter density plot. Most sample points are concentrated along the 1:1 diagonal line, indicating that the SHAFTS model exhibits good estimation performance in the study area. Overall, RMSE of building height remains within approximately two floors, demonstrating satisfactory accuracy. Specifically, the RMSE values are 5.92 m for 2018, 6.80 m for 2021, and 7.03 m for 2024.

3.2. Spatiotemporal Evolution of LCZs

Figure 5a illustrates the spatiotemporal distribution of the LCZs at the neighborhood scale in the main urban area of Guangzhou City from 2018 to 2024. The classification results indicate that, between 2018 and 2024, the boundaries between built LCZ types and land cover LCZ types remained relatively stable overall. In compact built LCZ types, LCZ 1 and LCZ 2 are concentrated in Yuexiu District, radiating outward in a ring-like structure that forms a typical center–periphery pattern; LCZ 3 has the widest coverage, extending the urban core boundary southward, northward, and eastward from the core of Yuexiu District. The spatial distribution of open built LCZ types is relatively dispersed, With the units primarily concentrated in the central part of Huangpu District, exhibiting a horizontal clustering trend. In contrast, land cover LCZ types are distributed as follows: LCZ A and LCZ B are primarily located in Baiyun District and the northern part of Huangpu District, while LCZ D is concentrated in the northwestern part of Baiyun District and the eastern and western ends of Panyu District. The OA values for the LCZ maps were 92.49% in 2018, 88.37% in 2021, and 89.43% in 2024, indicating a high level of reliability for subsequent spatiotemporal analyses.

Figure 5b and Figure 5c respectively illustrate the dynamic evolution trends of different LCZ types in terms of quantity and proportion. Throughout the study period, a built LCZ types remained dominant, with their proportion increasing by 2.73%. High-rise LCZ types (LCZ 1 and LCZ 4) had the smallest proportion and exhibited relatively stable changes, with an overall increase of 1.82%. Mid-rise LCZ types (LCZ 2 and LCZ 5) exhibited significant growth, with their combined proportion increasing by 7.41%. Among them, LCZ 2 showed the largest individual increase, rising by 5.1%. The proportion of low-rise LCZ types (LCZ 3 and LCZ 6) decreased by 8.23%, with LCZ 3 showing the most significant decline, decreasing by 9.57% over the six-year period. As the proportion of built LCZ types increased, the proportion of land cover LCZ types decreased by 2.79%. Among these, the proportion of LCZ D decreased by 2.27%.

The LCZs in the main urban area of Guangzhou exhibited different transformation patterns at different times. From 2018 to 2021, the transformation process was characterized by systematic updates of mid- and low-rise buildings. The transformation of LCZ 3 was the most intense, with 461 units converting to LCZ 2 and LCZ 5, accounting for 83.2% of all conversion units in this category. Additionally, 50 units converted from LCZ D to LCZ 6, reflecting the ongoing expansion of the urban periphery. From 2021 to 2024, the transformation process became more gradual compared to the previous stage, primarily involving adjustments to mid-rise buildings. LCZ 5 was the most transformed type during this stage, with 121 units transitioning to LCZ 2. Meanwhile, the number of units transitioning from LCZ D to LCZ 6 increased to 69, indicating the ongoing encroachment of urban fringe development on natural land cover LCZ types.

3.3. Spatiotemporal Differences in Urban Thermal Environment

3.3.1. SUHII Classification

Table 5. SUHII classification results for 2018, 2021, and 2024.

SUHII	2018		2021		2024
	Count	Ratio	Count	Ratio	Count	Ratio
No SUHII	1142	23.98%	1506	31.62%	1426	29.94%
Very low	118	2.48%	90	1.89%	101	2.12%
Low	2151	45.16%	1588	33.34%	1774	37.5%
Moderate	747	15.68%	683	14.34%	670	14.07%
High	540	11.34%	787	16.52%	647	13.58%
Very high	65	1.36%	109	2.29%	145	3.04%

presents the classification statistics for the SUHII levels for the years 2018, 2021, and 2024. Based on this classification framework, Figure 6 illustrates the spatiotemporal evolution of SUHII levels. Overall, low-level SUHII dominates the study area, and it is primarily distributed in the peripheral regions. However, its proportion shows a fluctuating downward trend, with a cumulative decrease of 7.6%. Moderate-level SUHII is mainly concentrated on the edges of the urban core, with its spatial extent gradually shrinking over time and progressively transitioning to higher intensity levels. In contrast, high and very high levels SUHII exhibit a staged growth trend, forming concentrated distributions originally in the central part of the study area, then gradually expanding outward along the periphery, replacing previously low-value areas such as central Baiyun District and Huangpu District. Statistical results indicate that the total proportion of high and very high levels of SUHII increased by 6.11% from 2018 to 2021 but slightly declined by 2024, suggesting that the SUHII experienced a degree of fluctuation and adjustment.

3.3.2. Inter-LCZ Thermal Characteristics Variations of SUHII

The thermal differences indicate that built LCZ types generally exhibit higher SUHII values. Specifically, LCZ 2 and LCZ 3 show the most pronounced heat effects, recording the highest number of extreme SUHII values (Figure 7). Notably, the median SUHII of LCZ 3 increased from 2.43 °C to 3.11 °C. In contrast, LCZ 1 and LCZ 4 exhibited similar cooling trends, with median SUHII values decreasing by 2.20 °C and 1.94 °C, respectively. LCZ 5 and LCZ 6 displayed comparable thermal characteristics, with a slight upward trend in median SUHII over time.

In comparison, land cover LCZ types showed significantly lower median SUHII values and interquartile ranges (IQRs). LCZ A-B consistently maintained a negative median SUHII, which increased by 0.29 °C from 2018 to 2023. LCZ G had the lowest median SUHII among all types, with IQRs stably ranging between 2.22 and 3.10 °C, indicating the strongest thermal stability, with the median SUHII further decreasing to −3.0 °C. As the baseline for SUHII calculation, LCZ D exhibited greater fluctuations, with its median SUHII rising to 0.31 °C and then falling to −0.3 °C. Furthermore, the ridge plot (Figure 7b) analysis provides additional evidence of the distributional contrasts across LCZ types. In 2018 and 2024, LCZs 1–6 were predominantly above 0 °C, with ridge peaks also located above 0 °C, whereas LCZs A–B, D, and G were mainly below 0 °C, with ridge peaks consistently below the zero line. Interestingly, in 2021, LCZ D shifted to mostly positive values, with a ridge peak above 0 °C, while LCZ 1 showed the opposite pattern, being mostly below 0 °C, with its ridge peak also located below the baseline. This opposite dynamic suggests that the thermal responses of dense built-up areas and vegetated zones may fluctuate differently under certain urban or climatic conditions, highlighting the spatiotemporal heterogeneity in SUHI mechanisms.

3.4. Impact of Urban Morphology on SUHII in Various LCZs

Given the dominant influence of buildings on SUHII, this study focuses on the effects of morphology indicators on SUHII within built LCZ types, quantifying their impact using feature importance. To improve model accuracy, indicators exhibiting strong correlations, namely FLU and FAR, were excluded based on Pearson correlation analysis. Figure 8 presents the feature importance of each morphological indicator on SUHII across different LCZs for 2018, 2021, and 2024.

Based on the interannual median changes of SUHII shown in Figure 7, it is evident that the influence patterns of morphology indicators on SUHII vary markedly across different LCZs, with dominant factors differing among them. In LCZ 4, the importance of BH steadily increased from 0.101 in 2018 to 0.146 in 2024, while the importance of BR rose from 0.121 to 0.169; these increases correspond to the substantial decline in SUHII within LCZ 4. In LCZ 1, which also consists of high-rise buildings, the decline in SUHII was primarily driven by the rising importance of BR and BHCV. LCZ 2 exhibited relatively stable changes, with the importance of BSF gradually decreasing from 0.170 in 2018 to 0.156 in 2024, while the importance of SVF slightly increased from 0.154 to 0.169, corresponding to only a minor decline in SUHII. In contrast, in LCZ 3, the importance of PSF fluctuated downward from 0.286 to 0.219, whereas the importance of BSF steadily increased to 0.184. Under the dominant influence of these two indicators, the SUHII in LCZ 3 showed a continuous upward trend. The SUHII in LCZ 5 and LCZ 6 experienced an initial increase followed by a decline. In LCZ 5, the importance of PSF first decreased to 0.210 in 2021 before rising to 0.221 in 2024, while the importance of SVF also declined initially and then recovered to 0.148. In LCZ 6, the importance of PSF ultimately decreased to 0.222, whereas the importance of BR and BSF first declined and then rebounded over different years, finally reaching 0.188 and 0.155, respectively.

A comparative analysis across LCZ types reveals that PSF consistently exhibited the highest feature importance in most LCZ types across all years, although it displayed a clear downward trend. BSF generally ranked second, maintaining a relatively high level of importance in most LCZs, with a slight upward trend. Although the influence of the remaining morphology indicators was comparatively weak, some of them also demonstrated notable effects in certain LCZs. For example, in LCZ 4, the importance of SVF reached 0.209 in 2018, while the importance of BHCV reached 0.210 in 2021. In LCZ 6, the importance of BR was 0.187 in 2024, and the importance of BH reached 0.185 in LCZ 1 in 2021. Overall, the influence of urban morphological indicators on the thermal environment exhibits a phased evolutionary pattern. The temperature regulation mechanism, initially dominated by PSF, has gradually weakened, with the dynamic changes in SUHI increasingly driven by building-related factors such as BSF, BH, and BR.

4. Discussion

4.1. Improving LCZ Mapping Accuracy with Building Height Data

In this study, LCZ types were first adjusted through visual interpretation of high-resolution imagery, and representative samples were extracted to establish a localized classification framework for Guangzhou. Since building height is a key morphological parameter that captures urban three-dimensional structure [35,36,37], we reconstructed temporally consistent building height datasets using the SHAFTS model and integrated them with other morphological indicators into an improved GIS-based LCZ workflow.

Table 6 showed our LCZ maps achieved an OA of 88.37–92.49%, which is higher than those in most studies in China. Many previous products reported only moderate accuracies, which may partly be attributable to the lack of reliable vertical morphology. For instance, Ma et al. [105] demonstrated considerable variation across cities (74.38% in Shanghai to 91.03% in Nanjing), which may be related to the absence of explicit building height and building surface fraction, thereby making performance more sensitive to local density. The So2Sat GUL dataset achieved 79% OA in Guangzhou despite using Sentinel-1/2 data and deep learning, reflecting the limitations of rule-based classification without vertical data. Ren et al. [33] reported 80% OA for Guangzhou using Sentinel-1 DEM, further suggesting that reliable vertical information can enhance classification consistency.

In contrast, our results highlight the advantages of explicitly retrieved building height, which improves the separability of morphologically similar LCZ types and ensures temporal consistency across multiple years. Importantly, the approach remains effective even in compact urban environments. Xu et al. [106] observed that LCZ classification accuracy in Guangzhou was lower than in Wuhan due to its higher density. However, our relatively high accuracy in Guangzhou suggests that the integration of vertical morphology can help mitigate density-related challenges, underscoring its value in dense urban contexts.

To further benchmark performance, we compared our maps with So2Sat GUL, the first global LCZ dataset. As shown in Figure 9, our results align more closely with high-resolution imagery and reduce the fragmentation present in So2Sat GUL [107]. While So2Sat GUL distinguishes built versus natural zones reasonably well, its lack of building height information limits differentiation within built types. Figure 9 also illustrates the robustness of our framework in capturing block-scale changes: Site 1 (residential) and Site 2 (offices) show densification from open low-rise to compact high-rise, while Site 3 reflects redevelopment from low-rise housing to open high-rise. These examples underscore the ability of our approach to capture both city-wide structure and micro-scale dynamics, highlighting the critical role of three-dimensional morphology in accurate LCZ mapping.

Table 6. Comparison of LCZ classification accuracies with different BH acquisition method.

Study/Dataset	Study Area	BH Acquisition Method	OA
This study	Main urban area of Guangzhou	Explicit retrieval	88.37–92.49%
So2Sat GUL [107]	Global	Implicit height-related features from SAR backscatter and coherence	79%
Ma et al. [105]	Shanghai, Nanjing, Hangzhou	/	74%, 91.03%, 85.95%
Ren et al. [33]	More 50 Chinese cities	DEM derived from Sentinel-1 InSAR data	~80% (Guangzhou)

Table 6. Comparison of LCZ classification accuracies with different BH acquisition method.

Study/Dataset	Study Area	BH Acquisition Method	OA
This study	Main urban area of Guangzhou	Explicit retrieval	88.37–92.49%
So2Sat GUL [107]	Global	Implicit height-related features from SAR backscatter and coherence	79%
Ma et al. [105]	Shanghai, Nanjing, Hangzhou	/	74%, 91.03%, 85.95%
Ren et al. [33]	More 50 Chinese cities	DEM derived from Sentinel-1 InSAR data	~80% (Guangzhou)

4.2. Impacts of LCZ Transitions on the Urban Thermal Environment

In existing studies, most research has primarily focused on the temporal and spatial variations of the urban thermal environment and their associated influencing factors, while systematic investigations into the correspondence between evolution of LCZs and regional surface thermal conditions remain relatively limited [21]. Unlike previous studies, this research is based on multi-temporal LCZ maps and not only provides a detailed characterization of the thermal properties of different LCZ types but also offers an in-depth analysis of the relationship between LCZ transitions and SUHII evolution. This dynamic, spatially oriented approach provides a new perspective for understanding how urban morphology regulates the thermal environment.

The results indicate that the SUHII in land cover LCZs is generally lower than in built LCZs, which is consistent with observations from cities under different climatic contexts, such as Budapest [108], Phoenix [109], and Wuhan [110]. Some studies suggest that this phenomenon may be attributed to the evapotranspiration and shading effects provided by vegetation, which help to suppress increases in surface temperature to some extent [111]. This study confirms this mechanism and further highlights that LCZ G, due to its high specific heat capacity and continuous latent heat exchange, exhibits stable thermal characteristics and consistently low SUHII [112]. In contrast, built LCZs are primarily composed of impervious surfaces with high heat capacity and low albedo [113], enabling them to absorb and store substantial amounts of solar radiation. At the same time, concentrated anthropogenic activities significantly increase heat emissions, further intensifying the urban heat island effect [114]. Notably, this study found that the regulatory effect of vegetation has weakened to varying degrees across most built LCZ types. This suggests that, as urban morphology and building characteristics change, the thermal regulation mechanism originally dominated by the PSF is gradually diminishing, and urban building form parameters are becoming key drivers influencing the spatiotemporal evolution of SUHI.

This shift in mechanisms is reflected in distinct thermal characteristics across different built LCZ categories. For example, LCZ 2 and LCZ 3 have consistently exhibited relatively high SUHII, a pattern also supported by in situ observations in Guangzhou [115]. Contrary to previous studies that generally report LCZ 3 as having the highest SUHII [67,116], our study finds that the thermal intensity difference between these two types is not fixed but dynamically fluctuates over time. Meanwhile, LCZ 1 and LCZ 4 generally show relatively low SUHII across most years. This difference may be attributed to building height and spatial layout: high-rise buildings cast larger shadows, effectively reducing incoming solar radiation and lowering daytime temperatures [117]. The declining trends observed in LCZ 1 and LCZ 4 may also be related to increased vertical development, enhanced shading effects, and the implementation of rooftop and vertical greening measures [118,119], demonstrating the potential of urban morphological optimization to mitigate urban heat island effects.

In contrast, the SUHII in LCZ 3 and LCZ 6 shows an increasing trend, which may be related to the large-scale demolition and redevelopment of low-rise buildings [120]. Previous studies generally considered LCZ 6 to have the lowest SUHII [67]; however, our findings indicate that its thermal effect has intensified over time, ranking as the third highest among built LCZs in 2024. This may be attributed to the exposure of bare soil and gravel (with low thermal inertia) during demolition and reconstruction phases, which leads to rapid daytime surface heating [21,121]. In addition, vegetation loss associated with construction activities further reduces evapotranspiration capacity, exacerbating local SUHII.

4.3. Heterogeneous Effects of Urban Morphology on SUHI Across LCZ Types

Numerous studies have demonstrated that the relationship between SUHI and urban morphology exhibits significant spatiotemporal dynamics [12,22,102]. In this study, we employed the G-XGBoost model to construct a nonlinear explanatory framework, enabling a refined analysis of the differential contributions of morphology factors to SUHI across various LCZ types. Specially, while vegetation has been widely recognized for its cooling effects through evapotranspiration and shading [67], and PSF is generally considered the most important explanatory factor across LCZ types [67,122,123], LCZ 4 presents an exception. For example, the importance of PSF peaked at 0.393 in 2018 but dropped to fourth place in 2024, likely reflecting urban renewal and vegetation loss. In LCZ 4, this sharp decline can be largely attributed to reduced vegetation cover caused by building renewal. Since LCZ 5 was the only conversion source of LCZ 4, this transition further suggests that vegetation decline amplified the role of BH, a variable strongly linked to building form. Excluding LCZ 4, PSF typically followed a “decline–then–rebound” trajectory, corresponding to phased vegetation dynamics—initial construction-induced losses [121,124] followed by recovery through greening efforts. By contrast, BSF showed a generally increasing trend (except in LCZ 2), underscoring its growing role in surface heat absorption and storage. Notably, in 2018, the median SUHII of LCZ 2 was slightly higher than that of LCZ 3, with BSF showing similar importance in both types. However, in 2021 and 2024, BSF importance in LCZ 2 slightly declined relative to LCZ 3, with LCZ 3 exhibiting a higher median SUHII. Such subtle temporal differences indicate that even minor fluctuations in morphological contributions can drive SUHI intensity shifts—patterns often overlooked by conventional methods.

When building density was further considered, compact LCZ types exhibited a more balanced distribution of feature importance, suggesting that SUHI formation in these areas is jointly regulated by multiple factors. By contrast, open LCZ types were often dominated by one or two key drivers, with PSF usually ranking first, while the relative roles of other variables varied substantially across years and types without forming a stable pattern. In low-rise LCZs, the ranking of variable importance was more consistent overall, yet local temporal differences were evident. For example, in LCZ 3, BSF remained second to PSF in both 2018 and 2024, whereas BR contributed little; in LCZ 6, however, BR played a stronger role, ranking second to PSF in both years.

Overall, urban morphological evolution and spatial configuration can fundamentally reshape the mechanisms driving SUHI [122]. More importantly, they demonstrate that SUHI drivers not only fluctuate temporally but also reveal structural differences both across and within LCZ types. Compared with traditional approaches such as linear regression or generalized linear models [69,70], our results offer a more refined and dynamic perspective. The integration of LCZ classification with the G-XGBoost model enhances explanatory accuracy, allowing us to capture nonlinear and type-specific relationships often ignored by conventional methods. Furthermore, by quantifying the temporal evolution of morphological contributions within built LCZs [125], this study transcends previous single-pattern assessments of the morphology–SUHI relationship, providing more targeted and practical insights for urban planning and thermal environment regulation.

4.4. Strengths and Limitations

This study achieved several key methodological advances. Firstly, by employing the sensitivity-validated SHAFTS model, we successfully reconstructed multi-temporal building height datasets, effectively supplementing the key three-dimensional urban morphology features in LCZ classification. Meanwhile, the LCZ classification framework was constructed to be more applicable to the main urban area of Guangzhou by combining the regional urban morphology features, resulting in notable improvements in both classification accuracy and spatial adaptability. Secondly, the application of the G-XGBoost model enabled nonlinear modeling of the driving mechanisms of SUHI, addressing the limitations of traditional linear approaches in capturing spatial heterogeneity, and broadening the method for urban thermal environment studies. More importantly, this study systematically explored the multi-scale impacts of urban morphological evolution on thermal conditions by considering both inter- and intra-LCZ variations, offering a more nuanced perspective on the spatial response characteristics of SUHI.

Nevertheless, several limitations should be acknowledged. First, the accuracy of LCZ classification is highly dependent on the quality of building height estimation. Although the SHAFTS model demonstrates overall robustness, it tends to underestimate height in morphologically complex areas such as urban villages, thereby reducing the precision of local morphological identification. Secondly, this study used different reference building height datasets, which may affect RMSE comparisons and limit cross-year comparability. Moreover, the lack of direct field measurements means that the accuracy assessment relies solely on these reference datasets, which constrains the ability to better evaluate the precision of building height estimates and the resulting LCZ classification. Thirdly, the thresholds of several key parameters, such as BH and PSF, have limited theoretical validation, and their applicability has not been thoroughly assessed. Therefore, improving the scientific basis and adaptability of these parameterization methods remains a critical challenge in LCZ mapping practice. In addition, the temporal coverage of this study is relatively limited, involving only three time points. As a result, the observed thermal trends may be partially affected by data noise, weakening the stability of temporal inferences. Incorporating additional time-series data could improve the robustness of trend analysis. Finally, frequent summer cloud cover and rainfall in the study area limited the availability of satellite imagery suitable for LST retrieval. Future work should explore multi-source satellite data integration or spatiotemporal data fusion techniques to enhance the continuity and reliability of thermal environment monitoring. Despite these limitations, this study presents a replicable analytical framework for modeling the relationship between urban morphology and SUHI, and provides a clear direction for further improving LCZ classification accuracy and SUHI modeling performance.

5. Conclusions

In this study, building heights for 2018, 2021, and 2024 in the main urban area of Guangzhou were estimated using the SHAFTS model, providing critical height information for LCZ classification. LCZ maps generated by GIS-based methods revealed the evolution of urban morphology in detail. LST values were retrieved using the SMW algorithm to calculate SUHII, allowing for an analysis of thermal characteristic variations among different LCZ types. The G-XGBoost model was applied to quantitatively assess the contribution of urban morphology to SUHI across LCZ types in different years, thereby revealing the spatiotemporal heterogeneity in the driving mechanisms of SUHI. The main conclusions are as follows:

(1)

Building height estimation errors in the main urban area of Guangzhou were controlled within a range of 5.92–7.03 m, indicating excellent model performance. The LCZ maps improved by incorporating building height data showed a high degree of similarity to high-resolution satellite imagery, accurately capturing morphology changes during urban renewal.

(2)

Built LCZ types were predominant in the study area, with an average proportion of 76.53%. From 2018 to 2021, LCZ conversions were characterized by the renewal of low- and mid-rise buildings, with LCZ 3 exhibiting the most significant changes. From 2021 to 2024, LCZ conversions mainly involved morphological adjustments of mid-rise buildings, with LCZ 5 experiencing the greatest transformation during this period.

(3)

The SUHI effect exhibited evident temporal variations, with low- and moderate-level SUHII showing a decreasing trend, while high and very high levels SUHII showed an overall increasing trend with fluctuations. Stronger SUHI effects were observed in built LCZ types, particularly LCZ 2 and LCZ 3. In contrast, the median SUHII of LCZ G continued to decline, reaching −3.0 °C.

(4)

Significant spatiotemporal heterogeneity was observed in the influence of urban morphology on SUHII across different built LCZ types. Although the importance of PSF declined over time, it consistently remained the core driving factor for SUHII. BSF also showed a steadily increasing importance, indicating its significant impact. Other variables, such as BR, BH, SVF, and BHCV, were influenced by urban renewal disturbances, resulting in fluctuating impacts across different years and LCZs.

The analytical framework proposed in this study was able to accurately capture the dynamics of urban morphology during urban renewal and focused on exploring the complex interactions between urban morphology and SUHI effects across different LCZs over time. Block-scale analysis not only enabled the interpretation of how urban morphology evolution influences the thermal environment but also provided refined scientific support for urban planning, thereby contributing to effective thermal environment mitigation.