Decoding Tianjin: How Does Urban Form Shape the Diurnal Cycle of Surface Temperature?

Abstract

In the context of global climate change and the increasing severity of the urban heat island effect, it is particularly important to study the spatial variation mechanism of urban land surface temperature (LST). The LST data provided by ECOSTRESS offer a new perspective for deepening our understanding of the diurnal cycle and spatial variation of urban LST. In this study, based on a block scale, Tianjin is divided into nine block types, and a multi-scale geographic regression weighting (MGWR) model is used to comprehensively explore the relative contributions of urban 2D and 3D landscape indicators of different block types to the spatial changes in diurnal urban LST cycles. The results indicate that ① the thermal effect during the daytime is mainly influenced by the building density, while at night, it is more influenced by the building height and the heat retention effect; ② the building indicator and the water-body indicator had the most significant effect on surface temperature at different observation times; ③ the influence of urban morphology on land surface temperature shows significant spatial non-stationarity across different block types. This study enhances the understanding of the mechanisms driving urban heat island formation and provides a scientific basis for urban authorities to develop more effective urban planning and heat island mitigation strategies.

1. Introduction

Globally, rapid urbanization has led to a significant increase in the urban population. It is estimated that by 2050, 640 million people, or 67% of the total global population, are expected to live in urban areas [1]. Urbanization not only changes land use patterns but also increases anthropogenic heat source emissions, significantly affecting human thermal comfort in urban areas [2]. This growth has not only changed the geography of cities but has also exacerbated the urban heat island (UHI) effect, where urban areas are hotter than the surrounding rural areas [3]. At the same time, global climate change is further exacerbating these problems. Extreme heat events, such as heat waves, are becoming more frequent and intense as a result of climate change [4,5]. Studies have shown that exposure of the global urban population to extreme heat has increased by nearly 200% between 1983 and 2016 due to the combined effects of global warming and the urban heat island effect [3]. This increased exposure not only poses a threat to human health but also to the sustainability of cities. In addition, characteristics such as urban form and layout play an important role in the exposure of urban dwellers to extreme thermal events, with large temperature differences between different parts of the city, especially during hot weather, due to differences in building densities, green space coverage, road types, gradients, and wind speeds. For example, the commercial and transportation-intensive areas in the city center often become hot spots due to dense construction and scarcity of green space, while green spaces and parks in the periphery of the city may be relatively cool [6], and the microclimate within the city shows significant spatial non-stationarity characteristics. Regardless of how cities grow, ways of designing and managing land use based on spatial non-smoothness characteristics of LSTs within cities could be a key tool for adapting to ongoing climate change.

The current research on urban heat islands mainly focuses on two categories: canopy urban heat island (CUHI) [7] based on air temperature (AT) and surface urban heat island (SUHI) [8] based on LST. AT measured by stationary weather stations has the advantage of high reliability [9] and continuity [10], while large-scale air temperature data from stationary weather stations are useful for assessing the UHI effect and the impact of urbanization on regional climates [11]. However, because fixed weather stations are usually unevenly distributed, especially in densely built-up urban centers, this limits their ability to capture microclimatic differences within cities [12]. This is coupled with the fact that nighttime temperature data may underestimate the actual heat island intensity if the urban heat island effect is not adequately considered [13]. Compared to traditional ground-based measurement methods, remote sensing technology can significantly reduce costs because it does not require the deployment of a large number of ground stations in complex urban environments [14]. Meanwhile, LST acquired by remote sensing technology is valuable for urban climate research due to its advantages of high spatial resolution [15] and coverage [16]. Currently, the vast majority of remote sensing data used to study the urban thermal environment mainly include satellite data such as Landsat series [17], MODIS [18], and GaoFen [19]. Current research on urban LST focuses on different seasons [20] and long time scales [21], and there is a lack of research on the diurnal cycle of urban LST. The urban UHI phenomenon has large differences between day and night; during the day, urbanization affects surface temperature mainly by altering the surface energy balance [22]; at night, UHI intensity is related to urban geometry [23]. The urban energy balance follows a 24-h cycle, and its assessment must take into account the difference in energy exchange between daytime and nighttime, the urban heat island effect, the role of urban green spaces, the impact of urban design on energy efficiency, and the combination of modeling and observational data. Together, these factors determine the state of the urban energy balance in order to more accurately understand and improve energy use and environmental quality in cities [24]. Since the observation time of the above satellite data is fixed in the same study area, the data are not available in the studied LST diurnal cycle. Fortunately, NASA successfully launched ECOSTRESS (Ecosystem Spaceborne Thermal Radiometer Experiment on Space Station) to the International Space Station (ISS) on 29 June 2018 and completed the installation on 3 July 2018. ECOSTRESS is a multi-channel thermal infrared-imaging radiometer whose primary mission is to study and understand how climate change is affecting the Earth’s water and carbon cycles by observing temperature changes at the Earth’s surface from the space station [25]. The instrument consists of five thermal infrared spectral bands centered at 8.29, 8.78, 9.20, 10.49, and 12.09 μm [26]. Radiation in these wavelength ranges is able to effectively penetrate clouds and the atmosphere, directly reflecting temperature information at the surface. The LST product of ECOSTRESS provides high spatial resolution data of 70 m × 70 m, which makes it possible to capture subtle surface temperature variations, which is important for studying the surface temperature distribution [27]. In addition, the ECOSTRESS instrument operates over a ±25° swath and is capable of generating cross-track strips equivalent to about 400 kilometers on the surface, with each swath lasting 1.29 seconds, which further improves the capability and efficiency of the surface it observes [28]. In summary, the LST products provided by ECOSTRESS offer new tools and perspectives for exploring the temperature distribution and diurnal patterns of change within cities.

Most of the current analyses of the spatial patterns and drivers of the urban temperature environment use the grid as the basic unit. The gridded approach may not adequately capture micro-variations within the city, such as the specific layout of buildings and the differences in the thermal environment between different functional areas [29]. Second, the heat island effect in cities is not only determined by the LST but is also influenced by a variety of factors, such as socio-economics, demographics, and building characteristics. These factors may be oversimplified or neglected at the grid scale [30]. As an important statistical unit in urban planning and management, the role of blocks in the study of the relationship between thermal environmental factors and diurnal surface temperature should not be overlooked. By analyzing the distribution of building types, ages, roofing materials, and surrounding green spaces and water bodies within different blocks, it is possible to more accurately assess how these factors affect the thermal environment of the area [31]. In addition, block-scale studies can help identify and optimize “hot spots” in cities, i.e., areas with high temperatures due to human factors, which is important for targeted urban planning and improving the quality of life of residents [32].

The landscape and structure of a city affect the urban thermal environment in many ways. Land use/cover type (LULC) is an important factor affecting the urban thermal environment, the physical characteristics of the surface in urban areas with rapid population growth have been altered, and impervious surfaces and anthropogenic heat emissions have become major contributors to warming [33]. Urban canyon geometry plays an important role in the nighttime urban heat island effect [23]. This suggests that spatial layout in urban design, such as the arrangement of streets and buildings, has a significant impact on regulating the internal temperature of the city. Urban form also has a significant effect on the urban thermal environment; for example, low sky view factor (SVF) and high SVF usually mitigate the UHI effect, while medium SVF may exacerbate the UHI effect [6]. This suggests that the spatial structure within the city, such as the density of buildings and the way they are distributed, plays an important role in regulating the urban thermal environment. Meanwhile, meteorological factors, such as cloudiness, wind speed, and relative humidity [34], and three-dimensional morphological parameters, such as the area and height of buildings [35], are important variables that affect the urban thermal environment. These factors work together in the formation and development of urban thermal environments through different mechanisms, such as radiative exchange, evaporative dispersion, and turbulent exchange. Therefore, it is important to consider these factors in urban planning and design to improve the urban thermal environment and enhance the comfort and health of residents. In urban environments, blue–green space is used as a natural solution to reduce the temperature of the surrounding area by adding blue (e.g., rivers, lakes) and green (e.g., parks, meadows) spaces to create a relatively cool area that helps to mitigate the urban thermal environment. The cooling distance and cooling intensity of blue–green spaces, as typical cooling indicators, are influenced by a variety of factors, including the size [36], type [37], shape [38], vegetation cover [39], and the surrounding environment of the blue–green space.

Understanding the relationship between the above urban 2D and 3D landscape factors and LST is crucial for studying urban climate and developing sustainable management strategies. Data-driven approaches are challenging for understanding urban landscape characterization and climate dynamics. Current research methods include linear regression models [40], geographic regression weighted models (GWR) [41], multi-scale geographic regression weighted models (MGWR) [42], and machine learning algorithms [43,44,45]. In urban environments with significant microclimates, the variation in LST is influenced by a variety of 2D and 3D urban landscape factors, and the effects of vegetation, water bodies, and building density in different areas on surface temperature may be spatially non-stationary. In contrast, linear models usually assume that the parameters are constant throughout the study area, and machine learning models, while capable of handling nonlinear and complex data relationships, may require additional spatial processing to achieve performance comparable to GWR or MGWR [46]. The GWR and MGWR models are able to capture and analyze spatial non-stationarity, allowing for the regression coefficients to vary with geographic location, thus revealing spatial heterogeneity. At the same time, they emphasize local regression analysis [47], which is able to capture the local relationships of different regions in the city, which is very important in thermal environment studies. As an extended model of GWR, MGWR is able to better capture the complexity and diversity of the impact of urban characterization factors on LST by choosing the optimal bandwidth for each covariate and allowing different processes to operate at different spatial scales [48]. And it shows higher flexibility and accuracy when dealing with urban data with complex spatial structure. When studying the relationship between urban form and LST, the MGWR model can reveal the influence of urban form indicators on LST at different operational scales, which provides a more fine-grained guidance for urban planning and heat island effect mitigation. However, due to the low frequency of acquisition of ECOSTRESS data and the problem of discontinuous or missing data caused by atmospheric conditions, as well as the high computational complexity of the MGWR model, this combination is not widely used in the study of diurnal variations of LST at present.

In this study, we integrate ECOSTRESS satellite data into the GWR and MGWR models, and the better-performing models are used to explore the relative importance and spatial non-stationarity of the effects of urban 2D and 3D landscape metrics on urban diurnal LST for different block types. To ensure the accuracy of the results, we rigorously validated and pre-processed all datasets and adjusted key parameters for each method. This study aims to address the following concerns: (1) the spatial change pattern of urban diurnal LST; (2) the spatial distribution pattern of urban diurnal LST in different block types; and (3) the relative importance and spatial heterogeneity of the influence of urban two-dimensional and three-dimensional landscape factors on diurnal LST in different block types. This study can provide a scientific basis for the development of targeted urban planning and strategies to mitigate the heat island effect by the relevant departments in each block.

2. Materials and Methods

2.1. Study Area

Tianjin is a municipality directly under the central government and a is mega-city in China located in the northeastern part of the North China Plain, downstream of the Haihe River Basin, and east of the Bohai Sea. Tianjin has a warm, temperate, semi-humid monsoon climate with four distinct seasons. The average annual temperature ranges from 12 to 15 °C, with the highest average temperature in the city lasting from 90 to 140 days of summer and the highest temperature of the year in July. The average annual precipitation is 550–600 mm, with June–August precipitation accounting for about 75% of the total annual precipitation. The monsoon prevails in Tianjin, with an average annual wind speed of 2 to 4 m/s, mostly from the southwest (Figure 1).

As of October 2023, Tianjin has 16 districts with a total area of 11,966.45 km². As of the end of 2023, Tianjin had a resident population of 13.64 million, with an urbanization rate of 85.49%. Since entering the 21st century, Tianjin has experienced significant urbanization expansion, and the central city area has expanded significantly [49]. This urban expansion has led to an increase in urban construction land and a concentration of buildings, and Tianjin thus faces the problem of urban heat island effect [50].

The Outer Ring Road of Tianjin’s central city was built in 1987, and as the core area of the city within the line, with an area of about 427 km², it has accumulated a large amount of economic, cultural, and commercial activities over a long time. With its high building density and dense population, it is a major area of manifestation of the urban heat island effect and microclimate change. At the same time, the area within the Outer Ring Road is a transportation hub of Tianjin with well-developed public transportation and infrastructure, plus the area contains alternating portions of old and new urban areas with unique historical and modern architectural features. This makes the region show complex interactive effects of climate change, urban green space, and building density and can provide strong data support for understanding the multidimensional impacts of microclimate change during urban development. Therefore, this paper focuses on the central city of Tianjin within the Outer Ring Road, an area that can serve as a representative sample of other high-density urban centers for broader urban microclimate research.

2.2. Data Sources and Preprocessing

2.2.1. ECOSTRESS LST Data

ECOSTRESS LST data were obtained from NASA’s AppEEARS website (https://appeears.earthdatacloud.nasa.gov/ (accessed on 24 June 2024)) using the ECOSTRESS Tiled Surface Temperature and Emissivity Instantaneous Level 2 Global 70 m (ECO_L2T_LSTE) version 2 data product. It provides atmospherically corrected land surface temperature and emissivity (LST&E) values from five thermal infrared bands. The product is derived using a physically based temperature and emissivity separation (TES) algorithm with a spatial resolution of 70 meters. ECO_L2T_LSTE Version 2 data products are provided directly in cloud-optimized GeoTIFF (COG) format, with each band distributed as a separate COG. Obtaining high-quality images at different points in time on the same day is very difficult due to the high rainy weather and thick cloud cover during the warmer months of June through September. Moreover, in months with heavy rainfall, this may lead to missing or incomplete data as well as cause bias in temperature values, thus affecting the results of the temperature analysis and weakening the scientific validity of the study. In addition to this, the temporal repetition period of ECOSTRESS in the same region is 3 to 5 days, and the period of availability of the ECO_L2T_LSTE data of Tianjin is after 24 January 2022. Therefore, in this study, ECOSTRESS LST data for five seasons (May to September) with higher temperatures from 2022 to 2023 were selected and filtered to finalize the selection of high-quality imagery for four different date study times, namely 1 May 2023 at 08:26, 30 September 2023 at 13:30, 8 May 2024 at 21:55, and 22 May 2022 at 00:59. Since ECOSTRESS LST data are acquired at different times on different dates (

Table 1. Weather conditions corresponding to the ECOSTRESS LST used in this study.

Date	Time (GMT + 8)	Wind Direction	T	RRR	H	Tn	Tx	Impact of Weather Conditions on LST
1 May 2023	08:26	south wind	18.5	no rainfall	>2500 m, or no clouds	10.6	23.8	Consistent favorable weather conditions for obtaining LST data
30 September 2023	13:30	north wind	24.8	no rainfall	>2500 m, or no clouds	14.2	26.5
8 May 2024	21:55	southwest wind	24.3	no rainfall	>2500 m, or no clouds	11.7	29.6
22 May 2022	00:59	southwest wind	23.2	no rainfall	>2500 m, or no clouds	18.0	35.0

Data are from the historical data of Tianjin Observatory 54527 on the RP5 website (https://rp5.ru/ (accessed on 26 August 2024)). T refers to the atmospheric temperature at 2 m above ground level (°C); RRR refers to the amount of precipitation in millimeters; H refers to the altitude of the bottom of the lowest cloud; Tn refers to the minimum air temperature during the past period of time (up to 12 h); Tx refers to the maximum air temperature during the past period of time (up to 12 h).

) and differences in weather conditions may have an impact on data accuracy, this study utilizes the hour-by-hour LST products acquired by the Copernicus Global Terrestrial Service (https://land.copernicus.eu/global/products/lst (accessed on 24 June 2024)) to Quantify LST Differences for Correcting ECOSTRESS LST Data. In this study, we collected hourly LSTs available during the 1 May to 30 September 2022 and 1 May to 30 September 2023 study periods. Although the acquired ECOSTRESS LST data and the Copernicus Global Terrestrial Service (CGTS) LST product will fluctuate somewhat at the same measurement time, a correction of the ECOSTRESS LST data can be performed by using the average LST diurnal cycle of the study time period as a reference baseline by which the overall trend can be represented, using the following formula [40,51,52]:

$$LST(t)=LS{T}_{ECO}(d,t)+T_{Global\_m}(t)-T_{Global}(d,t)$$

(1)

where $LST(t)$ denotes the $LST$-corrected for time t, $LST(d,t)$ denotes the ECOSTRESS LST obtained for date $d$ and time $t$, $T_{Global}(d,t)$ denotes the LST of date $d$ and time $t$ at the Copernicus Global Land Service, and $T_{Global\_m}(t)$ denotes the average LST at time $t$ over the entire study period at the Copernicus Global Land Service.

2.2.2. Block Demarcation Methods

The study of specific blocks helps to reveal the complexity and diversity of urban microclimates [53]. In this study, urban blocks are manually delineated using road data and administrative district boundary data in the central city of Tianjin. Roads data were obtained from the Digital Earth Open Web Platform (https://open.geovisearth.com (accessed on 25 June 2024)), including primary, secondary, tertiary, and quaternary roads, railroads, and provincial highways, and administrative-boundaries data were obtained from the Resource and Environment Science Data Platform (https://www.resdc.cn/ (accessed on 25 June 2024)). A total of 1796 blocks were delineated, and it should be noted that blocks with areas < 1 ha or BDs equal to zero were merged using Arcgis 10.8 in order to accommodate the different resolutions of the various data sources and to avoid fragmentation, so this number is much smaller than the number that would have been achieved if the area had been delineated using only roads.

According to the previous classification of blocks [40], BH and BD were used as the criteria for block classification (

Table 2. Block classification criteria.

Classifications	BD(%)	BH(m)
LRLD	BD < 0.15	BH < 10
LRMD	0.15 ≤ BF < 0.25	BH < 10
LRHD	BF ≥ 0.25	BH < 10
MRLD	BD < 0.15	10 ≤ BH < 25
MRMD	0.15 ≤ BF < 0.25	10 ≤ BH < 25
MRHD	BF ≥ 0.25	10 ≤ BH < 25
HRLD	BD < 0.15	BH ≥ 25
HRMD	0.15 ≤ BF < 0.25	BH ≥ 25
HRHD	BF ≥ 0.25	BH ≥ 25

), and urban blocks were divided into nine categories (Figure 2): 95 low-rise low-density (LRLD), 59 low-rise medium-density (LRMD), 31 low-rise high-density (LRHD), 128 mid-rise low-density (MRLD), 364 mid-rise medium-density (MRMD), 688 mid-rise high-density (MRHD), 30 high-rise low-density (HRLD), 129 high-rise medium-density (HRMD), and 272 high-rise high-density (HRHD) blocks.

2.2.3. Data on Urban Landscape Indicators

Sentinel-2 satellite imagery derived from Copernicus Data Spatial Ecosystems (https://dataspace.copernicus.eu (accessed on 20 June 2024)), with a spatial resolution of 10 m, was used to perform object-based land classification using eCognition 9.0 software, which identified five feature classes: vegetation, water bodies, impervious subsurface (other than buildings), buildings, and bareland. In order to reduce the redundancy between landscape pattern indices, landscape pattern indices such as aggregation index (AI), largest patch index [5], and edge density (ED) were selected in this study, and the landscape pattern indices of each type of features were calculated based on a moving-window algorithm using Fragstats 4.0 software. The building data were obtained from the Fudan University team [54], with a spatial resolution of 10 meters, and contain building height (CNBH) data and building footprint data, from which building density (BD) can be calculated. Anthropogenic heat flux data were obtained from the Spatio-Temporal Tripolar Environmental Big Data Platform (http://poles.tpdc.ac.cn/ (accessed on 15 September 2024)) at a spatial resolution of 500 m. Population data were obtained from the WorldPop platform (https://www.worldpop.org/ (accessed on 3 September 2024)) on 16 September 2020 and were used to calculate the spatial aggregation pattern of the population. Wind data were obtained from the National Centers for Environmental Information (https://www.ncei.noaa.gov/ (accessed on 16 September)), hourly wind data for 2022 and 2023 were obtained through the Integrated Surface Dataset (Global), and the ventilation metric Frontal Area Index was calculated using wind and building data [55]. As a coastal city, Tianjin’s sea breeze system can indirectly affect the urban heat island effect by changing air flow patterns. It should be noted that, in this study, before applying the GWR and MGWR models, the variance inflation factor (VIF) test was used to assess the covariance of the independent variables, and those with a VIF value of greater than 10 were excluded, and the retained 2D and 3D cityscape metrics were used as independent variables (

Table 3. Two-dimensional and three-dimensional cityscape metrics used in this study.

Typology	Norm		Unit	Formula	Reference
2D	Landscape components	Percentage of water area (PW)	%	$P_i={\displaystyle \frac{S_i}{S_{Aj}}}$, where $S_i$ denotes the size of patch type (class) $i$ in each block, $S_A$ denotes the total area of each block $j$.	[52]
		Percentage of impervious surface area (PI)
		Percentage of building footprint area (PB)
	Landscape patterns	Vegetation aggregation (AI_v)	%	$AI=\left[{\displaystyle \frac{g_{ii}}{\max \to gii}}\right](100)$, where $g_{ii}$ denotes the number of like adjacencies (joins) between pixels of patch type (class) $i$ based on the single-count method. $\max \to gii$ denotes the maximum number of like adjacencies (joins) between pixels of patch type (class) i (see below) based on the single-count method.	(C3)Aggregation Index (https://fragstats.org/index.php/fragstats-metrics/patch-based-metrics/aggregation-metrics/c3-aggregation-index (accessed on 17 October 2024))
		Water aggregation (AI_w)
		Bareland aggregation (AI_ba)
		Vegetation largest patch index (LPI_v)	%	$LPI={\displaystyle \frac{\underset{j=1}{\overset{n}{\max (a_{ij})}}}{A}}(100)$, where $a_{ij}$ denotes the area (m2) of patch $ij$, and $A$ denotes the total landscape area (m2).	(C3)Largest Patch Index (https://fragstats.org/index.php/fragstats-metrics/patch-based-metrics/area-and-edge-metrics/c3-largest-patch-index (accessed on 11 September 2024))
		Water largest patch index (LPI_w)
		Largest patch index of impervious surface (LPI_i)
		Largest patch index of building footprint (LPI_bu)
		Bareland largest patch index (LPI_ba)
		Edge density of water (ED_w)	m/ha	$ED={\displaystyle \frac{E}{A}}(10000)$, where $E$ denotes the total length (m) of edge in landscape, and $A$ denotes the total landscape area (m2).	(L4)Edge Density (https://fragstats.org/index.php/fragstats-metrics/patch-based-metrics/area-and-edge-metrics/l4-edge-density (accessed on 11 September 2024))
		Edge density of impervious surfaces (ED_i)
		Bareland edge density (ED_ba)
	socio-economic factors	Size of population(POP)	N
		Anthropogenic heat flux (AHF)	W/m2
3D	Meteorological factor	Frontal area index (FAI)	-	$FAI={\displaystyle \frac{A_{(\theta )}}{A_{plane}}}$, where $A_{(\theta )}$ denotes the projected area of a building in a particular wind direction, and $A_{plane}$ the area of the calculation unit.	[56]
	Urban canyon parameters data	Building height (BH)	m	$BH={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n(H_i\times S_i)}}{{\displaystyle {\sum }_{i=1}^n{S}_i}}}$, where $H_i$ the height of building $i$, and $S_i$ denotes the footprint of building $i$.	[40]
		Sky view factor (SVF)	-	$SVF=1-{\displaystyle {\sum }_{i=1}^N{\sin }^2{\beta }_i}({\displaystyle \frac{{\alpha }_i}{{360}^{°}}})$, where $N$ denotes the number of azimuthal directions sampled, ${\alpha }_i$ denotes the height angle of the shade in the direction $i$, and ${\beta }_i$ denotes the horizontal angle of the occlusion in the $i$ direction.	[57]

).

2.3. GWR and MGWR Models

GWR is a statistical method for modeling spatially non-stationary relationships. The core idea is to take into account the spatial relationships between different geographic locations and adjust the regression model by assigning different weights to each observation to better capture the spatial heterogeneity of the data. The following is the formula for GWR [58]:

$$y_j={\beta }_{j0}+{\displaystyle \sum _{k=1}^p{\beta }_{jk}{x}_{jk}+{ϵ}_j}, j=1,\dots ,n$$

(2)

where $y_j$ denotes the dependent variable at location $j$, ${\beta }_{j0}$ denotes the intercept term at location j, $x_{jk}$ denotes the $k$th explanatory variable at location $j$, ${\beta }_{jk}$ denotes the local regression coefficient of the $k$th explanatory variable at location $j$, and ${ϵ}_j$ denotes the random error term associated with location $j$.

MGWR is an extended GWR model that captures spatial heterogeneity by allowing different variables to have different optimal bandwidths, thus providing finer spatial analysis capabilities [42]. The MGWR model can be expressed as follows:

$$y_j={\beta }_{jbw0}+{\displaystyle \sum _{k=1}^p{\beta }_{jbwk}{x}_{jk}+{ϵ}_j}, j=1,\dots ,n$$

(3)

where ${\beta }_{jbw0}$ denotes a specific bandwidth for each independent variable.

3. Results

3.1. Local Climatology and Spatial Distribution of Diurnal LSTs

Figure 3 shows the climatic series of monthly mean temperature and precipitation from January 2019 to October 2024 for the study area. From the figure, it can be seen that there is a clear seasonal variation in temperature and precipitation in the study area. The average monthly temperature reaches its highest in summer (May to September) and drops to its lowest in winter (December to February). Precipitation likewise peaks during the summer months, with notable peaks especially in July and August, and decreases significantly during the winter months (November through February). This seasonal characterization indicates that the region has a warm and humid climate in the summer and a cold and dry winter. Selecting May to September as the study period can capture the peak period of temperature and precipitation and can reflect the regulation of the thermal environment by urban multidimensional characterization factors more obviously.

The LSTs for the four observation times in a day are shown in Figure 4. During the daytime (08:26 and 13:30), the mean LST values were 29.049 °C and 38.631 °C, respectively, with the highest temperature occurring at 13:30 (Figure 4c), reaching 46.25 °C, and the temperature difference at this moment also reached its maximum at 15.3 °C. The temperature and temperature difference at 08:26 were the second highest (Figure 4b), with 36.17 °C and 11.82 °C, respectively. The high daytime temperature zones are mainly sporadically distributed in built-up areas, which may be related to the high heat capacity of urban building materials and the higher intensity of sunshine, while the low value zones are mainly located in rivers and green spaces, highlighting the role of water bodies and vegetation cover in cooling. Compared to the daytime, the LST at night (21:55 and 00:59) had eased as the intensity of sunshine and radiation decreased, with mean values of 27.632 and 24.329 °C and temperature differences of 7.28 and 9.32 °C, respectively, with the lowest temperature, which was only 18.11 °C, occurring at 00:59 a.m. (Figure 4a). The sub-low temperature time-point was 21:55 (Figure 4d), when the minimum temperature was 22.8 °C, while the temperature difference reached a minimum of 7.28 °C at this time. Areas of high nighttime temperatures are contiguous in the central part of the built-up area, with green space remaining the main area with low values. The LST of the water column at 00:59 decreased further as the radiation decreased further during the night. Overall, there were significant differences in the distribution patterns of the thermal environment between daytime and nighttime, with microclimatic environments evident throughout the study area.

3.2. Diurnal Change Patterns of Urban LST in Different Blocks

The LSTs for different blocks at the four observation times are shown in Figure 5. As a whole, there are similarities in the thermal patterns of the blocks during daytime and nighttime, although there are differences in the LST of the blocks at different times of the day.

During the day (08:26 and 13:30), LST was significantly higher than at night (21:55 and 00:59). In particular, at 13:30 (Figure 4c), the average LST reached its highest in all blocks, where the highest value was 42.665 °C. During the daytime, high-density low-rise residential (LRHD) blocks generally exhibited higher LSTs, reaching 29.187 °C (08:26) and 39.005 °C (13:30), respectively, and were the blocks with the highest daytime LST averages. Meanwhile, the hot areas during the daytime are mainly concentrated in high-density blocks (e.g., LRHD blocks, MRHD blocks, and HRHD blocks), a phenomenon that suggests that the thermal effect during the daytime is mainly influenced by building density. The LRLD blocks had the lowest average LST at 08:26 (28.536 °C), while the lowest average LST occurred in the HRLD blocks at 13:30 (37.759 °C). This variation may be related to building heights, green space coverage, and shading in different blocks. Higher LST in high density areas during the day suggests that increased building density may lead to higher temperatures. In addition, there is greater temperature dispersion during the day, especially at 13:30, with a wider range of temperature variations in high-density areas, showing enhanced temperature variability between blocks.

The maximum temperatures during the night (21:55 and 00:59) were significantly lower, with maximum values of about 28.642 °C and 26.375 °C at 21:55 (Figure 4d) and 00:59 (Figure 4a), respectively. This temperature variation reflects significant differences in insolation intensity between day and night. At night, the HRHD blocks had the highest LSTs of 27.806 °C and 24.623 °C, while the LRLD blocks had the lowest LSTs of 27.163 °C and 23.441 °C, respectively. This suggests that high-rise, high-density building areas have higher temperatures due to the slow release of heat accumulated during the day at night. Low-rise, low-density blocks with more open space dissipate heat more quickly, helping to lower nighttime temperatures. At night, temperatures were more concentrated and less variable, especially at 21:55 (Figure 4d), with less temperature variability across all blocks, suggesting a more balanced distribution of nighttime heat among different blocks. The high-temperature areas at night are concentrated in high-rise blocks (e.g., HRHD, HRLD, etc.),while the low-temperature areas are concentrated in low-rise blocks (LRHD, LRMD, and LRHD, etc.) and thus are more affected by building heights and heat retention effects at night.

Overall, this study found significant differences in LST across blocks at different times of day, with differences in thermal patterns between daytime and nighttime reflecting the different roles of building density and height in the formation of the thermal environment. Hot areas during the day are concentrated in high-density blocks, while, at night, they are more concentrated in high-rise blocks. These findings highlight the influence of building density and height on the urban thermal environment at different times of the day, especially the role of heat retention effects at night, and provide an important basis for further understanding and optimizing the distribution of the thermal environment in urban planning.

3.3. Performance of GWR and MGWR Models

In this study, by selecting R² and AIC values as the criteria for evaluating the model’s performance, in general, a higher R² indicates that the model explains a larger proportion of the variance in the dependent variable (LST) and the model fit is better; a lower AIC value means that the model implies a better balance between model complexity and goodness-of-fit. As shown in Figure 6, the values of R² for the GWR and MGWR models range from 0.045 to 0.779 and 0.323 to 0.814 (mean 0.344 and 0.610), respectively, indicating that it is possible to explain 34.4% and 61.0% of the effect of urban characterization factors on LST on average for the two models, respectively. The AIC values for GWR took a range of 70.386–1960.800 (mean 560.001), while the AIC values for MGWR took a range of 20.257–1858.443 (mean 503.852). The MGWR model always outperforms the GWR model.

There are significant differences in the performance of different models at different observation times. The mean value of R² for the GWR model reaches a maximum value of 0.375 at 13:30, followed by 08:26 (0.362), then 00:59 (0.353), and it reaches a minimum value of 0.284 at 21:55; whereas the mean value of R² for the MGWR model reaches a maximum value of 0.696 at 00:59, and at 08:26 and 13:30, respectively, the R² were 0.646 and 0.621 and reached a minimum value of 0.590 at 21:55. The mean value of AIC for GWR reached a minimum value of 554.650 at 13:30, followed by 556.149 and 557.115 at 00:59 and 08:26, respectively, and a maximum value of 572.091 at 21:55, whereas the mean value of AIC for MGWR reached a minimum value of 472.180 at 00:59, followed by 500.128 and 512.919 at 08:26 and 13:30, respectively, and reached a maximum value of 530.179 at 21:55.

Overall, the MGWR can better explain the degree of influence of the independent variables on LST, so this paper chooses the MGWR model for subsequent analyses, and it should be noted that the bandwidth ranges need to be manually selected due to the small sample sizes of a few blocks such as LRHD block.

3.4. Spatial Heterogeneity in the Effects of Urban 2D and 3D Landscape Factors on Diurnal LST Across Block Types

First, the spatial autocorrelation hypothesis was tested using Global Moran’s I. As shown in

Table 4. The results of the Global Moran’s I spatial autocorrelation test for each block and time dimension.

Time	LRLD	LRMD	LRHD	MRLD	MRMD	MRHD	HRLD	HRMD	HRHD
00:59	0.559	0.779	0.586	0.671	0.706	0.733	0.628	0.738	0.730
08:26	0.563	0.663	0.792	0.681	0.662	0.716	0.528	0.589	0.669
13:30	0.522	0.512	0.696	0.505	0.687	0.670	0.536	0.599	0.610
21:55	0.617	0.554	0.501	0.526	0.648	0.592	0.587	0.745	0.652

, the Global Moran’s I values for all observation times and blocks were positive and higher than 0.5, indicating that LSTs in the study area have significant positive spatial autocorrelation. This implies that LST exhibits distinct spatial clustering characteristics in different blocks and time points, i.e., high- and low-temperature regions tend to cluster spatially, respectively. In particular, the spatial aggregation effect of LST is more pronounced in high-density blocks (e.g., MRHD blocks and HRHD blocks), reflecting the more concentrated heat island effect in these areas. In addition, the spatial clustering phenomenon is more significant at 00:59 and 08:26 over time, suggesting that the spatial distribution of LSTs exhibits dynamically changing spatial dependence across time.

There were significant differences in the effects of urban characterization factors on LST for different block types at different observation times, reflecting the role of complex urban features in regulating the thermal environment (Figure 7). Overall, the building indicators (PB, FAI, BH, SVF, LPI_bu) and water indicators (PW, AI_w, LPI_w, ED_w) had the greatest impact on the thermal environment of the blocks, followed by the impervious surface (other than buildings) indicators (PI, LPI_i, ED_i) and the bare ground indicators (AI_ba, LPI_ba, ED_ba), and the vegetation indicators (AI_v, LPI_v, POP) and socio-economic indicators (AHF, POP) had the least significant direct impacts on LST in each block.

In the LRLD blocks, the top three key influences at 00:59 were FAI, PB, and ED_w, with regression coefficients of 0.353, −0.326, and 0.206, respectively, suggesting that ventilation conditions and building-specific gravity were positively correlated with LST, while the boundary complexity of the water body was positively correlated with LST. AHF, ED_w, and LPI_v were significant factors affecting LST at 08:26, with regression coefficients of 0.309, 0.305, and −0.226, respectively, which suggests that anthropogenic heat fluxes and the degree of water-body boundary fragmentation had a positive effect on LST, while vegetation patch dominance had a negative effect on LST. At 13:30, the top three key factors were AHF (0.329), ED_w (0.284), and ED_i (−0.277), similarly reflecting the fact that anthropogenic heat fluxes and the degree of water-body boundary fragmentation were positively correlated with the thermal environment and that the degree of fragmentation of impervious surfaces was negatively correlated with LST, while PB (0.503), AI_ba (0.407), and ED_i (0.325) were the key factors at 21:55, highlighting that building-specific gravity, aggregation of bare ground, and boundary complexity of impervious surfaces had a positive effect on LST.

In LRMD blocks, the top three key factors at 00:59 were PI, AI_v, and LPI_i, with regression coefficients of 0560, −0.389, and −0.355, respectively, implying that the weight of impervious surfaces exacerbates the elevation of LST, whereas the degree of vegetation agglomeration and the dominance of bare ground at this time are negatively related to LST. At 08:26, PI (0.663), LPI_i (0.340), and ED_w (0.298) were the key factors influencing LST, and similarly, the specific gravity of impervious surfaces, bare ground dominance, and boundary complexity of water bodies were positively associated with LST. At 13:30, the most influential factor was PI (0.684), followed by AI_v (−0.362), and again PW (0.313), which implies that the weight of impervious surfaces and buildings has a positive effect on the thermal environment, whereas the agglomeration of vegetated landscapes is negatively correlated with LST. AI_ba (−0.304), LPI_bu (0.300), and PB (−0.289) at 21:55 were the key factors at this time, suggesting that the dominance of architectural patches had a positive effect on LST and that the aggregation of the bare landscape and the proportion of buildings had a negative effect on LST.

For the LRHD blocks, ED_w, LPI_i, and LPI_ba were the key factors affecting LST at 00:59, with regression coefficients of −0.706, −0.680, and −0.673, respectively, reflecting the fact that fragmentation of water-body patches, impervious surfaces, and dominance of bare ground were negatively associated with LST. At 08:26, LPI_w (0.864), BH (0.714), and PB (0.662) had a positive effect on the thermal environment. PB (0.709), BH (0.701), and LPI_v (0.692) at 13:30 were positively correlated with LST. At 21:55, LPI_bu, PB, and POP were the key factors with regression coefficients of 0.531, 0.431, and 0.428, respectively, which indicates that the dominance and weight of architectural patches as well as the number of population are positively correlated with LST.

In the MRLD blocks, PI, ED_i, and SVF were the key influencing variables of LST at 00:59, with regression coefficients of 0.321, 0.243, and −0.195, respectively, implying that the proportion of impervious landscapes and the edge complexity had a positive effect on the LST, while sky openness was negatively correlated with the LST. At 08:26, PI (0.366) was positively correlated with the thermal environment, while LPI_w (−0.258) and LPI_i (−0.245) were negatively correlated with LST. At 13:30, AI_ba, ED_w, and AI_w were the top three key factors affecting LST, with regression coefficients of 0.23, −0.215, and 0.164, respectively, which suggests that the agglomeration of bare land and water-body landscapes has a positive effect on LST, and edge complexity of water-bodypatches is negatively related to LST. By 21:55, LPI_bu (0.416), LPI_w (0.401) are the key factors positively correlated with LST, and LPI_ba (−0.302) is the key variable negatively correlated with thermal environment.

In MRMD blocks, POP (0.300), ED_w (0.284), and AI_w (−0.233) were the key variables affecting LST at 00:59, reflecting that population size and the degree of edge fragmentation of the water body had a positive effect on LST, whereas the degree of aggregation of water patches was negatively correlated with LST. At 08:26, POP, AI_ba and ED_ba were the key factors, and the regression coefficients were 0.142, −0.121 and 0.118, respectively, indicating that the number of population and the degree of aggregation of bare land were positively correlated with the LST, and the edge fragmentation of bare land patches was negatively correlated with the LST. At 13:30, PI (0.216), POP (0.131) and ED_w (0.11) were significant variables positively associated with LST. POP, AI_w and ED_w were the key variables affecting LST at 21:55, with regression coefficients of 0.361, −0.231 and 0.215, respectively, implying that the population gauge and the degree of edge fragmentation of water patches were positively correlated with LST, and the degree of clustering in the water landscape was negatively correlated with LST.

For MRHD blocks, FAI (0.287), PI (−0.147), and LPI_ba (−0.110) were the key factors at 00:59, indicating that ventilation conditions, impervious surface specific gravity, and dominance of bare ground patches were negatively correlated with LST. FAI (0.290), POP (0.180) and LPI_ba (−0.104) were the key variables affecting LST at 08:26, indicating that ventilation conditions and dominance of bare ground patches negatively affected LST, and population had a positive effect on LST. At 13:30, FAI (0.357) and LPI_i (0.125) were the key variables with positive effects on LST, while BH (−0.123) was the key factor negatively associated with LST. At 21:55, FAI (0.105), LPI_ba (−0.1101) and ED_ba (0.083) were the key factors, implying that ventilation conditions and dominance of bare ground landscape were negatively correlated with LST, while the edge complexity of bare ground was positively correlated with LST.

In the HRLD blocks, LPI_ba, ED_w, and AI_w were the key factors affecting the thermal environment at 00:59, with regression coefficients of −0.888, 0.803, and −0.780, respectively, indicating that the dominance of bare ground patches and the degree of agglomeration of the water body were negatively related to LST, while the edge complexity of the water-body landscape had a positive effect on LST. At 08:26, AI_v (−0.564) and AI_w (−0.8) are the key variables negatively affecting the thermal environment, and LPI_v (−0.787) is the key factor positively correlated with LST. At 13:30, PI (0.734), LPI_bu (−0.782), and FAI (0.705) were the key variables affecting LST, indicating that the weight of impervious surfaces was positively associated with LST, and that the dominance of built-up patches and the ventilation conditions had a negative effect on LST. At 21:55, ED_ba (0.839), LPI_ba (−0.696), and LPI_i (0.688) were the key variables affecting LST, suggesting that edge complexity and dominance of impervious surface patches on bare ground were positively correlated with LST, while bare ground dominance was negatively correlated with LST.

For HRMD blocks, LPI_v (−0.822), FAI (0.484), and AI_v (0.329) played a key role in the effect of LST at 00:59, implying that water-body landscape dominance and ventilation conditions had a negative effect on LST, while the degree of aggregation of the vegetated landscape had a positive effect on LST. At 08:26, POP (0.375), ED_w (0.338), and LPI_w (−0.337) were the key factors, indicating that the population size and the edge complexity of the water body were positively correlated with the LST, while the dominance of the water-body patches was negatively correlated with the thermal environment. At 13:30, LPI_w (−0.560), PB (0.396), and POP (0.330) were the key variables, implying that the dominance of water-body patches was negatively correlated with the LST, whereas the building weight and population size were positively correlated with the LST. At 21:55, POP (0.428) remained the key factor positively associated with the thermal environment, while LPI_w (−0.233) and LPI_bu (−0.214) were the key variables with a negative effect on LST.

Among the HRHD blocks, POP (−0.306), ED_w (−0.226), and LPI_i (0.151) were the key variables affecting LST at 00:59, indicating that population size and edge complexity of the water body were negatively correlated with LST and that dominance of impervious surface patches was positively correlated with LST. At 08:26, ED_ba, LPI_ba, and POP were the key variables affecting the thermal environment, with regression coefficients of −0.365, 0.314, and 0.145, respectively, which implies that the complexity of the bare land landscape edge was negatively related to LST, while the dominance of the bare land patches and the size of the population had a positive effect on LST. At 13:30, POP (−0.387), LPI_w (−0.277), and BH (−0.254) were the significant factors with negative effect on thermal environment. At 21:55, LPI_w (−0.190), PW (0.144), and PI (−0.142) were the key variables affecting LST, indicating that the dominance of water-body patches and the weight of impervious surfaces were positively correlated with LST, while the weight of the water-body landscape was negatively correlated with LST.

3.5. Analysis of Regression Coefficients of Dominant Factors Affecting Diurnal LST in Different Block Types

The spatial distribution of the regression coefficients of the 2D and 3D urban characterization factors that have a primary effect on the LST in different blocks at different observation times is shown in Figure 8. It should be noted that the primary factor is the maximum regression coefficient factor that is selected under the condition of satisfying at p < 0.05 (or p < 0.1), and if there is no factor with p < 0.05, then the factor is screened with p < 0.1 as the condition.

In the LRLD blocks, at 00:59, the effect of FAI on LST increases roughly from the center to the perimeter, with regression coefficients ranging from −0.319 to 0.918, and increased windward floor space in the northwest area and increased ventilation in the southeast area are beneficial in reducing LST. The regression coefficients of AHF at 08:26 and 13:30 ranged from 0.087 to 0.554 and 0.313 to 0.341, respectively, the spatial heterogeneity of the effect of AHF on LST at 13:30 was smaller compared to that at 08:26, and the reduction in anthropogenic heat fluxes was favorable to reduce LST. At 21:55, the PB regression coefficient ranges from −0.773 to −0.109, and increasing the weight of building patches is beneficial in reducing LST.

In LRMD blocks, the regression coefficients of PI at 00:59, 08:26, and 13:30 were 0.489–0.616, 0.284–0.787, and 0.156–0.788, respectively, the spatial distribution non-stationarity of the positive correlation of PI to LST showed a gradual increase at the above three observation times, and the reduction in the proportion of impervious surfaces was conducive to the reduction in LST. At 21:55, the AI_ba regression coefficient (−0.365–−0.259) increased from west to east, and increasing the aggregation of bare ground patches was conducive to reducing LST.

For LRHD blocks, the degree of spatial heterogeneity in the effect of LPI_ba (−0.727–−0.635) on LST was small at 00:59, when increasing the dominance of bare ground patches favored LST reduction. BH (0.606–0.895) is the dominant factor influencing the thermal environment at 08:26, with the degree of influence increasing from south to north, suggesting that lowering the height of the building is conducive to lowering the LST. LPI_v (0.550–0.914) was the dominant factor at 13:30, and the regression coefficient showed a spatial pattern of increasing from north to south. At 21:55, the POP impact level is 0.323–0.503, increasing roughly from south to north, and reducing the population size at this time would be beneficial in mitigating the thermal environment.

For the MRLD blocks, the extent of PI influence on LST at 00:59 and 08:26 was 0.238–0.385 and 0.348–0.392, respectively, and there was less spatial heterogeneity in the extent of PI influence across the regions at these two observation times. The extent of the effect of AI_ on LST at 13:30 ranged from 0.197 to 0.261, with small spatial differences in the regression coefficients. The regression coefficients for LPI_bu as the dominant factor influencing LST at 21:55 range from 0.278 to 0.564, with the influence increasing from south to north.

In the MRMD blocks, POP was the dominant factor for LST at 00:59, 08:26, and 21:55, with regression coefficients of 0.285–0.312, 0.112–0.175, and 0.354–0.377, respectively. And at 13:30, PI (0.143–0.283) was the dominant factor. Meanwhile, the spatial distribution of the degree of influence of POP and PI on LST is smoother in the corresponding time period, and reducing the population size and the proportion of impervious surface can effectively reduce LST.

In the MRHD blocks, the extent of FAI influence on LST at 00:59, 08:26, 13:30, and 21:55 was 0.280–0.298, 0.2180.526, 0.231–0.726, and −0.267–0.369, respectively. It is clear that the positive correlation of FAI on LST also shows a pattern of first strengthening and then weakening for most of the MRHD blocks during the above four time periods, and that optimizing ventilation conditions is beneficial for reducing LST.

For the HRLD blocks, the spatial smoothness of the degree of influence of FAI (0.639–0.786) on LST was higher at 00:59, and optimizing ventilation conditions facilitated a significant reduction in LST. At 08:26, the effect of LPI_w (−0.744–−0.496) on LST was enhanced from north to south, and increasing the dominance of water-body patches was in favor of decreasing LST. At 13:30, the regression coefficient of LPI_bu ranges from −0.813 to −0.724, with high spatial smoothness, and increasing the dominance of building patches is beneficial in reducing the LST at this time. The degree of influence of AHF (0.439–0.849) on LST at 21:55 showed a spatial pattern of enhancement from north to south, and the reduction in anthropogenic heat fluxes was beneficial to mitigate LST.

In the HRMD blocks, the regression coefficients for FAI at 00:59 ranged from 0.370 to 0.575, with the magnitude of the effect increasing roughly from south to north. The range of regression coefficients for POP was 0.308–0.407 and 0.380–0.498 at 08:26 and 21:55, respectively, with high spatial smoothness. At 13:30, LPI_w (−0.580–−0.548) has a lower degree of spatial non-stationarity on LST.

In the HRHD blocks, the regression coefficients of POP were divided into −0.848–0.511 and −0.847–0.405 at 00:59 and 13:30, respectively, and the degree of POP’s influence on the LST at these two observation times showed a pattern of change that was increasing from the middle to the perimeter, and the spatial non-stationarity was high.

ED_ba (−0.431–−0.297) and PW (0.126–0.160) were the dominant factors at 08:26 and 21:55, respectively, and both had less spatial heterogeneity in the degree of influence on LST. For most of the area within the blocks, LST reduction can be facilitated by increasing the fragmentation of bare ground edges and decreasing the specific gravity of the water body at 08:26 and 21:55, respectively.

4. Discussion

4.1. Block Scale, Urban Characterization Factors, and MGWR Model

In most of the similar studies, the research scale mainly focuses on grid cells, ignoring the subtle effects of urban functional zoning, building layout, and transportation network [20,30]. This study breaks away from the limitations of traditional grid cells and uses the block cell scale for analysis (

Table 5. Relevant literature on the relationship between urban characterization factors and diurnal LST.

Study Area	Climate	Basic Research Units	Observation Time	Landscape Type	LST Data	Research Methods	Block Types	LST Characteristics	Dominant Factors Influencing LST (SUHI)
Fuzhou [40]	subtropical monsoon climate	blocks	02:12 07:04 10:18 19:32	BD, NDVI, NDISI, MNDWI, BH, SVF, FAR, FAI	ECOSTRESS LST	Pearson’s correlation and stepwise regression analysis	LRLD LRMD LRHD MRLD MRMD MRHD HRLD HRMD HRHD	LST(10:18) > LST(19:32) > LST(07:04) > LST(02:12)	LRLD: MNDWI (+) (02:12), NDISI(+) (07:04), NDISI (+) (10:18), FAI(+) (19:32) MRMD: MNDWI(+) (02:12), NDVI(-) (07:04), NDVI(−) (10:18), FAI(−) (19:32) HRHD: MNDWI(+) (02:12), NDVI(−) (07:04), FAR(−) (10:18), NDVI(−) (19:32)
Fuzhou [52]	subtropical monsoon climate	grids(90, 150, 270, 390, 570 m)	02:12 07:04 10:18 17:35	PI, PB, PV, PW, PD, ED, Cohe, LPI, RD, AHF, PopS, BH, NBH, SVF	ECOSTRESS LST	RF	-	LST(10:18) > LST(17:35) > LST(07:04) > LST(02:12)	270m:PW(02:12), LPI_BGS(07:04), LPI_B(10:18), PI(17:35)
Beijing [57]	temperate monsoon climate	blocks	10:42 14:13 22:32 03:00 07:20 19:06	PLAND, LPI, AI, BD, BH, TD, TH, SVF	ECOSTRESS LST	BRT	-	Basically LST (natural geography) < LST (artificial surface)	BD(10:42), TD(14:13), BH(22:32), BH(03:09), BH(07:20), SVF(19:06)
Beijing [58]	temperate monsoon climate	blocks	10:42 14:13 22:32 03:09	PLAND, LSI, AI, BD, BH, SVF, FAI	ECOSTRESS LST	RF BRT	LRB MRB HRB	(10:42)MRB > HRB > LRB (14:13)MRB > LRB > HRB (22:32)HRB > MRB > LRB (03:09)HRB > MRB > LRB	LRB:PLAD_v(10:42), PLAND_v(14:13), PLAND_i(22:32), FAI(03:09) MRB: PLAD_v(10:42), PLAND_v(14:13), BH(22:32), BH(03:09) HRB:BD(10:42), BD(14:13), FAI(22:32), LSI_i(03:09)
Beijing [59]	temperate monsoon climate	grids	06:10 10:42 14:13 17:14 19:52 22:57 00:59 04:04	PV, PW, BD, ED, PD, BH, SVF, POID, Albedo	ECOSTRESS LST	ANN GBM MARS MLR RF	-		BD(+) (10:42), PV(−) (14:13 to 17:14), POID(+) (19:52 to 22:57), Albedo(−) (00:59 to 06:10)
Dalian [60]	temperate monsoon climate	blocks	day night	NDVI, NDBI, BD, FAR, AH, URL, SVF, AR	Landsat 8 LST	MGWR	built-up types, non-built-up types	LST is higher during the day than at night in all divisions except LCZG	LST-Daytime: SVF(−), FAR(−), AH(−), NDVI(−), POI(+), NL(+), DFC(+), NDBI(+), BD(+) LST−Nighttime: SVF(−), BD(−), AH(−), NDVI(−), NDBI(−), DFC(+), POI(+), NL(+), BD(+)
the Guanzhong region [61]	humid sub tropical (Cwa) and semi-arid (BSk) climate	pixel	10:22 11:50 14:13 14:30 16:23 16:42 16:55 17:32 23:44 23:57 04:25 06:25	MAT, Prec, Urban size, Srad, Wind, POP, UDEM, GDP, PI, NDVI	ECOSTRESS LST Landsat ASTER MODIS	Pearson’s correlation	-	Higher LST in urban areas (especially Xi’an and Xianyang urban areas), lower LST in vegetated areas	Night SUHI: NDVI(−), UDEM(−), Prec(−), Wind(−), PI(+), MAT(+)
Tianjin (our study)	temperate monsoon climate	blocks	00:59 08:26 13:30 21:55	PB, PI, PW, AI, LPI, ED, BH, SVF, AHF, POP, FAI	ECOSTRESS LST	MGWR	LRLD LRMD LRHD MRLD MRMDMRHD HRLD HRMD HRHD	LST(13:30) > LST(08:26) > LST(21:55) > LST(00:59)	LRLD: FAI(±)(00:59), AHF(+)(08:26), AHF(+)(13:30), PB(−)(21:55) MRMD: POP(+)(00:59), POP(+)(08:26), POP(+)(13:30), POP(+)(21:55) HRHD: POP(±)(00:59), ED_ba(−)(08:26), POP(±)(13:30), PW(+)(21:55)

). Linear elements such as roads and rivers are used as boundaries, and each block is considered a separate thermal zone. The block unit is used as a basic unit in urban planning, and its thermal environment characteristics are closely related to residents’ living comfort, energy consumption, and the formation of the heat island effect [31]. Some studies have also used blocks as the unit of study but ignored the heterogeneity of block categories on thermal patterns [48]. In this paper, blocks are classified and discussed according to building height and density, which can better reflect the differences in the thermal environment of different blocks within the city and provide a strong theoretical basis for improving urban planning, optimizing spatial layout, and mitigating the heat island effect [32]. So, it makes sense to use block units for the analysis.

In this study, 2D factors (e.g., landscape components, landscape pattern index, socioeconomic factors) and 3D factors (e.g., meteorological factors, urban canyon parameters) were used to explore the influence of the urban thermal environment, and the combination of the 2D and 3D factors was analyzed, which was able to reveal the complexity of the urban thermal environment in a more comprehensive way and to identify the key factors affecting the urban surface temperature. First, the topography and structure of the city affect the thermal environment in many ways. In urban areas with rapid population growth, the physical properties of the ground surface are significantly altered, and the expansion of impervious surfaces and anthropogenic heat emissions have become the main drivers of urban warming [33]. Second, urban form has an important impact on the thermal environment. The urban heat island effect at night is significantly influenced by the geometry of urban canyons [23]. The spatial layout in urban design, such as the arrangement of streets and buildings, plays a key role in regulating the internal temperature of the city [6]. In addition, meteorological factors (e.g., wind speed, cloudiness, and relative humidity) regulate the distribution of heat through mechanisms such as radiative exchange, evaporative dispersion, and turbulence [34], while urban form factors (e.g., building area and height) determine the intensity and extent of these processes [35], which together influence the formation and development of the thermal environment. Therefore, combining 2D and 3D factors can fully reflect these multidimensional interaction mechanisms and reveal the complexity of the urban thermal environment in greater depth. Finally, blue–green spaces play a key role in mitigating urban thermal environments. By increasing the area of vegetation and water bodies, the blue–green space is able to effectively reduce the temperature of the surrounding area, creating a relatively cool area. The cooling effect of such spaces is limited by a variety of factors, such as the size, shape, and vegetation cover of the space [36,37,38]. Therefore, it is essential to incorporate a 2D landscape characterization of blue–green spaces in thermal environmental studies, which can help to further understand their role in urban temperature regulation. In summary, this multi-dimensional analysis method provides a scientific basis for urban planning, helps to optimize spatial layout, alleviates the urban heat island effect, and improves the quality of life of the residents, which has significant rationality and application value.

While machine learning and linear analysis are more widely used in exploring the relationship between urban form and surface temperature, they ignore spatial non-stationarity [40,43,44,45]. In this study, special attention is paid to the study of spatio-temporal non-stationarity, GWR and MGWR models are used to analyze the effect of city characteristic factors on diurnal LST, and it is found that the MGWR outperforms the GWR. By choosing the optimal bandwidth for each covariate and allowing different processes to operate at different spatial scales, MGWR is able to better capture the complexity and diversity of the impact of urban characterization factors on LST [48]. And it shows higher flexibility and accuracy when dealing with urban data with complex spatial structure. In studying the relationship between urban form and LST, the MGWR model is able to reveal the influence of urban form indicators on LST at different bandwidth scales, which provides a more fine-grained guidance for urban planning and heat island effect mitigation. Therefore, it is reasonable to use the MGWR model to study the effect of urban characterization factors on diurnal LST at the city block scale.

4.2. Effect of Urban Characterization Factors on Diurnal LST in Different Blocks

This study analyzes the effect of 2D and 3D urban characterization factors on the thermal environment in nine blocks at four observation times: 00:59, 08:26, 13:30, and 21:55. We found that the building indicators (PB, FAI, BH, SVF, LPI_bu) and the water indicators (PW, AI_w, LPI_w, ED_w) had the greatest impact on the thermal environment in each block, probably due to the fact that the higher number of these two types of indicators makes their impact on the thermal environment more direct. Impervious surface (other than buildings) indicators (PI, LPI_i, ED_i) and bare ground indicators (AI_ba, LPI_ba, ED_ba) are next in importance. Vegetation indicators (AI_v, LPI_v) and socio-economic indicators (AHF, POP) had the least significant direct impacts on the LST in the blocks, probably due to the fact that there were fewer of these two types of indicators and less socio-economic variations in the study area, which made the correlation between these indicators and the thermal environment less significant.

In this study, it was found that during the daytime, high-temperature areas are mainly concentrated in high-density blocks (e.g., LRHD, MRHD, and HRHD), where the high-density spatial structure restricts air circulation and the natural cooling process at ground level, and the presence of high building mass enhances UHI [40], a phenomenon that suggests that the thermal effect during the daytime is mainly affected by the building density. The high-temperature areas at night are mainly concentrated in high-rise blocks (e.g., HRHD, HRLD, etc.), and the low-temperature areas are mainly concentrated in low-rise blocks (LRHD, LRMD, and LRHD, etc.). The heat retention effect of the high-rise blocks makes the heat accumulated during the daytime more slowly released at night, indicating that the nighttime is more affected by the height of the buildings. The above findings are consistent with existing studies [57].

It is worth noting that building height does not always increase LST, and the effect of BH on LST is different during daytime and nighttime; it is negatively correlated with daytime temperatures but increases nighttime temperatures, consistent with existing research [58]. Therefore, even the same urban characterization factor may have opposite effects on the thermal environment at different observation times and block types, which is consistent with existing literature findings [40]. In addition to BH, other urban characterization factors with dynamic variability were found in this study. First, FAI had mainly positive effects on the thermal environment at all observation times but was positively or negatively correlated with LST at 00:59 for the LRLD blocks and 21:55 for the MRHD blocks, with mainly negative correlations in the northern region and the opposite in the southern region, which may be due to the differences in the ventilation conditions of the different regions of the city, which in turn affects the performance of the thermal environment at night [62]. Similarly, POP is usually positively correlated with the thermal environment, especially in high-density blocks, where dense population is often accompanied by more buildings, vehicles, and human activities, which in turn leads to more heat accumulation; however, HRHD blocks at 00:59 and 13:30 are positively or negatively correlated with LST, with a negative correlation in the central region and the opposite in the surrounding area, and the phenomenon may be related to several factors, mainly including urban structure, population distribution and time differences, etc. Since this paper uses fixed population data rather than population size characteristics extracted according to each observation time, the spatial agglomeration characteristics of the population may be somewhat different from the actual situation. PW mainly plays a cooling role for different blocks at all observation times because water bodies have a high specific heat capacity to absorb and store heat, especially during the daytime through the evaporative cooling effect, and water bodies can significantly reduce the surrounding temperature [63]. However, at 21:55 PW had a positive effect on the thermal environment, due to the fact that during the night the cooling effect of the water body diminishes and may instead exhibit a heat retention effect. In addition, LPI_bu mainly exerts a positive effect on the thermal environment of each block at different observation times, which is due to the fact that the heat absorbed and reflected by the buildings will increase the temperature of the surrounding air, leading to an increase in LST; however, it exerts a negative effect on the HRLD blocks at 13:30 hrs, which may be due to the fact that the open space and the green space in the HRLD blocks provide a stronger cooling effect to offset the heat accumulation of the buildings. effect, or the layout and microclimatic characteristics of the blocks may make the thermal effect of the buildings produce a cooling effect instead. AI_ba usually has a positive effect on the thermal environment, for example, it is positively correlated with the MRLD blocks at 13:30 because the bare ground surface usually absorbs heat faster and lacks cooling moderators such as vegetation, trees, or bodies of water [37]; however, AI_ba is negatively correlated with the LRMD blocks at 21:55, which may be due to the rapid heat dissipation effect of the bare ground at night that makes the local temperature drop. The above illustrates that the influence of urban characteristic factors on LST is not only complex but also interactive and dynamic, with large spatial non-smoothness in the thermal environmental influences of each block at different observation times, depending on the density of the block, the layout of the buildings, and the way of combining natural landscapes (e.g., water bodies and vegetation).

4.3. Implications for Urban Planning

Based on the influence of multidimensional factors on thermal patterns in different blocks under different observation times, it provides a scientific basis that helps to optimize spatial layout, mitigate the urban heat island effect, and enhance the quality of life of residents, which has significant rationality and application value. For LRLD blocks, anthropogenic heat flux should be reduced, dominance of vegetation patches increased, and impervious surface fragmentation increased during the day, while at night, ventilation should be optimized, building weight increased, bare ground landscape agglomeration and impervious surface edge complexity reduced, and the totality of the water body’s landscape should be improved both day and night. In LRMD blocks, impervious landscape dominance as well as water fragmentation and specific gravity should be reduced during the day and the aggregation of vegetated patches should be increased; reducing the maximum building patch size and increasing the aggregation of bare ground and vegetated landscapes with the specific gravity of building patches is conducive to the mitigation of nighttime thermal environments; and the specific gravity of impervious surfaces should be reduced in both the day and nighttime. For LRHD blocks, mitigation of the daytime thermal environment can be achieved by increasing water-body patch edge complexity and impervious surfaces as well as bare ground dominance, decreasing building dominance and population size; whereas at night, water and vegetation patch dominance and building heights should be reduced; and both day and night can be achieved by decreasing the building weight to mitigate LST. In MRLD blocks, the proportion of impervious surfaces needs to be reduced both day and night to mitigate LST; during the day, the maximum patch area of waterbodies and impervious surfaces and the fragmentation of waterbody landscapes should be increased, and the aggregation of bare ground and waterbody landscapes should be reduced; thermal conditions can be mitigated by increasing the area of the sky view and the maximum patch area of bare ground, by decreasing the fragmentation of impervious surfaces, and by dominating the patch area of buildings and waterbodies. In MRMD blocks, population size and water-body patch fragmentation should be reduced during both day and night; reducing the aggregation and fragmentation of bare ground patches and the proportion of impervious surfaces can be effective in reducing LST during the day, and water-body slab aggregation should be enhanced during the night. For MRHD blocks, ventilation should be optimized and the share of bare ground dominant patch area should be increased both day and night; higher building heights, smaller population sizes, and a share of impervious surface dominant patches can mitigate the thermal environment during the day; and a higher share of impervious surfaces and lower bare ground fragmentation can contribute to a lower nighttime LST. In HRLD blocks, increasing the maximum patch weight of bare ground mitigates the diurnal thermal environment; during the day, it reduces the weight of impervious surfaces and the maximum patch area as well as the fragmentation of bare ground patches, optimizes ventilation, and reduces the weight of the maximum patch area of buildings. For HRMD blocks, the thermal environment can be mitigated both day and night by reducing the population size and increasing the proportion of dominant patch area of water bodies; during the day, the fragmentation of water landscapes and the proportion of buildings should be reduced; and during the night, the proportion of the largest patch area of vegetation and buildings can be increased, the ventilation conditions can be optimized, and the agglomeration of vegetated landscapes can be reduced. In HRHD blocks, LST can be reduced during the day by increasing the fragmentation of bare ground patches and the maximum patch specific gravity of the water body, as well as building heights, and decreasing the dominance of bare ground patches; and during the night, it can be reduced by increasing the fragmentation and dominance of water-body patches and the specific gravity of impervious surfaces and by decreasing the maximum patch specific gravity of impervious surfaces and the specific gravity of the water body in order to mitigate the thermal environment.

4.4. Limitations and Outlook

Although this study reveals the diurnal variation pattern of LST and its influencing factors in different blocks of the city, there are still some limitations. First, ECOSTRESS data are acquired from different dates, and although they have been corrected using geostationary satellites, the effects of weather and other uncertainties may still not be completely eliminated. Second, the relatively low resolution of the ECOTRESS data, anthropogenic heat flux data, and demographic data may not be able to capture the specific effects of influencing factors on LST in microclimatic environments in a nuanced way when capturing fine-scale heterogeneity for the MGWR, which may affect an accurate assessment of the actual magnitude of the impacts of these variables. Third, the population data used in this paper are fixed data, and future studies should extract data characterizing the spatial agglomeration of the population under the corresponding observation time. Fourth, this study only analyzed the LST changes at several key time points, while the dynamics of the urban thermal environment is more complex; future studies can combine the more time-continuous ECOSTRESS satellite data to further explore the diurnal change rule of LST. In addition, this study is mainly based on the existing 2D and 3D landscape factors, and future studies can introduce more socio-economic factors and meteorological factors, such as traffic flow and humidity, to further analyze the effects of these factors on LST. Meanwhile, as urban climate change intensifies, future research should also focus on the potential trends in LST changes under different climate scenarios to provide a basis for developing more resilient urban planning strategies.

5. Conclusions

In this study, we explored the effects of 2D and 3D urban characterization factors on diurnal LST in different blocks through MGWR, and the results were as follows: the highest LST for the four observation times of the day occurred at 13:30, followed by 08:26, the next lowest temperature at 21:55, and the lowest at 00:59. High-temperature zones during the day are concentrated in high-density blocks (e.g., LRHD blocks, MRHD blocks, and HRHD blocks), a phenomenon that suggests that the thermal effects during the day are primarily influenced by building density. The high-temperature areas at night are concentrated in high-rise blocks (e.g., HRHD, HRLD, etc.),while the low-temperature areas are concentrated in low-rise blocks (LRHD, LRMD, and LRHD, etc.) and thus are more affected by building heights and heat retention effects at night. We found that the building indicators (PB, FAI, BH, SVF, LPI_bu) and water indicators (PW, AI_w, LPI_w, ED_w) had the greatest degree of influence on the thermal environment of the blocks, with the impervious surface (other than buildings) indicators (PI, LPI_i, ED_i) and the bare ground indicators (AI_ba, LPI_ba, ED_ba) being the next most significant, and the vegetation indicators (AI_v, LPI_v) and socio-economic indicators (AHF, POP) had the least significant direct impacts on LST in each block. We also find that the effects of urban characterization factors on LST exhibit significant spatial non-stationarity across different observation times and block types, reflecting the moderating effect of complex urban structures on the thermal environment. The results of this study can provide theoretical references for analyzing the spatial heterogeneity of the effects of urban multidimensional factors on diurnal LST and mitigating the thermal environment of similar high-density urban center areas.