Seasonal Variations of Daytime Land Surface Temperature and Their Underlying Drivers over Wuhan, China

: Rapid urbanization greatly alters land surface vegetation cover and heat distribution, leading to the development of the urban heat island (UHI) effect and seriously affecting the healthy development of cities and the comfort of living. As an indicator of urban health and livability, monitoring the distribution of land surface temperature (LST) and discovering its main impacting factors are receiving increasing attention in the effort to develop cities more sustainably. In this study, we analyzed the spatial distribution patterns of LST of the city of Wuhan, China, from 2013 to 2019. We detected hot and cold poles in four seasons through clustering and outlier analysis (based on Anselin local Moran’s I) of LST. Furthermore, we introduced the geographical detector model to quantify the impact of six physical and socio-economic factors, including the digital elevation model (DEM), index-based built-up index (IBI), modiﬁed normalized difference water index (MNDWI), normalized difference vegetation index (NDVI), population, and Gross Domestic Product (GDP) on the LST distribution of Wuhan. Finally, to identify the inﬂuence of land cover on temperature, the LST of croplands, woodlands, grasslands, and built-up areas was analyzed. The results showed that low temperatures are mainly distributed over water and woodland areas, followed by grasslands; high temperatures are mainly concentrated over built-up areas. The maximum temperature difference between land covers occurs in spring and summer, while this difference can be ignored in winter. MNDWI, IBI, and NDVI are the key driving factors of the thermal values change in Wuhan, especially of their interaction. We found that the temperature of water area and urban green space (woodlands and grasslands) tends to be 5.4 ◦ C and 2.6 ◦ C lower than that of built-up areas. Our research results can contribute to the urban planning and urban greening of Wuhan and promote the healthy and sustainable development of the city.


Introduction
Cities are gathering places for human activities and survival. With the development of cities and their population growth, urban management and planning have to confront an increasing number of issues, including significant urban climate effects. The urban heat island (UHI) is one of the most significant urban climate effects. It refers to the phenomenon that causes the atmospheric and surface temperatures in urban areas to be higher than in the surrounding rural areas [1,2]. The UHI effect has been linked to a series of negative impacts of cities, including the deterioration of air and water quality, indirect economic losses, reduced comfort, and increased mortality rate [3][4][5]. It has become a research hotspot in urban meteorology, urban planning, urban geography, and urban ecology [6].
UHI can be divided into "surface" and "air" (or "atmospheric") UHIs based on the ways and heights that they are formed [7]. Air UHI refers to UHI effects in the canopy factors by using LST retrieval from TIR measures. For instance, Cao et al. [48] examined the effect impervious surfaces (IS) spatial patterns have on LST in Wuhan, China. They clarified the dynamic urbanization process of Wuhan and found that urbanization had a positive effect on LST but a negative effect on Enhanced Vegetation Index (EVI) [49]. Huang et al. [50] analyzed summer spatial temporal negative relationships of LST with NDVI, VF (vegetation fraction), and ISF (impervious surface fraction) and found a consistently strong negative relationship between LST and NDVI, as well as VF, and a positive relationship between LST and ISF. Shen et al. [46] generated high spatial resolution summer LST data for Wuhan for 1988-2013 and observed a decline in high temperature areas in the old city area and inter-annually stable relationships between distributions and land cover types. Huang et al. [47] proposed an indicator called Land Contribution Index (LCI) to better quantify the relationship between urban land-use types and UHI effect. Li et al. [51] obtained the land covers and LST maps of Wuhan from Landsat-5 images in 2000, 2002, 2005, and 2009, discussed the distribution of land-use/cover change and LST variation, and analyzed the correlation between population distribution and LST values in residential regions. Selecting Wuhan as a case study, Wang et al. [52] explored the climatic effect of landscape patterns and found that temperature contrast between surrounding landscape patches has a major influence on LST.
Although these studies focused on the effects of geographic environmental factors and human activity on LST variations, there is still a lack of studies that comprehensively compare factors and combinations of variables. In addition, most studies used correlation or regression analysis and focused little on quantitative analysis of multi-factors and their interactions. In comparison, the geodetector method can reveal the driving force behind a geographic phenomenon by detecting its spatial stratified heterogeneity. This method has been extensively used in the social sciences [53] and human health research [54] and is being popularized in natural [55] and environmental [56] sciences. The core assumption is that if a certain independent variable has a significant impact on a certain dependable variable, then there is relatively good agreement between the spatial distribution patterns of the independent and dependent variables [57]. Therefore, in this study, we conducted a multi-factor attribution analysis of LST variations in Wuhan using the geodetector method, with the seasonal spatial distribution of Wuhan's LST derived from Landsat 8 OLI/TIRS.
The present research aims to analyze the seasonal spatial distribution of Wuhan's LST by retrieving Landsat 8 images and identify the main driving factors affecting its cold and hot poles and their interactions. Furthermore, the differences in LST among cultivated lands, forest areas, grasslands and buil-up lands were quantitatively analyzed. The results of our study can provide a reference and scientific guidance to formulate urban development plans and effective mitigation of the UHI of Wuhan.

Study Area
Wuhan City (113 • 41 -115 • 05 E, 29 • 58 -31 • 22 N) is the capital of Hubei Province and is located east of the Jianghan Plain and the middle reaches of the Yangtze River. It is a metropolitan city in the central China area, and it is also a significant transport interchange [42]. The Yangtze River and its largest tributary, the Hanjiang River, converge in the city. The rivers in the city are intertwined with lakes and ports. Wuhan is called "a city with 100 lakes", as the water area accounts for almost a quarter of the city's total area [58]. The study area belongs to a subtropical humid climate with an average annual temperature of 15.8-17.5 • C according to Köppen climate classification [42]. By the end of 2019, Wuhan had 13 districts with a total area of 8569.15 km 2 , a permanent resident population of about 11.21 million, and a regional GDP of 1.62 trillion yuan [59]. Seven adjacent central districts constitute its main urban area and account for 11.2% of the total area ( Figure 1). Similar to Shanghai, Beijing, and Guangzhou, Wuhan has witnessed rapid urbanization in the past 30 years. Wuhan is recognized as one of the hottest "stove cities" in China [46,60,61]. Especially in summer, its daytime LST ranks at the top among China's major cities, seriously affecting people's living comfort. Therefore, it is urgent to monitor the LST in Wuhan and find its driving factors to improve its thermal comfort.

LST Retrieval from Landsat Images
There are several algorithms for retrieving LST, including the radiative transfer equation (RTE) method [62], the single channel algorithm [63,64], and the split window algorithm [63,65]. However, these methods rely on near-real-time atmospheric profile data for when the satellites pass through the study area, which are hard to obtain. Therefore, similar to many other UHI studies [46,66], we employed a method in which only the NDVI and the Top of the Atmosphere (TOA) spectral radiance are needed. The parameter information in the header file can help acquire the TOA spectral radiance and reflectance of the Landsat-8 OLI/TIRS data. Adopting the conversion formula, the spectral radiance values of the Landsat thermal infrared bands were converted to at-sensor brightness temperatures under the assumption of uniform emissivity [67]: where T sensor is the effective at-sensor brightness temperature in kelvin (K), L λ is the TOA spectral radiance in W/(m 2 ·sr·µm), and K 1 and K 2 are the calibration constants. For Landsat-8 TIRS, K 1 is 774.89 W/(m 2 ·sr·µm) and K 2 is 1321.08 K for band 10 (http: //landsat.usgs.gov/Landsat8_Using_Product.php).
The Tsensor values obtained above were referenced to a black body, which is quite different to the properties of real objects. Therefore, it was necessary to correct the spectral emissivity, and the LST based on satellite brightness temperature (Tsensor) was computed by the following equation [68]: where LST is the LST in K; T sensor is the black body temperature, and also the satellite brightness temperature, in K; λ is the wavelength of the emitted radiance in meters; α = 1.438 × 10 −2 mK; and ε is the surface emissivity. For ε, water, (NDVI < 0) was assigned a value of 0.9925, urban impervious areas and bare soil (0 = <NDVI < 0.15) were assigned a value of 0.923 [69], and vegetation (NDVI > 0.727) was assigned a value of 0.986 [70]. Otherwise, there was a modeling relationship with the NDVI values through the following equation [71]: NDVI was derived as follows: where RED is the red band and N IR is the near-infrared band.
Considering that the atmospheric environment is different between urban and rural areas [72,73], and in order to obtain accurate NDVI data, atmospheric correction is necessary to convert the TOA reflectance to surface reflectance [46,74]. In this study, we used the MODTRAN (Moderate Spectral Resolution Atmospheric Transmittance Algorighm And Computer Model) based FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) model embedded in ENVI software for the atmospheric correction [46,75].

Landsat Images
We obtained all the available Landsat 8 OLI/TIRS images over Wuhan City during 2013-2019 with a maximum of 10% cloud cover (Table A1). All the Landsat 8 data were downloaded from the United State Geological Survey Earth Explorer website (http:// earthexplorer.usgs.gov/); each set of Landsat data included three adjacent paths (path/row: 122/39, 123/38 and 123/39). The time of data acquisition was approximately 10:50 a.m. Greenwich Mean Time (GMT) in Wuhan.

Spatial Pattern analysis
Spatial autocorrelation means that the value of a variable at one location is correlated with the values of the same variable at its neighboring locations [76]. Spatial autocorrelation can be used to test whether the spatial distribution of LST presents clustering or discrete characteristics [76]. Cluster analysis can be used to find high and low temperature clusters, allowing study of the location and distribution characteristics of Wuhan's hot and cold poles. Spatial autocorrelation is usually measured by Moran's I, which includes the global and Anselin Local Moran's I [27]. The Anselin Local Moran's I calculates each sampling location to reveal its degree of spatial autocorrelation, and it is characterized by identifying spatial clusters of elements with high or low values, as well as identifying spatial outliers, which was described as follows [24]: where Z i is the value of the variable Z at location i; Z is the average value of Z with the sample number of n; Z j is the value of the variable Z at all the other locations (where j = i); σ 2 is the variance of variable Z; and w ij is a weight that can be defined as the inverse of the distance d ij among locations i and j. The weight w ij can also be determined using a distance band: samples within a distance band are given the same weight, while those outside the distance band are given the weight of 0. Anselin Local Moran's I statistic distinguishes between a statistically significant (0.05 level) cluster of high values (high-high clusters), cluster of low values (low-low clusters), outlier in which a high value is surrounded primarily by low values (high-low outliers), and outlier in which a low value is surrounded primarily by high values (low-high outliers) [23]. For this study, we employed the Anselin Local Moran's I statistic to identify the spatial clusters and spatial outliers of Wuhan's LST at the significance level of p < 0.05 in the four seasons of a year with the Cluster and Outlier Analysis tool in ArcMap 10.2 [77].

Geodetector Analysis
Geodetector is a new spatial statistical method mainly applied to identify the impact factors of spatial heterogeneity and the mechanisms through which they act [78][79][80]. It contains four detectors: a factor detector, risk detector, interaction detector, and ecological detector [57,81]. This study mainly used the factor detection and interaction detection of the geodetector to detect the driving mechanism of LST changes in Wuhan, and it quantified the impact of various factors on its change.
Factor detector detects the spatial heterogeneity of the dependent variable and the power of the determinant of the independent variables on the dependent variable. Factor detector is measured by q [78,80]: where h = 1, . . . , L is the layer of independent variable X, and N h and N are the number of sample units in layer h and the total region, respectively. σ 2 h and σ 2 are the variance in the h layer and the variance in the region. SSW is the sum of the spatial variance of each layer, SST is the total variance of Y in the region, L is the layer number of factor X. The q lies in the range [0, 1], and q = 1 indicates that Y is completely determined by X, while q = 0 indicates there is no association between Y and X [57,80]. A larger q indicates a stronger explanatory power of the independent variable X to attribute Y.
The interaction detector is used to identify the interaction between influencing factors [57,82], that is, to evaluate the explanatory power of the combined (increased or decreased) and independent effects of the impact factors on LST. In the first step, the q of two impact factors on LST is calculated; then, the q of the interaction of the impact factors is calculated, and q (X1), q (X2), and q (X1 ∩ X2) are compared.
The 1 km grid GDP and population data of the study area in 2015 were obtained from the Resource and Environment Science and Data Center, Chinese Academy of Sciences (http://www.resdc.cn). For altitude, we used the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global 30-m digital elevation model (DEM) V2 data. To characterize urban density, we used the index-based built-up index (IBI) proposed by Xu [83], which is defined as: where MIR is the medium infrared band and N IR is the near-infrared band.
To estimate the presence of water and vegetation, the modified normalized difference water index (MNDWI) and NDVI (Equation (4)) were calculated separately. MNDWI was derived as follows [84]: where GREEN is the green band and SW IR is the short-wave infrared band. Figure 2 shows the spatial distribution of each LST driving factor.

Land-Cover Data
The land-use data of Wuhan in 2019 were obtained from databox.store (www.databox. store), which is a Chinese company dedicated to providing geographic big data products and services to government departments, universities, research institutes, enterprises, and other users. The Land-Use and Land-Cover Change (LUCC) classification system was established by the Institute of Geography of the Chinese Academy of Sciences, and it includes six primary types (croplands, woodlands, grasslands, water areas, built-up lands, and unused lands) and 25 secondary types. In the present study, we used five classes: croplands (includes paddy field and dry land) (52.6%), woodlands (includes forestland, shrubs, sparse woodland, orchards, etc.) (9.1%), grasslands (refers to all kinds of grasslands with a coverage of more than 5% of herbaceous plants) (0.8%), built-up lands (includes urban land, rural residential land, industrial and mining land, and traffic land) (16.5%) and water areas (include rivers, lakes, and beaches) (20.1%). A total of 99.1% of Wuhan city was analyzed ( Figure 3). The remaining 0.9% were mainly unused areas. We extracted several statistical data from the LST of each land cover to detect the differences between different land covers, including minimum, maximum, mean, and standard deviation. In addition, kurtosis and skewness are also used to describe the spatial distribution of LST [29]. Skewness is a statistic to study the symmetry of data distribution [85]. By measuring the skewness coefficient, we can determine the asymmetry degree and direction of LST distribution. Kurtosis is a statistic that studies the steepness or smoothness of data distribution [85]. By measuring the kurtosis coefficient, we can determine whether the LST distribution is steeper or smoother than the normal distribution.

Spatial Distribution of LST
There are significant differences in the maximum, minimum, and average temperatures in Wuhan's LST among the four seasons ( Figure 4). The highest LST in spring is 43.8 • C, the average is 24.8 • C, and the lowest is 4.1 • C; the highest LST in summer is 56.1 • C, the average is 33.2 • C, and the lowest is 12.9 • C; the highest LST in autumn is 40.3 • C, the average is 23.1 • C, and the lowest is 13.2 • C; the highest LST in winter is 24.0 • C, the average is 11.6 • C, and the lowest is 1.9 • C. The standard deviation observed in spring (2.8 • C) and summer (3.2 • C) are higher than in autumn (1.9 • C) and winter (1.5 • C) ( Table 1). A heat pole is shown in all four seasons in the main urban area, in the periphery and on the banks of the Yangtze River. In contrast, the coldest areas are found in the water area and forest zone.

Spatial Patterns of LST
As shown in Figure 5, the spatial distribution of low and high temperature areas of Wuhan's LST varied greatly in different seasons. The "high-high" cluster areas had the highest average LST in each season of the year, the highest in summer (38.4 • C), followed by spring (29.0 • C) and autumn (26.4 • C), and the lowest in winter (13.5 • C) ( Table 2). These areas were larger in spring and summer, which was reflected in the proportion of area (8.40% and 11.88%, respectively), than in autumn and winter (3.47% and 5.27%, respectively). They were connected into a compact area with the Yangtze River as the symmetry axis, with a roughly symmetrical distribution, and they extended from the city center to the surrounding areas. The accumulation of high temperatures was more obvious in spring and summer. By comparing the LUCC map (Figure 3) of Wuhan City, it can be found that the "high-high" cluster areas of LST matched the spatial distribution of the built-up areas. The average LST of "low-low" cluster areas was the lowest in each season, the highest was in summer (29.0 • C), and the lowest was in winter (9.1 • C). The "low-low" cluster areas were mainly located in the major lakes, forest lands, and grasslands, indicating that the water area and green vegetation cover had a good cooling effect. Especially in the northeastern part of Wuhan, a large area of concentrated forest was a low-temperature region during the four study periods. Additionally, this area was also a high-altitude area in Wuhan. The "low-low" cluster areas was the largest in spring (15.10%) and the smallest in winter (8.02%). On the whole, Wuhan's LST was more likely to cluster in summer, and the Anselin local Moran's I showed that a total of 23.81% of the region showed significant spatial autocorrelation, which was the smallest in winter, only 13.29%.

Quantifying the Contribution of Impact Factors to LST
In this section, we analyzed the influence of six impact factors, including DEM, IBI, MNDWI, NDVI, population, and GDP ( Figure 2) on the distribution of Wuhan's LST.
Firstly, the factor detector was used to determine the influence of each impact factor on LST change. It can be seen from Figure 6 that the q of all impact factors ranged from 0.008 to 0.342, but their impact on the four seasons was not completely consistent. In general, MNDWI has the greatest impact on LST, which to a certain extent proves the importance of the water body in urban temperature regulation. This is followed by NDVI, which reflects that high-quality urban development cannot be separated from vegetation (including both natural and artificial vegetation). The third most important impact factor was IBI. Wuhan, as one of the largest cities in China, currently has a built-up area of 1419 km 2 , and urbanization is still increasing. The contribution of IBI to LST is expected to continue to increase.
The influence of each impact factor on the spatial distribution of LST was detected by using the interactive detector (Table 3). Overall, the interaction of any two impact factors on the change of LST was greater than one. For example, the q of the interaction between NDVI and MNDVI on LST change in spring was 0.470, which had the greatest influence on LST change, while the interaction between DEM and all socio-economic factors had a significantly greater influence on LST than DEM alone (0.420).
In terms of seasons, the interaction of NDVI and MNDWI was the largest (0.470) in spring, followed by the interaction between IBI and MNDWI (0.455). The interaction between IBI and MNDWI contributed the most to the variations of Wuhan's LST in summer with the interaction of 0.439, which was followed by the interaction of IBI and NDVI (0.340). The interaction of IBI and MNDWI (0.411) was the largest in autumn, which was followed by the interaction between DEM and MNDWI (0.371). The interaction between GDP and MNDWI (0.310) was the largest in winter, which was followed by the interaction between MNDWI and IBI (0.308).

Land Cover Analysis of LST
In this section, we quantified the difference of LST over the five land-cover types. Figure 7 shows the distribution of LST values for each land cover in the four seasons.  (Table 4). This kind of land covers exhibited a broader thermal amplitude in the spring and summer, which is detectable in the standard deviation (1.6 • C and 1.2 • C, respectively), than in winter and autumn (2.5 • C and 3.0 • C, respectively). On an interannual scale, LST in the built-up area also presented the largest average thermal amplitude (24.8 • C), which was obtained by subtracting the mean LST in summer (36.9 • C) from the mean LST in winter (12.1 • C). The LST of cropland was lower than that of the built-up area in each season. The average LST of these two land covers was very close in winter, with only 0.1 • C difference, but the difference increased to 2.0 and 3.7 • C in spring and summer, respectively. Another characteristic of cropland was the high kurtosis in each season of the year, indicating that the LST distribution over croplands was more concentrated and had some outliers. The LST of most cropland pixels was below average, and the temperature difference was large. In addition, the long tail of the LST was on the right, which indicated that a small portion of the cropland pixels had a higher LST. This is probably due to the wide altitude range of croplands distribution (between 0 and 550 m) as well as different types of cultivated land and different farming systems, resulting in a large temperature difference among the agricultural lands. The average LST of woodland in the four seasons was lower than that of several land covers analyzed before, forming a low-temperature zone in Wuhan. This gap was more obvious in spring and summer, especially in summer when it was 5.7 • C lower than the built-up area, resulting in the largest temperature difference recorded between different land covers in the same season. This is considered to be normal. Although the forest land in Wuhan is widely distributed from 0 to 800 m, it is mainly distributed in the high-altitude area northwest of Wuhan. Due to the influence of a vertical thermal gradient, the temperature of forest land could be expected to be lower than that of the built-up area and cropland. However, this difference began to decrease in autumn, and it was very small in winter (1.0 • C). On the inter-annual scale, the thermal amplitude of woodland was the smallest (20.1 • C). The mean LST of grassland was higher than that of woodland and less than that of construction land in four seasons. It was very close to agricultural land, but it exhibited a lower temperature, which showed the cooling function of grassland to a certain extent. In addition, the thermal amplitude of grassland in each season was very small, which can be seen from the standard deviation. The last land-use type analyzed was water area, which was another low-temperature area in Wuhan City. The performance of the average LST of the water area in the four seasons was very similar to that of woodland. It had the lowest temperature in all seasons, creating the smallest average LST record for a year (10.2 • C). However, the Yangtze River in winter was an exception. The average temperature of the Yangtze River was higher than that of all land covers, including the water area, with a temperature of 12.4 • C. A possible explanation for this phenomenon is the fact that the Yangtze River's waterway is busy with water transportation and high anthropogenic heat emission, which increase the temperature of the Yangtze River.
Although there are thermal differences between different land covers, because they are not at the same altitude, their comparison is not straightforward. To better compare the thermal differences between the urban green spaces (woodlands and grasslands) and the built-up areas in Wuhan, and to eliminate the errors caused by the vertical decline rate of temperature, we further analyzed the seasonal performance of LST for woodlands, grasslands and built-up lands in the main urban area within the altitude range of 0-100 m (Figure 8). In addition, we excluded agricultural land because to emphasize the thermal differences within cities and the potential impact that such thermal differences may have on urban settlements and living conditions. In addition, water bodies were also removed to highlight the cooling effect of green spaces in Wuhan. The LST of the urban green spaces (woodland and grassland) in Wuhan City tended to be 2.6 • C lower than that in the built-up area in spring and summer ( Table 5). The mean difference between the three land covers was very small in winter (0.8 • C). This means that the region was generally colder in winter within the same altitude range, with an average temperature of 11.8 • C. However, this phenomenon ended in the spring when the temperature began to gradually warm up. The thermal amplitude of built-up land increased higher than that of woodland and grassland (27.1 • C). Although the mean LST of woodland and grassland decreased, the difference between woodland and built-up land increased to 2.9 • C. In summer, the mean LST of woodland and grassland decreased further, and the difference between woodland and built-up land increased further. At this time, we detected the maximum temperature difference of 3.4 • C. In autumn, the thermal amplitude of each land-cover type began to fall, and the average difference between woodland and built-up land also shrank to 1.9 • C. Table 5. Descriptive statistics of the density plots in Figure 8.

Effects of Impact Factors on LST
In this study, we used Landsat-8 OLI/TIRS data to examine the relationship of DEM, IBI, MNDWI, NDVI, population, and GDP with LST in Wuhan, China. The results from factor detector show that all these factors have effects on LST, although the explanatory power of each factor is different. Due to the complexity of geographic processes, controlling factors often do not act independently but instead act collectively [57,86]. The extent to which the effects of controlling factors can be combined remains an unresolved issue [86]. This study examined the interaction between any two of the factors affecting LST distribution using the geodetector method to determine its pattern. The results show that the interactions between MNDWI, IBI, and NDVI were the main factors influencing the variations of Wuhan's LST in each season. Previous studies [43,[87][88][89] have confirmed that physical factors such as water and vegetation play an important role in mitigating SUHI by using the remote sensing index (MNDWI and NDVI). Similarly, there is no hightemperature agglomeration in Wuhan's water area and forest land. Peng et al. [90] found that DEM shows a great impact on LST in Western Sichuan Plateau, China. However, our results show that DEM has no significant impact on Wuhan's LST. This difference could be due to the dominant topography of the Western Sichuan Plateau being mountainous land, most having altitudes of 2000-3000 m, accounting for 83.4% of the total area, while Wuhan is mainly located in a plain area, and the area with elevation below 100 m accounts for 95.5% of its total area.
With the acceleration of urbanization and industrialization, the influence of socioeconomic factors on LST variations is becoming more and more important [88,91]. The impact of socio-economic factors, such as anthropologic heat releases and build-up intensity, on the spatial distribution pattern of SUHI were confirmed in China's 32 major cities [92]. On the whole, among the three social factors (IBI, Population, GDP) used in this study, IBI has the strongest explanatory power to LST, whether it is the single explanatory power or interactive explanatory power with other factors. Previous studies also showed the importance of IBI (or normalized difference built-up index (NDBI)) on LST variations [89,91]. The larger the area of urban construction land, the higher the NDBI and LST [93]. The enhancement of LST by impervious surface (IS) was more pronounced in individual cities than in urban agglomerations [94]. Therefore, high-temperature areas in Wuhan are mainly concentrated in built-up land. A remarkable feature of the winter thermal environment in Wuhan was the high temperature in a section of the Yangtze River (Figure 4), which is similar to the results by Guo et al. [95] for the winter surface thermal environment of Nanjing City (China). This phenomenon cannot be explained by the large specific heat capacity of the water bodies alone, because there were no high temperatures in Hanjiang River, Donghu Lake, and Tangxun Lake. Considering its important economic and shipping status, the Yangtze River has become the world's largest comprehensive transportation channel, with railways, expressways, pipelines, and ultra-high voltage power transmission lines [96]. Wuhan port is one of the 28 major ports in China and the largest public terminal operator in Hubei Province. There are a large number of industrial and transportation facilities along the river [97]. Many studies have also shown that anthropogenic heat emissions are deteriorating the thermal environment of water bodies [98][99][100]. Therefore, the thermal environmental anomalies in the Wuhan section of the Yangtze River in winter may be caused by anthropogenic thermal emissions related to transport heat release. In a word, the three main impact factors and other impact factors jointly affect the spatial aggregation and distribution pattern of Wuhan's LST.
Our study also shows that the driving factors of LST pattern in Wuhan are complex and seasonal, not static [47], because the most powerful explanatory factors are different in each season (see Section 4.3). Similar results were found in Shenzhen City (China) by Peng et al. [43], who showed that the most important factors affecting spatial heterogeneity of LST in summer and transition seasons were NDBI and NDVI, respectively, while in winter, they were construction land percentage and NDVI.

The Role of Different Land Covers in Conditioning LST
With respect to the combined analysis of the different land covers and LST estimated by Landsat, several cities, such as Shenyang, China [101] and Kuala Lumpur metropolitan area, Malaysia [102], show patterns similar to Wuhan, with built-up land displaying the highest temperatures among the land-use types, the lowest being recorded in the water area. As we all know, the specific heat capacity of water is higher than that of soil, which has a certain regulating effect on temperature. As an inland city, the large water area is a distinctive feature of Wuhan. It could be found from Table 4 that the LST of water areas in Wuhan city tends to be 5.4 • C lower than that in the built-up area in spring and summer, reflecting that the water area plays an important role in cooling Wuhan. Le et al. [103] used Landsat 8 images to study the relationship between urban landscape patterns and the thermal environment in Longquan, which is a national forest city in China. They found that the lowest temperature area was in forest land, which was slightly different from our research results. This may be because the forest land of Longquan occupies 87.4% of the land area, while the water area only accounts for 0.8%; in contrast, the water area of Wuhan accounts for 20.1% of the total area.
The temperature difference between built-up area and forest land is more significant in spring and summer, while they reach a minimum in winter, which was consistent with the results of the Rotterdam (Netherlands) metropolitan area obtained by Van Hove [104] and the results provided by Lemus-Canovas [29] for Barcelona (Spain). This is mainly because in spring and summer, the daytime is long, the height of the sun at noon is high, and the ground receives more solar radiation, which increases the heating differences between different types of surfaces of the city. The SUHI study in Italian metropolitan cities also found that the daytime SUHI increased significantly in summer, especially in inland cities, by increasing the size of areas with low tree cover densities in the metropolitan core (or decreasing areas with low tree cover densities outside the metropolitan core), further increasing its intensity when the impervious density grew [33]. This study revealed that urban green space has a clear impact on the reduction of the urban thermal environment and can effectively mitigate the UHI [66]. For three megacities in Southeast Asia-Bangkok (Thailand), Jakarta (Indonesia) and Manila (Philippines)-the average LST of the green space was found to be about 3 • C lower than that of the impervious surface, which is close to our results (2.6 • C), highlighting its role in weakening the UHI [6]. Here, an increase in NDVI tends to mean a gradual decrease in UHI [105]. In contrast, the city of Jaipur (India) has witnessed considerable growth in built-up area at the expense of greener patches over the last decade; as a result, its UHI has been strengthened and extended to new regions [106].

Conclusions
This study examined the spatial variation of daytime land surface temperatures in Wuhan from 2013 to 2019 and the driving factors affecting its cold and hot poles. The main results are as follows: (1) LST in Wuhan shows spatial aggregations, with high temperature areas mainly distributed in the built-up areas and low temperature areas mainly distributed in water and woodland areas. In summer, during the study period, the average temperature difference among different land covers rose to 6.2 • C, while in winter, this average difference dropped to 1.9 • C. (2) Physical and socio-economic factors and their interactions were responsible for the patterns of LST differences in Wuhan. Among them, MNDWI, IBI, and NDVI, relating to the distribution of water areas, built-up, and green spaces and their interactions, were the main factors affecting temperature differences. (3) In spring and summer, the water area and urban green spaces (woodland and grassland) registered an LST of up to 5.4 • C and 2.6 • C lower than the built-up lands, respectively. From this perspective, water area and urban green spaces plays an extraordinary role in improving the thermal comfort of the city and preventing the occurrence of high temperature heat stress.
This study can help to objectively manage the relationship between urban built-up areas, green spaces, and forest lands in urban planning and construction.