Exploring the Variations and Influencing Factors of Land Surface Temperature in the Urban Agglomeration on the Northern Slope of the Tianshan Mountains

Abstract

Changes in land surface temperature (LST) can have serious impacts on the water cycle and ecological environment evolution, which in turn threaten the sustainability of ecosystems. The urban agglomeration on the northern slopes of the Tianshan Mountains (UANSTM) is located in the arid and semi-arid regions of northwest China, with an extremely fragile ecological environment and sensitive to climate change. However, studies on the LST of the UANSTM have not received much attention. Therefore, this study explored the spatial distribution pattern, fluctuation characteristics, and influencing factors of the LST of the UANSTM from 2005 to 2021 based on MODIS time series LST data and the geo-detector model with optimal parameters. The results show that the UANSTM is dominated by medium- and high-temperature classes, with high- and extremely high-temperature classes clustered in Turpan City. The daytime and nighttime LST patterns are significantly different, with a typical “daytime cold island and nighttime heat island” feature in the oasis region. During 2005–2021, LST fluctuated greatly in the northwestern part of the UANSTM, with LST showing an increasing trend during both daytime and nighttime, and the warming rate was more intense during daytime than nighttime. The increasing trend of LST in Urumqi, Changji Hui Autonomous Prefecture, Shihezi, and Wujiaqu is very significant and will remain consistent in the future. Precipitation, DEM, and AOD are the most important influencing factors of LST in the UANSTM, where precipitation and DEM are negatively correlated with LST, and AOD is positively correlated with LST. Land cover factors (LULC, NDVI,, and NDBSI) are the next most influential, and socioeconomic factors (NTL, GDP, and POP) are the least influential. The results of this study can provide a scientific reference for the conservation and sustainable development of the ecological environment of the UANSTM.

1. Introduction

Land surface temperature (LST) is a key factor affecting the surface–atmosphere energy exchange and water cycle, and it is also an important manifestation of the surface energy balance [1,2,3]. With the rapid advancement of global urbanization, a large number of impervious surfaces dominated by cement and asphalt have destroyed the original natural landscape and caused changes in surface parameters such as surface albedo and emissivity, which make the LST changes in urban areas more obvious [4,5,6,7]. The change of urban LST will not only reduce the comfort of people’s living environment and endanger human health, it will also have a serious impact on the local climate and eco-environment evolution of the city [8,9]. The urban agglomeration is the space carrier for the development of regional economy, industry, and urbanization to a certain advanced level, which has become the most prominent manifestation of global urbanization [10,11]. At present, many large urban agglomerations have been formed in China, such as Beijing-Tianjin-Hebei [12], Yangtze River Delta [13], Pearl River Delta [14], and Chengdu-Chongqing [15] urban agglomerations. Therefore, it is very necessary to understand the spatial distribution pattern, changing trend, and influencing factors of LST in urban agglomerations for protecting the eco-environment and sustainable economic development.

In recent years, a large number of scholars have found fruitful research results in the spatial-temporal distribution pattern [16,17] and influencing factors [18,19,20,21] of LST, and many spatial statistical models [22,23] and landscape ecology theories [24,25,26] have also been introduced into the study of the evolution and influencing mechanism of LST. Previous studies have indicated that land cover is considered a key factor affecting LST [27]. The impervious surface will increase the sensible heat flux of the surface and increase the heat radiation energy outward from the surface, which will lead to the increase of LST [28]. Water and green space are the natural radiators of the city. Vegetation can effectively reduce the LST through transpiration and shading [29]. Water has a high specific heat capacity and can remove a large amount of heat through evapotranspiration, so it can keep a lower temperature [30]. The normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and normalized differential build-up and bare soil index (NDBSI) can accurately characterize land cover information, and they have shown a good linear relationship with LST [31,32]. In addition, some climatic, socioeconomic, and topographic factors also have an assignable impact on LST [33,34]. For example, DEM and gradient have a significant negative linear correlation with LST [35]. Population density (POP), gross domestic product (GDP), nighttime lights (NTL), aerosol optical depth (AOD), and precipitation also have strong correlations with LST [31,36,37,38].

In previous studies, Landsat series data have often been used to analyze changes in the spatial distribution pattern of LST based on one or several sensor transit moments. Although Landsat data have a high spatial resolution, the temporal resolution is low and often limited by the cloud coverage conditions at the moment of sensor transit, making it difficult to provide continuous and stable observations of large-scale LST. In studies of influencing factors, simple linear regression models [23,39,40] are also mostly used as a research method to explore the relationship between single or multiple influencing factors and LST. However, the above methods cannot fully reveal the nonlinear relationship between each influencing factor and LST, nor can they analyze the effect on LST when different influencing factors act together. In addition, the selection of influencing factors mostly prefers the land cover factors and lacks consideration of climatic, socioeconomic, and topographical factors. Therefore, in this study, MODIS LST data with high temporal resolution were selected to analyze the spatial-temporal change trend of LST, and the geo-detector model (GDM) [41] was applied to explore the effects of various influencing factors on LST. The GDM can quantitatively reveal the linear and nonlinear relationships between each influencing factor and LST, and measure the impact on LST when different influencing factors act together [42].

This study was conducted in the urban agglomeration on the northern slope of Tianshan Mountains (UANSTM). The UANSTM is the most urbanized and most densely populated and industrialized area in Xinjiang, China [43]. With one-fifth of Xinjiang’s population and half of its GDP, it has an irreplaceable role in Xinjiang [44]. However, since the 1980s, urbanization has led to the deterioration of the eco-environment, and a large amount of ecological land has been converted into construction land, and the desert-oasis ecotone has continued to shrink [45,46]. In addition, the current studies on LST of urban agglomeration are still focused on coastal and developed regions, while very little attention has been paid to the study of urban agglomerations in inland arid and undeveloped regions, which is an important reason why the UANSTM has been chosen as the study area.

Therefore, to address the above issues, the main objective of this study is to analyze the spatial distribution pattern, fluctuation characteristics, and change trends of LST in the UANSTM based on MODIS LST data with high temporal resolution, and to analyze the effects of multiple influencing factors on LST using the geo-detector model. This study will help planners to provide decision-making and support for improving the quality of human habitat and surface thermal environment issues.

2. Study Area and Data Source

2.1. Study Area

The UANSTM is an emerging urban agglomeration in the inland arid region of northwest China [47]. It is located at the northern foot of the Tianshan Mountains and the southern edge of the Junggar Basin in Xinjiang, and it includes nine cities: Urumqi, Wujiaqu, Changji Hui Autonomous Prefecture, Turpan, Shihezi, Karamay, Kuitun, Usu, and Shawan (Figure 1). The UANSTM is a typical oasis urban agglomeration with scarce water resources, fragile eco-environment, and sensitivity to climate change [48]. Compared with urban agglomerations in developed countries and coastal China, the UANSTM is particularly uncompetitive in terms of urbanization, resources, and eco-environment carrying capacity [44,49]. The Chinese government has attached great importance to the eco-environment issues in recent years. To enhance the competitiveness of the UANSTM and promote its sustainable development transformation, it is necessary to search for the eco-environment problems in various aspects. LST, as an important parameter in the eco-environment, needs urgent attention. Therefore, under the influence of intense human activities, it is of very high application value to understand the changing characteristics and influencing factors of LST in the UANSTM.

2.2. Data Collection and Preprocessing

2.2.1. LST Data

The LST data for this study used all MODIS Aqua LST products MYD11A2 (version 6) covering the UANSTM from 2005 to 2021, which required track numbers “h23v04” and “h24v04” for mosaicking, with a total data volume of 1564 remote sensing images (https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 15 April 2022). Each image contains “LST_Day_1km” and “LST_Night_1km” bands, which are daytime and nighttime LST data, respectively. The MYD11A2 product is calculated from the thermal infrared bands of MODIS 31 and 32 channels using a split-window algorithm with a spatial resolution of 1 km. This product is a composite LST product obtained by averaging 8 days of clear-sky LST, which can effectively reduce the effect of cloudiness, and the product has been validated with a series of accuracies and has been widely used in regional- or global-scale LST studies [50,51,52,53]. We calculated the average daytime and nighttime LST images for each year from 2005 to 2021 for the UANSTM and counted some details of the images, as shown in

Table 1. Details of the annual mean LST images of the UANSTM from 2005 to 2021.

Time	Maximum	Minimum	Average	Standard Deviation	Time	Maximum	Minimum	Average	Standard Deviation
2005 daytime	46.50	−13.10	23.25	8.82	2013 nighttime	10.57	−25.12	−0.25	4.09
2005 nighttime	12.64	−27.38	−1.28	4.34	2014 daytime	41.36	−13.52	23.14	8.38
2006 daytime	46.19	−12.35	24.13	8.40	2014 nighttime	9.59	−25.68	−1.41	4.18
2006 nighttime	12.85	−26.75	−0.45	4.23	2015 daytime	43.60	−12.25	24.31	8.67
2007 daytime	41.75	−12.80	23.80	8.48	2015 nighttime	10.92	−22.72	0.19	4.16
2007 nighttime	10.99	−27.53	−0.61	4.20	2016 daytime	44.01	−12.14	23.59	8.79
2008 daytime	42.70	−12.83	24.66	8.46	2016 nighttime	12.72	−28.19	0.03	4.21
2008 nighttime	10.53	−29.97	−0.72	4.23	2017 daytime	43.37	−12.08	24.29	8.77
2009 daytime	41.92	−13.00	24.20	8.67	2017 nighttime	13.84	−28.11	−0.04	4.38
2009 nighttime	10.17	−26.78	−1.19	4.16	2018 daytime	41.88	−13.38	23.03	8.55
2010 daytime	42.46	−13.24	22.30	8.57	2018 nighttime	9.87	−29.60	−1.67	4.07
2010 nighttime	9.77	−27.17	−1.31	4.09	2019 daytime	42.60	−13.81	24.09	8.48
2011 daytime	44.43	−12.87	23.15	8.74	2019 nightime	11.10	−25.49	−0.38	4.26
2011 nighttime	9.86	−27.21	−1.38	4.20	2020 daytime	43.61	−12.30	24.45	8.41
2012 daytime	44.03	−14.09	23.07	8.87	2020 nighttime	11.04	−27.48	−0.96	4.22
2012 nighttime	10.37	−28.24	−1.76	4.30	2021 daytime	43.25	−12.73	24.63	8.65
2013 daytime	46.35	−12.54	24.63	8.60	2021 nighttime	11.09	−24.54	−0.75	4.18

Note: the unit in Table 1 is °C.

.

2.2.2. Influencing Factors’ Data

Based on previous studies and data availability, we considered four types of influencing factors: land cover, climate, socio-economic, and topographic factors, and finally selected 10 influencing factors, including: land use/land cover (LULC), NDVI, NDBSI, precipitation, AOD, NTL, GDP, POP, DEM, and gradient (

Table 2. The details of influencing factors’ data.

Types of Influencing Factors	Factors	Initial Resolution/Resampling Resolution	Time	Access
Land cover factors	LULC	1000 m/1000 m	2019	https://zenodo.org/, accessed on 19 April 2022
	NDVI	1000 m/1000 m	2019	https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 15 April 2022
	NDBSI	500 m/1000 m	2019	https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 15 April 2022
Climatic factors	Precipitation	1000 m/1000 m	2019	http://data.tpdc.ac.cn/, accessed on 16 April 2022
	AOD	1000 m/1000 m	2019	https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 15 April 2022
Socioeconomic factors	NTL	500 m/1000 m	2019	https://eogdata.mines.edu/products/vnl/, accessed on 15 April 2022
	GDP	1000 m/1000 m	2019	https://www.resdc.cn/, accessed on 15 April 2022
	POP	1000 m/1000 m	2019	https://www.resdc.cn/, accessed on 15 April 2022
Topographical factors	DEM	30 m/1000 m	-	http://www.gscloud.cn/, accessed on 16 April 2022
	Gradient	30 m/1000 m	-	http://www.gscloud.cn/, accessed on 16 April 2022

). The LULC data are a 30 m annual Chinese land cover dataset from 1990 to 2019 produced by Yang et al. [54] (https://zenodo.org/, accessed on 19 April 2022). We reclassified it into six types, including: cropland, woodland, water area, grassland, construction land, and unused land. The NDVI, AOD, and NDBSI data come from the MODIS vegetation index product MYD13A2, terrestrial aerosol optical depth product MCD19A2, and surface reflectance product MYD09A1, respectively (https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 15 April 2022). The calculation method of NDBSI is detailed in [47]. The GDP and POP data were from the Resource and Environmental Science and Data Center (https://www.resdc.cn/, accessed on 15 April 2022). The NTL data were from the NPP-VIIRS NTL monthly product produced by the Colorado School of Mines (https://eogdata.mines.edu/products/vnl/, accessed on 15 April 2022). The precipitation data were obtained from the monthly precipitation raster data of the National Tibetan Plateau Data Center (http://data.tpdc.ac.cn/, accessed on 16 April 2022). The DEM and gradient data were obtained from the Geospatial Data Cloud (http://www.gscloud.cn/, accessed on 16 April 2022).

We resampled all influencing factor data to a 1 km resolution to ensure that all data had the same spatial resolution. We calculated annual average rasters for NDVI, NDBSI, precipitation, AOD, and NTL to ensure that all data were annual averages.

3. Methodology

In this study, the spatial distribution pattern of LST was first analyzed by grading the LST using the standard deviation method. Next, the absolute variability, Sen’s slope analysis, Mann–Kendall trend test, and Hurst index were used to explore the fluctuation characteristics and change trends of LST. Finally, the effects of land cover factors, climatic factors, socioeconomic factors, and topographic factors on LST were analyzed based on the geo-detector model (the overall research flow is shown in Figure 2).

3.1. Land Surface Temperature Classification

To accurately describe the spatial pattern differences of LST, the standard deviation method was used to divide LST into five classes: extremely high temperature (EHT), high temperature (HT), medium temperature (MT), low temperature (LT), and extremely low temperature (ELT) (

Table 3. The classification standard of land surface temperature.

LST Classes	LST Range
EHT	T > μ + 1.5 std
HT	μ + 0.5 std < T ≤ μ + 1.5 std
MT	μ − 0.5 std ≤ T < μ + 0.5 std
LT	μ − 1.5 std ≤ T < μ − 0.5 std
ELT	T < μ − 1.5 std

Note: T is the LST value, μ is the mean, and std is the standard deviation.

). The standard deviation method can well-characterize the concentration and volatility of LST by combining the mean and different standard deviation multiples [40].

3.2. Absolute Variability (AV)

The AV reflects the absolute deviation between the data and the mean over a certain time period, which means that it can reflect the degree of fluctuation of the data over a certain time period [52]. In this study, AV was used to characterize the fluctuations of LST in the daytime and nighttime between 2005 and 2021. The expression for AV is as follows:

$$AV=\frac{{{\displaystyle \sum }}_{i=1}^n\left|x_i-x_{mean}\right|}{n}$$

(1)

where x_i is the LST observation in year i, x_mean is the average LST from 2005 to 2021, and n is the number of years (n = 17). The AV in Equation (1) is proportional to the fluctuation strength.

3.3. Trend Analysis Methods

3.3.1. Sen’s Slope Analysis

Sen’s slope analysis is a method used to estimate the changing trend of the data over a time series, which can effectively avoid the effect of missing values and outliers. Its fundamental principle is to first calculate the slope between all adjacent data in the time series data, and then take the median value of the slope as the value of the changing trend [55]. The basic equation is written as:

$$slope=median\left(\frac{LS{T}_j-LS{T}_i}{j-i}\right) , \forall j>i$$

(2)

where median is the median function, and LST_j and LST_i are the observed values at moments j and i in the time series. In Equation (2), when the slope < 0, it indicates a decreasing trend, and when the slope > 0, it indicates an increasing trend.

3.3.2. Mann–Kendall Trend Test

The Mann–Kendall trend test is a nonparametric test method that is often used in conjunction with Sen’s slope analysis to determine the significance of a changing trend [56,57]. As with Sen’s slope analysis, the Mann–Kendall trend test is insensitive to missing and outlier data. Its normal distribution statistic, S, is calculated as follows:

$$S={\displaystyle \sum }_{i=1}^{n-1}{\displaystyle \sum }_{j=i+1}^{n}sgn\left(LS{T}_j-LS{T}_i\right)$$

(3)

where LST_i and LST_j are the LST observations in year i and j, respectively, and sgn in Equation (3) is a symbolic function whose expression is presented in Equation (4):

$$sgn\left(LS{T}_j-LS{T}_i\right)=\left\{\begin{array}{c}1, LS{T}_j-LS{T}_i>0\\ 0, LS{T}_j-LS{T}_i=0\\ -1, LS{T}_j-LS{T}_i<0\end{array}\right.$$

(4)

From Equation (4), it can be seen that the value of the sgn function varies between −1, 0, and 1 based on (LST_j − LST_i). The variance, Var(S), of S can be calculated based on Equation (5) as follows:

$$Var\left(S\right)=\frac{n\left(n-1\right)\left(2n+5\right)}{18}$$

(5)

Based on S and Var(S) in Equations (3) and (5), the standard orthogonal statistical distribution, Z, of the Mann–Kendall trend test can be calculated as shown in Equation (6):

$$Z=\left\{\begin{array}{c}\frac{S-1}{\sqrt{Var\left(S\right)}}, S>0\\ 0 , S=0\\ \frac{S+1}{\sqrt{Var\left(S\right)}}, S<0\end{array}\right.$$

(6)

Referring to previous studies, this study uses 95% and 99% confidence intervals, which means that the changing trend is very significant when Z ≥ 2.58 and Z ≤ 2.58, significant when 1.96 ≤ Z < 2.58 and −2.58 < Z ≤ −1.96, and otherwise the trend is not significant. We superimposed the results of Sen’s slope analysis and the Mann–Kendall trend test to classify the changing trends of LST into six types (

Table 4. Condition and level definition of change trend.

Slope and Z Value	Changing Trends	Slope, Z Value and Hurst Index	Future Changing Trends
Slope < 0, Z ≤ −2.58	Very significantly decreased	Slope < 0, Z ≤ −2.58, H > 0.5	Sustained very significant decrease
Slope < 0, −2.58 < Z ≤ −1.96	Significantly decreased	Slope < 0, −2.58 < Z ≤ −1.96, H > 0.5	Sustained significant decrease
Slope < 0, −1.96 < Z < 1.96	Not significantly decreased	Slope < 0, −1.96 < Z < 1.96, H > 0.5	Sustained no significant decrease
Slope > 0, 1.96 < Z < 1.96	Not significantly increased	Slope > 0, 1.96 < Z < 1.96, H > 0.5	Sustained no significant increase
Slope > 0, 1.96 ≤ Z < 2.58	Significantly increased	Slope > 0, 1.96 ≤ Z < 2.58, H > 0.5	Sustained significant increase
Slope > 0, Z ≥ 2.58	Very significantly increased	Slope > 0, Z ≥ 2.58, H > 0.5	Sustained very significant increase
–	–	H ≤ 0.5	Unknown

).

3.3.3. Hurst Index

The Hurst index can effectively describe the persistence of changes in time series data. It reflects the interrelationship between before and after changing trends, that is, the past changing trends can affect the current changing trends, and the current changing trends can affect the future changing trends [58,59]. Therefore, the Hurst index can be used to analyze the future trend of LST. The basic principle of the Hurst index is as follows:

Define a time series as: {LST(t), t = 1, 2, … n}, and for any integer α ≥ 1, the average series is shown in Equation (7):

$${\overline{LST}}_{\alpha }=\frac{1}{\alpha }{\displaystyle \sum }_{t=1}^{\alpha }LST\left(t\right) \alpha =1,2,\cdots ,n$$

(7)

The cumulative deviation is calculated as shown in Equation (8):

$$X\left(t,\alpha \right)={\displaystyle \sum }_{t=1}^t\left(LST\left(t\right)-{\overline{LST}}_{\alpha }\right) 1\le t\le \alpha$$

(8)

Define the range of extreme difference, R(α), as in Equation (9):

$$R\left(\alpha \right)=\underset{1\le t\ge \tau }{max}X\left(t,\alpha \right)-\underset{1\le t\ge \tau }{min}X\left(t,\alpha \right) \alpha =1,2,\cdots ,n$$

(9)

Calculate the standard deviation, S(α), as shown in Equation (10):

$$S\left(\alpha \right)={\left[\frac{1}{\alpha }{\displaystyle \sum }_{t=1}^{\alpha }{\left(LST\left(t\right)-{\overline{LST}}_{\alpha }\right)}^2\right]}^{1/2} \alpha =1,2,\cdots ,n$$

(10)

Finally, the Hurst index is calculated by Equations (11) and (12):

$$\frac{R\left(\alpha \right)}{S\left(\alpha \right)}={\left(c\alpha \right)}^H$$

(11)

$$\log {\left(R∕S\right)}_n=a+H\times \log \left(n\right)$$

(12)

where H is the Hurst index, c is the relation constant, and a is the intercept. The H value can be obtained by fitting Equation (12) using the least squares method. When H > 0.5, it indicates that the future changing trend of LST is most likely to be the same as the present, when H < 0.5, it indicates that the future changing trend of LST may be opposite to the present, and when H = 0.5, it indicates that the future changing trend of LST is random. In this study, H < 0.5 is defined as the future changing trend is unknown. To understand the future changing trend of LST, we superimpose the Hurst index and the LST changing trend types to obtain the future changing trend types of LST (Table 4).

3.4. Geo-Detector Model Based on Optimal Parameters

The GDM is a spatial statistical method used to explore influencing factors based on spatial differentiation theory. The GDM measures the degree of correlation between independent variables and dependent variables by measuring the degree of spatial consistency between them, which can effectively avoid the multicollinearity problem between dependent variables [41]. This study used four modules of GDM: factor detector, interaction detector, risk detector, and ecological detector, to analyze the effect of each influencing factor on LST. Due to the availability of data, this study only used 2019 as an example to analyze the influencing factors.

When using GDM, continuous variables must be discretized, and the discrete method can have a large impact on the results. The commonly used data discretization methods are natural breaks, equal interval, quantile, standard deviation, and geometric interval. However, most of the previous studies rely on personal experience in choosing the discretization methods, which is rather subjective, while the GD package of R studio can select the optimal method based on q values to obtain the optimal parameters for GDM. Therefore, in this study, firstly, 2 km grid points were established according to the study area to extract the values of LST and each influence factor, and then the optimal discretization of continuous type variables and calculation of GDM results were performed by GD package (https://cran.r-project.org/web/packages/GD/, accessed on 20 April 2022).

3.4.1. Factor Detector

The factor detector measures the impact of the independent variable on the dependent variable by detecting the extent to which the dependent variable explains the spatial heterogeneity of the independent variable, and this degree of impact is quantified using the q value. The expression is as follows:

$$q=1-\frac{{{\displaystyle \sum }}_{h=1}^L{N}_h{\sigma }_h^2}{N{\sigma }^2}$$

(13)

where the range of q is [0,1]. A larger q value indicates that the influencing factor has a greater impact on the LST. h = 1, 2, …, L is the classification or partition of the independent or dependent variable, N_h and N are the number of cells in layer h and the whole region, respectively, and ${\sigma }_h^2$ and σ² are the variance of the dependent variable in layer h and the whole region, respectively.

3.4.2. Interaction Detector

The interaction detector identifies interactions between different influencing factors by assessing the magnitude of the explanatory power of the LST when influencing factors X1 and X2 act together. The identification method is to first calculate the q values of the two independent variables X1 and X2 on LST: q(X1) and q(X2), and the q values of the two independent variables on LST when they interact: q(X1∩X2), and then compare q(X1), q(X2), and q(X1∩X2) to classify the different interaction types (

Table 5. Types of interactions and judgment criteria.

Interaction Types	Judgment Standard
Weaken, nonlinear	q(X1∩X2) < Min(q(X1), q(X2))
Weaken, univariate	Min(q(X1), q(X2)) < q(X1∩X2) < Max(q(X1), q(X2))
Independent	q(X1∩X2) = q(X1) + q(X2)
Enhance bivariate	q(X1∩X2) > Max(q(X1), q(X2))
Enhance, nonlinear	q(X1∩X2) > q(X1) + q(X2)

Note: Min(q(X1), q(X2)) is q(X1), q(X2), both to take the minimum. Max(q(X1), q(X2)) is q(X1), q(X2), both to take the maximum. q(X1) + q(X2) is q(X1), q(X2), both to sum.

).

3.4.3. Risk Detector

The risk detector is used to judge whether the mean values of dependent variables in different independent variable intervals are significantly different, and the t statistic is used to judge:

$$t=\frac{{\overline{M}}_{h=1}-{\overline{M}}_{h=2}}{{\left[\frac{Var\left({\overline{M}}_{h=1}\right)}{n_{h=1}}+\frac{Var\left({\overline{M}}_{h=2}\right)}{n_{h=2}}\right]}^{1∕2}}$$

(14)

where ${\overline{M}}_h$ is the average value in the h interval, n_h is the number of samples in the interval h, and Var() is the variance function.

3.4.4. Ecological Detector

The ecological detector is used to compare whether the influence of the two influencing factors X1 and X2 on the spatial distribution of the dependent variable is significantly different, measured by the F statistic:

$$F=\frac{N_{X1}\left(N_{X2}-1\right)SS{W}_{X1}}{N_{X2}\left(N_{X1}-1\right)SS{W}_{X2}}$$

(15)

where N_X1 and N_X2 denote the sample sizes of influences X1 and X2, and SSW_X1 and SSW_X2 denote the sum of variances within the X1 and X2 layers, respectively.

4. Results

4.1. Spatial Distribution Pattern of LST

To avoid chance, we calculated the average values of daytime and nighttime LST from 2005 to 2021 for the urban agglomeration on the northern slopes of Tianshan to analyze the spatial distribution pattern of LST during the daytime and nighttime (Figure 3). From the results, it is clear that the daytime LST ranged from −12.61 to 42.64 °C. The area proportions of HT and EHT were 33.28% and 2.23%, respectively, both mainly clustered in Turpan city. The area proportions of LT and ELT were 18.38% and 8.64%, respectively, mainly distributed in the high-altitude areas of the Tianshan Mountains in the southwestern edge and central part of the study area, as well as in the oasis area in the central part. The LST at night ranged from −25.58 to 10.61 °C. The proportions of HT and EHT areas were 20.43% and 5.98%, respectively, and both were equally clustered within Turpan city. The area of EHT almost doubled at night compared to daytime, while the area of HT significantly decreased. The area proportions of LT and ELT at night were 11.01% and 7.30%, respectively. During the nighttime, the area and spatial distribution of the ELT were almost the same as during the daytime. However, the area of LT was significantly reduced, mainly because the large area of LT in the oasis region in the central part of the study area changed to MT and HT at night. In general, the LST classes are dominated by MT and HT in the UANSTM, both during daytime and nighttime. The high-temperature classes (EHT and HT) were always clustered in Turpan city. In the oasis area of the study area, there was a significant difference in LST classes between daytime and nighttime.

4.2. The Fluctuation Characteristics of LST

We analyzed the fluctuation characteristics of daytime and nighttime LST of the UANSTM from 2005 to 2021 by absolute variability (AV), and higher values of AV represent more intense fluctuations of LST. From Figure 4, the AV ranged from 0.27 to 5.18 °C with a mean value of 1.09 °C during the daytime and from 0.26 to 2.89 °C with a mean value of 0.59 °C during the nighttime, indicating that LST fluctuations are more intense during daytime than nighttime. The daytime high AV areas are mainly clustered in Karamay, Shawan, and Changji Hui Autonomous Prefecture in the northwestern part of the study area, especially in Karamay, indicating strong LST fluctuations in these areas during the daytime. The nighttime high AV areas are mainly clustered in Usu, Karamay, Shawan, and parts of Changji Hui Autonomous Prefecture and Urumqi, especially in Usu, indicating that the LST fluctuations are stronger in these areas during the nighttime.

4.3. The Changing Trend of LST

4.3.1. The Changing Trend of LST from 2005 to 2019

The results of the Sen’s slope analysis (Figure 5) showed that the trend of LST during the daytime in the UANSTM ranged from −0.65 to 0.96 °C-yr⁻¹ during 2005–2021, with a mean value of 0.05 °C-yr⁻¹, and the proportion of the area with an increasing trend was 76.83%, while the proportion of the area with a decreasing trend was only 23.17%. The nighttime LST trend ranged from −0.22 to 0.49 °C-yr⁻¹ with a mean value of 0.03 °C-yr⁻¹, and the proportion of the area with an increasing trend was 86.59%, while only 13.41% of the area had a decreasing trend. In general, the UANSTM showed an increasing trend during both daytime and nighttime, and the increasing trend was stronger during daytime than nighttime.

The results of the Sen’s slope analysis and the Mann–Kendall trend test were superimposed, and the trends were classified into the following six levels according to Table 4: “Very significantly decreased”, “Significantly decreased”, “Not significantly decreased”, “Not significantly increased”, “Significantly increased”, and “Very significantly increased”. From the results (Figure 6), the trend of LST in most of the area of the UANSTM was insignificant (“Not significantly decreased” and “Not significantly increased”) in both daytime and nighttime, and only a small portion of the area had a significant or very significant trend (“Very significantly decreased”, “Significantly decreased”, “Very significantly increased”, and “Significantly increased”). During the daytime, the proportion of the area with an insignificant trend was 84.29%, while the proportion of the very significant and significant area was only 14.25%. Among them, the proportions of “Significantly increased” and “Very significantly increased” were 9.55% and 5.39%, respectively, mainly in Urumqi City, Wujiaqu City, Changji Hui Autonomous Prefecture, and Shihezi City, and the proportions of “Significantly decreased” and “Very significantly decreased” were 0.22% and 0.55%, mainly in Kuitun City. During the nighttime, the proportion of areas with insignificant trends was 93.43%, and only 6.57% of the areas were very significant and significant. The proportions of “Very significantly increased” and “Significantly increased” were 5.19% and 1.18%, respectively, with a discrete distribution and small clusters in Urumqi, the eastern part of Changji Hui Autonomous Prefecture, and Turpan City. The proportions of “Very significantly decreased” and “Significantly decreased” areas did not exceed 0.1%, and there were no obvious clustering areas.

4.3.2. The Future Changing Trend of LST

From the results of the Hurst index (Figure 7), during the daytime, the percentage of the area with sustained future trend (Hurst > 0.5) was 50.44%, and the percentage of the area with an unknown future trend (Hurst ≤ 0.5) was 49.56%. During the nighttime, the proportion of the area with a sustained future trend was 66.85%, and the proportion of the area with an unknown future trend was only 33.15%.

The Hurst index was coupled with the trend level, and the future trend of LST was classified into seven levels based on Table 4: “Sustained very significant decrease”, “Sustained significant decrease”, “Sustained no significant decrease”, “Sustained very significant increase”, “Sustained significant increase”, and “Sustained no significant increase”. During the daytime, 13.49% of the area showed a sustained decrease in the future trend of LST (“Sustained very significant decrease”, “Sustained significant decrease”, and “Sustained no significant decrease”), and 36.95% of the area showed a sustained increase (“Sustained very significant increase”, “Sustained significant increase”, and “Sustained no significant increase”). Among them, “Sustained very significant increase” and “Sustained significant increase” are mainly distributed in Urumqi, Wujiaqu, Changji Hui Autonomous Prefecture, and Shihezi, which means that the LST in these regions will continue to significantly and consistently increase in the future. During the nighttime, the proportion of the area that sustained an increase was 62.33%, while the proportion of the area that sustained a decrease was only 2.73%. The proportion of the area that “Sustained no significant increase” was 56.02%, which means that in the future, although the LST in most of the areas at night has a sustained increasing trend, the increasing trend is not significant (Figure 8).

4.4. Analysis of the Influencing Factors of LST

4.4.1. Discretization of Continuous Variables

When using the geo-detector model for influencing factor exploration, the independent variables are required to be type values. In this study, all variables except LULC are continuous type variables, thus requiring discretization of NDVI, NDBSI, precipitation, AOD, GDP, POP, DEM, and gradient data. We first defined five discretization methods, including: natural breaks, equal interval, quantile, standard deviation, and geometric interval methods. Then, we selected the optimal discretization method and the number of discretization intervals for daytime and nighttime based on the maximum q value by using the optidisc (optimal discretization for continuous variables and visualization) function in the GD package of R.

Table 6. The optimal discretization method and number of discretization intervals for continuous variables.

Continuous Variables	Daytime		Nighttime
	Method	Number of Intervals	Method	Number of Intervals
NDVI	standard deviation	7	standard deviation	7
NDBSI	natural breaks	7	standard deviation	5
Precipitation	natural breaks	6	natural breaks	6
AOD	standard deviation	7	standard deviation	7
NTL	quantile	7	quantile	7
GDP	quantile	6	quantile	7
POP	quantile	7	quantile	7
DEM	standard deviation	7	standard deviation	5
Gradient	standard deviation	7	standard deviation	7

shows the selected discretization methods and the number of discretization intervals for each influence factor during the daytime and nighttime.

4.4.2. Impact of Each Influencing Factor on LST in the UANSTM

The factor detector results show the degree of impact of each influencing factor on LST. From the results (Figure 9), the impacts of each influence factor on LST differed between daytime and nighttime, and the impacts were greater in the daytime than in the nighttime.

In terms of the influencing factors, regardless of daytime or nighttime, precipitation had the greatest influence on LST, followed by DEM and AOD, and the q values of these three influencing factors were above 0.5 in the daytime and above 0.4 in the nighttime, which were obviously higher than other influencing factors. This indicates that precipitation, DEM, and AOD were the main factors influencing LST in the UANSTM. In terms of the types of influencing factors, regardless of daytime or nighttime, climatic factors (precipitation and AOD) had the greatest influence on LST, with average q values of 0.6476 and 0.5074 for daytime and nighttime, respectively, followed by topographic factors (DEM and gradient) and land cover factors (LULC, NDVI, and NDBSI). The socioeconomic factors (NTL, GDP, and POP) had the least effect on LST, with mean q values of 0.1747 and 0.0838 for daytime and nighttime, respectively.

4.4.3. Interaction between Influencing Factors

In this study, the interaction detector was used to detect the combined effect of two influencing factors on LST and the interaction relationship (Figure 10). In both the daytime and the nighttime, the influencing factors only showed two types of bivariate enhancement and nonlinear enhancement, there were no independent influencing factors, and the combined impact of the two influencing factors was greater than the impact of a single influencing factor on LST.

In the daytime, only the interaction relationships of NTL ∩ DEM and NTL ∩ gradient showed nonlinear enhancement, while the rest showed bivariate enhancement. The interaction of precipitation with other influencing factors was very strong, with q values above 0.7. The largest q value for precipitation ∩ LULC (q = 0.8325) indicated that the interaction between precipitation and LULC had the greatest effect on daytime LST in the UANSTM. In the nighttime, the nonlinear enhancement relationship only occurred between socioeconomic factors (NTL, GDP, and POP) or between socioeconomic factors and other influencing factors. The interaction effect of precipitation with other influencing factors remained very strong. The interaction q value between precipitation and DEM was the largest (q = 0.7481), which means that the interaction between precipitation and DEM had the largest effect on nighttime LST in the UANSTM.

4.4.4. Risk Interval Analysis of Each Influencing Factor

In this study, we used the risk detector module to analyze the risk intervals of each influencing factor during daytime and nighttime in the UANSTM. The risk detector can count the average LST within each discretized interval for each influence factor, and we defined the interval with the highest average LST as the risk interval. From the results of the risk detector (Figure 11), precipitation and DEM were negatively correlated with LST, while AOD was positively correlated with LST, both during the daytime and the nighttime. The risk interval of LST for precipitation, DEM, and AOD during daytime occurred in the intervals of 97–664 mm for precipitation, −227–−48.5 m for DEM, and 0.285–4 for AOD. The nighttime risk interval was within the intervals of 97–588 mm for precipitation, −227–163 m for DEM, and 0.285–4 for AOD. From the LULC perspective, the LULC daytime risk interval occurred in the unused land type because the unused land is basically desert and the daytime temperature is very high. The nighttime risk interval occurred in the construction land type because the desert cools quickly at night, while the temperature is higher in the construction land due to various human activities. The mean LST fluctuated with NDVI, NDBSI, and gradient. The NDVI and gradient daytime and nighttime risk intervals were the same, both occurring in the 0.01–0.06 interval for NDVI and the 2.43–7.75° interval for gradient. NDBSI daytime and nighttime risk intervals differed little, occurring in the intervals 0.612–0.647 and 0.607–0.648, respectively.

4.4.5. Differences in the Impact of Influencing Factors on LST

The ecological detector can detect whether the influence of each influencing factor on the spatial distribution of LST is significantly different. The results showed (Figure 12) that, regardless of daytime or nighttime, most of the influencing factors had significant differences in their effects on LST, and only a few influencing factors had no significant differences. Among them, the effects of precipitation and DEM were not significantly different. That is to say, the influence mechanisms of precipitation and DEM on LST were similar. From the spatial distribution, precipitation and DEM had a high consistency in spatial distribution, where the higher the DEM, the higher the rainfall, and from the results of the risk detector (Figure 11), it can be seen that there was a negative linear relationship between precipitation and DEM with LST, where the higher the precipitation and DEM, the lower the LST. Therefore, there was no significant difference between the effects of precipitation and DEM on LST. The socioeconomic factors were very closely related to each other, and their spatial distribution was highly consistent, with high-value regions of NTL indicating that their GDP and POP were also high. Therefore, there was no significant difference between their effects on LST.

5. Discussion

5.1. Spatial Distribution Patterns and Changing Trend of LST

Turpan is an extremely hot region in China with a mean interannual LST maximum above 75 °C [60,61]. This is consistent with our findings, where we found that Turpan is the main concentration of HT and EHT in both the daytime and nighttime. The main reason is that most of Turpan is a desert, which leads to direct solar radiation and fast warming in the daytime. In addition, Turpan is a typical graben basin, resulting in poor air mobility and slow heat dissipation. The cold island phenomenon in oases was first proposed in [62]. The UANSTM is a typical oasis urban agglomeration, and both the spatial distribution pattern of LST and the risk detector results of LULC indicate a cold island phenomenon during the daytime and a heat island phenomenon during the nighttime in the oasis region. This is consistent with the results of [63] based on MODIS LST data and meteorological station information for the Tarim Basin. However, some studies have shown that in some developed urban agglomerations in China, such as the Beijing-Tianjin-Hebei urban agglomeration [51] and the Pearl River Delta urban agglomeration [33], they exhibit a strong heat island phenomenon in both the daytime and the nighttime. The main reason for this difference is that the periphery of the oasis is mostly desert, bare land, and low-coverage grassland, which have low specific heat capacity and warm up fast by direct solar radiation in the daytime. The interior of the oasis has more artificial blue and green landscapes, which have a high specific heat capacity and warm up slowly. Therefore, in the daytime, the temperature inside the oasis is lower than that at the periphery of the oasis, while the opposite is true in the nighttime [64]. In addition, in arid regions, agriculture is dominated by irrigated farming, and some studies have shown that irrigation can enhance the cold island phenomenon in oases [65,66].

Since the 1960s, the biggest feature of global climate change has been large-scale warming [67]. However, intermittent global warming has also been found, which could also explain the fluctuation of the interannual mean LST to some extent [68]. In this study of change trends, we found that LST exhibited an increasing trend in both the daytime and the nighttime, with an average increase of 0.05 °C per year in the daytime and 0.03 °C per year in the nighttime. This is consistent with the change trend of LST in the whole of China. The authors of [53] explored the change trend of LST in China from 2003 to 2019 and found that China’s LST showed an increasing trend in both the daytime and the nighttime, and the daytime was stronger than the nighttime. It is noteworthy that there was a significant increasing trend of daytime LST in Urumqi, Wujiaqu, Changji Hui Autonomous Prefecture, and Shihezi during 2005–2021, and the same trend will remain in the future. This may be due to the following two reasons: (1) The “Urumqi-Changeji-Wujiaqu” metropolitan area and Shihezi have developed rapidly in recent years, and strong human activities and intense land use changes have led to the increase in LST. (2) The “Urumqi-Changeji-Wujiaqu” metropolitan area and Shihezi are the key development areas in the development plan of the UANSTM and will continue to lead the development of the Tianshan north slope urban agglomeration in the future [44].

5.2. Influencing Factors of LST

Based on the results of the GDM, we found that precipitation has the greatest impact on the LST of the UANSTM. Water shortage has been the most central problem limiting the development of the Xinjiang region [43]. Water is a natural heat sink due to its high specific heat capacity and slow warming, and precipitation can directly carry away a large amount of heat from the surface and increase surface runoff. There was also a significant correlation between precipitation and vegetation cover, so precipitation can also indirectly affect LST by influencing vegetation [69,70]. We found that AOD also has a strong impact on the LST of the UANSTM. Aerosol particles can influence the absorption of solar short-wave radiation at the surface and the radiation exchange with the atmosphere, which leads to changes in LST [71]. The Gurbantunggut desert to the north of the UANSTM is one of the major global sources of sand and dust aerosol emissions, which, together with the large amount of anthropogenic aerosol particle emissions, has led to serious aerosol particle pollution in the UANSTM [72,73]. The influence of DEM on LST is second only to precipitation, which may be related to the large difference in altitude between the UANSTM. The difference between the highest and lowest elevations in the UANSTM is above 5000 m. In the high-altitude region of the Tianshan Mountains, which is covered with snow all year round, LST is extremely low, while Turpan, which has the lowest altitude, has the highest LST in China. We also found that socioeconomic factors (NTL, GDP, and POP) have very little impact on LST in the UANSTM. This is contrary to the findings of [33] for the PRD urban agglomeration, who found that NTL, GDP, and POP have a very significant effect on LST. The reason for this discrepancy may be that the UANSTM, as an urban agglomeration in the northwestern frontier of China, has a wide spatial extent but a low urbanization rate, a relatively backward economic level, and a sparse population compared with other developed urban agglomerations, resulting in a smaller impact on LST.

5.3. Research Limitations and Future Work

Any geographical element has scale effects, and LST is no exception [74]. This means that there may be significant differences in study results at different temporal and spatial scales. This study deeply analyzed the spatial distribution pattern, changing trend, and various influencing factors of LST in the UANSTM, and provided a theoretical basis for the prevention and management of urban thermal environment problems, but there are still some limitations. The first is that this study analyzed the trend of LST only from the interannual scale, and in the future, we need to consider the variation on the seasonal scale, which will be more instructive for agricultural production [75]. Second, although the MODIS data have high temporal resolution, the spatial resolution is poor, which makes it difficult to reflect the details of changes when studying the LST within cities. Therefore, combining high spatial resolution data with downscaling fusion of MODIS data to obtain LST data with high spatial and temporal resolution will be an important direction for future urban surface thermal environment studies [76,77]. Third, although we considered 10 influencing factors in terms of climate, land cover, socioeconomics, and topography, LST is influenced by multiple factors, and more influencing factors need to be considered in future studies. Finally, we only used the geo-detector model to explore the influencing factors, although it possesses higher accuracy than the ordinary linear model. However, other methods may work better, such as random forest regression, logistic regression, and geographically weighted regression models [78,79,80].

6. Conclusions

This study analyzed the spatial distribution pattern of LST based on MODIS LST data, with the UANSTM as the study area. Absolute variability, Sen’s slope analysis, and the Mann–Kendall trend test were used to analyze the fluctuation characteristics and change trends of LST from 2005 to 2021, and the future change trends of LST were explored with the Hurst index. Finally, the influence of 10 factors, such as LULC, NDVI, NDBSI, and precipitation, on LST was analyzed by using the geo-detector model based on the optimal parameters. The detailed findings of this study are as follows:

1.

In terms of the spatial distribution pattern, both daytime and nighttime, the LST classes of the UANSTM are dominated by MT and HT, with EHT and HT clustered in Turpan city. As far as the diurnal differences are concerned, in the oasis region, the cold island feature is observed during the daytime and the heat island feature is observed at night.

2.

During 2005–2021, the northwestern part of the urban agglomeration on the northern slopes of the Tianshan Mountains fluctuated greatly, especially in Karamay and Usu. The LST showed an increasing trend in both the daytime and the nighttime, with an increase rate of 0.05 and 0.03 °C/yr⁻¹, respectively. The increasing trend of LST in Urumqi, Changji Hui Autonomous Prefecture, Shihezi, and Wujiaqu was very significant during the daytime and still showed a significant increasing trend in the future, which needs to be investigated.

3.

The climatic and topographic factors of precipitation, DEM, and AOD are the main factors affecting the LST of the UANSTM, and they all have q values above 0.5 during the daytime and above 0.4 during the nighttime. The effects of land cover factors (LULC, NDVI, and NDBSI) was the second most important, and socioeconomic factors (NTL, GDP, and POP) had the least influence on LST. The interactions between the influencing factors all showed enhancement types (nonlinear enhancement and bivariate enhancement), and one of the two influencing factors that showed the nonlinear enhancement type must be a socioeconomic factor.

4.

Precipitation and DEM showed a negative linear correlation with LST, while AOD showed a positive linear correlation with LST. The mean LST values for each interval of precipitation and DEM decreased as the values within the interval increased, and the risk interval for precipitation and DEM occurred in the lowest value interval, while the mean LST values for each interval of AOD increased as the values within the interval increased, and the risk interval for AOD occurred in the highest value interval.