Air Temperature Estimation with MODIS Data over the Northern Tibetan Plateau

Near-surface air temperature (T _a) is one of the key factors used to monitor climate variations (IPCC, 2013). T _a, which is observed in-situ at meteorological stations, is processed into gridded data to evaluate climate models and predict climate changes. For example, the National Oceanic and Atmospheric Administration/National Climatic Data Center aggregate monthly global surface temperature anomaly data into 5^° grid boxes (Smith et al., 2008). T _a plays an important role in heat transfer, the moisture cycle, and the land surface energy balance (Savage et al., 2009; Ma et al., 2014; Wang and Zeng, 2015). Moreover, the difference (T _d) between surface temperature (T _s) and T _a (hereafter, T _d), which is a significant indicator of the land-air mass exchange and energy balance, is one of the fundamental parameters used to study land-atmosphere interactions (Ma et al., 2006b, 2011).

T _a data have a high temporal resolution, and are widely used to study climate variations. However, these measurements and gridded T _a data suffer from relatively lower spatial resolutions and uncertainty in areas where the terrain is complex and whose effect on climate is vital, especially on the Tibetan Plateau. Furthermore, insufficient knowledge regarding the spatial distribution of T _a influences land-atmosphere interaction research, climate change and model simulations (Ren et al., 2014). Therefore, methods are urgently needed to estimate the spatial distribution of T _a at a high spatial resolution.

Figure 1.  The (a) location of the study area and (b) distribution of sites in the northern Tibetan Plateau.

(Hijmans et al., 2005) produced high-accuracy global air temperature and precipitation data with latitude, longitude and elevation as variables using the thin-plate smoothing spline algorithm. With the limitation of sparse sampling points and no consideration of soil moisture and vegetation coverage, the method exhibits considerable uncertainty in estimating T _a. (Ren et al., 2010) constructed time series of monthly temperature anomaly gridded data from 1951 to 2006 over the Tibetan Plateau using an area-weighted average method; however, the subsequent spatial distribution was unsatisfactory. With the development of remote sensing techniques, using T _s data from satellites to estimate T _a has become a feasible approach, and these estimations have shown good results. For instance, Moderate Resolution Imaging Spectroradiometer (MODIS) T _s data can be used for linear regression estimations of daily maximum and minimum T _a (Mostovoy et al., 2006). (Good, 2015) developed a dynamic multiple linear regression model for estimating daily maximum and minimum T _a using daily maximum and minimum T _s, vegetation fraction, distance from the coast, latitude, urban fraction, and elevation as predictors. (Sun et al., 2005) described a new approach for estimating T _a from T _s based on thermodynamics, but the method required various parameters (aerodynamic resistance, air density, etc.) that are difficult to obtain. Then, (Sun et al., 2014) developed a new algorithm for estimating mean T _a using a combination of MODIS day-night T _s and enhanced vegetation index data, and they achieved good results.

Moreover, various attempts to accurately estimate T _a have been made. (Nieto et al., 2011) estimated T _a with Meteosat Second Generation/Spinning Enhanced Visible and Infrared Imager data using a calibrated Temperature-Vegetation Index algorithm, and improved the precision of T _a estimation by estimating the maximum normalized difference vegetation index (NDVI) based on the observed T _a for different vegetation types. (Zhang et al., 2011) established regression models considering solar declination to estimate the daily maximum, minimum and mean T _a, concluding that nighttime T _s was optimal for estimating these temperatures. (Hengl et al., 2012) produced gridded temperature maps via spatiotemporal regression-kriging, and the average accuracy of mapping the temperature was 2.4^°C. Similarly, (Kilibarda et al., 2014) proposed a mapping framework for estimating the daily mean, minimum and maximum T _a using spatiotemporal regression-kriging, and the average accuracy was 2^°C.

The present study expands upon estimates of T _a at high spatial resolutions, and its main objective is to propose a new method for estimating T _a based on MODIS T _s, NDVI, and elevation data over the northern Tibetan Plateau (NTP), where few meteorological stations are distributed. Spatial distribution maps of monthly mean T _a and T _d were calculated and analyzed over the NTP. To assess the performance of this methodology, the estimation models were tested using statistical methods, and estimated T _a values were compared with T _a values measured in situ and T _a maps based on Institute of Tibet Plateau Research, Chinese Academy of Sciences (ITPCAS) reanalysis data (He and Yang, 2011).

The NTP (Fig. 1), whose latitude and longitude are 31^°-36^°N and 90^°-95^°E, is located between Mount Kunlun and Mount Nyainqen Tanggula on the Tibetan Plateau. The dominant vegetation types are alpine grassland and meadow, and the average elevation is above 4 km. A sub-frigid plateau zone can be found in the area, and the main distinguishing features of the climate are alpine-cold, oxygen-deficient and arid ecosystems, as well as intensive solar radiation and long sunshine duration. A dry climate with lower temperatures and strong winds occurs from November to March. From May to September, the climate is relatively warm and humid, and is the period when 80% of the annual precipitation occurs and vegetation development is at its strongest. The land surface is dry and covered by grassland and sensible heat flux is dominant before the onset of the monsoon. The region is covered by meadows and dominated by latent heat flux during the onset phase of the monsoons (Li et al., 2011).

Six field weather stations and seven service meteorological stations are distributed in this area. All stations except Amdo and BR, due to data insufficiencies, were used to perform the regression analysis. Most of these stations are distributed along the Qinghai-Tibet railway (Fig. 1b), and the basic information for all stations is listed in Table 1.

2.2.1. In-situ observations

Two types of in-situ observations were used in this study. The first included field weather station data from the Nagqu Plateau Climate and Environment (NPCE) station on the NTP. This dataset included observations of meteorological elements such as T _a, humidity, air pressure, T _s, wind speed and direction, soil temperature and moisture, precipitation, and solar radiation, at the NewD66, D105, NPAM, BJ and MS3608 sites (Ma et al., 2006). The observational time intervals of the NPCE data were 10 min, 30 min or 1 h. The T _a data from NPCE, the temporal range of which was from 2001 to 2013, were gathered at an observation height of 1.05 m using an HMP45D temperature and humidity sensor, which had an observation accuracy of 3%. Furthermore, the in-situ T _a data were processed into monthly mean T _a data to meet the needs of this study. Additionally, meteorological station data from the China Meteorological Data Sharing Service System (http://cdc.nmic.cn) were used in this analysis. Observations from 756 meteorological stations from 1951 to 2014 in China were collected, and monthly mean T _a data from 2001 to 2013 at the WDL, TTH, Ando, NQ, BG, SX and BR sites on the NTP were used (see Fig. 1b). Particularly, in order to distinguish the meteorological station Anduo from the field weather station Amdo, the former was written as Ando.

2.2.2. MODIS products

The T _s data and NDVI data applied in this study were all from MODIS, aboard the Terra satellite. The MODIS products used in this paper included MOD11C3 V5 (MODIS/ Terra Land-Surface Temperature/Emissivity Monthly Global 0.05 Deg CMG) and MOD13A3 V5 (MODIS/Terra Vegetation indices Monthly L3 Global 1 km SIN Grid). These products cover the period from 2001 to 2013 and are provided by LAADS (NASA Level 1 and Atmosphere Archive and Distribution System: http://ladsweb.nascom.nasa.gov).

For consistency, the coverage area and spatial resolution of the T _s and NDVI data were processed by the MODIS Reprojection Tool (MRT), as shown in Table 2.

2.2.3. Auxiliary data

Two additional datasets——a 1 km resolution Digital Elevation Model (DEM) of China and the Chinese Meteorological Forcing Dataset——were also employed in this study. These datasets were provided by the Cold and Arid Regions Sciences Data Center at Lanzhou (http://westdc.westgis.ac.cn). The first dataset was used to estimate T _a in the following GWR model and the second, whose spatial and temporal resolutions were 0.1^° and 3 h, respectively, were used to validate the estimated T _a on the NTP. The Chinese Meteorological Forcing Dataset is a set of near-surface meteorological and environmental factor reanalysis data developed by ITPCAS, based on Princeton reanalysis data, Global Land Data Assimilation System data, Global Energy and Water Cycle Experiment-Surface Radiation Budget radiation data, and Tropical Rainfall Measuring Mission precipitation data. The ITPCAS reanalysis data contained seven meteorological variables: T _s, air pressure, specific humidity, wind speed, ground precipitation rate, downward shortwave radiation, and longwave radiation. Additionally, ITPCAS reanalysis T _a data were used in this study, and were processed into monthly T _a data to compare them with the estimated T _a data.

Some statistical tests and indicators were used to evaluate the models of estimating T _a. First, two statistical tests, the F-test and T-test, are introduced. The F-test, a significance test for general linear equations, tests the significance of a linear relationship of a model in terms of sample observations. The F-test model is calculated as follows: $$ Y_i=\beta_0+\beta_1X_{1i}+\beta_2X_{2i}+\cdots+\beta_kX_{ki}+u_i . $$ We must test whether the parameter β_k is equal to zero. According to the principles and procedures of statistical hypothesis testing, the original hypothesis and alternative hypothesis are as follows: \begin{eqnarray*} &H_0:\beta_1=0,\beta_2=0,\cdots,\beta_k=0;&\\ &H_1:\beta_j(j=1,2,\cdots,k)\hbox{ are not all zero}.& \end{eqnarray*} The F-stat, where \begin{eqnarray*} F=\dfrac{{\rm ESS}/k}{{\rm RSS}/(n-k-1)}\\[-2mm] \end{eqnarray*} obeys the F-distribution with (k,n-k-1) degrees of freedom under the condition that the original hypothesis H₀ is tenable. Additionally, \(\rm ESS=\sum(\hat{Y}_i-\bar{Y})^2\) and \(\rm RSS=\sum(Y_i-\hat{Y})^2\). \(\hat{Y}_i\) represents fitted values and \(\bar{Y}\) is the average value of actual values Y_i. The variable n stands for the sample number, and the variable k represents the number of regressions. Then, the F-value is compared with the F_α value at a given significance level α. If F>F_α(k,n-k-1), the original hypothesis is rejected. If F≤ F_α(k,n-k-1), the original hypothesis is accepted. The F-stat is the statistical value of the F-test. Nevertheless, for a multiple linear regression model, the significance of the overall linear relationship cannot explain the significance of the effect of every explanatory variable on the explained variable. Therefore, significance tests must be performed for the explanatory variables to demonstrate their effects on the model. The T-test is a significance test for variables in a multiple linear regression model. The original hypothesis and alternative hypothesis of the T-test are as follows: \begin{eqnarray*} H_0:\beta_j=0,\\ H_1:\beta_j\neq0. \end{eqnarray*} The absolute value of T is compared with T_α/2 at a given significance level α. If |T|>T_α/2(k,n-k-1), the original hypothesis is rejected. If |T|≤ T_α/2(k,n-k-1), the original hypothesis is accepted. The T2-stat is the statistical value of the T-test for the coefficient of the NDVI, and the T3-stat is the statistical value of the T-test for the coefficient of elevation.

The following additional statistical indicators were also used. The determination coefficient R², also called the goodness-of-fit, tests the fitting degree of sample observations from the perspective of an estimation model; the higher the value, the better the fitting degree of the model. However, the R² value increases with an increase in the number of explanatory variables. To eliminate the effect of the number of variables on the goodness-of-fit, the adjusted determination coefficient R _Adj² is proposed. In addition, the root-mean-square error (RMSE), also known as the standard error, is used to explain the dispersion degree of the sample observations. The estimation of error variance, σ², represents the variance in the error caused by random factors. Moreover, pattern correlation is based on the Pearson product-moment coefficient of the linear correlation between two variables, which are the values of the same variables at corresponding locations on two different maps. The anomaly correlation used in this study is a special case of pattern correlation, for which the variables being correlated vary from some appropriately defined mean, most commonly a climatological mean. This correlation indicates the similarity between the estimated T _a and the T _a from the ITPCAS reanalysis data.

Regression analysis is a type of statistical analysis technique that determines the quantitative relationship between two interdependent variables. Ordinary linear regression (OLR), a linear regression model that contains one explanatory variable, seeks to reveal the linear relationship between an explained variable and an explanatory variable, and the mathematical model is as follows: $$ \mathcal{Y}=\beta_0+\beta_1\mathcal{X}+\varepsilon, $$ where β₀ and β₁ are unknown parameters, called the regression constant and regression coefficient, respectively. ε is the random error.

The OLR model established above is a global model, and the value of the regression coefficient estimated from it, which is the mean value of the entire study area, cannot reflect the real spatial distribution. The geographically weighted regression (GWR) model is an expansion of the OLR model, with the spatial locations of data used in a linear regression equation. The GWR model is constructed via regression of the observed T _a values (in ^°C) with the temporally and spatially collocated T _s (in ^°C), NDVI, and elevation (Ele, in m) data. The regression formula is as follows: $$ T_{\rm a}=a_0+a_1T_{\rm s}+a_2{\rm NDVI}+a_3{\rm Ele}. $$ a₁,a₂ and a₃ are the regression coefficient of T _s, NDVI and Ele, respectively; a₀ is a constant term. These selected variables are all significantly correlated with T _a (Good, 2015). Other variables, such as wind speed, solar zenith angle (SZA) and albedo, were not added to the model, as SZA and albedo were closely related to T _s and NDVI, respectively. Additionally, including these variables did not improve the model. Furthermore, due to the absence of surface data, wind speed data were not applied to the estimation.

Figure 2.  Temporal variation of the monthly MODIS T _s and T _a at station BJ from 2001 to 2012.

4.1.1. Temporal variations of monthly T s and T a values at each station

The temporal variations of the monthly T _s and T _a values at 11 stations were analyzed, and the trends were similar at all stations. Particularly, BJ station was taken as an example. The results showed that the annual and interannual variations of T _a were similar to those of T _s at this site (Fig. 2). In terms of the annual variations, T _s and T _a started to rise gradually in January, peaking in July. Then, they exhibited a decreasing trend. The temporal variations of T _s and T _a at BJ were roughly consistent, and the annual variations of T _s exhibited double-peak variations in some years, with T _s values greater than T _a.

4.1.2. OLR analysis

Based on the time-series analysis, the OLR analysis was performed using the in-situ monthly T _a and T _s values at the 11 stations, and the results are shown in Table 3. The determination coefficients (R² and R _Adj²) at these 11 stations were all above 0.85 (0.88-0.93), and the RMSE values were relatively low (1.75^°C-2.65^°C).

Next, the OLR analysis was separately performed using the in-situ T _a and T _s values in each month at these stations, and the results, including six statistical indicators, are displayed in Table 4. From the perspective of F-statistics (p<0.01), the simple linear regression model based on T _s was statistically significant for estimating T _a. The determination coefficients (R² and R _Adj²) in each month were between 0.40 and 0.80, and they were smaller than 0.60 in February, March and August. In addition, except from April to September (RMSE < 1.0^°C), the accuracy of these models for estimating monthly T _a was not high, with RMSE values between 1.0^°C and 2.10^°C. This result suggests that T _s can be a crucial variable for estimating T _a, but other important variables may not be considered in the models. To estimate T _a more accurately, some other variables that potentially influence T _a should be added to the models.

Table 5 shows the results of the GWR model in each month. According to the analysis, the regression equations of the models used to estimate T _a in each month all passed the F-test with a significance level of 0.01 and were considered significant at that level. In terms of the T-test of the significance of regression coefficients, the variables NDVI and Ele had remarkable effects on the models at the 0.01 significance level. Furthermore, the determination coefficients (R² and R _Adj²) in these 12 months, which were within the range of 0.80 to 0.96, were all larger than 0.80. The RMSE values, which were less than 1.0^°C between April and December, were all less than 1.12^°C. These results indicated that the variables T _s, NDVI and Ele displayed excellent abilities to explain changes in T _a. Compared with the result of the OLR model, the GWR model, which added NDVI and Ele, greatly improved the estimation accuracy of T _a. Due to the consideration of fewer factors, the estimation accuracy of the OLR model was relatively low. Therefore, the GWR model with high accuracy was used to estimate T _a over the NTP.

4.3.1. Comparison of estimated T a and in-situ measured data

To evaluate the effect of the GWR model, observed T _a data from 2013 at the 12 stations mentioned above were selected to compare with the estimated T _a data. The results are shown in Fig. 3 in the form of a scatter plot, which indicates that the estimated T _a values calculated by the GWR model were in good agreement with the observed T _a data measured at weather stations, with R² values greater than 0.98 at each station, except for BR. In addition, the estimated T _a values were lower than the observed T _a values at the BR site, and the estimated T _a values were slightly larger than the observed T _a values at the WDL and TTH sites. The estimated T _a values were similar to the observed T _a values at other sites. For a more detailed comparison, a statistical index of absolute error was selected to evaluate the consistency between the estimated and observed T _a values. The absolute error values between the observed and estimated T _a values in 2013 are provided in Table 6. As shown in the table, the absolute error values at the BR site were the largest (4^°C-10^°C), and the values in November at NewD66 and in March, April and September at WDL were between 2.0^°C and 4.0 ^°C. Otherwise, the value was within 2^°C in each month at each site. The estimation accuracy of the GWR model was relatively low in the southeastern NTP area.

Figure 3.  Comparison between the observed T _a and estimated T _a at all sites in 2013.

As discussed above, the difference between the estimated and measured T _a values at BR was relatively large, and it may have been influenced by the varying topography and underlying surface. According to the annual variation of T _s over the NTP (Huang et al., 2016) in the vegetation growing season, T _s values over the NTP are relatively high overall and can peak in July and August. However, T _s values at the BR site were relatively low in July and August, which was inconsistent with the T _s variation pattern on the NTP. Moreover, the NDVI values at BR were extracted abnormally by MODIS, and NDVI values were far below those at other stations, even in August. This result suggests that NDVI values extracted by MODIS are unable to accurately reflect the vegetation pattern over the southeastern NTP. In addition, the BR site is located in the southeastern area of the NTP, which has a relatively low elevation, complex topography and dense vegetation (Fig. 1). Besides, the T _s and NDVI data at the BR site were all from MODIS data. Therefore, it was concluded that the applicability of the MODIS T _s extraction algorithm was poor in the southeastern area of the NTP, and T _s extraction data, which exhibited higher errors, failed to show the variation in T _s. Thus, relatively larger deviation occurred at the BR site.

Figure 4.  Temporal variation of the estimated T _a values and observed T _a values at four stations.

Figure 5.  The 12-year monthly mean spatial distribution of the estimated T _a values from GWR models (a1-d1) and measured T _a values from ITPCAS reanalysis data (a0-d0) in different seasons (units: ^°C; a0 and a1 are for January; b0 and b1 are for April; c0 and c1 are for July; d0 and d1 are for October).

Figure 6.  Spatial distribution of monthly mean T _a from 2001 to 2012 over the NTP (units: ^°C).

Figure 7.  Spatial distribution of monthly mean T _d from 2001 to 2012 over the NTP (units: ^°C).

Four weather stations distributed from north to south over the NTP were selected to carry out a temporal comparison between the estimated and observed T _a values. The temporal variations in the estimated and observed T _a values at these four stations are presented in Fig. 4, which reveals that the estimated T _a was in good agreement with the observed T _a temporally, especially at D105 and BJ. The estimated T _a values were slightly higher than the measured T _a values at WDL station, which is located on the northern NTP, and the estimation error varied from -1.87^°C to 3.96^°C. A more significant error may have existed in the southeastern NTP area, and the estimation of high temperatures was better than that of low temperatures at the SX site.

4.3.2. Comparison of estimated T a from GWR models with T a from ITPCAS reanalysis data

The spatial distribution of the estimated T _a from GWR models was also compared with the T _a data from the ITPCAS reanalysis data created by (He and Yang, 2011). This dataset has been widely used in climate research, with high precision and quality. The 12-year monthly mean spatial distributions of the estimated T _a values from the GWR model and measured values from the ITPCAS reanalysis data in different seasons are illustrated in Fig. 5, which shows that the estimated and measured values agreed spatially. Besides, in order to compare with the estimated T _a data (Figs. 5a1-d1), the T _a data from the ITPCAS reanalysis data were processed into monthly mean T _a values from 2001 to 2012 (Figs. 5a0-d0). Two statistical variables, including the pattern correlation coefficient and mean bias, were selected to quantitatively analyze the relationship between the estimated and measured T _a values. According to the analysis results of the pattern correlation coefficients and mean biases in Table 7, we can conclude that the spatial distribution of the estimated T _a was most similar with the T _a from ITPCAS reanalysis data in January, while the similarity was relatively low in April. Moreover, the estimated T _a values from the GWR model were smaller than the T _a values from the ITPCAS reanalysis data in January, April and October, and larger in July. In particular, as can be seen from Fig. 5, compared with T _a values from the ITPCAS reanalysis data, the T _a estimation values were relatively low in the southwestern area for all seasons.

Overall, the spatial distribution patterns of the estimated T _a agreed well with those of T _a from the ITPCAS reanalysis data. Because of the single underlying surface and higher latitude in the area north of the Tanggula Mountains, the spatial agreement between the estimated T _a and the measured T _a in this region was better than that in the area south of the Tanggula Mountains. In the southwestern NTP area, the estimated T _a values were lower than the measured T _a values. From the analysis above, it is clear that the spatial distribution of T _a can be successfully described by the MODIS data using the GWR model discussed in this paper.

The spatial patterns of monthly mean T _a and T _d from 2001 to 2012 over the NTP are separately presented in Figs. 6 and 7. One thing to be noted is that the T _a is an estimated T _a from GWR models. As shown in Fig. 6, the T _a values were all below 0^°C in January; in particular, the minimum reached -29.28^°C and the maximum -5.0^°C. In July, T _a values fluctuated between -0.53^°C and 14.0^°C. The variational range in the summer (from May to October) was -15.92^°C-14.0^°C. From October to December, the 0^°C isothermal level gradually declined from an elevation of 4-5 km, reaching a minimum value at an elevation of 3200 m in December. Then, T _a remains below 0^°C in January. The 10^°C isothermal level gradually began to increase from an elevation of 3200 m in May, peaking at an elevation of 4-5 km in July. Furthermore, T _a on the southern slope of the Tanggula Mountains was obviously higher than that on the northern slope. The T _a to the east of the Qinghai-Tibet Railway was higher than that to the west, and T _a showed an increasing trend from northwest to southeast.

Similarly, as shown in Fig. 7, the variational range of T _d was -6.52^°C-6.0^°C in January and -6.32^°C-6.0^°C in July. Excluding the southeastern area of the NTP, where T _d values were under 0^°C, T _d values in other areas were all above 0^°C in winter (from November to April). These results are similar to the conclusions made by (Zhang et al., 2007). In summer, the elevation of the area below 0^°C rose to 4-5 km, and T _d values to the south of the Tanggula Mountains were between -2.0^°C and 0^°C. Based on the map of frozen ground on the Tibetan Plateau produced by (Li and Cheng, 1996), the spatial distribution of T _d over the NTP was closely related with the distribution of seasonally frozen ground. In summer, when the seasonally frozen ground melts, negative T _d values were apparent in this area.

The temporal and regression analysis of MODIS T _s and observed T _a over the NTP conducted in this study showed that the temporal variations in T _a were similar to those of T _s, and T _a displayed good agreement with T _s. Therefore, MODIS T _s data can be considered a crucial factor in estimating T _a over the NTP. A new method for estimating monthly T _a from MODIS data is described in this paper, based on in-situ measurements, MODIS T _s, MODIS NDVI and DEM data. Monthly T _a values from 2001 to 2012 over the NTP were accurately estimated by the GWR method, with an estimation accuracy (R _Adj²>0.97; RMSE = 0.51^°C-1.12^°C) that was higher than that of the OLR method.

Observed T _a data in 2013 at 12 stations were selected to compare with the collocated estimated T _a data extracted from the regional dataset. Excluding station BR, which is located on the southeastern NTP, and particular months at stations NewD66 and WDL, the absolute error values associated with estimating the monthly T _a at other stations in all months were all less than 2^°C, with R² values greater than 0.98. Moreover, the estimated T _a values were in good agreement with the observed T _a values temporally. Through comparison with the T _a data provided by the ITPCAS reanalysis data created by (He and Yang, 2011), the estimated T _a data calculated using the GWR model were able to describe the spatial distribution of T _a over the NTP.

Based on the estimated results, the spatial distribution characteristics of the monthly mean T _a and T _d from 2001 to 2012 over the NTP were as follows. T _a values on the northern slope of the Tanggula Mountains were lower than those on the southern slope. According to the analysis of the 0^°C and 10^°C isothermals, T _a in the region with an elevation of 4-5 km was able to reach 10^°C in the summer, and T _a in the region with an elevation of 3200 m dropped to 0^°C in the winter. In addition, except in the southeastern area of the NTP, T _d values in other areas were all above 0^°C in the winter.

Figure 8.  Map of frozen ground on the Tibetan Plateau.

Some studies have proven that MODIS T _s values over the Tibetan Plateau share similar variational characteristics with T _a values in time and space, and that they exhibit a good linear relationship (Yao and Zhang, 2013). The approach proposed in this study is a regression-based method. This approach is feasible and has a relatively high estimation accuracy of T _a. In theory, the GWR model, which was created from in-situ observational data and MODIS data, can be applied to sparsely observed areas. The thermodynamic-based method presented by (Sun et al., 2005) is not practical because it requires input data that are not readily available. However, although the GWR model has a high estimation accuracy associated with estimating T _a, because it considers several factors that influence the variations in T _a, the model was constructed over surfaces largely covered by alpine grassland and meadows. Therefore, they can only be used to estimate T _a in a small area over the Tibetan Plateau. Moreover, the spatial resolution of 0.05^° meets the requirements of T _a estimation in large-scale areas, but it cannot satisfy the demands of T _a estimation in small-scale mountain areas. Therefore, the model requires supplementation and improvement before application over the entire plateau, which will be the focus of our next study.