Comparative Analysis of Three Near-Surface Air Temperature Reanalysis Datasets in Inner Mongolia Region

Abstract

Near-surface air temperature is important for climate change, agriculture, animal husbandry, and ecosystems undergoing climate warming in Inner Mongolia. Land surface reanalysis products feature finer spatial and temporal resolutions, that can provide important data support for the determination of crop growth limits, grassland biomass growth, and desertification research in Inner Mongolia. In this study, 119 in situ observed sites were collected to compare and evaluate the performance of near-surface air temperature in three reanalysis products from 2018 to 2020 in Inner Mongolia. The three reanalysis products included three widely used products derived from the European Centre for Medium-Range Weather Forecasts (ECMWF) Fifth Generation Land Surface Reanalysis (ERA5-Land), and U.S. Global Land Data Assimilation System (GLDAS), as well as the latest reanalysis product from the High-Resolution Land Data Assimilation System reanalysis product by the China Meteorological Administration (HRCLDAS). Results are as follows: (1) The three reanalysis temperature products all reasonably reflect the characteristics of spatial and temporal changes in surface temperature in Inner Mongolia. Compared with ERA5L and GLDAS, HRCLDAS is more consistent with the observed results. (2) For the evaluation period, HRCLDAS has a certain underestimation of temperature, while ERA5-Land and GLDAS have a significant overestimation of temperature. (3) During high-temperature processes, HRCLDAS is more accurate in simulating higher temperatures than ERA5-LNAD and can demonstrate the changes in high-temperature drop zones. The major conclusion of this study is that the HRCLDAS product demonstrates a relatively high reliability, which is of great significance for the study of climate, ecosystem, and sustainable development.

1. Introduction

As the most basic climate element, the near-surface temperature is an important indicator for measuring surface heat conditions and is also an indispensable factor in the study of climate change and related ecosystem characteristics, environmental change, disaster risk, and other research fields [1,2,3]. In the past century, the near-surface temperature has shown a significant upward trend, resulting in a huge impact. The sixth evaluation report of the IPCC points out that the global surface temperature has risen faster since 1970 than any other 50-year period. As climate warming intensifies, many changes in the climate system will be greater, and the intensity and frequency of extreme weather events such as high-temperature heatwaves and extreme droughts will increase [4,5,6]. Global warming will accelerate surface drying, leading to an increase in the severity and frequency of agricultural droughts, as well as potential changes in crop development cycles and yields [7,8,9].

Inner Mongolia spans the three major regions of Northeast, North China, and Northwest China. As the largest and most diverse ecological functional area in Northern China, the vegetation cover types from east to west are forest grassland, grassland, desert grassland, and desert, respectively. It is considered one of the most sensitive regions to global climate change [10,11,12]. Climate warming will lead to Inner Mongolia becoming a high-risk area for high temperatures, heatwaves, and extreme droughts, as well as a high-risk area for natural ecosystems and food production. Numerous researchers have found that the temperature rise rate in Inner Mongolia in the past 50 years has been significantly higher than the average annual warming rate in China and the world [13,14]. With the change in temperature, the boundary line of crops and the planting area of crops with different maturity periods in Inner Mongolia have undergone significant changes [15]. As pointed out by Qiao et al., the increase in accumulated temperature from 1959 to 2018 led to a significant northward and eastward expansion of the boundary lines for different types of spring corn, and the transition of spring corn from unsuitable planting areas to early and medium maturing types in the central and eastern regions [16]. Vegetation is also significantly affected by near-surface temperature. In the arid and semi-arid regions of Western Inner Mongolia, an increase in surface temperature can lead to an increase in soil evaporation and a decrease in soil moisture content. This means that the soil cannot provide sufficient water for grassland growth, leading to a decrease in productivity. At the same time, vegetation growth in higher altitude areas depends more on temperature [17,18,19]. Therefore, people are increasingly paying attention to the changes in near-surface temperature, and reliable temperature data can be used to study the trend of climate change in Inner Mongolia and the characteristics of grassland ecological environment changes.

However, most of the previous research data came from ground observation stations. Due to the vast area and harsh living environment of Inner Mongolia, there are relatively few observation stations, with a coverage rate of only 46.16% of near-surface temperature measurement points. There are large observation gaps in the eastern and western regions. The sparse and uneven distribution of meteorological stations inevitably leads to some uncertainties in climate change assessment [20]. Following the development of data fusion technology, high spatiotemporal resolution reanalysis datasets can be produced from the assimilation system by fusing station observations, numerical model forecasts, and satellite remote sensing data. These reanalysis products have the advantages of a broad coverage area and long time span, which can effectively address the problems of inhomogeneous temporal and spatial distribution of station observations [21,22]. The National Center for Environmental Prediction/National Center for Atmospheric Research (NCEP/NCAR), the European Center for Medium-Range Weather Forecasts (ECMWF), the Japan Meteorological Agency (JMA), the National Aeronautics and Space Administration (NASA) of the United States, and the China Meteorological Administration (CMA) can all provide long-term records of temperature reanalysis, such as datasets such as CFSR [23], ERA5 [24], JRA-55 [25], MERRA-2 [26], CRA-40 [27], etc. Since the successful development of the first-generation reanalysis dataset in the 1990s, a series of reanalysis data has become an important support for research in various disciplines such as atmospheric science and ecological geography, and has been widely used by scientific researchers in various studies such as climate change, disaster monitoring, and regional ecology [28,29,30,31].

Usually, the spatiotemporal resolution of near-surface elements is much higher than the requirements for high-altitude meteorological elements. Therefore, datasets generated by land surface assimilation systems that include near-surface temperature, soil temperature, and other elements are widely used in climate prediction and ecological cycle research [32,33]. The GLDAS V2.1 of the Global Land Surface Data Assimilation System, the ERA5-Land of the European Center for Medium-Range Weather Forecasts (ECMWF), and the HRCLDAS V1.0 dataset of the high-resolution land surface data assimilation system of the China Meteorological Administration have all been developed [34,35,36]. These three datasets differ in model structure, physical parameterization, input processing methods, and spatiotemporal resolution. The performance of ERA5-LAND and GLDAS has been widely evaluated and compared with station observations or other sources. Huang et al. evaluated the performance of ERA5 Land, GLDAS, and CLDAS in the Chinese Mainland by using the observation data of ground stations and found that ERA5-Land is superior to GLDAS, and there are significant differences between different regions [37]. Han, Yang, Zou et al. conducted more comparative studies on the Qinghai Tibet Plateau and the eastern coastal areas of China [38,39,40]. Although the near-surface air temperatures of GLDAS and ERA5-Land have been evaluated in limited areas, a comprehensive and detailed evaluation and comparison of these data in the Inner Mongolia region has not yet been conducted, especially with limited comparative research on HRCLDAS data and other land surface reanalysis data.

In this study, the accuracy and applicability of GLDAS, ERA5-Land, and HRCLDAS near-surface air temperature datasets in the Inner Mongolia Autonomous Region (IMAR) are evaluated from multiple perspectives based on observations collected at automatic weather stations in the IMAR. The conclusions of the evaluation will help researchers choose appropriate temperature datasets for studies on climate change, vegetation evolution, and ecological environment sustainability in the IMAR. Meanwhile, the results of the present study can also provide a reference for the improvement of these reanalysis datasets in areas with sparse observation sites.

2. Data and Methods

2.1. Study Regions

As shown in Figure 1, IMAR is located on the northern border of China, with a straight-line distance of over 2400 km from east to west and 1700 km from north to south, ranking third in terms of regional area in China. The climate is mainly temperate continental monsoon climate, with significant differences in annual temperatures within the region. The average annual temperature was −4.0 °C to 10.0 °C, the highest annual temperature was 29.7 °C to 41.1 °C, and the lowest annual temperature was −44.6 °C to −20.2 °C. The main meteorological disasters include low temperatures, sandstorms, droughts, rainstorms, high temperatures, snowstorms, and cold waves. The terrain of Inner Mongolia is mainly composed of grasslands and Gobi, accounting for about 1/4 of the total grassland area in China. It is an important agricultural and animal husbandry production base in China.

2.2. Data

The reanalysis products used in this study include GLDAS, ERA5-Land, and HRCLDAS for the period 2018–2020 and observations collected at 119 national weather stations in the IMAR. The 2 m air temperature data are extracted from the above datasets.

Table 1. Characteristics of datasets.

Datasets	Spatial Coverage	Spatial Resolution	Temporal Resolution	Data Type	Unit
GLDAS	180° W~180° E; 60° S~90° N	0.25°	3-hourly	Grid	K
ERA5-Land	180° W~180° E; 60° S~90° N	0.1°	Hourly	Grid	K
HRCLDAS	70°~140° E; 0°~60° N	0.01°	Hourly	Grid	K
NAWS	Over major land areas of Inner Mongolia Autonomous Region	119 stations	Hourly	Point	°C

lists the spatial and temporal resolutions of these datasets as well as the area covered by these data.

2.2.1. HRCLDAS Data

The CMA High-Resolution Land Data Assimilation System (HRCLDAS) was developed in the National Meteorological Information Center on the basis of CLDAS-V1.0 and CLDAS-V2.0 [33]. Multi-source data fusion and analysis technologies such as multi-grid variational analysis and model bias correction methods are employed to fuse multi-source data such as observations of automatic weather stations, retrievals of satellite digital elevation model, and numerical model outputs, etc., and produce 1 km resolution real-time fusion dataset in China. This dataset contains hourly ground precipitation, 2 m air temperature and humidity, 10 m winds, etc., at 1 km resolution. The spatial and temporal resolutions of the dataset are 0.01° and 1 h, spatial covering the area over (0–60° N, 70–140° E), time coverage from 2010 to present. At present, HRCLDAS data are publicly available to meteorological departments in various provinces, and other users only provide it upon request. Therefore, observations collected at national weather stations can be used to independently verify HRCLDAS.

2.2.2. ERA5-Land Data

ERA5 (the 5th generation of ECMWF Re-Analysis, ERA5) reanalysis product is the latest generation of reanalysis data produced by the Copernicus Climate Change Service (C3S) under the sponsorship of the European Union and operated by the European Center For Medium-Range Weather Forecasts (ECWMF) [24]. ERA5-Land is obtained from simulations of the revised land surface hydrological model HTESSEL and CY45R1, which are forced by the land surface atmospheric variables of ERA5. Compared with ERA5, ERA5-Land has a higher spatial resolution of up to 0.1° (9 km) and its temporal resolution is 1 h. The dataset covers the period from 1950 to present.

2.2.3. GLDAS Data

Global Land Data Assimilation System (GLDAS) is jointly developed by the National Aeronautics and Space Administration (NASA), the National Centers for Environmental Prediction (NCEP), and the National Oceanic and Atmospheric Administration (NOAA). It employs an advanced data assimilation technology to integrate ground-based observations and satellite remote sensing data and can provide a series of long-term gridded land surface parameters [34]. There are three versions of GLDAS, i.e., V2.0–2.2, among which V2.1 covers the period from 2000 to present and the data quality is relatively good. The present study employs near-surface air temperature data produced by GLDAS-2.1 Noah land module. The horizontal and temporal resolutions of this dataset are 0.25° and 3 h, respectively, and it covers the area of (60° S~90° N, 180° W~180° E).

2.2.4. Surface Meteorological Observation Data

Surface meteorological observations are 2 m air temperature for the period from 1 January 2018 to 31 December 2020 over the IMAR. The data are provided by the Meteorological Information Center of Inner Mongolia. To ensure the data quality and reliability, observations collected at 119 national weather stations that have undergone quality control are selected as the “true” values for evaluation. Figure 1 presents spatial distribution of national meteorological observation stations in Inner Mongolia.

2.3. Data Processing

Based on the latitude and longitude information of the 119 weather stations, the nearest neighbor interpolation method is employed to remap HRCLDAS, ERA5-Land, and GLDAS data to the stations to obtain sequences that can be compared with the observed true values. A total of 19,642,844 comparison samples were obtained. By calculating the related error indexes, comparison of accuracy between and evaluation of different reanalysis near-surface air temperature datasets are realized.

2.4. Evaluation Indexes

Correlation coefficient (COR), bias, root mean square error (RMSE) [37], and standard deviation (SD) [41] are used in this study as the evaluation indexes for overall accuracy as well as spatial and temporal statistical tests. The accuracy (T) [42] and RMSE are selected as the evaluation and statistic test indexes for high-temperature processes. These indexes are calculated as follows:

$$COR=\frac{{\displaystyle \sum _{i=1}^n(G_i-\overline{G_i})(O_i-\overline{O_i})}}{\sqrt{{\displaystyle \sum _{i=1}^n{(G_i-\overline{G_i})}^2}}\sqrt{{\displaystyle \sum _{i=1}^n{(O_i-\overline{O_i})}^2}}}$$

(1)

$$BIAS=\frac{1}{N}{\displaystyle \sum _{i=1}^n{G}_i-O_i}$$

(2)

$$RMSE=\sqrt{\frac{1}{N}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}(x_i}-\overline{x}{)}^2}$$

(3)

$$SD=\sqrt{\frac{1}{N}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}(G_{\mathrm{i}}}-O_{\mathrm{i}}{)}^2}$$

(4)

$$T_k=\frac{N{r}_k}{N{f}_k}\times 100\%$$

(5)

where O_i is station observation, G_i is the data to be evaluated interpolated to the station, N is the is the total number of samples participating in the test (number of stations), Nr_k denotes the number of corrected forecasts, Nf_k is the total number of forecasts, K = 1 indicates that the bias of forecast is equal to or less than 1 °C), k = 2 indicates that the bias of forecast is equal to or less than 2 °C.

3. Analysis of the Results

3.1. Evaluation of the Overall Accuracy

Figure 2 displays scatter plots of 2 m temperature of HRCLDAS, ERA5-Land, GLDAS, and NAWS true values as well as their linear fits. The total number of samples is 1,040,356. All three reanalysis products show a high accuracy with a coefficient of determination above 0.9. During 2018–2020, the mean temperature at the 119 stations was 6.0 ± 14.7 °C, and the mean temperatures of HRCLDAS, ERA5-Land, and GLDAS were 5.9 ± 14.8 °C, 6.2 ± 14.0 °C, and 6.4 ± 14.3 °C, respectively. The mean value of HRCLDAS is closest to the true value of station observations.

The overall accuracy test results are shown in

Table 2. Overall performance of HRCLDAS, ERA5-Land, and GLDAS temperature datasets from 2018 to 2020.

Evaluation Indexes	HRCLDAS	ERA5-Land	GLDAS
COR	0.94	0.87	0.85
BIAS (°C)	−0.19	0.18	0.41
RMSE (°C)	1.37	2.17	2.55

. In terms of correlation, HRCLDAS is better than ERA5-L and ERA5-L is better than GLDAS. The correlation coefficients of HRCLDAS, ERA5-L, and GLDAS with observations are 0.94, 0.87, and 0.85, respectively. In terms of bias, the bias of HRCLDAS is slightly negative, whereas ERA5-Land and GLDAS present positive biases, suggesting that HRCLDAS to a certain degree underestimates surface air temperature while ERA5-Land and GLDAS overestimate surface air temperature in the IMAR, and the overestimation is especially severe for GLDAS. Looking at RMSE, it is found that HRCLDAS has the lowest RMSE of 1.37°C, while the RMSEs of ERA5-Land and GLDAS both are larger than 2 °C. Generally, HRCLDAS performs best during the evaluation period and is significantly better than the other two reanalysis products, ERA5-Land performs the second best, and GLDAS has a relatively low accuracy in the IMAR.

3.2. Temporal Test

3.2.1. Diurnal Changes

Figure 3 displays the characteristics of errors for the three temperature reanalysis datasets at different times of the day. Figure 3a shows the three-year average temperature change curves at different times of the day. Overall, the change trends of near-surface air temperatures over times of the day for the three reanalysis datasets are basically consistent with that of the observations. For HRCLDAS and ERA5_LAND, their time series of air temperature closely follows the observed data, but GLDAS is significantly different from the mean air temperature observations. Figure 3b displays COR at different times of the day, which shows that the variation of COR for HRCLDAS is small with the COR value above 0.991 at all times, the COR for ERA5-Land displays a pattern of increasing–decreasing–increasing, and the COR for GLDAS also increases first and then decreases, reaching the lowest value of 0.978 at 2100 LST. Biases of the three reanalysis products at different times of the day are presented in Figure 3c, which indicates that HRCLDAS has positive biases at 0300 and 0600 LST and negative biases at all other times. The largest positive bias of 0.28 °C appears at 0300 LST, and the largest negative bias of −0.57 °C occurs at 1200 UTC. ERA5-Land only has a slight negative bias of 0.08 °C that occurs at 0900 LST, and the maximum positive bias of 0.32 °C appears at 21:00. GLDAS has positive biases at 0000, 1500, 1800, and 2100 LST and the positive bias is 1.83 °C at 2100 LST. It has negative biases at all other times with the largest negative bias of −1.12°C appearing at 0900 LST. Figure 3d shows RMSEs at different times of the day. The RMSEs are within the ranges of 0.94~1.80 °C, 1.93~2.64 °C, and 1.93~2.64 °C for HRCLDAS, ERA5-Land, and GLDAS, respectively. The largest RMSE occurs at 2100 LST for all three reanalysis products, which may possibly be attributed to the relatively poor simulation in the day-to-night transition period.

3.2.2. Daily Changes

The time series of error characteristics of daily average temperature in the IMAR from 2018 to 2020 are displayed in Figure 4. Figure 4a shows daily mean air temperature changes with time for the observations and the three reanalysis products. It is found that the temperature change trends of the three reanalysis products are basically consistent with that of observations. Daily CORs during the study period are displayed in Figure 4b, which shows that daily changes in the COR are small for HRCLDAS but large for ERA5-Land and GLDAS, especially for GLDAS, and the COR is within the range of 0.67~0.99. Figure 4c shows a time series of daily mean BIAS. It can be seen that the bias of HRCLDAS is basically negative and its range is within −0.5~0.1 °C. The bias of ERA5-Land is positive in winter, especially from December 2019 to February 2020, and surface air temperature is significantly overestimated. The bias of ERA5-Land is basically positive in spring, summer, and autumn. The negative bias of GLDAS prevails during February–March 2020 and December 2020, and the bias is positive at other times. The time series of daily mean RMSE is displayed in Figure 4d. Note that the RMSE of HRCLDAS is obviously lower than that of the other two reanalysis products. The daily mean RMSE of HRCLDAS is within the range of 0.44~2.68 °C. Compared to that of GLDAS, the RMSE of ERA5-Land is lower most of the time. In conclusion, for the daily mean air temperature, HRCLDAS is closer to the observations than ERA5-Land and GLDAS with lower RMSE, smaller bias, and higher COR.

3.2.3. Monthly Changes

Figure 5 presents monthly mean error characteristics over the period 2018–2020. The left panel of Figure 5 shows monthly COR changes for the three reanalysis products. The monthly COR for HRCLDAS is small with the value always above 0.996. ERA5-Land and GLDAS have a similar change pattern, i.e., the COR value gradually decreases with the increase in the month from April to July, and gradually increases with the increase in the month from January to April and from August to November. The middle panel of Figure 5 presents the monthly biases of the three reanalysis products compared to station observations. The bias of HRCLDAS is negative in all the months, and the bias of GLDAS is always positive in all the months. The bias of ERA5-Land is negative in April, May, June, August, September, and October and positive in other months. The monthly variation of the bias is the largest for ERA5-Land and the smallest for HRCLDAS. The largest bias occurred in December for all three reanalysis products. The right panel of Figure 5 displays monthly changes in RMSE. The RMSE of HRCLDAS is smaller than that of the other two products in all the months with a smaller monthly variation. The RMSE of ERA5-Land overall is smaller than that of GLDAS, while the RMSEs of the two products both exhibit a decreasing trend with month before August and an increasing trend with month after August.

3.2.4. Seasonal Changes

To further evaluate the accuracy of the three temperature reanalysis products in different seasons, Taylor diagrams (Figure 6) are produced to illustrate errors of the three products compared to station observations in spring (Figure 6a), summer (Figure 6b), autumn (Figure 6c), and winter (Figure 6d), respectively. The comparison indicates that the SDs of the three products are obviously smaller in summer than in the other three seasons. The SD values range between 5~6 °C in summer, 9.3~10.3 °C in spring and autumn, and 7~8.2 °C in winter. The correlations of HRCLDAS, ERA5-Land, and GLDAS with station observations are better in spring and autumn. The correlation of HRCLDAS with observations is the second best in summer and the worst in winter, while the opposite is true for ERA5-Land and GLDAS. The distances between the rhombus points and the X-axis in the coordinate show that the RMSE between HRCLDAS and station observations is less than 2 °C in all four seasons, while the RMSE of ERA5-Land is smaller than 2 °C in summer but larger than 2 °C in the other three seasons. The RMSE of GLDAS is larger than 2 °C in all four seasons. The RMSE is the largest in winter for all three products with the values of 1.73, 2.59, and 3.13, respectively. The colors of the rhombus points and the values corresponding to these different colors indicate that HRCLDAS underestimates temperature compared to station observations in all four seasons, whereas GLDAS overestimates temperature. ERA5-Land shows slight negative and positive biases in summer and autumn, while the negative and positive biases in spring and winter are obvious. HRCLDAS is closest to station observations in all four seasons, followed by ERA5-Land and the performance of GLDAS is the worst.

3.3. Spatial Test

3.3.1. Station Test

Analysis of spatial variations of correlations between individual station observations in IMAR and the three reanalysis products (Figure 7) reveals that the correlation between HRCLDAS and station observations is significantly better than the correlations of ERA5-Land and GLDAS with observations. Combining Figure 7a,d, it is found that the stations with a correlation coefficient between HRCLDAS and observations greater than 0.99 accounts for 99% of the total number of stations, while the stations with a correlation coefficient between ERA5-LAND and observations greater than 0.99 accounts for only 46.2% of the total stations (Figure 7b). Looking at the topography in the IMAR (Figure 1), stations with a high correlation coefficient are largely distributed in low-elevation and relatively flat terrain areas. As shown in Figure 7c, large differences exist in COR between GLDAS and observations at different stations, and the correlation coefficient is greatly affected by topography and underlying surface. The lowest value of COR is 0.95, which is located at the Southern Ordos Plateau. The highest value of COR is 0.99, which is concentrated in the Hulunbuir Grassland and the Xiliao River Alluvial Plain in Eastern Inner Mongolia. Overall, HRCLDAS is the best, and ERA5-Land is better than GLDAS.

Figure 8 presents spatial distributions of biases between the three reanalysis products and national weather station observations. As shown in Figure 8a,d, for HRCLDAS, the biases are within −1.0~1.0 °C at 94.9% of the stations and are mostly negative. For ERA5-Land (Figure 8b,d), the biases at the stations are within −1.7~1.83 °C, and large positive biases are concentrated in Northern Hulunbuir Grassland while negative biases are mainly distributed in the south of Daqingshan Mountains and the east side of the southern section of the Greater Khingan Mountains. Figure 8c,d indicate that most of the biases of GLDAS are positive with large positive values distributed in the desert and Gobi regions in Western Inner Mongolia, and the maximum value is significantly higher than that of the other two reanalysis datasets.

RMSE can further reflect the performance of the reanalysis products. As shown in Figure 9, the RMSE of HRCLDAS at most stations is lower than that of ERA5-Land and GLDAS. Figure 9a indicates that the RMSE of HRCLDAS is smaller than 2.0 °C at all the stations except the seven stations in the Greater Khingan Mountains and Yinshan Mountains. The RMSEs of ERA5-Land are concentrated over 1.5~3.0 °C (Figure 9b), while the spatial distribution of large RMSEs is consistent with that of large biases. The maximum RMSE of GLDAS can be up to 4.36 °C, and large RMSE values are mainly distributed in the northern part of the Hulunbuir Grassland, the eastern side of the southern section of the Greater Khingan Mountains, and deserts and Gobi areas in Western Inner Mongolia (Figure 9c). Figure 9d displays the histogram distributions of the number of stations with RMSE, which shows in a straightforward way that HRCLDAS performs better than the other two products.

3.3.2. Sub-Region Test

Inner Mongolia has a vast territory, and the topography and underlying surface vary greatly from west to east. Different cities have different geographical characteristics. According to the existing administrative divisions, Inner Mongolia has a total of 12 league cities, whose geographical positions and sizes are represented in different colors, as shown in Figure 10. Grouping statistics of the mean and error of each city can further demonstrate the spatial situation of the three reanalysis products (Figure 11). The table below Figure 11a lists average temperatures from observations and HRCLDAS, ERA5-Land, and GLDAS at each city. The histogram shows BIAS distributions for the three reanalysis datasets in different cities. The figure shows that the bias of HRCLDAS is positive in Wuhai and negative in all other cities, the bias of ERA5-Land is positive in all cities except Xingan, Huhehaote, Chifeng, and the bias of GLDAS is always positive except in Chifeng and Xingan. In Wuhai, all three reanalysis products overestimate temperature, and the overestimation is especially significant in HRCLDAS. The opposite is true in Xingan, where the temperature is underestimated in all three products, and the underestimation is most severe in GLDAS. The table below Figure 11b lists the correlation coefficients of HRCLDAS, ERA5-Land, and GLDAS with station observations in different cities. The correlation coefficient of HRCLDAS with observations is above 0.99 at all the 12 cities, the correlation coefficient of ERA5-Land with observations is above 0.99 in Xingan, Hulunbuir, Tongliao, and Xilingele and above 0.98 in the other 8 cities. The correlation coefficient of GLDAS with observations is above 0.98 in all the cities except Ordos City and Wuhai. The histograms of RMSEs of the three reanalysis datasets in different cities are displayed in Figure 11b, which shows that the RMSE of HRCLDAS is lower than that of the other two reanalysis products in all the cities. The RMSE of ERA5-Land is lower than that of GLDAS in all the cities except Huhehaote. The largest RMSEs of HRCLDAS and GLDAS both occur in Wuhai, while the largest RMSE of ERA5-Land is found in Hulunbuir. In summary, the three reanalysis datasets perform the best in surface air temperature simulation in Tongliao.

3.4. Test of High-Temperature Process

In meteorology, high-temperature weather is defined as the daily maximum temperature at a single station ≥ 35.0 °C. From 2018 to 2020 in Inner Mongolia, a total of 1191 stations experienced high-temperature weather. These high-temperature weathers are listed in

Table 3. Number of stations experiencing high temperatures from 2018 to 2020.

	≥40 °C	<40.0 °C and ≥37.0 °C	<40.0 °C and ≥37.0 °C	Total
2018	9	134	352	495
2019	6	59	241	306
2020	0	78	312	390

. Most of them are single-day or persistent regional high temperatures, and persistent whole-region high-temperature weather is relatively rare.

The high-temperature process that occurred in August 2018 affected a large area and many stations. Seven national weather stations experienced extreme high temperatures. Among them, the highest temperature in Balinzuoqi of Chifeng reached the historical record on August 3. Therefore, the high-temperature process on 2–3 August 2018 is selected as the case for further study. Since the temporal resolution of GLDAS is 3 h, which makes it hard to obtain daily maximum temperature, only HRCLDAS and ERA5-Land temperature datasets are selected for the applicability test of the high-temperature process in Inner Mongolia.

The actual maximum temperature from NAWS and that from the two reanalysis products during this high-temperature process is displayed in Figure 12. The red circle denotes the area where the maximum temperature is higher than ≥40.0 °C. This high-temperature process affected Xilingele, Chifeng, Tongliao, Alashan, and Southern Xingan. High-temperature weather with a surface air temperature above 40.0 °C occurred at Balinzuoqi and Balinyouqi of Chifeng. Figure 12b presents the spatial pattern of temperature from HRCLDAS, which shows that the area with a temperature ≥ 35.0 °C is consistent with observations, while the areas with a temperature ≥ 40.0 °C are mainly located in Chifeng and Tongliao. Inside the area denoted by the red circle, the temperature from HRCLDAS agrees well with observation at Balinzuoqi, but HRCLDAS underestimates the temperature at Balinyouqi. The overall performance of HRCLDAS is good. Figure 12c presents the spatial pattern of temperature from ERA5-Land during the high-temperature process. Compared to station observations and HRCLDAS, ERA5-Land severely underestimates temperature during this high-temperature process. The areas of high temperature in Alashan and Xilingele are smaller than the observations, the temperature within the area of the red circle is lower than 40.0 °C, and the area of temperature ≥ 40.0 °C is displaced from the observed high-temperature area.

Table 4. Two process-based reassessments of temperature data process maximum temperature quality assessment.

HRCLDAS			ERA5-LAND
Accuracy (≤1 °C)	Accuracy (≤2 °C)	RMSE	Accuracy (≤1 °C)	Accuracy (≤2 °C)	RMSE
83.19%	94.96%	0.86	35.29%	73.95%	1.79

lists the test result of the highest temperature accuracy and root mean square error based on the two reanalysis datasets. In terms of accuracy, the accuracy rates of HRCLDAS reach up to 94.96% and 83.19% with BIAS ≤ 2 °C and BIAS ≤ 1 °C, respectively. The accuracy rates of ERA5-Land are only 73.95% and 35.29% with BIAS ≤ 2 °C and BIAS ≤ 1 °C, respectively. In terms of RMSE, the RMSE of HRCLDAS is 0.86°C and the RMSE of ERA5-Land is 1.79 °C, indicating that the average error magnitude of HRCLDAS is smaller than that of ERA5-Land.

4. Discussion

Climate warming is rapidly changing the hydrological and ecological systems of land, and obtaining grid temperature data with high spatiotemporal resolution is essential for studying agriculture, animal husbandry, and the ecological environment in Inner Mongolia. The results of this study indicate that the HRCLDAS temperature product has the best performance compared to the observed data in Inner Mongolia, and ERA-LAND is superior to GLDAS, which is consistent with the results described by Huang et al. [37]. Although the hourly temperature product changes in HRCLDAS are most closely related to observation data from different locations in Inner Mongolia, we found that HRCLDAS underestimates the temperature of most locations, while ERA5-LAND and GLDAS products overestimate near-surface temperature, which is inconsistent with the results of ERA5-LAND’s temperature simulation ability in the Qinghai Tibet Plateau and southeastern coastal areas of China described by Huang [43] and Zou et al. [40]. This may be due to significant differences in terrain between different regions. Compared to the Qinghai Tibet Plateau and the southeastern coastal areas, the terrain in Inner Mongolia is relatively flat and the elevation fluctuations within the region are relatively low.

More research has shown that elevation and the terrain parameters affected by elevation are important reasons for the differences between the site and the reanalysis dataset [44,45]. While the reanalysis model topography could function to better represent the average elevation of grid cells, it inevitably smooths the observed elevation by interpolating DEM elevation to the reanalysis model grid scale [46]. In order to explore whether the elevation difference may affect the accuracy of the three types of data in Inner Mongolia, we extracted elevation values of different resolutions using observation stations and compared them with station observations, then observed and reanalyzed the elevation differences between the stations and the terrain. From the average DEM values of different grid resolutions extracted from 119 observation stations and the observed elevation of the stations (Figure 13), the difference between the grid terrain height values of HRCLDAS and the station elevation is small, with an elevation difference range of −52 m to 31.2 m, and the elevation difference of ERA5-Land is between −66.7 m to 310 m, while the highest value of GLDAS is 552.5 m. From the evaluation results in Inner Mongolia, it can be seen that the reanalysis data of ERA5 Land and GLDAS have varying degrees of overestimation relative to the observed temperature. This may be because the elevation background field values of ERA5 Land and GLDAS in Inner Mongolia are mostly greater than the station elevation values. Therefore, when using reanalysis temperature data at a smaller regional scale, height difference correction should be carried out to effectively reduce errors and improve the accuracy and applicability of reanalysis data, which is consistent with the research results of Zhang et al. [47,48,49].

Before the elevation difference, we also found that reanalysis data with different spatial resolutions have significant differences in temperature simulations in small-scale areas. For example, in the Chinese Mainland, ERA5 near-surface temperature is in good agreement with ground observation [50,51,52], while ERA-1nterim with coarse resolution significantly underestimated the observed temperature in the Eastern Tibet Plateau and Northern Asian alpine region [53]. Hu et al. studied the performance of five different resolution reanalysis temperature products in the permafrost region of the Qinghai Tibet Plateau, including CFSR, ERA-Interim, GLDASNOAH, MERRA, JRA-55, and CMFD, of which ERA-Interim has a better overall performance [54]. Sun et al. compared the daily maximum temperature performance of CLDAS, ERA5, and GLDAS in the Chinese Mainland, and found that the high-resolution CLDAS is significantly better than the other two products [55]. From the above research, it can be seen that the performance of high-resolution temperature data in intergenerational reanalysis data is usually better than that of low-resolution data in the same year, and high-resolution reanalysis temperature data can better depict small and medium-sized fine features. Unlike previous studies, this study explores the applicability of the latest 1 km spatial resolution HRCLDAS data released by the China Meteorological Information Center at the regional scale. HRCLDAS is currently the highest spatial resolution reanalysis temperature dataset in China. Through multiple comparative evaluations, it was found that HRCLDAS has better applicability in Inner Mongolia than ERA5 Land and GLDAS. The significant difference may be due to the use of multi-grid variational analysis in the generation of the HRCLDAS dataset, where the background and observation fields are gradually merged and assimilated from large to small scales, and the temperature data collected by automatic observation stations are integrated to ensure the quality of HRCLDAS temperature products in the Chinese region.

Due to the high spatiotemporal resolution, wide spatial coverage, and long time series of reanalysis temperature data, it has been widely used in standardized temperature index and agricultural drought index calculations, as well as forced data in crop models [56]. Xue et al. use the total precipitation and near-surface temperature data of ERA5 to calculate the self-calibrating Palmer Drought Severity Index (scPDSI) to assess the drought changes in the Belt and Road Area [57]. Based on ERA’s temperature and precipitation data, Zhu et al. calculated three indicators: Standard Precipitation Index (SPI), Standard Precipitation Evapotranspiration Index (SPEI), and Extreme Degrees Day (EDD) from 1979 to 2017 and found that high temperature and drought have a significant impact on the yield of wheat and corn [58]. The underestimation of temperature by HRCLDAS and the overestimation of temperature by ERA5 Land and GLDAS may bring greater uncertainty to the prediction or simulation of high temperature and drought events. If ERA5-Land is used to analyze the overall characteristics of high-temperature and drought composite events and diagnose the spatiotemporal evolution characteristics of high-temperature and drought composite events over a certain period of time, it may lead to higher frequency or intensity of high-temperature and drought events than actual results without correction or evaluation. That is why it is important to evaluate the errors between reanalysis and on-site observations.

This study reveals some important issues that differ from previous studies [24,27,35]. For example, GLDAS has a greater deviation at night and in the morning than at other times and exhibits a higher positive deviation, while ERA5-LAND has a positive deviation almost all day. In terms of season, the seasonal correlation between ERA5-LAND and GLDAS with observed values is better in spring and autumn, worst in winter, and second in summer. This may be because Inner Mongolia has a long winter and a high number of extreme low temperatures, and the ability of the two reanalysis temperature data to depict extreme temperatures is insufficient [35,36]. Based on the above discussion, we believe that although there are certain errors in the temperature reanalysis dataset of Inner Mongolia, they still have a certain degree of applicability and credibility in Inner Mongolia, because the distribution of observation points there is uneven, and the observation density is low. Therefore, these reanalysis datasets have certain reference values. It should be noted that although the quality of HRCLDAS is superior to the other two datasets based on careful evaluation of the three reanalysis datasets in this study, HRCLDAS also has certain limitations, such as only covering China and surrounding areas (0–60° N, 70–140° E), and currently only data from 2010 to present. These limitations result in HRCLDAS being only applicable to research within the Chinese region and being a relatively short time series. Compared with HRCLDAD, ERA5-LAND and GLDAS have longer time series, larger spatial ranges, and more elements. ERA5-LAND and GLDAS are more suitable for studying large-scale regions or conducting long-term climate change studies. Therefore, suitable reanalysis datasets should be selected based on different application scenarios and requirements.

5. Conclusions

Observation of the site, remote sensing products, and reanalysis datasets are the main datasets used by the scientific community to monitor temperature changes. The in situ measurements offer the most accurate atmospheric data but are limited by their sparse and uneven distributions in the Inner Mongolia region. Reanalysis of datasets is of great significance in filling unobserved data gaps at both temporal and spatial scales, but validation is needed before further research can be conducted. In this study, we evaluated the performance of the near-surface air temperature in Inner Mongolia produced from the ERA-LAND, GLDAS, and HRCLDAS reanalysis datasets based on in situ observations from 119 monitored during 2018—2020. The main conclusions are as follows:

(1)

Compared with the observation results, the three sets of reanalysis temperature products all reasonably reflect the characteristics of surface temperature changes in Inner Mongolia over time. The correlation coefficient between the three reanalysis temperature products and observations exceeds 0.85. GLDAS shows a warm bias most of the time, HRCLDAS shows a slight negative bias in time, ERA5-Land shows a warm bias in autumn and winter, and spring and summer show a cold bias.

(2)

The evaluation results at multiple spatial scales indicate that the accuracy and applicability of HRCLDAS are significantly superior to the other two datasets. The quality of reanalysis datasets varies at different observation stations and cities, and the quality of the three reanalysis datasets is significantly affected by terrain.

(3)

In terms of maximum temperature, HRCLDAS is closer to the observation results than ERA5 Land, and HRCLDAS exhibits more locally detailed features, but there are a few false reports. ERA5-Land exhibits smooth characteristics towards high temperatures, and its accuracy in reproducing high temperatures is lower than that of HRCLDAS.

In summary, there are significant differences in the evaluation results of the three reanalysis data temperature products in Inner Mongolia from 2018 to 2020. HRCLDAS has the highest accuracy in the Inner Mongolia region, followed by ERA5 Land, and GLDAS is relatively the worst. This study’s research outcomes furnish a crucial point of reference for investigating the correlation between agricultural drought, vegetation biomass, and climate in Inner Mongolia. Particularly notable are the implications arising from the impacts of elevated temperatures and cold spells on agricultural processes. Expanding on the study’s importance in agriculture management, it highlights the potential for refining crops’ phenology modeling through the use of observed hourly temperature data, ultimately leading to enhanced crop yields and quality. Moreover, these implications extend to the monitoring of frost events. The inadequacy of this study is that it only evaluates grid temperature data within a limited time frame, which may not accurately demonstrate the evaluation characteristics of certain climate states. In addition, the different spatiotemporal resolutions of the three grid temperature data can lead to certain limitations in the evaluation conclusions. In the future, the combination of improved assimilation methods and better quality observation data will improve the quality of the reanalysis dataset, and more observations and longer time series reanalysis data will be used to further prove the reliability of the results.