Evaluation of ERA-Interim Air Temperature Data over the Qilian Mountains of China

College of Geographical Science, Fujian Normal University, Fuzhou 350007, China Institute of Geography, Fujian Normal University, Fuzhou 350007, China Fujian Provincial Engineering Research Center for Monitoring and Accessing Terrestrial Disasters, Fujian Normal University, Fuzhou 350007, China State Key Laboratory of Subtropical Mountain Ecology, Fujian Normal University, Fuzhou 350007, China Institute of Meteorological and Climate Research, Karlsruhe Institute of Technology, Campus Alpine, Garmisch-Partenkirchen 82467, Germany China Institute of Water Resources and Hydropower Research, Beijing 100038, China Fujian Key Laboratory of Severe Weather, Fuzhou 350001, China


Introduction
e Qilian Mountains (QLM), located in the northwest of China and sited at the intersection between Gansu Province and Qinghai Province, are mountainous regions with complex terrain and climate conditions. e altitude is more than 4000 m in most of QLM, and the total area is over 19.5 × 10 4 km 2 [1][2][3]. e QLM is an important ecological security barrier to northwestern China and also an important water source for the Yellow River basin [4]. e the climate and glacier changes in QLM [6][7][8][9]. Jiang et al. [10] found that the annual variations of ice and snow are significant along elevations based on MODIS data. Tian et al. [9] also found that the glaciers in QLM have melted significantly in the past 20 years, which may be caused by rising temperatures. Chen et al. [11] found that the yearly air temperature lapse rate (TLR) was weaker at >3000 m elevations and the seasonal TLRs were more divergent, which further bolsters the evidence of elevation-dependent climate change. Wu et al. [12] found that there is a significant time lag between vegetation and temperature at different altitudes in Qilian Mountains. Niu et al. [13] found that air temperature increases with altitude from the north edges of Hosi Corridor to the north slope of Qilian Mountains. Qing et al. [14] found that the lapse rate is around 0.48°C/100 m in 2012 based on the four weather stations in the Hulugou watershed in QLM, which is smaller than the lapse rates in other regions of QLM. Cao et al. [15] found that the annual and seasonal temperatures exhibited unanimously fluctuating warming trends, especially in winter during the 1960-2014 period based on 19 meteorological stations on the south slope of QLM. Jin et al. [16] found the temperature gradient increases with elevation in the nonglacierized area of the Laohugou basin of QLM.
Compared with reanalysis data, the observed records have some shortcomings, such as short time series and uneven temporal/spatial distributions. Spatial interpolation methods such as Kriging interpolation and inverse distance weights interpolation are usually used to estimate those regions without station observation data [17]. However, different interpolation methods often lead to large errors due to the limitations of the spatial interpolation itself, such as the density and uneven distribution of sites. Reanalysis data have been widely used in the past two decades because of their high resolution and long-time series [18,19]. However, reanalysis data are sometimes unreliable due to their biases. (e.g., [17,19]) evaluated the CFSR, ERA-Interim, MERRA, and MERRA-2 reanalysis data sets and found that the reanalysis data are inappropriate for use as supplementary data for offshore wind resources assessment. Auger et al. [20] found that reanalysis data exhibit significant differences over particular regions based on a global ensemble mean of the four third-generation climate reanalysis models in 1981-2010. erefore, it is still a necessity to evaluate the quality and bias of reanalysis data.
ere are many studies regarding the evaluation of ERA-Interim in different regions such as Canadian Prairies [21], the Tibetan Plateau [22], and Portugal [23]. For example, Gao et al. [22] and Wang et al. [24] have verified that the reanalysis data could be an alternative to observations in areas lacking data because of their resolution and long-time series. Liu et al. [25] concluded that the researchers should be careful when using ERA-Interim precipitation and temperature data in complex topography areas. Gao et al. [22] concluded that ERA-Interim is generally reliable for climate change studies over the Tibetan Plateau. Gao et al. [26] concluded that evaluation of ERA-20CM data is helpful for potential users of reanalysis data to determine local climate change impact assessments in China. Gao et al. [19] also found that the monthly bias is around −3.5°C for ERA-Interim in the Tibetan Plateau. However, knowledge of the reliability of reanalysis data in QLM is quite limited. erefore, this analysis could provide a reference for reanalysis data application at the QLM.
is study uses temperature data obtained from 24 meteorological stations for the 1979-2017 time period to test the daily 2 m air temperature from ERA-Interim in QLM.
is important evaluation would answer the question that whether ERA-Interim temperature is suitable for local climate studies, especially in complex terrains. e paper is organized as follows. A short overview of ERA-Interim data, meteorological observations, and evaluation methods are given in Section 2.
e results and discussions are presented in Section 3, and finally the conclusions are given in Section 4.

ERA-Interim Data (T e ).
e ERA-Interim 2 m air temperature (00UTC, 06UTC, 12UTC, and 18UTC) data (T e ) were supplied by the European Centre for Medium Range Weather Forecasts (ECMWF). e spatial resolution is 0.25°× 0.25°, and the time step is 6 hours. e spatial range is 35.5°N-40.75°N, 93°E-104.5°E, and the period is from January 1979 to December 2017. ERA-Interim temperature data are calculated as daily mean temperature (T_mean), daily maximum temperature (T_max), and daily minimum temperature (T_min). Since the time of reanalysis data is UTC time, it matches the ground observations by using time difference conversion. e ERA-Interim modelled (point) height is calculated by dividing geopotential by gravity at each grid point (Table 1).

Observations (T o ).
Daily air temperature (T o ) and elevation information for 24 meteorological stations from daily data set V3.0 were provided by the China Meteorological Data Sharing Service System of National Meteorological Information Center (http://data.cma.cn/) ( Figure 1 and Table 1). e provider strictly controlled the quality of the observed data including 825 National Reference and Basic Stations (NRBS) over China from five aspects: climate limit value and allowable value checking, extreme value checking, internal consistency checking between timing value, daily average, and daily extreme value, time consistency checking, and space consistency checking. All the data are checked and corrected manually (http://data.cma.cn/data/cdcdetail/ dataCode/SURF_CLI_CHN_MUL_DAY_V3.0.html, last access: 06 February 2020). Based on this data set, the Chinese homogenized historical temperature data set (CHHT, version 1.0) was released in 2009 [27,28]. However, the CHHT was updated only to 2012 (http://data.cma.cn/data/cdcdetail/ dataCode/SURF_CLI_CHN_TEM_MUT_HOMO.html, last access: 06 February 2020). At present, the daily data set V3.0 is the most authoritative and reliable observation data in China, which were widely applied for climate studies [19,29,30]. e selection criteria for stations were long-term consecutive measurement with no gaps exceeding two consecutive weeks.   Table 1.
Daily air temperature includes daily mean temperature, daily maximum temperature, and daily minimum temperature.
We have to mention that ERA-Interim latitude/longitude grids are interpolated from a reduced Gaussian grid N128 (∼79 km). e grid point closest to the station is selected because the ERA-Interim grid point covers an area of 0.25°× 0.25°in which the station is located. In other words, the grid points represent the value for the whole grid. However, no matter what methods we use to interpolate ERA-Interim to individual site, it will definitely bring new errors.
e traditional interpolation methods such as Kriging, Inverse Distance Weighted, and Spline have certain shortcomings in complex mountain areas [17]. erefore, we used the variables from ERA-Interim gridded data without interpolation for a more objectively comparison. e expert (Dr. Florian Pappenberger) from the ECMWF also recommended the nearest grid point to the meteorological station in the private communication. We selected the grid point closest to the station because this ERA-Interim grid point covers the area of 0.25°× 0.25°w here the station is located. erefore, according to the longitude and latitude coordinates of 24 meteorological stations, the corresponding grid points are selected for comparison, which can avoid the error caused by multigrid spatial interpolation [17,31]. e 24 stations are located at different altitudes, ranging from 1000 m to 4000 m. e ERA-Interim grid point nearest to each station was extracted according to the coordinates of stations. Detailed station information is given in Figure 1 and Table 1. Elevation difference (ERA-Interim grid point height minus meteorological station height) is also listed in Table 1. In general, 23 stations have positive elevation difference, with the exception of station No. 11 having a negative elevation difference. e four seasons are defined as spring (March-May), summer (June-August), autumn (September-November), and winter (December-February).

Evaluation Methods.
In order to evaluate the quality the ERA-Interim data set, three criteria were computed based on comparison of the ERA-Interim and observed temperatures at the 24 meteorological stations: correlation coefficient (r), root-mean-square-error (RMSE), and bias.

Daily Temperature Comparisons.
e overall performance of ERA-Interim daily air temperature in 1979-2017, with respect to daily mean temperature (T_mean), daily maximum temperature (T_max), and daily minimum temperature (T_min), is listed in Table 2. e correlation coefficient (r) ranges from 0.956 to 0.996 for the three temperatures. e high r indicates that T e captures the daily observations very well. e biases of daily mean temperature range from −3.9°C to +2.8°C with an average value of −1.8°C.
e positive values indicate that T e is warmer than T o and the negative values reveal that T e is cooler than T o . e largest positive bias (+2.8°C) is found at station No. 11 (Wu Shaoling in Gansu Province), which is located in the east of QLM.
e grid height (2604 m) is much lower than the station Wu Shaoling (3045 m). e largest negative bias (−3.9°C) is found at station No. 15 (Tuo Le in Qinghai Province) situated at an elevation of 3460 meters in the northern QLM. However T e was modelled at the grid point height of 3936m. e second highest negative bias (−3.5°C) appears for station No. 6 (Gao Tai) which is at an altitude only of 1357 m, while the ERA-Interim grid height is much higher (2225 m). We selected another three typical stations: station No. 14, station No. 19, and station No. 20 as representative sites, which are distributed in the western and southern boundary of QLM. e daily biases are −1.0°C, −3.4°C, and −2.5°C for these three stations, respectively. Figure 2 shows the comparison of ERA-Interim daily reanalysis with observations of station No. 6, No. 11, No. 14, No. 15, No. 19, andNo. 20 in 1979-2017

Monthly and Seasonal Mean Temperature Comparisons.
Table S1shows the monthly bias between ERA-Interim reanalysis and observations. In general, the largest positive bias is found at station No. 11 while the largest negative biases are  e high values of r mean a good agreement between T e and T o at the seasonal scale. However, the seasonal variability that reflects the year-to-year fluctuations in the seasonal means is not accordant.

Warming Trends of ERA-Interim Temperature and
Observations. Figure 3 shows the aggregated annual temperatures from T e and T o as well as the temperature trends over the whole QLM. e linear increasing rate of T o is about +0.457°C/decade during the 1979-2017 period. e T e has an increasing trend of +0.384°C/decade, which means that ERA-Interim generally captures the warming trend well (Table 4). e trend difference between T e and T o is 0.076°C/ decade for annual mean temperature. e largest trend difference is found in winter (0.093°C/decade). ERA-Interim captures the trends in spring, summer, and autumn very well, with values of 0.077°C/decade, 0.071°C/decade, and 0.051°C/decade, respectively. e sparse observations in the western QLM are possibly responsible for the difference. Generally, T e is reliable for the warming trend detection over QLM because T e only has an averaged 0.073°C/decade trend difference against T o . However, the averaged RMSE (±2.7°C) should be given enough attention (i.e., bias correction) before T e is applied at local scale. Figure 4 shows the correlation of daily biases and elevation differences between T e and T o . Please note that bias and elevation difference were calculated by subtracting T o from T e . e correlation of determination (R 2 ) measuring the fit is 0.762, which reveals that the altitude differences between ERA-Interim grid points and meteorological stations cause the bias. erefore, it is possible to reduce the bias between ERA-Interim reanalysis and observations by using an elevation correction method, in order to improve the applicability of ERA-Interim [17,26].

Bias Analysis.
Spring, summer, autumn, and winter R 2 values are 0.837, 0.713, 0.725, and 0.381, respectively, and the annual correlation is 0.762 ( Figure 5), which suggests again that altitude differences are responsible for the biases. Particularly with respect to spring temperature, the elevation difference is the main error source, indicating that elevation correction could reduce the error and further improve the applicability of ERA-Interim reanalysis. Assimilation data error, model system error, and interpolation error are also possible. Table 5, Table 6, and Table S3 represent the averaged r, bias, and RMSE for different elevation groups. We divide elevation into five classifications: 1000-1500 m (5 stations),       T_mean is found for the 3000-3500 m group. e largest negative bias (−5.0°C) of T_max is found for the 2500-3000 m group, and the smallest negative bias (−3.4°C) of T_max is found for the 3000-3500 m group. e largest negative bias (−0.5°C) of T_min is found for the 2000-2500 m group, and the largest positive bias (+0.4°C) of T_min is found for the 3000-3500 m group. In general, RMSE increases with the rise of elevation group. e largest RMSEs of T_mean and T_min are found for the 3000-3500 m group, and largest RMSE of T_max is found for the 2500-3000 m group.
Seasonal averaged r, bias, and RMSE are distinguished among different elevation groups. Averaged r ranges from 0.695 to 0.980 for four seasons for different elevation groups. e largest negative bias (−2.9°C) in spring is found in the 2000-2500 m elevation group, and the smallest negative bias (−1.1°C) is found in the 3000-3500 m group. e largest negative bias (−2.5°C) in summer is found in the 2500-3000 m group, and the smallest negative bias (−1.8°C) is found in the 3000-3500 m group. e largest negative bias (−2.3°C) in autumn is found in the 2500-3000 m group, and the smallest negative bias (−1.2°C) is found in the 3000-3500 m group. e largest negative bias (−1.9°C) in winter is found in the 2500-3000 m elevation group, and the smallest negative bias (−0.3°C) is found in the 1000-1500 m group. In general, seasonal averaged RMSE shows an increasing trend with rising altitude. Annual averaged r ranges from 0.672 to 0.959. e largest negative bias (−2.3°C) is found in the 2500-3000 m elevation group, and the smallest negative bias (−1.3°C) is found in the 3000-3500 m group.

Conclusions
In this study, ERA-Interim temperatures (T e ) are compared with observations (T o ) from 24 meteorological stations over the Qilian Mountains of China (QLM) at multiple temporal scales. High daily correlations from 0.956 to 0.996 indicate that ERA-Interim could capture the cycle very well. e biases of daily mean temperature ranging from −3.9°C to +2.8°C (averaged −1.8°C) are mainly from altitude differences between ERA-Interim grid height and the elevation of observing stations (R 2 � 0.762). e biases of daily maximum temperature range from −7.0°C to +2.9°C with an average value of −4.0°C. e biases of daily minimum temperature range from −2.6°C to +2.8°C with an average value of −0.1°C. e average RMSEs of daily mean temperature, daily maximum temperature, and daily minimum temperature are ±2.7°C, ±4.7°C, and ±2.7°C, respectively. e results of this comparison indicate that ERA-Interim reanalysis should not be applied directly at the local scale for climatological and hydrological model purposes due to large biases.
Monthly biases are similar to daily biases. For monthly bias variability, the largest negative bias is in April (−2.5°C), and the smallest negative bias is in December (−0.8°C). e average correlations for spring, summer, autumn, and winter are 0.939, 0.910, 0.873, and 0.906, respectively, which suggest that that ERA-Interim can reproduce the interannual variability over the QLM. Spring, summer, autumn, and winter RMSE values' biases are ±2.6°C, ±2.4°C, ±2.1°C, and ±1.5°C, respectively, which also indicate that T e could not be used directly in scientific studies. e biases in temperatures are mainly due to altitude differences, especially during the spring months (R 2 � 0.837).
is suggests that elevation correction could reduce the error and further improve the applicability of ERA-Interim reanalysis. e relationship (R 2 ) of bias and elevation differences between seasonal observations and ERA-Interim reanalysis are 0.713, 0.725, and 0.381, for summer, autumn, and winter, respectively. e winter climate in mountain areas is more complex and changeable due to impact factors such as "cold lake." us, the correlation between temperature and elevation is relatively weak. For annual variability, an average correlation is 0.906 for all stations. e average annual RMSE is ±2.1°C, which also indicates that T e could not be used directly in scientific studies.
A significant warming trend (+0.457°C/decade) is detected over QLM based on observations over the 1979-2017 period. ERA-Interim can generally reproduce the warming trend at the rate of +0.384°C/decade. e largest warming trends are both detected in summer for the observations and ERA-Interim, +0.552°C/decade and +0.481°C/decade, respectively. e seasonal warming trend differences between observation and ERA-Interim are lower than 0.1°C/decade. Generally, ERA-Interim is reliable for warming trend detection over the QLM.
ERA-Interim has different performances at different altitudes. Generally, RMSE increases along higher elevations. e largest RMSE for daily T_mean and T_min is found in the 3000-3500 m elevation group, and for T_max it is found in the 2500-3000 m group. It indicated that ERA-Interim may be weaker for the regions that are higher than the ERA-Interim model height.
By now, this evaluation is limited to 24 meteorological stations' range in elevation from 1000 m to 4000 m. e analysis can be further extended by adding more observations in the surrounding areas. It should also be worth trying to investigate other meteorological elements of ERA-Interim reanalysis such as precipitation and humidity over the QLM.

Data Availability
e ERA-Interim data used in this paper are provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). Station data are provided at China Meteorological Data Sharing Service System of National Meteorological Information Centre (http://data.cma.cn/, last access: 06 February 2020).

Conflicts of Interest
e authors declare that there are no conflicts of interest in this paper. Advances in Meteorology 9