Effects of Initial Drivers and Land Use on WRF Modeling for Near-Surface Fields and Atmospheric Boundary Layer over the Northeastern Tibetan Plateau

To improve the simulation performance of mesoscale models in the northeastern Tibetan Plateau, two reanalysis initial datasets (NCEP FNL and ERA-Interim) and two MODIS (Moderate-Resolution Imaging Spectroradiometer) land-use datasets (from 2001 and 2010) are used in WRF (Weather Research and Forecasting) modeling. The model can reproduce the variations of 2m temperature (T2) and 2m relative humidity (RH2), but T2 is overestimated and RH2 is underestimated in the control experiment. After using the new initial drive and land use data, the simulation precision in T2 is improved by the correction of overestimated net energy flux at surface and the RH2 is improved due to the lower T2 and larger soil moisture. Due to systematic bias in WRF modeling for wind speed, we design another experiment that includes the Jimenez subgrid-scale orography scheme, which reduces the frequency of low wind speed and increases the frequency of high wind speed and that is more consistent with the observation. Meanwhile, the new drive and land-use data lead to lower boundary layer height and influence the potential temperature and wind speed in both the lower atmosphere and the upper layer, while the impact on water vapor mixing ratio is primarily concentrated in the lower atmosphere.


Introduction
Near-surface meteorological fields are the most important and basic elements for weather and climate research and provide crucial information for water resources, wind energy, and agricultural activities.For example, an accurate spatial and temporal description of the temperature field is essential for evaluations of water resources and ecosystems and is also an important input variable for hydrological models [1].At the same time, because it acts as a link between land surface and free atmosphere, the atmospheric boundary layer affects regional climate via land-atmosphere coupling of momentum, energy, water, and matter [2,3].
Observations and numerical simulations are the two most important methods used to acquire information from nearsurface meteorological fields and the atmospheric boundary layer structure.However, the distribution of stations is still sparse in regions with complex terrain such as the Tibetan Plateau, and the detailed spatial distribution of meteorological fields cannot be obtained from routine stations observations, including the complex flow field caused by topographic forcing, terrain obstruction, and funneling [4].Recently, with the development of mesoscale numerical models, the advantages of numerical simulations have been highlighted due to their ability to provide high-resolution data for near-surface meteorological fields and the atmospheric boundary layer structure.In addition, mesoscale models are commonly used to provide an in-depth understanding of the relevant physical process and mechanisms.The WRF (Weather Research and Forecasting) is a mesoscale numerical model that has been widely used since its development in 2000.Many researchers have studied the factors that affect the model performance and have found that the WRF simulation accuracy depends on spatial resolution, initial conditions, parameterization scheme, and driving data [5][6][7][8].

Advances in Meteorology
Reanalysis datasets provide initial and boundary conditions for mesoscale models, and the precision of these data directly influences the model performance.To date, few studies have focused on the impact of different driving data on WRF performance on the northeastern Tibetan Plateau, which is the headwater of many inland rivers in western arid and semiarid China.Thus, analysis of the impact of different reanalysis data on the WRF simulation of nearsurface meteorological fields and the atmospheric boundary layer structure is important for studies of land-atmosphere interaction in the region.
Additionally, the accuracy of land surface parameters, including topography, land use, vegetation cover, and soil type, also influences the modeled land surface processes and atmospheric boundary layer characteristics.These variables greatly influence the model performance and directly determine surface parameters such as albedo, emissivity, roughness, leaf area index, vegetation roots, impedance vegetation, soil porosity, and soil thermal conductivity [9].Previous studies primarily focused on parameterization schemes for land surface models [10,11].Recent studies have explored the effect of land surface data on the simulation accuracy of near-surface meteorological fields through the improvement of land surface data [12][13][14][15].In the WRF model, the default land-use information is taken from the 2001 MODIS-based land-use data, and, thus, we can investigate the possibility of improving WRF performance by applying newly acquired land-use data.
To better understand the applicability of the driving and underlying surface data for the WRF model, two reanalysis datasets (NCEP FNL and ERA-Interim) and two MODIS land-use datasets (from 2001 and 2010, resp.) were chosen to explore the possibility of improving model performance in the region.The outline of the paper is as follows.Section 2 introduces the data and methods.Sections 3 and 4 show the impacts of driving and land-use data on the nearsurface meteorological fields and atmospheric boundary layer, respectively.Thesummary and conclusions from the results are presented in Section 5.

Study Area and Ground Meteorological Observation Data.
The northeastern Tibetan Plateau (94 ∘ 39  -103 ∘ 27  E, 35 ∘ 51  -40 ∘ 31  N), with an elevation range from 758 m to 5725 m a.s.l., was chosen as the study area, as shown in Figure 1(a).This area contains the headwaters of many inland rivers (e.g., the Heihe and Sule rivers) and plays an important role in the hydrology and agriculture of the downstream arid region.A total of 34 meteorological stations are located in the study area, and these stations provide daily observations of the 2 m temperature, 2 m relative humidity, and 10 m wind speed.Among these stations, Arou station is located in the Heihe River watershed with an elevation of 3033 m.The observation systems at this station include a meteorological tower used to measure the gradients of meteorological variables such as surface temperature, humidity, wind speed, soil temperature, soil humidity, surface heat flux, and radiation flux at intervals of 10 minutes and an eddy covariance system that records     Advances in Meteorology [18] scheme as the subgrid-scale orography parameterization scheme compared with the CTRL experiment.

Effects of Driving and Land-Use Data on
Near-Surface Meteorological Fields  To summarize, the new initial field and underlying surface condition show improvement in modeling T2 and RH2, though not providing sufficient improvement in U10.

Analysis of the Improvement in Simulated Meteorological
Field.In this section, we take Arou station as an example to analyze how the model improves the near-surface meteorological field simulation under the new initial condition and land-use data.

Surface Radiation.
Figure 3 shows the time series of observed and simulated surface radiation (short wave (SW), reflected short wave (RSW), long wave (LW), and upward long ware (ULW)) in June at Arou station.The corresponding statistical analysis of the comparisons is shown in Table 3.In Figure 3(a), the modeled short-wave radiation agrees well with observations (with  of 0.81 to 0.87), and  in INTL experiment is larger than MODS experiment.Around 12:00 on June 3 (local time, all the time used in latter part is local time), the measured solar short-wave radiation decreases suddenly.This decrease might be caused by large cloud cover.The modeled SW in INTL experiment is closer to the observed values in this period, which means that the ERA-Interim driving data can successfully simulate the impact of cloud cover in this day.In Figure 3(b), four experiments simulate the reflected solar radiation well, with  of 0.82 to 0.89, because the simulated albedos are rather close to the observed value (Table 4).Due to the deviation of the modeled solar short-wave radiation, the simulation values of the RSW are much higher than the observed results at noon (e.g., on June 18 and 19).In Figure 3(c), the simulated deviation in the LW is greater than that of other radiation variables, with  of 0.59 to 0.62 and MB of −23.4 to −20.7 wm 2 .The underestimation of LW was also found in previous studies [19,20].In Figure 3(d), the variation in ULW is reproduced well, with  of 0.82 to 0.86, MB of 4.3 to 14.3, and NMB of 1.1 to 3.8.The MODS experiment has greater precision in simulation of ULW compared with CTRL and INTL experiments, due to lower emissivity as shown in Table 4 and the lower emissivity leads to smaller ULW.

Surface Energy.
Solar radiation is the most fundamental source of earth-atmosphere system energy.A portion of solar radiation is used to release sensible heat and latent heat to provide energy for transportation of the turbulent boundary layer, and the other portion is absorbed by the land surface via the heat transfer processes.Figure 4 shows the comparisons between observed and modeled net radiance (NR), sensible heat (SH), latent heat (LH), and soil heat flux (SHF) in the four experiments, and the corresponding statistical analysis is shown in Table 4.In Figure 4(a), the simulated NR agrees well with the observed values, with  of 0.84 to 0.87.However, on several days (e.g., June 2 at noon), the NR are larger than the measurements because of the overestimation of the solar short-wave radiation.In Figure 4(b), the variation tendency of the modeled SH is relatively consistent with the observations (with  of 0.74 to 0.77), and the negative values are simulated in all experiments (low-level inversion structure).However, the SH is significantly overestimated (MB = 43.6,34.9, 38.6, and 32.4,resp.), especially in the CTRL experiment where the largest deviation reaches 162 W⋅m −2 .This effect might be due to the limitation of the Monin-Obukhov similarity theory (MOST) used in the WRF.Recently, variational methods have been used to take into consideration both observations and MOST has been used to improve the sensible heat flux computations [5,21,22].In future research, we will explore this observation more fully.The observed albedo is 0.218.
In Figure 4(c), the model can simulate the diurnal variability of LH (with  of 0.83 to 0.85), and the LH is overestimated in four experiments, especially in CTRL experiment; the ME is equal to −32.5.In Figure 4(d), the modeled SHF are around 20 W⋅m −2 larger than the observations during the day and approximately 50 W⋅m −2 smaller than the observations at night.Therefore, the simulated SHF is obviously underestimated in experiments (ME = −6.3,−5.2, −3.5, and −2.2, resp.).This may be due to the modeled soil heat flux (SHF) values modeled at the soil surface, but the observations are measured at 5 cm depth in the soil.The heat storage effect at a depth of 0-5 cm reduces the diurnal range of soil heat flux.

Soil Temperature and Humidity.
In Figure 5(a), the model can reproduce the variation of soil temperature (ST) (R = 0.63, 0.68, 0.65, and 0.69, resp.) and the ST is overestimated with MB of 1 to 3.2, for example, during the period from June 16 to June 20.Among the four experiments, the simulations in the INMO experiment are closest to the observations; the CTRL experiment always displays the maximum bias due to the largest soil temperature.The simulation period in this study is relatively short, so the initial values of soil temperature and moisture have greater impact on the simulation compared with the new MODISbased land-use data.

Near-Surface Meteorological Fields.
A comparison of the modeled and observed near-surface meteorological elements is shown in Figure 6, and the corresponding statistical analysis is presented in Table 5.In Figure 6    In Figure 6(b), the four experiments approximately simulate the daily variation in the 2 m relative humidity, with  of 0.50 to 0.55.Specifically, RH2 is obviously underestimated in CTRL experiment (MB = −30.6),due to the overestimated temperature at surface under insignificant change of water vapor pressure.Compared with CTRL experiment, the simulation is improved in three sensitivity experiments, because of the lower values of modeled T2.Additionally, the higher soil moisture in sensitivity experiment increases the soil heat capacity and thermal conductivity and then augments the surface evaporation.The highest improvement in RH2 is in INMO experiment, followed by INTL experiment and then MODS experiment.
In Figure 6(c), for the 10 m wind speed, the simulated values are lower than the observations, and the simulation precision is very low in these four experiments (with  of 0.16 to 0.21).Previous studies have demonstrated that systematic biases exist in WRF modeling of wind speed in complex terrain areas [10].Therefore we designed another JIME experiment that uses the Jimenez scheme as  the subgrid-scale orography parameterization scheme.As the results in Figure 7(a) show, the CTRL experiment significantly overestimates the frequency of wind speeds below 1 m/s and underestimates significantly the frequency of wind speed between 2 and 4 m/s.Compared with the CTRL experiment, the JIME experiment greatly decreases the frequency of low wind speed and increases the frequency of high wind speeds, which is more consistent with the frequency of the observations of wind speed.As shown in Figure 7(b), the wind direction changes frequently and is directed mainly toward the south during the research period.The simulated wind direction is similar in the two experiments.Figure 8 shows the mean hourly wind speed at Arou station in June.
The two experiments accurately simulate the daily variation in wind speed.However, the simulated wind speeds are      ).The changes in potential temperature in the upper atmosphere are not consistent with the changes in the lower atmosphere.The impact on the water vapor mixing ratio is primarily found in the middle and lower atmosphere (Figures 10(b), 10(d), and 10(f)).Figure 11 shows the differences of vertical profile for the average U and W wind speed along 100.46 ∘ E. The disturbance for U wind speed is relatively larger than that of W wind speed.There is a clear extreme value area which can be observed in the lower atmosphere at approximately 40 ∘ N. Figures 12 and 13 present the average vertical profiles of the boundary layer at Arou station in June at 12:00 and 0:00, respectively.In Figure 12(a), for potential temperature (at 12:00), the boundary layer is warmest in the CTRL experiment, whereas the boundary layer is coldest in the MODS experiment.This observation suggests that the new land-use data have a greater impact on potential temperature compared with the ERA-Interim driving data.At 0:00, compared with the CTRL experiment, the simulated values in the MODS and INMO experiments are higher below 750 m.Generally, the differences for potential temperature The four experiments are able to accurately simulate the diurnal variation of the 2 m temperature and relative humidity.The ERA-Interim driving data and updated MODS-based land-use information improve the simulation in T2 and RH2, through the correction of overestimated surface net energy flux and underestimated soil moisture.Previous studies also pointed out that the WRF model is highly sensitive to soil moisture [23,24] and the ERA-Interim reanalysis data have a greater reliability of application in China compared with the NCEP data [25][26][27].However, both the new initial driving and underlying surface data do not lead to sufficient improvement for the 10 m wind speed due to the complex terrain of the Tibetan Plateau.Therefore, we designed another JIME experiment to analyze the effect of topography on the 10 m wind speed.The JIME experiment greatly decreases the frequency of the low wind speed and increases the frequency of the large wind speed, which is more consistent with the observations.With the ERA-Interim reanalysis and 2010 MODIS-based land-use data, averaged for the study area, the experiments result in a lower boundary layer height due to smaller net energy flux.Additionally, the lower T2 and higher RH2 in these experiments also make the BLH higher.For the potential temperature and wind speed, the new initial conditions and underlying surface influence the lower atmosphere as well as the upper layer, and the impact on the water vapor mixing ratio is primarily concentrated in the lower atmosphere.Generally, the difference among simulated results in different experiments at 0:00 is less than that at 12:00.
Land-Use Data.Two types of initial driving data are used in this research.The NCEP FNL data from the Global Data Assimilation System (GDAS) have a spatial resolution of 1 ∘ × 1 ∘ and a time interval of 6 h.The data include ground information and a total of 26 layers (1000 hPa to 10 hPa) of isobaric surface data.The ERA-Interim program began in 2006 and was intended to improve and gradually displace the ERA-40.The spatial resolution of ERA-Interim reanalysis data used in this study is 0.75 ∘ × 0.75 ∘ , and the time interval is 6 h.These data include ground information and a total of 37 layers (1000 to 1 hPa) of isobaric surface data.Figures1(c) and 1(d) show the U wind speed from the NCEP FNL and ERA-Interim reanalysis data, respectively, and we note existence of large differences between the two datasets.The default land-use data (Figure1(e)) of the WRF model are the 2001 MODIS-based land-use data provided by Boston University.However, in recent years, the land-use types have obviously changed under the influences of climate change and human activity.Therefore, we used the year 2010 of the 16-day synthesized MODIS enhanced vegetation index (EVI) data and days 223 to 230 (no clouds phase) of the 8day synthesized surface reflectance and SRTM DEM digital elevation data to remake the land-use dataset according to the classifications of the MODIS 2001 land use.The specific steps are described as follows.First, for the EVI and surface reflectance data, we used the principal component transform to compress the amount of data and calculated the degree of homogeneity of the first principal component obtained from the reflectance data using the gray level cooccurrence matrix.Then, we constructed a classification data matrix consisting of the EVI principal component, reflectance principal component, digital elevation, and homogeneous degree information.Next, we chose typical training areas for all types of surface objects based on high-resolution Google Earth data and converted them into training samples corresponding to the MODIS data.Finally, we used the decision tree classifiers constructed by the CART algorithm to perform computer classification and obtained the new land-use classification of the study area.The 2010 MODIS-based land-use data are shown in Figure 1(f).

Figure 5 (
b) shows the comparison of soil moisture between observed and simulated results.The experiments significantly underestimate the soil moisture, with MB of −0.18 to −0.14 and NMB of −47.3 to −36.8.The maximum underestimation occurs in CTRL experiment, and the simulation in INTL experiment is closer to the observation compared with the MODS experiment.
(a), T2 is well represented (R of 0.77 to 0.81) and overestimated in four experiments.The overestimation of T2 is due to overestimated net energy flux (NE, NE = SW + LW + LH) at surface.The observed mean NE is 681.8 wm 2 in June, while the larger NEs (730.4,716.5, 721.4,and 715.1) are for CTRL, INTL, MODS, and INMO experiments, respectively.Larger surface NE leads to higher T2, and the maximum overestimation of T2 occurs in CTRL experiment (NMB = 24.4%)and the minimum in INMO experiment (NMB = 18.5%), and the simulation precision in INTL experiment is higher (RMSE = 21.4)than MODS experiment (RMSE = 23.2).

Figure 7 :
Figure 7: Figures of probability density distribution: (a) wind speed and (b) wind direction.

Figure 8 :
Figure 8: Mean hourly wind speed at Arou station in June.

Figure 10 :Figure 11 :Figure 12 :
Figure 10: Vertical profile difference for the average potential temperature and water vapor mixing ratio along 100.46 ∘ E in June: (a) INTL-CTRL for potential temperature, (b) INTL-CTRL for water vapor mixing ratio, (c) MODS-CTRL for potential temperature, (d) MODS-CTRL for water vapor mixing ratio, (e) INMO-CTRL for potential temperature, and (f) INMO-CTRL for water vapor mixing ratio.

4. 2 .
Structure of the Atmospheric Boundary Layer.
Figure 10   shows the vertical profile difference for the average potential temperature and water vapor mixing ratio along 100.46 ∘ E. It can be observed that the impact of the new driving and land-use data is more complicated for potential temperature (Figures10(a), 10(c), and 10(e)
2.2.Model Setup and Experimental Design.The numerical experiments in this study were conducted using the Advanced Research WRF model Version 3.5.The WRF is a nonhydrostatic, primitive-equation, mesoscale meteorological model with advanced dynamics, physics, and numerical schemes (details of the model can be found at http://www .mmm.ucar.edu/).In this study, the model domains are twoway nested with 24 km (103 × 91 grids) and 8 km (118 × 91 grids) horizontal spacings (Figure1(b)).Each domain contains 28 vertical pressure levels with the top level set at 50 hPa.The scope of domain d02 is consistent with the study area shown in Figure1(a).The WRF physical parameterization schemes used in this research include the Purdue Lin microphysical parameterization, Rapid Radiative Transfer Model (RRTM) long-wave radiation scheme, Dudhia short-wave radiation scheme, Monin-Obukhov surface layer, Noah land surface, Mellor-Yamada-Janjic (MYJ) planetary boundary layer scheme, and Grell-Devenyi (GD) cumulus scheme.

Table 2 :
Statistics of 2 ( ∘ C), RH2 (%), and 10 (m/s) between the observation and model results. is the number of paired samples;  OBS and  MOD are the average values of observation and model results;  is the correlation coefficient between the observation and model results; MB and NMB are the mean biases between the observation and model results and the normalized mean bias between the observation and model results, respectively; RMSE is the root-mean-square error of observation and model results.
highest precision occurs in INMO experiment, whose  increases by 0.05 and MB reduces by 0.6 compared with CTRL experiment.Additionally,  in INTL experiment is 0.01 larger and MB is 0.1 K smaller than those in MODS experiment.For RH2, the largest underestimation occurs in CTRL experiment (MB = −18.1),and the three sensitivity experiments show somewhat improvement (with MB of −11 to −9.9).Compared with CTRL experiment,  increases by 0.06 and RMSE reduces by 7.2 in INMO experiment which presents the highest precision among four experiments.For wind speed, the model performance is worse than the above two meteorological parameters.The correlation coefficients between the observed and simulated values are quite low in the four experiments, and the observed mean wind speed is higher than the modeled values.

Table 4 :
Land surface information and parameters at Arou.

Table 5 :
Statistics of 2 m temperature 2, RH2, and 10 between the observation and model results at Arou station.
).The surface NE averaged in the study area is 731.2 wm 2 in CTRL, and the NEs are 717.9wm 2 , 715.3 wm 2 , and 710.7 wm 2 in INTL, MODS, and INMO experiments, respectively.Additionally, the variation in boundary layer height is related to turbulent transport, which is affected by the temperature and humidity conditions near the ground.A warmer and drier near-surface layer is conducive to the development of the boundary layer.As shown in Table2, the mean T2 is highest and the mean RH2 humidity is lowest in CTRL experiment, and the higher T2 and lower RH2 lead to larger BLH in CTRL experiment.