Effect of soil texture and hydraulic parameters on WRF simulations in summer in east China

This article investigated the effect of soil texture, i.e. 5‐min Food and Agriculture Organization soil texture (STFAO) and 30' Harmonized World Soil Database soil texture (STHWSD), and hydraulic parameters, i.e. US Department of Agriculture soil hydraulic parameters (SPUS) and China soil hydraulic parameters (SPCH), on the Weather Research and Forecasting (WRF) model simulation in summer in eastern China. Near‐surface meteorological fields from 365 automatic weather stations were used to evaluate the performance of WRF. Obvious differences between the two soil texture datasets and the two soil hydraulic parameter datasets were found. The simulated 2‐m temperature, 2‐m specific humidity, and 10‐m wind speed were improved significantly at a 95% confidence interval via bootstrap test when STHWSD was used. The agreement is weaker as SPUS was replaced by SPCH. Soil texture and hydraulic parameters affect surface energy partitioning and the distribution of precipitation. The influence of wilting point on WRF's performance is more significant than other soil hydraulic parameters.


Introduction
Compared to surface data, acquirement and verification of soil texture are more difficult, and studies on the effect of more representative soil texture on numerical simulations are relatively less. Five-min Food and Agriculture Organization (FAO) soil texture (outside United States, hereafter referred as ST FAO ) is widely used in numerical simulations. Gao et al. (2008) found a noticeable effect of the new soil texture generated from 1 : 100 000 resolution subgroup soil types on MM5 simulations over the Heihe River Basin compared to the default ST FAO . However, the high-resolution soil texture used in Gao et al. (2008) covers a limited region. Recently, the Harmonized World Soil Database (HWSD), established by the FAO and collaborators, provides soil information on 30' resolution over the world, from which high-resolution soil texture maps can be derived (hereafter referred as ST HWSD ). The impact of ST HWSD on numerical simulations needs to be evaluated.
Soil parameters play an important role in soil thermal and hydrological processes and have a significant effect on boundary layer and precipitation simulations (Horváth et al., 2009;Breuer et al., 2012). The most widely used soil parameter dataset in numerical models is the US Department of Agriculture (USDA) dataset (Cosby et al., 1984, hereafter referred as SP US ). As soil parameters are spatially heterogeneous, models used for regional studies prefer regional specific soil parameters (Horváth et al., 2009). Recently, Dai et al. (2013) developed a soil hydraulic parameter dataset for China based on the 1 : 1 000 000 soil map of China and 8595 representative soil profiles (hereafter referred as SP CH ). The effect of this dataset on numerical simulations is unclear and needs further research.
In this article, two sets of soil texture (ST FAO and ST HWSD ) and two sets of soil hydraulic parameter data (SP US and SP CH ) are used in the Weather Research and Forecasting (WRF) model to find a suitable soil dataset for mesoscale numerical simulation over eastern China.

WRF model run
WRF-ARW V3.3 was configured with Single-Moment 6-class microphysics scheme, the Kain-Fritsch cumulus parameterization scheme, the Rapid Radiative Transfer Model longwave and Dudhia shortwave radiation parameterization scheme, the Yonsei University planetary boundary layer parameterization scheme, and the Noah land surface model (LSM) parameterization scheme (Chen and Dudhia, 2001), which have been widely evaluated over China and other regions (Lo et al., 2008;He et al., 2013He et al., , 2014. Three nested domains with horizontal resolutions of 25, 5, and 1 km were used to reduce the spurious boundary effects on the interested inner domain (Figure 1(a)). The inner domain covers the Taishan Mountain and the southeast of the North China Plain with vegetation cover ranging from 20 to 80%. Vertically, there were 35 full eta levels extending to the model top at 50 hPa, with 16 levels below 2 km. One-month integration from 30 June 2012 to 31 July 2012 was conducted, of which 20% were sunny or cloudy days, 35% were shallow convection, and 45% were deep convection. The first 24 h run was disregarded as spin up. The time-step is 150, 30, and 6 s for the three domains respectively. The National Centers for Environmental Prediction (NCEP) Final Operational Global Analysis (FNL) data were used as driving field for WRF, because it has the same soil level as WRF, which is beneficial to decrease model error. Three model runs were performed: SIM1 with ST FAO and SP US , SIM2 with ST HWSD and SP US , and SIM3 with ST HWSD and SP CH . When driving field is not significantly coarser than the model resolution, analysis nudging is very useful for improving model performance (Stauffer and Seaman, 1994). In this work, analysis nudging was used for domain D01. The normal one-way nesting was used for SIM1. Results from D02 of SIM1, stored at every time-step, were used as the input for D03 of SIM2 and SIM3 to keep the same initial and boundary conditions as the base run (SIM1). Hourly data from D03 was used for evaluating the model performance. Three typical weather situations were selected to deeply investigate the effect of soil parameters on model results performance. The details were provided in Appendix A.

Soil texture and hydraulic parameters
Soil texture in the default dataset of WRF, ST FAO , is classified into 16 categories according to the USDA soil texture classification. Loam and clay loam are the main soil types in the study area, accounting for 50.3 and 49.4%, respectively (Figure 1(c)). For ST HWSD , loam, loamy sand, and sandy loam are the main soil types in the study area, accounting for 77.5, 10.9 and 6.3%, respectively (Figure 1(d)). ST HWSD has been described in our previous study (He et al., 2016). Large differences exist between ST FAO and ST HWSD , especially in the west and the north of the study domain. SP US was derived from 1448 soil samples in the United States, in which 78% of the samples are sand and 10% are silt (Cosby et al., 1984). SP CH was derived from 8595 soil profiles in China using pedotransfer functions (Dai et al., 2013). The most significant difference between SP CH and SP US is the wilting point ( w ), with 117% mean differences (Table S1, Supporting information). The differences in saturated soil moisture ( S ), field capacity ( f ), Campbell's porosity index (b), saturated soil water potential ( S ), and saturated soil moisture conductivity (k S ) are 6, 20, 9, 45, and 35% on average, respectively. The effects of soil texture and parameters are realized through LSM. A brief introduction of Noah LSM was provided in Appendix B.

Other land surface information
The resolution and accuracy of land surface information affect model results, especially the temperature . In this study, 500-m resolution moderate resolution imaging spectroradiometer (MODIS) land use in 2012, 1-km resolution vegetation fraction derived from MODIS Normalized Difference Vegetation Index in 2012, and 90-m resolution Shuttle Radar Topography Mission terrain were used instead of the default data supplied by WRF.

Observational data
Hourly 2-m temperature (T 2 ), 2-m specific humidity (Q 2 ), 10-m wind speed (WS 10 ), and precipitation (PREC) from the Jinan Meteorological Bureau were used to evaluate the model results. T 2 and Q 2 were recorded as instantaneous values every hour, while WS 10 and PREC were recorded as 2-min average and 1-h accumulation, respectively. The dataset include 356 automatic weather stations (AWS) with the best data quality in the study area (Figure 1(b)). Several statistic indices used for model evaluation and significant test were supplied in Appendix C.

Performance of WRF
The model's performance was evaluated by comparing the simulated results with the available observations at 356 AWSs. The statistics was provided in Table 1. WRF can well reproduce T 2 , with the index of agreement (IOA) exceeding 0.9, followed by Q 2 , WS 10 , and PREC with the IOA of 0.68, 0.46, and 0.34, respectively. The correlation coefficient between simulations and observations is significant according to t-test at 95% confidence interval. The performance of WRF in this study is comparable with previous studies (Carvalho et al., 2012;He et al., 2013). Based on one-way analysis of variance (p < 0.05), WRF significantly overestimates T 2 and WS 10 , and significantly underestimates Q 2 and PREC. The uncertainties of physical parameterizations, including land surface, boundary layer, and microphysics processes, are important factors affecting model performance (Carvalho et al., 2012;He et al., 2014). Apart from physical parameterizations, the deriving fields (i.e. reanalysis data or Global Climate Model (GCM) data) also affect model performance significantly (Ma et al., 2014). The causes of simulated bias will not be discussed here as it is out of the scope of our focus.

Effect of soil texture
Clay loam was replaced by loam in most of the western and the northern regions when ST HWSD was used instead of ST FAO (Figure 1), resulting in −28 to 56% changes in soil hydraulic parameters (Table S1). These changes affected surface evaporation and soil thermodynamics and hydrology (Equations (A1)-(A15)), and in turn the near-surface meteorological fields. The near-surface meteorological fields were improved in SIM2 except for PREC. Root mean square error (RMSE) decreased by 0.05 K, 0.08 g kg −1 , and 0.02 m s −1 for T 2 , Q 2 , and WS 10 , respectively and increased by 0.005 mm h −1 for PREC (Table 1). The changes of RMSE for T 2 , Q 2 , and WS 10 are significant at 95% confidence interval via the bootstrap test, while not significant for PREC. The details of significant test for statistical indices are listed in Table S2. The improvement of model performance related to soil texture and the accompanying parameters is more obvious during daytime (Figures 2 and S3, Supporting information). The relative changes of RMSE (defined in Equation (A22)) between SIM1 and SIM2 for three typical weather conditions (Table S3), imply that the impact of soil texture on T 2 is relatively small under shallow convection conditions, which may be related to model uncertainties in reproducing local convection. The use of ST HWSD improved Q 2 and WS 10 the least on sunny days.
Replacing ST HWSD with ST FAO involved changes of soil texture. Changes at the 356 AWSs include clay loam to loam (60%), loam to loamy sand (12%), loam to sandy loam (7%), clay loam to loamy sand (6%), and loam to sandy clay loam (5%), which makes an obvious increase of the quartz content (q tz ) and an obvious decrease of S and w (Table S1). The potential evaporation (E p ), the ground surface evaporation (E dir ), the vegetation transpiration (E t ), and the evaporation of intercepted water (E c ) from different model runs were compared in Figure 3 to investigate the effect of soil texture data on latent heat (LH). As can be seen from Figure 3 and Table S1, Sw and E dir increased and E p , volumetric soil water content ( ), and w decreased, these indicates that the decrease of w is responsible for the increase of E dir based on Equation (A10). The decrease of w and fw leads to the increase of soil moisture factor (F 4 ) and the decrease of the canopy resistance (R c ), and in turn the increase of the function of canopy resistance (B c ) based on Equations (A13)-(A15), which eventually results in the increase of E t (Equation (A11)). Compared to E dir and E t , E c and its changes are very small. The decrease of indicates that the water holding capacity of ST HWSD is much lower than that of ST FAO . Considering the high vegetation cover in the study area, the differences of monthly average E dir and E t between SIM1 and SIM2 ( Figure S4) indicate that transpiration plays the major role in flux differences between SIM1 and SIM2. Net radiation increases with the decrease of land surface longwave radiation relating to the decrease of surface temperature ( Figure S5) when albedo and emissivity do not change. The change of soil texture from ST FAO to ST HWSD increases q tz and S , and leads to the increase of thermal conductivity (K, Equations (A3)-(A5)), making soil heat transfer more easy. The changes of E dir and E t increase LH. Based on energy conservation, sensible heat (SH) decreases. The difference of cloudiness and precipitation between SIM1 and SIM2 result in differences in radiation and heat fluxes. The changes of heat fluxes are responsible for the decrease of T 2 and the increase of Q 2 . The decrease of turbulence mixing weakens momentum transfer and reduces WS 10 slightly. Soil texture affects the distribution of monthly accumulated PREC, but its impact on total PREC is negligible ( Figure S6). Conditions of deep and shallow convections were further investigated by comparing the LH and convective available potential energy (CAPE) before the occurrence of PREC. LH increases in most of the northern and northwestern areas while decreases in most of the southeast areas when ST HWSD is used instead of ST FAO (Figures 4  and S7). The increase (decrease) of LH results in the decrease (increase) of lifting condensation level, and increases (reduces) the CAPE. The change of atmospheric stratification finally alters the distribution of PREC. Compared to SIM1, the total PREC in SIM2 decreases by 8% in shallow convection, while increases by 1% in deep convection, which indicates that shallow convection is more sensitive to soil texture data than deep convection. In a word, the change of surface energy partitioning, related to the change of soil texture, improves the performance of near-surface meteorological fields, with significant improvement for T 2 , Q 2 , and WS 10 . The distribution of PREC changes slightly, with negligible effect on total PREC.

Effect of soil hydraulic parameters
The use of SP CH weakens WRF performance (Table 1 and Figure 2). The difference of RMSE is significant for T 2 , Q 2 , and WS 10 . Compared with SP US , S and w increase dramatically, while f decreases for loam, loamy sand, and sandy loam (Table S1), which are the major soil types for ST HWSD . The changes of S and w result in the decrease of Sw , and have a potential to increase E dir (Equation (A10)). However, the increase of w is more significant, which leads to the decrease of E dir (Figure 3(b)). Though there is a decrease of R c , the change of temperature decreases B c during daytime  via R r and results in a decrease of E t during daytime (Figure 3(c)). LH decreases due to the change of E dir and E t , and SH increases. The change of surface energy partitioning increases T 2 and reduces Q 2 ( Figure S3). The differences of monthly average E dir and E t between SIM2 and SIM3 ( Figure S4) indicate that differences in transpiration is the main contributor to the differences in heat fluxes due to the differences in soil hydraulic parameters. The impact of soil hydraulic parameters seems to be more important on sunny days than on convective days (Table S3). The standard deviations (STD) of soil hydraulic parameters for individual soil categories are large over China (Dai et al., 2013), which may affect model performance. The use of SP CH instead of SP US weakens WRF's performance, which may be related to the large STD and the uncertainties in SP CH . Local soil parameter table or observed soil parameter with high spatial resolution may be more appropriate for regional studies. On the other hand, SP US has been used and improved by a wide community. Its error in China may be counteracted by other errors in WRF and thus a better performance.
Further sensitivity tests were conducted to investigate the effect of soil hydraulic parameters. Six sensitivity tests under three typical weather conditions were performed using SIM3 as the reference run (Table S4). The RMSE of T 2 , Q 2 , and WS 10 from these sensitivity tests indicate that smaller S , f , and w values are more suitable for the study area ( Figure S8). Near-surface meteorological conditions are most sensitive to w , followed by f and S . This is related to the evaporation and transpiration processes (Equations (A10)-(A15)) in which w plays an important role. However, the importance of w may change with season, as transpiration is very low in winter. It is also related to the coefficients of variation for S , f , and w , which are 8, 20, and 28% respectively. w has been noted to be an important soil parameter in previous studies (Mölders et al., 2005;Ács et al., 2010).

Conclusions
In this study, the impact of two soil texture datasets (i.e. ST FAO and ST HWSD ) and two soil hydraulic parameter datasets (i.e. SP US and SP CH ) on WRF's performance in summer in eastern China was investigated. The vegetation cover in the study area ranges from 20 to 80% with an average of 46%. One-month simulation (July 2012), including 20% sunny or cloudy days, 35% shallow convection, and 45% deep convection, was analyzed. Compared to ST FAO , ST HWSD is more heterogeneous. Though the difference of near-surface meteorological parameters simulated using different soil texture datasets is quite low, the T 2 , Q 2 , and WS 10 were improved significantly when ST FAO was replaced by ST HWSD in WRF. Water holding capacity of ST HWSD is much lower than that of ST FAO . Significant difference between SP CH and SP US exists in w , with 117% mean differences. The differences in S , f , b, S , and k S are 6, 20, 9, 45, and 35% on average, respectively, between SP CH and SP US . The agreement is weaker as SP US was replaced by SP CH , which may be related to the large STD and the uncertainties of soil hydraulic parameters in SP CH . Considering the high vegetation cover, changes in transpiration related to the changes in soil parameters is the main contributor leading to the change of heat fluxes and meteorological parameters. Local soil parameter table or directly observed soil dataset with high spatial resolution may be more appropriate for regional studies. Soil texture and hydraulic parameters affect surface energy partitioning and the distribution of PREC. The influence of w on WRF's performance is more significant than other soil hydraulic parameters.
Except for the 1-month simulation, three typical weather situations, i.e. sunny day (1 July), deep convection (4 July), shallow convection (17 July), were selected to deeply investigate the effect of soil parameters on model results. Figure S1 shows the weather maps at 500 hpa at 0800 BT (Beijing Time) on 1, 4, and 17 July 2012 over China. The subtropical anticyclone covers the south of the Huaihe River on 1 July 2012. Cold air controls Jinan and the surrounding regions with significant cold advection and northwesterly wind at 500 hpa. This circulation pattern favors sunny and cloudless weather as manifested by the satellite cloud image ( Figure S2(a)). The trough line at 500 hpa is located near Jinan on 4 July 2012. The convergence zone and sufficient water vapor produce strong and deep convection ( Figure S2(b)). The subtropical anticyclone covers the east of China on 17 July 2012. A week cyclone is discovered in the study area at 500 hpa, which is beneficial for the occurrence of shallow convection. The satellite cloud image ( Figure S2(c)) depicts that some local shallow convections existed in the study area.

Appendix B: Noah LSM
Noah LSM is a state of the art one-dimensional LSM developed from OSU soil-vegetation model, including soil thermodynamics and hydrology processes directly affected by soil parameters. It has four soil layers. The depths of soil layers are 0.1, 0.4, 1, and 2 m, respectively.
Soil thermodynamics is controlled by the usual diffusion equation: where T and z are soil temperature and thickness, respectively. The volumetric heat capacity (C) and the thermal conductivity (K) are functions of volumetric soil water content ( ): C water , C ice , C air , and C soil are the volumetric heat capacities of liquid water (4.2 × 10 6 J m −3 K −1 ), ice (2.106 × 10 6 J m −3 K −1 ), air (1004 J m −3 K −1 ), and soil (2.0 × 10 6 J m −3 K −1 ), respectively. water is the moisture content of unfrozen soil. Kersten number (K e ) is a function of the saturated soil moisture ( S ) and . K dry is the dry thermal conductivity, which depends on soil density. Saturated thermal conductivity (K sat ) is controlled by the thermal conductivity of different components: K ice , K water , K quartz , and K other are the thermal conductivities of ice (2.2 W m −1 K −1 ), water (0.57 W m −1 K −1 ), quartz (7.7 W m −1 K −1 ), and other (2.0 W m −1 K −1 ), respectively. q tz is the quartz content in soil. Soil hydrology is described by the Richards equation: where s is the sink term, including precipitation, evaporation, transpiration, and runoff. Soil diffusivity (D) and hydraulic conductivity (k) are functions of and S : where b and k S are Campbell's porosity index and saturated soil moisture conductivity, respectively, is the soil water tension function depending on saturated soil water potential ( S ), and is formulated as: Evaporation and transpiration play an important role in soil hydrology processes and are the sum of the direct evaporation from the ground surface (E dir ), evaporation of intercepted water (E c ), and vegetation transpiration (E t ), which depend on the potential evaporation (E p ), calculated by the Penman-based energy balance method (Mahrt and Ek, 1984). The three components of evaporation are formulated as: where f represents the vegetation fraction, W c and S are the intercepted canopy water content and the maximum intercepted water content from precipitation or condensation, respectively, B c is the function of canopy resistance: where C h is the surface exchange coefficient, Δ depends on the slope of the saturation specific humidity curve, R r is a function of surface air temperature, pressure, and C h , R c is the canopy resistance, which is affected by factors related to solar radiation (F 1 ), vapor pressure deficit (F 2 ), air temperature (F 3 ), and soil moisture (F 4 ): where R cim is the minimum stomatal resistance, i and d i represent soil layer and soil layer thickness.

Appendix C: Evaluation methods and significant test
Terrain difference between grid point and the observational site has impact on temperature, especially in complex terrain area. Some studies make a correction on temperature using constant lapse rate, but in fact, a simple temperature lapse rate is often difficult to describe the spatial temperature structure and lapse rate has seasonal and diurnal variations. Zhang et al. (2009) used temperature lapse rate from simulations to account for the terrain difference, but it cannot represent the real temperature lapse rate and thus could result in unpredictable effect on model evaluation. Which layer should be used in temperature correction is still an open question as large difference in temperature lapse rate exists between the surface and the upper layer. In this article, correction is made via observed temperature lapse rate in mountainous areas with diurnal variation. Large temperature lapse rate appears in daytime, while small temperature lapse rate appears in nighttime. No lapse rate was used to correct the humidity, wind speed, and precipitation, as it is very complex and there is no standard method to do so. Six statistical indices, i.e. the index of agreement (IOA), the correlation coefficient (R), the standard deviation (STD), the root mean square error (RMSE), the mean bias (MB), and the mean error (ME), are used for model evaluation, as shown in Equations (A16)-(A21): where F and O are the simulated and the observed values, respectively, F andŌ are the mean simulated and observed values, respectively, x represents F or O, x represents F orŌ, N is the number of samples. Apart from the statistical indices mentioned above, the relative variation of RMSE (RV RMSE ) is used in this study: where the subscripts 1 and 2 represent different model runs.
A comparison of statistical indices between different simulations provides information on the effect of soil texture and parameters on WRF simulations. It is very important to determine whether the difference of statistical indices is significant or not. Confidence bounds for statistical indices were calculated using Matlab ® and the overlap of confidence bounds indicates that the difference is not significant. For R, the confidence bounds are based on an asymptotic normal distribution of 0.5 × log((1 + R)/(1 − R)). These bounds are accurate for large samples when the variable has a multivariate normal distribution. For other statistical indices, the bootstrap test was used to determine the confidence bounds. The lower and upper bounds for a 95% confidence interval were provided

Supporting information
The following supporting information is available: Table S1. Soil hydraulic parameters.          Figure S8. RMSE of T 2 (a), Q 2 (b), and WS 10 (c) for the six sensitivity tests and the reference run (SIM3).