Performance Evaluation of Four Different Land Surface Models in WRF

2. 1 WRF Model Configuration and Initialization

WRF is a next-generation, fully compressible Euler non-hydrostatic mesoscale forecast model with a runtime hydrostatic option. Two cores, Advanced Research WRF (ARW) and Non-hydrostatic Mesoscale Model (NMM), are available in WRF. The WRF-ARW core, used in this study, is based on the Fifth-Generation Penn State/NCAR Mesoscale Model (MM5). The model uses terrain following hydrostatic pressure coordinate system with permitted vertical grid stretching (Laprise, 1992). The horizontal grid in WRF is Arakawa-C grid.

The microphysics scheme used in the simulation is WSM6 that is extension of previous WSM5 scheme. Detailed description of the WSM6 can be found in the work of Hong and Lim (2006). The RRTM longwave and Dudhia shortwave radiation schemes were used in the simulation. The slope inclination and shadow effect by terrain obstacles are considered in the Dudhia’s radiation scheme and used in this study. The vertical sub-grid-scale fluxes were solved using Yonsei University planetary boundary layer (YSU-PBL). More information about YSU-PBL can be found in work of Hong et al. (2006).

The meteorological initial and boundary conditions for WRF obtained from National Centers for Environmental Prediction (NCEP) are final analysis (FNL) data with 1.0°×1.0° horizontal resolution and 6 hour time interval. The mother domain (D01) used in the simulation has 27 km grid size. The domain is centered at 38°N, 126°E and covers the whole Korean Peninsula, Japan and part of East China (not shown). The horizontal resolutions of D02, D03 and D04 are 9 km, 3 km and 1 km, respectively. The innermost fine domain (D05) is centered at Haean Basin and consists of 54 columns and 57 rows of 0.3×0.3 km grid cells (Fig. 1c). The five domains interact with each other through a twoway nesting strategy. The vertical structure includes 36 layers. Topography and land-use data were interpolated from ASTER Global Digital Elevation Model (ASTER DEM) and Korea Ministry of Environment (KME) with an appropriate spatial resolution for each domain. The KME medium-category land-use classification was used to represent dominant vegetation types (Fig. 2b). Additionally, the KME land-use layer was updated by ground mapping data-sets from TERRECO campaign (Fig. 2c). The rest of regions out of South Korea were filled by United States Geological Survey (USGS) land-use data-set (Fig. 2a).

Fig. 1.

Design of model domains: (a) Domain D02-D05, (b) domain D04-D05 and (c) domain D05. The location of flux station (triangle) and tethered balloon experiment (cross) are pointed out in the middle of the basin.

Fig. 2.

Land-use layers used in WRF simulation. (a) USGS (b) previous KME (c) modified KME with TERRECO ground mapping data-sets.

2. 2 Land Surface Schemes

The 5-layer thermal diffusion LSM (TD LSM) is based on the 5-layer soil temperature model used in MM5 mesoscale model. The energy budget consists of radiation, sensible and latent heat flux. TD LSM does not predict soil moisture. The soil moisture is fixed with a land-use type and a season dependent constant value. The 5 layers are 1, 2, 4, 8 and 16 cm thick (Skamarock et al., 2008). The model does not have a snow scheme.

The Noah LSM is developed from OSU LSM, described by Chen and Dudhia (2001). The model includes diurnal dependent Penman potential evaporation (Mahrt and Pan, 1984), multilayer soil model of Mahrt and Pan (1984) and canopy model of Pan and Mahrt (1987). The LSM predicts soil moisture and temperature in 4 layers. The layer thickness is 10, 30, 60 and 100 cm from top to bottom. The model is able to predict canopy moisture, water-equivalent snow depth and soil ice (Skamarock et al., 2008).

The Rapid Update Cycle LSM (RUC LSM) includes multilevel soil model with 6 default levels. The number of levels can be increased. The model solves heat diffusion, Richards diffusive and gravitational motions for soil moisture transfer (Richards, 1931). In the cold season phase changes of soil moisture are considered (Smirnova et al., 1997). Special feature of RUC LSM is a thin layer spanning the ground surface that includes half of the first atmospheric layer and half of the top soil layer is used to solve the energy budgets. The heat storage in the layer is determined by contribution of soil fluxes and atmospheric fluxes (Smirnova et al., 1997). In RUC LSM, the latent heat flux is directly affected by vegetation through incorporation of free water from canopy and evapotranspiration. The vegetation processes are treated similarly to Noah LSM following concept developed by Pan and Mahrt (1987).

The Pleim and Xiu LSM is based on a set of five partial differential equations for soil temperature and soil moisture in two layers (1-cm surface layer and 1-m root zone layer) and canopy moisture (Pleim and Xiu, 1995).

2. 3 Observation

The research site was located in Haean Catchment, Kangwon province, South Korea. Haean is an intensively used agricultural region surrounded by forested mountains. The dominant crop types in Haean are rice paddy, radish, bean and potato.

Basic meteorological variables were observed at 2.5 m above ground level with an Automatic Weather Station (AWS, WS-GP1, Delta-T Devices Ltd., UK) located at a rice field (38°17ʹ27ʺ N, 128°07ʹ50ʺ E, 457 m a.s.l.). Measured variables include air temperature, wind speed, wind direction, relative humidity, precipitation, and global radiation.

Located close to the AWS an eddy-covariance measurement system (38°17ʹ28ʺ N, 128°07ʹ52ʺ E) was installed on a 2-m-high mast, to obtain the turbulent atmospheric fluxes of sensible and latent heat and carbon dioxide (CO₂). The system was equipped with an ultrasonic anemometer (USA-1, Meteorologische Messtechnik GmbH, Germany) and a fast-response openpath H₂O/CO₂ gas analyzer (LI-7500, LI-COR Inc., USA) both working at a sampling frequency of 20 Hz. The eddy-covariance software package TK3 (Mauder and Foken, 2011), developed by the Department of Micrometeorology, University of Bayreuth, post-processed the high-frequency raw data according to all international agreed procedures. As result correct 30-minute aggregated sensible and latent heat and CO₂ fluxes are available. The detailed calculating and correcting strategy of TK3 could be found in Mauder and Foken (2011) and Foken et al. (2012).

The period from Sep. 23 to 26 (DOY 266 to 269), 2010, was picked out as Golden Days for the validation of models. During this period it was either sunny or cloudy with radiation fog events in the morning. The rice paddies were in the late growing season before harvest with a canopy height of 0.88 m and a leaf area index of 0.68 m² m^-2. The field was irrigated with a water depth of 2 cm. For further information about the field observation, see Zhao et al. (2011).

A multi-step outlier check procedure (Zhao and Lüers, 2012) was applied on both meteorological variables and turbulence fluxes to ensure the high quality of data. as an enhanced quality control procedure. Briefly speaking, this procedure filters those data as outliers that fall beyond the physical thresholds, or have poor quality flags (Foken et al., 2004; Foken and Wichura, 1996), or occur during a bad instrument status, or fail in the statistical check. A quality-controlled data-set is used for the validation of the models.

Tethered balloon soundings were performed from the basin floor during clear undisturbed nights on 24^th and 25^th of September 2010 (Fig. 1). Since the Haean basin is very close to the demilitarized zone between North and South Korea any radio signal based data transmission was forbidden. Therefore, instead of common radio-sonde, a lightweight probe with internal memory and consisting of temperature/humidity sensor and pressure sensor was used.

3. 1 Temperature

The modeled near surface temperatures (at 2 m above surface) were compared with the measured in Fig. 3. All four models performed quite well with IA higher than 0.9. The best results were given by WRF-Noah with the highest IA (0.95) and lowest MB (0.29°K) (Table 1). Significantly underestimated temperature can be seen in the early morning on 24^th of September when the maximal temperature difference between the predicted and observed temperature reached 3.15°K. Similar temperature drop was produced by WRF-TD (Fig. 3a) and the maximal difference reached 2.8°K. Another significant deviation produced by WRF-Noah was on 24^th September at night. The model overestimated the temperature by 3.4 K. The low RMSE (0.089) indicates that both errors are obviously non-systematic and except these two deviations the WRF-Noah performed sufficiently. Other models tended to overestimate the temperature especially during the daytime periods. This is most evident in case of WRF-RUC. All models (including WRF-Noah) seem to be skeptic to nocturnal cooling during the first half of nights.

Fig. 3.

Comparison of near surface temperature results with observation.

Table 1.

Results of statistical evaluation of LSMs performance.

The potential temperature profiles were compared with tethered balloon soundings. The underestimated nocturnal cooling can be seen in the potential temperature profiles as well (Fig. 4). At 20:00, the most realistic simulation seems WRF-Noah. However, the profile does not change much between 22:00 and 02:00. An interesting finding was, the successful prediction of the fog layer formation using WRF-RUC which might be related to the overestimated evaporation expressed as latent heat energy flux in Fig. 6. The overestimated evaporation compensates the radiative cooling deficit (Fig. 3) and allows the saturation in lower model’s layers. The shading effect of fog layer is then well simulated and the profiles of potential temperatures are in closer agreement with the observation (Fig. 5). This can be very well seen in Fig. 4 at 9:00 when the WRFRUC follows the temperature inversion in the fog layer while other WRF-LSMs already predicted the well mixed layer.

Fig. 4.

Observed and modeled vertical profiles of potential temperatures on 24th-25th September 2010.

Fig. 5.

(a) Comparison of Observed and modeled hourly sensible heat fluxes. (b) relation between observed and modeled values for four WRF-LSMs.

Fig. 6.

(a) Comparison of Observed and modeled hourly latent heat fluxes. (b) relation between observed and modeled values for four WRF-LSMs.

3. 2 Surface Energy Fluxes

The sensible (SHF) and latent heat fluxes (LHF) are compared with measured fluxes in Fig. 5. and Fig. 6. The gaps of LHF, during night time are caused by the bad working condition when water droplets condensed on the light path of the gas analyzer during fog events.

The modeled heat fluxes by WRF-TD and WRF-PX are very similar. Both of them significantly overestimated SHF and underestimated LHF during the day time. The WRF-TD has the highest RMSE among all models for SHF (Table 1). Better agreement is found between predicted and measured SHF for WRF-RUC and WRF-Noah. The WRF-RUC underestimated the SHF during the day time, although smaller RMSE (28.19) was found compare to WRF-Noah (33.15). For LHF the WRF-Noah showed best results with slight overestimation. The RMSE of WRF-Noah is higher than WRF-TD, but small RMSEs indicates that this was caused by non-systematic error. On other hand WRFTD has greater RMSEs compare to RMSEu and underestimated the LHF continuously. The second best IA was given by WRF-RUC but it also produced the greatest RMSE (89.28) and positive MB (51.7 W m^-2). The overestimation is evident during the noon and the maximal LHF.