The parameterization of thermal roughness length z0h plays a key role in land surface modeling. Previous studies have found that the daytime land surface temperature (LST) on dry land (arid and semiarid regions) is commonly underestimated by land surface models (LSMs). This paper presents two improvements of Noah land surface modeling for China's dry-land areas. The first improvement is the replacement of the model's z0h scheme with a new one. A previous study has validated the revised Noah model at several dry-land stations, and this study tests the revised model's performance on a regional scale. Both the original Noah and the revised one are driven by the Global Land Data Assimilation System (GLDAS) forcing data. The comparison between the simulations and the daytime Moderate Resolution Imaging Spectroradiometer- (MODIS-) Aqua LST products indicates that the original LSM produces a mean bias in the early afternoon (around 1330, local solar time) of about −6 K, and this revision reduces the mean bias by 3 K. Second, the mean bias in early afternoon is further reduced by more than 2 K when a newly developed forcing data set for China (Institute of Tibetan Plateau Research, Chinese Academy of Sciences (ITPCAS) forcing data) is used to drive the revised model. A similar reduction is also found when the original Noah model is driven by the new data set. Finally, the original Noah model, when driven by the new forcing data, performs satisfactorily in reproducing the LST for forest, shrubland and cropland. It may be sensible to select the z0h scheme according to the vegetation type present on the land surface for practical applications of the Noah LSM.
 Arid and semiarid regions in China are mainly located in the north and the west and occupy more than half of the country. These regions are characterized by bare soil and grasslands. Arid and semiarid regions in China have experienced or are experiencing significant environmental changes, such as the enhanced drying trend in northern China [e.g., Wang and Zhai, 2003; Ma and Fu, 2006], enhanced warming signal [e.g., Chase et al., 2000; Ren et al., 2005], grassland degradation and desert extension [e.g., Fu and Wen, 2002], discharge decrease in the Yellow River [Yang et al., 2004; Tang et al., 2008], and drought-area expansion in principal farming areas [Wang and Zhai, 2003]. Understanding the interactions between land and atmosphere and their response to climate change will help people to confront the severe environment problems there. For instance, numerical modeling by Xue  suggested that the desertification in the Mongolian and the Inner Mongolian grasslands would reduce areal rainfall through modification of the surface energy balance. The physical representation of land-atmosphere interactions plays a key role in improving the predictability of weather and climate [Pielke et al., 1999; Chen et al., 2001; Koster et al., 2004; Los et al., 2006]. However, issues concerning the exchange efficiencies of energy and water vapor between the land surface and the atmosphere still remain poorly understood [Chen and Zhang, 2009].
 In arid and semiarid regions, surface upward longwave radiation plays an important role in the surface radiation budget; sensible heat flux (H) and ground heat flux (G0) dominate the surface energy budget because of the lack of water replenishments. It is crucial to simulate the land surface temperature (LST), because it modulates both the radiation budget by determining upward longwave radiation (L↑) and the surface energy budget (SEB) by changing ground-air and soil temperature gradients. However, daytime LSTs in arid and semiarid regions are typically not well simulated in current land surface models (LSMs): Hogue et al.  found that the Noah model tends to underestimate the LST and overestimate H during the daytime at semiarid sites; LeMone et al.  also found this phenomenon and pointed out that modifying the thermal roughness length (z0h) scheme is helpful to reproduce the observed LST and H; Chen and Zhang  further examined the observed surface exchange coefficient for heat (Ch) over various surface vegetation types and suggested the coefficient in the z0h scheme should depend on vegetation type; Yang et al. , using data archived in the Coordinated Enhanced Observing Period (CEOP) program [Koike, 2004], pointed out that several operational global circulation models (GCMs) systematically underestimate the diurnal range of surface-air temperature differences, particularly in arid and semiarid regions, since the heat transfer resistances are underpredicted. Yang et al.  further evaluated three off-line LSMs (i.e., Common Land Model, Simple Biosphere Model, version 2, and Noah) and confirmed that the three models significantly underestimate the daytime LST because of the underestimation of heat transfer resistances under dry conditions. Trigo and Viterbo  argued that an underestimation of the diurnal cycle of modeled brightness temperature in clear-sky conditions during daytime can be attributed to the problems in the model surface-to-boundary-layer coupling. These results further argue that a proper representation of the z0h or Ch for dry land is crucial in modeling the LST as well as H. Much of the literature has focused on the parameterization of z0h [e.g., Sheppard, 1958; Brutsaert, 1982; Betts and Beljaars, 1993; Zilitinkevich, 1995; Malhi, 1996; Zeng and Dickinson, 1998; Kanda et al., 2007]. For the evaluation of some of these schemes against observations, the readers are referred to the work by Yang et al. .
 On the basis of the work of Chen et al. , the goal of the present study is to extend our modeling work over dry-land surface from the local point scale to a regional scale. More specifically, we intend to improve the performance of the Noah model for the arid and semiarid regions in China.
 In addition to a realistic representation of land processes in a model, the LST modeling in China encounters three additional challenges. The first is the shortage of the forcing data to drive the model. The China Meteorological Administration (CMA) has not released its own reanalysis data set that is suitable to drive LSMs at the regional scale, whereas available global data sets (e.g., National Centers for Environmental Prediction (NCEP) reanalysis) have evident biases in China [e.g., Ma et al., 2008, 2009]. Wang and Zeng  pointed out that using a reanalysis-based data set developed by Sheffield et al.  tends to produce too-wet soil in northeastern China and too-dry soil in northwestern China, but these deficiencies can be significantly reduced after the in situ measured precipitation data are introduced into the forcing data. In this study, we use two forcing data sets to drive the LSM. They are the Global Land Data Assimilation System (GLDAS) forcing data set and a newly developed forcing data set (hereinafter ITPCAS (for Institute of Tibetan Plateau Research, Chinese Academy of Sciences) forcing data), whose descriptions will be presented in subsection 3.1.2.
 The second challenge is how to validate the gridbox-based simulation results on a regional scale. Currently, reliable observed LST and turbulent flux data are available at quite a few individual sites. In order to overcome this problem, the Moderate Resolution Imaging Spectroradiometer (MODIS) level 3 0.05° LST products are used to evaluate the model performance.
 The third challenge is the evaluation method. In the point-scale cases, we can directly measure many variables, such as soil moisture and soil temperature and set the model parameter values to the observed ones, so that the evaluation can be conducted by strictly comparing the simulated variables against their respective observations. However, it is impossible to do this on a regional scale, and hence we evaluate the model performance by computing the error metrics over all grid cells in the entire study region rather than doing this pixel by pixel.
 The rest of the paper is as follows: Section 2 briefly introduces the Noah model and the new z0h scheme; section 3 introduces the forcing data, validation data, model parameter data sets, and the numerical experiments' design. The simulations are evaluated against the daytime MODIS/Aqua LST products in section 4; section 5 discusses the evaluation against MODIS/Terra LST products, the choice of z0h scheme in practical applications, and the comparison between GDLAS and ITPCAS radiation fluxes. The conclusions are given in section 6.
 Noah was developed based on the Oregon State University (OSU) LSM, which includes a two-layer soil model with thermal conduction equations for soil temperature and a diffusive form of Richardson's equation for soil moisture. The Noah model and its predecessors continue to benefit from a series of improvements, especially after being adopted by NCEP and AFWA [e.g., Chen et al., 1996; Koren et al., 1999; Ek et al., 2003]. A more detailed overview of the physics lineage is presented by Ek et al. . The present study employs the Noah version 2.7.1, and it has four soil layers (with depths of 10, 30, 60, and 100 cm from top to bottom), a single-canopy layer, and a single snow layer; please refer to detailed model physics by Chen and Dudhia .
 Concerning the energy exchanges between land surface and atmosphere in arid and semiarid regions, the SEB equation can be written as
In the radiation budget equation (1a), Rnet is the net radiation, S↓ and L↓ are the downward shortwave and longwave radiation, respectively, Tsfc is the land surface temperature (i.e., LST), σ is the Stefan-Boltzmann constant, and α and ɛ are the surface albedo and the ground surface emissivity, respectively. In the energy budget balance, equation (1b), H is the turbulent sensible heat flux, LE is the turbulent latent heat flux, and G0 is the surface soil heat flux. In dry conditions, LE accounts for a minor proportion, and H and G0 are thus the dominant terms on the right-hand side of equation (1b).
H is calculated through the bulk heat transfer equation:
where ρ is the air density, cp is the specific heat capacity of air at constant pressure, u is the wind speed, θair is the air temperature adjusted adiabatically for the height above the surface, and θsfc is the counterpart at the surface.
 The surface soil heat flux is written as
where kT is the soil thermal conductivity that is a function of soil water content (Θ) and soil properties, T1 is the soil temperature in the uppermost layer, and h1 equals half of the first layer depth. In dry conditions, Θ is very small with negligible temporal variations for most of the time. Therefore, the value of kT can be assumed as a constant value at a specific site. As a result, G0 depends mainly on the modeled Tsfc.
2.2. Improvement of z0h Scheme
Ch plays an important role in calculating both LST and H in arid and semiarid regions, which can be obtained through the Monin-Obukhov similarity theory and depends on z0h and the aerodynamic roughness length (i.e., z0m). The z0m often varies quite little with time, and it is typically prescribed according to the vegetation height in large-scale modeling. Then, z0h becomes the determining factor in the calculations of Ch and thus for the LST and H, so that it needs a proper parameterization in LSMs. It is worth noting that the parameterization of z0h plays a more important role in the daytime LST simulation than in the nighttime one. In the daytime, the excess resistance, because of the difference between z0m and z0h, is comparable to the aerodynamic resistance, but the excess resistance in the nighttime is much less than the total resistance because of the stability increase [Yang et al., 2008]. Furthermore, the soil heat flux in the nighttime is a major controller of the SEB, and thus the simulation accuracy of the surface temperature highly depends on the accuracy of soil thermal inertia in addition to the heat transfer resistance. Therefore, we put emphasis on the simulation of LST and H in the daytime.
Chen et al.  conducted a sensitivity study using several widely used z0h schemes in the framework of the Noah model. Their results showed that the simulated LST and SEB is very sensitive to these schemes. For instance, the simulated maximum LST and H using currently available z0h schemes can differ by more than 20 K and 200 W m−2, respectively.
 In the original Noah model, the z0h scheme developed by Zilitinkevich  takes the following form:
where k is the von Kármán constant and normally equals 0.4, Re* is the roughness Reynolds number (defined as Re* = z0mu*/ν, where u* is the friction velocity and ν is the fluid kinematical viscosity), and Czil is an empirical coefficient specified as 0.1 by Chen et al. . After comparing the Noah-simulated Ch against the observed values based on a total of 12 flux data sets, Chen and Zhang  pointed out that Ch tends to be overestimated for short vegetation and underestimated for tall vegetation by using Czil = 0.1 in Noah. Moreover, they suggested that the value of Czil should be vegetation dependent, and they related this value to the canopy height through an empirical regression method. However, such a relationship needs further assessment over different vegetation types. In both Noah version 2.7.1 and the GLDAS, Czil = 0.075 is prescribed, and this default value is used in this study. A discussion about the choice of the Czil value for dry-land surfaces is given in subsection 5.2.
 In the revised Noah model, another scheme was applied to replace the z0h scheme by Chen et al. . This new scheme was developed and validated by Yang et al. [2002, 2008]. They used a turbulence-related height (hT) to parameterize z0h. This height separates the fully turbulent layer and the transitional layer and can be determined by the critical Reynolds number (Recrit):
where Recrit is equal to 70.
 The ratio of hT to z0h is parameterized with both the frictional velocity and a temperature scale (θ*, ≡ H/(ρcpu*)) and the final form of z0h is
Yang et al.  evaluated several widely used schemes against turbulent flux data collected at seven dry-land stations. They indicated that a common feature of z0h over dry lands is its diurnal variation, which has been reported in several earlier studies over bare soil and grassland [Verhoef et al., 1997; Sun, 1999; Ma et al., 2002; Yang et al., 2003]; such a diurnal variation cannot be reproduced by commonly used parameterization schemes, but the scheme of Yang et al. manages to reproduce it, largely because of its ability to capture the diurnal course of θ*. This scheme also performs well for the flux parameterization over a glacier surface [Guo et al., 2011]. After this scheme was implemented into the SiB2 model, the biases in the simulated daytime LST were much reduced at two alpine desert sites [Yang et al., 2009]. Furthermore, Chen et al.  validated this scheme in the framework of the Noah model at several dry-land sites, confirming its advantages over several other schemes. In this paper, we examine the performance of the revised Noah LSM with this scheme over the arid and semiarid areas of China.
3. Data and Methods
3.1. Forcing Data
3.1.1. GLDAS Forcing Data
 The 3 hourly, 0.25° × 0.25° forcing data from GLDAS [Rodell et al., 2004] version 1, which is a combination of several data sets, is used in this paper. The near-surface wind speed, air temperature, specific humidity, and surface pressure are provided by the National Oceanic and Atmospheric Administration (NOAA) Global Data Assimilation System (GDAS) atmospheric reanalysis data, which is the operational global atmospheric data assimilation system of NCEP. The precipitation data are derived through the spatial and temporal disaggregation of NOAA Climate Prediction Center Merged Analysis of Precipitation (CMAP) fields, which merged the satellite (infrared and microwave) and gauge observations. The AFWA Cloud Depiction and Forecast System II (CDFSII) data are used to calculate downward radiation fluxes at the ground surface.
3.1.2. ITPCAS Forcing Data
 For comparison, a new forcing data set for China is also used to drive the Noah model. The new data set was developed by the hydrometeorological research group at the Institute of Tibetan Plateau Research, Chinese Academy of Sciences (hereafter ITPCAS) [He, 2010]. The temporal and spatial resolutions of the ITPCAS forcing data are 3 hourly and 0.25° × 0.25°, respectively. The observations collected at 740 operational stations of the CMA are merged into the corresponding Princeton meteorological forcing data [Sheffield et al., 2006] to produce near-surface air temperature, pressure, wind speed, and specific humidity. The height of air temperature and specific humidity is 1.5 m above ground; the height of wind speed is 10 m above ground. Three precipitation data sets are combined to produce the precipitation field, including the Tropical Rainfall Measuring Mission (TRMM) 3B42 precipitation products [Huffman et al., 2007], precipitation observations from 740 operational stations, and the Asian Precipitation – Highly Resolution Observational Data Integration Toward Evaluation of the Water Resources (APHRODITE) precipitation data [Yatagai et al., 2009]. The downward shortwave radiation was derived by correcting the Global Energy and Water Cycle Experiment – Surface Radiation Budget (GEWEX-SRB) [Pinker and Laszlo, 1992] shortwave radiation data set with reference to radiation estimates from CMA station data using a hybrid radiation model [Yang et al., 2006]. The downward longwave radiation is calculated by the model of Crawford and Duchon  based on the produced near-surface air temperature, pressure, specific humidity, and downward shortwave radiation.
 The ITPCAS forcing data are still being updated and evaluated. Several independent data sets, including a quality-controlled downward shortwave radiation data set at 95 CMA radiation stations, the high-resolution data collected through GEWEX Asia Monsoon Experiment-Tibet (GAME-Tibet) and the data collected through the CEOP Asia-Australia Monsoon Project-Tibet (CAMP-Tibet) have been used to evaluate the forcing data [He, 2010]. In subsection 5.3 of this paper, both the GLDAS and ITPCAS radiation fluxes are directly compared against field radiation measurements.
3.2. Validation Data
 In this paper, the retrieved MODIS LST products of version 5 in 2003 are used as the ground “truth” to examine the performance of the revised Noah. There are many types of MODIS LST products with different spatial and temporal resolutions. Here, the level 3 0.05° gridded products (MOD11C1 and MYD11C1) recommended by the algorithm development team are taken, since the algorithm used to generate these two data sets is the physically based day-night method and has an advantage over the split-window method for the 1 km LST products [Wan et al., 2002].
 The MOD11C1 product is produced for the Terra satellite, whose overpass times are around 10:30 A.M. (local solar time) in descending mode and 10:30 P.M. (local solar time) in ascending mode. The MYD11C1 LST product is produced for the Aqua satellite, whose overpass times are around 1:30 P.M. (local solar time) in ascending mode and around 1:30 A.M. (local solar time) in descending mode. Because the MYD11C1 and MOD11C1 LST products are derived from sensors aboard different satellites, these two data sets are used separately to evaluate the performance of the revised Noah and the original one. In this paper, the daytime MODIS/Aqua MYD11C1 LST product is adopted as the primary validation data, since its time (around 1:30 P.M., local solar time) corresponds better to the time when relatively high LST values occur in a diurnal course. Details concerning the usage of MODIS/Terra MOD11C1 LST product will be discussed in subsection 5.1.
 In order to make a reliable evaluation of the simulated areal LST, MODIS LST values are used only if their quality-control (QC) flags are labeled as 00, indicating that the mean error of these LST values is less than 1 K and the data should have a good quality (refer to MODIS LST User's Guide for details). Moreover, a two-step procedure is used to make a matched comparison between the MODIS LST and the simulated one. First, the fine-resolution (0.05°) MODIS LST values are aggregated to the coarse-resolution (0.25°) ones corresponding to the grid cells on which the simulation is conducted. It is obvious that each coarse grid cell contains 25 fine-resolution MODIS LST pixels. The aggregation step is performed only when the number of fine-resolution LSTs with high-quality values in each coarse grid exceeds 21. Second, given the fact that the simulation time never precisely matches the MODIS overpass time, the simulated LST values cannot be directly applicable; in this paper, we derive the LST value at the desired overpass time simply from the two most adjacent simulated values (based on a linear interpolation).
3.3. Soil, Vegetation, and Other Input Data Sets
 The vegetation and soil as well as other parameter data sets used in this study are exactly the same as for the settings in producing the GLDAS 0.25° product of version 1 (i.e., GLDAS_NOAH025SUBP_3H). In other words, the vegetation type map is resampled from a static, 1 km resolution, land cover class that was produced by the University of Maryland (UMD) based on observations from the Advanced Very High Resolution Radiometer (AVHRR) aboard the NOAA-15 satellite. The monthly greenness vegetation fraction climatology data set is derived based on the 5 year (1985–1990) AVHRR Normalized Difference Vegetation Index (NDVI). The percentages of sand, silt, and clay components were horizontally resampled to the 0.25° grid based on the 5 s resolution global soils data set developed by Reynolds et al. . The NCEP quarterly albedo climatology and maximum snow albedo are both produced based on AVHRR data. The other input parameter data set is the land-sea mask produced by the UMD. These model parameter data sets are available from the NASA Land Information System website and GLDAS website.
3.4. Design of Numerical Experiment
 The total land-cover types, on which the simulation experiments are made, add up to 14 according to the UMD vegetation map. To facilitate our evaluation, these vegetation types are further categorized into five broader classes corresponding to primary climatic regions in China (Figure 1), including bare soil, crop, grass, shrub, and forest. This paper is concerned with the performance of the Noah model in arid and semiarid regions. As shown in Figure 1, the arid regions, whose land covers mainly include desert in the northwest and alpine desert in the west Tibetan Plateau, are labeled as bare soil and the semiarid regions are labeled as grassland, including grassland, alpine steppe, alpine meadow, and sparse shrubland in various subclimatic regions.
 We conducted the land surface modeling experiments from January 2000 to December 2003. The temporal and spatial resolutions are 0.5 hourly and 0.25°, respectively. Four simulation experiments with different configurations of the models (the original LSM and the revised one) and the forcing data sets (the GLDAS and the ITPCAS) are designed to test the model performance. The aims of these experiments are twofold. The first is to check whether improvements are achieved by the revised Noah (subsection 4.1). The second is to investigate whether the ITPCAS forcing data can improve the model performance (subsection 4.2).
4.1. Modeling Improvement by z0h Scheme Replacement
 In order to determine the effect of using the different z0h schemes as introduced in section 2.2, the widely used GLDAS forcing data are first applied to drive both the original Noah and the revised one. After QC and aggregation, a total of 5171 coarse-resolution daytime (around 1:30, local solar time) Aqua LST values are obtained in 2003 over the bare soil land-cover type in China. Figures 2a and 2c present the scatterplots between the simulated LST values and the corresponding aggregated LST values, and Figures 2b and 2d show the histograms for their differences. Figure 2b shows that the histogram for the original Noah evidently deviates from the zero point and shifts to the negative direction of the axis, indicating negative mean biases. Similarly, Figure 2d shows that the revised model also underestimates the LST, but the magnitude is less than that derived from the original model. The statistical metrics are listed in Table 1. The mean bias and root-mean-square error (RMSE) for the original Noah are −6.6 K and 9.0 K, respectively, whereas the revised model reduces the mean bias by 3 K and the mean RMSE by about 1 K. Nevertheless, the simulation errors using the revised model are still large and R2 values can even become lower.
Table 1. Determination Coefficient (R2), Mean Bias (BIAS), and Root Mean Square Error (RMSE) for Simulated LST Against Daytime (Around 1:30, Local Solar Time) MODIS/Aqua MYD11C1 LST Productsa
The sample number for bare soil is 5171, and the sample number for grassland is 16,852. BIAS and RMSE are defined as BIAS = , RMSE = , where N is the sample number.
Figure 3 shows similar plots, but for grassland. A total of 16,852 aggregated coarse-resolution Aqua LST values over this land-cover type are applicable. The results are very similar to those over bare soil. As indicated in Table 1, the mean bias and the mean RMSE are reduced by 3.4 K and 1 K, respectively.
 The above assessments argue that the revised Noah can evidently reduce the negative biases in the simulated LST values and the performance of the z0h scheme is critical for improving the modeling ability. However, the remaining errors are still large. We note the fact that there is always an abnormal “branch,” denoted by the gray circle in the scatterplots, implying that factors other than the choice of z0h scheme may cause this phenomenon to interfere with the evaluation. The forcing data are the main suspects, as will be seen in the next subsection.
4.2. Modeling Improvement by ITPCAS Forcing Data
 It is well known that inaccuracies existing in the forcing data may have substantial impacts on the simulation of land surface energy partitioning. To identify potential inaccuracies, a numerical experiment is conducted. The GLDAS and ITPCAS forcing data are both used to drive the revised Noah model, and the results are compared against the aggregated daytime (around 1:30, local solar time) Aqua LST.
Figure 4 indicates that the use of ITPCAS forcing data can improve the modeling accuracy of the revised Noah against the Aqua aggregated LST on both bare-soil and grassland land-cover types. The mean bias values are notably decreased by 2.6 K on bare soil and by 2.8 K on grassland when the ITPCAS forcing data are used. The other statistical metrics (i.e., R2 and RMSE) are also improved, as shown in Table 1. The shapes for the error distribution in Figures 4b and 4d become sharper and narrower than in Figures 2d and 3d; the abnormal branches in the scatterplots also vanish, which will be further discussed in subsection 5.3. It is also found that the biases decrease when the original Noah model is driven by the ITPCAS forcing data (not shown).
 In general, the modeling of LST is much improved for dry land when the ITPCAS forcing data are used, suggesting that the uncertainties in the ITPCAS forcing data are smaller than those in the GLDAS forcing data in China. This is not surprising because the ITPCAS forcing data benefit from merging the information from 740 CMA operational stations, which are not used in GLDAS forcing data. Their difference will be further discussed in subsection 5.3.
4.3. Impacts on Surface Energy Budget Modeling
 The modeling of each component in the SEB is also sensitive to the choice of the z0h scheme in the Noah model. Chen et al.  showed that the revised Noah model not only improves the LST simulation but also improves the accuracy of SEB components at several flux sites. It is very difficult to examine the gridded simulation results against field measurements because of the lack of such data. However, some hints about the improvements of simulation of the energy components may be obtained by comparing the results from the original Noah and those from the revised model when driven by the ITPCAS forcing data.
Figure 5 shows the corresponding histograms for the differences between the early afternoon Rnet, H, LE, and G0 simulated by the original Noah and those simulated by the revised Noah on bare-soil surfaces (Figures 5a–5d) and on grassland (Figures 5e–5h). The ITPCAS forcing data are used to drive both simulations. As shown in Figure 5, the original Noah produces averages of Rnet and H that are higher by about 24 W m−2 and 45 W m−2, respectively, compared with the revised Noah over bare-soil surfaces. The LE, which accounts for a minor proportion in dry conditions, is also slightly larger by about 4 W m−2. As a result, the G0 is lower by about 27 W m−2 on average for the original Noah. The results for grassland are similar.
 The differences are consistent with the underestimation of LST by the original Noah, which leads to lower upward longwave radiation (thus overestimated Rnet) and a lower soil temperature gradient (thus underestimated G0). All these can be attributed to the overestimation of daytime Ch by the original Noah. The overestimated Ch represents the magnified coupling strength between the land surface and the atmosphere, which will pump more sensible heat flux (overestimated H) and latent heat flux (overestimated LE, because of using the same Ch) to heat the atmosphere, and consequently the land surface is cooling down (underestimated LST).
5.1. Results by Using MODIS/Terra MOD11C1 LST Data
 In this subsection, the daytime (around 10:30 A.M., local solar time) MODIS/Terra MOD11C1 LST product is also applied to identify the modeling improvements by replacing the original z0h scheme as well as using the new forcing data. A total of 26,174 aggregated coarse-resolution Terra LST values in 2003 are qualified on bare-soil surfaces, and the number for grasslands is 70,887. The statistical metrics for both bare soil and grassland are listed in Table 2. As shown, the mean bias for the original Noah forced by GLDAS data is about −3.0 K on bare soil surfaces. This value is reduced by about 2 K by the revised Noah and is further reduced by 0.8 K after ITPCAS forcing data are used. Similar results are obtained for grassland.
Table 2. Similar to Table 1 but Against the Daytime (Around 10:30, Local Solar Time) MODIS/Terra MOD11C1 LST Producta
The sample number for bare soil is 26,174, and the sample number for grassland is 70,887.
 It is noted that the biases for the simulated LST values against the daytime (around 1:30 P.M., local solar time) MODIS/Aqua MYD11C1 LST are larger than those against the daytime (around 10:30 A.M., local solar time) MODIS/Terra MOD11C1 LST. This may be due to the fact that the Aqua ascending overpass time is closer to the moment when high LSTs appear during a day, which tend to be underestimated most notably by LSMs in dry-land conditions [e.g., Yang et al., 2009; Chen et al., 2010].
5.2. The z0h Scheme Choice for Practical Application
 The scheme of Yang et al. was initially developed based on the data sets from three alpine steppe sites [Yang et al., 2002] and evaluated against the data sets from seven dry-land sites [Yang et al., 2008]. Then, it was implemented into LSMs and was proved to be effective in reproducing the surface fluxes as well as LSTs over various ground types, including alpine steppe, grassland, and deserts [Yang et al., 2009; Chen et al., 2010]. This study further confirms its good performance on the regional scale. On the basis of these strict assessments, the authors suggest that this scheme can be applied in the parameterization of the surface fluxes over bare soils and sparsely vegetated surfaces.
 In order to thoroughly examine the new scheme, we also applied it over vegetated surfaces. Against the MODIS/Aqua MYD11C1 LST product, the revised Noah forced by ITPCAS data produces mean biases of 3.4, 5.4, and 2.3 K for forest, cropland, and shrubland, respectively (not shown), while those forced by the original Noah are 0.4, 1.1, and −0.4 K, respectively (see Table 3). The poor performance of the revised Noah model for high vegetation surfaces is expected, because the scheme of Yang et al. is essentially based on an analogy between natural flow and that over a smooth flat plate [Yang et al., 2002]. Therefore, the applicability of the revised Noah model should not be extended to other surfaces with high vegetation.
Table 3. Statistical Metrics for the Simulated LST by the Original Noah Against the Daytime MOD11C1 (or MYD11C1) LST for Forest, Shrubland, and Croplanda
The sample numbers for forest, shrubland, and cropland are 25,259, 9917, and 20,528, respectively, when compared with MOD11C1 LST products, and 6482, 2687, and 4801, respectively, when compared with MYD11C1 LST products.
 The performance of the original scheme [Zilitinkevich, 1995] is also examined when Czil is set to different values. We evaluated the choice of Czil at two experimental dry-land sites where reliable forcing data are available. The two sites are Audubon and Shiquanhe. The former is located in Arizona, and its surface is characterized by brown sparse grass during the simulation period. The data were collected through the AmeriFlux network (see http://public.ornl.gov/ameriflux/). The latter is located in the western Tibetan Plateau with an elevation of 4279 m above sea level, and its surface is almost bare soil. The measurements were collected through the GEWEX Asian Monsoon Experiment-Tibet (GAME-Tibet) [Koike et al., 1999]. For the detailed introduction of the two sites and the model parameter settings, the readers are referred to the work by Chen et al. . Czil is set to different values varying from 0.01 to 3.0 with an increment of 0.01. For each value, the simulations are conducted and the mean bias and RMSE of simulated Tsfc, Rnet, and H are calculated, as shown in Figure 6. It is indicated that when Czil increases to a certain point in its interval, with the value 0.075 as the reference (the default value of Czil in the Noah model), the simulation of Tsfc, Rnet, and H can be improved. However, it is not straightforward to find a robust and unique Czil value in order to well simulate all three variables at the two sites (see Figure 6). This reveals the practical deficiency of the original scheme over dry-land conditions, especially for regional scale modeling.
 On the other hand, two simulation experiments are designed to examine the performance of the original Noah over forest, shrubland, and cropland, in response to the GLDAS forcing data and the ITPCAS forcing data. Table 3 lists the statistical metrics for the simulated LST values over forest, shrubland, and cropland against both MOD11C1 and MYD11C1 LSTs, indicating that the original Noah gives reasonable R2, bias, and RMSE by using the ITPCAS forcing data. This result further adds weight to the previous analysis in subsection 4.2, suggesting that the original Noah model, driven by the ITPCAS forcing data, can reasonably simulate LST for shrubland, cropland, and forest.
 In a word, this paper shows that Zilitinkevich's scheme with Czil = 0.075 performs satisfactorily over surfaces covered by forest, shrub, and crop, whereas the Noah LSM performs considerably better with the scheme of Yang et al. than without it for bare-soil and short-vegetation surfaces. Therefore, it may be sensible to select the z0h scheme according to the vegetation type present on the land surface in the Noah model.
5.3. Comparison Between GLDAS and ITPCAS Radiation Forcing Data
 As mentioned in subsection 4.2, there are abnormal branches in Figures 2 and 3, which means that huge differences exist between the simulations and the observations. We found 175 bare-soil grids and 335 grassland grids that have large biases (>25 K), including the abnormal branches, and all these grids were located in the northwest part of the Tibetan Plateau, as shown in Figure 7. In this area, the elevation is very high and the meteorological stations are very sparse. By using the ITPCAS forcing data, the numbers of grids with such large biases are reduced to 6 and 16, respectively (not shown). This may suggest that the GLDAS forcing data have more severe deficiencies in the northwest part of the Tibetan Plateau than in other places, and these are reduced by merging the observations from sparse meteorological stations in the ITPCAS data set.
 Owing to the key role of incoming radiation fluxes for modeling surface energy partitioning, the quality of the daily averaged radiation fluxes of GLDAS and ITPCAS forcing data is evaluated against that of measured fluxes at four typical dry-land sites: Amdo, Gaize, Tongyu, and Dunhuang stations. Among them, Amdo, Gaize, and Tongyu are CEOP reference sites. Dunhuang station is maintained through the Field Experiment on Interaction between Land and Atmosphere in Arid Region of Northwest China (NWC-ALIEX) [Zhang et al., 2005].
 The Amdo station is on an alpine meadow, located in central Tibet where the climate is much affected by the Asian monsoon. The Gaize station is an alpine desert site located in western Tibet, where the climate is mainly controlled by the midlatitude westerlies. The Dunhuang station is located in the Gobi Desert in northwest China, where the ground is covered by pebbles and brown sands. The climate there is very dry, with a mean annual precipitation of about 39 mm. The Tongyu station is located in northeast China, with an annual mean amount of precipitation around 400 mm. The coverage of grass is about 60% in summer and less than 40% in the dry period [Liu et al., 2006]. The available shortwave radiation and longwave radiation data in 2003 at the Amdo, Gaize, and Tongyu sites are used for assessment. At the same time, the available data for 2004 are used at the Dunhuang site.
Table 4 gives the statistical metrics for daily averaged GLDAS and ITPCAS radiation fluxes against the observed data. It shows that the daily averaged GLDAS shortwave radiation has been clearly overestimated at the four sites, whereas the biases (BIAS) for ITPCAS shortwave radiation are dramatically decreased by about 22, 12, 63, and 26 W m−2 at Amdo, Gaize, Dunhuang, and Tongyu, respectively. The RMSE and R2 are also much improved for ITPCAS shortwave radiation. The statistical metrics for daily averaged ITPCAS longwave radiation also show some improvements, as shown in Table 4. These results suggest that the errors in ITPCAS radiation fluxes are obviously less than those in GLDAS data over dry lands of China. Therefore, the modeling, when driven by the ITPCAS forcing data, produces better results over these regions.
Table 4. The Statistical Metrics for Daily Averaged Downward Shortwave Radiation (S↓) and Downward Longwave Radiation (L↓) of GLDAS and ITPCAS Forcing Data Against the Measured Fluxes at Four Dry-Land Sites
BIAS (W m−2)
RMSE (W m−2)
 In this study, simulations of land surface temperature (LST) are improved over China's dry-land areas in two ways. One is replacing the original z0h scheme in the Noah LSM with the scheme of Yang et al. [Yang et al., 2008]. The second is driving the model using the newly developed ITPCAS forcing data set. The impact is shown by evaluating the results against the MODIS LST products (MYD11C1 and MOD11C1). The following findings and recommendations are presented.
 First, the revised Noah model can significantly reduce the mean bias of the simulated daytime LST on China's dry land. Previous studies have found that LSMs often underestimate the daytime LST on dry land. At present, both the original Noah and the revised one are driven by the widely used Global Land Data Assimilation System (GLDAS) forcing data. The evaluation indicates that the original Noah produces a mean bias of about −6 K in the early afternoon, while the revised one reduces the mean bias by 3 K, when compared with the daytime (around 1:30, local solar time) MODIS/Aqua LST products.
 Second, the usage of the ITPCAS forcing data can further reduce the mean bias produced by the revised Noah by more than 2 K against the daytime (around 1:30, local solar time) MODIS/Aqua LST product. Because the ITPCAS forcing data benefits from merging information of 740 CMA operational stations, it is more accurate than the GLDAS forcing data, as illustrated by the comparison of the radiation data.
 Third, the modeling of the surface energy budget may be also improved by applying the revised Noah and the ITPCAS forcing data on dry land. The original Noah model commonly underestimates the daytime LST, which leads to overestimated net radiation, underestimated ground soil heat flux, and overestimated sensible heat flux, whereas the revised Noah improves the modeling of LST and the surface energy budget.
 Finally, it is recommended that the z0h scheme should be selected according to the vegetation type present on the land surface in the Noah LSM. The original Noah performs satisfactorily over surfaces covered by forest, shrubland, and cropland, while the scheme of Yang et al. can be adopted in the parameterization of the surface heat fluxes over dry land. Another feasible way to address this issue is to combine the schemes for bare soils and vegetated surfaces by taking their areal fractions into consideration [Su et al., 2001; Zeng et al., 2005]. For a dual-source LSM that separately considers heat transfer from canopy and under-canopy, the new scheme may be applied to both bare soils and under-canopy heat transfer.
 This work was jointly supported by Open Fund (ZW9KFJJ017) from the State Key Laboratory of Remote Sensing Science, Global Change Program of Ministry of Science and Technology of China (2010CB951703), Science Fund from Institute of Desert Meteorology (Sqj2010003), and NSFC grant (40875009). GLDAS forcing data and model parameter data sets used in this study were acquired as part of the mission of NASA's Earth Science Division and archived and distributed by the Goddard Earth Sciences (GES) Data and Information Services Center (DISC).