Performance of HIRLAM in a Semiarid Heterogeneous Region: Evaluation of the Land Surface and Boundary Layer Description Using EFEDA Observations

1. Introduction

The parameterization of land surface and atmospheric boundary layer (ABL) processes has made considerable progress during the last decade. Advanced surface schemes have recently been incorporated in the major numerical weather prediction (NWP) models. A series of comparison and validation studies (Betts et al. 1993, 1998, 2001; Holtslag et al. 1995; Holtslag and Ek 1996; Yucel et al. 1998; Bringfelt et al. 1999; and others) have contributed significantly to this progress. These previous comparisons have all one feature in common in that the observational sites are located in fairly homogeneous areas, with only small variations in topography and gross vegetation type. All but the Arizona site of Yucel et al. (1998) (who do not provide a detailed study of boundary layer development) and the Sahelian site of van den Hurk et al. (2000) (who perform an offline validation of the land surface scheme) are furthermore located in semihumid or humid climate zones. Yet they all identify problems related to the transport of moisture. An interesting question is, therefore, how the models behave in arid or semiarid climate zones.

It is the objective of this work to evaluate the surface and ABL performance of the NWP High-Resolution Limited Area Model (HIRLAM, version 5.2; Unden et al. 2002) for a semiarid environment with heterogeneous land use. A consolidated dataset from the European Field Experiment in a Desertification-threatened Area (EFEDA; Bolle et al. 1993) is used for this purpose. The EFEDA area is located at an elevation of 700 m and is surrounded by mountains at distances of several tens of kilometers. The surface evapotranspiration is inhomogeneous on scales of several tens of kilometers because of partly irrigated land use. EFEDA is one of a series of land surface experiments that were conducted in a wide range of ecosystem and climate zones (Jochum et al. 2000). These field experiments represent a concentrated effort to document a general circulation model (GCM) grid-scale volume and to study the aggregation of surface exchanges from subgrid scale to landscape units.

Each land surface experiment involves detailed ground-based measurements (micrometeorology, soils, vegetation) at selected sites, research aircraft with in situ and remote sensing instrumentation, aerological soundings, and a variety of satellite observations, and thus provides a complete dataset for model evaluation. The First International Satellite Land Surface Climatology Project (ISLSCP) Field Experiment (FIFE; Sellers et al. 1992) was the first land surface experiment to be used for NWP model evaluation. For that purpose, a single mean time series of surface and near-surface parameters was generated by Betts and Ball (1998). Betts et al. (1993) compared this time series with a 24-h forecast time series obtained from the European Centre for Medium-Range Weather Forecasts (ECMWF) operational model at the grid point next to the FIFE area. They found a number of systematic errors associated with overpredicted incoming solar radiation, too slow ground-surface interaction, surface and boundary layer drying out due to deficiencies in the formulations for soil moisture and surface evaporation, the roughness length for heat, and the description of entrainment at the boundary layer top.

Based on these results and further studies, improved land surface parameterizations were developed (Viterbo and Beljaars 1995) and introduced in the ECMWF operational model and used in the ECMWF 15-yr ReAnalysis Project (ERA-15). Betts et al. (1998) evaluated the performance of ERA-15 by comparing the corresponding model output time series with the same FIFE dataset. They found that some of the earlier model biases were removed, but still identified model errors and areas for further model improvement, such as the longwave radiation code and the incorporation of a seasonal cycle of vegetation and ground storage. An area of concern remains also the diurnal cycle of boundary layer development and the transport of moisture.

In a similar way and using the same FIFE dataset, Betts et al. (1997) evaluated the performances of the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) reanalysis model and the NCEP Eta Model and contributed to improving the land surface parameterization in those models. Yucel et al. (1998) compared data from the NCEP Eta Model with longer time series of surface data from two observational networks, located in a semihumid, continental climate (Oklahoma and Kansas) and in a semiarid climate (Arizona). They again find systematic biases in the model solar radiation and in parts of the surface formulation.

Betts et al. (2001) used data from the old black spruce site of the Boreal Ecosystem–Atmosphere Study (BOREAS; Sellers et al. 1997) in Manitoba, Canada, for comparison with the ECMWF and NCEP–NCAR reanalysis models. Here they find again shortcomings of the surface and boundary layer parameterizations related to radiation, soil processes, evaporation, boundary layer development, and moisture. The work led to improved land surface parameterization for high-latitude forests.

Bringfelt et al. (1999) evaluated the tile approach (Avissar and Pielke 1989; Koster and Suarez 1992) as implemented in an experimental high-resolution version of HIRLAM. Their reference observations are from three sites of different land use (agricultural and forest) in the Northern Hemisphere Climate Processes Land-Surface Experiment (NOPEX; Halldin et al. 1999). They find that model results depend critically on a realistic description of the vegetation physiology and the initial soil moisture and that the tile approach works well for NOPEX.

We evaluate here the performance of the operational HIRLAM version 5.2 radiation, land surface, and boundary layer description for a semiarid environment. The output of the fully coupled 3D model is compared with a comprehensive observational dataset from EFEDA (Bolle et al. 1993). Distributed micrometeorological surface stations, radiosondes, flux aircraft, and airborne lidar provide a unique validation dataset of the diurnal cycle of surface and boundary layer processes. The purpose of this comparison is twofold. First, the results will provide indications for further model improvement. This corresponds to the classical model validation. Second, the information on model output errors enables us to assess uncertainties arising from the use of four-dimensional data assimilation (4DDA) and forecast data in practical applications, such as impact studies and the derivation of biogeophysical parameters from satellite data. The observational dataset, the model versions, and the evaluation strategy are described in section 2. The comparison starts with the atmospheric boundary layer structure and profiles (section 3). The surface thermodynamic cycle is investigated in section 4, and the surface radiation and energy budgets are compared in section 5. Section 6 summarizes the conclusions.

2. Validation layout

The HIRLAM system is a complete NWP system including 4DDA with analysis of observations and a limited-area short-range forecasting model (Unden et al. 2002). It is currently used operationally by seven European Weather Services (Denmark, Finland, Iceland, Ireland, the Netherlands, Norway, Spain, and Sweden). HIRLAM uses the Interaction Soil–Biosphere–Atmosphere (ISBA) surface scheme (Noilhan and Planton 1989; Mahfouf and Noilhan 1996; Rodríguez et al. 2003) and the tile approach to represent fractions of land use classes present within a single grid element (Avissar and Pielke 1989; Koster and Suarez 1992). The boundary layer scheme is based on a prognostic turbulent kinetic energy model together with a diagnostic length scale according to Cuxart et al. (2000). The turbulent surface fluxes are calculated from the standard bulk formulation using differences between the surface and lowest model layer following Louis (1979) and Louis et al. (1982). The fast radiation scheme is derived from Savijärvi (1990).

The operational version 5.2 of HIRLAM is used for the validation. Its horizontal grid resolution is 0.5°. The analysis of upper-air variables is based on an optimum interpolation method (Unden et al. 2002). A research version of HIRLAM [as currently used by the Spanish Weather Service Instituto Nacional de Meteorologia (INM)] serves to analyze details at higher spatial resolution (0.25°). Unless explicitly mentioned, the results shown in this paper refer to the 0.5° version. The comparison focuses on the 6-h forecast (fc06). Differences in prognostic variables introduced by the analysis increment are discussed whenever they are significant.

The observational datasets were collected during the EFEDA field phase in June 1991 (Bolle et al. 1993). A large sample of 22 case studies of anticyclonic conditions typically encountered in the area during the summer months is available. The 23 June 1991 is most densely covered by observations (Michels and Jochum 1995). The sample composite is almost identical to the individual 23 June observations (Jochum et al. 2004, manuscript submitted to J. Appl. Meteor.). Therefore, the individual case study approach is adopted in this work. The diurnal cycles of surface and ABL processes are compared in detail with observations of 23 June 1991. The performance evaluation at seasonal and annual time scales is the subject of a follow-up study that uses data from two permanent observing stations in the same area.

The EFEDA experimental area is located in the region called La Mancha, part of the Castilian high plateau in the southeast of Spain, at an average elevation of 700 m above mean sea level. It extends about 60 km in north–south direction and 80 km in east–west direction, respectively. The area is generally flat, with elevation variations of up to 100 m, but surrounded and influenced by mountain ranges. The distance to the mountains is 35 km from the south, 45 km toward the northeast, and 40 km toward the northwest. Three intensive observation sites (supersites) were selected as to represent the major Mediterranean ecosystems found in the area.

Figure 1 shows the location of the two main supersites, Tomelloso (o) and Barrax (+), along with the locations of operational radiosonde stations. The symbols mark the center points of a 10 km by 10 km square of each supersite. These center points coincide with the high-resolution radiosonde launch sites in either case. A regular 1° by 1° grid (solid lines) serves to give a rough orientation about the HIRLAM grids. The two different resolutions of HIRLAM used here, 0.5° and 0.25°, would subdivide each 1° by 1° grid element into 4 and 16 smaller elements, respectively. Table 1 summarizes the model grid dimensions.

This comparison focuses on the Tomelloso site, which is representative of semiarid conditions and thus would be expected to be well reproduced by the model. The Barrax site is characterized by two additional sources of moisture, one at the ground from irrigation and one aloft from the Mediterranean sea breeze penetrating inland in the afternoon. The Barrax comparison is used occasionally, whenever it provides additional insight. Both represent typical Mediterranean land use. For each site, the comparison is performed in column mode, that is, aggregated site observations against model output at the corresponding grid points. The site-aggregated surface dataset for Tomelloso is taken from Linder et al. (1996). The site-aggregated dataset for Barrax was developed on the basis of Linder et al. (1996) and the vegetation classification of Calera Belmonte (2000). The procedure and results from aggregating surface fluxes to various NWP model grid elements are described by Jochum et al. (2004, manuscript submitted to J. Appl. Meteor.). The radiosonde and aircraft datasets are described in Michels and Jochum (1995) and Jochum (1993).

Because of the presence of mountainous terrain at model scales, the selection of model grid elements for comparison needs great care. The principal and obvious idea is to have the observational site inside the model grid cell, which in practice matches the nearest-gridpoint concept of Betts et al. (1993, 1998). Both FIFE and BOREAS, however, are located within extended areas of similar characteristics and of small topographic variations. In contrast, the EFEDA area itself is rather flat, but is surrounded and influenced by mountain ridges, which come closest to the area from the south (35 km) and east. In a pragmatic way, we have still identified the nearest grid point to the location of each observational set of profiles (see Table 2 for all relevant coordinates), while monitoring the differences to a set of adjacent grid points, in order to assess the influence of the model orography. All points included in the comparison are located within the same climate regime. The model grid elements, which include the Tomelloso and Barrax sites, also contain part of the southern mountains. The model grid elements covering the northern part of the EFEDA area include more mountain ranges toward the north and east. The model uses a mean orography, which is consistent with the current mountain drag formulation. This results in a notable difference in elevation between grid element and site, depending on the model type and resolution. For Tomelloso, for example, the HIRLAM 0.5° grid surface is at 100 m above ground level.

3. Boundary layer structure and profiles

The daytime diurnal cycle of 23 June 1991 was densely covered with aircraft observations. Radiosondes were launched roughly every 2 h at Tomelloso and Barrax. The day was almost cloud free with weak winds. Michels and Jochum (1995) performed a detailed analysis of the diurnal cycle of the ABL structure at both sites, as obtained from all available observations (radiosoundings, aircraft ascents, airborne lidar).

At Tomelloso (Fig. 2), the stable nocturnal boundary layer (NBL) is about 500 m deep and remains until very late in the morning (1100 UTC). The mixed layer (ML) grows rather slowly until then. Its rapid growth phase starts only after complete erosion of the NBL and is furthermore slowed down by the slightly stable stratification of the very deep (3 km) residual layer (RL). The ML reaches its full depth by late afternoon (1730 UTC) only, when finally the RL has been eroded as well. The maintenance of the RL for most of the day is rather unusual, but was observed quite often during EFEDA (Michels and Jochum 1995). A possible reason is the occurrence of cirrus cloud reducing the incoming solar radiation. Also, the slightly stable stratification of the RL is relevant, since more energy is needed to remove a stable RL than a neutral one. The RL would be slightly stable for the combined effect of two processes. On one hand, entrainment of warm air at the top of the ML will continue for a short while after the ML has decoupled from the surface in the evening. On the other hand, nighttime cooling due to aerosol and water vapor occurs in the NBL and lower RL.

At Barrax (not shown; see Michels and Jochum 1995), the NBL is twice as deep and takes until noon to be eroded. The ML is growing more slowly and reaches a lower final depth. Both features are due to the larger amounts of water vapor available from the irrigated fields. Moisture advected by the sea breeze penetrating inland in the afternoon inhibits further ML growth here. The RL is similar to the one at Tomelloso.

The comparison of the vertical profiles of potential temperature at Tomelloso (Fig. 2) shows small differences between model forecast and observations. The profiles from analysis (not shown) and forecast are very similar. Given the individual sonde measurement errors, differences in temperature and specific humidity exceeding 0.8 K and 1 g kg⁻¹, respectively, are considered to be significant. At 0600 UTC, the model NBL temperature is about right, whereas the model RL is slightly (1 K) too cold. At 1200 noon, the ML temperature is reproduced very well. However, the model does not capture the weakly stable RL and, therefore, produces an ML top at about 740 hPa, which is almost 100 hPa too high. By 1800 UTC, the model ML top is about right, but the ML is 1 K too cold. There are no observations available at midnight. These differences in potential temperature are just above the limit of significance. Taking into account that each comparison involves a single radiosonde profile (which gives a “snapshot” only) and that the results for other fair-weather days (not shown here) are similar, it is concluded that the model performs well on daytime potential temperature.

In the 0600–1200 UTC period, the model-diagnosed warming (Fig. 3a) of the lower ABL is too high, especially at the lowest grid point (by 0.3 K h⁻¹). This period is characterized by erosion of the NBL and incipient growth of the ML, so a combination of factors could contribute to the overprediction. Part of the rapid ML growth phase occurs in the afternoon period 1200– 1800 UTC. Therefore, the overall observed ML warming during that period is stronger in the lower layers. This feature is not captured by the model, which gives rise to a slight underprediction (0.1 K h⁻¹) of lower ML warming. The RL warming/cooling is reproduced fairly well in both periods.

For moisture, the situation is different. At 0600 UTC, the radiosonde shows a complex layered structure (Fig. 2), with a deep relatively dry (4 g kg⁻¹) RL and a moist NBL (6–7 g kg⁻¹) on top of the shallow drier incipient ML. The model forecast (at all times again similar to analysis) comes close to the NBL moisture (0.5 g kg⁻¹ higher), but is unable to resolve the rest of the layers. This means also that the incipient ML is too moist. At 1200 UTC, the model gives an ML top between 800 and 760 hPa, with humidity slightly decreasing upward. The radiosonde-derived ML extends only up to 850 hPa, while the rest is RL. The model ML average of specific humidity (7 g kg⁻¹) is about 1.5 g kg⁻¹ (27%) too moist. At 1800 UTC, the model ML is again about 1.5 g kg⁻¹ too moist. The ML top around 660 hPa agrees reasonably well with observations.

The observed drying or moistening varies vertically according to the layered structure (Fig. 3b). During the morning period (0600–1200 UTC) the erosion of the NBL leads to drying of the lowest layers. The lower RL is moistening by entrainment, while the upper RL is strongly drying by entrainment of very dry free atmosphere (FA) air. The model is able to reproduce the general shape of the vertical profile in this case, even though the RL top is too low. During the afternoon period (1200–1800 UTC), the model ML dries out more than observed, which helps to partially offset the low bias in the 1200 UTC humidity profile (such that the 1800 UTC profile is less biased). The seemingly good agreement of the upper-RL moistening is due to another compensating effect rather than to the model's ability to reproduce the corresponding physical processes. This moistening peak results from the final ML growth going on in the early afternoon, when the remnants of moist RL air are entrained down into the ML. As explained above, the RL erosion takes a large part of the afternoon. The model does not resolve the moist RL, neither at 1200 nor at 1800 UTC. The model-produced moistening peak results rather from the underpredicted RL moisture at 1200 UTC.

Figure 4 shows the comparison of model forecast profiles at adjacent grid points. The Tomelloso gridpoint elevation is 755 m above mean sea level (MSL), the Barrax grid point is at 891 m MSL. Other than altitude effects, there are small differences in moisture (0.3–0.6 g kg⁻¹), but not in temperature. The comparison to observations for the Barrax grid point is slightly more favorable because of the enhanced moisture observed there. However, this cannot be attributed to the model performance, since the major moisture source (irrigation) is not represented in the model.

The general conclusion from the ABL comparison is that two types of differences occur. One is the general high bias in ML moisture. The other is in the vertical structure itself, since the model does not resolve the RL properly. The roots for the latter might be found in the vertical model resolution, which is insufficient for this type of landscape, and in the vertical diffusion and entrainment parameterization (see also Lenderink and Holtslag 2000). Part of the layered moisture structure observed in the RL is probably due to mesoscale circulations of topographic origin (Miao et al. 2003), which obviously are not resolved by the model.

The high bias in ML moisture could originate from various sources. The ML moisture is locally determined by the state and fluxes at the surface and at ML top, but is also conditioned by mesoscale dynamics and the large-scale flow field. The surface flux bias identified in section 5 below would produce a low bias in ML moisture, rather than the observed high bias. The entrainment fluxes cannot be directly validated. The unresolved vertical moisture layers suggest a potential error source there. The remaining candidate error source is the moisture data assimilation, where the range of relevant scales is not resolved in the model, but also not in the operational observations used in the assimilation. Problems in the analysis of moisture in regions of complex topography and close to coastlines have been reported on various occasions (e.g., Trenberth and Guillemot 1998) and remain unresolved to date (Dee and da Silva 2003).

In the present case, the only operational observations available for the moisture analysis are the radiosonde data at the locations shown in Fig. 1. First, standard radiosonde observation levels cannot capture the observed complex ABL vertical structure. Second, the two nearest stations to the EFEDA area are located in very different climate regimes (Madrid in dry continental; Murcia in semihumid Mediterranean climate). Figure 5 shows the analysis moisture profiles at the observation points Madrid (Ma) and Murcia (Mu; where the analysis scheme successfully corrects an observation error at 850 hPa).

The example provides evidence of a similar (albeit smaller) high bias as observed at the Tomelloso grid point, which lies in between these two stations and which is in the same climate and elevation regime as Madrid. This suggests shortcomings of the analysis scheme (which we do not investigate here). On the other hand, the spatial (horizontal and vertical) sampling of the observing system alone would explain part of the deficiencies in the analyzed moisture fields in this area. Along similar lines, control runs with a 3D variational assimilation scheme (instead of optimum interpolation as used in our reference version) do not improve the results.

4. Surface thermodynamic cycle

The analysis of screen-level variables uses univariate statistical interpolation (Rodríguez et al. 2003). Input data are the 2-m temperature and relative humidity from SYNOP reports. The model diagnostic package calculates the 2-m air temperature and humidity by means of interpolation between the surface (temperature and wetness) and the lowest model level (about 30 m above ground). The special interpolation technique used for this purpose preserves the surface-layer similarity profiles (Geleyn 1988). Model data are compared with the site-aggregated screen-level surface observations. Based on the assessment of the individual measurement errors and the intrasite variability, differences in temperature and specific humidity exceeding 0.8 K and 1 g kg⁻¹, respectively, are considered to be significant and, thus, indicative of model biases.

The observations of 2-m temperature at Tomelloso (Fig. 6a) show a typical diurnal cycle of a fair-weather summer day in a dry continental climate. We find a narrow (2 h) minimum in the very early morning before sunrise (0445 UTC), followed by 6 h of rapid warming until almost noon. The warming then slows down to lead into a broad (4 h) afternoon temperature maximum. The evening cooling sets on about 1 h before sunset (1949 UTC). The Barrax 2-m temperature is generally lower, except for the early part of the rapid warming phase. After 1800 UTC it is further reduced by the onset of irrigation. The model analysis output at the four forecast intervals cannot be expected to resolve details of the diurnal cycle. At both sites (Barrax not shown), the model forecast reproduces well the midnight and 0600 UTC temperatures, although the model obviously misses the minimum. The latter is thought to be connected to the land surface and boundary layer scheme (see also Vogelezang and Holtslag 1996). The model performs well for the rapid warming phase, but then stops short of reaching the observed maximum. The 1800 UTC temperature is slightly too low. This means that the morning surface warming slows down and stops earlier in the model than in reality. The observed high temperature amplitude of the day (17 K) cannot be fully reproduced by the model output (which does not provide daily minimum and maximum temperatures), but is obviously strongly underestimated. Intermediate 3-h forecast data (circles in Fig. 6a; not part of the routine output) confirm this conclusion.

The 2-m specific humidity diurnal cycle (Fig. 6b) exhibits a more complex behavior. We observe two maxima, one in the early-morning hours and one in the late evening till midnight. The daytime drying is strongest in the late morning, during the rapid ML growth phase, and again in the late afternoon, coinciding with the moment when the plants have closed their stomata. Because of the presence of moisture from irrigated fields, the surface at Barrax keeps moistening or stays at almost constant specific humidity until late afternoon, when some drying sets on. The model reproduces fairly well the daytime dry state and the weak drying from noon to 1800 UTC. At night, however, the model overestimates the moistening between 1800 and 2400 UTC and then keeps moistening until 0600 UTC, where it reaches a maximum that is 2.5 g kg⁻¹ above the observed morning maximum.

The physical processes contributing to the daytime thermodynamic cycle can be understood from the (θ, q) mixing lines shown in Fig. 7, where θ is 2-m potential temperature. The observations show a decoupling of surface warming and drying processes during the early-morning warming period (0600–1000 UTC). To a lesser extent, this applies also to the late-morning drying period (1000–1200 UTC) (where slight warming does occur) and to the afternoon warming period (where some drying occurs). The model output obviously does not resolve these individual periods. Model warming and drying appear as coupled processes with strong drying and strong warming between 0600 and 1200 UTC (points 2 to 3 in Fig. 7). Less drying and very little warming occurs between 1200 and 1800 UTC (points 3 to 4). Recurring again to the (nonroutine) 3-h forecast data (points B at 0900 and C at 1500 UTC for Tomelloso in Fig. 7), we see evidence of two distinct, but still fully coupled, regimes in the morning phase: the first with stronger warming, and the second with stronger drying. In contrast, we see one continuous coupled regime at Barrax (points b and c).

The observed decoupling of warming and drying processes is essentially due to the complex and inhomogeneous moisture fields in the ABL and above (Michels and Jochum 1995). During the early-morning warming period, there is very little surface evaporation at Tomelloso. With the slowly growing ML, moist air is entrained from the overlying NBL. This contributes to a slight moistening of the surface until the shallow NBL is eroded. This is then the start of the rapid ML growth phase, where entrainment of very dry air from the RL contributes to strong surface drying. Entrainment of RL air does not enhance warming, because the RL and ML have about the same potential temperature at that time. When the ML has grown to its maximum top (equal to the RL top), entrainment brings in warm, but not very dry, air from the free atmosphere aloft. This again leads to a weak decoupling of warming and drying.

The surface warming/cooling and drying/moistening is generally driven by the heat and moisture exchange at the surface and by entrainment processes reaching downward from the ML top to the surface (e.g., Mahrt 1991). Airborne lidar observations and bivariate conditional sampling of airborne fluxes show that entrained air almost reaches the surface layer (Michels and Jochum 1995).

As discussed in the context of the ABL profiles, the model is obviously not able to resolve the observed vertically layered moisture fields and probably does not produce strong enough entrainment. Its daytime surface thermodynamic cycle starts at 0600 UTC, potentially too warm and much too moist. The warming rate is less than observed, which offsets the wrong start and helps the model to catch up with the observed thermodynamic state at noon. During the afternoon forecast period, it seems that the model again does not reproduce the entrainment contribution to warming and drying.

Table 3 gives a summary overview of the major findings of this section.

5. Surface radiation and energy budgets

The model output surface fluxes are time integrated over the past 6-h forecast period at each forecast time (0000, 0600, 1200, and 1800 UTC). The observations consist of time series of half-hourly flux averages. Figure 8 shows the observed diurnal cycles of the surface radiation and energy budgets, respectively. For comparison with model output data, 6-h averages were derived from each time series of surface flux observations. Figure 9 and Tables 4 and 5 show the comparison of the resulting 6-h average fluxes from model and observations. The reference point for the comparison is again Tomelloso. Results for Barrax are shown and used whenever they help to gain additional insight.

a. Incoming solar radiation and aerosol load

The model overestimates the incoming solar radiation by 10% in the morning and 17% in the afternoon (Table 5 and Fig. 9a). This is surprising for the sign of the bias and even more for its magnitude. The model atmosphere is too moist and slightly too cold (section 3; Fig. 2). Therefore, one would expect a reduced shortwave (SW) transmission in the model. With a more realistic atmospheric moisture profile, the bias would even be larger. Betts et al. (1993) found a similar (10%) high bias in the T-106 L-19 cycle 39 version of the ECMWF model as compared to FIFE data. With the introduction of aerosols and the prognostic cloud scheme in the ERA-15 reanalysis scheme, the incoming shortwave radiation performed rather well for the FIFE conditions (Betts et al. 1998). Comparison of the ERA-15 data with the same EFEDA–Tomelloso observations (not shown here) does indeed show a better agreement of the incoming solar radiation (5% low in the morning; 10% high in the afternoon).

About 1–2/10 of thin cirrus cloud was observed in the area on most days. The model grid point Tomelloso does not show any cloud cover on 23 June, whereas at the Barrax grid point it gives 0.11 cirrus cloud cover at noon. The lack of cloud cover at the model grid point would explain about 2% of the bias.

In order to explore further potential sources of the SW high bias, details of the HIRLAM SW radiation scheme (Sass et al. 1994) were analyzed. The scheme is based on the parameterization of Savijärvi (1990), whose terminology is used here. For cloud-free grid elements, the incoming solar radiation at the surface (watts per meter squared) is expressed as

View Expanded

where aa and as are constants (see below), u is the column-integrated (linearly) pressure-scaled water vapor (centimeters), and α is the SW albedo of the ground. The “solar constant” S (watts per meter squared) is approximated as

View Expanded

where d is the running date from 1 January, and h the local sun-height angle.

The terms in Eq. (1) represent the reduction of the top-of-the-atmosphere solar radiation (S sinh) by ozone UV and visible absorption in the stratosphere (term I), water and tropospheric ozone and CO₂ absorption (II), Rayleigh scattering (III), and backscattering from the reflected beams (IV). The constants aa = 1.2 and as = 1.25 serve to include the effects of aerosol absorption and scattering, with values selected from empirical fits to data in “continental industrialized areas during normal visibility” (Savijärvi 1990).

EFEDA field visibility records and airborne lidar observations (Jochum 1993; Kiemle et al. 1995) suggest that the atmosphere was hazy and that visibility was low on most days. Therefore, the aerosol parameterization constants are not expected to be adequate for the area. We have calculated the individual terms of Eq. (1) from the corresponding model data. Terms I and IV are very small (below 3%). Terms II and III contribute about 20% and 7%, respectively. Enhancing the aerosol absorption coefficient aa from 1.2 to 1.5 increases the contribution of the corresponding term to 26%, which would result in reducing the high bias by about 6%.

Some of the aerosol may also be of marine origin, especially in the afternoon (brought into the area by the sea breeze penetrating inland from the Mediterranean coast in the afternoon). This would contribute to enhanced scattering as well, increasing the aerosol scattering coefficient as to 1.9 (Paltridge 1973) and, thus, augmenting term III by 2%–3%.

6. Conclusions

Observations from the European Field Experiment in a Desertification-threatened Area (EFEDA) were used to evaluate the performance of the radiation, land surface, and boundary layer description of HIRLAM in semiarid conditions. Analysis and 6-h forecast data of the fully coupled three-dimensional model were compared with the comprehensive dataset of a case study representative for a sample of 22 days of anticyclonic conditions.

The model surface, soil, and boundary layer are found to be too moist and slightly too cold during most of the diurnal cycle. The model radiation and surface energy budgets are biased toward more humid conditions.

Summarizing the overall results of the comparison and assigning priorities (based on bias magnitudes and impacts) to the observed discrepancies, we obtain the following picture:

1. Assignment of too dense vegetation leads to biased surface fluxes and net radiation. With adequate land-use class assignment, the ISBA surface scheme performs well.

2. The incoming solar radiation is overestimated, which results from the aerosol parameterization in the radiation scheme. Modification of the aerosol parameters to account for lower visibility reduces the bias significantly.

3. Large errors (up to 50%) in the ABL humidity profiles result from deficiencies in the moisture assimilation.

4. The residual layer is not resolved properly and the entrainment contribution to surface and ML warming and drying is not reproduced adequately because of the predominant effect of vertical model resolution, which is insufficient for this type of landscape, and associated deficiencies in the vertical diffusion and entrainment parameterization.

5. A soil type with too high clay percentage might generate too much soil moisture, although the soil moisture assimilation effects would need to be assessed here as well.

6. The high bias in net longwave radiation is probably related to the excess atmospheric moisture and the underestimated surface temperature.

Immediate improvement of the model performance can be expected from the use of a land-use and soil classification (with its associated physiographic database) adapted to Mediterranean landscapes, in combination with the use of aerosol parameters in the radiation scheme, that account for the typically higher aerosol load of arid and semiarid environments.

An adequate land-use classification for Mediterranean landscapes would need to account for sparse dry canopies as well as for two classes of irrigated crops (spring-irrigated and summer-irrigated), with a physiographic database reflecting their different phenological cycles. The classification of Calera Belmonte (2000) is a regional-scale example of the approach. The classification of Champeaux et al. (2000) follows a similar approach, based on the seasonality of the Normalized Difference Vegetation Index (NDVI) in more general terms and covering all western Europe. The use of this classification in HIRLAM is currently being tested at the Spanish Weather Service.

Most of these conclusions are expected to be valid for other NWP models as well. In fact, the ongoing comparison of ERA-15 and ERA-40 data with the same observational data gives similar results. Our results are in line with other NWP validation studies, which have often focused on the land surface scheme. Major NWP centers are currently upgrading their land-use classification and physiographic database (e.g., van den Hurk et al. 2000; Gustafsson et al. 2003; Masson et al. 2003) to account for a wider variety of land-use classes and for more realistic seasonal variations in key canopy parameters. Irrigated crops, however, have not yet been separated in seasonal classes.

The radiation scheme is more specific for HIRLAM. Other schemes might require a different analysis and solution, although the higher aerosol load prevailing in (semi-) arid environments would generally need to be accounted for. An improved aerosol parameterization for Mediterranean areas could draw on aerosol climatologies (e.g., Holben et al. 2001) or use satellite-derived aerosol datasets (e.g., Yu et al. 2003) that are recently becoming available.