The total downwelling shortwave (SWD) and longwave (LWD) radiation and its components are assessed for the limited-area version of the Global Environmental Multiscale Model (GEM-LAM) against Atmospheric Radiation Measurements (ARM) at two sites: the southern Great Plains (SGP) and the North Slope of Alaska (NSA) for the 1998–2005 period. The model and observed SWD and LWD are evaluated as a function of the cloud fraction (CF), that is, for overcast and clear-sky conditions separately, to isolate and analyze different interactions between radiation and 1) atmospheric aerosols and water vapor and 2) cloud liquid water. Through analysis of the mean diurnal cycle and normalized frequency distributions of surface radiation fluxes, the primary radiation error in GEM-LAM is seen to be an excess in SWD in the middle of the day. The SWD bias results from a combination of underestimated CF and clouds, when present, possessing a too-high solar transmissivity, which is particularly the case for optically thin clouds. Concurrent with the SWD bias, a near-surface warm bias develops in GEM-LAM, particularly at the SGP site in the summer. The ultimate cause of this warm bias is difficult to uniquely determine because of the range of complex interactions between the surface, atmospheric, and radiation processes that are involved. Possible feedback loops influencing this warm bias are discussed. The near-surface warm bias is the primary cause of an excess clear-sky LWD. This excess is partially balanced with respect to the all-sky LWD by an underestimated CF, which causes a negative bias in simulated all-sky emissivity. It is shown that there is a strong interaction between all the components influencing the simulated surface radiation fluxes with frequent error compensation, emphasizing the need to evaluate the individual radiation components at high time frequency.
The surface radiation budget (SRB) is one of the main controls on key surface variables, such as temperature, soil moisture, snow cover, and evaporation rates. A systematic bias in the simulated SRB can lead to errors in any of these variables, with the potential for subsequent error propagation throughout the simulated climate system. With respect to simulating anthropogenic climate change and feedbacks involving cloud–radiation interactions, it is important that the fundamental processes controlling the SRB in a given model are accurately simulated. Because the simulated SRB is mainly controlled by downwelling shortwave (SWD) and longwave (LWD) radiation, it is therefore highly dependent on the representation of cloud amounts, microphysical processes, and cloud–radiation interactions. Because of their extreme complexity, cloud–radiation interactions are highly parameterized in present-day models. As mentioned in the Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (Randall et al. 2007), large differences exist between climate models in their simulated cloud–radiation feedbacks, which are the main source of uncertainty in climate model sensitivity to a doubling of atmospheric CO2 (Bony and Dufresne 2005; Soden and Held 2006). To make reliable estimates of future climate conditions, it is crucial that further improvements are made in our ability to simulate the fundamental physics controlling the SRB.
A number of studies have evaluated simulated cloud amounts and SRB in climate models, often at the climatological scale and with different modeling tools, such as global climate models (GCMs), regional climate models (RCMs), or single-column models (SCMs). GCMs are valuable tools for studying cloud–radiation interactions (Cess et al. 1996; Norris and Weaver 2001; Walsh et al. 2002; Weare 2004; Stephens 2005; Martin et al. 2006; Williams et al. 2006), because feedbacks (e.g., surface radiation–surface evaporation–cloud formation) can develop in an internally consistent manner within the model. This can help in improving the main feedback loops controlling the SRB. However, GCMs over a given region can suffer from circulation errors, often with origins that are remote to the region of study, which makes it difficult to evaluate the simulated SRB against point observations. With RCMs, the simulated large-scale meteorology can be partially constrained to follow the observed evolution through the application of analyzed lateral boundary conditions (LBCs), while still leaving freedom for local interaction between the model parameterizations and the resolved dynamics. As a result of the constraints resulting from the application of analyzed LBCs, simulated RCM processes can be compared to point observations in a common thermodynamic/dynamic phase space (Hogan et al. 2001; van Meijgaard and Crewell 2005).
The limited-area version of the Global Environmental Multiscale Model (GEM-LAM; Côté et al. 1998; Zadra et al. 2008) is presently being evaluated for use as a new operational RCM for regional climate change projection over Canada. As part of a group effort on evaluating and developing the different GEM-LAM components, Markovic et al. (2008, 2009) have identified summer biases over North America for GEM-LAM for the SWD and LWD, coupled respectively with an underestimate of the cloud fraction (CF) and a warm 2-m temperature bias. They also show that cloud cover and cloud-free SWD biases often compensate to result in an accurate SWD for all-sky conditions.
Markovic et al. (2008) focused their analysis on central North America. It is important that models be evaluated over a wide range of climate conditions to test the parameterization transferability as discussed in Takle et al. (2007) and Morrison and Pinto (2006). Therefore, we concentrate on two sites from the Atmospheric Radiation Measurement (ARM) Program, with high-quality observations of cloud and radiation but radically different climates: the southern Great Plains (SGP) site in the central United States and the North Slope of Alaska (NSA) site in Barrow, Alaska. A methodology is developed to fully utilize the cloud and radiation observations from these two surface sites in order to evaluate in greater detail the cloud and radiation processes simulated by GEM-LAM. The objective of this work is to gain a better understanding of the origin of the biases found in the aforementioned references and to facilitate a more globally applicable approach to parameterizing, in particular, the cloud microphysical terms that strongly influence simulated radiation fluxes.
The SRB biases may arise from errors in the simulated cloud fraction, position, geometry, cloud water content, and phase, as well as in the assumed cloud optical properties. Errors in clear-sky conditions (or clear portions of a partially cloudy grid box), may arise from the different inputs to the radiation scheme, that is, temperature, water vapor, aerosols, and trace gases. To help understand and quantify the contribution of these errors, the model cloud fraction is first evaluated. It is expected that the simulated cloud cover plays a key role in determining the overall biases in SRB. We then evaluate the SWD and LWD separately for all-sky, clear-sky, and overcast conditions. Analyzing the results in the clear-sky and overcast conditions are essential to evaluating error contributions that are independent of the cloud-fraction error (e.g., cloud water content in overcast conditions or integrated water vapor in clear skies). This analysis is done at annual, seasonal, and diurnal time scales. Such a methodology can help to isolate specific processes (e.g., cloud physics, soil moisture control) or inputs (e.g., aerosols, albedo) requiring improvement.
The paper is organized as follows: in section 2, the model, observations, and evaluated variables are described. Section 3 presents a comparison between model results and observations, beginning with a brief evaluation of the large-scale meteorology simulated by GEM-LAM at the two ARM sites (section 3a). This is followed by an analysis of the CF and surface radiation fluxes in sections 3b and 3c. SRB is then split into clear-sky and overcast conditions to analyze the individual components controlling the total SRB (sections 3d and 3e) in more detail. Section 4 contains a discussion of the main results and recommendations for future work.
a. Model description and integration
GEM-LAM employs a two-time-level, semi-Lagragian, fully implicit advection scheme and a one-way lateral boundary nesting strategy following Davies (1976). Surface albedo and surface fluxes of heat, moisture, and momentum are calculated over four surface subtypes [land, water, sea ice, and land ice see Bélair et al. 2003a,b]. Subgrid-scale turbulent fluxes are calculated using an implicit vertical diffusion scheme, with prognostic turbulent kinetic energy (TKE) and a mixing length based on Bougeault and Lacarrère (1989) and Bélair et al. (1999). GEM-LAM uses a prognostic total cloud water variable with a bulk-microphysics scheme for nonconvective clouds. Fractional cloudiness is based on a relative humidity threshold, which varies in the vertical (Sundqvist 1988). Individual cloud layers are assumed to overlap in the vertical using a maximum random cloud overlap. The deep convection scheme is that of Kain and Fritsch (1990, 1993), whereas a Kuo transient scheme is used for shallow convection (Kuo 1965; Bélair et al. 2005). The radiation scheme is due to Li and Barker (2005) and employs a correlated k-distribution (CKD) method for gaseous transmission, with nine frequency intervals for longwave and four frequency intervals for shortwave radiation. Although the longwave spectrum and the near-infrared portion of the shortwave spectrum are treated using the CKD method, the rest of the shortwave spectrum is dealt with in frequency space with UV-C, UV-B, UV-A, and photosynthetically active radiation separately considered. The scheme treats the following gases interactively: H2O, CO2, O3, N2O, CH4, CFC11, CFC12, CFC113, and CFC114. The radiative effect of the background aerosols is included based on the climatology of Toon and Pollack (1976). This simple climatology specifies maximum aerosol loading at the equator and a decrease toward the poles, with different values for continents and oceans. The separation of total cloud water into liquid and solid is based on the local air temperature ranging from all ice at −40°C to all liquid at 0°C (Rockel et al. 1991). The liquid and solid effective radii (reff,liq and reff,sol) range from 4 to 17 μm (liquid) and from 20 to 50 μm (solid), parameterized as a function of the local cloud liquid or ice water content (Lohman and Roeckner 1996).
The model was run with a horizontal resolution of 0.5° and 53 vertical levels, extending up to 10 hPa. The model time step was 1800 s, while the radiative transfer is done every three time steps (1.5 h). In between the radiative time steps, the LW fluxes and heating rates are constant, whereas the SW fluxes and heating rates are corrected for the change in solar angle. Two geographically separate integrations were made for the 1998–2004/05 period, both of which employ observed sea surface temperatures (SSTs) and sea ice, and are derived from the Atmospheric Model Intercomparison Project (AMIP; Hurrell et al. 2008) dataset, for the lower boundary conditions, and the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40), for the lateral boundary conditions. The two integration domains (shown in Fig. 1) are each centered on one of the ARM observation sites. The choice of the two sites is due to the different climate regimes sampled at the sites. The SGP site is dominated by convection during the summer, whereas during winter the midlatitude synoptic weather systems are dominant. For the NSA site, while it experiences year-round cloudy conditions, multilayered liquid or mixed phase clouds are dominant during the summer, whereas in winter the mixed phase and low-level ice clouds dominate (Intrieri et al. 2002; Shupe et al. 2005; Curry et al. 1996).
b. Evaluated variables
We evaluate the model and observed SWD and LWD as a function of CF, that is, for overcast and clear-sky conditions, separately. Clear-sky conditions are determined when the CF is less than 10%, whereas overcast conditions are for a CF of 90% or more. This categorization is done separately for the model and observations, and then the evaluation is done on a set of overcast or clear-sky cases as a climatological analysis across a large range of common conditions. Note that the model cloud fraction is a true cloud fraction, in contrast to the observed retrievals, which provide effective cloud fractions. For this reason, and in order to increase the dataset available for evaluation so that robust statistics can be achieved with respect to model performance, 10% and 90% are chosen as the thresholds rather then 0% and 100%.
To evaluate the individual inputs to the radiation scheme, we first evaluate some of the components of the atmospheric water cycle, comparing the modeled and observed CF, liquid water path (LWP), integrated water vapor (IWV), and precipitation. The ice water path (IWP) would complete this evaluation, however, the available IWP observations seemed inconsistent at the time of our analysis. LWP and IWV are restricted to nonprecipitating periods because of the unreliability of the microwave radiometer when the instrument is wet. Thus, modeled LWP and IWV are also filtered to exclude cases when precipitation is greater than 0.25 mm over a 3-h period. The sensitivity of the simulated LWP to the threshold defining precipitation removal is assessed in section 3e.
We compare modeled variables at the grid point nearest to the relevant observation site. All of the variables are averaged or accumulated over 3-h intervals to reduce the representativity error between a single-point observed variable and a modeled gridbox-mean variable (van Meijgaard and Crewell 2005; Hogan et al. 2001). The period of comparison is from 1998 to 2004 for SGP and from 1998 to 2005 for NSA.
The seasonal and diurnal cycles are the two biggest forced modes of variability in the climate system; we therefore analyze the mean diurnal cycle of SWD and LWD to identify systematic errors within the diurnal cycle that may contribute to seasonal-mean errors. We also use 3-hourly mean frequency distributions to compare modeled quantities, such as LWP and precipitation, to observations, in order to verify that seasonal-mean results do not result from higher time frequency error cancellation. Frequency distributions can also indicate under which meteorological/climate regimes the model differs, most often from observations.
To complete our analysis, we plot 3-hourly mean covariability plots of SWD and LWD versus LWP. This is done for overcast conditions for both the model and observed quantities, and allows us to assess whether the model captures the underlying physical relationships of the cloud–radiation interaction controlling the simulated SRB.
c. Observation datasets
For the two sites, the observations were obtained from the ARM archive (available online at http://www.arm.gov). Table 1 lists all of the datasets used, along with a quoted observational accuracy when reported. For the SGP site, all observations are extracted for the Central Facility (CF1) when available, and, if not, they were extracted from the extended facility (E13). For the NSA site, all observations come from Barrow (C1). For the two sites, CF observations are available from many different sources. We compared five different estimates for SGP and four estimates for NSA (different instruments or analysis methods) for a common period to evaluate whether each method detected the same CF and thereby determine a range of uncertainty in the CF observations. Figure 2 shows the mean diurnal cycle of 3-hourly mean cloud observations for summer and winter separately. The period for SGP is from June 2000 to December 2003, and for NSA it covers only the year 2004. The five datasets for SGP are the CF derived from the shortwave radiation analysis of Long et al. (1999; Long; stars), the total sky imager (TSI; squares; see Kassianov et al. 2005), the microbase cloud-radar dataset (MB; circles; see Mather and McFarlane 2004), the International Satellite Cloud Climatology Project (ISCCP; crosses; see Rossow and Schiffer 1991; Rossow and Schiffer 1999), and the Vaisala ceilometer (VCEIL; diamonds; see Lonnqvist 1995). For the NSA site, the four datasets are from the Vaisala ceilometer (diamonds), the microbase cloud-radar dataset (circles), the ISCCP satellite data (crosses), and the micropulse lidar (MPL; plus signs; see Welton and Campbell 2002).
Figure 2 shows that for SGP the Vaisala ceilometer generally underestimates CF compared to the other observations, whereas ISCCP, which is representative of a 30 km × 30 km area, tends to overestimate CF for December–February (DJF), compared to the other observations. For the ceilometer, the summer underestimate is likely explained by the maximum detection height of 7.5 km, which leads to an underdetection of upper-troposphere, optically thin clouds (Lonnqvist 1995). For ISCCP, the winter differences may arise from the documented problems satellites have in distinguishing winter season low-level clouds, where discrimination between a low-level cloud and a snow-covered surface is difficult in the visible wavelengths, whereas discrimination between cloud-top infrared emission and surface emission is complicated because of the frequent presence of a low-level thermal inversion (Key and Barry 1989; Schweiger and Key 1992). The Long, total-sky imager and microbase CF generally agree within 5%–15% at SGP for both seasons. For this reason, we used these three datasets in our analysis, averaging the three datasets every 3 h when all were available. If one or two datasets are not available at a given time, the datasets that are available are used as the observed CF. For the NSA site, the Vaisala ceilometer seems to match the microbase dataset more closely (clouds that are generally located at a lower altitude at NSA compared to SGP means that the 7.5-km height limit of the ceilometer is less of a problem), whereas the micropulse lidar seems to underestimate CF during the summer season, compared to the other datasets. Based on the close agreement between the ceilometer and the microbase datasets, we decided to average the 3-hourly CF from these two datasets to provide the observed CF used in our analysis. The reader is reminded that the CF observations do not agree, and to some extent this level of disagreement should be viewed as an observational uncertainty (of the order ±15%) that varies with the seasons. This level of accuracy should be borne in mind when specific cloud–radiation parameters are analyzed, and this indicates the critical importance of accurate cloud-fraction observations. Furthermore, it can also be noted from Fig. 2 that there is either no significant diurnal cycle in the cloud fraction at these sites and seasons or that the observation error masks it.
a. Large-scale meteorology
This section gives a brief overview of the model’s ability to reproduce large-scale meteorology at the two observation sites. We do this to confirm that the simulated atmosphere generally follows the observed evolution, allowing cloud–radiation processes to be evaluated against observations in a common thermodynamic phase space. We also make a preliminary analysis of the simulated 2-m temperatures at the two sites, in order to later relate SRB errors to such a key variable.
As a measure of the large-scale synoptic variability, in Fig. 3 we plot the 3-day mean surface pressure and IWV for the model grid box, collocated with each observation site and the same observed quantity. We choose one representative summer and winter season from the 7–8 yr of analysis for each of the observation sites; the other seasons are generally similar. For surface pressure, the synoptic variability is well reproduced by the model at NSA during the summer [June–August (JJA)] and winter (DJF) seasons, with only occasional small biases. For SGP, the variability is well reproduced by the model during the winter. Larger differences are seen at SGP during the summer season, as might be expected when the model atmosphere is less constrained by the LBCs (Lucas-Picher et al. 2008). There is no apparent systematic surface pressure bias in any of the seasons or locations. Once the model is corrected for the altitude difference at SGP, it accurately reproduces the observed amplitude of surface pressure for both sites and seasons. For IWV, the model (without any correction applied, because the effect of the altitude difference derived from maximum surface mixing ratios observed at the site is less than 0.6 kg m−2) reproduces the observed variability better in winter at both sites; this is also likely because of the stronger control by the LBCs in winter.
In Fig. 4, we present the observed and simulated mean diurnal cycle of 2-m temperature, for JJA and DJF seasons at both the SGP and NSA sites. These are averages of over 7 yr for SGP and 8 yr for NSA. At SGP, GEM-LAM has a warm bias of ≈5°C through out the diurnal cycle for JJA, although the actual amplitude of the diurnal cycle is well captured. At NSA, in the summer season a nocturnal warm bias of ≈1°C increases to 3°C during the afternoon period. During the winter season, the model reproduces the diurnal evolution of 2-m temperature at both locations quite accurately. The SGP in JJA temperature error is local to central North America (Markovic et al. 2009) and does not appear to be linked to major circulation errors. A comparison at SGP in JJA between the observations and models shows an underestimate in soil water content of 0.2 m3 m−3 [for an observed summer mean of 0.3 m3 m−3; see Schneider et al. 2003]; a 2-m relative humidity (RH) underestimate from 20% to 35%, along with a clear positive bias in the sensible heat flux of ≈50 W m−2 in the middle of the afternoon; and a commensurate negative bias in the surface latent heat flux (Cook 2007). The exact origin of this significant summer warm and dry bias is difficult to pinpoint and may arise from model shortcomings in any part of the complex range of interactions involving surface and soil processes, the atmospheric water cycle, cloud microphysics, and radiation. In this article, we focus on a part of this interacting chain, namely, the representation of atmospheric radiation and its associated processes. We do this to better understand how errors arise in this area, and thereby focus our attention on key radiation-related parameterizations that require improvement in GEM-LAM. A full determination of the underlying causes of the near-surface warm bias requires a more wide-ranging model evaluation and extensive sensitivity tests, involving all land and atmospheric parameterizations in the model. This is beyond the scope of the present study.
b. Cloud fraction
Figure 5 shows the mean annual cycle of CF at SGP and NSA. The observed CF is an average of the estimates (Long, TSI, and MB for SGP and Vceil and MB for NSA) shown in Fig. 2. We remind the reader that a degree of uncertainty exists regarding the absolute accuracy of the observed CF (of ±15%). GEM-LAM generally underestimates CF, ranging from ≈10% at SGP to ≈20% at NSA in JJA. In the winter season, GEM-LAM overestimates CF at NSA by ≈25%–30%. It is well established that most observational platforms have difficulty in detecting the optically thin clouds that may be quite frequent at NSA in the winter. Wyser and Jones (2005) showed that by filtering modeled clouds to preclude all clouds with an optical thickness of less than 0.5, the resulting model CF was reduced by ≈20%–25% in the winter season over the Arctic. Furthermore, Karlsson et al. (2008) determined that the minimum cloud optical thickness detection limits for the Advanced Very High Resolution Radiometer (AVHRR) satellite are 1.0 and 3.0 for low-level clouds at night and twilight, respectively. Therefore, the NSA winter cloud bias in GEM-LAM should be treated with some caution. Figure 6 shows a normalized frequency distribution of 3-hourly mean CF occurrences at SGP and NSA. Although the general shape of the frequency distribution is well captured by GEM-LAM, the small underestimate in SGP clouds appear mainly because of an overestimate of clear-sky (CF ≤ 10%) occurrences, along with an underestimate of fractional cloud occurrences; this suggests of an inability to simulate weakly forced convective clouds in the summer over SGP. At NSA, the JJA underestimate in the mean CF is more a result of an underestimate of the occurrence of overcast conditions (CF ≥ 90%).
The direct radiative effect of the CF underestimate at SGP in both seasons and at NSA in JJA should be an underestimate of LWD and an overestimate of SWD. However, CF is not the only factor influencing the diurnal cycle of SWD and LWD. The clearest example of this can be seen in the next section, where a model overestimate of LWD is seen during summer at SGP, even though CF is underestimated.
c. All-sky surface radiation fluxes
In this section, we compare simulated and observed mean annual and mean diurnal cycles of SWD and LWD for JJA and DJF, respectively. This analysis is done for all-sky conditions. A more detailed analysis follows in sections 3d and 3e, where CF is used to isolate the separate roles of water vapor or cloud liquid water on surface radiation. Figure 7 shows the mean annual cycle of SWD and LWD at both SGP and NSA. GEM-LAM overestimates SWD during the spring and summer by ≈15–20 W m−2 at SGP, whereas this overestimate is concentrated only in the summer season at NSA; however, it reaches ≈40 W m−2, with an opposite bias of ≈20 W m−2 only 2 months earlier. This is likely caused by a similar behavior in the mean annual cycle of CF at NSA (Fig. 5), which exhibits a positive bias of ≈20% in April and a negative bias of ≈20% in June. LWD is slightly overestimated during summer at SGP (≈10 W m−2) and underestimated by a similar magnitude in winter. At NSA, summer and fall LWD are underestimated by ≈10 W m−2, whereas winter shows a positive LWD bias of a similar magnitude. At NSA, the LWD biases are consistent with the CF biases shown in Fig. 5. The quoted observational accuracy of SWD and LWD are ±10 and ±5 W m−2, respectively.
To better understand the source of errors in the SWD and LWD annual cycles, we begin by constructing mean diurnal cycles of SWD and LWD from both the model and observations. Figure 8 shows the diurnal cycle of the 3-hourly mean SWD and LWD for summer (JJA) and winter (DJF) at both sites. The overestimate of seasonal-mean SWD at both SGP and NSA is clearly associated with a developing SWD overestimate in the middle of the day, with a ≈40 W m−2 maximum overestimate for SGP in JJA, ≈30 W m−2 for SGP in DJF, and ≈50 W m−2 for NSA in JJA.
Positive and negative biases seen in the mean annual cycle for LWD are also visible through the diurnal cycle during JJA and DJF. For SGP in JJA, the overestimate is maximum (≈15 W m−2) in early afternoon and stays positive for the rest of the day. The near-surface positive temperature bias at SGP in JJA, shown in Fig. 4a, will contribute significantly to this positive LWD bias. The difference in near-surface temperature between the model and observations at SGP during the summer season (a warm bias of ≈5°C, as shown in Fig. 4) will lead directly to a ≈20–30 W m−2 difference in LWD, assuming that surface downwelling LWD emanates mainly from near-surface emission and applying a value of 0.73–0.85 for the near-surface atmospheric emissivity (Swinbank 1963; Chen et al. 1991). This LWD overestimate, resulting from near-surface thermal errors, is partially balanced by a cloud underestimate at SGP in JJA (see Fig. 5a), which acts to reduce the total-sky emissivity in GEM-LAM compared to the observations, and through a compensation of LWD errors, reduces the all-sky LWD bias at SGP in JJA to ≈15 W m−2. For SGP in DJF and NSA in JJA, the LWD biases also show a diurnal cycle with a maximum underestimate of ≈12 W m−2 before local noon, whereas for NSA in DJF, the overestimate of 7 W m−2 is constant through the diurnal cycle but is within the observational uncertainty (Shi and Long 2002).
d. Surface radiation fluxes for clear-sky conditions
In this section, we analyze surface radiation fluxes for clear-sky conditions only (CF ≤ 10%). This allows us to compare the representation of LWD and SWD fluxes that are isolated from the confounding effects of either CF or cloud–radiation parameterization errors.
Figure 9 shows the mean diurnal cycle of SWD and LWD for clear-sky conditions, and this should be compared to Fig. 8, which shows the same quantities for all-sky conditions. The overestimate of the total SWD for all-sky conditions is significantly reduced when only clear-sky conditions are considered. For SGP in JJA, the all-sky SWD overestimate of ≈40 W m−2 is reduced to below 20 W m−2 around local noon, whereas, for SGP in DJF, the all-sky SWD overestimate now becomes an underestimate of ≈20 W m−2 around local noon. Thus, the simulated SWD in clear-sky conditions does not appear to be the main cause of the overestimate of the all-sky SWD seen in Fig. 8, in fact the clear-sky errors sometimes act in an opposite sense to the all-sky errors.
For LWD in clear-sky conditions at SGP, the same biases as LWD for all-sky conditions are seen, suggesting that the LWD biases are largely controlled by near-surface temperature errors. Nevertheless, an analysis of near-surface temperatures at SGP in JJA shows that both simulated and observed overcast and all-sky conditions are generally colder than clear-sky situations (not shown). The all-sky LWD is made up of clear-sky, overcast, and partially cloudy contributions. Based on the differences in near-surface temperatures, one might expect the all-sky LWD to be lower than that seen for clear-sky conditions. This is not the case at SGP in JJA, where all-sky LWD is marginally higher than clear-sky LWD, indicating that the presence of clouds increases the total atmosphere emissivity enough to more than balance the reduction in LWD resulting from cooler near-surface temperatures. A similar picture is also seen at SGP in DJF, with cloudy situations contributing to a higher LWD for all-sky conditions than in clear skies, even though the near-surface temperatures are lower in the all-sky situations. For NSA in JJA, the simulated LWD in clear-sky conditions shows an overestimate of the diurnal cycle resulting in a maximum overestimate of ≈15 W m−2 in the late afternoon. This clear-sky error is also likely tied to the near-surface warm bias at NSA in JJA, which peaks in the late afternoon (Fig. 4). In terms of the all-sky LWD at NSA in JJA, the clear-sky overestimate is not seen because of the significant underestimate in CF (peaking at ≈25% in the afternoon, not shown), leading to an overall underestimate of all-sky emissivity that completely offsets the thermal contribution to the clear-sky LWD. The final result is a negative bias in the all-sky LWD at NSA in JJA. For NSA in DJF, the all-sky LWD positive bias of ≈7 W m−2 increases to ≈15–20 W m−2 in clear-sky conditions.
Atmospheric water vapor is one of the principal controls on the surface radiation budget in clear-sky conditions. Figure 10 presents the mean diurnal cycle of IWV. GEM-LAM underestimates the IWV throughout the diurnal cycle for SGP in JJA by ≈3.5 kg m−2 (≈10%). Sensitivity tests with a simple 1D radiative transfer model using standard atmospheres shows that such an underestimate of IWV can cause an overestimate of ≈2–3 W m−2 in SWD and an underestimate of ≈6 W m−2 in LWD. Figure 9 shows that the warm bias in surface temperature for SGP in JJA outweighs the underestimate in IWV, resulting in a positive bias in the clear-sky LWD. For SGP in DJF and NSA for both seasons, GEM-LAM reproduces the observed diurnal cycle of IWV quite accurately, close to the quoted observational uncertainty of ≈0.7 kg m−2 (Turner et al. 2007). These small underestimates of IWV may explain some of the SWD and LWD clear-sky errors in the model.
To verify whether the radiative transfer scheme is able to properly represent the interaction between LWD/SWD and IWV, we examined the observed and simulated covariability plots between IWV and SWD/LWD for clear-sky conditions (not shown). GEM-LAM accurately represents the relationship between increasing IWV and increasing LWD (decreasing SWD). However, offsets (as seen in Fig. 9) are also seen, suggesting that errors likely lie outside of the direct IWV treatment and that thermal errors, aerosols, or trace gas may contribute. For SGP in DJF, one probable cause of the negative SWD bias is an excess of aerosol loading. At this location, the parameterization of Toon and Pollack (1976) gives a broadband aerosol optical depth (AOD) centered at 550 nm of 0.169 year-round. Observations (Shi and Long 2002) show a strong seasonal cycle of AOD at 500 nm from 0.063–0.095 for DJF to 0.170–0.199 for JJA. Offline radiation tests suggest that halving the aerosol loading for typical midlatitude winter conditions increases clear-sky SWD by ≈10–20 W m−2. This comparison could not be done for NSA because of a lack of observations.
e. Surface radiation fluxes for overcast conditions
In this section, we present surface radiation fluxes for overcast conditions, that is, when CF is ≥90%. In doing this, we aim to isolate SRB errors arising solely from cloud microphysics and the overcast radiation parameterization, although clear-sky conditions above and below cloud cover do influence these findings somewhat. This is particularly true for LWD and the subcloud layer temperature errors. Figure 11 shows the mean diurnal cycle of SWD and LWD for overcast conditions. The overestimate seen in Fig. 8 for SWD (all-sky conditions) is now amplified for SGP in JJA (with a maximum overestimate of ≈150 W m−2 in the middle of the day). At SGP in DJF, the overcast SWD error is of a similar magnitude and sign to the all-sky errors (≈30 W m−2). At NSA in JJA, the all-sky SWD positive bias of ≈30 W m−2 (Fig. 8) in the afternoon becomes a negative bias of ≈70 W m−2 when only overcast conditions are considered, suggesting considerable problems in the representation of cloud and solar radiation interaction.
For LWD in overcast conditions, GEM-LAM has a similar but less pronounced bias for SGP in JJA to that seen in Fig. 8 for all-sky conditions, again indicating the balancing effect of the underestimated CF and the overestimated near-surface temperatures in relation to LWD. For all of the other seasons and sites the simulated LWD is close to the observational uncertainty; although, it is noteworthy that small negative biases in LWD (seen in Fig. 8 for SGP in DJF and NSA in JJA) become small positive biases for overcast conditions and vice versa for NSA in DJF.
To better understand the SWD and LWD overcast errors, we directly evaluate the cloud liquid water amounts simulated by the model, as well as analyze the covariability between SWD/LWD and cloud liquid water. Cloud water is an important factor controlling SWD and LWD for overcast conditions, influencing the radiation fluxes in a number of ways. The absolute quantity of cloud water controls both cloud albedo/absorptivity in the solar range and emissivity in the infrared range (Stephens and Webster 1981), although the partitioning into solid and liquid phase is also important, primarily for cloud–SWD interactions (Liou 1992). Finally, the actual droplet/crystal scattering and absorptivity/emissivity characteristics must be accurately represented, generally through a realistic estimate of both the droplet and ice crystal effective radii. The latter are highly parameterized in GEM-LAM, as in many climate models (see section 2a), and are often very difficult to evaluate or even constrain because of a lack of observations.
GEM-LAM underestimates the mean diurnal cycle of LWP for nonprecipitating events by ≈20 and 60 g m−2 for SGP in DJF and NSA in JJA, respectively (Fig. 12). The underestimate of LWP at SGP in DJF is consistent with the overcast SWD positive bias seen in Fig. 11. For NSA in JJA, the negative bias in LWP is definitely not consistent with the negative bias in overcast SWD. Similarly, the large positive SWD bias at SGP in JJA does not appear to be due to a systematic underestimate of LWP. Because cloud albedo is a nonlinear function of LWP (Stephens and Webster 1981; Slingo 1989), it is not sufficient just to analyze the mean diurnal cycle of LWP, it is also necessary to establish how GEM-LAM simulates the range of variability of LWP at a higher time frequency (i.e., the underlying LWP statistics that make up the mean values). In doing this, we may also better understand the physical processes in the model, leading to the underestimate of LWP noted for SGP in DJF and NSA in JJA. To analyze LWP variability, we utilize a normalized frequency histogram approach, as used previously by van Meijgaard and Crewell (2005). We bin all observed and simulate 3-hourly mean LWP and normalize each bin by dividing by the total number of LWP occurrences in either the entire observation or model dataset. This is done for nonprecipitating events [defined as precipitation ≤0.25 mm (3 h)−1], and the histograms are made separately for JJA and DJF at both the NSA and SGP sites. Results are presented in Fig. 13. One should be aware of the logarithmic profile of the ordinate in Fig. 13, which influences the absolute size of LWP errors depending on whether they are at the lower or higher end of the distribution. The quoted uncertainty for the LWP observations is 20–30 g m−2 (Turner et al. 2007); thus, the first bin in our analysis encompassing the range of 0–15 g m−2 is uncertain as to whether it represents clear-sky conditions.
For all cases, GEM-LAM overestimates the frequency of occurrence of LWP in the 0–15 g m−2 range. This is particularly true for NSA in JJA. For SGP in JJA, GEM-LAM also overestimates the frequency of occurrences for LWP ≥ 165 g m−2. In contrast, GEM-LAM underestimates the frequency of occurrence of higher LWP amounts for SGP in DJF (≥45 g m−2), leading to a reduced cloud albedo and excess SWD, as seen in Fig. 11. At NSA, LWP (≥15 g m−2) is underestimated in both seasons, which should lead to an excess SWD in GEM-LAM under overcast conditions; however, this is opposite to what is seen in Fig. 11 (we return to this point later). The histogram statistics are consistent with the mean LWP diurnal cycle (Fig. 12), which presented negative biases in LWP for the SGP in DJF and NSA in JJA cases and a smaller negative bias for NSA in DJF.
One possible cause of an underestimate of LWP is that GEM-LAM reasonably simulates the total cloud water (ice and liquid); however, it has a poor fractional separation of the total water into the two respective phases. Unfortunately, we did not have access to reliable IWP observations to evaluate the simulated IWP. Hence, in Fig. 12, we present the simulated total condensed cloud water (CWP = LWP + IWP; dash–dotted line) for nonprecipitating events. For SGP in DJF and NSA in JJA, inclusion of the simulated IWP does not appear to greatly change the findings that LWP, and now CWP, are underestimated in GEM-LAM. This does not preclude the possibility that IWP values are underestimated in GEM-LAM, but it does show that the total cloud water in the model atmosphere is underestimated for SGP in DJF and NSA in JJA. In contrast, an incorrect separation of CWP into LWP and IWP may be the possible cause of the underestimate of LWP for NSA in DJF.
Another possible cause of an underestimate of LWP is that precipitation removal of cloud water occurs at too low LWP values in the GEM-LAM microphysics scheme. Figure 13 presents a simulated LWP that is filtered using different precipitation thresholds; the original threshold of 0.25 mm (3 h)−1 (red line) is the same threshold used for the observations (blue line). The light-blue line shows LWP occurrences when a threshold of 1 mm (3 h)−1 is used for filtering GEM-LAM results, whereas the pink line shows LWP values in GEM-LAM for all events, irrespective of rain occurrence. At both SGP in DJF and NSA in JJA, the inclusion of light precipitation events results in a distribution of LWP closer to that of the observations, whereas the inclusion of LWP for all cases, including all precipitation events, is now very close to the observed distribution. Inclusion of LWP when the model is precipitating increases the relative occurrences of LWP ≥ 75 g m−2. For NSA in DJF, the inclusion of LWP for all cases of precipitation has an important impact only for classes of LWP ≥ 135 g m−2. These results suggest that GEM-LAM, particularly for SGP in DJF and NSA in JJA, produces precipitation at a too-low threshold of liquid water content (and hence prevents the existence of high-LWP nonprecipitating clouds). Precipitation is an efficient sink for cloud water, and therefore it is likely responsible for the too-efficient removal of liquid water from the simulated clouds. This type of problem is shared by a number of other models (van Meijgaard and Crewell 2005).
Figure 14 shows the 3-hourly accumulated precipitation frequency distribution for model and observations. The overestimate of simulated precipitation in the range of 0.5–4 mm (3 h)−1 for SGP in DJF and NSA in JJA, confirms the tendency for LWP to be too efficiently removed by precipitation and suggests the need to modify the autoconversion term in the model microphysics.
The LWP analysis for SGP in JJA and NSA in JJA (which is accurate for SGP and an underestimate for NSA) shown in Figs. 12 and 13 are not able to explain the overcast SWD errors seen in Fig. 11 (a positive bias at SGP in JJA and a negative bias at NSA in JJA), hence further analysis is required. In Fig. 15, we analyze the covariability between LWP and SWD. Three-hourly mean values of LWP are plotted against 3-hourly mean SWD fluxes for both the model and observations. SWD is normalized by the solar zenith angle, with a maximum angle of 85°.
Figure 15 shows that for a given LWP amount, the simulated SWD for SGP in JJA is systematically higher than that observed, particularly for clouds with LWP values of less than 50 g m−2, with a maximum bias greater than 200 W m−2. This positive bias in the relationship between SWD and LWP directly relates to the positive bias seen in the overcast SWD diurnal cycle (Fig. 11). Furthermore, GEM-LAM slightly overestimates the absolute LWP amounts during the afternoon at SGP in JJA (not shown), which will cancel out some of this positive bias for the afternoon period, resulting in a smaller overcast SWD bias through error compensation. For SGP in DJF, a similar, albeit smaller, positive bias in GEM-LAM SWD can also be seen compared to observations extending to clouds with larger LWP values. Without observed IWP, the cause of this bias cannot be determined and is the subject of future research. At NSA in JJA, GEM-LAM better represents the observed covariability between SWD and LWP, except for an underestimate of ≈50 W m−2 in the range of 30–100 g m−2. This result, combined with the LWP underestimate for NSA in JJA, cannot explain the entire overcast SWD underestimate seen in Fig. 11. NSA in DJF is not shown because of a lack of SWD in LWP observations in overcast conditions.
The SWD–LWP covariability figures are done for nonprecipitating events because of the limitations of the microwave radiometer. In contrast, the SWD diurnal cycle for overcast conditions (Fig. 11) is made for all occurrences of CF ≥ 90%, irrespective of precipitation occurrence. For consistency we applied the same precipitation filter to the overcast SWD diurnal cycle for NSA in JJA. Figure 16 shows that when precipitating cases are removed for both the model and observations (dashed lines), the underestimate in SWD for overcast conditions in GEM-LAM is reduced from ≈70 to ≈25 W m−2. This bias now agrees more closely with the SWD–LWP covariability plot for NSA in JJA (Fig. 15), which shows an underestimate of SWD for clouds with LWP values of 30–100 g m−2. A possible cause of this underestimate of solar transmissivity for optically thin clouds may be that the effective radius calculation is tuned for midlatitude conditions and requires an increased degree of flexibility to represent polar clouds, where the dominant liquid and solid effective radii may be systematically larger than at midlatitude regions. When precipitation is included in the overcast SWD mean diurnal cycle, the negative bias becomes larger, presumably resulting from the inclusion of more cloudy occurrences in the model compared to the observations, because the model overestimates the frequency of occurrence of precipitation (Fig. 14).
Figure 17 shows that GEM-LAM accurately reproduces the rapid increase to saturated emissivity as a function of LWP at SGP during both seasons. The constant positive bias of ≈20 W m−2 seen for LWD in SGP in JJA for all ranges of LWP, again arises from the warm near-surface temperature bias, which influences overcast LWD through clear-sky emission below cloud. For NSA in JJA, GEM-LAM also captures the rapid increase in emissivity in the lower range of LWP and the succeeding saturation. For NSA in DJF, the increase of LWD for low LWP is more visible and GEM-LAM tends to overestimate this by ≈30 W m−2. This overestimate in LWD may also just be a reflection of errors in clear-sky emission below the cloud base, because a similar bias was seen in the LWD in IWV covariability for NSA in DJF (not shown). The near-surface temperature error for NSA in DJF is quite small (Fig. 4) and cannot explain this bias. The LWD error is therefore likely either because of a poor representation of water vapor emission at very low temperatures and dry conditions or as a result of the misrepresentation of the LWD emission for other trace gases or aerosols. Recent observations (Shupe et al. 2006) suggest that the liquid fraction in mixed phase Arctic clouds tends to approach zero at temperatures of −25° to −30°C. If this is the case, then the GEM-LAM parameterization, in order to split total water into liquid and ice fractions, will overestimate the liquid fraction at cold temperature. All other things being equal, this would contribute to an overestimate of LWD from winter clouds at NSA (as seen in the figure).
The general underestimate of LWP (seen in Fig. 12) has little impact on LWD for SGP (in both JJA and DJF), and NSA in JJA in overcast conditions because the simulated clouds, even with a negative bias in LWP, remain at an emissivity of unity. LWP errors do have a larger impact for NSA in DJF, where the amount of LWP is very small and an underestimate of LWP will lead to a negative bias in cloud emissivity. One should bear in mind that clouds at NSA in DJF likely contain a significant amount of ice water, and because of observation problems, we are unable to gauge the contribution of IWP errors to LWD, which are likely largest for NSA in DJF.
4. Discussion and conclusions
In this paper, we have evaluated the cloud–radiation interaction in the limited-area version of GEM, concentrating on the total surface radiation fluxes, as well as evaluating the various process-level terms that contribute to errors in the total surface radiation flux. This has been done using observational data from two ARM sites (SGP and NSA) with radically different climates. The large-scale meteorology (defined here by a 3-day mean surface pressure and IWV variability) is accurately simulated at both sites, allowing a comparison of simulated cloud–radiation to point observations in a common thermodynamic/dynamic parameter range.
Comparison of the mean annual cycle and mean diurnal cycle of SWD showed that GEM-LAM generally overestimates SWD at all sites. GEM-LAM overestimates LWD at SGP in the summer, whereas it underestimates LWD in the winter. For NSA, the model has relatively small LWD biases that lie within the range of the observational uncertainty. Different processes contribute to these biases: CF errors influence the all-sky radiation, whereas aerosols and IWV are the primary terms controlling clear-sky solar radiation, and both IWV and near-surface temperatures influence the clear-sky LWD. In overcast conditions, LWP amounts and their interaction with radiation is the primary variable, particularly with respect to SWD.
GEM-LAM generally underestimates CF, except at NSA for the winter where it appears to have a positive bias, although the latter error should be treated with caution because of known observational limits with respect to winter Arctic clouds. The underestimate in CF is a strong contributing factor to the overestimate of all-sky SWD and the underestimate of LWD. Nevertheless, we have shown that errors in clear-sky and overcast radiation fluxes also influence the all-sky radiation and sometimes result in error cancellation. A clear example of this is the overestimate of SWD for NSA in JJA, which results from an underestimate of simulated CF, a correctly simulated clear-sky SWD, and an underestimate of simulated overcast SWD (i.e., when clouds are present they are too reflective), partially balancing the CF underestimate. Moreover, the excess clear-sky LWD is partially balanced with respect to the all-sky LWD by an underestimated CF, leading to a negative bias in simulated all-sky emissivity at NSA in JJA.
The general underestimate of LWP in GEM-LAM only influences SWD in overcast conditions for SGP in DJF and LWD for overcast conditions at NSA in DJF. An overestimate of the occurrence of light precipitation in GEM-LAM appears to be the main cause of the negative bias in LWP at SGP in DJF and NSA in JJA, pointing to a need to improve microphysical conversion processes in the model.
GEM-LAM represents the covariability between IWV and LWD in clear-sky conditions quite accurately (not shown), with a positive bias in LWD linked to a near-surface warm bias at SGP in JJA. For overcast conditions at SGP in JJA, the positive bias in LWD, as a function of LWP (Fig. 17a), also mainly results from the near-surface warm bias. At NSA in DJF, there is no obvious thermal explanation for the clear-sky and all-sky LWD biases, suggesting instead a possible erroneous contribution from aerosols or trace gases. With respect to the covariability of SWD–LWP, biases at SGP for optically thin clouds (Figs. 15a,b) seem to be the main cause of the large positive bias in the mean diurnal cycle of overcast SWD. As yet, we do not have a clear explanation for this bias, which may be related to an incorrect separation of LWP and IWP to an incorrect treatment of cloud effective radii or cloud optics, or even a sampling problem between observations and model. GEM-LAM uses a plane-parallel radiative transfer scheme, whereas a more sophisticated 3D treatment of clouds might improve the treatment of cloud–solar radiation interaction.
Tests on the parameterization of these variables are needed to improve the covariability of SWD as a function of LWP. Moreover, an opposite bias in the SWD–LWP relationship was seen for optically thin clouds at NSA (surface SWD values are too small in the model for a given LWP value, when LWP < 30 g m−2), suggesting that a more flexible approach to the parameterization of cloud effective radii is required, which can more faithfully capture the differing underlying controls on cloud droplet/crystal size at various geographic locations (e.g., background aerosols loadings). Similar conclusions were made by Wyser et al. (2008) in their analysis of eight RCM-simulated cloud–radiation interactions over the Arctic Surface Heat Budget of the Arctic Ocean (SHEBA) site.
In summary, the primary SRB error in GEM-LAM is the excess solar radiation reaching the surface in the middle of the day. This problem is common to SGP in both seasons and to NSA in the summer. At SGP, this all-sky error is partly caused by an underestimate of cloud amounts, which is compounded by a large overestimate of SWD in the middle of the day during overcast conditions. The latter error is linked to an overestimate of solar transmissivity in optically thin clouds (LWP < 50 g m−2) at SGP during both seasons. Without complementary information on observed IWP, the absolute cause of this error remains to be determined. In contrast, at NSA in JJA the impact of a CF underestimate is partially offset in terms of the all-sky SWD by the cloud solar transmissivity being too low (i.e., too little SWD reaching the surface in overcast conditions). This error is partly explained by an underestimate of solar transmissivity in optically thin clouds (LWP < 30 g m−2), and it is amplified in comparison to observations of overcast radiation by the inclusion of precipitating events because the model overestimates the frequency of occurrence of precipitation. An underestimate of LWP does not contribute significantly to errors in the overcast LWD at SGP or NSA during the summer because the simulated clouds still having an emissivity of unity, even with a negative LWP bias.
Although the emphasis of this paper has been on surface radiation fluxes, a reasonable question to ask is how much of the near-surface warm bias seen in the model at SGP is a direct result of the underestimate of the cloud fraction and the excess SWD in overcast conditions? Because of the strong coupling and multiple interactions between surface–soil processes, boundary layer and convective mixing, atmospheric moisture profiles, and subsequent cloud and radiation, it is difficult to uniquely determine the underlying cause of such an error. There is little doubt that both an underestimate of the cloud fraction and an overestimate of the overcast solar transmissivity will contribute to the development of a warm bias. It is not straightforward, to be sure, that these errors cause the warm, dry bias or rather are just a response to an initial surface error, which is then amplified by the response of clouds and radiation. One possible cause of the warm bias at SGP in JJA might be an underestimate of the surface albedo. A comparison of observed (Shi and Long 2002) and simulated surface albedo at SGP and NSA (not shown) shows that, in general, GEM-LAM has a realistic value of the surface albedo at both sites. The model does fail to accurately represent the observed diurnal cycle of surface albedo, which is linked to the changing solar zenith angle. Nevertheless, the deviations of albedo at SGP in JJA are certainly too small (ranging from 0. to −0.02) to explain the warm bias seen in Fig. 4.
With respect to the cloud fraction underestimate, at this stage, it is not possible to uniquely determine the source of this error, which may lie at any point in the range of interactions outlined above. With respect to the SWD errors in overcast conditions, because these occur solely for overcast skies (CF > 90%) for both the model and observations and there appears to be no systematic underestimate of LWP in the model at SGP in JJA, we feel confident that the fundamental source of this SWD bias does lie in the parameterization of the cloud–solar radiation interaction. Such a positive SWD bias will drive a response at the surface of warming and drying. Without extensive model sensitivity tests, we are unable to determine whether this is the primary error in GEM-LAM that is responsible for the development of the surface warm, dry bias, or whether other errors also play a fundamental role. This question is being addressed with a new set of GEM-LAM simulations, for the same domain and period, where the Interactions between Soil, Biosphere, and Atmosphere (ISBA) land surface scheme has been replaced by the Canadian Land Surface (CLASS) scheme (Verseghy 2000), which employs a more physically based treatment of the land surface.
Clearly, there is strong interaction between all of the components influencing the simulated SRB with frequent error compensation. Nevertheless, some fundamental weaknesses in the cloud and clear-sky radiation parameterizations in GEM-LAM have been identified. Improving these processes will improve the overall physical realism of the model. We suggest that the evaluation approach outlined in this article be used in a more general evaluation of cloud–radiation processes, in a wider number of RCMs. Furthermore, the procedure could be extended to other ARM sites with similar observational availability, however, from different climate regimes (e.g., the ARM Tropical Western Pacific site).
The data were obtained from the Atmospheric Radiation Measurement (ARM) Program sponsored by the U. S. Department of Energy, Office of Science, Office of Biological and Environmental Research, Environmental Sciences Division. This study was supported by funding from the Mathematics of Information Technology and Complex Systems (MITACS; Grant 61851). The authors wish to thank Dr. Bernard Bugas and Dr. Ayrton Zadra for their helpful discussions and Ms. Katja Winger for all of her technical contributions.
Corresponding author address: Danahé Paquin-Ricard, UQAM, Département des Sciences de la Terre et de l’atmosphère, 201 Président Kennedy, Montreal QC H2X 3Y7, Canada. Email: email@example.com