1. Introduction
The general circulation of the atmosphere is driven by differential radiative heating and cooling. To model atmospheric processes correctly, it is necessary to adequately represent radiative processes. One measure of success commonly used in general circulation models is the radiative balance at the top of the atmosphere (e.g., Hartmann et al. 1986). However, compensating errors may combine to produce acceptable top-of-the-atmosphere radiative fluxes, while surface conditions or heating profiles remain incorrect. In the context of numerical weather prediction in particular, surface temperature, cloud cover, and associated radiative processes need to be accurately represented on short time scales, not only in the longer-term mean. Thus, it is necessary to investigate and address biases and compensating errors in the surface radiation, particularly the downward radiative fluxes, which are directly linked to cloud amount and properties. This study will focus on the shortwave component as previous evaluations have suggested a bias in the surface irradiance over land in the European Centre for Medium-Range Weather Forecasts (ECMWF) model.
In 2002, a systematic assessment of surface radiation at the Atmospheric Radiation Measurement (ARM) Southern Great Plains (SGP) site found that the ECMWF model overestimates surface irradiance under all conditions, even clear sky (Morcrette et al. 2008). The shortwave radiation scheme in the model has since been updated with the McRad package (Morcrette et al. 2008), which has significantly improved clear-sky shortwave radiation (Iacono et al. 2008). Cheinet et al. (2005) investigated a month-long summer period and found that the ECMWF model appeared to be missing fair-weather shallow cumulus clouds that are frequently observed at the site and suggested that this lack of cloud cover may contribute to the shortwave bias. Subsequently, much effort was spent on improving and evaluating the model’s representation of boundary layer clouds (Köhler et al. 2011; Ahlgrimm and Köhler 2010). Yet, a model bias in the surface irradiance persists at the SGP location. While previous results were based on limited time periods (a few months; Morcrette 2002; Cheinet et al. 2005), this study takes advantage of the long record of available observations from the SGP site to systematically assess how much the fair-weather cumulus regime contributes to the surface irradiance bias in a recent operational cycle of the ECMWF model (CY36R4). Furthermore, substantial improvements have been made to the ECMWF model since the earlier studies (see Bechtold et al. 2008; Jung et al. 2010), which make it timely to revisit the issue of shortwave biases.
The concept of compositing by meteorological regime or cloud type in order to establish a link between space or time mean model biases and model parameterizations has been explored previously (Tselioudis and Jakob 2002; Jakob 2003). Kollias et al. (2007) propose a classification based purely on the observed cloud and precipitation fields at the SGP site. The usefulness of the resulting cloud climatology for model evaluation is shown in Tselioudis and Kollias (2007). The authors conclude that while the ECMWF model produces the correct amount of boundary layer clouds in the annual mean, this is the result of compensating errors with an underestimate of wintertime boundary layer cloud amount and an overestimate in summer. The model also lacks midlevel clouds, particularly in multilayer situations.
The first aim of this study is to conclusively assess the contribution of the fair-weather cumulus regime to the ECMWF model’s surface irradiance bias. Unlike previous studies, a longer time series will be used to give more confidence in the results. Second, the contribution to the radiation bias from other cloud types will be examined using a simplified cloud classification similar to that in Kollias et al. (2007). Rather than comparing cloud amount and frequency of various cloud types, samples of a given cloud type will be ranked by their contribution to the radiation bias.
The observations and model data used in this study are described in section 2 while section 3 describes the model’s surface irradiance bias at the ARM SGP site. In section 4, the impact of fair-weather cumulus clouds on surface radiation in the ECMWF model is assessed based on the composite of 146 days with observed fair-weather cumulus clouds (Zhang and Klein 2010) at ARM SGP. Section 5 describes the use of a cloud classification method to identify the regimes that contribute most strongly to the model bias. Section 6 provides the conclusions.
2. Observations and model data
Observational data for clouds, radiation, and meteorological state at the SGP site are available in an easy-to-use and consistent format in the Climate Modeling Best Estimate (CMBE) product (Xie et al. 2010). This product includes hourly averaged cloud fraction profiles from the Active Remotely-Sensed Cloud Locations (ARSCL) product based on the millimeter wavelength cloud radar (MMCR) and micropulse lidar (MPL) retrievals, as well as surface radiation measurements. The observational record used in this study spans the years 1997–2009. In this period, 146 days with fair-weather cumulus are identified through a combination of an automatic selection algorithm based on the observed cloud fraction profile in combination with a manual screening of Total Sky Imager pictures. This method was first applied by Berg and Kassianov (2008), then adapted by Zhang and Klein (2010) to exclude days with precipitation within a 50-km radius. The time series used here was kindly provided by Zhang and Klein and is an extension to the record used in their 2010 publication with additional screening of satellite images to exclude days influenced by weather systems in the vicinity.
ECMWF model data from the operational forecasts are stored hourly for the grid box centered nearest to the ARM SGP site, as well as averaged over the four grid boxes nearest the location. This provides an easily accessible long-term record of model data for comparison with observations. These data are used in sections 3 and 5. A drawback of this record is that the frequent upgrades to the operational forecast model at ECMWF result in a time series from a variety of model cycles. It would be impractical to rerun the complete span of time with the most recent model version to produce a consistent record. However, a recent cycle (CY36R4) was rerun for the 146 shallow cumulus days with 91 vertical levels and horizontal grid resolution of 40 km (triangular truncation at wavenumber 511 to the model’s spherical harmonics). The model is initialized at 1200 UTC on the previous day and output from forecast hours 18 to 41 (approximately 0–23 h local time) at the grid box centered nearest the SGP site are used for comparison with observations. These data are used in section 4.
The cloud fraction profile from the model is the instantaneous cloud fraction on each model level from the model grid box nearest the ARM site, or averaged over the four nearest grid boxes. The ARSCL profile on the other hand is derived by averaging observations from vertically pointing active sensors operating at high temporal frequency over an hour-long period. Thus, the cloud fraction representative of the area associated with a model grid box is compared to a temporal average at one point. This is commonly done (e.g., Illingworth et al. 2007), yet adds an element of uncertainty to the comparison. The analysis of the operational data record (sections 3 and 5) was performed with both the nearest-box and the four-box average data records. While individual hourly samples differ slightly, statistics for the long-term records are very similar, and the choice of record does not affect the conclusions drawn from this study. Only one set of figures (for the nearest gridbox record) will be shown.
The ARSCL retrieval cannot always distinguish cloud water species from precipitation, particularly in the ice phase, such that the observed “cloud fraction” would be more accurately titled a “hydrometeor fraction.” To take the contribution from precipitation to the model’s hydrometeor fraction into account, a threshold (10−5 kg m−2 s−1) is applied to the model’s precipitation fluxes. Cloud overlap is calculated offline using the generalized overlap assumption (Räisänen et al. 2004) splitting the grid box into 112 subcolumns and assuming a decorrelation length of 2.13 km for cloud fraction. This is consistent with the subcolumn approach and overlap assumptions in the model’s shortwave radiation routine. Based on ground and satellite observations, Shonk and Hogan (2010) suggest the use of a latitudinally dependent decorrelation length for cloud fraction. The value of this model parameter (invariant throughout the year) falls within the range observed by Mace and Benson-Troth (2002) for the fall and winter seasons at ARM SGP. After cloud overlap has been determined, any precipitation fluxes above the threshold are maximally overlapped with the cloud.
This new hydrometeor fraction is then used to assign cloud types in the model. Ideally, a full radar/lidar forward model should be run on the modeled clouds to estimate which part of the precipitation contributes to the detectable hydrometeor fraction. However, the hydrometeor profile here is used only to determine the base and vertical extent of cloud layers by searching for the lowest and highest level where the cloud fraction exceeds a threshold (2%), and results are not very sensitive to the threshold values chosen.
3. The surface irradiance bias
Shown in Fig. 1 is the diurnal cycle composite of surface irradiance [or downward surface shortwave radiation (SWDN)] from the multiyear operational model record (2004–09) for all-sky conditions, together with the observed shortwave flux from the CMBE product. In the multiyear composite, the maximum bias between modeled and observed radiation is around 50 W m−2 at noon, or 23 W m−2 averaged over all samples and times. Only daytime samples, as identified by model shortwave radiation above 1 W m−2, and samples with good-quality observed shortwave radiation and cloud profile are included. It is the origin of this multiyear mean bias that this study is investigating.
Multiyear (2004–09) all-sky diurnal composite of surface irradiance at the ARM SGP site. The model is shown in black, observations from the CMBE product in gray. Only daytime samples (modeled SWDN exceeding 1 W m−2) with good-quality coincident observations are included.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Multiyear (2004–09) all-sky diurnal composite of surface irradiance at the ARM SGP site. The model is shown in black, observations from the CMBE product in gray. Only daytime samples (modeled SWDN exceeding 1 W m−2) with good-quality coincident observations are included.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Multiyear (2004–09) all-sky diurnal composite of surface irradiance at the ARM SGP site. The model is shown in black, observations from the CMBE product in gray. Only daytime samples (modeled SWDN exceeding 1 W m−2) with good-quality coincident observations are included.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Shown in Fig. 2 are scatterplots of the observed and modeled surface irradiance for clear-sky situations between January 2005 and December 2009. The new shortwave radiation package “McRad” went into operations on 5 June 2007 (Morcrette et al. 2008). The left panel shows results from the operational model record previous to the shortwave radiation update, while the right panel shows results following the update. As the time periods before and after the update are similar, samples sizes are also comparable. The mean surface irradiance bias for all clear-sky samples dropped from approximately −11.7 to −3.2 W m−2. A clear-sky sample is identified as an hour-long period over which the MMCR and MPL-derived total cloud cover indicates less than 0.1% cloud cover in conjunction with a model total cloud cover of less than 0.1%. The cloud fraction threshold chosen here (0.1%) for the identification of clear-sky samples is stricter than the one used above (2%) to identify cloudy layer base and top heights. This is a deliberate choice to avoid contamination of the clear-sky assessment by samples with very low cloud fractions, while requiring a more substantial amount of cloud to mark the base or top of a cloud layer. The clear-sky bias shown in Fig. 2 is opposite in sign to the all-sky bias in Fig. 1. In addition, the all-sky bias from 2008/09 only (after the radiation upgrade) has similar magnitude to that from the pre-McRad period. Hence, the positive all-sky bias must be related primarily to cloudy conditions.
Observed vs modeled hourly average clear-sky surface irradiance. (a) For the period from January 2005 through May 2007 with old shortwave radiation package. (b) For the period from June 2007 through December 2009 with new McRad radiation package. Linear fit in gray.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Observed vs modeled hourly average clear-sky surface irradiance. (a) For the period from January 2005 through May 2007 with old shortwave radiation package. (b) For the period from June 2007 through December 2009 with new McRad radiation package. Linear fit in gray.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Observed vs modeled hourly average clear-sky surface irradiance. (a) For the period from January 2005 through May 2007 with old shortwave radiation package. (b) For the period from June 2007 through December 2009 with new McRad radiation package. Linear fit in gray.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
4. The fair-weather cumulus regime
A diurnal composite of surface irradiance from observations and the model for the 146 fair-weather cumulus days is shown in Fig. 3a. The model’s clear-sky SWDN is shown for reference as a dashed line. Figure 3b shows the modeled and observed shortwave cloud forcing, in reference to the modeled clear-sky radiation, while Fig. 3c shows the model’s composite radiation bias. In the diurnal cycle composite, the observed shortwave cloud forcing peaks around 90 W m−2. The model is able to reproduce this peak well. A small bias in the model’s SWDN is evident particularly in the morning and evening with changing sign throughout the day. Early in the day, too much of the modeled shortwave radiation reaches the ground, while SWDN is underestimated in the afternoon. The bias is on the order of 10–15 W m−2, or about a fifth of the magnitude of the cloud’s impact in reducing incoming shortwave radiation. The measurement uncertainty of a single observation (including uncertainties associated with radiometer calibration and measurement system installation, operation, and maintenance) is approximately 10 W m−2 (or ±6%; Stoffel 2005). Averaging over many observations will minimize the random error component, thus the observation error for the diurnal composite will likely be smaller than the estimate above.
(a) Surface irradiance at ARM SGP site, composited over 146 fair-weather cumulus days from (gray) observations and (black) model. Also shown is (black dashed) composite model clear-sky surface irradiance. (b) Surface shortwave cloud forcing from (gray) observations and (black) model with respect to model clear-sky radiation. (c) Composite surface irradiance bias (model − obs).
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
(a) Surface irradiance at ARM SGP site, composited over 146 fair-weather cumulus days from (gray) observations and (black) model. Also shown is (black dashed) composite model clear-sky surface irradiance. (b) Surface shortwave cloud forcing from (gray) observations and (black) model with respect to model clear-sky radiation. (c) Composite surface irradiance bias (model − obs).
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
(a) Surface irradiance at ARM SGP site, composited over 146 fair-weather cumulus days from (gray) observations and (black) model. Also shown is (black dashed) composite model clear-sky surface irradiance. (b) Surface shortwave cloud forcing from (gray) observations and (black) model with respect to model clear-sky radiation. (c) Composite surface irradiance bias (model − obs).
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
A diurnal composite of observed and modeled cloud fraction for the 146 days is shown in Fig. 4. Qualitatively, the model captures the fair-weather cumulus clouds well. But despite the reasonable cloud forcing seen in Fig. 3b, the model appears to underestimate the low cloud fraction by about half. The modeled low clouds’ development begins about an hour later than observed and clouds persist longer throughout the afternoon. This corresponds well with the changing sign of the SWDN bias in Fig. 3c, which suggests that there is some information in the bias signal.
Diurnal composite of (a) observed and (b) modeled cloud fraction for 146 fair-weather cumulus days at ARM SGP.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Diurnal composite of (a) observed and (b) modeled cloud fraction for 146 fair-weather cumulus days at ARM SGP.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Diurnal composite of (a) observed and (b) modeled cloud fraction for 146 fair-weather cumulus days at ARM SGP.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
In Fig. 5, the projected total cloud cover (TCC) is shown in Fig. 5a as solid lines. The dashed lines in the same figure show the composite of a subset of 50 days without high clouds during the day (0900 to 1700 local time), such that the projected total cloud cover is coming solely from low clouds. The same underestimate of cloud fraction as seen in Fig. 4 is evident in the total cloud cover as well. The subset of days without upper cloudiness shows that this discrepancy is linked to the fair-weather clouds, rather than to differences in upper-level clouds.
Diurnal composites of (left) total cloud cover, (middle) cloud amount when present, and (right) cloud frequency of occurrence. Model data in black and CMBE observations in gray. Dashed lines show results for a subset of 50 fair-weather cumulus days without observed high clouds above. A threshold of 2% total cloud cover was chosen to determine whether a sample (1-h period) is cloudy or clear.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Diurnal composites of (left) total cloud cover, (middle) cloud amount when present, and (right) cloud frequency of occurrence. Model data in black and CMBE observations in gray. Dashed lines show results for a subset of 50 fair-weather cumulus days without observed high clouds above. A threshold of 2% total cloud cover was chosen to determine whether a sample (1-h period) is cloudy or clear.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Diurnal composites of (left) total cloud cover, (middle) cloud amount when present, and (right) cloud frequency of occurrence. Model data in black and CMBE observations in gray. Dashed lines show results for a subset of 50 fair-weather cumulus days without observed high clouds above. A threshold of 2% total cloud cover was chosen to determine whether a sample (1-h period) is cloudy or clear.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
The middle panel of Fig. 5 shows the cloud amount-when-present (AWP) for hours with cloud cover exceeding 2%, while the right panel shows the frequency with which clouds occur. Evidently, the discrepancy in the composite cloud fraction between model and observations stems primarily from an underestimate of cloud occurrence in the model. This is confirmed by a visual inspection of the model’s cloud cover whereby 17% or 25 out of the 146 days are free of low clouds. These cloud-free days must be compensated for by cloudy days with excessive cloud forcing in order to produce a reasonable composite SW cloud forcing.
Observations of liquid water path (LWP) at the ARM site are available from a two-channel microwave radiometer retrieval (MWRRET; Turner et al. 2007). Shown in Fig. 6a (gray curve) is the normalized distribution of LWP for samples retrieved every 20 s between 0730 and 1830 local time for the fair-weather cumulus days. Using the product’s quality control flags, clear or bad quality samples have been removed. Despite the screening, there are still many samples with LWP below zero. Based on the retrieval’s uncertainty of approximately 20 g m−2, the frequency distribution includes only samples with values above 20 g m−2. Shown in black is the corresponding curve from the ECMWF model. To show the in-cloud liquid water path as the shortwave radiation scheme would see it, the model’s overlap routine is run offline to produce subcolumn profiles of in-cloud liquid water contents (LWC) and LWP. A function of the form suggested in Eq. (1) of Räisänen et al. (2005) is applied to the in-cloud liquid water contents on subcolumns to represent inhomogeneity within the cloud. The shape of this function is invariant, and its effect is to add variance (i.e., increase sample numbers with very low or high LWP while decreasing sample numbers with moderate LWP). It is evident that the model produces relatively fewer samples with low LWP (<100 g m−2) than observed, but overestimates the occurrence of mid- and high LWP. This is consistent with the overestimated shortwave cloud forcing of the model’s fair-weather cumulus clouds. The impact of the inhomogeneity function is positive in this case (vs assuming homogeneous in-cloud LWC), as it increases sample numbers with low LWP. Using the offline cloud overlap routine, we can assess the impact of changes in the model’s cloud liquid water content on the in-cloud LWP. If the model cloud liquid water is reduced by a half to a third, the fit to the observed LWP distribution would significantly improve.
(a) Normalized frequency distribution of (gray) observed and (black) modeled in-cloud liquid water path for 146 fair-weather cumulus days at ARM SGP. Only samples with LWP above 20 g m−2 are included. (b) Normalized frequency distribution of (gray) observed and (black) modeled liquid effective radius for fair-weather cumulus days.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
(a) Normalized frequency distribution of (gray) observed and (black) modeled in-cloud liquid water path for 146 fair-weather cumulus days at ARM SGP. Only samples with LWP above 20 g m−2 are included. (b) Normalized frequency distribution of (gray) observed and (black) modeled liquid effective radius for fair-weather cumulus days.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
(a) Normalized frequency distribution of (gray) observed and (black) modeled in-cloud liquid water path for 146 fair-weather cumulus days at ARM SGP. Only samples with LWP above 20 g m−2 are included. (b) Normalized frequency distribution of (gray) observed and (black) modeled liquid effective radius for fair-weather cumulus days.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
The Microbase product combines cloud radar, microwave radiometer, and radiosonde observations to retrieve cloud microphysical properties (Dunn et al. 2011). Keeping in mind that this algorithm, using LWP as input, is subject to the same uncertainty as discussed above, it appears that the model overestimates the liquid effective radius of fair-weather cumulus clouds by a factor of 3 (Fig. 6b). The observations suggest that effective radii around 4 μm are most common, while the model produces radii of 11–12 μm. The model’s secondary peak at 4 μm is artificial and a result of an imposed lower limit in the radiation code. The model’s overestimate of effective radius would act to lessen the cloud’s shortwave forcing, thus it is most likely the overestimated liquid water path that is the primary reason for the excessive shortwave cloud forcing when cloud is present.
In summary, the fair-weather cumulus regime, as represented by the 146 days investigated, contributes only weakly to the multiyear mean all-sky bias in the model’s surface irradiance. However, this is achieved through compensating errors: cloud-free days compensate for days with good cloud fraction, but overestimated cloud reflectance. The model’s in-cloud liquid water path is too large in these clouds and likely the main cause for the radiative bias. Better agreement might be expected if the shallow convection scheme were triggered more frequently, but the resulting clouds contained less water.
5. Cloud classification to determine source of shortwave bias
The fair-weather cumulus regime was found to contribute only weakly to the overall surface irradiance bias, so there is still a question of what other cloud regimes may be contributing to the multiyear bias, or even if any individual regimes may be identified as contributing more to the bias than others. To explore this question, each hourly cloud fraction profile between 2004 and 2009 is classified by cloud type. Our approach is similar to the classification described and used in Tselioudis and Kollias (2007) and Kollias et al. (2007). However, instead of comparing occurrence and cloud fraction profile associated with each cloud type, the SWDN bias associated with each paired sample (consisting of an observed and coincident modeled cloud profile) is examined. The classification is performed following these steps:
A total of 23 906 hourly samples, consisting of a modeled and coincident observed cloud fraction profile and the associated SWDN bias, are available over the 6-yr period, 2004–09.
Cloud types are assigned independently to each cloud fraction profile (modeled and observed) based on cloud-base height and depth of cloud. Cloud boundaries are determined for a 2% cloud fraction threshold. Table 1 lists cloud-base and thickness limits for the six cloud types considered, which may exist in combination (e.g., low clouds in combination with high clouds in the same profile).
The matched samples are labeled based on the combination of observed and modeled cloud type (e.g., model low cloud; observed low and high cloud).
Samples with the same cloud-type combination are grouped to form a category.
The SWDN biases for all samples within a category are added to form the accumulated (net) SWDN bias for this category.
The SWDN biases from all samples with negative bias within a category are added to form the negative accumulated SWDN bias for that category. This step is repeated for samples with positive bias to calculate the positive accumulated SWDN bias.
- Using the accumulated SWDN bias for all samples as a benchmark, the contribution of each category to this bias is considered by comparing the category’s net SWDN bias to the all-sample bias. As the accumulated bias from all 23 906 samples is a very large number, it has been scaled by the mean SWDN bias of 22.78 W m−2 in Fig. 7 and Table 2:
This scaled quantity expresses the contribution of a category (in W m−2) to the multiyear mean SWDN bias of 22.78 W m−2.
A shortlist of categories that contribute at least 2% to the all-sample bias is created. This list is shown in Table 2. Categories with similar cloud types have been grouped together.
Cloud-base and -thickness criteria for classification of cloud layers in vertical cloud fraction profiles. Note that heights are measured above mean sea level (i.e., the cloud-base height thresholds above are 318 m lower when considering height above ground).
The accumulated contribution to the multiyear SWDN bias is shown, scaled by the mean bias of 22.78 W m−2. The thick black bar shows the net contribution from all daytime samples, as well as from the “deep,” “low,” and “clear” groups, as described in the text. The relative contribution to the mean bias is shown in percent. The thin bar shows the extent to which samples with negative and positive bias contribute to the net for each group. The number of samples within each of the groups is given in parentheses. The sample occurrence for each group throughout the year is shown as histograms in the left half of the figure.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
The accumulated contribution to the multiyear SWDN bias is shown, scaled by the mean bias of 22.78 W m−2. The thick black bar shows the net contribution from all daytime samples, as well as from the “deep,” “low,” and “clear” groups, as described in the text. The relative contribution to the mean bias is shown in percent. The thin bar shows the extent to which samples with negative and positive bias contribute to the net for each group. The number of samples within each of the groups is given in parentheses. The sample occurrence for each group throughout the year is shown as histograms in the left half of the figure.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
The accumulated contribution to the multiyear SWDN bias is shown, scaled by the mean bias of 22.78 W m−2. The thick black bar shows the net contribution from all daytime samples, as well as from the “deep,” “low,” and “clear” groups, as described in the text. The relative contribution to the mean bias is shown in percent. The thin bar shows the extent to which samples with negative and positive bias contribute to the net for each group. The number of samples within each of the groups is given in parentheses. The sample occurrence for each group throughout the year is shown as histograms in the left half of the figure.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Accumulated SWDN bias from various cloud-type combinations identified in the modeled and observed cloud fraction profiles. Values are scaled by the mean SWDN bias over all samples (i.e., the accumulated SWDN bias of a category is divided by the total accumulated bias over all samples and multiplied by the mean bias over all samples; 22.78 W m−2).
To make rapid progress in reducing the surface irradiance bias, it is useful to focus first on those categories that are either very common, or where individual samples have large biases and where samples biases have consistently the same sign. The accumulated SWDN bias will highlight these categories.
Several groups of categories stand out in the short list compiled in Table 2. As previously discussed in section 3, clear-sky samples contribute to a negative bias. While the magnitude of this bias is small, the number of samples in this category makes the contribution from this group significant for partially balancing the positive bias from cloudy situations. Clouds with large vertical extent, including combinations of deep cloud, a thick cloud originating at midlevel (with or without low cloud beneath), contribute strongly to the positive multiyear mean bias (the “deep group”). Last, the “low cloud group,” consisting of various combinations of observed and modeled low cloud or clear cases, is a strong contributor to the positive bias. Other categories involving thin midlevel clouds also make the shortlist, but are less robust. The classification of thin midlevel clouds is more sensitive to the treatment of precipitation in the hydrometeor fraction profiles.
Figure 7 (top row) shows the net, negative, and positive contributions to the bias from all samples. Positive biases outweigh negative biases by a ratio of about 2:1, resulting in a net positive accumulated bias. The corresponding numbers can be found in the bottom row of Table 2. Also shown in the figure is the distribution of samples throughout the year. For the “all samples” category, the longer length of day during summer is reflected in increased sample numbers during this season. Shown beneath in the figure are contributions to the bias from the deep, low, and clear groups. The deep and low groups each explain just over a quarter of the accumulated bias. The remaining bias listed as “residual” in Table 2 is due to clouds that fall into a multitude of categories. With the classification used here, it is difficult to identify a systematic link between these clouds and the remaining bias. In this study, the focus is on the contribution from the low clouds. Unlike samples from the deep group, which are systematically biased positive, the low group contains samples with both positive and negative biases. Further separation of this group into low cloud with broken (<90%) and overcast (≥90%) total cloud cover reveals that in overcast situations the model is systematically biased positive (Fig. 8e), while for broken cloud cover the bias is predominantly negative (Fig. 8c). The model frequently fails to produce overcast conditions when observed, instead producing broken clouds (Fig. 8f). Often, the model remains cloud free when low clouds are observed (Fig. 8a). The opposite case (i.e., clouds in the model when none are observed; Fig. 8b) is less common. Not surprisingly, the model’s surface irradiance exceeds observed values when clouds are absent (Fig. 8a) or fraction underestimated (Fig. 8f). However, even when both model and observations agree on overcast conditions, a positive bias remains (Fig. 8e). When model and observations agree on broken cloud cover, the bias is negative (Fig. 8c), consistent with the conclusions drawn from the fair-weather cumulus days. Indeed, the majority of samples with broken clouds are from the summer season.
Accumulated contribution to the multiyear mean bias from samples with observed and modeled low clouds. Notation as in Fig. 7. The first two bars show misses (model clear, observed low cloud) and false alarms (modeled low cloud, observed clear). The following four bars show contributions from cases with coincident observed and modeled low cloud, but sorted according to broken cloud cover (<90% total cloud cover) or overcast conditions (≥90% total cloud cover).
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Accumulated contribution to the multiyear mean bias from samples with observed and modeled low clouds. Notation as in Fig. 7. The first two bars show misses (model clear, observed low cloud) and false alarms (modeled low cloud, observed clear). The following four bars show contributions from cases with coincident observed and modeled low cloud, but sorted according to broken cloud cover (<90% total cloud cover) or overcast conditions (≥90% total cloud cover).
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Accumulated contribution to the multiyear mean bias from samples with observed and modeled low clouds. Notation as in Fig. 7. The first two bars show misses (model clear, observed low cloud) and false alarms (modeled low cloud, observed clear). The following four bars show contributions from cases with coincident observed and modeled low cloud, but sorted according to broken cloud cover (<90% total cloud cover) or overcast conditions (≥90% total cloud cover).
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
The frequency distribution of in-cloud LWP for these broken cloud cases (Fig. 9a) is comparable to the fair-weather cumulus days (Fig. 6a), and the same conclusions apply: the model overestimates in-cloud LWP in broken low clouds. Figure 9b shows the distribution for cases where model and observations agree on overcast low cloud conditions, most commonly found during the spring and fall. Here, the situation is reversed: the model produces low LWP too frequently and underestimates the occurrence of high LWP. As the overlap scheme is run offline to produce in-cloud estimates of cloud liquid water contents and path, we can experiment with increasing and decreasing cloud liquid water for a better match to the observed distributions. A reduction of cloud liquid water to a third of its original value provides a much better fit for the broken cloud cases (Fig. 9c). The inhomogeneity function applied to the cloud liquid water acts to shift samples to the tails of the distribution. While this is beneficial for the broken cloud case where low-LWP samples are underestimated, it is detrimental for the overcast cloud case. Here, the assumption of a more homogeneous cloud liquid water content provides a better fit with the observed distribution. As a sensitivity test, Fig. 9d shows the distribution for overcast clouds where cloud liquid water is increased by a factor of 1.5 and is assumed to be homogeneous throughout the cloud.
Normalized frequency distributions of (gray) observed and (black) modeled in-cloud LWP at ARM SGP. (a) For samples with broken (<90%) low clouds in model and observations. (b) For samples with overcast (≥90%) low clouds in model and observations. (c) As in (a), but with model cloud liquid water contents reduced by a factor of 0.33. (d) As in (b), but with cloud liquid water contents increased by a factor of 1.5 and homogeneous in-cloud liquid water.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Normalized frequency distributions of (gray) observed and (black) modeled in-cloud LWP at ARM SGP. (a) For samples with broken (<90%) low clouds in model and observations. (b) For samples with overcast (≥90%) low clouds in model and observations. (c) As in (a), but with model cloud liquid water contents reduced by a factor of 0.33. (d) As in (b), but with cloud liquid water contents increased by a factor of 1.5 and homogeneous in-cloud liquid water.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
Normalized frequency distributions of (gray) observed and (black) modeled in-cloud LWP at ARM SGP. (a) For samples with broken (<90%) low clouds in model and observations. (b) For samples with overcast (≥90%) low clouds in model and observations. (c) As in (a), but with model cloud liquid water contents reduced by a factor of 0.33. (d) As in (b), but with cloud liquid water contents increased by a factor of 1.5 and homogeneous in-cloud liquid water.
Citation: Monthly Weather Review 140, 11; 10.1175/MWR-D-11-00316.1
These sensitivity tests suggests that future parameterization changes should aim to increase cloud liquid water in overcast low clouds and decrease cloud liquid water and effective radius in broken cloud situations over land, which are dominated by shallow convective processes. The impact of these changes on surface radiation will have to be assessed.
6. Conclusions
As a result of improvements in the shortwave radiation and cloud schemes over recent years, the ECMWF model’s representation of the fair-weather cumulus regime is in qualitatively good agreement with observations from the ARM SGP site. While the cloud radiative forcing of the model’s clouds is overestimated, this is compensated by occasional days when the model does not produce clouds. These compensating errors need further investigation, yet the fair-weather cumulus regime, as represented by the 146 days selected from the observed CMBE record, does not contribute significantly to the multiyear mean bias in surface irradiance that is found at the model grid box centered nearest the SGP site. A simple cloud classification used in conjunction with the model surface irradiance bias identifies low clouds as significant contributors to the multiyear mean bias. Tselioudis and Kollias (2007) found that the ECMWF model produces too little boundary layer cloud in winter, but too much in summer. Considering that overcast low clouds (stratus) are more common in winter, while broken low clouds (fair-weather cumulus) are found predominantly in summer, the SWDN biases associated with the low cloud category here are consistent with the conclusions drawn in Tselioudis and Kollias (2007). Previous studies using spaceborne observations to evaluate the frequency of occurrence and cloud amount over ocean indicate a lack of cloud cover in the stratocumulus regime and a compensation between frequency of occurrence and cloud amount-when-present in the trade regime (Ahlgrimm et al. 2009; Ahlgrimm and Köhler 2010). These compensations appear to be common themes in the model. By addressing the problem at the ARM SGP site, it may be possible to reduce biases globally.
The opposing biases produced by broken and overcast low cloud samples explain why recent changes and qualitative improvements to the ECMWF’s shallow convection scheme were not reflected in the surface shortwave radiation fields: reducing the bias for broken cloud conditions addresses only part of the problem. Unless the bias observed during overcast conditions is addressed at the same time, the lack of compensation will lead to an overall worse result. Future developments will have to address how to reduce the shortwave cloud forcing and liquid water path for summertime shallow convective clouds over land, while at the same time increasing cloud cover and liquid water path for overcast cases in the fall, winter, and spring seasons. A more sophisticated cloud inhomogeneity function that is able to distinguish between broken and overcast situations promises to improve agreement with observed in-cloud liquid water paths.
Observations from the Routine ARM Aerial Facility (AAF) Clouds with Low Optical Water Depths (CLOWD) Optical Radiative Observations (RACORO) campaign (Vogelmann et al. 2012) provide a highly suitable in situ dataset characterizing overcast and broken low clouds around the ARM SGP site, from late winter into early summer. These observations will provide more accurate cloud microphysical observations for future assessment of model clouds with low optical depth.
The focus here has been on low clouds. However, “deep” and “thin midlevel” clouds were also identified as main contributors to the SWDN bias, and the shortlist of cloud-type categories discussed here can only explain about half of the multiyear SWDN bias. Clearly, other cloud types need further investigation as well. By allowing any and all combinations between individual cloud types, similar hydrometeor profiles may be scattered across marginally different cloud-type categories, and thus fail to stand out as a distinct category. A stricter classification with fewer categories might produce more distinct groups of cloud classes that failed to make the shortlist here. On the other hand, the hydrometeor fraction profile alone may not contain enough information to differentiate between similar categories, some of which contribute to the bias, while others do not. A more sophisticated cloud classification, such as used in Marchand et al. (2009), which links cloud profiles with meteorological conditions in the area surrounding the SGP site likely holds more promise when trying to establish a connection between model parameterizations and radiation biases for more complicated situations with deep and multilayer clouds.
Based on the new insights described in this paper, it will be possible to carry out targeted sensitivity studies to improve cloud occurrence and radiative properties by examining the formulations of the shallow convection trigger, mass transport, and cloud microphysical properties. The combination of a simple cloud classification with the model surface irradiance bias as described in this paper has proved to be a valuable tool for model assessment to discover compensating errors and is a necessary step to guide future parameterization improvements.
Acknowledgments
This study was funded through a fellowship of the Atmospheric System Research program under Grant DE-SC0005259. Kind appreciation to Drs. Yunyan Zhang and Steven Klein for making available the list of dates with observed fair-weather cumulus clouds at the ARM SGP site. Also thanks to Anton Beljaars and other colleagues for helpful discussions and to the three anonymous reviewers for their constructive comments.
REFERENCES
Ahlgrimm, M., and M. Köhler, 2010: Evaluation of trade cumulus in the ECMWF model with observations from CALIPSO. Mon. Wea. Rev., 138, 3071–3083.
Ahlgrimm, M., D. Randall, and M. Köhler, 2009: Evaluating cloud frequency of occurrence and cloud-top height using spaceborne lidar observations. Mon. Wea. Rev., 137, 4225–4237.
Bechtold, P., M. Köhler, T. Jung, F. Doblas-Reyes, M. Leutbecher, M. Rodwell, F. Vitart, and G. Balsamo, 2008: Advances in simulating atmospheric variability with the ECMWF model: From synoptic to decadal time-scales. Quart. J. Roy. Meteor. Soc., 134 (634), 1337–1351.
Berg, L., and E. Kassianov, 2008: Temporal variability of fair-weather cumulus statistics at the ACRF SGP site. J. Climate, 21, 3344–3358.
Cheinet, S., A. Beljaars, M. Köhler, J.-J. Morcrette, and P. Viterbo, 2005: Assessing physical processes in the ECMWF model forecasts using the ARM SGP observations. ECMWF ARM Rep. 1, 25 pp.
Dunn, M., K. Johnson, and M. Jensen, 2011: The microbase value-added product: A baseline retrieval of cloud microphysical properties. Tech. Rep. DOE/SC-ARM/TR-095, DOE Office of Science Atmospheric Radiation Measurement (ARM) Program (United States), 34 pp.
Hartmann, D., V. Ramanathan, A. Berroir, and G. Hunt, 1986: Earth radiation budget data and climate research. Rev. Geophys., 24 (2), 439–468.
Iacono, M., J. Delamere, E. Mlawer, M. Shepard, S. Clough, and W. Collins, 2008: Radiative forcing by long-lived greenhouse gases: Calculations with the AER radiative transfer models. J. Geophys. Res., 113, D13103, doi:10.1029/2008JD009944.
Illingworth, A., and Coauthors, 2007: Continuous evaluation of cloud profiles in seven operational models using ground-based observations. Bull. Amer. Meteor. Soc., 88, 883–898.
Jakob, C., 2003: An improved strategy for the evaluation of cloud parameterizations in GCMs. Bull. Amer. Meteor. Soc., 84, 1387–1402.
Jung, T., and Coauthors, 2010: The ECMWF model climate: Recent progress through improved physical parametrizations. Quart. J. Roy. Meteor. Soc., 136 (650), 1145–1160.
Köhler, M., M. Ahlgrimm, and A. Beljaars, 2011: Unified treatment of dry convective and stratocumulus-topped boundary layers in the ECMWF model. Quart. J. Roy. Meteor. Soc., 137 (654), 43–57, doi:10.1002/qj.713.
Kollias, P., G. Tselioudis, and B. Albrecht, 2007: Cloud climatology at the Southern Great Plains and the layer structure, drizzle, and atmospheric modes of continental stratus. J. Geophys. Res., 112, D09116, doi:10.1029/2006JD007307.
Mace, G., and S. Benson-Troth, 2002: Cloud-layer overlap characteristics derived from long-term cloud radar data. J. Climate, 15, 2505–2515.
Marchand, R., N. Beagley, and T. Ackerman, 2009: Evaluation of hydrometeor occurrence profiles in the multiscale modeling framework climate model using atmospheric classification. J. Climate, 22, 4557–4573.
Morcrette, J., 2002: Assessment of the ECMWF model cloudiness and surface radiation fields at the ARM SGP site. Mon. Wea. Rev., 130, 257–277.
Morcrette, J., H. Barker, J. Cole, M. Iacono, and R. Pincus, 2008: Impact of a new radiation package, McRad, in the ECMWF integrated forecasting system. Mon. Wea. Rev., 136, 4773–4798.
Räisänen, P., H. Barker, M. Khairoutdinov, J. Li, and D. Randall, 2004: Stochastic generation of subgrid-scale cloudy columns for large-scale models. Quart. J. Roy. Meteor. Soc., 130 (601), 2047–2067.
Räisänen, P., H. Barker, and J. Cole, 2005: The Monte Carlo independent column approximation’s conditional random noise: Impact on simulated climate. J. Climate, 18, 4715–4730.
Shonk, J., and R. Hogan, 2010: Effect of improving representation of horizontal and vertical cloud structure on the earth’s global radiation budget. Part II: The global effects. Quart. J. Roy. Meteor. Soc., 136 (650), 1205–1215.
Stoffel, T., 2005: Solar infrared radiation station (SIRS) handbook. Atmospheric Radiation Measurement Program Tech. Rep. ARM TR-025, 29 pp.
Tselioudis, G., and C. Jakob, 2002: Evaluation of midlatitude cloud properties in a weather and a climate model: Dependence on dynamic regime and spatial resolution. J. Geophys. Res., 107, 4781, doi:10.1029/2002JD002259.
Tselioudis, G., and P. Kollias, 2007: Evaluation of ECMWF cloud type simulations at the ARM Southern Great Plains site using a new cloud type climatology. Geophys. Res. Lett., 34, L03803, doi:10.1029/2006GL027314.
Turner, D., S. Clough, J. Liljegren, E. Clothiaux, K. Cady-Pereira, and K. Gaustad, 2007: Retrieving liquid water path and precipitable water vapor from the Atmospheric Radiation Measurement (ARM) microwave radiometers. IEEE Trans. Geosci. Remote Sens., 45 (11), 3680–3690.
Vogelmann, A. M., and Coauthors, 2012: RACORO extended-term aircraft observations of boundary layer clouds. Bull. Amer. Meteor. Soc., 93, 861–878.
Xie, S., and Coauthors, 2010: Clouds and more: ARM climate modeling best estimate data. Bull. Amer. Meteor. Soc., 91, 13–20.
Zhang, Y., and S. Klein, 2010: Mechanisms affecting the transition from shallow to deep convection over land: Inferences from observations of the diurnal cycle collected at the ARM Southern Great Plains site. J. Atmos. Sci., 67, 2943–2959.
