1. Introduction
Clouds are important modulators of global and local climate. It follows that clouds might also have a large effect on the present climate warming through positive and negative feedbacks. Dufresne and Bony (2008) analyzed cloud feedbacks on the temperature difference between an equilibrium climate with present-day greenhouse gas concentrations and an equilibrium 2 × CO2 climate. Output from 12 coupled ocean–atmosphere general circulation models (GCMs) was analyzed. All models produced a positive cloud feedback (0.75 K, on average) on the temperature increase, but the intermodel standard deviation of the feedback was 0.5 K. This was 70% of the intermodel standard deviation in the total temperature increase. Apparently, a large part of the model-induced uncertainty in climate predictions stems from uncertainty in cloud feedback (see also Randall et al. 2007, p. 636).
It seems reasonable to assume, though impossible to prove, that those models, which are better in simulating the present-day climate, may be expected to produce more accurate future climate predictions. This calls for evaluation of model simulations of the present-day climate with a focus on cloud-related parameters in view of the apparent large intermodel variability in cloud feedback. In such evaluations, one may distinguish between properties of the clouds themselves, like cloud cover and cloud-top temperature, and the important effects of clouds on climate, like the radiative fluxes at the top-of-the atmosphere (TOA) and precipitation. In recent years important methodological progress was made within the research area of model cloud evaluation. Williams and Tselioudis (2007) introduced a cloud-clustering method and applied this technique to compare cloud properties of six GCMs with International Satellite Cloud Climatology project (ISCCP) satellite data. The approach favors the evaluation of processes related to particular cloud regimes. Modeled radiative fluxes at the surface, the TOA, and three pressure levels were evaluated with Clouds and the Earth’s Radiant Energy System (CERES) data and related to the underlying cloud properties by Su et al. (2010). The use of radiative fluxes at more than one level enabled the analysis of cloud height and the vertical distribution of cloud amount. Whereas most satellite data are collected by polar-orbiting satellites and therefore resolve the diurnal cycle only poorly or not at all, data from geostationary satellites like Meteosat provide an excellent description of the diurnal cycle. Roebeling and van Meijgaard (2009) compared modeled cloud properties over Europe with data derived from Meteosat’s Spinning Enhanced Visible and Infrared Imager (SEVIRI), but they did not exploit the TOA radiative fluxes from Meteosat’s Geostationary Earth Radiation Budget (GERB) sensor. Allan et al. (2007) used the GERB fluxes and two cloud masks derived from SEVIRI data for model evaluation.
In the present paper, we discuss an evaluation of the Regional Atmospheric Climate Model version 2 (RACMO) operated at the Royal Netherlands Meteorological Institute (KNMI). The model domain encloses Africa and parts of the surrounding oceans, and the time frame is the month of July 2006. Model-predicted cloud properties [cloud cover, cloud-condensed water path (CWP), i.e., the sum of cloud liquid water path (LWP) and cloud ice water path (IWP), cloud-top temperature (CTT), and effective droplet radius (reff)] are compared to cloud properties retrieved from SEVIRI data. Simulated TOA fluxes (reflected shortwave, outgoing longwave, and net total) are evaluated with GERB data. We also exploit ISCCP cloud cover data and microwave LWP data from the University of Wisconsin climatology. We will argue that the GERB observations are relatively accurate and that in several parts of the domain differences between modeled and observed TOA fluxes are larger than the uncertainties in the observations. We will therefore refer to those differences as “model biases, overestimations, and underestimations” and use the term “model evaluation” when modeled and observed TOA fluxes are compared. The situation is different for the cloud properties retrieved from remote sensing measurements. These retrievals generally have relatively large error bars, which in addition are difficult to quantify, while in some occasions the size of the error bars cannot be estimated at all. Consequently, model–observation differences in cloud properties may be substantial while still not exceeding the error bars. In those cases we will use terms like “model–observation differences” and “model–observation comparison,” which implies that the differences could be due to inaccuracies in both the model and the observations. However, the simultaneous model to observation comparison of cloud properties and TOA fluxes will allow defining some constraints on model biases of cloud properties.
Three aspects make our setup particularly suitable for testing parameterizations of the physical processes in climate models. First, a Regional Climate Model (RCM) forced at the lateral boundaries by observations or pseudoobservations [e.g., the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalyses as in this work] is more suitable for this purpose than a GCM. As a result of such forcing, differences between simulated and observed large-scale dynamics are relatively small. Hence, the largest part of the model–data differences can be ascribed to shortcomings in the physical components of the model. Our focus on Africa is the second asset for testing physical parameterizations. At midlatitudes synoptic systems form an essential component of the climate, so the quality of model simulations depends largely on the correct simulation of the dynamics of these systems. In Africa, such systems are almost absent as testified by the relatively large persistence of the weather and physical processes have a dominant influence on the weather. Finally, analysis of diurnal cycles is a good test of the simulation of atmospheric physical processes like convection and the formation of clouds and precipitation since these occur typically on time scales shorter than a day.
In the present paper, we will go beyond a standard evaluation. Though RACMO was able to reproduce the observed position of various cloud and climate regimes with considerable accuracy, some important model–data differences were found. To reduce these differences, sensitivity model runs, in which parameterizations were altered, were made. Results of some successful experiments will be presented and provide indications for model improvement. The structure of the paper is as follows. In section 2, we will outline the model and the satellite data. Results of the model–data comparison, both for the initial model setup and for model runs with changes in parameterizations, will be discussed in section 3. Section 4 contains the conclusions and a discussion.
2. Model and satellite data
a. GERB TOA fluxes
The Geostationary Earth Radiation Budget 2 (GERB-2) instrument (Harries et al. 2005) has provided datasets of the reflected shortwave TOA flux and the outgoing longwave TOA flux (OLR) since 26 March 2004. In May 2007, it was succeeded by a similar instrument, GERB-1. The GERB instruments are unique in the sense that they are the only earth radiation budget sensors on-board geostationary satellites (Meteosat). They are positioned over the equator near the prime meridian. In fact, each GERB instrument consists of an array of 256 sensors aligned in the north–south (NS) direction, which all scan in the east–west (EW) direction. The sensors measure shortwave and total radiances and longwave radiances are obtained as the difference. According to an analysis of the individual components of the retrieval procedure, the theoretical accuracy of the radiances is estimated at 1.99% for shortwave radiance and 0.9% for longwave radiance (Clerbaux et al. 2009), at one standard deviation. Radiances are converted into fluxes using angular-dependence models based on CERES data for the shortwave and based on radiative transfer computation for the longwave. This conversion is likely to incur additional uncertainty on the order of 10 W m−2 for shortwave fluxes and 5 W m−2 for OLR fluxes (Harries et al. 2005). Clear-sky OLR fluxes over the Atlantic Ocean were compared by Allan et al. (2005) with calculations made with a numerical weather prediction (NWP) model. Biases were smaller than about 5 W m−2. A thorough comparison of GERB-2 radiances and fluxes with CERES data was carried out by Clerbaux et al. (2009). GERB OLR fluxes were found to be 1.3% lower than the CERES fluxes. GERB shortwave fluxes were 7.5% higher than the corresponding CERES fluxes, which is more than expected on the basis of the theoretical accuracy of the two instruments. It is unknown which of the two instruments, GERB or CERES, is closer to the truth.
The field-of-view of the raw GERB data measures 68 km (EW) × 38 km (NS) at nadir. Several products are derived from these raw data. For the present study we exploited GERB-2 high-resolution (HR) images. This product is presented on a grid with a spacing of 3 × 3 SEVIRI pixels (i.e., 9 km × 9 km at nadir) and is provided every 15 min as instantaneous values at the time of the SEVIRI observations. To produce the HR data, finescale measurements of the radiances from SEVIRI are combined with a GERB product of lower resolution as described in Dewitte et al. (2008). At that lower resolution radiances are conserved during this process so uncertainties in the HR product are as described before. The motivation for using the data with the highest resolution stems from the requirement to produce images of the TOA clear-sky albedo αclr, a parameter that is currently not provided by the GERB community, although a procedure using cloud masks has been proposed in Futyan and Russell (2005). We computed αclr with a procedure that has the advantage of not using any auxiliary data such as a cloud mask. In summary, we made subsets of the albedo observations, with each subset consisting of all observations (maximum = 31) for a certain pixel and a certain time of the day. The second lowest value was then sorted out and determined the clear-sky albedo. The entire procedure is described in detail in the appendix. It is important to notice that there is no day-to-day variation in the assigned clear-sky albedos and that, due to selecting the second lowest value from each subset, the clear-sky albedos represent, for local conditions, a relatively pristine atmosphere. The uncertainty in this product was assessed by computing the difference between the clear-sky albedos computed from, respectively, the fourth and the second lowest value of each data subset.
b. SEVIRI cloud properties
The SEVIRI are placed on the same Meteosat satellites as the GERB sensors and images are delivered at the same rate (4 h−1). However, SEVIRI has a much higher spatial resolution, namely 3 km for nadir view and even 1 km for the high-resolution channel. SEVIRI measures radiances in narrowbands, of which three are located in the shortwave part of the spectrum and eight in the longwave part. For the present study we employed the following products derived from the SEVIRI data:
The cloud mask (Derrien and Le Gléau 2005) developed by the Nowcasting Satellite Application Facility (SAFNWC): the mask is produced with a multispectral threshold method. It describes each pixel as clear, cloud contaminated, or cloud covered. In accordance with Roebeling and van Meijgaard (2009), we converted these qualifications to fractional cloud covers of 0, 0.5, and 1, respectively, to obtain the “standard cloud cover dataset.” We also produced two other cloud cover datasets, converting cloud-contaminated pixels to cloud covers of 0 and 1, respectively. The difference in cloud cover between one of these two “extreme” datasets and the standard dataset is a high estimate of the uncertainty in cloud cover due to quantifying cloud-contaminated pixels. However, other issues like subpixel cloud inhomogeneity also contribute to the total uncertainty in cloud cover, but a solid estimate of the total uncertainty has not been made. The SEVIRI cloud mask was validated with 3.5 yr of ground-based observations in Schulz et al. (2010). Monthly mean biases were found to be substantial, for example, approximately +0.10 when averaged over all ocean stations and varying between −0.20 and 0.00 when averaged over all tropical land stations. Root-mean-square errors were not provided.
The reff, cloud LWP, and cloud IWP derived from reflectances in two visible/near-infrared (VNIR) SEVIRI bands with the so-called Cloud Physical Properties (CPP) algorithm, described in Roebeling et al. (2006) and based on Nakajima and King (1990): retrieval is confined to cases with solar zenith angles smaller than 72°, and for individual pixels the retrieved cloud water is always entirely assigned to either the liquid or the ice phase. For the purpose of model evaluation we employed the sum of LWP and IWP, which we refer to as CWP. Validation studies of reff and IWP can only be performed with in situ airplane measurements, which makes the sample size far too small for proper validation of these products. Also, methods that have been developed to retrieve IWP from ground-based measurements are only sufficiently accurate for thin cirrus clouds and therefore not applicable in validation studies of IWP of, for example, frontal systems and deep convective systems. LWP from CPP was validated for two stations in Europe and found to have biases of approximately 5 g m−2 for both stations (Roebeling and van Meijgaard 2009). However, to the authors’ knowledge no published report on LWP validation for clouds over the oceans and clouds over tropical or subtropical land exists, and the error estimations for clouds over midlatitude land areas cannot simply be transferred to other cloud regimes. VNIR LWP estimates of oceanic clouds lacking ice crystals can be compared with estimates from passive microwave sensors. Recently, numerous papers dealing with such comparisons have been published (e.g., Greenwald 2009; Seethala and Horváth 2010). Generally, substantial differences were found, for example, relative differences in 2-yr mean values for overcast conditions up to 60% by Greenwald (2009). No clear picture of the causes of the discrepancies has yet emerged (Bennartz et al. 2010), but many authors agree on the following points: (i) microwave values are likely positively biased for partly cloudy skies, (ii) owing to methodological problems VNIR values tend to increase with solar zenith angle, and (iii) differences between VNIR and microwave values tend to increase with decreasing cloud amount. We investigated the accuracy of the SEVIRI LWP retrievals in the stratocumulus region off the coast of Angola by testing them for radiative closure. The essence of radiative-closure calculations is to use observed cloud properties as input to a radiative transfer model (RTM). If the computed TOA albedos then match with the observed TOA albedos, so-called closure is achieved. We tested for radiative closure by selecting samples of SEVIRI LWP and reff and of GERB TOA albedo from a 2° latitude × 2° longitude box within the stratocumulus region. All samples are for the time interval between 1200 and 1300 UTC from our July 2006 datasets and represent overcast conditions according to the SEVIRI observations. Calculations were performed with the Doubling Adding KNMI (DAK) (Kuipers Munneke et al. 2008), a detailed RTM, setting reff and the solar zenith angle equal to their mean values (5.5 μm and 40°, respectively), and taking cloud height and thickness from local in situ measurements (Keil and Haywood 2003). The DAK computations of the TOA albedo for five different values of LWP show excellent agreement with the GERB measurements (Fig. 1), especially since the deviations of the albedo observations from the DAK calculations can largely be explained by variations in observed reff. Given the accuracy of the GERB observations (≈0.01) and the RTM for cloudy atmospheres (see Wang et al. 2011), we estimate the uncertainty of SEVIRI LWP for the stratocumulus region to be 20% or less. However, this uncertainty estimate cannot be transferred to other cloud regimes. In conclusion, the current state of knowledge does not allow making a meaningful and credible uncertainty estimate of LWP retrievals over the African continent. Over the surrounding oceans the differences in LWP found by the VNIR and microwave methods may be considered as indicative of the uncertainty in the SEVIRI product and for the stratocumulus regimes the uncertainty in LWP is 20% or less.
CTT: the algorithm for the computation of CTT is based on the combination of SEVIRI signals in infrared channels with NWP forecasts of temperature and humidity profiles [see the Satellite Application Facility on Climate Monitoring (CM-SAF) algorithm theoretical basis document (ATBD) available online at http://www.cmsaf.eu]. Cloud-top height has been validated (Derrien and Le Gléau 2010) with 1 yr of observations based on lidar and radar signals in Palaiseau (France). Standard deviations are approximately 1 km, both for opaque and semitransparent clouds. On the assumption of a temperature lapse rate of −7 K km−1, this corresponds to an uncertainty estimate in CTT of 7 K.
Scatterplot of TOA albedo against LWP for overcast cases in the stratocumulus region off the coast of Angola. Triangles represent satellite observations (from SEVIRI and GERB) of both quantities between 1200 and 1300 UTC during July 2006. Dots represent DAK radiative transfer calculations; see text for details.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Scatterplot of TOA albedo against LWP for overcast cases in the stratocumulus region off the coast of Angola. Triangles represent satellite observations (from SEVIRI and GERB) of both quantities between 1200 and 1300 UTC during July 2006. Dots represent DAK radiative transfer calculations; see text for details.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Scatterplot of TOA albedo against LWP for overcast cases in the stratocumulus region off the coast of Angola. Triangles represent satellite observations (from SEVIRI and GERB) of both quantities between 1200 and 1300 UTC during July 2006. Dots represent DAK radiative transfer calculations; see text for details.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
c. RACMO
The RACMO is a hydrostatic limited-area atmospheric model employing semi-Lagrangian dynamics. The model is used for regional climate modeling (van Meijgaard et al. 2008) and was developed at KNMI by porting the physics package of the ECMWF Integrated Forecasting System (IFS) into the forecast component of the High-Resolution Limited-Area Model (HIRLAM) numerical weather prediction model, version 5.0.6 (de Bruijn and van Meijgaard 2005). In this paper, we apply an upgraded version of RACMO based on ECMWF cycle 31r1, almost catching up with cycle 31r2 used in the ERA-Interim project (Uppala et al. 2008). Cloud processes in RACMO are described by prognostic equations for cloud fraction and condensed water (liquid water and ice). The distinction between the liquid and the ice phase is made as a function of temperature. Cloud forming and dissolving processes are considered subgrid scale and hence parameterized; however, large-scale transport of cloud properties is accounted for on the resolved scale. Sources and sinks of cloud fraction and cloud condensate are process oriented and physically based, in contrast to the more commonly applied statistical approach. Total 2D cloud cover is obtained from the vertical profile of cloud fraction by assuming random-maximum overlap within a model grid box. The importance of the overlap assumption resides in its influence on the radiative transfer calculations. Details of the cloud parameterizations are described on the ECMWF Web site (http://www.ecmwf.int/research/ifsdocs/CY31r1/PHYSICS/IFSPart4.pdf and references therein).
For the purpose of this study, RACMO is operated at a horizontal resolution of 50 km × 50 km and a vertical mesh of 40 layers with the top layer at 10 hPa and the bottom layer at 10 m above the surface. The model domain, counting 222 points in the zonal direction and 192 points in the meridional direction, fully encloses the domain of evaluation. The initial model state, consisting of the atmosphere and the land surface/soil state, is taken from the ERA-Interim reanalysis archive. Starting from the initial state, continuous integrations are made for the month of July 2006, that is, the model state evolves freely. However, two types of boundary conditions are imposed. First, atmospheric forcing at the lateral boundaries is taken from subsequent ERA-Interim reanalyses of wind, temperature, and humidity at a 6-h time interval. Second, at the ocean surface, sea surface temperatures and sea ice fraction are prescribed from ERA-Interim at a 6-h time interval. Land surface characteristics like vegetation type and coverage, surface roughness length due to vegetation, and surface albedo are derived from the ECOCLIMAP dataset, version 1 (Champeaux et al. 2003). This description constitutes the reference run (EXP0) of the present study. In addition to EXP0, a number of sensitivity runs has been carried according to the description in Table 1. Their results will be discussed in Section 3.
List of RACMO experiments.
In recent years, RACMO has been used for present-day climate simulations and future climate predictions of Europe, Africa, Antarctica, and Greenland and was found to perform relatively well. Within the European Union (EU)-ENSEMBLES project, RACMO was found significantly better in simulating the present-day climate of Europe than any of the other 14 RCMs contributing to the project (Christensen et al. 2011). Also, for the West African region RACMO provides favorable results for precipitation relative to results from other RCMs obtained within ENSEMBLES (Paeth et al. 2011).
d. Computation of grid box values
In a first step, the hourly (15 min in the case of GERB) satellite data were spatially aggregated to the boxes of the RACMO grid (0.44° × 0.44°). GERB radiative fluxes and most of the SEVIRI cloud properties were simply averaged within the RACMO grid boxes. Gridbox values of model reff and model and satellite CTT were computed in a different way.
Model reff is a multilevel variable. Therefore, gridbox values were obtained by averaging reff over all model layers, weighing each layer’s contribution by its liquid water path (the product of the liquid water content and the geometric depth of the layer). Since the liquid water content of clouds tends to increase toward the cloud top (e.g., Martin et al. 1994), our averaging method tends to put more weight on the upper layers of the clouds. The satellite-retrieved reff is also more representative of the upper part of a cloud than of the lower part of a cloud, but the issue is complicated since the vertical weighing factors depend on the near-infrared wavelength used for the retrieval (Chen et al. 2008).
To compare modeled to observed CTT, subgrid-scale variability in CTT needs to be considered. In RACMO subgrid-scale variability in CTT exists because cloud fractions are defined at each vertical model level with the relative position of the clouds in the horizontal dimension determined by the overlap assumption (random maximum in the case of RACMO). Regarding SEVIRI, each RACMO grid box contains a large number of SEVIRI pixels (167, on average), so the possible existence of observational subgrid-scale variability is obvious. We defined CTT25 as follows: 25% of the model grid box have a CTT lower than CTT25. To compute model CTT25 the predicted layer cloud fraction was transferred into a cloud cover profile as viewed from satellite, that is, a cloud cover profile gives the total cloud cover at each level that would be seen from a satellite if there were no clouds below the level in question (van Meijgaard et al. 2001). In this process the cloud cover of each layer is multiplied by its emissivity. As a result the cloud cover profiles increase monotonically downward. CTT25 is then extracted as the atmospheric temperature at the level where cloud cover reaches 0.25, where we implicitly assumed that temperature decreases monotonically with altitude. The satellite-derived CTT25 for a RACMO grid box was computed in a straightforward way by ranking all CTTs of the SEVIRI pixels within a box from the lowest value to the highest value and taking out the 25th-percentile value from the row. If more than 75% of a RACMO grid box were clear (after correction for emissivity) and/or more than 75% of the SEVIRI CTT pixel values were missing, missing values were assigned to the grid box. For this study we also computed CTT05 in an analogous way.
e. Computation of monthly means
For many analyses we exploited monthly means. Computation of monthly means from hourly and 15-min values is mostly trivial. In this paper “monthly albedo” refers to the ratio of the monthly mean reflected and incoming TOA shortwave fluxes, and net total radiation is defined as shortwave incoming radiation minus total outgoing radiation. Owing to missing data, computation of monthly means from model output and satellite data is not always trivial. The following cases of missing data are discriminated:
The satellite products CWP and reff are not computed for solar zenith angles beyond 72°, so they are confined to part of the daylight period. To make a proper comparison of model with satellite CWP and reff, RACMO CWP and reff values for solar zenith angles greater than 72° were also dismissed for the calculation of time averages (Fig. 2 and Table 3) or SEVIRI CWP was corrected with independent data of the diurnal cycle (Fig. 7 and Table 4).
Cloud-top temperature (CTT25) is not computed if less than 25% of a model grid box is cloudy, and reff is not computed for model grid boxes with no cloud liquid water. This applies both to the satellite data and to the model output. The monthly mean is the average over the available samples.
A few of the 256 GERB sensors have been damaged by exposure to direct sunlight. This results in rows of missing data, which correspond to latitudes around 40°S. A north–south oscillation around this latitude due to the inclination of the satellite occurs.
Radiance-to-flux conversion introduces too large errors in the GERB shortwave flux for reflection angles near the glint angle. Therefore, GERB shortwave data are missing for reflection angles within 15° from the glint angle and, in the case of clear-sky ocean scenes, for reflection angles within 25° from the glint angle. We filled the resulting data gaps by making a linear interpolation of the albedo time series.
A small percentage of the images is missing (0.6% for GERB and 3.0% for SEVIRI).
July 2006 (a)–(c) albedo, (d)–(f) means of OLR, (g)–(i) net total radiation, (j)–(l) cloud cover, (m)–(o) CWP, (p)–(r) CTT, and (s)–(u) reff. (left) Satellite data, (middle) RACMO output for the reference run (EXP0), and (right) the difference between model and observations (positive when RACMO value is higher than the satellite-derived value). Black areas correspond to missing values. CWP and reff values are averages over all model output and observations for solar zenith angles less than 72°. The blue arrow shows the position of the transect analyzed in Fig. 7. Rectangles delimit subregions on which some evaluations presented in this paper focus.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 (a)–(c) albedo, (d)–(f) means of OLR, (g)–(i) net total radiation, (j)–(l) cloud cover, (m)–(o) CWP, (p)–(r) CTT, and (s)–(u) reff. (left) Satellite data, (middle) RACMO output for the reference run (EXP0), and (right) the difference between model and observations (positive when RACMO value is higher than the satellite-derived value). Black areas correspond to missing values. CWP and reff values are averages over all model output and observations for solar zenith angles less than 72°. The blue arrow shows the position of the transect analyzed in Fig. 7. Rectangles delimit subregions on which some evaluations presented in this paper focus.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 (a)–(c) albedo, (d)–(f) means of OLR, (g)–(i) net total radiation, (j)–(l) cloud cover, (m)–(o) CWP, (p)–(r) CTT, and (s)–(u) reff. (left) Satellite data, (middle) RACMO output for the reference run (EXP0), and (right) the difference between model and observations (positive when RACMO value is higher than the satellite-derived value). Black areas correspond to missing values. CWP and reff values are averages over all model output and observations for solar zenith angles less than 72°. The blue arrow shows the position of the transect analyzed in Fig. 7. Rectangles delimit subregions on which some evaluations presented in this paper focus.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
3. Evaluation
a. Whole domain
The overall pattern is well simulated by the reference run (EXP0; Fig. 2). Within the model domain the ITCZ is perhaps the most striking feature. It is characterized by deep convection, often organized in mesoscale convective systems (MCSs), and a frequent occurrence of cirrus mainly formed as anvils from convective updrafts (see Sassen et al. 2009). According to the observations, in July 2006 the ITCZ runs from west to east across the domain, mainly just north of the equator. Its width is a mere 5–10° over the Atlantic Ocean, but it then widens gradually in eastern direction over the continent to reach a maximum north–south extent of 15–20° near 30°E, and it finally disappears abruptly east of the Ethiopian Highlands. Compared to the areas to its north and south, the ITCZ has larger cloud cover, CWP, albedo, and precipitation and lower OLR and cloud-top temperature. Since the higher reflected shortwave flux compensates for the lower OLR, the ITCZ is less striking in the maps of net total radiation. All these features as well as the geographic location of the ITCZ are well simulated by the model.
The north–south variations outside the ITCZ, for example, over the southern Atlantic, are also well captured by the model. There is a minimum in cloud cover, CWP, and albedo just south of the ITCZ and a corresponding maximum in the OLR and the cloud-top temperature. Toward the southern edge of the model domain cloud cover, CWP and albedo then gradually increase whereas the OLR, cloud-top temperature, and net total radiation gradually decrease.
The model is also well able to simulate the spatial variation in OLR across the Sahara and the Arabian Peninsula and the maximum in Mesopotamia. Moreover, relatively small topographic features like the Ethiopian Highlands and the north–south-extending mountain chain on Madagascar have similar effects on both satellite and model variables, though the magnitude of the local extremes may differ considerably.
Despite these similarities, important features show up in the difference plots on the right-hand side of Fig. 2. Some of these features are present in large parts of the model domain, so they are not linked to specific regions or cloud regimes. These are the following:
Many regions have positive biases in OLR, which are larger than the error in the data (≈6 W m−2).
Compared to the observations, the model tends to have higher cloud cover over land and smaller cloud cover over the oceans. To judge whether the model–observation differences might actually be called model biases, they are compared with the uncertainty in the observations plotted in Fig. 3a. The latter was taken as the difference between cloud covers computed with cloud-contaminated pixels counting as cloud cover values of 1.0 and 0.5, respectively. This observational uncertainty due to quantifying cloud-contaminated pixels reaches values larger than 0.15 within most of the ITCZ and is smaller than 0.06 in most regions outside the ITCZ. We then produced another plot (Fig. 3b) of the model–observation difference in cloud cover, masking out all model grid boxes where absolute differences in cloud cover were substantial (an arbitrary threshold of 0.10 was set) but did not exceed the observational uncertainty. As a result the remaining absolute differences greater than 0.10 can be classified with larger likelihood as model biases. They occur mainly in the western part of the continental ITCZ and in the region of stratocumulus off the coast of Angola. The model–observation differences in cloud cover can be compared with Roebeling and van Meijgaard (2009), who evaluated a previous version of RACMO for Europe and found that model calculations of cloud cover underestimated the observations by approximately 0.20, both over land and over ocean.
Modeled CWP tends to be larger than observed CWP. Notable exceptions with negative simulated minus observed CWP are the ITCZ over the Atlantic Ocean, some continental regions along the southern margin of the ITCZ, and the region of the stratocumulus fields off the coast of Angola. In view of the unknown uncertainty in the observations, the differences cannot be ascribed with certainty to the model. Positive model–observation differences of CWP are in broad agreement with the 30% CWP difference between RACMO and SEVIRI found by Roebeling and van Meijgaard (2009) for Europe and with a generally positive difference between microwave LWP and the 40-yr ECMWF Re-Analysis (ERA-40) over the oceans (O’Dell et al. 2008).
July 2006 monthly mean cloud cover: (a) uncertainty in the observations due to the uncertainty in quantifying cloud-contaminated pixels and (b) difference in cloud cover between reference model run and observations. In (b) all model grid boxes with absolute differences in cloud cover larger than 0.10 but less than the observational uncertainty were masked out (black).
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 monthly mean cloud cover: (a) uncertainty in the observations due to the uncertainty in quantifying cloud-contaminated pixels and (b) difference in cloud cover between reference model run and observations. In (b) all model grid boxes with absolute differences in cloud cover larger than 0.10 but less than the observational uncertainty were masked out (black).
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 monthly mean cloud cover: (a) uncertainty in the observations due to the uncertainty in quantifying cloud-contaminated pixels and (b) difference in cloud cover between reference model run and observations. In (b) all model grid boxes with absolute differences in cloud cover larger than 0.10 but less than the observational uncertainty were masked out (black).
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
b. Sahara and southern Africa
Both the Sahara and southern Africa between 10° and 24°S, as defined by the rectangular areas in Fig. 4c (SAH and SAFR), had very low mean cloud cover during July 2006 (Table 2). Despite the scarcity of clouds, considerable differences between model and satellite albedo exist (Figs. 2a–c). Compared to the GERB albedo, the monthly model albedo of the Sahara rectangle is 0.051 lower, and the model albedo of the southern African region is 0.040 higher. Both values exceed the observational uncertainty (≈0.01). Moreover, the observed spatial albedo pattern of the Sahara is not mimicked by RACMO (Fig. 4a), which produces a spatially too uniform pattern.
July 2006 mean biases in clear-sky albedo for (a) the reference model run (EXP0) and model runs with the surface albedo prescribed from (b) annual MODIS data (EXP1) and (c) July MODIS data (EXP2). Panel (c) also shows the abbreviations of areas used for regional analyses (SAH = Sahara, SAFR = southern Africa, and TRCUM = Trade wind cumulus).
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 mean biases in clear-sky albedo for (a) the reference model run (EXP0) and model runs with the surface albedo prescribed from (b) annual MODIS data (EXP1) and (c) July MODIS data (EXP2). Panel (c) also shows the abbreviations of areas used for regional analyses (SAH = Sahara, SAFR = southern Africa, and TRCUM = Trade wind cumulus).
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 mean biases in clear-sky albedo for (a) the reference model run (EXP0) and model runs with the surface albedo prescribed from (b) annual MODIS data (EXP1) and (c) July MODIS data (EXP2). Panel (c) also shows the abbreviations of areas used for regional analyses (SAH = Sahara, SAFR = southern Africa, and TRCUM = Trade wind cumulus).
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 regional means of all-sky albedo, clear-sky albedo, and cloud cover for the Sahara (16°–30°N, 8°W–32°E) and southern Africa (10°–24°S, 15°–34°E). For the clear-sky albedo the spatial standard deviation is given between parentheses.
Figure 4a shows the difference between the simulated clear-sky albedo and the clear-sky albedo derived from the GERB data. Within the two considered regions these differences are almost identical to the differences in the total albedo (Fig. 2), both in magnitude (Table 2) and spatial variation. Apparently, the cause of the discrepancies must be sought in parameters or processes that determine the clear-sky albedo. We therefore examined whether the differences could be reduced by changing the surface albedo fields. In the RACMO reference run (EXP0) the albedo was prescribed as a function of surface type, of which the spatial variation was, in turn, given by the ECOCLIMAP dataset. In the Sahara most pixels are designated as desert but only two types of desert (bright and dark) are distinguished. This appears to be insufficient to describe the real spatial variability.
We therefore replaced the ECOCLIMAP-derived surface albedos by MODIS-inferred surface albedos. More precisely, we employed the broadband (0.3–5.0 μm), black-sky, 5-yr mean (2000–04) albedo fields from the dataset described in Moody et al. (2008). These data represent the snow-free surface, have a resolution of 1 min, and were aggregated onto the RACMO grid. In a first sensitivity run (EXP1, see Table 1), annual mean MODIS surface albedos were prescribed. As a result, the areal mean biases in clear-sky albedo diminished, namely, from −0.044 to −0.029 for the Sahara and from 0.041 to 0.019 for southern Africa (see Table 2). The residual biases can be compared to the uncertainty in the clear-sky albedo, which we estimated as the difference between the clear-sky albedo computed from, respectively, the fourth and the second lowest value of the albedo for each image time and pixel (see section 2a). This uncertainty (0.004 for the Sahara and 0.002 for southern Africa) appeared to be much smaller than the residuals, so the uncertainty in the clear-sky albedo does not explain the residuals. It is striking that most residual differences are negative in the Sahara and positive in southern Africa. Also, due to replacing the ECOCLIMAP by the MODIS albedo maps, the simulation of the spatial pattern of the clear-sky albedo becomes significantly better (cf. Fig. 4b with Fig. 4a), as expressed in the much better agreement between simulated and observed spatial standard deviation (Table 2).
To test whether the residual differences could be attributed to the fact that the MODIS albedo field of EXP1 was not specific for July, we replaced the annual mean MODIS albedo by the 12–27 July 5-yr mean MODIS surface albedo field (EXP2). Clear-sky albedos of EXP1 and EXP2 are almost identical (cf. Fig. 4c with Fig. 4b, see also Table 2). Apparently, seasonal variations in vegetation do not explain much of the residuals. Other possible causes of the residuals will be discussed in section 4.
Replacing the surface albedo fields of EXP0 by those of EXP2 caused changes in the regionally mean modeled TOA reflected shortwave radiation of 7 W m−2 (Sahara) and −5 W m−2 (southern Africa). The effect on the OLR was negligible (absolute magnitude ≤ 1 W m−2).
c. ITCZ
While the oceanic ITCZ region is one of the best-simulated regions in terms of the TOA fluxes, simulated fluxes within the continental ITCZ differ much more from observed values (Table 3). Within the rectangular area 7°–12°N, 11°W–24°E (ITCZ in Fig. 4c), monthly mean net total radiation is, on average, 40 W m−2 less than the observed value. Though the magnitude of this underestimation does not exhibit an east–west gradient across the region, the cause of the underestimation differs between the western (from Guinea to West Nigeria) and the eastern part (Nigeria to Central African Republic). In the western part both the shortwave-reflected flux (by 19 W m−2) and the OLR (by 22 W m−2) are too high. In the eastern part the model–data difference in net total radiation is almost entirely due to an overestimation of the OLR (by 36 W m−2), which is more severe than in the western part. The albedo overestimation in the western part (by 0.043) can be explained qualitatively by a model overestimation of cloud cover in this part (by 0.24), which exceeds the uncertainty estimate of cloud cover (0.17, Fig. 3a). In the eastern part the small albedo bias (+0.011) corresponds to an insignificant model–observation difference in cloud cover (+0.13 compared to an uncertainty of 0.21). The model overestimation of the OLR (by 31 W m−2 for the whole region) is inconsistent with the model–observation difference of cloud cover (+0.17), which has the wrong sign. The same holds for the model–observation difference in CWP (+32%), where this statement is only valid for the period of the day with CWP observations (≈0800–1600 LT), when the OLR bias is 28 W m−2. However, the OLR bias can qualitatively be explained by the bias in cloud-top temperatures. While model and satellite CTT05, representing the highest clouds, are found hardly different, modeled CTT25 exceeds the satellite CTT25 differing by 34 K, which is far beyond the uncertainty in the satellite observations (7 K). This difference cannot be explained by modeled coverage by high clouds (0.31; in the model high clouds are defined as clouds above the level where pressure equals 45% of the surface pressure), which is almost equal to the value inferred from ISCCP observations (0.29; ISCCP defines high clouds as clouds above 440 hPa). We conclude that, though in RACMO the fractional coverage by high clouds matches the observed values well, these clouds are far too thin or, equivalently, their thermal emissivity is far too small. This is similar to the conclusion by Klein and Jakob (1999) that in the ECMWF model high clouds of midlatitude baroclinic systems are optically too thin.
July 2006 means of selected variables for the continental ITCZ (7°–12°N, 11°W–24°E). Model and satellite CWP are limited to solar zenith angles less than 72°.
Figure 5 compares monthly mean diurnal cycles of modeled (EXP0) and satellite-retrieved albedo, OLR, cloud cover, and CWP. Except for a difference in the mean, simulated and observed diurnal cycle of cloud cover agree quite well. However, Fig. 5d shows large discrepancies in CWP even though the SEVIRI data do not allow the model–observation comparison of the full diurnal cycle. According to the observations, CWP decreases until around local noon and increases afterward, which means, in combination with the small amplitude of the diurnal cycle of cloud cover, that convective clouds tend to become thinner during the morning and thicker during the afternoon. The model diurnal cycle of CWP is almost inverted and suggests that clouds become thicker during the morning and early afternoon (≈0700–1400 LT) and decrease in thickness afterward. These findings are in broad agreement with those from other studies. According to Roebeling and van Meijgaard (2009), maximum convection is predicted earlier by RACMO than it is observed by SEVIRI while da Rocha et al. (2009) report that many climate models simulate a precipitation maximum in the tropics that occurs several hours earlier than observed. This suggests a failure in the parameterization of deep convection that is common to many models. Support for the hypothesis that the model–observation difference in the diurnal cycle of CWP can be ascribed to the model is found in the consistency between the model–observation difference in CWP and the biases in the TOA fluxes. The OLR overestimation discussed in the previous paragraph is present during the entire day but it is subdued during the early afternoon, which matches with the model overestimation of CWP during approximately the same time interval. The same consistency can be seen in the diurnal cycle of the albedo. The clear-sky albedo curves almost coincide, so biases in the diurnal cycle of the albedo are due to clouds, and the model underestimation (overestimation) of the albedo during the morning (afternoon) nicely corresponds with the negative (positive) model–observation difference of CWP during the same part of the day. It is important to note that overestimation of VNIR-derived CWP at high solar zenith angles (Seethala and Horváth 2010) may also induce an artificial diurnal cycle in the observations, which will likely contribute to the difference between simulated and observed CWP.
July 2006 mean diurnal cycles of (a) the albedo and the clear-sky albedo, (b) the OLR, (c) cloud cover, and (d) CWP within the continental ITCZ (7°–12°N, 11°W–24°E). Black curves represent the satellite data and red curves output from the RACMO reference run. The observed total albedo is missing within time periods just after sunrise, just before sunset and when the reflection angle is close to the glint angle. In these cases radiance-to-flux conversion causes too much uncertainty.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 mean diurnal cycles of (a) the albedo and the clear-sky albedo, (b) the OLR, (c) cloud cover, and (d) CWP within the continental ITCZ (7°–12°N, 11°W–24°E). Black curves represent the satellite data and red curves output from the RACMO reference run. The observed total albedo is missing within time periods just after sunrise, just before sunset and when the reflection angle is close to the glint angle. In these cases radiance-to-flux conversion causes too much uncertainty.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 mean diurnal cycles of (a) the albedo and the clear-sky albedo, (b) the OLR, (c) cloud cover, and (d) CWP within the continental ITCZ (7°–12°N, 11°W–24°E). Black curves represent the satellite data and red curves output from the RACMO reference run. The observed total albedo is missing within time periods just after sunrise, just before sunset and when the reflection angle is close to the glint angle. In these cases radiance-to-flux conversion causes too much uncertainty.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Frequency distributions of the albedo, OLR, cloud cover, and CWP within the ITCZ rectangle are plotted in Fig. 6. We confined the analysis to those samples for which the local solar zenith angle was less than 72°, the threshold for the computation of CWP from the satellite data. Therefore, all frequency distributions in the four panels are based on samples that are coincident in time and space and restricted to the largest part of the daylight cycle. For all four variables the most important difference between the modeled and the observed distributions is that extreme values are less frequent in the model than in the observations and that, consequently, values around the median of the distributions are more frequent in the model than in the observations. This behavior is also found, though it is not visible in the graphs, for high CWP. An exception is the finding that overcast conditions are more frequently simulated by the model than seen in the observations. So the model tends to have less variability than the observations, which also occurred for CWP in a RACMO run for Europe (Roebeling and van Meijgaard 2009).
Frequency distributions of the observed and simulated (EXP0) (a) albedo, (b) OLR, (c) cloud cover, and (d) CWP within the continental ITCZ (7°–12°N, 11°W–24°E). Satellite values were first aggregated to the RACMO grid boxes before being counted. Negative values of cloud cover and CWP represent cloud-free grid boxes. For all variables only those samples were considered for which the local solar zenith angle was less than 72°. Averages over all samples are given between parentheses.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Frequency distributions of the observed and simulated (EXP0) (a) albedo, (b) OLR, (c) cloud cover, and (d) CWP within the continental ITCZ (7°–12°N, 11°W–24°E). Satellite values were first aggregated to the RACMO grid boxes before being counted. Negative values of cloud cover and CWP represent cloud-free grid boxes. For all variables only those samples were considered for which the local solar zenith angle was less than 72°. Averages over all samples are given between parentheses.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Frequency distributions of the observed and simulated (EXP0) (a) albedo, (b) OLR, (c) cloud cover, and (d) CWP within the continental ITCZ (7°–12°N, 11°W–24°E). Satellite values were first aggregated to the RACMO grid boxes before being counted. Negative values of cloud cover and CWP represent cloud-free grid boxes. For all variables only those samples were considered for which the local solar zenith angle was less than 72°. Averages over all samples are given between parentheses.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Consistency is found in the fact that the model shows too little variability in albedo, cloud cover, and CWP. As far as the distribution of OLR is concerned, the reference model run almost completely misses values below 200 W m−2 (3%), which make up 22% of the observations. Obviously, very cold cloud tops with high emissivity are far too rare in the model, causing the overestimation in the mean OLR.
d. Stratocumulus and trade wind cumulus over the Atlantic Ocean
1) Observations and uncertainty in observations
The third climate regime with substantial model–data differences is the climate of large parts of the Atlantic Ocean outside the ITCZ. We focus on the part between the equator and 25°S, which is largely covered by shallow, broken cumulus clouds (also called trade wind cumulus clouds). This regime does not exist over the ocean off the west coast of Angola (roughly between 8° and 20°S and between 0° and the coast), where more or less continuous stratocumulus cloud fields dominate. The transition between these two regimes (de Roode and Duynkerke 1997) is due to a westward rise in SST and a westward decline in subsidence velocities. The observations are in broad agreement with the partitioning of the cloud climate into these two regimes as demonstrated by the east–west transects along 15°S in Fig. 7. SEVIRI cloud cover is close to 1.0 within the first approximately 1000 km off the coast of Angola and then gradually decreases toward the west. SEVIRI CWP also decreases toward the west but only starts to do so clearly at approximately 2000 km off the coast. The albedo decreases almost linearly from east to west within the first approximately 3000 km off the coast. In the western part of the transect all three variables (cloud cover, CWP, and albedo) level off. Regarding the first 1000 km off the coast, we like to note here that the constant cloud cover and the westward CWP increase seem to be inconsistent with the observed eastward increase in the albedo. Qualitatively this inconsistency could be explained by the observed coastward decrease in reff. Indeed, for constant LWP and cloud cover, smaller droplets lead to a higher cloud albedo. We finally note that east–west variations in OLR are small.
July 2006 (a) monthly albedo and (b) monthly means of OLR, (c) cloud cover, (d) CWP, and (e) reff along an east–west transect at 15°S. The Angolan coast is situated at 0 km. Averages were computed across a 200-km-wide zone. Different curves represent the satellite observations (black) and results obtained from four RACMO runs (colors). The gray-shaded area in (c) shows the uncertainty in SEVIRI cloud cover due to quantifying cloud-contaminated pixels. Because SEVIRI CWP is only retrieved for solar zenith angles less than 72°, it was converted to full-time CWP by means of the monthly mean diurnal cycle from the UWisc LWP climatology (multiplication factors between 1.01 and 1.14). Both model and satellite reff are restricted to the mentioned part of the day.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 (a) monthly albedo and (b) monthly means of OLR, (c) cloud cover, (d) CWP, and (e) reff along an east–west transect at 15°S. The Angolan coast is situated at 0 km. Averages were computed across a 200-km-wide zone. Different curves represent the satellite observations (black) and results obtained from four RACMO runs (colors). The gray-shaded area in (c) shows the uncertainty in SEVIRI cloud cover due to quantifying cloud-contaminated pixels. Because SEVIRI CWP is only retrieved for solar zenith angles less than 72°, it was converted to full-time CWP by means of the monthly mean diurnal cycle from the UWisc LWP climatology (multiplication factors between 1.01 and 1.14). Both model and satellite reff are restricted to the mentioned part of the day.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 (a) monthly albedo and (b) monthly means of OLR, (c) cloud cover, (d) CWP, and (e) reff along an east–west transect at 15°S. The Angolan coast is situated at 0 km. Averages were computed across a 200-km-wide zone. Different curves represent the satellite observations (black) and results obtained from four RACMO runs (colors). The gray-shaded area in (c) shows the uncertainty in SEVIRI cloud cover due to quantifying cloud-contaminated pixels. Because SEVIRI CWP is only retrieved for solar zenith angles less than 72°, it was converted to full-time CWP by means of the monthly mean diurnal cycle from the UWisc LWP climatology (multiplication factors between 1.01 and 1.14). Both model and satellite reff are restricted to the mentioned part of the day.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
As stated before (section 2b) it is difficult to estimate the uncertainty in SEVIRI cloud cover, and it is even impossible to make a solid estimate of the uncertainty in SEVIRI CWP, with the exception of the stratocumulus region. We therefore also plotted July 2006 mean ISCCP cloud cover and LWP derived from microwave sensors [from the University of Wisconsin (UWisc) climatology, see O’Dell et al. 2008) in addition to the SEVIRI values. Note that CWP from SEVIRI can be compared to LWP from microwave sensors because most clouds over the considered transect are water clouds. In addition, Fig. 7c shows the uncertainty range in cloud cover due to assigning cloud cover values to cloud-contaminated pixels as the gray shaded. Over most of the transect, cloud cover from SEVIRI is offset by values between +0.05 and +0.15 with respect to cloud cover by ISCCP, with fairly good agreement (difference < 0.05) in the region of the stratocumulus. We regard the high values of cloud cover near the coast (the stratocumulus region) as rather accurate (uncertainty < 0.1), because of the negligible uncertainty range in the SEVIRI values, because of the small difference between SEVIRI and ISCCP, and also because we assume that cloud cover estimates are most accurate when cloud cover is either 0 or 1. Compared to the differences in cloud cover between SEVIRI and ISCCP, the differences in LWP between SEVIRI and UWisc are much larger in a relative sense. This holds especially in the region of the trade wind cumulus clouds (SEVIRI CWP 0.32 × UWisc LWP within most westerly 2000 km of the transect), whereas in accordance with Seethala and Horváth (2010) the agreement is closer in the stratocumulus region (SEVIRI CWP 0.83 × UWisc LWP within the first 1000 km off the coast).
2) Evaluation of reference model simulations
Simulation of the stratocumulus regime with the RACMO reference run (EXP0) is poor. Right off the west coast of Angola, the albedo bias reaches its maximum absolute value across the entire domain (Fig. 2c) while simulated cloud cover is definitely too low (Figs. 2j–l). Along the transect, maximum mean simulated cloud cover is only approximately 0.6 while the SEVIRI and ISCCP data show a maximum cloud cover of almost 0.95. Moreover, whereas the observations show stepwise changes in cloud cover, CWP, and albedo right at the coast, the model computes a more or less linear increase in cloud cover, CWP, and albedo in westward direction within the first approximately 1000 km off the coast.
Table 4 compares the model with the observations within the most westerly 2000 km of the transect, where trade wind cumulus prevails. In the reference run (EXP0), the albedo is overestimated by 0.055 (28%) within this part of the transect. This overestimation could, in principle, be explained by a positive bias in simulated cloud cover, but simulated cloud cover is smaller than SEVIRI cloud cover (by 0.12) and also slightly smaller than ISCCP cloud cover (by 0.03). Model cloud cover also falls outside the error bars of the SEVIRI estimate. Hence, a positive bias in cloud cover is unlikely. Alternatively, the positive albedo bias might be explained by a positive bias of cloud optical thickness, which, in turn, could be caused by either an overestimation of CWP or an underestimation of reff in the simulations or a combination of these two factors.
July 2006 means within the most westerly 2000 km of the Atlantic transect of Fig. 7. Both model and satellite reff are restricted to solar zenith angles less than 72°. The same holds for SEVIRI CWP but this variable was converted to full-time values by means of the diurnal cycle from the UWisc LWP climatology.
In the model, reff is computed, based on Martin et al. (1994), from liquid water content under the assumption that droplet concentration is constant (51 cm−3 over the ocean; 321 cm−3 over land). We computed the vertical average for each model grid box and hour, weighing each layer’s contribution by its liquid water path. Figure 7e shows the monthly mean of the vertically averaged reff along the transect, and Table 4 averages those values for the trade wind cumulus regime along the transect (11.3 μm). This is close to the SEVIRI mean value (11.0 μm) and agrees well with in situ airborne measurements made in unpolluted air over the ocean near Tasmania (Boers et al. 1998). Admittedly, the method of vertically averaging reff of each model layer does not match with the implicit vertical weighing of reff in the satellite observations (see section 2d), which makes the comparison not conclusive. However, in view of the good agreement between modeled reff and both satellite and in situ measurements of reff, we deem it likely that modeled values of reff in the region of the trade wind cumulus clouds are realistic. That leaves overestimation of CWP by the model as the most plausible cause of the albedo bias in this region.
Indeed within the most westerly 2000 km of the transect, modeled CWP (69 g m−2) is much larger than the SEVIRI value (22 g m−2). In view of the above-mentioned arguments, the apparent excellent agreement between the model and UWisc LWP (65 g m−2) must be questioned. Overestimation of microwave retrieved values of LWP for broken clouds is confirmed by Seethala and Horváth (2010). On the other hand, the ratio between modeled and SEVIRI CWP (3.1) seems too large to be consistent with the relatively small model–observation albedo ratio (1.28), despite the nonlinear relation between CWP and albedo (Fig. 1). This suggests that SEVIRI CWP of the trade wind cumulus clouds is too low.
Diurnal cycles of cloud amount and CWP of the trade wind cumulus region (rectangular area TRCUM in Fig. 4c) are shown in Fig. 8. Observations and simulation agree that the amplitude of the diurnal cycle of cloud cover is small (0.11 and 0.05, respectively). Figure 8a also shows that the largest part of the monthly mean difference between modeled and SEVIRI observed cloud cover is of nocturnal origin. Since the SEVIRI CWP data cover only the portion of the daylight period when the sun is more than 18° above the horizon, we added the monthly mean microwave-based LWP cycle from the UWisc climatology. The model and the two types of satellite observations all agree that the minimum in CWP occurs around local noon. According to the UWisc observations the maximum is reached between 0200 and 0400 LT. In the model the maximum occurs somewhat later (near 0700 LT). The amplitude is small in both the microwave observations (11 g m−2) and in the model simulations (6 g m−2). A similarly good simulation of the diurnal cycles in this region was obtained with the Met Office global forecast model as reported by Allan et al. (2007).
July 2006 mean diurnal cycles of (a) cloud cover and (b) CWP for the southern Atlantic (0°–25°S, 10°–30°W). Black curves represent the satellite data, and colored curves represent output from four RACMO runs.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 mean diurnal cycles of (a) cloud cover and (b) CWP for the southern Atlantic (0°–25°S, 10°–30°W). Black curves represent the satellite data, and colored curves represent output from four RACMO runs.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
July 2006 mean diurnal cycles of (a) cloud cover and (b) CWP for the southern Atlantic (0°–25°S, 10°–30°W). Black curves represent the satellite data, and colored curves represent output from four RACMO runs.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
3) Model experiments
We performed several experiments that improved the simulations in the region of the stratocumulus clouds (see Table 1). In a recent update of the ECMWF model (Cy32r3), the amount of turbulent mixing was reduced where the reduction increased with height above the surface, starting from no change at the surface (Bechtold et al. 2008). This adaptation leads to less entrainment of dry air at the top of stratocumulus layers and therefore to higher cloud covers. Bechtold et al. (2008) found that this led to a better simulation of cloud cover of the stratocumulus fields in the subtropical subsidence regions off the western coasts of the continents. We incorporated this modification into RACMO and performed a sensitivity experiment (EXP3). In the largest part of the domain, the impact of the modification is small as substantiated in Figs. 7 and 8, and Table 4 for the trade wind cumulus region of the southern Atlantic. However, off the coast of Angola (Fig. 7), cloud cover, CWP, and albedo all increased considerably with respect to the reference run, resulting in a better agreement of cloud cover with the observations. On the other hand, in EXP3 simulated CWP rises rapidly in western direction from zero over the continent to values that are, between 500 and 1500 km from the coast, higher than both the SEVIRI and the microwave radiometer observations. In view of the relatively good agreement between the two types of satellite observations and the estimated uncertainty in SEVIRI CWP in this area (≤20%), we interpret this model–observations difference in CWP as a model overestimation. A CWP overestimation would be consistent with an albedo overestimation, but this is not the case. Also, the simulated albedo gradient has the wrong sign in the stratocumulus region.
Qualitatively these two albedo discrepancies (overestimation and wrong sign of the gradient) might be caused by the differences between modeled and observed reff in the stratocumulus region. While simulated reff hardly varies across the transect, observed values decline eastward within the last 1500 km toward the coast to reach a minimum value of approximately 5 μm. Indeed, smaller values of model reff would increase the model albedo. Hence, a decrease in reff toward the coast would lead to an increase in albedo in the same direction. To determine the effect of the observed spatial variation in reff on the albedo, the SEVIRI reff field of Fig. 2g was imposed on the model in a further sensitivity experiment (EXP4). As expected, albedos of the stratocumulus cloud fields increased, and a better match with the observed albedo was obtained. It is interesting to note that the albedo increase due to imposing the field of observed reff is more or less constant within the first 1000 km off the coast. This is caused by two competing effects: the change in reff increases toward the coast causing enhancement of the effect in eastern direction, whereas CWP and therefore optical thickness decrease toward the coast causing mitigation of the effectiveness of the change in eastern direction. We like to add here that the modification of EXP4 had nonnegligible effects in some other parts of the model domain, for example, parts of the continental ITCZ.
In a final sensitivity experiment (EXP5), we set the model inhomogeneity factor finh equal to 1.0. In normal mode this factor has a value of 0.7 and is used to convert model CWP into a virtual CWP for the computation of radiative fluxes. This factor (≤1) accounts for the fact that horizontal inhomogeneities within clouds cause a reduction in the areal mean cloud albedo and an increase in areal mean cloud transmissivity as compared to the albedo and the transmissivity of horizontally homogeneous clouds. The model value of 0.7 is applied to all cloud types, but this value is likely much too low for the relatively homogeneous stratocumulus fields (de Roode and Los 2008). Figure 7 shows that, as expected, by setting finh = 1.0, the TOA albedo increases along the entire transect, which brings the modeled albedo closer to the observations near the Angolan coast. Admittedly, simulated and observed albedo in the stratocumulus region still show notable disagreement in terms of the zonal gradient. This might be linked to the disagreement between the zonal gradients of modeled and observed CWP in the same area.
In EXP5 the simulated albedo of the trade wind cumulus regime is also raised with respect to EXP4, which is undesired. However, in this regime, where clouds are much more inhomogeneous than in the stratocumulus regime, finh = 0.7 is perhaps more realistic than finh = 1.0. Obviously, finh should not be a model constant but should be prescribed as a function of cloud cover and/or other resolved cloud properties.
We conclude that the discussed modifications (reduced turbulent mixing above the boundary layer, prescription of observed reff, and finh = 1) all lead to a better simulation of the stratocumulus fields off the coast of Angola, both in terms of cloud cover and TOA albedo. However, gradients of the simulated albedo and of CWP as well as the absolute value of modeled CWP still show substantial differences with the observations.
e. Representativeness of July 2006
To finish, Fig. 9 presents model–data differences of net total TOA radiation for two months, namely, July 2006 (left) and June 2005 (right). The resemblance of these two graphs is striking. We also produced figures of the model biases in albedo and OLR for June 2005 (not shown) and compared these to results for July 2006. Again the resemblance was striking. Hence, the resemblance of the net total TOA radiation is not due to compensating effects in the short and longwave contributions to the flux. Since the TOA fluxes are crucial model variable, which integrate much of the variations in the cloud parameters that we evaluated, we conclude that our results are not specific for July 2006, but that they are likely typical for boreal summer months.
Mean (a) July 2006 and (b) June 2005 net total TOA radiation difference between the reference simulations (EXP0) and the GERB observations.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Mean (a) July 2006 and (b) June 2005 net total TOA radiation difference between the reference simulations (EXP0) and the GERB observations.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
Mean (a) July 2006 and (b) June 2005 net total TOA radiation difference between the reference simulations (EXP0) and the GERB observations.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
4. Conclusions and discussion
In this paper we reported on an evaluation of regional climate model simulations of Africa and parts of the oceans surrounding this continent for a boreal summer month (July 2006). We simultaneously analyzed cloud properties and TOA radiative fluxes since the latter are strongly related to clouds. By exploiting the products retrieved from the measurements made by SEVIRI and GERB, two sensors on-board the geostationary Meteosat satellite, the diurnal cycle was very well resolved. Such a setup is particularly suitable for testing parameterizations of physical processes in the RCM.
In large parts of the model domain, differences between monthly mean simulated and observed TOA fluxes exceeded the uncertainty in the observations, so these differences are significant and can be referred to as “model biases.” We focused our investigations on regions with substantial and widespread biases in the TOA fluxes, the first being the Sahara and parts of southern Africa, where clear skies prevail. Here, albedo biases were traced back to biases in the clear-sky albedos. Using MODIS surface albedos (EXP1 and EXP2) instead of the ECOCLIMAP albedos from the reference run (EXP0), the model albedo bias was considerably reduced, though residual model–data differences in clear-sky albedo remained (Fig. 4). Since the relative magnitude of the residuals varied considerably across the domain, a calibration error can be discarded as the only cause of the residuals. It is striking that the residuals are mostly negative in the Sahara and mostly positive in southern Africa. This hemispheric difference suggests that the clear-sky albedo residuals could partly be due to errors in the angular-dependence models used to convert GERB radiances into fluxes. Such errors were also mentioned in Allan et al. (2005). Also, interannual variations in the surface albedo might to some extent explain the residuals in clear-sky albedo. While the MODIS dataset is, in line with the purpose of a climate model, a climatology (5-yr mean; 2000–04), the model simulations are for a single year (2006) only. Hence, differences between the albedo for 2006 and the climatology, probably mainly caused by interannual variations in vegetation, contribute to the differences between modeled and observed clear-sky albedo.
We also looked in some detail at the simulation of the stratocumulus fields off the coast of Angola, where the observations of cloud amount and CWP and relatively accurate. In the reference run, TOA albedo, cloud amount, and CWP were all underestimated by the model. Model performance in this region was then improved by the following modifications to the code: a reduction of the amount of turbulent mixing, prescription of cloud effective droplet radii from SEVIRI data, and setting the cloud inhomogeneity factor equal to 1. These suggestions for model improvement might be useful for the development of other climate models but before implementing them their effect needs to be analyzed for other climate regions. An example of an adverse effect of one of the modifications that appeared to be beneficial for the stratocumulus region is the undesired increase in the calculated TOA albedo of the trade wind cumulus regions when the cloud inhomogeneity factor is set equal to 1.
Since within the continental ITCZ the difference between modeled and observed CTT25 clearly exceeded the uncertainty in the observations, we concluded that model CTT25s were too high. This forms a good explanation for the model bias in OLR in the same region. By comparing simulated cover by high clouds and CTT05 with satellite observations, we inferred that too thin high clouds were the most likely cause of the high OLRs and CTT25s in the ITCZ. We intend to confirm this result in future by comparing our model simulations with CloudSat/Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) observations.
In contrast to the reasonable accuracy of cloud cover and CWP in clear-sky regions like the Sahara and southern Africa and the region of the semipermanent stratocumulus clouds, in many other regions of the domain SEVIRI cloud amount is burdened with a large error bar and CWP and reff uncertainty estimates do not exist at all. In addition, our estimate of the uncertainty in cloud cover, related to quantifying cloud-contaminated pixels, forms a lower estimate of the true uncertainty. Because of these difficulties to assess observational errors, model-observation differences in CWP and reff in the regions of the trade wind cumulus clouds and in the ITCZ were termed “differences” instead of “biases.” This makes the comparison of modeled with observed cloud cover in these regions more complicated. Where the model–observation difference exceeds the uncertainty in the observations, like in the western part of the ITCZ, the model–observation difference likely reveals a model bias. Where the opposite occurs, like in the eastern part of the ITCZ, no conclusions regarding the difference can be drawn. So, in many cases we were not able to conclusively evaluate cloud properties on the basis of a comparison between model computations and observations themselves. However, the synergy of the simultaneous evaluation of TOA fluxes and cloud properties enabled us to infer some indications about the validity of the cloud properties.
A good example of this kind of inference is the evaluation of the trade wind cumulus clouds. In theory the model overestimation of the albedo of this region must be caused by either a model overestimation of cloud cover or a model overestimation of cloud optical thickness, which in turn could be due to an overestimation of CWP or an underestimation of reff. A combination of these potential causes is possible as well, of course. We argued that model overestimation of cloud cover and underestimation of reff is unlikely, which leaves overestimation of CWP as the most plausible cause of the albedo bias. Indeed, modeled CWP is larger than SEVIRI CWP but modeled CWP is almost equal to microwave LWP (most cloud water is liquid in the considered region). The inference is that the microwave measurements themselves are probably too high, which is in accordance with the findings of Seethala and Horváth (2010). We like to add here that correct simulation of the trade wind cumulus regime is found to be a problem for many climate models. Williams and Tselioudis (2007) evaluated six GCMs and report that none of the GCMs did a good job at simulating this cloud regime.
We finally like to remark that the retrieval of cloud parameters from radiances measured by satellites is inconsistent with the ways these properties are determined from model output. In the future, we plan to determine model cloud parameters also with a SEVIRI simulator, so that such inconsistencies are eliminated and a more proper comparison of the model output with the observations can be made.
Acknowledgments
We like to thank two anonymous reviewers, whose comments and suggestions contributed significantly to this paper. Rainer Hollmann (DWD) kindly provided us the SEVIRI data. We are also very grateful to the following KNMI colleagues for their assistance: Bert van Ulft wrote much of the necessary RACMO software, Rob Roebeling and Piet Stammes stimulated us with their discussions, Ping Wang performed valuable radiative transfer calculations, and Gerd-Jan van Zadelhoff and Maarten Sneep supported us with their knowledge of IDL. This study was carried out within the framework of SRON project EO-073 (Synthesis of active and passive measurements for climate model improvement; SYNTHESIS).
APPENDIX
Computation of the Clear-Sky Albedo
This appendix describes the details of the method used to compute the clear-sky albedo product from the GERB TOA albedo observations. After collecting all TOA albedo measurements for a pixel, the following steps are made (see Fig. A1):
Divide all samples into subsets. Each subset consists of the data for a certain image time of the day, resulting in a maximum of 31 samples per subset.
Select the second lowest value
Count ni, the number of values within 0.04 from
Discard
Consider all remaining
- As the daylight curve of αclr must be a smooth curve, noise is suppressed by computing αclr,j for each image time j as the average of all
Determine the quality of the clear-sky albedo by computing the root-mean-square of the differences (rmsd) between αclr,i and
In case the rmsd is above a threshold (0.01) or N below a threshold (200), no clear-sky albedo is assigned to the pixel in question. Note that this applies to the complete daytime cycle of the pixel. This masking out occurs mainly in regions that are more or less continuously covered by clouds, for example, the ITCZ. It is this step of the method that favors high-resolution data because the chance of having a clear sky increases with increasing spatial resolution.
All GERB albedos for a particular HR pixel in Mozambique as a function of solar time (circles). For each time of the day the preliminary clear-sky albedo, that is, the second lowest albedo, is marked by a triangle. The smooth curve shown by the line is computed from these points and represents the final clear-sky albedo. The length of the bars corresponds to the number of albedos at each image time within 0.04 from the preliminary clear-sky albedo ni. For this pixel the rmsd = 0.005 and N = 433 (see text). The surface corresponding to this pixel is covered by dark vegetation.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
All GERB albedos for a particular HR pixel in Mozambique as a function of solar time (circles). For each time of the day the preliminary clear-sky albedo, that is, the second lowest albedo, is marked by a triangle. The smooth curve shown by the line is computed from these points and represents the final clear-sky albedo. The length of the bars corresponds to the number of albedos at each image time within 0.04 from the preliminary clear-sky albedo ni. For this pixel the rmsd = 0.005 and N = 433 (see text). The surface corresponding to this pixel is covered by dark vegetation.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
All GERB albedos for a particular HR pixel in Mozambique as a function of solar time (circles). For each time of the day the preliminary clear-sky albedo, that is, the second lowest albedo, is marked by a triangle. The smooth curve shown by the line is computed from these points and represents the final clear-sky albedo. The length of the bars corresponds to the number of albedos at each image time within 0.04 from the preliminary clear-sky albedo ni. For this pixel the rmsd = 0.005 and N = 433 (see text). The surface corresponding to this pixel is covered by dark vegetation.
Citation: Journal of Climate 24, 15; 10.1175/2011JCLI3856.1
REFERENCES
Allan, R. P., A. Slingo, S. F. Milton, and I. Culverwell, 2005: Exploitation of Geostationary Earth Radiation Budget data using simulations from a numerical weather prediction model: Methodology and data validation. J. Geophys. Res., 110, D14111, doi:10.1029/2004JD005698.
Allan, R. P., A. Slingo, S. F. Milton, and M. E. Brooks, 2007: Evaluation of the Met Office global forecast model using Geostationary Earth Radiation Budget (GERB) data. Quart. J. Roy. Meteor. Soc., 133, 1993–2010.
Bechtold, P., M. Koehler, 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 scales. Quart. J. Roy. Meteor. Soc., 134, 1337–1351.
Bennartz, R., P. Watts, J. F. Meirink, and R. Roebeling, 2010: Rainwater path in warm clouds derived from combined visible/near-infrared and microwave satellite observations. J. Geophys. Res., 115, D19120, doi:10.1029/2009JD013679.
Boers, R., J. B. Jenssen, and P. B. Krummel, 1998: Microphysical and shortwave radiative structure of stratocumulus clouds over the Southern Ocean: Summer results and seasonal differences. Quart. J. Roy. Meteor. Soc., 124, 151–168.
Champeaux, J. I., V. Masson, and F. Chauvin, 2003: ECOCLIMAP: A global database of land surface parameters at 1 km resolution. Meteor. Appl., 12, 29–32.
Chen, R., R. Wood, Z. Li, R. Ferraro, and F. L. Chang, 2008: Studying the vertical variation of cloud droplet effective radius using ship and space-borne remote sensing data. J. Geophys. Res., 113, D00A02, doi:10.1029/2007JD009596.
Christensen, J. H., E. Kjellström, F. Giorgi, G. Lenderink and M. Rummukainen, 2011: Assigning relative weights to regional climate models: Exploring the concept. Climate Res., 44, 179–194.
Clerbaux, N., and Coauthors, 2009: Comparison of GERB instantaneous radiance and flux products with CERES Edition-2 data. Remote Sens. Environ., 113, 102–114.
da Rocha, R. P., C. A. Morales, S. V. Cuadra, and T. Ambrizzi, 2009: Precipitation diurnal cycle and summer climatology assessment over South America: An evaluation of Regional Climate Model version 3 simulations. J. Geophys. Res., 114, D10108, doi:10.1029/2008JD010212.
de Bruijn, E. I. F., and E. van Meijgaard, 2005: Verification of HIRLAM with ECMWF physics compared with HIRLAM reference versions. HIRLAM Tech. Rep. 63, 39 pp.
de Roode, S. R., and P. G. Duynkerke, 1997: Observed Lagrangian transition of stratocumulus into cumulus during ASTEX: Mean state and turbulence structure. J. Atmos. Sci., 54, 2157–2173.
de Roode, S. R., and A. Los, 2008: The effect of temperature and humidity fluctuations on the liquid water path of non-precipitating closed cell stratocumulus clouds. Quart. J. Roy. Meteor. Soc., 134, 403–416.
Derrien, M., and H. Le Gléau, 2005: MSG/SEVIRI cloud mask and type from SAFNWC. Int. J. Remote Sens., 26, 4707–4732.
Derrien, M., and H. Le Gléau, cited 2010: Validation report for “cloud products” (CMa-PGE01 v3.0, CT-PGE02 v2.0 & CTTH-PGE03 v2.1). [Available online at http://www.nwcsaf.org/indexScientificDocumentation.html.]
Dewitte, S., L. Gonzalez, N. Clerbaux, A. Ipe, C. Bertrand, and B. De Paepe, 2008: The Geostationary Earth Radiation Budget Edition 1 data processing algorithms. Adv. Space Res., 41, 1906–1913.
Dufresne, J.-L., and S. Bony, 2008: An assessment of the primary sources of spread of global warming estimates from coupled ocean–atmosphere models. J. Climate, 21, 5135–5144.
Futyan, J. M., and J. E. Russell, 2005: Developing clear-sky flux products for the Geostationary Earth Radiation Budget experiment. J. Appl. Meteor., 44, 1361–1374.
Greenwald, T. J., 2009: A 2 year comparison of AMSR-E and MODIS cloud liquid water path observations. Geophys. Res. Lett., 36, L20805, doi:10.1029/2009GL040394.
Harries, J. E., and Coauthors, 2005: The Geostationary Earth Radiation Budget project. Bull. Amer. Meteor. Soc., 86, 945–960.
Keil, A., and J. M. Haywood, 2003: Solar radiative forcing by biomass burning aerosol particles during SAFARI 2000: A case study based on measured aerosol and cloud properties. J. Geophys. Res., 108, 8467, doi:10.1029/2002JD002315.
Klein, S. A., and C. Jakob, 1999: Validation and sensitivities of frontal clouds simulated by the ECMWF model. Mon. Wea. Rev., 127, 2514–2531.
Kuipers Munneke, P., C. H. Reijmer, M. R. van den Broeke, G. König-Langlo, P. Stammes, and W. H. Knap, 2008: Analysis of clear-sky Antarctic snow albedo using observations and radiative transfer modeling. J. Geophys. Res., 113, D17118, doi:10.1029/2007JD009653.
Martin, G. M., D. W. Johnson, and A. Spice, 1994: The measurement and parameterization of effective radius of droplets in warm stratocumulus clouds. J. Atmos. Sci., 51, 1823–1842.
Moody, E. G., M. D. King, C. B. Schaaf, and S. Platnick, 2008: MODIS-derived spatially complete surface albedo products: Spatial and temporal pixel distribution and zonal averages. J. Appl. Meteor. Climatol., 47, 2879–2894.
Nakajima, T., and M. D. King, 1990: Determination of the optical thickness and effective particle radius of clouds from reflected solar radiation measurements. Part I: Theory. J. Atmos. Sci., 47, 1878–1893.
O’Dell, C. W., F. J. Wentz, and R. Bennartz, 2008: Cloud liquid water path from satellite-based passive microwave observations: A new climatology over the global oceans. J. Climate, 21, 1721–1739.
Paeth, H., and Coauthors, 2011: Progress in regional downscaling of West African precipitation. Atmos. Sci. Lett., 12, 75–82, doi:10.1002/asl.306.
Randall, D. A., and Coauthors, 2007: Climate models and their evaluation. Climate Change 2007: The Physical Science Basis, S. Solomon et al., Eds., Cambridge University Press, 589–662.
Roebeling, R. A., and E. van Meijgaard, 2009: Evaluation of the daylight cycle of model-predicted cloud amount and condensed water path with observations from MSG SEVIRI. J. Climate, 22, 1749–1766.
Roebeling, R. A., A. J. Feijt, and P. Stammes, 2006: Cloud property retrievals for climate monitoring: Implications of differences between Spinning Enhanced Visible and Infrared Imager (SEVIRI) on Meteosat-8 and Advanced Very High Resolution Radiometer (AVHRR) on NOAA-17. J. Geophys. Res., 111, D20210, doi:10.1029/2005JD006990.
Sassen, K., Z. Wang, and D. Liu, 2009: Cirrus clouds and deep convection in the tropics: Insights from CALIPSO and CloudSat. J. Geophys. Res., 114, D00H06, doi:10.1029/2009JD011916.
Schulz, J., and Coauthors, cited 2010: Annual validation report of the CM-SAF. CM-SAF Doc. SAF/CM/DWD/VAL/OR5. [Available online at www.cmsaf.eu.]
Seethala, C., and Á. Horváth, 2010: Global assessment of AMSR-E and MODIS cloud liquid water path retrievals in warm oceanic clouds. J. Geophys. Res., 115, D13202, doi:10.1029/2009JD012662.
Su, W., A. Bodas-Salcedo, K.-M. Xu, and T. P. Charlock, 2010: Comparison of the tropical radiative flux and cloud radiative effect profiles in a climate model with Clouds and the Earth’s Radiant Eenergy System (CERES) data. J. Geophys. Res., 115, D01105, doi:10.1029/2009JD012460.
Uppala, S., D. Dee, S. Kobayashi, P. Berrisford, and A. Simmons, 2008: Towards a climate data assimilation system: Status update of ERA-Interim. ECMWF Newsletter, No. 115, ECMWF, Reading, United Kindgom, 12–18.
van Meijgaard, E., U. Andræ, and B. Rockel, 2001: Comparison of model predicted cloud parameters and surface radiative fluxes with observations on the 100 km scale. Meteor. Atmos. Phys., 77, 109–130.
van Meijgaard, E., and Coauthors, 2008: The regional atmospheric climate model RACMO, version 2.1, KNMI Tech. Rep. 302, 43 pp.
Wang, P., W. H. Knap, and P. Stammes, 2011: Cloudy-sky shortwave radiative closure for a Baseline Surface Radiation Network site. J. Geophys. Res., 116, D08202, doi:10.1029/2010JD015141.
Williams, K. D., and G. Tselioudis, 2007: GCM intercomparison of global cloud regimes: Present-day evaluation and climate change response. Climate Dyn., 29, 231–250, doi:10.1007/s00382-007-0232-2.
