Precipitation Climatology in an Ensemble of CORDEX-Africa Regional Climate Simulations

An ensemble of regional climate simulations is analyzed to evaluate the ability of 10 regional climate models (RCMs) and their ensemble average to simulate precipitation over Africa. All RCMs use a similar domain and spatial resolution of ;50 km and are driven by the ECMWF Interim Re-Analysis (ERA-Interim) (1989–2008). They constitute the first set of simulations in the Coordinated Regional Downscaling Experiment in Africa (CORDEX-Africa) project. Simulated precipitation is evaluated at a range of time scales, including seasonal means, and annual and diurnal cycles, against a number of detailed observational datasets. All RCMs simulate the seasonal mean and annual cycle quite accurately, although individual models can exhibit significant biases in some subregions and seasons. The multimodel average generally outperforms any individual simulation, showing biases of similar magnitude to differences across a number of observational datasets. Moreover, many of the RCMs significantly improve the precipitation climate compared to that from their boundary condition dataset, that is, ERA-Interim. A common problem in the majority of the RCMs is that precipitation is triggered too early during the diurnal cycle, although a small subset of models does have a reasonable representation of the phase of the diurnal cycle. The systematic bias in the diurnal cycle is not improved when the ensemble mean is considered. Based on this performance analysis, it is assessed that the present set of RCMs can be used to provide useful information on climate projections over Africa.


Introduction
There is a growing interest from a wide range of decision makers for climate change information on regional to local scales and at high spatial resolution. The primary tool for providing future climate projections is atmosphere-ocean general circulation models (AOGCMs), which simulate climate changes under a range of future greenhouse gas emission scenarios. Present-day AOGCMs have spatial resolution of order 100-250 km. Despite rapid advances in high-performance computing, computational demands still limit the use of highresolution (50 km or better) AOGCMs for multicentury simulations, while a few AOGCM simulations have been run for short time slice periods at such high resolutions (e.g., Oouchi et al. 2006). Because of the dual constraint of providing an ensemble of projections over long periods, AOGCMs cannot fulfill the requirements of high spatial detail required by many users and are therefore generally supplemented with statistical or dynamical downscaling to produce future climate projections at regional scales.
Over the past decade, several international projects have applied regional climate models (RCMs) to generate high-resolution, multimodel ensembles of future climate projections by downscaling output from AOGCMs. These include the Prediction of Regional Scenarios and Uncertainities for Defining European Climate Change Risks and Effects (PRUDENCE) (Christensen et al. 2007) and ENSEMBLE-Based Predictions of Climate Changes and their Impacts (ENSEMBLES) (van der Linden and Mitchell 2009) for Europe; the North American Regional Climate Change Assessment Program (NARCCAP) (Mearns et al. 2009) for North America; Europe-South America Network for Climate Change Assessment and Impact Studies in La Plata Basin (CLARIS-LPB) (Menéndez et al. 2010) over South America and ENSEMBLES-African Monsoon Multidisciplinary Analyses (AMMA) (van der Linden and Mitchell 2009) for West Africa. Each of these projects has made significant contributions to downscaling efforts over their specific region, but there has been very limited international coordination of such projects and therefore limited transfer of knowledge between projects and regions. A new initiative-the Coordinated Regional climate Downscaling Experiment (CORDEX) Jones et al. 2011), sponsored by the World Climate Research Programme-aims to fill this gap by coordinating international efforts in regional climate downscaling. CORDEX is organized in a similar manner to the Coupled Model Intercomparison Project phase 5 (CMIP5) for global model simulations, with predefined regions, grids, experiment protocols, output variables, and output format facilitating easier analysis of possible future regional climate changes, not only by the scientific community, but also by end-user communities at regional and local levels.
Twelve regions, covering the majority of the populated land areas worldwide, plus both the Arctic and Antarctic, have been defined within CORDEX. An initial focus in CORDEX will be on Africa, which is particularly vulnerable to climate change and has a low adaptive capacity. For example, water supply and food security are of critical importance in Africa; we therefore emphasize the quality of RCM-simulated precipitation as a major driver for both sectors over Africa. Many studies have examined the performance of a single RCM, driven by the reanalysis, in simulating precipitation over different parts of Africa: the fifthgeneration Pennsylvania State University-National Center for Atmospheric Research Mesoscale Model (MM5) (Tadross et al. 2006), the Weather Research and Forecasting model (WRF) (Patricola and Cook 2010), the Regional Climate Model, version 3 (RegCM3) (Afiesimama et al. 2006;Anyah and Semazzi 2007;Pal et al. 2007;Sylla et al. 2009Sylla et al. , 2010, and the Regional Model (REMO) (Paeth et al. 2005;Haensler et al. 2010). However, few studies have focused on simulated precipitation from an ensemble of different RCMs, as, for example, Druyan et al. (2010) and Paeth et al. (2011) have recently done for West Africa. All studies indicate that RCMs are able to capture the main climatological features of precipitation but with quite different levels of accuracy depending on the model, region, and season, with the multimodel average generally outperforming individual RCMs (Paeth et al. 2011).
In this study we present the first results of the CORDEX-Africa project by evaluating an ensemble of 10 RCM simulations, covering the entire African continent and driven by the European Centre for Medium-Range Weather Forecasts (ECMWF) Interim Re-Analysis [ERA-Interim (ERAINT)]. The performance of the individual RCMs and their ensemble average is documented in detail at a range of time scales. We begin by evaluating simulated seasonal mean climatologies, followed by a more detailed analysis of seasonal precipitation associated with the West African monsoon (WAM). We then evaluate the simulated annual and diurnal cycles of precipitation over a number of different African regions characterized by different climate regimes. In next section we first present a brief description of the data, models, and methods.

a. Model simulations
An ensemble of 10 different RCMs is utilized for this study. A full list of the participating RCMs and details of their dynamics and physical parameterizations is presented in Table 1. All simulations are performed at ;50-km resolution over the same CORDEX Africa domain ( Fig. 1) and use ERA-Interim (1989-2008 as lateral boundary conditions. No nudging toward ERAINT was applied in any of the RCMs within the model domain. One model-Action de Recherche Petite Echelle Grande Echelle (ARPEGE)-is a global variable resolution model. ARPEGE uses a high-resolution region, within its continuous global domain, coincident in resolution and geographic area with the RCMs. Within this region no nudging toward ERAINT was applied; however, outside the high-resolution region, nudging was applied to keep the inflow into the high-resolution region comparable to the boundary data specified for the other RCMs. With respect to spectral nudging of an RCM solution toward the driving data at large wavelengths (von Storch et al. 2000), this technique is well established for midlatitude regions, with some theoretical understanding of which wavelengths should be nudged and at what altitudes (Alexandru et al. 2009). This is not the case in the tropics, and it may be more difficult to formulate given the stronger role of surface forcing and multiscale convection in driving large-scale circulations. We therefore chose to preclude spectral nudging from the experimental design, pending further work in this area. The ERAINT sea surface temperature (SST) is used as the ocean boundary conditions in 8 of the 10 RCMs, while RegCM uses the National Oceanic and Atmospheric Administration (NOAA) weekly optimum interpolation SST and the Canadian Regional Climate Model (CRCM) uses the Atmospheric Model Intercomparison Project, phase 2 (AMIP2) SST. For analysis we collect 3-hourly total precipitation, averaged for 0000-0300, 0300-0600, . . . 2100-0000 universal time (UTC), which is then also averaged into daily and monthly means.

b. Observations and reanalysis
One of the main problems in evaluating RCM simulations over Africa is a lack of high-quality observation datasets at suitable temporal and spatial resolution. In recent decades satellite measurements have partly improved the situation, and here we use two satellite-based precipitation products with 0.258 spatial and 3-hourly temporal resolution: the Tropical Rainfall Measuring Mission (TRMM 3B42 version 6, 1998(TRMM 3B42 version 6, -2010Huffman et al. 2007) and the Climate Prediction Center morphing technique (CMORPH, 2003-10;Joyce et al. 2004). CMORPH represents 3-hourly precipitation estimates for the same time intervals as the RCMs, while TRMM precipitation averages are shifted by one and onehalf hours, representing approximately the 22: [1][2][3][4][19][20][21][22]. Several studies report that the timing of peak rainfall during the day in satellite products, including infrared (IR) sensor data (both TRMM-3B42 and CMORPH in our case), may be delayed by ;3 h compared to radar/microwave satellite precipitation estimates. This is because IR sensors do not directly detect rain, but rather brightness temperature at the top of convective clouds (Kikuchi and Wang 2008;Yamamoto et al. 2008). However, Pinker et al. (2006) show that for a site in Nigeria, CMORPH is exactly in phase with ground-based observations. To resolve this issue, we have additionally analyzed the TRMM-3G68 precipitation, which includes precipitation rates from a precipitation radar, microwave imager, and their combination (ftp://trmmopen.gsfc.nasa.gov/pub). In contrast to some earlier studies, we find no significant difference in the daily phase of precipitation over Africa between the two TRMM datasets and therefore use only the CMORPH and TRMM-3B42 as primary observations for analysis of the diurnal cycle.
To estimate observational uncertainties, we include in our analysis four gridded observational datasets with coarser temporal and spatial resolution than TRMM and CMORPH. The Global Precipitation Climatology Project (GPCP, version 1.1, 1998(GPCP, version 1.1, -2010 satellite-gauge combination covers the entire African domain with a 18 spatial and daily temporal resolution (Huffman et al. 2001). Three gauge-based datasets are available at 0.58 spatial and monthly temporal resolution: the University of Delaware (UDEL version 2.01, 1901(UDEL version 2.01, -2008Legates and Willmott 1990), the Climatic Research Unit at the University of East Anglia (CRU version 3.1, 1901(CRU version 3.1, -2009Mitchell and Jones 2005), and the Global Precipitation Climatology Centre (GPCC version 5, 1901(GPCC version 5, -2009Rudolf et al. 2010). Each dataset covers different periods. We use three analysis periods for different aspects of the African precipitation climatology and try to include as many observational datasets as possible in each analysis. The periods are 1990-2008 (1989 is spinup) for interannual variability analysis, 1998-2008 (limited by TRMM and GPCP) for observational uncertainty/seasonal mean/ annual cycle, and 2003-08 (limited by CMORPH) for the diurnal cycle. Although some of these periods are rather short, we find that the obtained results are insensitive to the period chosen, for example, longer periods-1990-2008 instead of 1998-2008 for the seasonal mean (without GPCP and TRMM) or 1998-2008 instead of 2003-08 for the diurnal cycle (without CMORPH)-do not change the results.
Precipitation from ERAINT is also compared to the RCM results and observations. We note that ERAINT precipitation is a forecast product, and there are several ways to derive ERAINT precipitation (e.g., different spinup, base time, and forecast steps) that can lead to different precipitation estimates (Dee et al. 2011). Two methods (with and without spinup) to derive the 3-hourly and daily precipitation have been tested, and we have found only minor differences in the results obtained with these two methods. Given the uncertainty as to which method is most appropriate, we derive ERAINT precipitation by the simplest method, without spinup: 3-hourly precipitation uses the base times 0000 and 1200 and forecast steps 3, 6, 9, and 12 h, while daily precipitation uses base times 0000 and 1200 and forecast steps of 12 h. Monthly means are calculated by averaging the daily means.

c. Methods
Most RCM simulations are performed on the same rotated grid with 0.448 horizontal resolution. This is used as a reference grid, and all datasets on other grids are remapped to this reference grid. The TRMM and CMORPH datasets are aggregated from the native 0.258 grid onto the reference grid by a first-order conservative remapping (Jones 1999), while all other datasets, with similar or coarser resolution than 0.448, are remapped through bilinear interpolation.
The climatological annual cycle, calculated from daily data, has a significant amount of high-frequency noise due to the relatively short sampling period and large day-to-day precipitation variability. To highlight the slowvarying annual cycle, we apply a finite impulse response filter (Smith 2002) with a 50-day cutoff period, which is approximately equivalent to the widely used 30-day moving average but with better frequency response. The filter length is chosen equal to the cutoff period in days.
The climatological diurnal cycle is estimated from 3-hourly precipitation. A common way to determine the phase and amplitude of the diurnal cycle based on discrete data is to approximate it by the diurnal and semidiurnal (Fourier) harmonics (Dai et al. 2007;Jeong et al. 2011). As the harmonic analysis is not the only available method and may be not the best approach, especially if the diurnal cycle is nonharmonic (Yang and Smith 2006), we also fit a cubic spline to the 3-hourly precipitation. Although the results based on the harmonic and cubic spline approaches are quite similar, on average the cubic spline gives a better description of the diurnal cycle, capturing smaller-scale features. We have tested several precipitation thresholds in calculating the diurnal cycle, by including in our analysis only days with total daily accumulations greater than 1, 2, and 3 mm day 21 . Use of these different thresholds results in different amplitudes of the diurnal cycle but does not affect the phase. We therefore decided to impose an accumulation threshold of 1 mm day 21 to determine which days to be included in our calculation of the diurnal cycle to exclude nonrepresentative dry days.

Observational uncertainties
We first evaluate the spread among the different observational datasets over Africa to get an estimate of the observational accuracy available for model evaluation. All the gridded products agree quite well with respect to large-scale precipitation patterns, but significant deviations can occur locally (Gruber et al. 2000;Fekete (1980) Tiedtke (1989) Radiation scheme Morcrette (1990) Fouquart and Bonnel (1980); Mlawer et al. (1997) Kiehl et al. (1996 Ritter and Geleyn (1992) Turbulence vertical diffusion Mellor and Yamada (1982) Louis ( indicates that TRMM has a large dry bias over southern tropical Africa compared to GPCP, with relative differences of as much as 50% in some regions. The gauge-based precipitation datasets and GPCP show much better consistency, although some significant differences can still be seen on smaller spatial scales. Similar results hold for other seasons (not shown), with a common tendency for TRMM to have less precipitation than GPCP and the gauge products to be in better agreement. ERAINT precipitation is heavily biased (wet) compared to differences among the observations. The large difference between TRMM and GPCP is explained by the fact that both products are adjusted to large-scale monthly precipitation from gauge networks but use different gauge analysis products. TRMM (version 6) is adjusted to the old GPCC gauge analysis ''monitoring product'' (version 2) for 1998-April 2005 and to the Climate Anomaly Monitoring System (CAMS) gauge analysis product (Xi et al. 1996) thereafter (Huffman and Bolvin 2011). GPCP (version 1.1) is scaled to the GPCC version 4 product using the full analysis up to 2007 and the monitoring product thereafter (Huffman et al. 2009). Globally, the shift from the old GPCC to the newer version 4 causes a 7% increase in precipitation over land in GPCP, with a pronounced increase over Africa (Huffman et al. 2009). The spread among the pure ground-based products (UDEL, CRU, GPCC), which partly share the same gauge stations, may arise from different processing algorithms and different levels of station availability in given time periods, especially in regions with very few or no stations (Yin and Gruber 2010). For example, a detail analysis of the number of gauge stations included in the GPCC product shows that almost no stations report observations over central Africa (primarily Angola, Democratic Republic  (1989) Tiedtke (1989) Kain andFritsch (1990, 1993) Gregory and Rowntree (1990); Gregory and Allen (1991) Kain (2004) Kain and Fritsch (1990); Kuo (1965) Fouquart and Bonnel (1980) Morcrette et al. (1986); Giorgetta and Wild (1995) Savijä rvi (1990); Sass et al. (1994) Edwards and Slingo (1996) Dudhia (1989); Mlawer et al. (1997) Li and Barker (2005) Eddy-diffusivity (first-order K) mass flux approach  Zadra et al. (2008) of the Congo, Tanzania, and Mozambique; see Fig. 3), the region with maximum rainfall in JFM, for all months in the 1998-2008 period. In these regions it is difficult to quantify uncertainties among the observational datasets in terms of which one is more realistic. Because of its use of a more recent gauge dataset for large-scale scaling, we decided to use GPCP as the reference dataset for the evaluation of the seasonal mean and annual cycle, despite its relatively coarse resolution. For evaluation of the diurnal cycle both TRMM and CMORPH are utilized, although the reader should keep in mind that the current version of TRMM may underestimate precipitation amounts. One possible approach for future analysis may be to simply apply an ensemble average to all observational datasets as a reference while retaining an estimate of the interobservational spread, at least for evaluation of the seasonal mean over land.

Seasonal mean precipitation
Figures 4 and 5 show GPCP precipitation for the July-September (JAS) and JFM seasons and differences from GPCP in the other observational datasets, the 10 RCM simulations, as well as ERAINT and the RCM multimodel ensemble. In JAS (Fig. 4) the rain belt associated with the intertropical convergence zone (ITCZ) reaches its most northerly location, stretching from the mountains of Darfur in East Africa to the Guinea Highlands and downstream into the Atlantic. Compared to GPCP, UDEL and GPCC are in good agreement with some tendency for smaller precipitation amounts, while TRMM has a large underestimate over East Africa and the Guinean Highlands and an overestimate over the Cameroon Highlands. As expected, the RCM biases are very diverse and vary significantly across the individual simulations. All RCMs capture the large-scale patterns of JAS precipitation (not shown) but have larger biases than the spread across the GPCP, UDEL, and GPCC products. Nevertheless, biases in many of the RCMs are of comparable magnitude to differences between the TRMM and GPCP observations. The largest differences in JAS are found in West Africa over the Guinean Highlands and downstream, with some models (HIRHAM, REMO, WRF, CRCM) strongly overestimating precipitation, one (ARPEGE) strongly underestimating, and one [the Consortium for Small-Scale Modeling (COSMO) model in climate mode (CCLM)] showing large, oppositely signed biases in a dipole pattern. Almost all RCMs have a large positive precipitation bias northeast of Madagascar. This is an area where significant observed precipitation interacts with the RCM lateral boundary and may, therefore, reflect problems in representing convection in the RCM boundary relaxation zones, where artificial dynamical imbalances can lead to low-level convergence and excessive convection. Oceanatmosphere coupling may also improve the simulated precipitation over the Indian Ocean, as shown by Ratnam et al. (2009) in coupled and uncoupled experiments with RegCM3.
In the JFM season, the ITCZ reaches its most southerly location (Fig. 5), with maxima in southern tropical Africa and Madagascar. Most of the conclusions for JFM are similar to JAS: UDEL and GPCC are in better agreement with GPCP than TRMM, which has a dry bias over land. The RCMs biases vary significantly between individual simulations, with biases again generally larger than across the observation datasets. One common feature is a significant negative bias over Madagascar in most of the RCMs relative to GPCP, which is also seen in the majority of the alternative observation datasets. Many of the RCMs again show positive precipitation biases close to the eastern boundary, collocated with an observed precipitation maximum.
For both seasons the RCM ensemble average outperforms nearly all the individual models, with biases comparable to differences between the observational datasets. Figures 4 and 5 clearly show that much of this good performance of the ensemble mean results from the cancellation of oppositely signed biases, a number of which may reflect precipitation placement errors. The ensemble mean substantially improves the ERAINT precipitation (see also Paeth et al. 2011). Moreover, many individual RCMs have more accurate precipitation than the ERAINT data, which were used as lateral and surface boundary conditions.
To evaluate the seasonal progression of the ITCZ, we use a simple method to approximate its position by calculating the geographical position of maximum precipitation in the center of the rain belt and the 1 mm day 21 mean intensities on either flank of this maximum. Figure 6 plots the ensemble mean, seasonal mean, and precipitation bias with respect to GPCP for all four seasons, together with the approximated position of the ITCZ. While individual RCMs have localized placement errors in terms of the ITCZ seasonal progression, through the cancellation of these errors, the RCM ensemble mean has an extremely accurate estimate of the location of the ITCZ and the maximum precipitation within the ITCZ in all four seasons.
As well as representing seasonal mean climatological precipitation, it is important that RCMs capture the observed interannual variability in precipitation, particularly the variability associated with large-scale modes of the coupled climate system, such as El Niñ o-Southern Oscillation (ENSO) forced modes. Figure 7 displays time series of the spatially averaged, seasonal mean normalized precipitation for three regions: northern West Africa (7.58-158N, 108E-108W), southern central Africa (108-08S, 108-258E), and East Africa (158-08S, 308-408E) as illustrated in Fig. 1. Over West Africa interannual precipitation variability is in very good agreement across the observational datasets. All of them show the same extremely wet or dry years relative to the 1990-2008 climatology, with a small difference in magnitude in some years. ERAINT does not reproduce the observed interannual variability over this region and has an artificial trend not seen in observations. The RCM ensemble average substantially improves on ERAINT, more closely following the observed variations, albeit failing to reproduce 1994, 1995, and 2008. Spread across the observations is larger in East Africa, where one of the observational dataset can show an extremely wet/dry year while others show the same year close to the climatology, for example, 1994. Both FIG. 2. GPCP mean JFM precipitation (1998-2008 and differences between other gridded precipitation products and GPCP. the ERAINT and the ensemble mean, in general, follow the observational envelope but both or one of them fail to reproduce some wet/dry years when the observations are in agreement (1992,2000). Observational uncertainties are even larger in central Africa compared to East Africa, and for many individual years it is impossible to conclude whether it is a wet or dry year. In addition, the CRU precipitation shows no interannual variability at all in 2002-08 in contrast to the GPCC and UDEL. Such large uncertainties among the CRU, UDEL, and GPCC datasets, most likely related to there being almost no observational stations in central Africa (see Fig. 3), preclude us from evaluating the interannual precipitation variability over this region.

West African monsoon rainfall
It is a challenging task for both global and regional climate models to simulate the range of complex processes that make up the WAM (Cook and Vizy 2006;Druyan et al. 2010;Xue et al. 2010;Ruti et al. 2011). In Fig. 8 we plot the observed and simulated annual cycle of precipitation averaged between 108E and 108W. Such a figure depicts the latitudinal progression of the main WAM rainfall, as well as its seasonal intensification and decay. GPCP and TRMM represent the well-known behavior of the WAM: maximum rainfall at the end of May/beginning of June over the Gulf of Guinea (08-68N) followed by an abrupt northward jump from the Gulf of Guinea to the Sahel, with the establishment of a new maximum around 108N in August and a southward retreat thereafter (Sultan and Janicot 2000;Janicot 2009). While the temporal and spatial details are generally similar in both datasets, TRMM exhibits higher rainfall intensities over the Gulf of Guinea, south of 58N, and slightly lower intensities over the Sahel. ERAINT reproduces the large-scale features of the WAM rainfall but differs in rainfall intensities and, in particular, fails to propagate precipitation far enough north in July-August, resulting in a significant dry bias between ;108 and 188N. The individual RCMs reproduce the WAM rainfall with varying degrees of accuracy. Most RCMs capture the two rainfall maxima, over the Gulf of Guinea and the Sahel, although the positioning, intensity, and duration of these maxima differ across the models. Seven of the models [Rossby Centre regional atmospheric model (RCA), HIRHAM, CCLM, REMO, Royal Netherlands Meteorological Office (KNMI) Regional Atmospheric Climate Model (RACMO), Providing Regional Climates for Impacts Studies (PRECIS), and WRF] simulate a northward jump in precipitation from the Gulf of Guinea to the Sahel region, at roughly the correct time of the year, although with varying degrees of accuracy. Four models (HIRHAM, REMO, CCLM, and PRECIS) show a second maximum over the Gulf of Guinea in November that resembles a ''southward jump'' of the WAM, while a more gradual retreat is evident in the observations. Three of these RCMs (HIRHAM, REMO, and CCLM) strongly overestimate rainfall maxima both in May-June and October-November. Some models show very weak maxima: CCLM in Sahel, RegCM3 over the Gulf of Guinea, or no maximum at all as in ARPEGE in the Sahel. Three models (RCA, ARPEGE, and CRCM) do not propagate WAM precipitation sufficiently far north. The multimodel average smoothes these diverse biases and presents the best simulated WAM precipitation. Despite some overestimation of rainfall in May-June, the shape and intensity of the annual cycle is quite accurate, with some evidence of a slight underestimate of the northward extension and the existence of an erroneous secondary jump/maximum in October-November. The RCM ensemble mean WAM rainfall is a significant improvement over that associated with the ERAINT driving data. onset is that most RCMs overestimate precipitation, indicating a too-early start of the rainy season by about one month (e.g., RCA, ARPEGE, and PRECIS). The ensemble mean reproduces the common tendency of the RCMs to overestimate precipitation during the onset but otherwise accurately represents the observed annual cycle, although again we emphasize error cancellation across individual RCMs leads to the accurate ensemble mean estimate. Again, several RCMs and particularly the ensemble mean improve on ERAINT, which has large dry bias during the rainy season. In the bottom row of Fig. 9, we reproduce the same spatial mean, annual cycle plots, but now the full range of the RCM results is plotted in gray shading, while the 25th-75th percentile range is shown in yellow. For West Africa (bottom, right) removing the outlier models significantly improves the accuracy of the ensemble precipitation.

Annual cycle
In central Africa the distinctive feature is the bimodal structure of the annual cycle, with maxima in November and March-April (middle column). Similar to West Africa, both GPCP and TRMM have the same shape of the annual cycle with slightly higher intensities in GPCP. All RCMs capture the basic shape of the annual cycle, on average underestimating the first maximum and overestimating the second. Nearly all the RCMs simulate precipitation more accurately than ERAINT, which FIG. 6. The ensemble mean precipitation bias relative to GPCP for four seasons [including April-June (AMJ) and October-December (OND); color] and the approximate position of the ITCZ in GPCP precipitation (bluewhite line) and in the RCM ensemble mean (red dashed line).The ITCZ position is defined as the maximum precipitation in the center of the rain belt and 1 mm day 21 on its flanks. has too-intense rainfall throughout the year. The ensemble mean and the 50% most accurate RCMs closely follow the observed annual cycle.
Over East Africa (right column), where maximum rainfall occurs in the boreal winter, ERAINT overestimates precipitation, while individual RCMs again show various behaviors grouped around the observations, with the ensemble average again leading to an accurate representation of the annual cycle.

Diurnal cycle
The diurnal cycle is a prominent forced mode of atmospheric variability that can significantly modulate regional, near-surface thermal, and hydrological regimes. Many studies have shown that global and regional climate models generally have a poor representation of the diurnal cycle of precipitation, suffering from biases in the frequency of precipitation occurrence, precipitation intensity, and the timing of rainfall during the day (e.g., Dai et al. 1999;Dai and Trenberth 2004;Shin et al. 2007;da Rocha et al. 2009;Jeong et al. 2011). In the tropics, the main features of the diurnal cycle have been well documented from ground-and satellite-based observations and include a late afternoon/evening maximum over land and an early morning maximum over open ocean (Dai 2001;Yang and Slingo 2001;Yang and Smith 2006;Dai et al. 2007). Figure 10 plots the local time of maximum rainfall in the diurnal cycle for the JAS season, with the black lines indicating the extent of the ITCZ for this season, as depicted in Fig. 5. While there is a fair amount of small-scale variability, both TRMM and CMORPH indicate rainfall maxima between late afternoon and midnight, or even later. The near-midnight maximum is related to the fact that the majority of precipitation produced in the ITCZ region during this season is produced by mesoscale convective systems. These are usually initiated around 1700-1800 local solar time (LST) but precipitate at maximum intensity later in the day, during their mature phase (McGarry and Reed 1978;Hodges and Thorncroft 1997). ERAINT precipitates too early, with the widespread occurrence of maximum precipitation intensity around local noon. The majority of RCMs exhibit the same out-of-phase diurnal cycle, with only RCA and CRCM capturing to some degree the observed phase of the diurnal cycle.
The diurnal cycle of precipitation is sensitive to the convective parameterization used (Dai et al. 1999;Liang et al. 2004). One possible explanation for the more realistic diurnal cycle in RCA and CRCM is that both models employ the Kain-Fritsch (KF) convection scheme (Kain and Fritsch 1990;Bechtold et al. 2001), while ERAINT and most of the other RCMs have different schemes (see Table 1). Liang et al. (2004) evaluated the diurnal cycle of precipitation over the United States in a regional model and found the KF scheme to be more accurate than the Grell scheme (Grell 1993) in representing late afternoon rainfall peaks when moist convection is controlled by near-surface forcing. However, it is not clear why WRF, which also has the KF scheme, precipitates too early.
A more detailed analysis of the diurnal cycle, spatially averaged over West, central, and East Africa is presented in Fig. 11. In West Africa for the JAS season, TRMM and CMORPH show basically the same shape and amplitude of the diurnal cycle, with an evening to midnight maximum. CMORPH generally has higher precipitation intensities than TRMM. Only one model-CRCM-captures both the shape and amplitude of the observed diurnal cycle, while RCA has a too-flat diurnal cycle over this region. In agreement with Fig. 10, convective precipitation in all other models and ERAINT is triggered too early, between 0800 and 1000 LST with maximum precipitation rates generally occurring around local noon. REMO has the latest maximum of this group, with peak intensities at 1500 LST. The ensemble mean can partly correct the amplitude of the diurnal cycle, which is very different among the individual simulations but not the phase, with a maximum between 1300 and 1400 LST. Similar behavior is evident over both central and East Africa for FIG. 7. Seasonal mean normalized precipitation over (top) northern West Africa (7.58-158N, 108W-108E), (middle) southern central Africa (108-08S, 108-258E), and (bottom) East Africa (158-08S, 308-408E). Gray shading is the full range of the RCM results, and yellow shading is the 25th-75th percentile range. the DJF season (Fig. 11), where the observed rainfall maximum occurs between 1700 and 1800 LST. Again, CRCM has the most accurate diurnal cycle, with RCA having a more pronounced diurnal cycle over these two regions, although with peak intensities 2-3 h too early. ERAINT and all other RCMs again have maximum rain rates around local noon, which is also seen in the ensemble average.
One approach, to better understand the diurnal cycle of precipitation, is the use of frequency-intensity plots (FI) based on 3-hourly data. Figure 12  rates from ;0800 to 1500 LST (the red area on the figure) and a temporally collocated minimum in the occurrence of heavier rain rates in excess of 1 mm 3 h 21 . Between ;1800 and 2100 LST, both observations now indicate a minimum in the occurrence of light/zero rain rates and a relative maximum in intensities greater than ;0.5 mm 3 h 21 .
ERAINT is almost a mirror image of TRMM, with completely the opposite phase; the highest frequency of light/zero rainfall between 1800 and 0800 LST; and the maximum occurrence of intense precipitation around local noon. Four RCMs-HIRHAM, RACMO, CCLM, and REMO, which all employ different versions of the Tiedke convection scheme-show similar diurnal intensity distributions to ERAINT, which also has the Tiedke scheme employed in the ECMWF model used to generate ERAINT. Of the three models employing the KF scheme (CRCM, RCA, and WRF), only CRCM captures the main features of the observed FIs, while RCA has a tooflat distribution and WRF, somewhat puzzlingly, has a distribution very similar to models employing the Tiedke scheme. All other RCMs with different convective schemes (ARPEGE,RegCM3,and PRECIS) have FIs similar to the Tiedke family of models, with the highest frequency of light/zero rainfall between late afternoon and morning and the maximum occurrence of intense precipitation centered at local noon. Figure 12 emphasizes the main RCM deficiency in simulating precipitation over Africa remains a poor representation of the diurnal cycle. Models employing the KF scheme appear to do better in this regard; two possible factors contributing to this improvement may include (i) the use of an advanced convective trigger function and (ii) that entrainment and detrainment processes are responsive to environmental conditions, through the buoyancy sorting approach (Kain and Fritsch 1990). More research is planned to understand the factors behind the good representation of the diurnal cycle in the CRCM model and differences seen relative to RCA. These findings may help improve the diurnal cycle in some of the other RCMs.

Summary and conclusions
We present a first evaluation of the precipitation climatology over the African continent in an ensemble of regional climate simulations performed within the CORDEX-Africa project. The ensemble consists of 10 different RCMs all run at ;50-km resolution, on a common grid, and driven by ERA-Interim for the period 1989-2008. Performance of the individual models and the ensemble average is evaluated in detail for different aspects of African precipitation, namely, seasonal means, West African monsoon rainfall, and annual and diurnal cycles.
Several gridded precipitation products, based on ground and/or satellite-derived observations, are included in the FIG. 11. Diurnal cycle of precipitation over (left) northern West Africa (7.58-158N, 108W-108E), (middle) southern Central Africa (108-08S, 108-258E), and (right) East Africa (158-08S, 308-408E) for the 2003-08 period. Gray shading is the full range of the RCM results, and yellow shading is the 25th-75th percentile range. evaluation to encompass uncertainties in the observational data. Intercomparison at seasonal mean time scales reveals that two satellite products-TRMM and GPCP-can differ by as much as 50% over large areas due to the adjustment of large-scale satellite estimates to different gauge products. Differences between pure gauge products-CRU, UDEL, GPCC-and the combined satellite-gauge dataset GPCP are, on average, smaller than differences between TRMM and GPCP. It is not possible to derive a thorough estimate of the quality of the gauge products as they are not independent of each other. Furthermore, there are very few or no reporting gauge stations over large parts of Africa. Interannual precipitation variability over regions of low surface observational density is very different across the gauge products, although climatological estimates for sufficiently long periods are more consistent. An ensemble mean approach, similar to that employed with climate models, may partly improve the situation, by reducing random errors in the observational products.
On average all RCMs capture the main features of the seasonal mean rainfall distribution and its annual cycle, although significant biases can be found in individual models depending on region and season. The fact that individual model biases vary considerably in space and time has also been found in earlier multimodel studies for West Africa (Druyan et al. 2010;Paeth et al. 2011). The multimodel average generally, but not always, outperforms many of the individual models with biases of similar magnitude to differences between the observational datasets. Such good performance is mostly a result of the cancelation of opposite-signed biases across the models. These findings are similar to those of Paeth et al. (2011) and emphasize the importance of using multimodel ensembles for simulating African climate. Nevertheless, many of the individual RCMs do improve the precipitation climate compared to their boundary condition dataset, ERA-Interim, which has biases significantly larger than the spread among the observations. Similar improvements over West Africa were also found by Paeth et al. (2011). Druyan et al. (2010) showed that for 4-yr simulations, initialized every year, all participating RCMs exhibited a positive precipitation bias, consistent with the bias in the boundary condition data-National Centers for Environmental Prediction-Department of Energy(NCEP-DOE) Reanalysis 2. This may be related to the short simulation period. The consistent improvement seen for the RCMs in this study indicates the potential advantage of applying high-resolution regional models which, by resolving small-scale processes, can improve on the large-scale boundary condition data in terms of societally important variables, such as precipitation. Individual RCMs simulate WAM precipitation with differing levels of accuracy. Most capture the WAM rainfall maxima, over the Gulf of Guinea in late Mayearly June and in the Sahel region in August, although the positioning, intensity, and duration of these maxima differ across models. A number of the RCMs show a tooearly onset of the WAM, but the majority captures its rapid northward progression into the Sahel by early July. Several models exhibit a second rainfall maximum over the Gulf of Guinea in November that is not supported by observations. As for the annual cycle, the multimodel ensemble mean generally outperforms individual RCMs, providing a significant improvement over the ERA-Interim WAM precipitation, which fails to propagate sufficiently far enough north.
The most challenging task over Africa remains the diurnal cycle of precipitation, with the majority of RCMs and ERA-Interim simulating a diurnal cycle completely out of phase with observations. The main problem appears to be related to the formulation of the convective parameterizations employed in the models, with a subset of the models that employ the Kain-Fritsch scheme showing a somewhat improved representation of the phase of the diurnal cycle. A thorough understanding of why these models show an improved diurnal cycle is deferred to a later study. When the ensemble mean is considered, the amplitude of the diurnal cycle is significantly improved but not the systematic bias in the phase. As projected climate changes may be sensitive to the diurnal precipitation characteristics (e.g., Lynn et al. 2007), model improvements, such as an improved diurnal cycle, may lead to more reliable future projections.
To summarize, most of the RCMs capture the main details of the precipitation climatology over Africa, although individual models do exhibit substantial biases depending on region, season, and evaluation metric. Overall, results of the first CORDEX evaluation over Africa are encouraging. For many of the RCM groups, this is the first experience of dynamical downscaling over Africa, where land-convection interactions play a leading role in defining details of the precipitation climate. The group of RCMs employed in this study will soon perform transient climate change projections within CORDEX, forced by coupled GCM results from CMIP5. This study provides the underlying assessment of the ability of these RCMs to describe African precipitation processes that are key to the management of freshwater resources and food security in the region.
Finally, we would emphasize that the ensemble average should not be viewed as the expected outcome in CORDEX-generated climate projections, although it outperforms most of the individual RCMs. Rather, we emphasize the need for a large ensemble of projections (as many as possible) to develop estimates of likelihood of occurrence based on agreement across the range of plausible AOGCM-RCM-simulated climate changes.