In this study, we evaluate a high-resolution regional climate model simulation over the southern African region. The southern African region has been reported to be one of the most affected regions by climate change in the world. In the 4th Assessment Report (AR4) of the Intergovernmental Panel on Climate Change (IPCC), a general warming accompanied by a decrease of precipitation is predicted for the region. The global climate models used in the AR4 projected a temperature rise for the region of about 3.5 K until the end of the 21st century (compared to the period 1980–1999), a decrease of annual mean precipitation of 4% and especially a strong decrease in winter rainfall by > 20% (Christensen et al., 2007). These severe changes will have significant environmental consequences for the region, which is known to be a biodiversity hotspot (Midgley et al., 2002; Thuiller et al., 2006).
To assess the impacts of climate change on the environmental conditions, detailed climate change information is needed for the region. Due to the amount of computer power required, climate estimates from global climate models are so far only available on a coarse resolution of about 200–300 km and therefore less suitable to be used for process studies on the local or regional scale. Qualified tools to receive the needed detailed information are statistical (von Storch et al., 1993; Hewitson and Crane, 2006) or dynamical (Giorgi, 1990; Laprise, 2008) downscaling methods. Dynamical downscaling is based on the application of regional climate models (RCMs) to derive high-resolution regional climate information on a selected region that covers the area of interest. With the application of RCMs climate, information can be downscaled to a kilometre scale. So far, most of the RCM-downscaling studies were conducted at horizontal grid spacings between 50 and 15 km (Giorgi, 2006). However, in some regions of the world even more detailed simulations are available (Jacob et al., 2008).
Due to the surplus of gained information, RCMs have been applied several times over southern Africa in the past. To assess the diurnal precipitation cycle over the region the MM5 RCM was used (Tadross et al., 2006). The same model was applied to investigate the role of vegetation on the regional climate (MacKellar et al., 2009). The RegCM3 RCM was used to study the rainfall variability over southern Africa (Kgatuke et al., 2008). Nevertheless, even a significant amount of RCM studies already exists over the region; most of them are restricted to very short-time periods of a couple of seasons.
Additionally to the above-mentioned process studies, several RCMs also have been applied over the region to downscale future climate projections (Arnell et al., 2003; Tadross et al., 2005). Furthermore, the variable-resolution global model CCAM was applied for regional climate change scenario simulations over subequatorial Africa (Engelbrecht et al., 2009). Nevertheless, none of these simulations have been conducted on a horizontal resolution higher than about 50 km. For the use in impact models climate information on this scale might not be appropriate.
Within the framework of the BIOTA South project (www.biota-africa.org; Krug et al., 2006), we are applying the Max Planck Institute for Meteorology (MPI-M) regional climate model REMO over the southern African region at a horizontal resolution of about 18 km for transient climate change projection from 1960 to 2100. This will be the first long-term climate projection that has ever been performed over the region at such a high horizontal resolution. The REMO model has already been successfully applied over several other regions of the world (Europe: Jacob et al., 2007; South America: Silvestri et al., 2009; South Asia: Roy et al., 2008; Saeed et al., 2009). Over tropical and northern Africa REMO has been applied to assess the future climate and also to estimate the effects of land use changes on the climate (Paeth and Thamm, 2007; Paeth et al., 2009).
In this study, we evaluate the capability of REMO to reproduce the predominant climate characteristics of the southern African region, as shown in Figure 1, with a focus on South-West Africa (SWA) and the main BIOTA North-South Transect (BNST). For this hindcast simulation we downscale almost 50 years of ERA40 reanalysis and operational analysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF) to 18 km horizontal resolution. So far, no long-term dynamical downscaling study has been conducted at a comparable resolution for the southern African region. The main challenge the model has to face in this region is to simulate its unique rainfall characteristics, with a strong seasonality on the one hand and the change between semiarid climates and desert climates on the other hand. Due to the fact that REMO so far has not been applied over the southern African region, this evaluation of the hindcast simulation will indicate the quality of the model results and is seen as an essential first step before the start of the transient climate change projection. Furthermore, due to its high resolution, the data is also used in the framework of the BIOTA South project as a supplement to the existing climate station data.
The article is organized as follows: In Section 2, a short description of the model and the simulation setup as well as the observation data, used in this study, is given. In Section 3, the results of the REMO simulations for the full domain and especially for SWA and BNST are presented. In Section 4, the remaining model deficits and the added value of the application of a RCM over the southern African region are discussed. Finally, a short summary and conclusion section is presented.
2. Model description, simulation setup and observation data
The simulations are conducted with the three-dimensional hydrostatic limited-area atmospheric model REMO (Jacob and Podzun, 1997; Jacob, 2001) in the version 5.7. The model is based on the ‘Europamodell’, the former numerical weather prediction model of the German Weather Service (Majewski, 1991). In its original version the physical parameterizations of REMO were based on those of the global climate model ECHAM4 (Roeckner et al., 1996). The prognostic variables of REMO are surface pressure, horizontal wind components, temperature, specific humidity and cloud water.
Lateral boundary conditions can be taken from analysis/reanalysis data as well as from global climate models. A relaxation scheme following Davies (1976) is used to adjust the prognostic variables towards the boundary forcing in a zone of eight lateral grid boxes. The parameters representing the vegetation and land surface characteristics in the model are taken from the land surface parameter (LSP2) data set (Hagemann, 2002). Thermal properties of the soil are parameterized according to ECHAM 4 (Roeckner et al., 1996). Soil temperatures are calculated for five discrete soil layers with zero heat flux at the bottom (10 m). The heat diffusion in the soil solely depends on the heat capacity and the heat conductivity of the soil and therefore is independent of the soil water content. Water infiltration into the soil is calculated by the improved Arno scheme (Hagemann and Gates, 2003) that separates rainfall and snowmelt into surface runoff and infiltration. The soil hydrology is represented by a bucket scheme, in which lower boundary is constraint by the mean rooting depth of the grid box.
2.2. Simulation setup
The simulation is conducted at a horizontal resolution of approximately 18 km. The model domain for this simulation is presented in Figure 1. To achieve this high resolution, the double nesting procedure was applied, in which the boundary forcing for the high-resolution simulation is generated by an intermediate resolution simulation. The lateral boundary forcing for the current high-resolution simulation was obtained from a sub-Saharan REMO-simulation (south of 10°N) at a resolution of about 50 km, which was forced with ERA40 reanalysis data (Uppala et al., 2005) and ECMWF operational analyses data. The update frequency of the lateral boundaries of the model was set to 6 h. The simulation period covers the years from 1958 to 2007.
2.3. Observational data
To evaluate the performance of REMO over the region, detailed observational data are required. For South Africa, a gridded daily precipitation data set at a horizontal resolution of 10 km was compiled (hereafter referred to as HCPD; Hewitson and Crane, 2005). Daily temperature data based on about 900 stations are also available for the region (Schulze and Maharaj, 2003). Additionally, the globally available monthly data set of the Climate Research Unit (CRU VS.2.1) for precipitation and temperature was used (New et al., 2002) as well as a gridded global precipitation climatology available at a horizontal resolution of 0.25° from the Global Precipitation Climatology Centre (GPCC; Schneider et al., 2008). The regional circulation climatology is taken from ERA40 reanalysis data (Uppala et al., 2005). In general, the analysis of the model data is performed for the period from 1960 to 2000. However, the South African gridded station data set for precipitation was only available for the period 1960–1999 and the GPCC data set was only available as a climatological mean for the period from 1951 to 2000. Although both data sets deviate from the standard analysis period, we decided to still use them, as their high horizontal resolution implies more regional details than the 0.5° CRU data set.
3. Evaluation of the model results
3.1. South African domain
The southern African region is mainly a semiarid region with rainfall ranging from almost no rain in the arid zone along the west coast to > 1000 mm/year in the humid mountainous region in the south-east (Figure 2b). Rainfall shows a strong seasonality with the majority of the region receiving its rainfall in the summer season (October–March). Exceptions to this are the South-Western Cape (SWC) having its rainfall maximum in the winter season and a part of the South African south coast that receives rainfall throughout the year. Rainfall patterns are influenced by both tropical systems (mainly in the northern, central and south-eastern parts of the domain) as well as frontal systems with prevailing westerly flows (mainly in the SWC region). During the summer season the central region is generally influenced by a weak heat low, which tends to produce convective rainfall systems over the moister eastern parts of the domain. In winter the situation mainly changes to a predominant anticyclonic circulation caused by a large single high pressure cell.
Figure 2 shows that REMO is able to reproduce the spatial rainfall patterns but tends to have too much rainfall towards the more humid regions mostly in the south-east and north-east. Mean wet bias in these parts of the domain range between 40 and 80%. However, over the dryer regions in the western areas REMO is in good agreement with the observations (in the range of ± 30%); only along the coast line a strong overestimation is visible. However, this costal wet bias should not be overstated as it is partly an artefact of the interpolation. One possible explanation for the wet bias over the humid parts of the domain can be found in the simulation of the atmospheric circulation patterns over the region, which could lead to an increased moisture input into the domain. Figure 3 shows the mean sea level pressure (MSLP) and near surface winds of REMO (for the 18 and 50 km simulation) and ERA40 as climatological mean over the 1960–2000 period for the summer and winter seasons. The 50-km simulation, which is used as forcing for the 18-km simulation, is included in the analysis to assure that the large-scale features are preserved in the REMO simulations. Compared to ERA40 data the predominant circulation patterns are captured by the model for both seasons and both resolutions in a satisfactory manner. Nevertheless, REMO tends to simulate a slightly lower MSLP in the summer heat low region as well as a slightly higher MSLP in winter high pressure cell in both cases. Additionally, the centre of the winter high pressure zone is slightly shifted towards the north-east in the 50-km simulation. The overestimation of the heat low and the winter high is slightly larger in the 18-km simulation, showing about 1.5 hPa (∼1 hPa in the case of the 50-km simulation) difference to ERA40 in the winter season and about 1 hPa (∼0.25 hPa) difference in the summer in the respective core pressure regions. With respect to the overestimation of summer precipitation of the 18-km simulation especially the former is important. The overestimation of the summer heat low connected with an enhanced inflow of moist air originating from the warm Indian Ocean could explain the overestimated summer precipitation over the south-eastern parts of the domain. This tendency of REMO to simulate a too high moisture transport into the domain should be considered when applying the model for climate change simulations over the region.
3.2. South-West Africa and BIOTA North-South Transect
The focus of this study lies on the performance of REMO over the SWA region and along the BNST. SWA is characterised by the cold upwelling regions of the Benguela ocean current and the Namib coastal desert system. Rainfall in the region shows a strong north–south gradient. In the rather humid regions in the north, tropical systems during the summertime sometimes carry a substantial amount of rain. The southern tip of SWA is characterised by winterly westerly flows and winter rainfall. The centre is generally rather dry, with arid conditions along the Namib coastal desert system. Especially in the arid regions in the centre, coastal fog and dew is one of the most important moisture sources. Furthermore, the SWA region is affected by a strong variability in rainfall, often connected to anomalies in the sea surface conditions (Reason et al., 2002; Rouault et al., 2003; Muller et al., 2008). Large-scale circulation features are often superimposed by local circulation features, induced by the land–sea contrast and the complex orography, often leading to locally very different climate states.
Figure 4 shows the climatological annual cycle of weighted area mean temperature and precipitation for Namibia and SWC. We use a weighted mean to account for differences in the gridbox area, which is related to the distance to the equator and therefore slightly changes over the domain. The mean temperature cycle is represented very well in REMO. However in terms of absolute values REMO tends to overestimate the 2-m air temperature compared to the observations. The warm bias in the simulation is generally increased in the dryer regions of Namibia, with an annual mean bias of about 1.5 K. The seasonal variations of precipitation with predominantly summer rain conditions in the north and winter rain in SWA are captured very well by the model. In addition, the mean precipitation amount for both regions is almost perfectly represented in REMO. Table I summarizes the simulated mean values and the respective bias of the model for both regions and both variables.
Table I. Summary of simulated annual mean values and respective annual mean bias compared to observations for Namibia and SWC. The range of the bias is due to several observational data sets—refer to Figure 4
Annual mean (K)
Absolute bias (K)
Annual mean (mm/month)
Relative bias (%)
To evaluate REMO along BNST, we calculated a zonal mean over all land points for the BNST region as indicated in Figure 1. For BNST the performance of REMO in comparison to observations is very similar as described above (Figure 5). On one hand, the simulated 2-m air temperatures are generally too warm, with a maximum bias of about 3.5 K around 28°S (Figure 5a). On the other hand, precipitation is simulated for both summer (Figure 5b) and winter (Figure 5c) seasons in good agreement with observations. Only the summer rain in the north is slightly overestimated by the model, resulting in a relative wet bias of about 20% at this part of BNST. This bias can be attributed to the previous mentioned overestimation of the summer heat low. Nevertheless, the seasonal distribution of rainfall is simulated in almost complete agreement to the one observed. Figure 5(d) shows the percentage of rainfall occurring in the DJF season, with almost no discrepancy of the simulated to the observed rainfall distribution. We further used the climate-type classification after Köppen & Trewartha (see de Castro et al., 2007 for details on the allocation of the climate types) as an integrated measure to investigate the performance of REMO along BNST. We find that in the southern part of BNST the observed climate types are well reproduced by the model (Figure 5e). In the northern part, REMO indicates a more humid climate type than is observed. This finding is inline with the wet bias depicted in Figure 5(b).
We further investigated if the climate along BNST changed during the investigated period, by calculating the trends (Figure 5f and g). For the 2-m air temperature, a general warming is indicated in the observations. REMO also simulates the temperature increase along the whole BNST, but especially in the northern part the simulated warming is about 0.7 K less than the observed one. For rainfall, the model and observations deviate in the direction of the trend. Although REMO indicates a moistening along the whole BNST, observations agree with the model in the southern part but indicate a drying in the northern part of BNST.
However, the above-discussed trends in precipitation and temperature are rather small compared to the strong inter-annual variability and cannot act as a proof for a clear trend towards a dryer climate. Applying a student t-test to the data, we found that along the transect the majority of the observed trends over the investigation period are not significant on a 95% confidence interval. Although none of the temperature trends seems to be statistically significant, a general increase in precipitation in the southern part of BNST (where the model and observations agree) is supported by the statistical analysis (Figure 5g). However, the contrary behaviour of REMO and CRU in the north of BNST does not seem to be a reliable feature and could also be caused by a slight difference in the inter-annual variability between model and data.
In Figures 6 and 7, the anomalies for precipitation and temperature for three regions along BNST are displayed. The time series show a noticeable variation in temperature and precipitation. Especially the extreme wet years in the mid-1970s are pronounced at all three locations. Moreover, these extreme wet years seem to be a turning point in the southern part of BNST, with rather dry years before and wet condition afterwards. Observed annual mean temperatures for those wet years are generally lower than the long-term average. This colder period is especially prominent in CRU data set for the north and central part of BNST. REMO is generally able to simulate the strong variability in the rainfall. The highest correlation between REMO and observations is in the South, having an anomaly correlation coefficient of 0.65 for the CRU data set and 0.73 for the stations (HCPD data set). In the north, the anomalies are less well captured. Focusing on the last third of the anomaly time series of the northern part of BNST, REMO shows 8 of 17 years wetter than average conditions, whereas the CRU data for the same period consists of only 4 wet years. This discrepancy can lead to the previous indicated contrary trend behaviour of REMO and CRU in that part of BNST.
The extreme wet years in the mid-1970s, which are connected to anomalies of the Indian Ocean sea surface temperatures and ENSO (Washington and Preston, 2006) as well as the switch in rainfall in the southern parts of the transect is well captured by the model. In the case of temperature, REMO simulates a pronounced inter-annual variability for all three cases as it is also observed. The anomaly correlation coefficients for temperature range between 0.43 for the northern and central parts and 0.62 in the south when compared to the CRU data set and are slightly higher for the stations. Furthermore, the link between the extreme wet years and a lower than average temperature is very well represented in the northern part. For the central and southern part, where the colder period is less pronounced, the model, however, simulates warmer than average years.
Although there is an agreement between the simulated and observed anomalies as described above, the consensus of the time series could partly also be random. To test the significance of the simulated inter-annual variability, we compared for each region the rainfall and temperature anomaly produced by REMO to a set of 10 000 randomly compiled time series, each following a normal distribution. The random time series were generated with FORTRAN by using the inbuilt random number generator. Due to the fact that the output of the random number generator in FORTRAN is equally distributed, a transfer function based on the central limit theorem was applied to receive normally distributed random values. To assure the same range of variability in the resulting normal distribution, we kept the standard deviation of the equally and normally distributed time series in the same range. Each of the resulting random time series was normalized to its maximum anomaly. For each timestep we then calculated the difference between the normalized observed (CRU) and the randomly compiled anomaly. To not only consider the differences in the anomalies, but also the direction (positive vs negative anomalies), we introduced a further constraint for the case that the direction of the anomalies did not agree. For these cases the estimated anomaly difference was doubled. Finally, the deviation of each random time series was summarized and the resulting cumulative deviations were scaled according to the maximum cumulative deviation. The same procedure already described for the 10 000 randomly compiled time series was applied to the REMO time series. In Figure 8, the scaled distribution of the cumulative deviations between the anomalies and the random time series as well as the cumulative deviation of REMO are depicted. For rainfall the difference between REMO and the observations is significantly smaller than the one of the random time series. This finding indicates that the inter-annual rainfall variability simulated by REMO is a consequence of the fact that the regional rainfall dynamics are well captured by the model and that the match between REMO and observations is not based on a coincidence. For temperature, the anomalies in the southern and central part are significantly better captured by the model, compared to the random time series. In the northern part, the temperature anomalies in REMO are not closer to the observed anomalies than the random anomalies. This result is inline with the general finding that there are still some deficits in REMO in simulating the 2-m air temperature. As the observed anomalies do not directly follow a normal distribution, we repeated the experiment with a set of 10 000 time series each equally distributed. In the case of equally distributed time series, REMO results are even better (not shown). Especially in the northern part, REMO temperature anomalies then scale very close to the 5th percentile of the random experiments.
In the previous section, we have compared the simulation results of REMO to a set of several observations. We found that REMO adequately simulates the predominant patterns of precipitation. Especially in the focus region of this study, in SWA and along BNST, the REMO results are very close to the observations; particularly the simulation of the mean rainfall conditions is almost in perfect agreement. One of the deficits of the model seems to be the warm bias in the temperature. Although the seasonal temperature cycle is well reproduced by the model, an absolute warm bias compared to observations is persistent. An analysis of the spatial patterns revealed that there is a linkage between warm bias and the model's soil moisture content. To identify the link between the soil moisture and the temperature bias, we analysed a monthly time series covering the period 1960–2000 of gridded REMO soil moisture and temperature bias (with respect to CRU data). We divided the soil moisture data into several bins and identified for each gridbox and month the corresponding warm bias. Later, for each soil moisture bin, the mean warm bias and the standard deviation was calculated and plotted. The result of this analysis is generally that the model seems to strongly overestimate the temperature in regions with lower soil moisture, than in wetter regions (Figure 9a). Based on these findings, we speculate that the resulting warm bias might be caused by two factors. First, the warming could be a consequence of the parameterization of the heat transfer in the soil. In the current model version, soil wetness is not affecting the thermal properties of the soil, so that they solely depend on the soil types. This could theoretically lead to a too large heat transfer into the soil, which would result in too warm surface temperatures and therefore also influence the 2-m air temperature.
Second, the warm bias of the REMO results over rather dry regions could also be a consequence of too low soil water content in REMO, which would have an effect on the latent heat transport. A correlation analysis applying the above-described method to a time series of REMO evaporation data revealed a significant (at the 99th confidence level) linkage between the observed warm bias and the evaporation fluxes, with a correlation coefficient of the respective bin means of 0.96 (Figure 9b). Generally, it is obvious that the warm bias in the REMO simulation is strongest in regions with low soil water availability and therefore related to this with a rather low evaporation rate. These findings support the theory of the availability of soil moisture as a control element for the energy budget. Unfortunately, neither soil moisture observations nor evaporation measurements are available for the region that are comparable to REMO output. Therefore, it only can be speculated that the available soil moisture is too low in REMO. Nevertheless, some support for the previous argumentation can be gained by comparing the REMO cloud cover to the CRU cloud cover data set, which finds the simulated cloud cover in REMO significantly lower than the one of the CRU data set (Figure 9d). Even if we do not suppose that local water recycling is the only reason for cloud formation in the region, we clearly find a linkage between evaporation rates and cloud cover in the REMO results that encourages the argumentation (Figure 9c). Furthermore, we conclude that a too low cloud cover itself could lead to a warm bias in the simulation, as the incoming radiation and therefore the potential available energy is increased. The close connection in the model between soil moisture, evaporation and cloud cover on the one hand and a bias in the 2-m air temperature on the other can also be seen in the above-discussed temperature anomalies for the three regions. Although in the north, the temperature decrease connected to the extreme wet years is captured in REMO, the model fails to simulate this link for the central and especially for the southern region (Figure 7). When we analyse the anomaly of simulated soil moisture, evaporation and cloud cover, we clearly find that in the north (cold period captured by REMO) these parameters are maximum for the wet period, whereas for the south, none of these three parameters are significantly higher than in other years (Figure 10). To improve the simulation of the 2-m air temperature, we are currently working on the parameterization of both the thermal as well as the hydrological properties of the REMO soil. Some further simulations with a changed parameterization of the soil properties will be done in the near future.
4.1. Added value by the application of a RCM
Generally, the purpose of applying a RCM is to improve the representation of the regional to local climate compared to the more coarsely resolved global climate models. The enhanced resolution of a RCM with its better representation of the topography and vegetation patterns allows for a better description of mesoscale atmospheric dynamics, especially along mountains or coastal regions (Laprise, 2008). Still, it sometimes remains difficult to proof the added value of a RCM. In our case, we can clearly identify an added value, which we gain by the downscaling of globally available reanalysis data. As already mentioned in Section 3, REMO reproduces the observed rainfall along BNST in both seasons very well with only a slight overestimation of the summer rainfall in the northern parts (Figure 5b). One of the challenges along BNST is the correct simulation of the seasonal distributed rainfall, with winter rainfall occurring around SWC. When we focus on SWC, the REMO simulation clearly shows an added value compared to the ERA40 forcing data. While in the ERA40 reanalysis data rainfall is underestimated by about 60–70%, REMO rainfall is well in the rainfall range defined by the observations (Figure 11). This clearly identifies the benefit of using a regional climate model especially in regions with complex topography and climate characteristics.
5. Summary and conclusions
In this study, we have evaluated the performance of the regional climate model REMO over the southern African region with a focus on SWA and the main BNST. We find that the model ably reproduces the observed patterns of temperature and precipitation. Especially, the precipitation characteristics in SWA and along BNST are captured in a very satisfactory manner by the model. Both the seasonal rainfall distribution and the inter-annual rainfall variability are represented by the REMO results. Furthermore, compared to the forcing ERA40 reanalysis data, a distinct added value in the form of a significantly better representation of the seasonal rainfall amounts around the southern African cape region could be achieved. This finding highlights the need for high-resolution regional climate simulations, especially with regard to the application of the model results as input for impact models in follow-up studies. On the basis of the promising results of this study, we conclude that REMO is capable to reproduce the major climate characteristics of the southern African region and therefore can be applied to generate the first transient long-term (1960–2100) high-resolution climate change projection over the region.
We thank our colleagues from the Terrestrial Hydrology Group as well as from the REMO Group for continuous support and discussions on this topic. A special thanks goes to Jan O. Haerter for reviewing the manuscript. Furthermore, we thank the Federal Ministry for Education and Research (BMBF, project number 01LC0624B) for funding the work.