GCMS
The individual members of the QUMP ensemble are referred to as HadCM3Q0–16, where HadCM3Q0 is the unperturbed member (the parameters values are the same as those used by the standard HadCM3 GCM) and the perturbed members Q1–16 are numbered according to the value of their global climate sensitivity, thus Q1 has the lowest global average temperature response to a given increase in atmospheric \(\hbox {CO}_2\), and Q16 the highest. From here on, these models are referred to simply as Q0–Q16. To downscale a GCM ensemble of this size with an RCM would be highly resource intensive. We therefore employ a method outlined in McSweeney et al. (2012) to sample from the ensemble in order to select a subset which represents a similar range of outcomes as the full ensemble.
RCMs
The regional configuration of the Met Office Hadley Centre Climate model, HadRM3P (Jones et al. 2004), was run for the period from December 1949 to December 2100 for whole of Africa using the domain defined by the Coordinated Regional climate Downscaling Experiment (CORDEX) project (Giorgi et al. 2009). The HadRM3P configuration for these simulations has a resolution of 50 km, with 19 vertical atmospheric levels and includes MOSES 2.2 (Met Office Surface Exchange Scheme version 2.2), a tiled land surface scheme (Essery et al. 2001) with 4 soil levels. The chosen global QUMP ensemble members, which were selected using the methodology outlined in Sect. 2.5, provide the boundary conditions for the RCM simulations. In all of the ensemble members the SRES A1B scenario (Nakicenovic et al. 2000) is used to represent future emissions; this scenario contains no mitigation and represents only one of several possible futures considered in the 4th assessment report of the IPCC (Meehl et al. 2007b).
Observations
Validating models’ simulations against observations can be a challenging exercise in Africa. Data coverage is generally sparse and the observational record often show significant discontinuity in times. While there is no simple solution to this lack of data we tried to address the issue by looking at a number datasets which use different techniques and data sources. The observed datasets used are detailed Table 1.
Table 1 Observational datasets used for validation of regional model simulations for Africa
The African Great Lakes
The African Great Lakes are an important feature of Africa and are crucial in representing the climate of the region. In HadRM3P and MOSES2.2 there is no specific lake model and therefore the model makes certain assumptions when the lakes are set to be inland water or sea points. A limitation of this particular configuration of the regional model is that lakes are assumed to be at sea level, and lake surface temperatures are interpolated from the nearest sea point. This results in a warm bias in the lake surface temperatures, and subsequently excessive evaporation. In order to alleviate the problem in these simulations, two actions are taken; first the Great Lakes are set to land points in the domain orography which means that they are at the correct height above sea level, but are maintained as water by the land-sea mask. Secondly, the lake surface temperatures in the SST ancillary files must be corrected from the values that were interpolated from sea points, using lake surface temperature observations.
We use the (night-time) climatological lake mean temperatures for each month for Lake Nyasa (Malawi), Tanganyiki and Victoria from the ARCLake project v1.1.2 (MacCallum and Merchant 2010, 2011), covering the period 1995–2009. The biases in the ancillaries are calculated from the difference between the observed annual temperature cycle and the model annual temperature cycle over a baseline period (1961–1990) for the unperturbed QUMP run (Q0), rounded to the nearest 0.5 K. These biases are then applied to each model run over their entire time period, which assumes all model runs have a common bias. This process is illustrated in Fig. 1; the black curve shows the annual cycle of observations and the yellow curve shows the annual cycle of the original Q0 ancillary. Once bias corrected to the observations (light-blue), the model ancillary is much closer to the observed ARCLake mean temperatures. As an example, the uncorrected (red) and corrected (dark-blue) annual cycle from another model run (Q2) is also included, illustrating that the correction derived for Q0 also improves the lake surface temperatures in this model run. A key assumption made here is that the bias correction applied will remain relevant into the future, i.e. that the difference between the true lake mean temperatures (as provided by ARCLake) and the temperatures interpolated from the nearest sea point will remain the same in a future climate. However, given that the bias between the model ancillaries and the lake mean temperatures from ARCLake is large, almost 3° in some cases, the application of the bias correction is necessary to ensure that the current and near future climate is represented correctly.
Selection of driving GCM runs
In order to reduce the computational requirements, only a sub-set of the 17-member QUMP ensemble was downscaled from the global models. To identify the most informative selection we adopt the procedure outlined in McSweeney et al. (2012).
First we eliminate the ensemble members that perform poorly in simulating the key features of the current African regional climate.
Once this operation is completed we select, from those remaining, the sub-set that best captures the range of responses in temperature and precipitation simulated by the 17 QUMP ensemble members.
In order to select the most appropriate sample the broad range of climatic regimes that occur across Africa must be considered. For this reason, as well as validating the QUMP ensemble projections against temperature and precipitation data for the whole of Africa, we also present results for nine geographical sub-regions that were chosen to represent the different climatic regimes across Africa. The climatic regions are shown in Fig. 2.
The coordinates that have been used to define the Africa region and the other climatic sub-regions are illustrated in Fig. 2 and given in Table 2.
Table 2 The Western longitude (W), Eastern longitude (E), Northern latitude (N) and Southern latitude (S) of the sub-regions
Validation of the African climate simulations
To validate the performance of the models, we compared the observed and simulated annual cycles of temperature and precipitation ( Fig. 3, and the geographical patterns of precipitation and 850 hpa winds in the simulations to those in observed datasets for the period 1961–1990 (not shown). The annual cycle of temperature for the whole of Africa suggests that the models capture the seasonal cycle of temperature realistically, although the majority slightly over-estimate temperatures between May and September (Fig. 3, top left).
Most models also capture the different seasonal cycles of temperature in the sub-regions although for some there is a greater spread in the simulations (e.g. Southern Africa in October), Model Q16 tends to be consistently the warmest model, and lies apart from the other models, and Q4 the coolest. The temperatures for Central Sahel, Fig. 4 (top left) and East Sahel, Fig. 4 (middle left) are generally under-estimated by most of the models for the period between April and June. In general the ensemble captures the annual cycle of rainfall for many of the regions of Africa shown here (Fig. 3, right column), however again there are differences in spread between ensemble members for different regions.
The simulations capture the main rainy season in the Sahelian regions in JAS, although the rainy season begins 2–3 months too early in most of the models for Central and East Sahel, and the range of magnitudes of wet-season rainfall is large. Rainfall in the western Tropical region arrives in the correct seasons, but is systematically too large, to a varying degree depending on the particular ensemble member. The simulations of precipitation for some of the sub-regions do not compare that well with observations, for example the northern Africa region seasonal cycle is not captured at all (Fig. 3, middle right) and though the two wet seasons observed in Kenya (Fig. 5, bottom right) are simulated by the ensemble, the first [March, April, May, (MAM)] is under-estimated by all of the ensemble members and the second [September, October, November (SON)] is over-estimated by some.
However, modelling the climate of Africa is a challenge, as highlighted in the IPCC 4th assessment (Solomon et al. 2007) , which noted excess rainfall over southern Africa of over 20 % on average in 90 % of the GCMs assessed and a tendency for the Inter-Tropical Convergence zone to be displaced towards to equator. In addition, several of the GCMs had no representation of the West African Monsoon at all (Meehl et al. 2007b). Also, given that the amounts of precipitation that occur in some of these sub-regions is very small it is helpful to compare the geographical patterns of precipitation with observations.
For the seasons June, July, August and September (JJAS) and December, January, February (DJF) the large scale patterns are generally captured by all the ensemble members (Figs. 4, 5) however many over-estimate the magnitude over central southern Africa particularly during DJF. In Fig. 4 the lower sensitivity models (Q1–Q5) tend to match the magnitude of the observed DJF precipitation climatology more closely than the higher sensitivity models (Q15 and Q16). The timings, and geographical location of wet periods and regions, however, are realistic.
Figures 4 and 5 show the precipitation for Africa for the seasons JJAS and DJF respectively. The large scale patterns are generally captured by all the ensemble members, however many over-estimate the magnitude of the precipitation over central southern Africa particularly during DJF. In Fig. 5 the lower sensitivity models (Q1–Q5) tend to match the magnitude of the observed DJF precipitation climatology more closely than the higher sensitivity models (Q15 and Q16). The timings, and geographical location of wet periods and regions, however, are realistic.
The circulation simulated by the model at 850 hPa has been compared with ERA40 (Uppala et al. 2005). As with the precipitation maps the models generally reproduce prevailing circulation patterns, including the direction of the trade winds (both north-east and south-east) (see supplementary information for more details). During JJAS the region of higher wind-speeds over the Horn of Africa (referred to as the ‘Somali Jet’) are also captured. However there is some variation between the ensemble members in the magnitude of the Somali Jet, with Q2, Q3, Q6 and Q7 matching the observations more closely than the other ensemble members. The direction of the DJF trade winds are also captured in most of the ensemble members e.g. Q8, Q9, Q11 and Q13; however the magnitude of the winds over the Sahel and southern Africa are slightly over-estimated in most of the ensemble members. Of all the ensemble members Q3 is the closest match to the observed climatology for the magnitude of DJF wind-speed. The near-surface temperature and sea surface temperature patterns (not shown here) in general compare well with the CRU observations and HadISST datasets respectively. However some of the ensemble members, particularly the higher sensitivity ones (Q9–Q16) do overestimate the temperatures in regions where temperatures are high. The mean sea level pressure patterns (also not shown) for the ensemble members also compare well with observations. Our validation of the 17 models shows that while all the models capture the broad seasonal and geographical pattern in key climate features, the range in magnitudes of features such as seasonal rainfalls, and the realism of those magnitudes, varies from across the models. However, it is not straightforward to identify a subset of models that perform better or worse across the whole region—models that do least well in some regions tend to be the most realistic in another. Our approach, therefore, is to select the sub-set based mainly on representing the spread of future climate outcomes across the regions. When making this decision, however, we take into account the shortcomings of some of the models. For example, where two models project similar characteristics of change in the future, we can use the validation information to choose to include the better performing model. On this basis Q1, Q3, Q4, and Q16 were discarded and not considered further in this analysis because the seasonal cycle of both precipitation and temperature do not compare as well with observations as other ensembles in the largest number of regions. In the following analysis we consider the spread of models with respect to temperature and precipitation changes to make the final selection of ensemble members (see Sect. 2.5.2).
Choosing a selection to represent the spread in QUMP outcomes
The final selection of ensemble members for Africa involves identifying the models which represent the range of the full ensemble in their change in precipitation (\(\varDelta P\)) and temperature (\(\varDelta T\)) for Africa and the key climatic sub-regions (see Table 2) for the A1B scenario between the 1970s and the 2080s. We average over 30 year time periods centred on these decades, in order to partially compensate for natural climate variability. This analysis takes the form of scatter plots which are shown for each region and season in Figs. 6, 7 and 8. There is no particular model that consistently shows the largest change in precipitation for all regions throughout the year e.g. for Kenya in DJF (Fig. 8, top) the largest change in precipitation is seen in Q14 but this model is not always the wettest model for the other seasons for this region; for example, Q14 is close to the ensemble mean for Kenya in JJA (Fig. 8, 3rd row). Q14 is also one of the driest models for some sub-regions, for example, some seasons (MAM, JJA, SON) in the West Sahel (Fig. 6 3rd column). On this basis the extremes of the ensemble distribution are classified in terms of which models consistently have the largest positive or negative change in precipitation across all the sub-regions and seasons. Therefore using this scoring system Q9 represents one of the wettest and Q0 represents one of the driest models in the range of the ensemble (but this does not mean these are the wettest and driest models in all sub-regions and seasons).
Temperature
Although the models are numbered 1–16 according to their global temperature response, regional responses will vary. Temperature response is more consistent across the regions and the seasons than the precipitation response, with, as expected, the higher response models tending to capture the warmer end of the range (Q13, Q14, and Q16 tend to have the largest temperature response across the regions and seasons) while the lower-response models, tend to indicate smaller temperature responses (Q1, Q2, Q3 tend to be coolest). Therefore on the basis that, of the lower response models, Q1 and Q3 do not validate as well as Q2 (and Q0 which also has a low regional temperature response) Q0 and Q2 are selected to represent the colder end of the range. At the warmer end of the range, Q16 has already been discounted on the basis of validation results, thus Q13 and Q14 are selected to represent this part of the range.
Rainfall
There is no particular ensemble member that consistently shows the largest change in precipitation for all regions throughout the year e.g. for Kenya in winter (DJF) the largest change in precipitation is seen in Q14 but this model does not then feature as the wettest model for the other seasons for this region; for example, Q14 is close to the ensemble mean for southern Africa and one of the driest models for western tropical Africa. On this basis the extremes of the ensemble distribution are classified in terms of which models consistently have the largest positive or negative change in precipitation across all the sub-regions and seasons. Therefore using this scoring system Q9 captures the wettest and Q0 captures the driest end of the ensemble range (but this does not mean these are the wettest and driest models in all sub-regions and seasons). On the basis of this analysis we conclude that a sample which reproduces important characteristics of current the African and Kenyan climates and represents the spread in projected outcomes produced by the QUMP ensemble consists of the following models: Q0, Q2, Q9, Q13 and Q14.
Table 3 Full run name of the regional model simulaitons, the name that was used to refer to these simulaitons in the text and the member of the QUMP ensemble member that has been used to drive them