(a) Available GCM data
The best tool to predict climate variation is a GCM, and here we concentrate on the latest climate model projections for West Africa and surrounding area. In preparation for the IPCC fourth Assessment, multi-model, multi-ensemble and multi-emission-forcing simulations have been made available, and some of these now include daily resolution output, essential for understanding changes in extremes. This is a significant advance on previously available climate data, which in general has been averaged to give only mean monthly diagnostics. The GCMs that we use here are those for which daily temperature and precipitation data is provided by the modelling centres for the IPCC fourth Assessment report (see acknowledgements). The models that we use are fully coupled ocean–atmosphere models from four climate research centres of the Geophysical Fluid Dynamics Laboratory, USA, the Centre for Climate System Research's ‘Model for Interdisciplinary Research on Climate’, Japan, the Meteorological Research Institute, Japan, and the National Centre for Atmospheric Research's ‘Parallel Climate Model’, USA. The model versions, respectively, are GFDL CM2.0 (hereafter, GFDL), MIROC3.2 medres (MIROC), MRI CGCM2.3.2a (MRI) and NCAR PCM1 (PCM). For each model, we have four simulations available driven by different external forcings: one driven by historical forcings appropriate for the period 1971–1989, and three for the years 2081–2099 taken from the end of runs forced by SRES scenario emissions. In increasing order of severity, these scenarios for the twenty-first century are B1, A2 and A1B, as described by Nakićenović et al. (2000). We also have monthly mean model data for 1901–1998 (1902–1998 for MRI). During the twentieth century, all four models are driven by estimates of changes in greenhouse gas concentration, the direct effect of tropospheric sulphate aerosols, volcanic aerosols and solar irradiance. In addition, the MIROC model is driven by changes in stratospheric ozone concentration (SO), the indirect effects of tropospheric sulphate aerosols (IS), black carbon aerosols (BC) and land use (LU), GFDL is driven by BC, SO and LU and PCM is driven by IS and SO. The estimates of future change as given by the SRES scenarios include forcings of greenhouse gases and sulphate aerosols.
The use of four GCMs allows a first estimate of inter-model differences in projections of future changes in daily rainfall and temperature. In particular, we are interested to compare the uncertainty generated by model formulation to the uncertainty generated by future emission scenarios.
(b) Mean rainfall behaviour
provides an illustration of the strong interannual rainfall variability of the region between 3.75 and 21.25°N and between 16.875°W and 35.625°E, based on gridded observations of the twentieth century. The observations are land-based gauge data taken from the Hulme dataset for 1900–1998 interpolated onto a 3.75° longitude by 2.5° latitude grid (Hulme 1992
; New et al. 2000
). The rainfall values are calculated for the growing season in each year (defined as July, August and September), and ‘binned’ into intervals of 0.25
mm. The blue bars are the years 1971–1989 which correspond to a period of repeated droughts; the vertical line differentiates the two years of 1983 and 1984, which suffered particularly severe drought conditions attracting major international action. It is of importance to note that this large area is designed to provide an initial assessment of model capability for the region, and it covers significant regional variation. For particular subsets of the area there is larger interannual variability; for instance the Sahel region in 1983 and 1984 suffered extreme drought, with rainfall far smaller than the mean for that location than might be inferred from .
Figure 1 Mean rainfall (mmday−1) during the growing season, as derived from the Climate Research Unit (CRU; University of East Anglia, Norwich, UK) gridded climatology and for the region of latitude 3.75–21.25°N, and longitude (more ...)
The identical statistics from the four twentieth century GCM runs are presented in . Three distinct inter-model differences are apparent. First, the absolute values vary significantly between the models, with the MRI simulation being particularly dry. Second, the spread of values capturing their estimates of interannual variability is small. However, the third issue, which is of particular interest, is that there is no prediction of the intense and prolonged drought, which began in the early 1970s. This is common to all coupled models we analyse here and suggests that the drought period may not be a consequence of external forcings, but instead is due to natural variability occurring through atmospheric coupling to long time-scale variations in oceanic temperature profiles. We note, however, that the failure of models to reproduce the drought period could also be down to a particular shortcoming common to GCM oceanic descriptions whereby modelled SST change is under-sensitive to alterations in atmospheric CO2
concentration. A useful set of GCM simulations to create would be long control runs capturing oscillations of the coupled climate system when forced by mean atmospheric greenhouse gas concentrations applicable to period 1971–1989. Comparison with the existing control runs appropriate to pre-industrial atmospheric concentrations would allow an assessment of whether anthropogenic emissions up to that period raised the probability of SST patterns occurring as observed during the two decades of drought. The simulations presented here do, however, fit the current opinion that the causes of the drought are a manifestation of internal variability rather than response to greenhouse forcing. Atmosphere-only GCMs forced by known SSTs alone can reproduce a significant part of the decadal drought signal (Giannini et al. 2003
). Such coupled ocean–atmosphere oscillations may be further enhanced over the region by land ecosystem feedbacks, as proposed by Charney (1975)
and illustrated with a coupled model by Zeng et al. (1999)
Figure 2 Transient GCM estimates of precipitation for the region, season and years as the observed data given in (note different vertical scales). None of the simulations capture the observed mean drying for the period in blue (years 1971–1989). (more ...)
shows the change in seasonal rainfall predicted by the four GCMs, between the contemporary period and the end of the twenty-first century. a and , give mean monthly rainfall for 1971–1989. We divide the tropical north African region into two latitudinal subareas that we define as ‘Sahel’ (11.25–16.25°N) and ‘Soudan’ (4.75–11.25°N). c and are mean monthly output for 2081–2099 averaged across the three SRES scenarios. In all panels, the black line represents observations for 1971–1989. We selected this period for comparison as it covers a period of drought.
Figure 3 Mean rainfall by month for the four GCMs. These correspond to an averaging period of 1971–1989 (a,b) and 2081–2099 and averaged across all scenarios (c,d). Two regions are presented (both with the same longitude variation as used in (more ...)
All the models show a distinct seasonal cycle that qualitatively mimics the observed cycle which is due to the seasonal migration of the inter-tropical convergence zone (ITCZ). We see that the ‘wetter’ models from are, generally, wetter for all months. The important result is that there is relatively little change in the monthly model mean predictions between the 1971–1989 and 2081–2099 periods, averaged across these two large areas. However, this may conceal large changes in the spatial patterns and extremes of rainfall.
shows the mean percentage change in growing season precipitation for all models and all emission scenarios for 2081–2099 with respect to 1971–1989. For all scenarios, the GFDL model shows a drying, while the other GCMs show a small wetting. There is less consistency across scenarios, but there is a small drying on average (last column).
Mean percentage increases in growing season precipitation in the tropical northern Africa region for each model and scenario combination in 2081–2099 with respect to the observed 1971–1989 climatology.
In , we compare and contrast year-on-year variability across future scenarios (a) and the four models (b) for July–September precipitation in the tropical northern Africa. The blue histogram in both panels represents observed rainfall for 1971–1989 with which we nudge all model data so that the 1971–1989 means are the same as those observed. We do this here, because we are interested in capturing future changes in precipitation, rather than overall model performance, which we described in .
Figure 4 Mean growing season precipitation, divided into bins of 0.25mmday−1. Both plots are normalized to a total of 19 years. (a) The coloured lines are averages across all four GCMS for the three SRES scenarios for the period 2081–2099. (more ...)
In a, we see that, despite small increases in mean growing season precipitation for three of the models, larger numbers of extreme dry and wet years are expected for all scenarios—especially A1B and A2. Large increases in the number of expected dry years are a concern for agriculture. Inspecting b, however, we see that although all four models show increases in rainfall extremes (solid lines), much of the increase in dry years comes from the GFDL model, which shows large mean drying. Another issue is the use of 1971–1989 as the climatology period. While this does allow the comparison of future climate changes to contemporary climate, the 1971–1989 period was unusually dry for the twentieth century (). If this is the consequence of natural variability, as the models suggest (), and not a forced change in climate, then the nudged anomalies of will predict too many dry years. Nevertheless, the models still predict that there will be more dry years in future with respect to their own contemporary climates.
(c) Short time-scale extremes
, we found that although the four GCMs predict relatively little change in future mean rainfall over tropical northern Africa, they do imply the possibility of significant increases in the number of drought years. In addition, there may be important changes in climate limited to small spatial and temporal scales that do not manifest themselves in monthly mean, area averaged data, but could also damage crop yields. Here, we investigate model predicted changes in gridded daily precipitation and temperature data. These quantities are of direct importance to physiological response, but to date, have rarely been available as diagnostics from large-scale climate models. The IPCC data centre has provided access to new GCM output, where these diagnostics are available.
shows the mean and standard deviation of changes in the number of ‘dry weeks’ in the four GCMs. Rather than choose a real threshold known to cause crop damage, we define a dry week as a 7 day period that falls in the bottom 10% of 7 day periods in the 1971–1989 run of each model. We do this because the models have the tendency to ‘drizzle’ on short time-scales, rather than produce focused storms, and plants experience precipitation via local soil moisture properties that are not well represented. Even so, the change in the percentage of dry weeks below a given threshold will give an indication of how we might expect the frequency of crop damage to change in future.
Figure 5 Model predictions of changes in frequency of low rainfall. For each climate model during the period (1971–1989) and for the growing season defined as July–September, we calculate the average daily precipitation threshold below which 10% (more ...)
a shows future changes in the number of dry weeks during the growing season for tropical northern Africa across all models and scenarios. At almost all points there is an increase in the number of dry weeks, although this pattern shows significant spatial variability. Further, the variance between models is large (see b). In , the overall mean change across all models and all scenarios for growing season rainfall is small. This implies the larger percentage changes in dry weeks corresponds to fewer but more intense storms in a future climate.
shows the number of days during the 92 day growing season when the mean temperature is greater than 33
°C (a temperature above which cereal crops suffer major physiological damage; Peter Craufurd, personal communication). All the model data are nudged so that mean monthly temperatures during 1971–1989 are the same as those observed (New et al. 2000
shows the number of ‘hot days’ occurring in the models during the 1971–1989 period. b
shows increases in the number of hot days for most of the tropical northern Africa region. In particular, shows hot days in the more populated southern part of the region, where there were none during 1971–1989. c
shows large inter-model and inter-scenario differences in predicted changes in number of hot days.
Figure 6 Model predictions of changes in number of ‘hot days’. For each climate model during the period (1971–1989), and for the growing season defined as July–September, we calculate the number of days where the mean temperature (more ...)