On RCM-based projections of change in southern African summer climate



[1] Two regional climate models (RCMs) are used to downscale 10 years of control and 10 years of future (2070–2079) southern African climate, as simulated by the HadAM3 general circulation model forced with the A2 SRES emissions scenario. Changes in early and late summer season total rainfall, rain days and average surface temperature are presented for the projected future climate. The two RCMs indicate broadly consistent changes over the region as a whole. However, time- and location-dependent differences are apparent, especially in the simulated magnitude of change, due to different representations of each model's internal physics and local hydrological cycle.

1. Introduction

[2] Using Regional Climate Models (RCMs) for downscaling projections of climate change from General Circulation Models (GCMs) is an area of active research [Intergovernmental Panel on Climate Change (IPCC), 2001]. Furthermore, the scope of their application is wide, both in terms of the region and aspect of climate that is of interest [e.g., Huntingford et al., 2003; Boo et al., 2004; Pal et al., 2004]. Over North America Chen et al. [2003] compared climate change projections from two different RCMs nested within the same GCM, whereas over Europe the PRUDENCE project has completed an extensive range of simulations [Christensen et al., 2002]. Both studies conclude that many aspects of the RCM simulations are similar, though differences exist due to the RCMs different internal atmospheric physics and dynamics. In this note we similarly compare the results from nesting two RCMs within the boundary conditions from a single GCM over southern Africa – a domain especially vulnerable to climate change, and for which few RCM climate change studies have been undertaken [e.g., Arnell et al., 2003; Hudson and Jones, 2002]. The primary purpose is to indicate where downscaled projections are consistent between RCMs, and where they may be subject to the choice of a particular RCM.

2. RCMs and Forcing Data

[3] The two RCMs used in this study are the 5th generation PSU/NCAR mesoscale model (MM5 [Grell et al., 1994]) and the “Providing REgional Climates for Impacts Studies” model (PRECIS [Jones et al., 2004]) of the Hadley Centre, UK. Both RCMs are nested within 10 years of control and future integrations of HadAM3P, which was forced by SSTs from HadCM3 and the A2 SRES emissions scenario [Jones et al., 2004]. The control run for PRECIS spans the period 1970–1979 whereas the control run of MM5 is 1975–1984. More data is being generated and these control periods reflect the data available at the time of writing. Both future timeslices are for 2070–2079. Here we present the climatological response (mean difference between future and control periods) of the two models, which was not significantly affected by the two different control periods; similar mean differences were obtained when using the common control period 1975–1979. This is perhaps surprising given known southern African multi-decadal variability [e.g., Reason, 2000]. However, Figure 1 demonstrates the mean October-March HadAM3P temperature at 700 hPa, averaged over southern Africa (15–35°E, 35–10°S), for both the 30-year control and future periods. This variable is representative of several aspects of climate variability as it integrates changes due to radiative heating/cooling (through changes in cloud cover) and changes due to latent heating/cooling (through changes in precipitation), as well as advection heat. Although Figure 1 demonstrates there is multidecadal and interannual variability, which means that the choice of 10-year RCM simulation period will affect the results, this variability is small compared to the mean change projected for the late 21st century. Therefore, the future-control anomalies presented here are little affected by the difference in the two simulated control periods.

Figure 1.

HadAM3P October–March 700 hPa mean temperature (K) for the 1961–1990 and 2071–2100 periods. Temperatures are averaged for the region 15–35°E, 35–10°S.

[4] MM5 was configured at 50 km resolution, with 23 vertical levels and lateral boundaries 4–57E and 40–0S, with a buffer zone of 5 grid points. Over southern Africa during the peak summer months this model is known to simulate a lower than observed frequency of rain days, though realistic diurnal cycle of rainfall [Tadross et al., 2005]. The PRECIS model was configured at 0.44° horizontal resolution and nominally spanned 4W–62E and 42S–10N. For comparison, data from both model simulations were regridded to a common domain at 0.5° horizontal resolution.

3. Results

[5] The base performance of the MM5 and PRECIS models over the southern Africa domain has been assessed and shown to be credible by Tadross et al. [2005] and Hudson and Jones [2002]. Here we also show key attributes of each model which demonstrate fundamental differences in their internal physics and which are relevant to discussions concerning the differences in projected anomalies.

[6] Figure 2 demonstrates the bias (with respect to observations from the Climatic Research Unit [New et al., 2000]) in the mean number of rain days and mean rainfall intensity (mm) per rain day (calculated as: total rainfall/number of rain days). Each bias is calculated for the Oct–Mar summer season of the control run for each RCM. It should be noted that the RCM control runs include systematic biases introduced through the GCM forcing fields. Even so, Figure 2 illustrates differences in how each model simulates the southern African climate, with MM5 on average simulating less rain days and more intense rainfall than PRECIS. Further differences were noted in the convective fraction (cf) of total rainfall (averaged 15–35°E, 30–10°S, not shown) between MM5 (cf: 0.83) and PRECIS (cf: 0.92) for the same period. Given that both MM5 and HadRM3H (from which PRECIS originates) simulate close to observed seasonal rainfall totals [Tadross et al., 2005; Hudson and Jones, 2002], it is apparent that MM5 simulates a more intense hydrological cycle and greater fraction of large-scale rainfall than PRECIS.

Figure 2.

October–March mean bias (with respect to CRU) in simulated number of rain days a) MM5, b) PRECIS and mean rainfall intensity (mm) per rain day c) MM5, d) PRECIS for the control period.

[7] Both models simulate a rain day minimum located over southern Zimbabwe/Mozambique and northern South Africa (coincident with the minimum in cf, not shown). Later (Figure 4) it is demonstrated that the highest increases in rain days occur during late summer over this same region, suggesting that a model's control climate (and biases therein) may play a role in determining the magnitude of any anomalies that a model may simulate in the future climate [e.g., Arnell et al., 2003]. For example the number of rain days has an upper limit, which is already reached in the control climate of the PRECIS model over much of southern Africa during the peak summer months (not shown).

[8] The projected change (future – control) from both RCMs differed between the early summer season (Oct–Dec, OND) and late summer season (Jan–Mar, JFM). The total seasonal rainfall anomaly for both seasons is shown in Figure 3 with the projected differences significant at the 90% level (according to a students t-test) shaded. During OND both models predict drying over the tropical western side of the continent with MM5 indicating that the drying extends further south and PRECIS further east. During JFM there is an indication of drying in the west towards the tropics, with increases in total rainfall towards the east. These increases cover a larger statistically significant area in the PRECIS data but are of greater magnitude in the MM5 data (likely due to a hydrological cycle of greater intensity in MM5). Examination of the monthly data (not shown) indicated that these increases in rainfall in the east were mostly during January and February in both models.

Figure 3.

Simulated change (future-control) in seasonal rainfall (mm) during OND a) MM5, b) PRECIS and JFM c) MM5, d) PRECIS. Positive (negative) changes significant at the 90% confidence level are shaded light (medium) grey.

[9] Figure 4 shows the complementary change in the number of rain days for the two seasons, which in general reflects the pattern of change in total rainfall. However, the projected changes in rain days are statistically significant over a wider area during OND as well as being significant over the western side of the continent during JFM. This reflects increases in the frequency of positive 500 hPa geopotential height anomalies towards the west in the HadAM3P forcing data (as identified in observed and GCM projected future circulation data, see Hewitson et al. [2005]), indicating that changes in the GCM forcing at the lateral boundaries in this region is the dominant forcing of the simulated change in rainfall in both models. This also serves to highlight that consistent responses may be better detected in statistics of daily precipitation, as opposed to changes in the seasonal or monthly means. In some of the results presented here it appears that total rainfall may change little or increase, perhaps due to increases in intensity, which can act in an opposite manner to the reduction in rain days e.g. contrast Figures 3b and 4b over the Zambia/Zimbabwe/Mozambique border. Consequently the results suggest that the suppression of rainfall during mid-summer, due to the increase in frequency of high pressure anomalies in the west, is opposed by intensified convection towards the east, presumably due to increased thermal heating.

Figure 4.

Simulated change (future-control) in total rain days during OND a) MM5, b) PRECIS and JFM c) MM5, d) PRECIS. Positive (negative) changes significant at the 90% confidence level are shaded light (medium) grey.

[10] Figure 5 demonstrates the projected change in average surface temperature (1.5 m for PRECIS data and 2 m for MM5 data). All projected changes are significant at the 95% confidence level or higher. Both RCMs project average temperature changes in excess of 1°C with highest temperature changes in excess of 4°C during OND. During this period decreases in precipitation (Figure 3) are projected over some of the regions experiencing the highest temperature increases, which suggests that the increases in temperature may be partly associated with either a reduction in latent cooling or increase in incident shortwave radiation (due to decreased cloud cover) at the surface. The latter is influenced by convective cloud, which the differences in convective fraction noted earlier suggest may be simulated to varying degrees in the two RCMs. During JFM the highest increases in both models are found over the tropical western regions with the lowest increases towards the southeast. This pattern is approximately the inverse of changes in total rainfall in Figure 3 but is more clearly the inverse of the change in rain days (Figure 4). The spatial pattern of projected increases in diurnal temperature range (not shown) is similar to projected decreases in the number of rain days and vice versa. Assuming that rain day frequency is a useful proxy for days with cloud, this suggests that changes in incident shortwave are important and contribute to different projections of temperature change.

Figure 5.

Simulated change (future-control) in average surface temperature (°C) during OND a) MM5, b) PRECIS and JFM c) MM5, d) PRECIS. All changes are significant at the 95% confidence level or higher.

4. Discussion

[11] The results presented here indicate that, at the broad regional scale, there is reasonable consistency between RCM downscalings from one particular GCM, even though the RCMs may differ significantly in their representation of internal physics (for example, MM5 is a hydrostatic model, and PRECIS is a non-hydrostatic model) as well as in their base climate (as evidenced from the intensity with which they simulate the local hydrological cycle). This suggests that baseline biases in the models are not necessarily an indication that the model cannot capture the synoptically forced climate change signal from the GCM. Given that projected GCM precipitation changes are often inconsistent [IPCC, 2001] the results presented here suggest that some of this uncertainty may be reduced through the RCM downscaling. However, this proposition needs to be tested with other GCMs before any firm conclusions may be drawn.

[12] There are, however, notable differences between the RCM projections of change (especially the magnitude of change) when considered on finer regional scales and over particular periods of interest. Thus, at least for these two models, it would be premature to use the downscaling from a single RCM for geographically specific climate change impacts and adaptation activities. Rather, it reinforces the imperative that multiple RCMs, preferably encompassing a range of hydrological cycles, are considered when dynamically downscaling GCM projections of climate change. It is further apparent that changes in the statistics of rainfall, which often are of greater relevance to impact studies than changes in the mean, may reflect the climate change signal more clearly. For example, a reduction in rain days combined with a compensating increase in intensity and increases in temperature will likely lead to greater runoff, evaporation and drying of soils, which is important for agricultural planning.

[13] Notwithstanding the limitations above, the RCM downscaling does provide regional change indications that are in accord with past changes, and correlate with changes projected by the native GCM precipitation and empirically downscaled results [Hewitson and Crane, 2005]. In that Africa is arguably the region most vulnerable to climate change, and yet perhaps has the least research in this regard, these results provide a positive outlook for the development of research activities to directly meet the needs of the region.


[14] We are grateful for funding through the Assessments of Impacts and Adaptation to Climate Change program project AF07 and the South African Water Research Commission grant of project 1430 which supported this work. Many thanks also to R. Jones, D. Hassell, S. Wilson and R. Taylor for their help and guidance on using the PRECIS model and two anonymous reviewers for their helpful comments.