Climate change scenarios from a regional climate model: Estimating change in runoff in southern Africa



[1] This paper describes an analysis of different ways of constructing climate change scenarios using output from three climate models. It focuses on using the HadRM3H regional climate model applied across southern Africa and a macroscale runoff model operating at a scale of 0.5 × 0.5° to simulate river runoff. HadRM3H has a spatial resolution of 0.44 × 0.44° and is driven by boundary conditions from HadAM3H, a global atmosphere general circulation model with a spatial resolution of 1.875 × 1.25°. This, in turn, used sea-surface boundary conditions from HadCM3, a coupled global ocean-atmosphere general circulation model that operates at a spatial resolution of 3.75 × 2.5°. Sixteen climate scenarios were constructed from the three models, representing different combinations of model scale, whether the climate model simulations were used directly or changes were applied to an observed baseline, and whether observed or simulated variations from year-to-year were used. The different ways of deriving climate scenarios from a single initial climate model experiment result in a range in change in average annual runoff at a location of at least 10%, and often more than 20%. There is a clear difference in the large-scale spatial pattern of change in runoff from HadCM3 to HadRM3H. Many of the climate features in HadRM3H are already present in HadAM3H simulations, as would be expected from the experimental design. This suggests that for studies over a large geographic domain, an intermediate-resolution global climate model can produce useful scenarios for impact assessments. HadRM3H overestimates rainfall across much of southern Africa and so results in too much runoff: This leads to smaller estimates of future change in runoff than arise when changes in climate are applied to an observed climate baseline. It is concluded that under these circumstances it is preferable to apply modeled changes in climate to observed data to construct climate scenarios rather than derive these directly from the regional climate model simulations. Incorporating increases in interannual variability as simulated by HadRM3H leads to little change in simulated annual mean runoff. However, it has a larger impact on the frequency distributions of runoff, with extreme flows predicted to increase more than mean flows and even to increase in areas where the mean flow decreases. This demonstrates the importance of considering not only changes in mean climate but also climate variability.

1. Introduction: Climate Scenarios at the Local Scale

[2] Scenarios for climate change impact assessments are currently usually constructed from the output of climate models that simulate climate over the entire globe. These global models typically operate at a spatial resolution of several hundreds of kilometers, which is acceptable for the simulation of climate over the global domain but is too coarse for the output to be used directly in impacts models. Local and regional climate may not be well simulated, and the low spatial resolution of the climate model output means that it is not possible to construct realistic time series at the scales required by impacts models: Weather averaged over tens of thousands of square kilometers has different temporal characteristics to weather at a point or over a catchment.

[3] Climate change scenarios constructed from simulation experiments made with global climate models have therefore largely defined changes in mean monthly climate, which are then applied to observed local- or catchment-scale time series of climate data. The simplest of these scenarios assume that the change in local mean climate is equal to the change in the coarse-scale mean climate, and virtually all assume that the future pattern of variation in climate from year to year is the same as that of the present. However, in many cases the impacts of climate change are likely to be felt through changes in the year-to-year variability in climate as well as through changes in mean climate.

[4] There have therefore been many studies that have attempted to develop methods to “downscale” from the global climate model scale to a finer spatial scale (see Mearns et al. [2001b] for a review). In the most general terms, there are four types of methods:

[5] 1. The first method is simple interpolation of the climate model data to a finer spatial resolution. In this approach, changes in climate at a fine spatial scale (for example 0.5 × 0.5°) are interpolated from the climate model resolution using some form of interpolation procedure. This approach was used, for example, to create the UKCIP98 scenarios in the United Kingdom [Hulme and Jenkins, 1998]. The most sophisticated variants of this approach factor in topography or other controlling variables such as coastline when interpolating climate model output. The approach is simple, but it is difficult to apply to anything other than mean climate.

[6] 2. The second method is statistical downscaling using empirical relationships between coarse-scale and local climate. This approach develops empirical relationships between the two scales using observed climate data, and applies these relationships to simulated coarse-scale climate data: It assumes that the relationships remain constant as climate changes. There are many variants of this approach, using different characterizations of coarse-scale climate, and some can be used to construct time series at the local scale (refer to Giorgi et al. [2001] for details). However, in addition to the necessary assumption that the relationships between scales remain constant, it is difficult in practice to apply statistical downscaling over a very large geographic domain.

[7] 3. The third method is use of a regional climate model. This approach uses a regional climate model (RCM) nested within the coarse-scale global model to simulate climate over a region at a finer spatial resolution (typically of the order of 0.5 × 0.5°). This method was, for example, used to create the UKCIP02 climate change scenarios for the United Kingdom [Hulme et al., 2002]. The advantages of this approach are that the downscaling is based on physical principles rather than empirical relationships, that finer-scale scenarios can be constructed over a large spatial domain, that weather time series are simulated at scales much closer to those required by impacts models, and that credible estimates of change in interannual variability can be produced at the local scale. However, a regional climate model will not be able to improve on large-scale errors in the coarse-scale forcing and thus errors in the latter will be transmitted to the RCM (a problem also for statistical techniques). In comparison with the statistical techniques, running an RCM is computationally intensive.

[8] 4. The fourth method is use of a high- or variable resolution global atmospheric general circulation model (AGCM) in a “time-slice” experiment. This approach runs the AGCM for a defined time-slice, with just the sea surface temperature determined from a coarse-scale fully coupled atmosphere-ocean model (AOGCM) [Mearns et al., 2001b]. Variable resolution models are under development and have not yet been used to create scenarios. High-resolution time-slice AGCM simulations have, however, been run [see Giorgi et al., 2001], but not yet used to create scenarios. The use of a global model has the advantage that the remote effects of increasing resolution are captured and, in the case of uniform resolution models, of producing global higher-resolution data. A theoretical disadvantage is that the response in such a model may be inconsistent with the sea-surface boundary conditions. Practically, this approach also has the disadvantage that substantial retuning of the model for higher resolution will probably be required (e.g., J. M. Murphy et al., manuscript in preparation, 2003) and that it is computationally much more expensive than RCMs or coarse resolution AOGCMs.

[9] The aim of this paper is to implement and compare several ways of constructing climate scenarios for use in impacts models, focusing on scenarios for changes in mean climate and the interannual variability of climate. (Mearns et al. [2001b] distinguish between climate change scenarios, which define a change in climate, and climate scenarios that characterize a climate in absolute terms.) The use of a regional climate model (HadRM3H from the Hadley Centre) is the primary method of downscaling employed in this study (method 3 above). However, as a starting point for comparisons, methods 1 and 4 above are applied to the Hadley Centre's coupled atmosphere-ocean GCM, HadCM3, and the high-resolution atmosphere-only GCM, HadAM3H respectively. Whilst there have been several studies into the use of regional climate models to identify possible changes in local climate (refer to Appendix 10.3 of Giorgi et al. [2001]) and a few studies comparing downscaled climates [e.g., Murphy, 2000; Mearns et al., 1999a], there have actually been very few studies which have used regional climate models to create scenarios to drive impacts models [Mearns et al., 1999b, 2001a; Leung and Wigmosta, 1999; Bergstrom et al., 2001; Venalainen et al., 2001]. The current study uses output from the Hadley Centre climate models to drive a macroscale runoff model and investigate potential changes in streamflow by the 2080s across the whole of southern Africa.

2. Scenarios From Regional Climate Models: An Overview

[10] In principle there are two main ways in which regional climate models can be used to drive an impacts model.

[11] First, it is possible to feed the regional model inputs directly into the impacts model, either for one regional model grid cell or across the entire model domain. This may involve adding the impacts model as a module in the regional climate model. This approach means that the impacts model can exploit fully the fine temporal resolution of the regional climate model, and account for changes not only in mean climate but also interannual and intra-annual variability. However, if the regional model simulates local weather or climate inaccurately the simulated impacts will be inaccurate. Also, the approach means either that the regional climate model must be run whenever a new impacts assessment is undertaken, or that large volumes of high-resolution output data must be stored from the regional model to be used off-line (e.g., 30 years of daily data from just one variable at one level in the atmosphere across the HadRM3H southern African domain amounts to half a Gigabyte). This may be impractical.

[12] A more practical approach uses summary output from the regional climate model to construct climate change scenarios, which may be applied with observed climate data in the same way as scenarios derived from global climate models are currently applied. The major advantage of scenarios constructed from regional models over those based on global models lies in the dimensions of climate that can be represented.

[13] This study explores a number of ways of creating practical climate scenarios from regional model output, focusing on this second approach, which is likely to be most widely used in practice. The study concentrates on scenarios for changes in mean climate or year-to-year variability.

3. Models and Baseline Data

3.1. Baseline Climate Data

[14] Observed climate data were taken from the Climatic Research Unit global 0.5 × 0.5° 1961 to 1990 climate archive [New et al., 1999]. Potential evaporation was calculated using the Penman-Monteith formula with monthly mean temperature, humidity, and wind speed. Baseline net radiation was calculated from observed monthly mean cloud cover combined with daily estimates of maximum day length.

3.2. Climate Models

[15] The regional climate model, HadRM3H, operates at a spatial resolution of 0.44 × 0.44° (approximately 2500 km2), and in this application has been applied over a domain extending from about 45°S to the equator and from 5°E to 55°E (Hudson and Jones [2002a]; Figure 1 shows the grid over the land). The model has 19 vertical layers, and runs at a time step of 5 minutes. Its boundary conditions are derived in a two-stage process from HadCM3, a course resolution AOGCM, and HadAM3H, a higher-resolution AGCM. HadCM3, with a spatial resolution of 3.75 × 2.5° (approximately 75,000 km2 over southern Africa), 19 vertical levels in the atmosphere and a time step of 30 minutes, has been used to simulate a transient climate change over the period 1850 to 2100, using historical greenhouse gas forcings to 1990 and emissions scenarios thereafter [Johns et al., 2001]. The emission scenario used in the present study is the IPCC A2 SRES scenario [Intergovernmental Panel on Climate Change (IPCC), 2000]. HadAM3H is used to obtain an improved regional-level simulation over specific time periods of interest identified from the coupled model integration. It operates at a spatial resolution of 1.88 × 1.24° (approximately 23,000 km2 over southern Africa), has 19 vertical levels, a time step of 15 minutes and has been run for two 30-year slices within the 250 year HadCM3 simulation, namely years 1961–1990 and 2071–2100 [Hudson and Jones, 2002b]. The present-day simulation with HadAM3H uses observed sea surface temperatures. The climate change forcing for the experiment is derived from the A2 emissions scenario and a lower boundary condition of the changes in sea-surface conditions derived from HadCM3. These are then also used in parallel HadRM3H experiments that take lateral boundary conditions from HadAM3H. Figure 1 shows the respective model grids over southern Africa.

Figure 1.

Spatial scales of HadCM3, HadAM3H, HadRM3H and the macroscale runoff model.

[16] HadAM3H is an improved version of the atmospheric component of HadCM3. These improvements and the use of higher resolution were required to improve the baseline climate in HadCM3 for driving regional climate models [Hudson and Jones, 2002b]. The formulation of HadRM3H is identical to that of HadAM3H except for a few resolution-dependent aspects (R. G. Jones et al., manuscript in preparation, 2003). A more complete description of the RCM is provided by Hudson and Jones [2002a]. Climate scenarios from all three models are constructed and applied in this study: These fall into the first, third and fourth categories of downscaling summarized in the introduction.

[17] Figure 2 shows the simulated baseline (1961–1990) average annual temperature and precipitation across southern Africa from HadCM3, HadAM3H and HadRM3H, at their original spatial resolutions, together with the observed baseline temperature and precipitation. The broad patterns of climate are similar in the three simulations, although the spatial resolution increases from HadCM3 to HadRM3H. In each case, rainfall is too high across much of southern Africa south of 10°S, because the models produce too much convergence over the subcontinent in summer, the peak rainfall season [Hudson and Jones, 2002a, 2002b]. In parts of the region, including the east (primarily Mozambique) and Madagascar, rainfall is underestimated because the ITCZ is slightly too far north. The overestimation of rainfall across much of the region, together with an overestimation of cloud cover, means that simulated temperatures are too cool across large areas.

Figure 2.

Observed and simulated baseline average annual temperature and precipitation: HadCM3, HadAM3H and HadRM3H.

[18] Figure 3 shows the observed standard deviation and coefficient of variation (standard deviation divided by the mean: CV) of annual precipitation, together with standard deviation and CV as simulated by HadRM3H. HadRM3H produces standard deviations of a similar magnitude to observed, although it tends to overestimate variability over southern regions and underestimate over northern regions. The simulated relative variability (CV), however, is generally lower than the observed (except over southern Mozambique), because the model overestimates mean rainfall.

Figure 3.

Year-to-year variability in annual rainfall: standard deviation and coefficient of variation, 1961–1990, observed and simulated.

[19] Table 1 summarizes the climate variables extracted from the climate model output. The number of simulated rain-days was not used as they were very significantly overestimated across much of the region.

Table 1. Climate Variables and Climate Changes Used
Variable (Mean Monthly Value)Percentage or Absolute Change?
PrecipitationAbsolute and percentage (see Table 3 for full details)
Wind speedPercentage
Vapor pressureAbsolute (calculated from relative humidity and temperature)
Net radiationAbsolute (calculated from simulated net short-wave and net long-wave radiation)

3.3. Runoff Simulation Model

[20] Runoff was simulated at a spatial resolution of 0.5 × 0.5°, using a development of the macroscale runoff model described by Arnell [1999]. This model treats each 0.5 × 0.5° cell as a separate catchment and calculates the evolution of the components of the water balance at a daily time-step. Actual evaporation is a function of potential evaporation and soil moisture content: Soil moisture is replenished when precipitation exceeds actual evaporation and drainage from the soil, and is depleted by evaporation and drainage. The soil moisture storage capacity varies statistically across the cell/catchment, with the mean dependent on soil texture and vegetation rooting depth. “Quick response” runoff at any time occurs from the parts of the cell/catchment that have saturated soils at that time. “Slow response” runoff is a function of the amount held in deep soil and groundwater storages, which are filled by drainage from the upper soil layer. River runoff is not routed from one cell to another. The rate of potential evaporation varies with catchment vegetation, and vegetation also intercepts precipitation: The intercepted precipitation is assumed to evaporate. In the current application of the model there are 13 vegetation classes, of which 4 occur in the study region.

[21] The model operates with input time series (in this study 30 years long) of monthly precipitation, temperature, vapor pressure, net radiation and wind speed, and uses a stochastic procedure to disaggregate monthly rainfall to the daily scale with knowledge of the number of rain-days each month (the observed number of rain-days was used in all calculations). The disaggregated rainfall is rescaled to match the original monthly total. The model produces daily runoff, but monthly totals only are output.

[22] The runoff model is not calibrated on real data, and as a general rule tends to overestimate substantially streamflow in large rivers in dry areas, including much of southern Africa. This partly reflects model form, and particularly the assumption that rain falls on the entire 0.5 × 0.5° catchment, but also arises because the model (1) does not account for evaporation from either large or small wetlands, which are prevalent in some semi-arid areas, (2) assumes that all generated runoff reaches the river channel (which may not be the case in flat areas with disorganized drainage) and (3) ignores transmission loss along the river channel due to infiltration and subsequent evaporation: This is a characteristic feature of semi-arid and arid river channels. However, whilst the model does not simulate flow in large rivers well, it is likely to give better simulations of the amount of runoff generated within a catchment and thus potentially available for use.

[23] The runoff model calculates actual evaporation from potential evaporation and soil moisture deficit, rather than use actual evaporation data from the climate model, for a number of reasons. Climate model-derived actual evaporation depends not only on the drivers of potential evaporation, but also on the climate model representation of the effect of soil moisture deficits on evaporation and the simulated temporal and spatial pattern of rainfall. Actual evaporation can vary considerably through a month, and mean monthly actual evaporation cannot realistically be disaggregated down to the daily time step necessary for the simulation of the water balance: The drivers of evaporation, however, can be disaggregated. This means that the runoff model does not necessarily produce the same simulated runoff or evaporation as the climate model used to estimate input data.

[24] Three runoff indicators were considered, as summarized in Table 2.

Table 2. Streamflow Indicators
Average annual runoffMean runoff (in mm year−1) averaged over the 30 year simulation period
Coefficient of variation of annual runoff (CV)Standard deviation in annual runoff divided by the mean
10-year return period maximum monthly runoff (M10max)Estimated by fitting a GEV distribution to the largest monthly runoff total in each year: a measure of high flows

3.4. Reconciling Spatial Scales

[25] HadRM3H, HadAM3H and HadCM3 operate at different spatial scales, all of which are different to the 0.5 × 0.5° resolution used by the macroscale runoff model (Figure 1). All the spatial resolutions were mapped onto the 0.5 × 0.5° scale, because this is the resolution of the baseline climate data and catchment characteristic data used by the runoff model.

[26] The HadCM3 simulated climate data were “downscaled” to the 0.5 × 0.5° scale in two ways. First, it was assumed that all the 0.5 × 0.5° cells within each HadCM3 cell had the same climate, with no interpolation. Second, a bilinear interpolation procedure was applied to interpolate the HadCM3 simulated climate down to the 0.5 × 0.5° resolution. The HadRM3H and HadAM3H climate data were downscaled to the 0.5 × 0.5° scale by calculating for each 0.5 × 0.5o cell the area-weighted average of all the HadRM3H or HadAM3H cells overlapping with that cell.

3.5. Baseline Average Annual Runoff

[27] Figure 4 shows the modeled average annual runoff across southern Africa, using the observed baseline mean monthly climate data and the baseline mean monthly climate as simulated by HadAM3H and HadRM3H. The simulations based on HadAM3H and HadRM3H both overestimate runoff across central southern parts of the region, and also produce substantially greater estimates of runoff in the humid tropical areas, reflecting the overestimation of rainfall (Figure 2).

Figure 4.

Modeled baseline average annual runoff, using observed climate data and HadAM3H and HadRM3H simulations.

4. Climate Change Scenarios

4.1. Simulated Changes in Climate

[28] Figure 5 shows the change in average annual rainfall and temperature by the 2080s as simulated by HadCM3, HadAM3H and HadRM3H, at the original model resolutions and, for HadCM3, based on interpolated rainfall and temperature.

Figure 5.

Simulated change in average annual rainfall and temperature by the 2080s: HadCM3, HadAM3H and HadRM3H, and change in simulated CV of annual rainfall for HadRM3H.

[29] There are clear differences in the pattern of change in climate from HadCM3 through HadAM3H to HadRM3H, with the greatest apparent similarity in pattern between HadAM3H and HadRM3H (because HadRM3H is of course driven by HadAM3H). HadRM3H, however, simulates a smaller decrease in annual rainfall across Zimbabwe than HadAM3H and, unlike HadAM3H, an increase in southern Mozambique and western Madagascar. Also, the area of greatest relative increase in precipitation with HadRM3H is further south, through the eastern Congo basin and Tanzania, than with HadAM3H. The patterns of change in temperature and rainfall produced by both HadAM3H and HadRM3H are different to those produced by simple interpolation of the HadCM3-simulated climate, suggesting that the use of finer-resolution models with improved formulations adds regional detail which is not present in the coarser-scale simulations.

[30] HadAM3H produces slightly lower increases in temperature across southern Africa than HadCM3, and HadRM3H slightly lower still. The increase in potential evaporation is greatest with HadCM3 and lowest with HadRM3H, due to differences in the increase in temperature, a smaller increase in net radiation with HadAM3H and HadRM3H than HadCM3, and smaller increases in wind speed under HadAM3H and HadRM3H. The radiation changes could partially be due to the changes in cloud representations between HadCM3 and HadAM/RM3H. Changes in vapor pressure were similar between the three models.

[31] Climate change increases the CV of annual rainfall across much of the region (by over 40% across large areas (Figure 5)), but there are reductions in eastern Africa and central South Africa.

4.2. Construction of Climate Change Scenarios

[32] Table 3 summarizes the climate change scenarios derived from HadCM3 (original resolution and interpolated), HadAM3H and HadRM3H compared in this study. The scenarios vary in their treatment of both mean climate (apply changes to the baseline or use climate model output directly) and year-to-year variability in rainfall (assume present year-to-year variability throughout, or use simulated variability). Interannual variability in rainfall is indexed by the time series of percentage departures of monthly rainfall from the long-term monthly mean. An alternative measure, not considered in this study, would characterize interannual variability in terms of standardized anomalies (absolute difference from long-term mean divided by standard deviation). The study also does not consider year-to-year variability in temperature or other climatic variables, as these are less important for hydrological changes than year-to-year variability in rainfall. Also, the study assumes that day-to-day variability in rainfall remains unchanged.

Table 3. Definition of the Climate Change Scenariosa
Mean BaselineChange in Mean RainfallBaseline Interannual Rainfall VariabilityChange in Interannual Rainfall Variability
  • a

    C, constructed from HadCM3; I, constructed from HadCM3 interpolated down to 0.5 × 0.5°; A, constructed from HadAM3H; R, constructed from HadRM3H. Subscript refers to the scenario mean (observed or simulated): a, absolute change; p, % change. Superscript refers to the scenario variability (observed or simulated): y, change in variability; n, no change in variability.


[33] Scenarios Cobs-aobs-n, Iobs-aobs-n, Aobs-aobs-n and Robs-aobs-n apply absolute changes in multi-annual (30-year) monthly mean rainfall (subscript obs-a), temperature, vapor pressure and net radiation, and percentage changes in wind speed, derived from HadCM3, interpolated HadCM3, HadAM3H and HadRM3H respectively, to the observed 1961 to 1990 climate time series. The superscript obs-n denotes that the scenarios use observed variability and (obviously) that this is unchanged for the future period. Cobs-pobs-n, Iobs-pobs-n, Aobs-pobs-n and Robs-pobs-n differ by applying percentage changes in rainfall (subscript obs-p). Whilst climate modelers tend to show rainfall changes in absolute terms (usually in mm day−1), impacts modelers generally apply ratio or percentage changes in rainfall to observed baseline data. This is believed to be more appropriate than applying absolute changes because it reduces the problems caused by inaccurate model simulation of the baseline climate. Applying absolute changes keeps the variance of the observed time series constant but changes its coefficient of variation, whereas applying percentage changes alters the variance but maintains the coefficient of variation. Problems may arise when calculating percentage changes when the baseline rainfall is zero but the changed rainfall is not: In practice, this never happened across the study region with any of HadCM3, HadAM3H or HadRM3H. Problems arise in applying absolute changes if a reduction is greater than the baseline observed rainfall to be perturbed: This happened in at least one month in 2099 in HadRM3H, because HadRM3H overestimates rainfall across much of the region.

[34] Scenarios Asimobs-n and Rsimobs-n are different in that for the baseline climatology they use the multi-annual mean monthly climate data (described in Table 1) as simulated by HadAM3H and HadRM3H respectively (subscript sim), with the observed year-to-year variability in rainfall. In all these scenarios, year-to-year variability in rainfall is assumed to be the same in the baseline period and over 2071–2100.

[35] The remaining scenarios are all based on HadRM3H. Robs-asim-y applies changes in mean monthly climate in the same way as Robs-aobs-n, but uses the simulated (superscript sim-y) rather than the observed year-to-year variability in rainfall. Robs-psim-y applies percentage changes in rainfall. In both Robs-asim-y and Robs-psim-y the variability in year-to-year rainfall is different between the baseline and future climates. Robs-asim-n and Robs-psim-n are variants of Robs-asim-y and Robs-psim-y which assume that the simulated year-to-year variability in rainfall in the baseline continues into the future (superscript sim-n). Finally, scenario Rsimsim-y uses the simulated HadRM3H climate time series directly; Rsimsim-n assumes that the year-to-year variability in rainfall in the future is the same as in the baseline period.

[36] The regional climate model-based scenarios described by Whetton et al. [2001] for southeastern Australia correspond to Robs-pobs-n in this study, as indeed do the UKCIP02 scenarios [Hulme et al., 2002]. Mearns et al. [1999b, 2001a] compared the consequences of coarse-scale and regional climate model scenarios for crop yield in the Midwest of the United States. Their scenarios correspond to Cobs-pobs-n and Robs-pobs-n as defined in this study, although they were based on just five years of climate model simulation. Bergstrom et al. [2001] downscaled two global climate model simulations to Swedish catchments using a regional climate model, following the Robs-pobs-n approach used here. Leung and Wigmosta [1999] used regional climate model simulations directly to drive catchment hydrological models in the Pacific Northwest of the United States, corresponding to method Rsimsim-y in this study. They also downscaled the regional model output from its 90 km resolution to 100 m and 200 m in order to run the hydrological model in their two study catchments, using an orographic precipitation model: They used seven years of regional climate model data for the baseline run, and eight for the changed climate. Venalainen et al. [2001] also used regional climate model data directly (daily temperature) to estimate changes in soil frost depth in Finland. Again, this corresponds to method Rsimsim-y. Of all these studies, only Mearns et al. [1999b, 2001a] compared the different ways of constructing scenarios.

5. Effects on Changes in Runoff

5.1. Change in Average Annual Runoff

[37] Figure 6 shows the percentage change in average annual runoff across southern Africa by the 2080s, under all 16 scenarios considered: Note that the large percentage changes in runoff in dry southwestern Africa represent small absolute changes. The pattern of change broadly follows the pattern of change in precipitation, with larger areas showing a reduction in runoff than precipitation because of the effect of increased evaporation. Average annual runoff generally decreases south of 10°S by well over 20%, with small areas of increase in northern Tanzania and western Madagascar.

Figure 6.

Change in average annual runoff by 2071–2100, under the sixteen scenarios. See Table 3 for definitions of codes.

5.1.1. Climate Model Resolution and Formulation

[38] The primary effect of the different model resolutions used to construct the scenarios (compare Cobs-aobs-n, Iobs-aobs-n, Aobs-aobs-n and Robs-aobs-n or Cobs-pobs-n, Iobs-pobs-n, Aobs-pobs-n and Robs-pobs-n) is to produce different spatial patterns of change in runoff. As with the change in rainfall (Figure 5), the changes under HadAM3H and HadRM3H are different to those from both the original and interpolated HadCM3 scenarios, with reductions in runoff across the central part of southern Africa rather than increases, and greater increases in runoff in the north east of the region. This results from the different regional response to climate change in these models in which both increased resolution and the improved model formulation play a role.

5.1.2. Absolute and Percentage Change in Rainfall

[39] A comparison between, Cobs-aobs-n, Cobs-aobs-n, Aobs-aobs-n or Robs-aobs-n on the one hand with, Cobs-pobs-n, Iobs-pobs-n, Aobs-pobs-n or Robs-pobs-n on the other reveals the effects of the two different ways of applying the changes in rainfall. The broad spatial patterns are the same, but applying rainfall changes as percentage rather than absolute changes tends to reduce the magnitude of the change in runoff, particularly in southwestern southern Africa. This is because the climate models tend to overestimate rainfall in this area, and relatively large absolute changes are applied to low observed rainfall values: In fact, in many months, the perturbed rainfall is set to zero because the absolute reduction is greater than the original baseline value. The difference in change in runoff is greatest where the difference between observed and simulated baseline rainfall is greatest.

5.1.3. Using Climate Model Output Directly

[40] It was shown in Figure 4 how the use of model-simulated and observed climate data led to different estimates of the current amount and distribution of runoff across southern Africa. Although the broad pattern of change is similar whether changes are applied to observed baseline data or the simulated climates are used directly (compare Asimobs-n with Aobs-aobs-n and Aobs-pobs-n, and Rsimobs-n with Robs-aobs-n and Robs-pobs-n), there are some regional differences and, most importantly, the magnitudes of change are different. Using the climate model data directly generally results in smaller percentage changes in runoff than when changes in climate are applied to the observed baseline because the simulated baseline climate is more humid than the observed: A given percentage change in rainfall has a greater relative effect on runoff the lower the ratio of runoff to rainfall.

5.1.4. Using Interannual Variability From the Regional Climate Model

[41] A regional climate model has the potential to add credible fine-scale detail to simulations of year-to-year variability derived from the coarser-scale climate models. This leads to three related questions:

[42] 1. What are the effects on simulated baseline runoff of using the simulated interannual variability in rainfall rather than observed variability? In general, using the observed interannual variability produces more runoff, especially in dry regions. This is because in these areas higher-than-average rainfall results in very large increases in runoff, and mean runoff is heavily influenced by the few large “outlier” years. The simulated interannual variability of rainfall (CV) is lower than the observed variability (Figure 3), so produces less extreme high rainfalls.

[43] 2. What are the effects on the change in simulated runoff of using simulated rather than observed interannual variability? Use of the simulated interannual variability produces a greater reduction in annual runoff across much of the region than use of the observed variability (Robs-pobs-n compared with, Robs-pobs-n, Robs-aobs-n compared with Robs-aobs-n, and Rsimsim-n compared with Rsimobs-n), principally because the simulated baseline runoff is lower (the lower the ratio of runoff to rainfall, the greater the relative effect of a given relative change in rainfall).

[44] 3. What are the effects on the change in simulated runoff of incorporating changes in interannual variability? Incorporating change in variability has very little effect on the pattern of change when changes in climate are added to the observed mean climate (Robs-pobs-y compared with Robs-psim-n, and Robs-asim-y compared with Robs-asim-n), although the increasing CV of rainfall does lead to smaller reductions in central southern Africa and a larger area of increase in southern Mozambique. The difference in pattern of change between Robs-pobs-n (observed interannual variability) and Robs-psim-y (simulated interannual variability with change) is, however, largely due not to the change in variability but the use of simulated rather than observed interannual variability in rainfall.

[45] However, the effect of increased interannual variability is much greater when the HadRM3H data are used directly (compare Rsimsim-n and Rsimsim-y), and incorporating changes in variability leads to smaller reductions in runoff across the southwest.

5.1.5. Range of Change in Runoff

[46] Each of the sixteen different ways of creating climate change scenarios derived from the family of climate simulations produces a different pattern of change. The patterns are broadly similar, but magnitudes of change in runoff are different. Figure 7 shows the range in percentage change in average annual runoff across all sixteen scenarios, and across just the nine scenarios based on HadRM3H.

Figure 7.

Range in percentage change in average annual runoff, under the sixteen scenarios (left) and just the nine HadRM3H scenarios (right).

[47] When all sixteen scenarios are considered, there is a range of at least 20% in change in annual runoff across almost the entire region, a range of over 40% in the south, and an even larger range in change in east Africa (where there are different directions of change in runoff). The range across just the nine scenarios derived from HadRM3H is smaller, and is less than 10% across much of the study region. The range is still, however, greater than 20% across southwest southern Africa and in east Africa.

5.2. Change in Runoff Variability From Year to Year

[48] The previous section focused on mean annual runoff. The variability in runoff from year to year in southern Africa is amongst the highest in the world, and in individual years annual runoff may be several times greater than the mean, especially in the more arid regions. The top pair of maps in Figure 8 shows the spatial distribution of the CV of annual runoff over the baseline period, using the observed interannual variability in rainfall (left) and the HadRM3H-simulated interannual variability together with observed mean climate (right): The simulated variability generally produces a lower CV in annual runoff. Across the study region, the coefficient of variation of simulated annual runoff over the 1961–1990 baseline period ranges from 0.3 to 1.0, with a mean of 0.59.

Figure 8.

Coefficient of variation in annual runoff, with observed and simulated interannual variability in rainfall, and changes in the CV of annual runoff.

[49] The middle pair of maps in Figure 8 show the change in the CV of annual runoff with the change in mean rainfall applied in absolute (left: Robs-aobs-n) or percentage (right: Robs-pobs-n) terms. In both cases the year-to-year relative variability in rainfall remains unchanged, but both approaches result in a change in the CV of runoff. This is partly because of the non-linear relationship between rainfall and runoff, and partly because of the simultaneous changes in other climatic variables. Applying absolute changes produces larger changes in CV of runoff. The CV of annual runoff increases across most of southern Africa, except in the dry southwest and parts of the northeast of the region.

[50] A comparison between the middle right (Robs-pobs-n) and lower left (Robs-psim-n) maps shows the effect on the change in the CV of runoff of using the observed or the simulated interannual variability of rainfall: The differences are small, although there is a slight tendency for the CV of runoff to increase more in the north east when simulated interannual variability is used, and decrease by more in the southwest.

[51] The effect of changing the interannual relative variability in rainfall on the CV of annual runoff can be seen by comparing Robs-pobs-n (no change) with Robs-psim-y (change in variability). Incorporating the modeled changes in relative interannual variability in rainfall leads to substantial increases in the CV of annual runoff across virtually the whole of the region, over and above the increases resulting from a change in just mean climate.

5.3. Change in High Flows

[52] The percentage change in the magnitude of the 10-year return period maximum monthly runoff (M10max) under nine HadRM3H scenarios is shown in Figure 9. The patterns are not as smooth as those for change in average annual runoff because M10max is determined from a probability distribution fitted to the data.

Figure 9.

Change in the 10-year return period maximum monthly runoff by 2071–2100, using scenarios derived from HadRM3H.

[53] In general, the percentage change in M10max is smaller than the percentage change in average annual runoff. When absolute HadRM3H changes are applied to the observed mean climate, there is little difference in pattern of change in M10max whether the observed (Robs-aobs-n) or simulated (Robs-asim-y) interannual variability is used. However, when rainfall changes are applied as percentages, there are indications that the reduction in M10max in the driest parts of the region is smaller when simulated variability (Robs-psim-y) is used than when observed variability (Robs-pobs-n) is assumed.

[54] When the HadRM3H data are used directly (Rsimobs-n and Rsimsim-y) there is a larger difference between changes in M10max and average annual runoff, with a greater tendency for maximum flows to increase even where average runoff decreases.

[55] A more detailed indication of the effect of different ways of creating scenarios on estimated changes in high flow magnitudes is given in Figure 10, which shows the estimated frequency curves for one grid cell located in southern Botswana. The top left graph compares the baseline frequency curves with the frequency curves for the period 2071–2100 under scenarios Robs-aobs-n and Robs-pobs-n: The different ways of applying the change in rainfall produce different frequency curves. In this case, the future frequency curves are lower than that for the present climate. The top right graph shows the estimated baseline frequency curves assuming observed and simulated interannual variability of rainfall, with the observed climatological mean. The curve assuming simulated variability is lower than the other, primarily because the simulated variability series does not generate the very large flows seen using the observed variability. The lower left graph shows the frequency curves using simulated interannual variability in rainfall, with (Robs-psim-y) and without (Robs-psim-n) changes in this variability. The increase in interannual variability in rainfall simulated in southern Botswana leads to a steepening of the frequency curve, almost entirely because of one very large flood event.

Figure 10.

Frequency curves for maximum monthly runoff, for one grid cell in southern Botswana.

[56] The lower right panel of Figure 10 shows the frequency curves resulting from the direct application of the HadRM3H simulations, using the simulated interannual variability of rainfall (scenarios Rsimsim-y and Rsimsim-n, with and without change in interannual variability respectively). The HadRM3H simulations substantially overestimate rainfall in this region, and the baseline curve is therefore very different to that derived from observed climate data (top left). Climate change lowers the frequency curve, and in this instance increasing interannual variability has little effect on the frequency curve.

6. Discussion

[57] The previous section has shown that whilst there is a broad degree of consistency in the spatial pattern of change in runoff deriving from each of the scenarios based on the Hadley Centre family of models, there are substantial local differences and the magnitude of change is very variable. It is not possible to quantitatively evaluate each scenario individually, because there is nothing to compare against, but it is possible to examine the sensitivity of results to different decisions about scenario construction.

6.1. Effects of Moving From HadCM3 to HadRM3H

[58] There are differences in the subregional effects of climate change with the three climate models are used, even if the coarse-scale HadCM3 climates are interpolated to a resolution similar to that of HadRM3. Magnitudes of change are different and there are some areas, notably in eastern Zambia, where the direction of change is different. Mearns et al. [2001a] noted in their study of crop yields that coarse-scale and regional model scenarios could give very different local results. In other words, the regional climate model does more than simply interpolate the broad-scale changes in climate produced by the coarser-scale global model.

6.2. Effects of Moving From HAdAM3H to HAdRM3H

[59] Before running HadRM3H, it is first necessary to run HadAM3H to improve upon certain large-scale deficiencies in the HadCM3 simulations. It is clear from Figure 6 that the difference between HadCM3 (at the original resolution or interpolated) and HadAM3H is greater than the difference between HadAM3H and HadRM3H. This is because the HadAM3H boundary forcing applied to HadRM3H ensures greater consistency between these models than between HadAM3H and HadCM3 from the latter's sea-surface forcing. This implies that over a large geographic domain, such as a continent, it would be appropriate to use the intermediate resolution HadAM3H rather than run a higher-resolution regional model and so the fourth downscaling method outlined in the introduction appears feasible in practice.

6.3. Relative Effects of Applying Percentage or Absolute Changes in Rainfall

[60] Applying changes in rainfall as percentages or absolute values to observed baseline rainfall produces different changes in runoff. The differences are greatest where the climate model simulates the baseline rainfall least well. Both approaches produce changes not only in mean runoff, but also in variability from year to year. It is preferable to apply rainfall changes as percentages, because in approximately two thirds of the grid cells across southern Africa, simulated reductions in absolute monthly mean rainfall were greater than the observed baseline monthly mean rainfall in at least one month.

6.4. Effects of Using Climate Model Data Directly

[61] The primary argument against using climate model simulations directly in impacts models is that if the climate model simulates the baseline climate inaccurately, then these errors will feed through the impacts model and into the impact assessment. HadAM3H and HadRM3H simulate the baseline climate of southern Africa reasonably well, although both overestimate rainfall substantially in the southwestern part of the region. The results of this study showed that whilst the broad pattern of change in runoff was similar whether the simulated climates were used directly or the changes in climate were applied to the observed baseline, the magnitudes of change were different, particularly where the climate model simulated the baseline climate less accurately. These differences are particularly apparent when indicators of extreme flows, rather than average flows, are considered. With even more complicated impacts models (which, for example, routed streamflow through a reservoir system and calculated reliable yield) the differences between the two approaches, and the effects of bias in climate model simulation, would probably be even greater.

6.5. Effects of Using the Model-Simulated Coefficient of Variation

[62] Impacts assessments have conventionally applied changes in mean monthly climate to observed climate time series, thus assuming that the temporal structure of climate remains unaltered. Although this is recognized to be unrealistic (and many impacts are likely to be considerably affected by change in relative variability), this assumption has been forced on impacts modelers by the lack of credible information on changes in interannual variability. In part, this has been due to the coarse spatial scale of the global climate models used to define most scenarios, but even where regional models have been applied they have often only been run for short periods. HadRM3H has been run to produce 30-year simulations.

[63] In general, HadRM3H simulates an increase in the relative variability in rainfall from year to year, and adding the effects of this increase to the effects of changes in mean climate produces relatively higher streamflow. Simulated increases in runoff are therefore larger, and simulated reductions smaller. However, use of HadRM3H time series directly (whether applied to observed or simulated mean climate) in southern Africa is problematic, because HadRM3H underestimates the coefficient of variation in the rainfall interannual variability. Whilst this does not greatly affect estimated average runoff, it does affect frequency distributions and would affect even more water resources indicators that are sensitive to the frequency and timing of extreme flow conditions, such as reservoir reliability.

[64] One response to this problem is to apply changes in the coefficient of variation of monthly rainfall to the observed baseline time series (as done by Arnell [2003] in the UK), to produce a perturbed historical time series with altered relative variability. However, this approach cannot account for possible changes in the pattern of rainfall variability from year to year.

7. Conclusions

[65] This paper has investigated several ways of using regional climate model output to create climate scenarios characterizing changes in mean climate and interannual variability over a large geographic domain. The effects of these different methods on estimated changes in runoff across southern Africa have been evaluated, although it must be emphasized here that the aim of the exercise was not to estimate the effects of climate change in southern Africa but rather to compare the different scenario construction methodologies. All the scenarios were based on just one global climate model forcing, and other climate models could give different forcings and therefore regional changes in climate.

[66] The resolution of the climate model used to create climate scenarios clearly has an effect on the spatial resolution of the effects and impacts of climate change, and both the medium resolution global model and the high-resolution regional model produce different changes in climate, and hence runoff, than scenarios interpolated from the coarse resolution model. The higher-resolution changes are more plausible than the coarser-resolution changes interpolated down to the finer scale because they use physical processes to generate local climates reflecting local controls on climate. They are, however, ultimately only as reliable as the changed large-scale climate.

[67] The different ways of deriving climate scenarios from the HadRM3H regional climate model result in a range in change in average annual runoff by the 2080s of around 10% (greater in south-western Africa). This range is enlarged considerably when climate scenarios constructed from the driving HadCM3 (the coupled ocean-atmosphere model run in transient mode) and HadAM3H (the high-resolution atmosphere model run in a time-slice experiment) models are included. Although the broad patterns of change in runoff are similar, there are parts of the study region where the three climate models produce different directions of change.

[68] The difference in pattern of change in runoff between HadCM3 (whether at the original 3.75 × 2.5° resolution or interpolated) and HadAM3H (1.88 × 1.24°) is greater than the difference between HadAM3H and HadRM3H (0.44 × 0.44°). This results from the experimental design, which ensures much greater large-scale consistency between the latter two models. This suggests that HadAM3H, built as a model to address deficiencies in the simulation of large-scale climate features in HadCM3 and providing higher-resolution information, can provide very useful scenarios for studies of the impacts of climate change over a large geographic domain.

[69] Using the absolute climate as simulated by the regional climate model produces biased estimates of runoff (assuming that the runoff simulated from observed baseline data is accurate). Also, the changes in runoff indicators are different from those that result when a change in climate is added to an observed baseline climatology. In parts of southern Africa HadRM3H overestimates rainfall. This produces too much runoff and the ratio of runoff to rainfall is also overestimated. Because the effect of a given change in climate on runoff increases as the runoff/rainfall ratio declines, overestimating the runoff/rainfall ratio underestimates the effect of climate change. It is therefore recommended that regional climate model simulations are used directly only where the regional model simulates regional and local climate well. More generally, regional climate models can be used to construct scenarios defining a change in climate.

[70] Expressing a change in rainfall in absolute terms (mm day−1, for example) can produce different estimates of change in runoff than expressing the change in rainfall in percentage terms. The greater the “error” in the simulation of baseline rainfall, the greater the difference. In two thirds of the grid cells in the region, absolute reductions in rainfall were greater than the baseline average rainfall in at least one month. It is recommended that rainfall changes are applied as percentages, particularly where baseline rainfall is overestimated.

[71] Using the simulated, rather than observed, interannual variability in rainfall to drive the runoff model has little effect on simulated average annual runoff, either over the baseline period or 2071–2100. The two different sets of anomalies, however, do result in different frequency distributions of runoff, largely because the simulated coefficient of variability (CV) of rainfall is generally lower than the observed. Climate change generally increases the interannual variability (CV), and including this additional variability has little effect on simulated average runoff. However, it has a larger impact on extreme flows with predicted increases in frequency becoming larger and reductions smaller and increases predicted even when the mean runoff reduces. Because the climate model does not simulate the coefficient of variability of rainfall very accurately, it is not appropriate to use it in hydrological simulations. However, it is feasible to apply the simulated change in coefficient of variation in rainfall to the observed baseline time series to produce a perturbed time series with unchanged temporal structure but altered relative variability. Changes in extreme events are likely to be affected by changes in day-to-day variability as well as changes in year-to-year variability: Again, there are several possible ways of extracting such information from regional models, and a similar comparative analysis is required.

[72] In conclusion, regional climate models offer many potential advantages over coarse-scale global climate models when creating climate scenarios covering a large geographic domain: The spatial resolution is finer, capturing many of the local effects on climate caused by, for example, topography, and it is therefore possible in principle to place greater confidence in simulated time series. Regional models offer advantages over statistical downscaling techniques in that they are not so explicitly empirical and automatically produce consistent climates over an entire region. Regional models have therefore often been cited as the way forward for the construction of “improved” climate scenarios, and some effort is going into coupling “impacts” models to regional climate models. However, the results from this study suggest that unless the regional climate model simulates baseline regional climate accurately, it is better to use the regional model to estimate changes in climate (mean and variability) that can be applied to observed baseline data. Even so, it must be remembered that changes in climate as simulated by a regional climate model are strongly influenced by the driving climates as simulated by coarse resolution global climate models.


[73] This work was funded by the Hadley Centre for Climate Prediction and Research under the UK's Department for Environment, Food and Rural Affairs (DEFRA) Climate Prediction Programme (PECD 7/12/37). D. Hudson was working at the Hadley Centre on a subcontract for the duration of this research and would like to thank them for access to their facilities. She would also like to thank the University of Cape Town (Department of Environmental and Geographical Science) for granting her an extended period of special leave to pursue research in the United Kingdom during this period. The authors thank the anonymous referees for their comments.