Figure 2 shows the correlation between monthly streamflow simulations and observations for all stations across different sub-regions (see Table S2). The correlation values show large variations in the GHA region. Firstly, there is a large variation in the performance of all six streamflow simulations and the mean. The highest correlations (> 0.8) with observed streamflow were found in the Blue Nile and Baro basins (A), while the lowest correlations (all < 0.5) were in basins that are located in the southern part of the GHA region (F). Correlation values for the other stations largely vary between 0.5 and 0.8. The relatively low correlation for Makalal station (M0017) on the White Nile, South Sudan, is likely due to the flow regulation by Lake Victoria, which has not been included in the hydrological model. Most stations in the Upper Ghibe basin (C) and the Hombole station (AW031) in the Awash basin have a correlation value > 0.7. The other four stations in Awash basin (D) are influenced by the Koka Dam, showing low correlation due to the dam influence. Finally, stations in the Juba-Shabelle basin (E), which has a bimodal season, have low correlation values (< 0.6).
Secondly, the correlation values show large variations between the six streamflow simulations for each station in the region. In general, there is a better agreement between the ensemble simulations for basins with high correlation (e.g., in Blue Nile basin), and less agreement in basins where the correlation with the observed streamflow is low (e.g., Juba-Shabelle stations). Simulations produced with IPSL-CM5A-LR or IPSL-CM5A-MR forcings outperform the rest for the majority of the stations. In general, the mean of the six streamflow simulations has higher correlation than any individual run, with very few exceptions.
The KGE also shows a regionally varying performance with > 0.8 for stations in the Nile basin and < 0 in Tanzanian basins (Table S4). Overall, the monthly streamflow has a median KGE of 0.6. The time series plots of monthly flows (after bias correction) for selected stations are shown in Fig. S2.
Figure 3 presents the shape of the flow climatology as the percentage mean monthly contribution to the annual flow for six selected stations (see Fig. S3 for the remaining stations). In common with the streamflow correlations presented earlier, the shapes of the flow climatology vary across the sub-regions. The flow climatologies of the observed and simulated streamflow agree for the Eddiem station (G1684, 3A), with all simulations correctly indicating the peak flow in August. Overall, the IPSL-CM5A SST-forced simulations have the best shape agreement, while EC-EARTH and GFDL-ESM2M display the least agreement. The peak flow month (August) was also correctly captured by all simulations for Wabi station (Fig. 3c) in the Upper Ghibe river basin and Hombole station (Fig. 3d) in the Awash river basin.
However, for the remaining selected stations (Fig. 3), the modelled streamflow could not match either the shape of the observed climatology or the peak flow month. All simulations indicate the peak flow in September for Makalal station (Fig. 3b) in the White Nile, 1 month earlier that the observed peak flow. As indicated earlier, the flow at this station is highly regulated by Lake Victoria. All simulations underestimate the relative contribution of the peak to the annual flow for the Hombole station (Awash). For Afgooye station (Shabelle, Fig. 3e), there is large disagreement between simulations on both the flow seasonality and the relative monthly contributions. Three simulations (IPSL-CM5A-LR, IPSL-CM5A-LR and HadGEM2-ES) miss the MAM flows and indicate most of the total annual flow occurring during July–October. The other three simulations correctly capture the bimodality of the flow for this station, but they relatively amplify the flow contribution during the MAM (EC-EARTH and GFDL-ESM2M) or OND (GISS-E2-H SST) seasons. While the simulations appear to agree on the unimodality of the flow season for Ruvuma river in Tanzania (Fig. 3f), they disagree on the peak flow month, ranging from January (GFDL-ESM2M) to March (GISS-E2-H). All of those are earlier than the observed peak flow in April. The streamflow simulations fail to capture the high flows in March–May for all stations in Tanzania.
The overall streamflow evaluation highlights the challenge of climate impact assessment in the GHA region that has such a highly variable hydroclimate. This is especially evident in the southern part of the region, where the monthly correlation is low and the streamflow seasonality is not well captured. Even the high-resolution climate simulations have a mixed performance in this sub-region. Nevertheless, our findings further underscore that an accurate climate impact assessment requires locally relevant climate models.
Changes in mean flow
Figure 4 shows the percent change in long-term mean streamflow for 2040s compared to the baseline period for each simulation driven by the SSTs from the six CMIP5 models. The simulations show different sign and magnitude of streamflow change across the GHA region. The changes can generally be classified into three groups: (1) streamflow decrease in most parts of the GHA region (EC-EARTH and IPSL-SM5A-MR), (2) a large decrease in the southern (Tanzania) but a small change in northern parts of the GHA region (HadGEM2-ES SST) and (3) a decrease in the northern (Sudan and northern Ethiopia), increase in the equatorial and decrease in the southern parts of the region (IPSL-CM5A-LR, GFDL-ESM2M and GISS-E2-H).
The magnitude of streamflow changes varies across the different climate simulations. EC-EARTH produced the driest projection with more than 25% streamflow reduction in large parts of the GHA region, especially the Juba-Shabelle and Awash rivers. Conversely, GISS-E2-H and GFDL-ESM2M produced projections of more than 25% increase in equatorial part of the region and the Juba-Shabelle basin. It is worth noting that the disagreement between the EC-EARTH and GISS-E2-H on the direction of streamflow change in the Juba-Shabelle occurs despite the fact that both runs correctly capture the bimodality of the flow season in the basin. Mean streamflow (averaged over the six simulations) shows a − 10 to 5% change in Ethiopia, up to 25% increase in parts of the equatorial region and up to 25% decrease in the southern and western parts of the GHA region.
While the pattern generally remains the same through the twenty-first century, the magnitude of the streamflow change intensifies in 2080s (2070–2099, Fig. 5). Two simulations (EC-EARTH and IPSL-SM5A-MR) show the largest mean streamflow decrease (− 50 to − 80%) for Ethiopia and Sudan, resulting in substantial streamflow reductions in the Nile, Awash and Juba-Shabelle rivers. Two simulations (IPSL-CM5A-LR and GFDL-ESM2M) show large streamflow increases (up to 100%) in the equatorial regions, but decreases (> 50%) in large parts of Ethiopia and in Sudan. The HadGEM2-ES and GISS-E2-H forcings produce large flow increases in most parts of the GHA region but show slight decreases in central and northern Ethiopia.
The ensemble mean shows more than a 10% increase in the equatorial part of the region, but a decrease in Blue Nile and Shabelle basins. The flow increase is in agreement with previous studies that reported a significant increase in rainfall in the equatorial part of the region in 2080s using ensemble mean of 4 CMIP5 models with RCP8.5 scenarios including EC-EARTH and HadGEM2-ES (Endris et al. 2016).
Changes in seasonal flows
Figure 6 shows the projected streamflow change in the summer season (JJAS) for 2080s. As noted earlier, the JJAS season is the main rainy season in the northern part of the GHA region, which includes Blue Nile, Upper Awash, Ghibe and Baro river basins. Overall, the simulations show a large flow decrease for those basins in 2080s compared to the baseline period. All simulations indicate a 10–50% flow decrease in Awash, Ghibe and Juba-Shabelle basins. The EC-EARTH shows the driest future during the JJAS season in the whole GHA region. Similarly, the IPSL-CM5A-MR and IPSL-CM5A-LR produced substantial flow decrease in the northern part of the region including Ethiopia, South Sudan and Sudan. They, however, show substantial flow increase in equatorial regions with a dry JJAS season. The GFDL-ESM2M follows a similar trend. The remaining simulations (HadGEM2-ES and GISS-E2-H) show an increase in flow (> 25%) in large parts of the GHA region except central Ethiopia, where a flow decrease is projected. The decrease in the northern part of the GHA region is in agreement with the study of Endris et al. (2016), which reported a rainfall decrease during JJAS. The mean of the six simulations shows a flow decrease (ranging from − 5 to − 50%) in the major rivers in the region including Nile, Juba-Shabelle, Awash and Ghibe rivers. However, a modest to large flow increase (10–50%) is projected in the equatorial regions during the JJAS season.
The mean streamflow change during the JJAS season in 2040s has the same general pattern as in 2080s (Fig. S4). Parts of the GHA region with projected flow increase (decrease) in 2080s also have the same increase (decrease) in 2040s, but the magnitude of the percentage change is smaller in 2040s. Interestingly, the HadGEM2-ES leads to inconsistent increasing/decreasing trends across different parts of the region, specifically in Tanzania, Rwanda and Kenya.
Large disagreement between the six streamflow simulations is observed for projected streamflow changes in 2080s during the MAM season (Fig. 7). EC-EARTH again shows a drying trend in the MAM season in the majority of the GHA region with the eastern part of the region (Somalia, Kenya, eastern Ethiopia) having more than 50% flow reduction in 2080s. The IPSL-CMA-LR and IPSL-CMA-MR produce large increase (in some parts up to 100%) in the equatorial region where the MAM season is one of the two rainy seasons. Conversely, they show a reduced flow in eastern Ethiopia, affecting the flows in the Awash and Shabelle rivers. The other three simulations (GFDL-ESM2M, HadGEM2-ES and GISS-E2-H) indicate large flow increase during the MAM season across the GHA region. This can be seen, for example, from the large flow increase in the Nile River during the MAM season. The ensemble mean shows an increase in streamflow in the equatorial part and a decrease in Ethiopia. The percentage flow change (increase or decrease) is larger during the MAM compared to the JJAS season.
For the 2040s (Fig. S5), the flow increase/decrease trend in MAM season is consistent with the 2080s for three simulations. The EC-EARTH essentially indicates the same pattern and trend, while IPSL-CM5A-LR and GFDL-ESM2M retain the same pattern but with lower percentage changes in 2040s. The remaining three simulations show inconsistent change patterns. The HadGEM2-ES shows a large flow increase in the equatorial region in 2080s but decrease in 2040s; the GISS-E2-H shows a flow decrease in south-eastern Ethiopia in 2080s but large increase in 2040s; and IPSL-CM5A-LR shows a flow increase in the equatorial region (e.g. Kenya) in 2080s but decrease in 2040s compared to the baseline period. The mean of the six simulations shows a consistent flow trend (decrease in northern and increase in the equatorial parts) but with lower change magnitudes in 2040s compared to 2080s.
Figure 8 shows the percentage change in streamflow during OND, the short rainy season. The streamflow simulations show different flow trends during the OND season across the GHA region. Three simulations (GFDL-ESM2M, HadGEM2-ES and GISS-E2-H) show large streamflow increase (some exceeding 100%) in 2080s, while two simulations (IPSL-CM5A-LR and EC-EARTH) indicate a modest flow increase in western Kenya and parts of Somalia. Conversely, IPSL-CM5A-MR shows a substantial flow decrease in the equatorial region during the OND season. IPSL-CM5A-MR also shows an extensive drying trend during the OND season over large parts of the GHA region. The OND streamflow changes in 2040s (Fig. S6) have a consistent trend with those in 2080s. The ensemble mean shows a streamflow increase in the equatorial sub-region during the OND season.
Changes in extreme flows
Figure 9 presents the low (Q95) and high (Q05) flow magnitudes for the baseline period and the percent change in 2040s and 2080s. The Q95 and Q05 here were computed based on the streamflow mean for the six simulations. A small change (± 5%) is projected in the low flow magnitude in the 2040s in the GHA region, except in Tanzania where the low flow reduction is higher, ranging 10–25% (Fig. 9b). However, a higher reduction (5–25%) in low flow is projected in 2080s for the northern parts of the GHA region (including Ethiopia) and for Tanzania (Fig. 9c). This has important implications on the water availability in river channels during the dry seasons and may severely impact water supply in the rapidly growing populations of Ethiopia and Tanzania, which are expected to have a combined population of 0.55 billion (3.3 times the current) at the end of the twenty-first century (United Nations 2017a).
The relative change in high flow magnitude is more pronounced. It is projected to decrease by up to 10% in 2040s for major rivers such as the Blue Nile, Shabelle, Awash and Ghibe (Fig. 9e). Conversely, it is projected to increase (5–25%) in the equatorial parts of the region. In 2080s, the percent change in high flows is higher, with flows in the major rivers projected to decrease by up to 25% of the baseline, and the equatorial part will experience substantial flow increase (Fig. 9f). The high flow reduction in Ethiopia is consistent with the reduction in mean flow across all seasons presented earlier. The reduction in extreme high flows, which is the main source of groundwater recharge in East Africa and semi-arid environments (Döll and Fiedler 2008; Taylor et al. 2013), will negatively impact groundwater recharge in large parts of the GHA region.
SST influence on streamflow
Studies report that East African rainfall variability is linked to the SST variations in the tropics; specifically, the variations in western and central Pacific and Indian Oceans have a strong influence on the MAM and OND rainfall patterns in the GHA (Black 2005; Bhattacharjee and Zaitchik 2015; Funk et al. 2014). The MAM is the planting season in the region and low rainfall conditions during this period have led to food insecurity and malnutrition in the region. The GHA spring droughts are linked to the strengthening of the Walker circulation, whose variations are closely associated with El Niño/Southern Oscillation (Vecchi et al. 2006; Williams and Funk 2011). The relative warming of western Pacific and cooling of central Pacific intensifies the Pacific branch of the Walker circulation, creating low (high) pressure system in west (central) Pacific Ocean (Funk et al. 2013). Previous studies have shown that the strengthening of this east–west atmospheric circulation in the Pacific can lead to decreased MAM rainfall conditions in eastern Africa (Hoell and Funk 2014; Funk et al. 2015).
Here. we compare the SSTs from the CMIP5 models with Hadley Centre Sea Ice and Sea Surface Temperature dataset (HadISST) (Rayner et al. 2003) during the 1976–2005 period. The CMIP5 models have wide-ranging SST biases in different parts of the globe (Fig. S7). Figure 10 shows the area-average SST over the west Pacific (10° S–10° N, 110°–150° E) and Niño-4 (central Pacific) region (5° S–5° N, 160° E–150° W) and the difference between SST over the two regions. We focus on these regions due to the reported sensitivity of the MAM season rainfall in the equatorial GHA to the SST in the regions. EC-EARTH has consistently colder SST in west Pacific with a mean value of − 2.98 °C during the 1976–2005 period compared to the observation, while IPSL-CM5A-MR has warmer bias with mean 0.65 °C (Fig. 10a). Other SST values closely match the observation over the west Pacific (± 0.25 °C). Over the Niño-4 region, models have largely colder SST than the observation with long-term mean values ranging from an average of − 0.71 °C (IPSL-CM5A-MR) to − 1.72 °C (EC-EARTH). Only GISS-E2-H has slightly warmer SST (0.31 °C) than observed values over the Niño-4 region. The SST differences between the two Pacific regions (Fig. 10c) show larger differences between the models. EC-EARTH shows a consistently warmer central Pacific with a long-term mean of + 1.04 °C, which can weaken the Walker circulation associated with increasing MAM rainfall in the equatorial GHA region. All other models and the observed record (except in the early 1990s) show a warmer west Pacific.
In agreement with previous findings (Funk et al. 2014; Funk et al. 2015), the MAM streamflow appears to be linked to the SST difference between the Nino-4 region and western Pacific. Figure 11 presents the time series of the MAM streamflow and SST (Nino-4 minus western Pacific) for a selected station in the Juba-Shabelle river basin in the equatorial GHA region. Simulations with the lowest SST difference between Nino-4 and western Pacific SST have the lowest MAM streamflow in the basin, while the simulation with the highest SST difference (EC-EARTH) largely produced the largest MAM streamflow. However, the reason behind the sudden decline in EC-EARTH streamflow after 1995 is unclear.