The trends in the annual Q01 and Q90 data for three periods (from 1981 up to the end of early-, mid-, and end-century) are analysed for each impact model and driving GCM for all five basins and two RCP scenarios (RCP2.6 and 8.5 representing a mild and an extreme emission scenario) (Fig. 1). Overall, the direction of Q01 trends lack statistical significance except the identified end-century trends (and mid-century for the Lena basin), particularly for RCP8.5, while Q90 trends are significant for most basins for the mid- and end-century, particularly for RCP8.5 (only Lena shows significant trends for the mid-century for RCP2.6). It is important to note that for a given impact model the direction of future trends varies depending on the driving CM (as expected); however for a given climate model, the direction of trends from the impact models is consistent. Moreover, for Q01 and for some CM-HM combinations (e.g. for the Ganges, Lena and Niger basins under RCP2.6), the trends appear to be stronger for the early-century compared to the end-century, yet they tend to be significant only for the later periods. This temporal variability in the flow trends is subject to effects of joint precipitation and temperature temporal changes.
Overall, an increasing trend in high and low flows is shown for the Lena basin (increasing in significance with increasing emissions scenarios and future period; for Q90 the significance is increased with increasing future period only for RCP8.5) but in the Rhine basin only for the end-century for RCP2.6. For the Ganges basin an increasing trend is shown only for high flows; yet the strength is not increased. Decreasing trends in high and low flows are shown for the Rhine and Tagus under RCP8.5 for the mid- and end-century, while the pattern of trends for the Niger is highly uncertain. Although statistically significant trends are projected for the Niger River basin for the end-century for both RCP2.6 and 8.5, the trends for the latter show an alternation of positive (quite strong) and negative slopes depending on the GCM scenario. Again, this highlights the presence of uncertainty in the future climate projections. Of particular interest is the alternation from positive (although not statistically significant) to negative trend during the early- and end-century. This could be due to the temporal patterns of the climatic variables in the Niger basin and was also found with a different set of RCMs by Aich et al. (2015).
Changes in long-term means
Results for changes in Q10 and Q90 are presented in Figs. 2 and 3 for all five basins, four RCPs and three future periods. The relative difference in changes depends on the emission scenario, basin, and future period; however, in general, changes in Q10 and Q90 vary within similar ranges (as Figs. 2 and 3 show between −50 and 200 %), with the Ganges basin being an exception since changes in Q10 seem to be more significant than changes in Q90 in the mid- and end-century.
As shown by most of the model combinations, Q10 in the Ganges and Lena is expected to increase, while Q10 is generally shown to decrease in the Tagus. Results for the Rhine and Niger basins are uncertain with the potential changes being dependent on the CM. In particular, in the Ganges, Q10 is shown to increase by up to ~100 % at the end of the century. Although projected changes in the Ganges are largely variable between the driving GCM models, the direction of change is consistent between the HMs for a given GCM model. Consistently across projections, the Lena basin high flows are shown to slightly increase with increases of up to ~40 % at the end of the century. Lena’s flow regime is strongly influenced by snow/ice accumulation/melting processes, which within the HM context are heavily controlled by temperature. It is therefore expected that, given also the increase in precipitation, the temporal distribution of temperature has a major impact on this result. Conclusions are supported by the changes in the monthly Q10 values; see the impacts on seasonality for all basins and models in Fig. S1 in the Supplement.
For the Niger, results are highly variable with for example the MIROC-driven outputs projecting increases in Q10 from 100 to 300 %. A similar range in Q10 change for this basin was observed by Aich et al. (2014). The discrepancies between CM-driven results are amplified in the end-century; however, the MIROC-driven outputs are highly variable even between the impact models. This could be due to HM structural inability to represent the processes in the floodplain of the Inner Niger Delta, which has a remarkable impact on the flow regime (about 40 % of the inflowing water evaporates from the floodplain). The Rhine basin shows the smallest expected changes in high flows (between ~ ±15 %) in all RCPs and future periods. As expected, the projected changes are larger for RCP8.5 compared to other RCPs, but still smaller in comparison to the changes shown in the other basins. In the Tagus basin, for higher end emissions scenarios (RCP4.5 to 8.5), a reduction in Q10 is projected for the end-century following the general reduction in precipitation over the region. The pattern is more complicated in the early-century with some driving models projecting increases (IPSL and Hadley) and others projecting reductions.
For low flows, changes related to the different HMs and driving CMs seem to be highly variable. Overall, an increase in Q90 is projected for the Lena (all CM-HM combinations) and Niger basins (apart from the GFDL-driven and some cases of the IPSL-driven results in the mid- and end-century), while a reduction is shown for the Rhine (apart from very few CM-HM combinations in the end-century) and Tagus (apart from cases of the IPSL-driven results). Results for the Ganges are uncertain, yet most CM-HM combinations show a reduction for RCP8.5 (up to ~ −50 %).
In particular, the Ganges results seem to be more dependent on the HM, RCP and future period than CM. On the contrary, results for the snowmelt dominated Lena basin show a consistent increase in low flows (increasing in significance with increasing emission scenarios and future period). In the Niger basin, increased Q90 seems to be expected due to the increase in precipitation (between 5.3 and 12.2 % in the early- and mid-century according to Krysanova and Hattermann, this special issue). However, the variability between future changes is very high for the end-century. In the Rhine basin, the reduction in Q90 (up to −50 % in the end-century) seems to occur due to the joint effect of temperature and precipitation; a high increase in temperature (between 0.9 and 4.4 °C) and a relatively minor increase in precipitation (between 1 and 6.6 %); see Krysanova and Hattermann (this special issue). In the Tagus basin, most of the model runs indicate a reduction in Q90 of up to −50 % following the projected decrease in precipitation (between −0.9 and −27.3 %). Notable exceptions are the IPSL-driven projections that indicate an increase for some impact models, RCPs and future periods.
Figure 4 illustrates the Q10, Q90 anomalies plotted against Pmean anomalies for each year of the period 2006–2099 under RCP8.5. An increase in precipitation results in an increase in high and low flows; however this relationship is not linear and its strength varies between the basins in accordance to their runoff coefficient (Table 1 in the introductory paper by Krysanova and Hattermann). The Q10 to Pmean and Q90 to Pmean sensitivities vary significantly between basins with the Niger’s response being the most sensitive to precipitation. Results for the Niger show that an increase (reduction) of 25 % in precipitation will result in ~60 % increase (~ 50 % reduction) on average in Q10. Results for the Niger are particularly interesting with the Q10 to Pmean and Q90 to Pmean anomalies being clustered since the MIROC-driven results are different from the other GCM-driven results. This highlights how different the sensitivity of the results is to different CMs.
In the Lena basin, an increase (reduction) of 25 % in annual precipitation results in an ~30 % and 45 % increase (~ 25 % and 20 % reduction) on average of Q10 and Q90 respectively. As mentioned in Section 4.2, snow/ice accumulation/melting (which are heavily controlled by temperature) strongly affects the flow regime in this basin. A similar sensitivity to climate is observed for Q10 in the Rhine basin. The strength of the relationship is slightly lower than that for the Niger (yet adjusted R2 ≥ 0.25 (which indicates an acceptable degree of predictability) only for Q10), however the response to a 25 % increased annual precipitation is a 30 % increase in Q10. For the Ganges and Tagus an increase in mean precipitation of 25 % results in increases in Q10 of 50 and 35 %, respectively. However the Q10 to Pmean and Q90 to Pmean trends are not statistically significant (based on the adjusted R2 value). Uncertainties from the GCMs and impact models seem to mask these relationships, and therefore caution is needed regarding the interpretation of these results.
The analysis for high flows was repeated using the 90th percentile of precipitation (P10) instead of Pmean (Fig. S2 in the Supplement). The results were very similar to those above, however the strength of the relationships was slightly lower (slightly smaller adjusted R2 values), and hence conclusions remain the same.
Uncertainty in future projections
Figure 5 analyses uncertainty in future projections, as this can be decomposed into uncertainty from the driving CMs and HMs, and related to the basins’ climatic conditions for RCP2.6 and 8.5 (see explanation in Section 3.3). The basins are ranked based on their dryness and evaporative indices. The Niger basin has the warmest climate and hence the highest dryness index, followed by the Ganges and Tagus. The dryness indices are almost similar for the Rhine and Lena basins. Figure 5 shows that uncertainty (described by the mean standard deviation of Q10 and Q90 change) propagated from both CMs and HMs generally increases under increasingly dry climatic conditions. This is particularly emphasised for RCP8.5. The increase of uncertainties under dry conditions is related to the quantification of actual evapotranspiration, which is a dominant flux in such environments. The HMs have applied different algorithms for potential and actual evapotranspiration to close the water balance (see Table 3 in Krysanova and Hattermann, this special issue), yet there is still incomplete understanding of how much water is lost via evaporation and/or transpiration to the atmosphere, and hence increased uncertainty is related to that..
Starting for high flows, in the wetter regions (Lena and Rhine basins), the propagated uncertainty from CMs and HMs is of similar magnitude (standard deviation of change is ~10 % in both model sets). However, CM uncertainty dominates the spread in the projections of discharge in the dry regions. In the Ganges basin under both RCP2.6 and 8.5, CM uncertainty is three times higher than the uncertainty propagated from HMs; the standard deviation of Q10 change due to climate modelling is ~28 % for RCP8.5. In the Niger basin, CM uncertainty is about five times higher than HM uncertainty for both RCP2.6 and 8.5; the standard deviation of Q10 change due to climate modelling is ~110 % for RCP8.5.
On the contrary, results for low flows generally show uncertainty from HMs being greater or close to CM uncertainty; this could also be related to the generally poor HM performance for low flows (see Huang et al., this special issue). Although both HM and CM uncertainties increase with increased emission scenario, the ratio between the HM and CM uncertainties seems to increase with decreasing dryness index for RCP8.5. Overall, although propagated uncertainty from CMs is higher (smaller or equal) than uncertainty from HMs for high (low) flows, both sources should be carefully assessed in climate change impacts studies. This needs additional caution in the regions with high dryness index due to increased uncertainty from both sources. See also an analysis of uncertainty propagation into drought characteristics by Samaniego et al. (this special issue).
Comparing with other recent studies
These new results from a consistent ensemble of regional HMs can be compared with both global scale and catchment scale studies over the same catchments; however, there exist very few comprehensive HM ensemble studies at the catchment scale. Global scale flooding studies using ensembles of HMs tend to analyse less frequent events than the high flow extremes considered here, but similar trends were shown by Dankers et al. (2014) and Hirabayashi et al. (2013) for all rivers considered here. On the other hand, Roudier et al. (2015) showed projected increases in the 10- and 100-year floods in the Tagus River which is opposite to the result shown here for Q10.
Using the ISI-MIP ensemble of global HMs, Prudhomme et al. (2014) showed similar changes to low flows in the Lena, Rhine and Tagus Rivers, and similar uncertainty of impacts in the Niger and Ganges Rivers. Roudier et al. (2015) showed similar decreases in low flows for the Rhine and Tagus Rivers. More importantly, Prudhomme et al. (2014) suggested that the uncertainty in the ISI-MIP ensemble of global HMs could be reduced by improving process representation in these models, e.g. by analysing the HM performance when forced by observed climate. In our view, this corresponds to what was done in this study applying the regional-scale calibrated HMs.
The catchment-scale models used in ISI-MIP2 attempt to minimise bias in the calibration period, potentially reducing the uncertainties of future projections. This is the key difference between the regionally or locally calibrated HMs and the global (or continental scale) HMs; the latter can show considerable bias when compared to observed discharge (see article by Hattermann et al. (this special issue) and Roudier et al. (2015)). Indeed, Hattermann et al. (this special issue) showed that for the global HMs having much weaker performance in the historical period compared to the regional-scale HMs, there is a much larger spread in climate change impacts, which may lead to the conclusion that the uncertainty from calibrated models is smaller, given the assumption that model performance in today’s climate is an indicator of the model’s ability to react to changing climate (Le Vine 2016). This last issue is, however, still contested.
Limitations of this study
The uncertainties shown in this study are the known uncertainties coming from the emissions scenario, CMs and HMs. Uncertainties arising from the bias-correction used to link the CM output to the HM input are not considered, and may themselves be of a magnitude comparable to the CM or HM uncertainty (Hagemann et al. 2011; Eisner et al. 2012). Furthermore, the climate projections have not been dynamically downscaled and the bias-correction has been used to both correct and downscale in this case. It is also impossible to fully sample the range of possible outcomes from feasible CMs or HMs. Nevertheless the range of CMs-related uncertainty is large and accounts for a significant proportion of the known uncertainties (Hempel et al. 2013). The HM ensemble includes a wide range of hydrological modeling types covering various levels of complexity; however variability in plausible parameterisations of each model have not been considered, which could further increase the HM uncertainty shown. Here, all ensemble members are considered equally plausible. Probabilistic prediction methods applying weighting of different ensemble members have been tested, particularly regarding the CM choice (Kjellström et al. 2010), concluding that this may be possible for smaller scale studies (for example a river basin). This was tested for a HM ensemble by Roudier et al. (2015), where poorer performing HMs for the indices studied (e.g. low flows) were omitted from the ensemble for those indices.