Global implications of 1.5 °C and 2 °C warmer worlds on extreme river flows

Targets agreed to in Paris in 2015 aim to limit global warming to ‘well below 2 °C and to pursue efforts to limit the temperature increase to 1.5 °C above pre-industrial levels’. Despite the far-reaching consequences of this multi-lateral climate change mitigation strategy, the implications for global river flows remain unclear. Here we estimate the impacts of 1.5 °C versus 2.0 °C mitigation scenarios on peak flows by using daily river flow data from a multi-model ensemble which follows the HAPPI Protocol (that is specifically designed to simulate these temperature targets). We find agreement between models with regard to changing risk of river flow extremes. Moreover, we find that the response at 2.0 °C is not a uniform extension of the response at 1.5°, suggesting a non-linear global response of peak flows to the two mitigation levels. Yet committing to the 2.0 °C warming target, rather than 1.5 °C, is projected to lead to an increase in the frequency of occurrence of extreme flows in several large catchments. In the most affected areas, predominantly in South Asia, while region-specific features such as aerosol loads may determine precipitation patterns, we estimate that under our 1.5 °C scenario the historical 1-in-100 year flow occurs with a frequency of 1-in-25 years. At 2.0 °C, similar increases are observed in several global regions. These shifts are also accompanied by changes in the duration of rainy seasons which influence the occurrence of high flows.


Introduction
The Conference of the Parties of the United Nations Framework Convention on Climate Change (UNFCCC) in its Paris Agreement in December 2015 agreed to hold 'the increase in the global average temperature to well below 2°C above pre-industrial levels and to pursue efforts to limit the temperature increase to 1.5°C above pre-industrial levels'. The simplicity of these targets led parties to the treaty to embrace a common climate strategy with the aim of preventing the risks and impacts associated with unabated climate change (Hulme 2016, Rogelj et al 2016, Schellnhuber et al 2016.
Yet the assumption that a 2°C limit is a safe bar as far as climate-change related tipping points are concerned, has diverted attention away from the difference in impacts between worlds in which global temperatures are stabilized at 1.5°C instead of 2.0°C. Overlooking these thresholds may have important consequences given the sensitivity of high flows to radiative forcings (Milly et al 2002). In particular, early studies projected an increase in annual runoff and flow peaks in snow-dominated catchments as a result of climate change (Nijssen et al 2001). From here, several studies have projected increases in high flow and flood frequency in Southeast Asia and eastern Africa, although important shifts have also been calculated for other regions such as the northern Andes, North America, and eastern Europe (Okazaki 2012, Hirabayashi et al 2013, Dankers et al 2014, Koirala et al 2014. As such, the impacts of extreme river flows on human lives, socio-economic development, and monetary losses may be particularly important in a changing climate Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence.
Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. (Peduzzi et al 2009, Wake 2013, Arnell and Lloyd-Hughes 2014, Jongman et al 2014, Winsemius et al 2015, Arnell and Gosling 2016. More recently, by downscaling climate projections and finding the corresponding year in which various levels of atmospheric warming are exceeded, a positive increase in future flood risk has been projected (Alfieri et al 2017). While the future increment in flood risks is important when compared to its historical baseline, the impacts appear to be more important at 4.0°C if compared to the less evident shifts observed between 1.5°C and 2.0°C.
Yet, since climatological biases in current climate models experiments, such as the Coupled Model Intercomparison Project, remain of the order of 0.5°C these efforts may not be sufficient to differentiate the risk associated with these two levels of warming (Hulme 2016, Mitchell et al 2016. For the hydrological cycle, in particular, the distribution of global precipitation between CMIP scenario experiments does not respond uniformly between CMIP scenario experiments due to the changing role of non-CO 2 forcing over the 21st century and also due to the precipitation sensitivity to emission-scenarios. This, in turn, makes it difficult to differentiate whether detected differences result from additional warming or additional factors. So, extracting anomalies for 1.5°C and 2.0°C levels of warming from the traditional CMIP scenarios driven experiments may not be scientifically robust (Mitchell et al 2016).
With such concerns in mind, the HAPPI protocol was specifically designed to simulate the specified Paris Agreement temperature targets (1.5°C and 2.0°C) as precisely as possible by separating the impact of an additional approximately half degree of warming from uncertainty in climate model responses and internal climate variability (Mitchell et al 2017). So, unlike other multi-model exercises, HAPPI does not exclusively rely on prescribed radiative forcing protocols to project the future temperature evolution, but robustly constrains it using prescribed future sea surface temperatures (SSTs) as well. This effectively reduces the influence of model sensitivity and results in a narrowly defined range of future temperature targets. Thus, the HAPPI Protocol provides an improved framework for the analysis of the impacts of an additional half degree of warming.
In this study, we estimate the extent to which global peak river flows seen under a 1.5°C scenario differ from those under a 2°C scenario. In particular, we investigate where shifts in high flow occurrence would be more important under these climate scenarios. We follow the HAPPI experimental protocol, to use the multi-model-ensemble daily output of four atmosphere-only general circulation models (AGCMs).

HAPPI experiments and data
The HAPPI protocol provides three 10 year simulation periods with prescribed atmospheric forcing, seasurface temperature and sea-ice coverage. The three scenarios are: (1) the reference or historical period which is the 'current decade' from 2006 to 2015, (2) a future decade that is about 1.5°C warmer than preindustrial levels, and (3) a future decade that is about 2.0°C warmer than pre-industrial levels. In this protocol, the reference experiment chosen is 2006-2015 because it is the most recently observed period. This period also contains a range of different SST patterns over the decade, allowing for an assessment of how the ocean conditions vary on interannual timescales, including important El Niño and La Niña events.
For each scenario, we used the output of four HAPPI AGCMS: CanAM4 (100 ensemble members); CAM4-2degree (100 ensemble members); NorESM1-HAPPI (125 ensemble members), and MIROC5 (50 ensemble members). Each simulation (ensemble member) within an experiment differs from the others in its initial weather state. So, the use of 50-125 tenyear time slices really provides 500 years of data for the MIROC5 experiment, 1000 years for CanAM4 and CAM4-2degree; and 1250 years for NorESM1-HAPPI. Such extensive record (500-1250 years) in turn provides the basis for robust calculations. More details of the HAPPI protocol is found in the supplementary information is available online at stacks.iop.org/ERL/ 13/094003/mmedia.

Runoff
The AGCMs used in this study have a land surface scheme which is coupled to the overlying atmosphere. For two models, NCC/NorESM1-HAPPI and ETH/ CAM4-2degree, runoff was stored in the HAPPI archive. The other two models (MIROC/MIROC5 and CCCma/CanAM4) did not have a runoff-production scheme activated for use in the present experiment and so we applied a comparable runoffgeneration scheme to calculate runoff in those cases. This in turn enabled us to use those models alongside the members of the ensemble for which runoff data were available.
So, for these two AGCMs we calculated runoff by using a simple runoff production model designed to be comparable with the runoff-production models typically embedded within climate models. Our scheme uses a Rutter-Gash canopy formulation (Gash 1979) together with Penman-Monteith evaporation calculated using available radiation data (Monteith 1965). Soil moisture was accounted for using a two-layer model with saturation-excess runoff computed using a generalized TOPMODEL (Clark and Gedney 2008).
The snowpack model used here was based on a temperature-based model of accumulation and melt (Moore et al 1999, Hock 2003, Beven 2011. Snow accumulates when precipitation falls while temperature is below a threshold temperature T a . When temperature is above a threshold for melt, T m melting occurs at a rate proportional to the difference between the current temperature and T m . This conceptual model is widely used (Hock 2003, Zhang et al 2006, Rango and Martinec 2007, Beven 2011) and gives performance comparable with that of more parameterrich energy balance models, despite their greater complexity (Parajka et al 2010).
We acknowledge the existence of other more sophisticated runoff schemes yet ours is intended to be applied directly to atmospheric outputs resulted from large ensembles such as the HAPPI ensemble allowing us to maximize the utility of the HAPPI ensemble by generating runoff when it was not provided directly.

Runoff routing
Total runoff derived from NCC/NorESM1-HAPPI and ETH/CAM4 models or calculated from CCCma/ CanAM4 and MIROC/MIROC5 were routed by implementing the grid-based hydrological routing scheme as presented by Dadson et al (2011). This scheme is based on the discrete approximation to the 1D kinematic wave equation with lateral inflow (Oki and Sud 1998). This routing model routes flows along the path of steepest descent. Resistance to flow is handled via a wave celerity parameter. This algorithm has been widely used in various large-scale hydrological studies as well as various existing climate and land surface model configurations. Further details are available in Bell et al (2007). We also selected this scheme as it is computationally tractable when dealing with a global problem with many hundreds of ensemble members, as with the HAPPI experiment.
Also, in this scheme the river network data used was constructed by using the network tracing method (NTM) which is a vector-based network scheme that traces the path of river networks downstream (Olivera and Raina 2003). This method works by overlaying a mesh of a certain grid size over a fine-scale joined river network to determine the coarse-scale downstream cell. This method has the advantage that it can be used in areas where digitized river networks exist but no DTM is available and it is usually spatially closer to the base river network. As NTM has been found to outperform other raster-based methods, it has been widely used in previous routing studies (Olivera and Raina 2003, Bell et al 2007, Davies and Bell 2009). The resolution of the river network used here is half degree (0.5°×0.5°).
Also, the spatial resolution of runoffs given by each AGCM (NCC/NorESM1-HAPPI: 1.875°×0.625°; ETH/CAM4: 2.5°×1.875°, MIROC/MIROC5: 1.40°×1.40°, CCCma/CanAM4: 2.81°×2.81°) was downscaled to match the resolution of our river network. The downscaling was performed by applying a bilinear interpolation technique. Apart from its simplicity, this technique has been found to provide a more realistic spatial gradient instead of patches of same runoff values from an AGCM in cells with finer resolution ( The final results of our study are presented at the scale of the routing model (0.5°×0.5°). We acknowledge that this increase in the spatial representation of runoff may result in our overlooking or underrepresenting runoff processes that operate at finer scales, which are represented by using statistical distributions of sub-grid runoff generation. We note that our routing scheme does not account for human interventions and flow regulations such as dams and reservoirs, or changes in land use and river interventions. We acknowledge the importance of these anthropogenic interventions, but they are not the subject of our study. Models that incorporate complex anthropogenic processes do not typically offer computational simplicity appropriate for use with large ensembles. The comparison of derived river flows with observations derived from GRDC can be found in supplementary figure 5.

River flow extreme statistics
The change in high flow hazards between the historical and future conditions (1.5°C and 2°C warming scenarios) was calculated from the probability of the historical 100 year return period river flow magnitude to be exceeded at the 1.5°C and 2°C warming scenarios. We selected the 100 year return period as the reference river flow in order to keep consistency with previous hydrological studies ( Dankers and Feyen 2008, Hirabayashi et al 2008, Ward et al 2013. So, time series (500-1250 years) of simulated annual maximum daily river flows in the historical scenario (2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) were fitted to a two-parameter Gumbel distribution (Gumbel 1941). As a result, the magnitude of the 100 year return period river flow of the historical scenario was calculated at each grid and for each AGCM. Lastly, the return period of the historical 100 year river flow was calculated for the time series of the 1.5°C and 2°C warming scenarios.
A two-parameter Gumbel distribution was selected as it provides relatively stable distribution parameters showing adequate results in similar previous studies (Hosking and Wallis 2005, Dankers and Feyen 2008, Hirabayashi et al 2008, Hirabayashi et al 2013. Also, while these previous studies acknowledge that the type of extreme distribution function selected may impact the probabilities of extreme flows, it does not affect the estimation changes in the frequency and tendency of them, which is the aim of this study. Similarly, Hirabayashi et al (2013) used the probability plot correlation coefficient test to find that globally 76% of the AGCMs adequately fit to the Gumbel distribution (coefficient>0.96 at 95% of significance). Grids that did not fit this distribution were mostly found in arid regions which has also been found in similar studies (Hirabayashi et al 2008).

Results
We detect regions where the shift in occurrence of global historical extreme flows at 1.5°C and at 2.0°C differ as calculated by our ensemble of four AGCMs (figure 1). First, we identify regions that show an initial increase in the frequency of occurrence of historical high flows at 1.5°C and then progressively increases at 2.0°C. These regions include South America (Amazon catchment), central Africa, central-western Europe, the south of the US (Mississippi river area), central Asia, and Siberian catchments. In these regions, the median increase of the frequency of the current 1-in-100 year flow is once in 70-90 years at 1.5°C, and at 2.0°C this frequency increases to at least once in 50 years. Moreover, we find that in most of central and eastern China, the southern part of the Amazon catchment, the Blue Nile, and in northern India, the 1-in-100 year flow at 1.5°C ranges between 1-in-50 and 1-in-60 years, whereas at 2°C such an event ranges between 1-in-25 and 1-in-35 years.
Second, we identify areas where the frequency of the 1-in-100 year flow increases at 1.5°C but does not change appreciably between 1.5°C and 2.0°C. Notably these areas are located in South and Southeast Asia (south of the Yangtze river) and the Indus river basin. In these regions under a 1.5°C scenario, the current 1-in-100 year flow increases in frequency to 1-in-50 year return period. However, we also identify some areas, such as Southeast Asia, where the current 100 year flood occurs 1-in-25 years in the future. In North America and eastern Europe, the current 100 year return period flow increases in frequency only slightly, to once in 80 years at 1.5°C.
We also find some regions which see a decrease in the frequency of high flows between the present day and 1.5°C or 2.0°C or both. These regions are located in the northern part of South America, areas of the western coast of the United States (Colorado River) and Canada, South Africa (Orange River), Scandinavia, and in some parts of eastern Europe. In most of these regions, we note that the current 1-in-100 year flow decreases in frequency to approximately 1-in-150 years, with little further decrease at 2.0°C. Lastly, we find that in the area of the Murray-Darling basin in Australia, there is a shift from historical high flows not changing in a world 1.5°C warmer (projected return periods between 90 and 125 years), whilst under a 2.0°C scenario the return period decreases to approximately 1-in-70 years.

Physical characterization of high flows
In order to understand the physical reasons that induced shifts in high flow frequencies, we calculated the change in both the number of frost days and the number of days where precipitation has occurred between the two scenarios in comparison to the historical baseline (see figure 2). We find an important decrease in the number of frost days at 1.5°C which slightly increases at 2.0°C. This decrease is greater in North America, and central and eastern Europe (figures 2(a), (b)). This finding suggests that globally the role of snowmelt in the generation of high flows will decrease as snow accumulation reduces. Exceptions to this pattern are found in the South and East Coast of the US and in the south of South America. In these areas, particularly in the US, where we estimate an increase in the occurrence of high flows, snow dynamics and possibly longer periods of snow accumulation may be important drivers of high flow events.
The change in the number of days with precipitation between the two climate thresholds (figures 2(c), (d)) shows a diverse regional response. For example, in most of the regions where we estimate a progressive increase in the frequency of high flows between 1.5°C and 2.0°C, such as central-western Europe or South America, the number of days with precipitation typically decreases or does not importantly change at 1.5°C and also at 2.0°C. This is also accompanied by a decrease in average annual precipitation magnitudes at both scenarios (see supplementary figure 2). This result suggests that in these areas extreme precipitation, will likely dominate the occurrence of high flows, despite the reduction in the number of rain days. An important exception is observed in Siberia, where we see an increase in the frequency of high flows accompanied by a slight decrease in the number of days with precipitation.
In regions where the increase in high flow occurrences is important at 1.5°C but minor at 2.0°C, such as the south of Asia, we find that, at this first temperature target, the frequency of rain days decreases as well as total annual precipitation magnitudes. Yet at 2.0°C days with precipitation are either more common (especially in North America) or show a lower reduction compared with that detected at 1.5°C. Similarly, annual precipitation magnitudes show a slight increase when compared to the historical reference period. This finding suggests that, in these regions, the role of shorter wet seasons and storms in high flows, which at 1.5°C seems to be more important, will decrease at 2.0°C. So, in these regions, the difference between the two climate targets, and thus flow occurrences, may be associated with shifts in durations and characteristics of the wet season. In the regions where we observe a decrease in the occurrence of high flows, we also detect a decrease in the number of days with precipitation and annual precipitation magnitudes.

Multi-model consistency and biases in results
The change in high flow frequency at 1.5°C and 2.0°C, although consistent in sign between models across most parts of the globe, differs in specific regions (see figure 3). Globally, we find that 49% of land grid cells at 1.5°C and 43% at 2.0°C show a high consistency in the tendencies described above (consistency defined as an agreement in trajectory of three out of four models). This high level of consistency is seen particularly in regions where we find an increase in high flow frequency, such as the south of Asia, India, or East Africa. Conversely, in just about 7% of land grid cell at 1.5°C and 4% at 2.0°C just one model agrees with the tendency mapped above. Typical regions with lower consistencies can be found in northern Russia and parts of the tropics. In these areas our projections are subject to an important spread across the AGCM ensemble.
Additionally, the examination of the way each AGCM simulates the changes in high frequency helps to disseminate further model biases (supplementary figure 1). For instance, NCC/NorESM1-HAPPI generally projects greater increases in the frequency of occurrence of the reference high flow in the tropics. Thus, it simulates a more pronounced increase in high flow from the historical reference return time of 1-in-100 years to approximately 1-in-40 years at 2.0°C for larger areas (e.g., central South America and South Asia). Also, MIROC/MIROC5 and CCCma/CanAM4 show greater increases in reference extreme flow frequencies in northern regions. We also find that ETH/ CAM4-2degree generally produces lower river flows outside the tropics.
Moreover, by examining the way each AGCM represents precipitation and then runoff (supplementary figure 3 for precipitation and 4 for runoff), we estimate whether existing biases (such as the lower river flow in ETH/CAM4-2degree) are due to deficiencies in the AGCM, runoff generated, or the river routing model. Using zonal means, we find that each of the four AGCMs slightly overestimates mean precipitation in the tropics compared with the Global Precipitation Climatology Center dataset, GPCC (Becker et al 2011) (maximum precipitation bias ∼30%).
This tropical overestimation persists for mean runoff in MIROC/MIROC5 and, to a lesser extent, in CCCma/CanAM4 when compared with observed river flow data obtained from the World Meteorological Organization Global Runoff Data Center, GRDC (Fekete et al 2002). ETH/CAM4-2degree and NCC/NorESM1-HAPPI on the other hand, show an importantly reduced runoff in the tropics compared with the other two AGCMs. In contrast, extra-tropical overestimation of precipitation in MIROC/MIROC5 does not lead to equally overestimated runoff in those regions. It is therefore not possible to identify the root cause for biases in river flow with confidence, given the substantial discrepancies in the precipitation-to-runoff ratio in the models. However, the fact that the two models with lowest zonal average runoff also show the lowest river flow is indicative that the treatment of the water budget and the associated hydrological balance in the AGCMs is crucial in capturing the magnitude of the river flow. That said, we emphasize that biases in river flow are unlikely to affect the change in frequency of the return time between the different temperature scenarios. This is also shown in our sampling uncertainty estimates derived from resampling our dataset 1000 times for the ten largest catchments globally (see table 1).

Magnitude of projected changes in the 100 year river flow
In contrast to the important increase in high flow occurrences, described above, we lastly find a diverse response in the change of flow magnitudes between the two temperature scenarios for the ten largest catchments in the world (see table 1). In catchments where we do not detect a large change in the frequency of occurrence of the historical 1-in-100 years return period flow such as the Amazon and Murray-Darling, we find a reduction of up to 5% in mean annual river flow under the 1.5°C scenario. In contrast, some mean river flows such as those at the outlets of northern catchments (Mackenzie and Ob') show important increases of up to 8% (at 2.0°C) despite quasiconstant changes in frequency of occurrence of peak flows. At the Yangtze river, apart from the increase in the frequency of occurrence of extreme river-flows as presented above, there is an important increase in the mean discharge at both, 1.5°C-2.0°C. Also, we note that for the Mississippi and the Ganges rivers the small change in mean river flow is contrasted by an important increase in high flow frequencies. Also, we note that regarding flow magnitude, the change from 1.5°C to 2.0°C levels of warming are not as clear as those detected for high river flow frequencies. However, we note that the combined fidelity of the Table 1. Projected change (expressed as return period) of the 1-in-100 year historical flow at both scenarios and estimated change in the mean river flow (%) from ten selected major global outlets. Uncertainty due to sample size is calculated by a bootstrap method. Uncertainty for the four AGCMs was averaged. Standard deviation is calculated from our ensemble of AGCMs.

Return period of historical high flow
Change in mean historical river flow (%) ensemble of our AGCMs and the river routing model to reproduce the annual cycle of the river flow, exemplified for a subset of major river catchments is limited in certain areas (see supplementary figure 5). The representation of the water budget and hydrological balance discussed in the section above may also be affecting the ability to represent changes in flow magnitudes.

Flow responses for alternative return periods
We note that although the four AGCMs ensemble disagree in maximum daily flow magnitudes, they are not sensitive to the choice of the return time as highlighted in figure 4, where we have plotted the return times for all scenarios and models for four different rivers which are affecting a large fraction of global population. At the Yangtze as well as at the Orange outlets, we find that the difference between 1.5°C and 2.0°C remains small for a wide range of return times (figures 4(a)/(d)). As mentioned above, the four models show similar tendencies when simulating changes in return periods of high flows for the selected regions, which suggests that our flow frequency results are indeed robust, despite biases in the magnitude of the river flow. The detected wet bias in runoff at certain latitudes in MIROC/MIROC5 reappears here, in that we find the river flow magnitude to be overestimated at the Yangtze and the Ganges (figures 4(a), (b)). Also, the difference in runoff between the two wettest and driest AGCMs is mirrored in the return time plots at the Upper Blue Nile (figure 4(c)), where we find considerably higher river flow magnitudes between the two model subsets. In other words, the regionally-overestimated precipitation (∼30% in NCC/NorESM1-HAPPI and CCCma/ CanAM4) translates directly into higher mean and extreme river flows.

Discussion
The results shown here agree with various of the previous findings that also project shifts in runoff,  suggest a general decrease in future high flows in central-eastern Europe. On the contrary, our results project an increase in high flow frequencies in this region (relatively good multi-model consistencies found here). Also, the nonlinearity found between current conditions and the two future warming scenarios might as well be related to counterbalancing effects due to the presence of different forcing agents that vary greatly over time. In both HAPPI future scenarios, the aerosol load is reduced to approximately one third of its current levels. This leads many regions to show higher rainfall increases at 1.5°C compared to what has been observed or simulated in similar experiments as high aerosol loads have been connected to a reduced strength of the hydrological cycle in several global regions (Bollasina et al 2011). This may be the case in the Asian monsoon region which is generally characterized with high aerosol loads and our results show an increase in high flow frequencies detected here. If the role these different drivers are playing is not appropriately analyzed, the attribution results can easily be misinterpreted. As such, it is also important to understand other region-specific features which may explain the different local flow responses between both scenarios.
Moreover, the consistent decrease in high flow frequencies found in our study for Scandinavia and South Africa has been previously estimated by Hirabayashi et al (2013) and Koirala et al (2014) for the RCP8.5 scenario. Similar projections were also found by Döll et al (2018) for the particular temperature targets agreed in Paris. Also, it is important to note that in some areas where consistently project an increase in high flow frequencies, various previous reports estimate a future decrease in mean river flow accompanied by major drought hazard at 2.0°C (Schewe et al 2014, Lehner et al 2017. These areas include, southern China (such as the Yangtze river), central Europe, and central South America. Thus, our findings suggest an intensification of hydrological extremes in these regions. Yet it is important to consider the additional role that that the reduction of aerosols, described previously, may have in showing wetter futures in our estimates in these regions.

Conclusions
By using the HAPPI Protocol, which is a framework designed to reduce the influence of model sensitivity in climate outputs, we analyze the impacts of an additional half degree of warming on high river flows, as agreed in Paris in 2015. We find that historical extreme flows (represented by 1-in-100 year return periods) occur more frequently at 1.5°C, but even greater additional changes are projected for many regions at 2.0°C. These regions include most of South America, central Africa, central-western Europe, the south of the US (Mississippi river area), central Asia, and Siberian catchments. We detect that in these regions, the median increase of the frequency of the current 1-in-100 year flow is once in 70-90 years at 1.5°C, and at 2.0°C this frequency increases to at least once in 50 years. Our result suggests that in these areas (with the exception of Siberia), this change is accompanied by shorter, and consequently more intense, rainy seasons. As such, committing to a 1.5°C level of warming reduces the chance of high flows in these regions. In other highly populated areas, such as South and Southeast Asia (south of the Yangtze river) and the Indus river basin, the current 1-in-100 year becomes more frequent at a 1.5°C level of warming (1in-50 year). Then the change from 1.5°C to 2.0°C is not importantly appreciable. Yet it is important to note that reductions in atmospheric aerosol loading represented in the HAPPI experiments lead to wetter futures in the areas discussed. As such, the implications that regional-scale processes may have in our results may be worth noting in a context of global mitigation targets. Lastly, while our analysis highlights limitations concerning regional and seasonal river flow projections produced using global models, we have demonstrated that our river flow results are robust with regard to changing risks tendencies. Thus, our results may be used to inform adaptation strategies and to guide assessments of socio-economic impacts at the regional or catchment scale.