Toward evaluating the effect of climate change on investments in the water resources sector: insights from the forecast and analysis of hydrological indicators in developing countries

The World Bank has recently developed a method to evaluate the effects of climate change on six hydrological indicators across 8951 basins of the world. The indicators are designed for decision-makers and stakeholders to consider climate risk when planning water resources and related infrastructure investments. Analysis of these hydrological indicators shows that, on average, mean annual runoff will decline in southern Europe; most of Africa; and in southern North America and most of Central and South America. Mean reference crop water deficit, on the other hand, combines temperature and precipitation and is anticipated to increase in nearly all locations globally due to rising global temperatures, with the most dramatic increases projected to occur in southern Europe, southeastern Asia, and parts of South America. These results suggest overall guidance on which regions to focus water infrastructure solutions that could address future runoff flow uncertainty. Most important, we find that uncertainty in projections of mean annual runoff and high runoff events is higher in poorer countries, and increases over time. Uncertainty increases over time for all income categories, but basins in the lower and lower-middle income categories are forecast to experience dramatically higher increases in uncertainty relative to those in the upper-middle and upper income categories. The enhanced understanding of the uncertainty of climate projections for the water sector that this work provides strongly support the adoption of rigorous approaches to infrastructure design under uncertainty, as well as design that incorporates a high degree of flexibility, in response to both risk of damage and opportunity to exploit water supply ‘windfalls’ that might result, but would require smart infrastructure investments to manage to the greatest benefit.


Introduction
Major infrastructure investments in virtually any sector require rigorous economic/financial analysis to ensure that expected returns justify investment, and that key risks that might jeopardize those returns are fully evaluated. Failure to consider changes in future climate impacts risks reliance on a faulty time series of future returns, but with most economic analysis incorporating discount rates on the order of 7-10%, decision making is heavily influenced by the net monetary flows of the first two decades-too short to reflect most effects of a changing climate. In this time perspective other variables are much more important. However, a bigger issue, in particular in developing countries, is whether water resources infrastructure investments that look economically attractive today are consistent with the best long-term development path. National governments and international financial institutions should consider, for example, whether a large multipurpose dam, with attendant irrigated agriculture, electricity dependent industry and related settlement patterns is sustainable in the face of long-term water challenges. Practical examples include options to invest in irrigation infrastructure in the Okavango basin in Botswana (World Bank 2010a); options to invest in high value irrigated agricultural production in parts of the Balkans, Central Asia, and the Southern Caucasus (Sutton et al 2013); and proposed hydropower investments in northern and western sub-basins of the Zambezi River basin in southern Africa (World Bank 2010b). In developed country contexts alternatives to large-scale infrastructure investments may be reasonable substitutes for infrastructure (e.g., water efficiency, input substitution, and other non-infrastructure related changes might be employed to maintain service levels). Nonetheless, while such alternatives may also play a role in developing country contexts, the general under-investment in large-scale infrastructure here (see Foster and Briceño-Garmendia 2010) suggests that long-lived infrastructure investments should continue to be proposed and thus require more rigorous analysis.
The best analyses of large-scale infrastructure include consideration of future climates and sensitivity analyses, but they are typically not tied to the specific, internally consistent scenarios of future precipitation and temperature changes that have been developed for climate change assessments (IPCC 2007), do not incorporate the full range of changes that could be associated with future climates and in particular do not adequately take into consideration the uncertainty with respect to future climates which is indicated by the full suite of climate models and emission scenarios of IPCC (2007) (e.g., Kuik et al 2008, Kirshen et al 2008, Ward et al 2010. It is now clear that the wide range of potential future climate and hydrologic outcomes suggest the use of planning tools such as robust decision-making (Lempert and Groves 2010), which focus on resilience to uncertain futures rather than optimization in relation to predicted futures and on methods of decision making for large-scale infrastructure that put a very high value on flexibility (De Neufville and Scholtes 2011).
In response to this growing need to evaluate the climate resilience of proposed development paths and related infrastructure investments, the World Bank has recently developed a method to evaluate the effects of climate change on six hydrological indicators across 8951 basins of the world. The indicators are designed for decision-makers and stakeholders to consider climate risk when planning water resources and related infrastructure-here we refer to risk as the product of severity (the magnitude of change) and frequency (the likelihood of change). These indicators reflect impacts of climate change (severity) on irrigation and drainage, large water supply and urban wastewater treatment, small water supply and rural wastewater treatment, flood protection, and river basin management and multipurpose infrastructure. To fully understand climate change as a risk factor, however, we are limited by an inability to attribute reliably the frequency (or probability) of alternative projections of climate change. The next best solution is to provide a representation of the breadth of future change across many plausible predictions of future climate. To accomplish this goal, the analysis examines relative changes from an historical baseline to three future periods for 56 GCM-SRES combinations available from the IPCC Fourth Assessment (IPCC 2007), enabling users to employ a risk-based approach to the effect of climate on investment plans. As described here, the results provide insights into key water resources challenges likely to arise in developing regions, including the prospect of much larger variability in key hydrological indicators in the poor countries least able to manage those risks.

Methods
Developing projections of hydrological indicators for 8951 world river basins under a wide range of possible future climate conditions presents challenges in characterizing baseline conditions (including the unit of analysis), projecting key climate variables, and developing hydrological indicators at the basin level. We review our methods for each of these three steps below.

Characterizing baseline conditions
The focus of this study is water resources planning and development at the regional and local level, and as such, the river basin was identified as the appropriate scale for this analysis. A key challenge then is determining an appropriate global definition for river basins. We rely on the USGS HydroSHEDS global basin definitions, based on a 1 km digital elevation model. We chose a combination of Level 3 and Level 4 basins from HydroSHEDS, in an attempt to roughly match basin size to the size of a typical GCM gridbox, in order to ensure the results were not over-specified relative to the scale of GCM results. Nonetheless, the Level 3 and 4 basins defined in this study vary significantly in size, ranging from approximately 2500 km 2 , which is similar to a grid cell of 0.5 • × 0.5 • , to more than 62 500 km 2 , which is similar to a grid cell of 2.5 • × 2.5 • . For climate data, we rely on a 30-year historical baseline , with the goal of projecting to future 30-year periods centered on three future eras: the 2030s, 2050s, and 2080s. Baseline precipitation and temperature data for the 1961-1990 baseline was taken from the University of East Anglia's Climate Research Unit (CRU) TS 2.1 data set, which provides monthly data at a 0.5 • × 0.5 • resolution.

Projecting key climate variables from a suite of GCM/SRES combinations
Projecting changes in climate variables from GCM simulations has often involved downscaling approaches, but both statistical and dynamical downscaling have well-studied uncertainties (Kerr 2011), and the time and costs of these computationally intensive approaches rarely allow the use of more than a few GCMs. Our goal in this work is to characterize the broadest possible range of 'not implausible' climate futures, as defined by the currently available set of GCM-SRES combinations. The only practical approach for a global analysis is to use projected changes in temperature and precipitation for 56 GCM-SRES combinations at their native resolutions. These native resolution changes were mapped onto a 0.5 • × 0.5 • grid, and then combined with the corresponding 0.5 • × 0.5 • grid of CRU baseline modeled data. Basin-scale aggregation was then achieved using GIS software to overlay basin boundaries with the 0.5 • × 0.5 • grids, and then aggregating cells based upon their weighted area in each basin. This approach was designed to capture the range of potential climate change impacts at a higher resolution without downscaling the GCMs themselves, thereby achieving a balance between precision and accuracy.
Note that the 56 climate projections represent the full range of available models for the B1, A1B, and A2 Special Report on Emissions Scenario (SRES) scenarios evaluated in the Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report (AR4). There are 17, 22, and 17 GCM runs, respectively, available for the three emissions scenarios, leaving a total of 56 GCM-SRES combinations. These three SRES scenarios were chosen because they are generally in the middle range of the marker SRES scenarios identified by the IPCC, and are the most commonly used emissions scenarios for impact and adaptation assessments.
To compare across GCMs, we converted GCM modeled baselines and projections into decadal average monthly changes relative to the model baseline by subtracting the modeled baseline from the projected values to produce delta temperature and precipitation derived from the archived CMIP3 IPCC AR4 outputs. For each GCM-SRES combination, these relative changes for the decades of the 2030s, 2050s, and 2080s, were then coupled with the 30 year CRU historical dataset to generate three 30 years absolute monthly projections representative of potential future conditions in decades of the 2030s, 2050s, and 2080s.

Translate trends from climate models into hydrological indicators
Basin-scale runoff is a key component of the six hydrological indicators. To model changes in runoff, this study employed CLIRUN-II: a hydro-climatic modeling framework with components that model, PET, Snow Water Balance, and soil moisture. Potential evapo-transpiration (PET) is a necessary input into runoff modeling as well as irrigation water requirements. CLIRUN-II uses the Modified Hargreaves method (Allen et al 1998, Droogers andAllen 2002).
The runoff modeling component is a two-layer, onedimensional conceptual rainfall-runoff model that simulates natural runoff with six calibration parameters (Strzepek and McCluskey 2010). This class of model requires natural runoff data to calibrate the model over an historic period.
While global databases of gauged flow are available (e.g., WMO 2012) there is no corresponding database of natural flows to use in assessing the performance of this procedure at global scale. McMahon et al (2007) are developing a global natural flow database based on statistical characteristics of natural flow and recreating natural flows from gauged flow, but this effort is limited in scope and not appropriate for our application. Hydrologists have taken an alternative approach using global gridded databases of climate time series and using hydrologic models to simulate natural flows. The Global Runoff Data Centre (GRDC) has developed a composite runoff database that combines simulated water balance model runoff estimates with monitored river discharge (Fekete et al 2002). This data set consists of average monthly runoff values for each cell at a 0.5 • × 0.5 • global land grid. We calibrated the model by minimizing the squared deviation between the 12 monthly GRDC runoff values and the 12 monthly averaged CLIRUN-II model outputs from the 10-year simulation period, which was chosen to best represent the decade used to generate the 12 months of GRDC runoff data. The limitations of using a modeled 'natural' runoff for calibration and having only monthly average values add uncertainty to the results. Other issues with the GRDC data that add to uncertainty in the analysis include: (1) there are large areas (especially in dry regions) that do not have gauge data, (2) the time period of available gauge data varies by station, therefore the resulting monthly discharge regimes are not fully consistent, (3) the historical climate data used in the water balance model (WBM) of the GRDC data set is not the same that was used in the CLIRUN-II model analysis, and (4) the data set is only provided for 12 average monthly values, not for a full time series. Additional uncertainty also exists in the choice of CLIRUN and its model uncertainty. Based on multi-model assessments, Haddeland et al (2011) and Schewe et al (2013) report that differences between hydrological model results are also a major source of uncertainty.
CLIRUN-II produces a 30-year time series of monthly hydro-climatic variables that are used in calculation of six hydrologic indicators 4 : (1) mean annual runoff (MAR); (2) river basin yield; (3) annual high flow (q10), or 10%  (q90), or the 90% exceedence flow; (5) baseflow or the sustained flow in a river basin resulting from groundwater runoff; and (6) reference crop water deficit, which is the crop water demand less available precipitation.
As crop modeling and analysis of agricultural water use at the global basin scale were well beyond the scope of this work, we employ a simplified version of the water deficit index approach (Woli et al 2008) to estimate reference crop water deficit. For a given basin-specific growing season, this formulation reduces to the sum of monthly PET minus precipitation for those months in which PET exceeds precipitation. For a more detailed investigations of the impact of climate change on irrigation water demand for a range of GCMs, see Konzmann et al (2013).

Results
The result of our analysis is a dataset that provides six hydrological indicators for over 8000 basins worldwide, for up to 56 alternative climate futures. The methodology and data set has been utilized by the World Bank in a number of cases for example for a policy note on adaptation options in Botswana (World Bank 2010a), for a policy note on adaptation options for the Sava River basin and for a multi-sector investment opportunity analysis in the Zambezi River basin (World Bank 2010b). A dataset of this size could easily overwhelm users, so the data also includes a user-friendly interface that allows for analysis at the country and regional level, with mapping products and statistical representations of output, such as box and whisker diagrams. The full data set and interface can be accessed at the World Bank Climate Knowledge Portal, by pointing on a map 5 . In this section, then, we first provide a summary overview of our global results, and then outline three observations from our analysis of the results.

Overview of GCM ensemble mean results
Figure 1 provides an overview of the mean changes in MAR and reference crop water deficit from the baseline to the 5 See for example, the following: http://sdwebx.worldbank. org/climateportal/index.cfm?page=country impacts water& ThisRegion=Africa&ThisCcode=KEN 2050s across the GCMs run for the A2 SRES scenario-the A2 scenario was chosen for presentation because it was also used in the World Bank Economics of Climate Change study (World Bank 2009). Regionally, model results suggest that, on average, MAR will decline in southern Europe; most of Africa; and in southern North America and most of Central and South America. Asia, most of North America, and the Pacific Islands are projected to experience increases in water availability. These general patterns hold for the q10 and q90 indicators as well. Mean reference crop water deficit, on the other hand, combines temperature and precipitation and is anticipated to increase in nearly all locations globally due to rising global temperatures. The most dramatic increases in crop water deficit are projected to occur in southern Europe, southeastern Asia, and parts of South America.
As part of our evaluation of these mean results, we compared our MAR projection to those from another recent analysis (Milly et al 2005). Figure 2 compares 2050s MAR projections of the current study to Milly et al (2005), each using the same set of the GCMs under the A1B SRES scenario. Although the results differ in several locations such as parts of South America and Australia, the general pattern is very similar globally.
3.1.1. Observation 1: hydrological indicators show a clear regional pattern that intensifies and grows less certain over time. For each of the World Bank regions, figure 3 provides boxplots of per cent changes in MAR from baseline to the 2030s, 2050s, and 2080s across the 17 A2 GCMs. The World Bank region results are population-weighted averages of basin-level values, grouped into regions based on basin centroids. The clear regional trends in MAR become more pronounced and less certain over time, illustrating the widely different challenges in water resources planning in different parts of the world. For example, planning for the projected increases in MAR and q10 in the Europe and Central Asia region poses vastly different challenges for infrastructure development than planning for the anticipated reductions in MAR and q90 in the Middle East and North Africa (MENA). Our data suggest that these differences are much greater at the catchment level. It is important to note, however, that in some regions, the direction of change in MAR become more certain over time. For example, within the MENA region, changes in  MAR are both positive and negative in the 2030s, whereas by the 2080s, almost all models project a decrease.
These results, as presented, suggest overall guidance on regions in which to focus water infrastructure solutions that could address future runoff flow uncertainty. The full dataset is much richer, however; the country and basin-level results provide insights at a finer geographic scale, but remain consistent with the geographic scale of results from GCMs. Nonetheless, the indicators do not support project-level analyses. Concerns over whether a particular hydropower investment may face substantial reductions in future flow, for example, require a yet more detailed site-specific analysis that incorporates engineering considerations that could be adopted to adapt to changes in flow. In addition, because our results indicate that the full range of available GCMs span a wide range of hydrologic outcomes, they suggest that project-level analyses may require a new method of decision-making for water infrastructure that puts a very high value on flexibility (De Neufville and Scholtes 2011).
3.1.2. Observation 2: uncertainty in projections of MAR and high runoff events is higher in poorer countries, and increases over time.
Our results also suggest that lower-income countries will face greater uncertainty in future hydrological conditions, particularly mean annual runoff and 10% exceedence flows (q10). Figure 4 displays the inter-quartile range for each country of per cent changes in MAR from the baseline to the 2080s across the 17 A2 GCMs (at left), and boxplots of per cent changes in MAR from baseline to the 2080s across the A2 GCMs for countries within each income category (at right). Income region boxplots are population-weighted averages of basin-level values, grouped spatially based on basin centroids. World Bank per capita income categories include lower (<$1005); lower middle ($1006-$3975), upper middle ($3976-$12 275), and high (>$12 276). Although uncertainty increases over time for all income categories, basins in the lower and lower-middle income categories are forecast to experience dramatically higher increases in uncertainty relative to those in the upper-middle and upper income categories. Strzepek and Schlosser (2010) find similar results for 2050 and the A2 GCMs when analyzing climate change impacts on the Climate Moisture Index. Figure 5 displays the relationship between income and uncertainty in projected country-level MAR and q10. The figure plots per capita country income against the IQR of projected percentage changes across the 17 A2 GCMs for the basins in that country (aggregated based on population; the size of each marker corresponds to the population of each country). All trends are statistically significant (p < 0.001), and steepen over time. Note that both the larger and smaller population countries appear to follow these trends. This result is not surprising, as precipitation is much more variable in low income countries currently, but our work shows that trend will be exacerbated by climate change.  Although the observation 1 results indicates more water runoff in general, the Observation 2 results suggests more uncertainty about the amount, and in particular for poorer countries, who are least prepared to manage uncertainty for reasons related to information, institutions, and infrastructure. First, poor countries have less knowledge about current and future climate. Second, poor countries seldom have the regulatory and institutional capacity (including the capacity for cross sectoral collaboration) to deal with uncertainty. (WMO 2013, Sivakumar et al 2011. Third, poorer countries often (though with many notable exceptions in regard to water storage infrastructure) have less water infrastructure, an investment which can serve as an effective response to uncertainty. In policy terms, then, this result underscores the need for an analytical approach to investment evaluation that focuses on uncertainty (e.g. robust decision-making, see Lempert and Groves (2010)) and on practical solutions (e.g., construction standards, concrete investments) that are flexible (see De Neufville and Scholtes (2011)).
3.1.3. Observation 3: uncertainty in projections of reference crop water deficit is higher in wealthier countries, and increases over time.
Interestingly, our analysis suggests that while the uncertainty in MAR and high runoff events increases with income, the opposite trend exists in projections of uncertainty in reference crop water deficit over time, as illustrated in figure 6. This trend appears to be more pronounced than for uncertainty in MAR. While this may appear to be a contradiction, as both measures consider temperature and precipitation forecasts, MAR is more dependent on precipitation outcomes, while reference crop water deficit is more dependent on temperature for the PET component, and also exhibits a threshold effect (when precipitation exceeds PET, deficit is 0). To the extent that  higher income countries are in higher latitudes, then, what appears to be at work is temperature outcomes exhibit higher variability in higher latitudes, while precipitation outcomes exhibit higher variability in lower latitudes. Additional work is underway to evaluate the robustness of this outcome.
We also conducted analyses of mean reference crop water deficit (rather than uncertainty) for the A2 scenario results versus income, and found no relationship between our projections with either income or with per cent of land area irrigated by country, suggesting that it is only the uncertainty in projections which vary with income. The result is potentially good news for poor agriculturally oriented countries, and presents a challenge for the agriculture sector in wealthier countries, particularly in areas where adding traditional water storage infrastructure has proven difficult owing to environmental concerns.
We also examined the relationship between the projected reference crop water deficit with climate change and the percentage of agricultural land that is currently irrigated across countries-in this case we forecast that countries with the highest current irrigation penetration also tend to face the highest increases in reference crop water deficit. This relationship is presented in figure 7 for the three future eras. The relationship is not as strong as for other results presented here, but does suggest that areas currently equipped for irrigation may face particular challenges related to increased crop water demand. Some of those issues could be resolved by altering crop choice, improving basin level and/or farm level water use efficiency, or increasing allocations to the agriculture sector (where possible). All of those measures, however, will require good information and advance planning to address.

Limitations
There are several key limitations to this analysis. First are the limitations of any hydrological study relying on climate change projections, namely (1) the assumptions, model physics, and parameterization of the GCMs; and (2) the unpredictability of future development pathways and the resulting scenarios for emissions of greenhouse gases, land use changes, and other factors influencing climate change; and (3) fundamental uncertainties in the impact of climate change on the hydrologic cycle and water resources and the modeling hereof.
In addition, there are several uncertainties which stem directly from using rainfall-runoff models in global climate change studies. These lumped models tend to be relatively simple, and often require a minimum amount of input in order to reduce both the uncertainty associated with inputs and the possibility of compounding errors. Their performance also relies heavily upon the quality of the calibration process, which is driven by the quality of the naturalized runoff inputs. Where the GRDC inputs are actually gauged flows, CLIRUN-II is being calibrated to human influenced flow rather than naturalized flow. Yet another issue is that because both the GRDC and CRU datasets tend to include too few extreme events (runoff and weather, respectively), there is a good chance that extreme events are under-represented in the CLIRUN-II results.
In terms of input data, both the CRU and GRDC datasets have additional uncertainties. Climatological station data is not always available for every time and place, an issue that tends to be more common in developing countries where station coverage is often poor. When and where weather records are not available, the CRU team uses an interpolation method to fill in missing data. Interpolation accuracy is of particular concern in areas with significant variation in elevation, and the accuracy of the original station data, in itself, is a source of notable uncertainty.

Discussion
The results presented here are designed to provide a sense of the value of hydro-indicators developed through this work; the real value rests in the value of these indicators to inform project planning, using a consistent and broad set of results. Infrastructure project design will nonetheless continue to require much more detailed hydrologic analyses. For example, climate change is expected to alter the seasonal pattern of precipitation, with the result that water can be in short supply at exactly the time it is needed most, during the high power demand or agricultural growing season. Higher temperatures also lead to more rapid evaporation from reservoirs, already a major consumptive use of water in many basins, and potentially more rapid evaporation from wetland areas such as those that characterize some areas, such as the Kafue flats region of the Zambezi River basin in southern Africa. These finer scale project-level assessments require a greater spatial and temporal scale than we can achieve with an indicators approach.
At a minimum, the enhanced understanding of the uncertainty of climate projections for the water sector that this work provides strongly support the adoption of rigorous approaches to infrastructure design under uncertainty, as well as design that incorporates a high degree of flexibility, in response to both risk of damage and opportunity to exploit water supply 'windfalls' that might result, but would likely require infrastructure to manage to the greatest benefit. In addition, it may make sense to not only consider changes to infrastructure investment levels, project design, and project operating rules, but also to consider non-infrastructure alternatives that can be effective 'in-the-moment' adaptations to changing climate. Despite the well-established infrastructure investment gap in many developing countries (SOFRECO Consortium 2011, Vivid Economics 2012, Foster and Briceño-Garmendia 2010, noninfrastructure alternatives may in specific situations postpone the need for some new climate-sensitive infrastructure. Certainly, climate change is not the only driver behind the need for more rigorous evaluations-issues of governance, institutional capacity, and the need for education and outreach to support wise use of infrastructure investments continue to be important as well.
The work also suggests a number of improvements that could be made in future efforts. First, there is a need for better hydrometeorological data-in particular in poor countries. The benefits of better data will be realized not only in the planning phases of these projects, but also in the operational phases. A better understanding of current variability may be at least as important as improving the physics in the GCMs, particularly when it is made clear that current water infrastructure is poorly adapted to current climate, let alone future climate risks. Improvements are particularly needed in both precipitation monitoring and understanding of naturalized runoff flows. Second, as noted above, a clear short-term need while data are enhanced and GCMs improved is focus in the near term on better planning models and practices for managing 'deep uncertainty'. Third, efforts are needed to mainstream what we have learned into the policies, planning and practice of vulnerable countries and the international finance community.
Finally, our results in figures 4 and 5 in particular provide a new insight about the relationship between water resources, climate, and country-level income, which deserves further attention. There is already a substantial and growing literature linking the temperature component of climate to country-level income (Acemoglu et al 2002, Dell et al 2009, and suggesting that the temperature component of climate may provide an indicator of future impacts of climate change (Horowitz 2009). Our work is prospective, concluding that lower-income countries that are least able to manage uncertainty in water availability are likely to face the greatest challenges in this area as a result of climate change. The results also suggests a subtle but potentially powerful factor in development research as well, that not only water availability but the level of certainty in water availability may be a key component of development success (Brown et al 2008), which is deserving of further exploration.