Climate change impact on river flow extremes in the Upper Blue Nile River basin

Extreme hydrological characteristics and available water resources, both in terms of quantity and quality, are significantly influenced by environmental changes, such as climate, land use, river engineering, construction of reservoirs, and mining activities. Among the influences, climate change and land use change are two essential factors controlling the hydrological behavior of catchments, such as river discharge (Dupas et al. 2016; Meresa 2018), water quality (Towler et al. 2010), and hydrological extremes (Taye et al. 2015). In particular, climate change has a huge influence on hydro-meteorological extremes of sub-Saharan countries (Chaney et al. 2014; WMO 2015). Zhang et al. (2016) compared the potential impacts in the near future river flow using projected land use patterns and hypothetical climate scenarios. The result proves that land use changes in the near future period induce slight (non-significant) reductions in groundwater discharge and surface runoff, whereas climate change produces pronounced increases in the river flow. This shows that the joint hydrological impacts are similar to those solely induced by climate changes. This is because the global climate has been changing and it will continue changing in future due to atmospheric greenhouse gases, aerosols, and human modifications of the land surface, thus significant climate change is expected in the future (IPCC 2013). According to the Intergovernmental Panel on Climate Change (IPCC 2013), the atmospheric concentration of CO₂ has increased from 345 ppm in 1950 to 405 ppm in 2011, and is expected to reach 463–640 ppm by 2050 and 800–1,313 ppm by 2100. IPCC further indicates that the global average air temperature has increased over the 21st century by about 0.9 ± 0.6 °C, and this increase is the largest of any century during the past 1,000 years. Depending on the different emission scenarios, the IPCC projects a further increase of global air temperature in the range of 1.1 to 4.8 °C. Similarly, precipitation intensity and variability has been varied in space and time (IPCC 2013). Closely linked to changes in atmospheric temperature and radiation balance, a number of components of the hydrological cycle can be affected, such as changing precipitation patterns, intensity and extremes, increasing atmospheric water vapor, increasing evaporation and changes in soil moisture and runoff (Vrac & Naveau 2007; Chen et al. 2011; Mawada et al. 2012). As a consequence, there is growing evidence worldwide of the changing characteristics of stream flows and extremes (Gelfan et al. 2015; Meresa et al. 2016). Therefore, this requires a better understanding and the need to quantify the change in projected extremes of precipitation, temperature, and runoff indices.

Much research has been carried out globally to explore, understand, and characterize the hydro-meteorological extremes (Towler et al. 2010; Hattermann et al. 2014; Byrne & O'Gorman 2015; Taye et al. 2015; Meresa et al. 2016, 2017). Similarly, hydrological and meteorological extremes are highly reported and in great detail in a European and American context. Many of the studies have pointed out that an increase in temperature will intensify the hydrological cycle and intense precipitation will increase, and Meresa et al. (2016) conclude that this is a precondition favorable for the development of hydro-climatic extremes and increase of hydro-meteorological drought. Indeed, Hattermann et al. (2012) point out that intense precipitation has increased worldwide. Taye et al. (2015) also mention that the increasing temperature projections indicate that potential evapotranspiration may simultaneously increase and lead to reduction in streamflow. Therefore, there is a growing need for information on climate change impacts on hydrological extremes (Taye et al. 2011; Te Linde et al. 2011; Hattermann et al. 2014; Meresa et al. 2016) and related damages. Haong et al. (2016) drew the conclusion that increasing precipitation leads to intensified extremes globally. However, according to the IPCC (2001), such impact of climate change on flood and drought is more serious in the sub-Sahara, South Africa, and eastern Asia. Similarly, the indirect consequences will increase, including risks to human safety and extreme hydrological events causing economic losses; these costs are rising exponentially (Dessu & Melesse 2013), threatening sustainable development. Management and planning in water resources and extremes refer to the time scale of decades and spatial characteristics (Meresa 2018), hence, measures implemented now should already take into account possible future climate change impacts on hydrology and water resources (Hattermann et al. 2012). There is, then, a need to improve the scientific understanding of changes and patterns in extreme hydrological events in the context of global climate change and regional climate changes to inform planning and management of water resources, disaster protection and alleviation at local/catchment scale. Also, the climate model simulation outputs are not always in good accordance with observation time series of precipitation and temperature (Solomon 2007; Osuch et al. 2015; Meresa et al. 2016). This means there may be some biases associated with the models and examinations; for instance, sometimes, heavy rainfall and number of heavy rain days (as well as their magnitude) are not well reproduced by RCMs (Taye et al. 2015). Bias correction is therefore necessary in order to improve the input data for hydrological process models, and can also be used as an input to other climate impact studies, such as water resource, agriculture, and environments (Piani et al. 2010; Eisner et al. 2012; Teutschbein & Seibert 2012).

Notably, all earlier studies are based on the SRES emission scenarios (Nakicenovic et al. 2000), which were used in the Coupled Model Intercomparison Project phase 3 (CMIP3). These scenarios, which only include non-intervention scenarios, have recently been replaced by the Representative Concentration Pathways (RCPs) scenarios (Van Vuuren & Riahi 2011; Stocker et al. 2013), resulting in a broader range of climate change. These most recent climate change scenarios (i.e., the CMIP5) are not yet routinely used to assess the hydrological impacts in the UBNRB. The CMIP5 scenarios also exhibit important improvements, both in terms of the GCMs' technical development (Knutti & Sedláček 2013) and the efficiency to reproduce historic climate conditions (Hasson et al. 2016). These important improvements and updates are highly relevant and require one to update the hydrological projections for the UBNRB. In this study, we will do this update and reflect whether the CMIP3 uncertainties relating to the hydrological signal will be reduced as well.

If hydrological models are applied within a single study, model calibration often results in reliable simulations of the past; however, the influence of model choice and model calibration on the simulation of climate and land use change impacts remains unclear, even if uncertainties (e.g., soil (Bossa et al. 2012) and climate are considered. For instance, from a theoretical viewpoint, a conceptual model represents the underlying hydrologic and land surface processes in greater detail than statistical models (Beven 2001). However, more parameters and greater calibration efforts are required as the degree of conceptual representation of relevant processes in a model increases.

We investigate the impacts of climate change in hydrological extremes on selected sites from the main Blue Nile River basin, Ethiopia. In particular our aims are: (i) to compare and evaluate three hydrological models (HMTS, GR4 J, and HBV) with two objective functions (one for low flow and the other for high flow) for climate change impact study; (ii) evaluation of climate models for impact analysis; and (iii) to quantify the changes and patterns in extreme indices derived from projected precipitation, temperature, and runoff.

In this paper, we address these knowledge gaps in understanding the UBNRB's hydrological extremes under climate change. Data input, study area, and the hydro-climatic characteristics of the selected catchments are briefly described. Three conceptual hydrological models were set up and calibrated for the selected river basins from the whole UBNRB, and we selected a set of seven climate change experiments for five GCMs and two RCPs from the CMIP5 and performed a downscaling and bias correction on the climate model output. Future changes in precipitation, temperature, and runoff changes were quantified, and we quantified changes in hydrological extremes, focusing on both extreme low and high flows. Finally, we summarize the results and conclude our research findings.

The UBNRB originates in the highlands of Ethiopia. About 75% of the Nile's waters originate in Ethiopia and Eritrea while the majority of the river's water is used in the Sudan and Egypt. The Blue Nile basin lies in the west of Ethiopia, between 7°45′ and 12°45′N, and 34°05′and 39°45′ E, and it covers an area of 199,812 km², with a total perimeter of 2,862 km (Figure 1). UBNRB accounts for almost 17.1% of Ethiopia's land area, about 50% of its total average annual runoff, and 25% of its population. The Abbay River has an average annual runoff of about 50 billion cubic meters (BCM). The rivers of the UBNRB contribute, on average, about 62% of the average Nile total at Aswan Dam. There are 16 main sub-basins in the river basin. However, in this study, we select six of them, as indicated in Figure 1 and Table 1, with their corresponding aerial coverage.

The Blue Nile has diverse climates, ranging from semi-arid desert in the lowlands to humid and warm (temperate) in the southwest. The climate of Abbay basin is dominated by an altitude ranging from 590 meters to more than 4,000 meters. The influence of this factor determines the rich variety of local climates ranging from hot to desert-like climate along the Sudan boarder with mean temperature of the coldest month above 18 °C, to temperate on the high plateau, and cold on the mountain peaks, with mean temperature in the warmest month below 10 °C. Annual rainfall varies between about 800 mm and 2,220 mm with a mean of about 1,420 mm. The region has three major climatic seasons: the dry season (Bega) from October to January; the short rainy season (Belg) from February to May; and the long rainy season (Kiremt) from June to September. Mostly the climate in the study area is controlled by three major weather systems, which are the ITCZ (Inter Tropical Convergence Zone) that drives the monsoon rainfall during the wet season (June–September), the Saharan anticyclone that generates the cool and dry northeasterly winds in the dry season (October–February), and the Arabian highlands that produce thermal lows in the hot season (March–May) (Seleshi & Zanke 2004).

The daily precipitation, daily air minimum and maximum temperature records for the study area were collected from the National Meteorological Agency (NMA) for about 30 stations which cover the entire selected river basins for the period 1971 to 2012 (Figure 1). Hydrological data for 1971–2012 of six major rivers were collected from the Ministry of Water Resources (MoWR) hydrology department, Ethiopia (Table 1). We calibrated and validated six sites/catchments, selected from different hydro-climatic conditions of the region. The study area has a season of highly variable rainfall from February to June, a wet season from July to September, and dry season from October to January. Whereas around 50% of the annual rainfall is received during the three-month wet season, rainfall distribution is highly varied and long dry periods are common. Figure 1 shows observed mean monthly pattern on temperature and precipitation for the selected catchments of the UBNRB during the time period 1971–2012. The information shown in Figure 1 is very important to assess and evaluate how climate has varied and changed in the past. The monthly mean observed precipitation and mean daily air temperature data can be plotted to show the observed climate and seasonality by month, for specific years, and for precipitation and temperature.

The input rainfall and PET data for hydrological models were calculated as weighted average time series from point measurements using the Thiessen polygon method. The Hamon method (Hamon 2016; Lu et al. 2005) was used for estimating PET. This method estimates potential evapotranspiration based on mean daily air temperature and central catchment elevation. Due to lack of data for all the required climatic inputs, the Hamon method of estimating PET with limited data was applied in the research.

The Coordinated Regional Climate Downscaling Experiment (CORDEX) was designed to produce improved regional climate change projections as part of the IPCC Fifth Assessment Report (AR5) and to standardize the generation and evaluation of regional climate projections across multiple modeling centers. As part of the CORDEX project, several regions were established with explicitly defined domains and model resolutions. This study relies on the CORDEX dataset, which contains regionally downscaled CMIP5 climate projections (Taylor et al. 2012) across the world. Therefore, we used mean daily precipitation, daily maximum air temperature, daily minimum air temperature, and daily average air temperature from three greenhouse gas concentration paths adopted by the IPPC for its Fifth Assessment Report (AR5). Given the relatively large number of GCMs under CMIP5, we first did a model selection by reviewing literature on GCM performance, and then we selected those GCMs that better reproduce historical temperature and precipitation conditions, implying their suitability to be used in the region. Thus, generally, we selected seven climate models that are produced by a combination of general circulation models and regional climate models (GCMs and RCMs) for this study (Table 2). For each GCM/RCM, we extracted climate data for three different RCP scenarios, namely, RCP2.6, RCP4.5, and RCP8.5. RCP4.5 is a medium to low scenario, assuming a stabilization of radiative forcing to 4.5 W/m² by 2100 (Thomson et al. 2011). RCP2.6 is a lower radiative forcing level for the lowest emission scenarios (Van Vuuren & Riahi 2011). RCP8.5 is a high radiative forcing scenario, assuming a rising radiative forcing leading to 8.5 W/m² by 2100 (Riahi et al. 2011). By selecting from the lower-range to mid-range and a high-end scenario, we expected to capture a reasonable range in climatic and hydrological extreme projections for the UBNRB, Ethiopia.

The experiment is carried out as follows: the hydrological models used here are three conceptual daily rainfall–runoff models. They are first calibrated against observed streamflow data and then driven using historical and future climate data with the same optimized parameter values to model the historical and future runoff. The future climate series are obtained by scaling the historical data, informed by outputs from seven climate models that contain a combination of RCMs and GCMs. The modeled future and historical runoff are then compared to estimate the climate change impact on runoff. The range of results from the seven GCMs and from the four hydrological models is then compared to assess the relative uncertainties.

It is known that GCMs/RCMs have biases and correction of these biases is necessary prior to use of the outputs (Christensen et al. 2008). Bias correction is the process of adjusting GCMs/RCMs output; mainly temperature and precipitation depending on discrepancies between observed and modeled results over the period of observation.

In this study, we used a distribution mapping method developed by Piani et al. (2010) to correct the bias of climate model outputs. The idea of distribution mapping is to correct the distribution function of RCM-simulated climate values to agree with the observed distribution function. This can be done by creating a transfer function to shift the occurrence distributions of precipitation and temperature (Piani et al. 2010; Rojas et al. 2011).

It is important to mention that for the climate model evaluation, only one station in each catchment was used for correcting the bias from each climate model.

The choice of a hydrological model for extreme hydrological process modeling in a given basin is a challenge for the hydrologist and water resource specialist. The available information is of great importance. In this study, based on the information we had about the river basin, we were oriented towards using lumped conceptual hydrologic models. Three hydrological models were used, namely, two lumped conceptual models (GR4 J and HMETS) and one semi-distributed conceptual model (HBV). The rationale for using three models displaying a range of model structures was that it would provide a clearer understanding of the importance of this particular factor in modeling extremes. Table 3 outlines the general properties of the hydrological models used in this study. In Table 3, P, Tmean, Tmax, and Tmin represent precipitation (P), mean temperature (Tmean), maximum temperature (Tmax), minimum temperature (Tmin), respectively.

The HBV model is a conceptual rainfall–runoff model of catchment hydrology which simulates discharge using temperature, rainfall, and potential evaporation. The model was developed for runoff simulation and hydrological forecasting (Blöschl et al. 2007). The advantage of HBV is that it covers the most important runoff generating processes by quite simple and robust structures where topographic and climate parameters serve as driving forces. Also, HBV does not require extensive datasets. The HBV model (Bergström 1976) has been applied in numerous studies, e.g., to compute hydrological forecasts, for the computation of design floods, or for climate change studies. HBV has been applied in more than 40 countries all over the world. It has been applied to areas with such different climatic conditions as, for example, Europe, South America, Asia, and Africa (Taye et al. 2015; Meresa et al. 2017).

The model consists of a precipitation routine representing rainfall and snow, a soil moisture routine determining actual evapotranspiration, overland flow and subsurface flow, a fast flow routine representing storm flow, a slow flow routine representing subsurface flow, a transformation routine for flow delay and attenuation, and a routing routine for river flow.

The GR4 J conceptual hydrological model has a parsimonious structure with four calibration parameters and has been frequently applied over hundreds of catchments worldwide (Demirel et al. 2013; Boukhaly et al. 2014; Meresa et al. 2016), with a broad range of hydro-climatic conditions from arid to semiarid and tropical to temperate catchments (Perrin et al. 2003). The GR4 J model requires only daily time series of temperature and precipitation and potential evapotranspiration as inputs (Table 3). The four parameters in GR4 J represent the maximum capacity of the production store (X1), the groundwater exchange coefficient (X2), the 1 day ahead capacity of the routing store (X3), and the time base of the unit hydrograph (X4). The upper and lower limits of these four parameters are selected based on previous works (Perrin et al. 2003; Thyer et al. 2009; Pushpalatha et al. 2011; Tian et al. 2012).

Production Store (X1) is storage in the surface of soil which can store rainfall. There are evapotranspiration and percolation in this storage. Groundwater exchange coefficient (X2) is a function of groundwater exchange which influences routing storage. Routing storage (X3) is the amount of water that can be stored in porous soil. Time peak (X4) is the time when the ordinate peak of flood hydrograph is created on GR4 J modeling. The ordinate of this hydrograph is created from runoff, where 90% of flow is slow flow that infiltrates into the ground and 10% of flow is fast flow that flows on the soil surface.

HMETS is a conceptual rainfall–runoff model developed at the Ecole de Technologie Supérieure (Brissette et al. 2010), which uses two connected reservoirs for the vadose and saturated zones. It has been used in multi-model averaging projects (Arsenault et al. 2015) and in climate change impact studies (Chen et al. 2011). The model simulates the basic hydrological processes: evapotranspiration, infiltration, snow accumulation and melting as well as flow routing to the catchment outlet. It is a MATLAB-based freeware, and has up to 20 parameters: ten parameters for snowmelt, one for evapotranspiration, four for infiltration, and five for upper and lower soil reservoirs. However, in this study, we used only ten parameters (one for evapotranspiration, four for infiltration, and five for upper and lower soil reservoirs). The HMETS model calibration is done automatically using the SCE-UA (Duan 2003) optimization technique. The model requires daily precipitation, maximum temperature, and minimum temperature as input to simulate river flow. A natural inflow or discharge time series is needed for proper calibration/validation.

The parameters of the hydrological models are optimized using the SCE-UA method, which is a global searching algorithm proposed by Duan et al. (1992) and has been used in many hydrological models. The SCE-UA method combines the direction-searching of deterministic and the robustness of stochastic, non-numerical methods to obtain a global optimal estimation (Duan et al. 1992). To make the implementation more convenient, Duan suggested some default values for the parameters of the SCE-UA method (Duan et al. 1992).

All models are verified on the same 30-year calibration and 10-year validation sets. We added a warm-up year at the beginning of the calibration and validation time series sets: this avoids unknown initial conditions having an impact on model performance and allows the models' internal state variables to adjust to appropriate values. The calculation of performance criteria does not include simulations during this warm-up period. Performance is thus evaluated over the same period for low flow and high flow hydrological variables.

A number of hydro-meteorological measures exists to describe the statistical properties of extreme hydrological and meteorological events. In terms of high flow, low flow, dry spell, wet well, and hot temperature, the following measures are considered in this study:

• Dry weather flow, MIN7, defined as the average annual 7-day minimum river flow.

• MLDD index is the number of consecutive dry days (i.e., maximum length of dry spell; RR <1 mm), where RR is daily precipitation.

• MLWD index is the number of consecutive wet days (maximum length of wet spell, maximum number of consecutive days with RR ≥1 mm), where RR is daily precipitation.

• ADMT index is the maximum value of daily maximum temperature.

• Average annual maximum/minimum daily river flow, MAX/MIN, defined as the average value of a time series consisting of annual maximum/minimum daily river flows.

• Coefficient of variation of annual maximum/minimum daily flows, CMAX/CMIN, a dimensionless measure of the variability of annual maximum/minimum river flow magnitudes.

• The number of high flows, NHF, i.e., flows greater than Q = α + 3* σ, where α and σ are the average and standard deviation of daily river flow series.

• The 90th/10th percentile flow, Q₉₀/Q₁₀, defined as the river flow, which is equaled or exceeded for 90%/10% of the period of record, which is determined from a flow duration curve.

These flow statistics are termed MAX, CMAX, NHF, Q10, MIN, CMIX, MIN7, and Q90 and define the flows exceeding 90 and 10% for the Q90 and Q10 flow from the flow duration curve, respectively. The first five indices belong to high flow characteristics of the basin and the last five represent low flow characteristics of the basin; they were computed to investigate how frequently these events occur.

To investigate climate change impact on the hydrological extremes over the period of 2071 to 2100, we first calibrated (1971–2000) and validated (2001–2010), and then we adopted each optimal hydrological model for future hydrological projections for our hydrological models in the reference period (1971–2010). The inputs of hydrological model, projected precipitation, temperature, and PET were derived from different GCMs/RCMs for RCP2.6, RCP4.5, and RCP8.5 emission scenarios after applying a distribution based quantile mapping (DQM) technique. The results of climate change impact on the hydrological extremes are presented in the following sections.

The bias is defined as long-term average difference between climate model output and observation. This shows how well the best available climate models capture the seasonal cycle of precipitation and temperature for the selected sites/catchments. The evaluation of climate model biases of precipitation and temperature are shown in Figure 2. Figure 2 shows the annual cycle of monthly biases of precipitation (left column) and temperature (right column). In this case, the error variability is indicated by box-and-whisker plots. The bias correction was evaluated by splitting the sample into two, one for calibration in the period 1971 to 1991 (dark) and the second part for validation in the period 1992 to 2012 (dark), together with the raw RCM bias of each period (darker (1971–1991) and darker (1992–2012)), respectively. This method of bias evaluation can help to understand the non-stationarity characteristics of precipitation in the area. The figure shows in the case of temperature, calibration in the period 1971 to 1991 (dark) and validation in the period 1992 to 2012 (dark), together with the raw RCM bias of each period (darker (1971–1991) and darker (1992–2012)), respectively.

For monthly timescale (Figure 2), the result evaluated in both calibration and validation period, corrected models have smaller biases and smaller bias ranges than the raw models. The bias correction successfully reduces biases during the calibration (1971–1991) and validation (1992–2012) period. This implies the bias correction technique performs reasonably well in preserving the variability and magnitude of mean monthly precipitation in most months, although there is a relatively insignificant model error. In all the sites and climate models, the larger bias was observed in the main rain season (June, July, August, and September) of the region. For instance, in Figures 2(a)–2(c) RCP4.5 scenario shows overestimation ranging from 1 mm to 9 mm for June to September and underestimation ranges from 1 mm/day to 7 mm/day for the months of April to May. In general, the raw RCM simulations of precipitation have modest biases (within −5 to +10 in RCP4.5 and −10 to 15% in RCP8.5) in the mountainous sites/catchments than the lowland sites, which show relatively higher bias in the lowland sites than mountainous sites. However, the biases behave differently for different emission scenarios (RCP2.6, RCP 4.5, and RCP 8.5), which indicates that the biases in precipitation are smaller in RCP2.6 compared to the other scenarios, and in RCP8.5 the biases are relatively larger. This is because of the lowest radiative forcing and substantial decline of greenhouse gas concentration in RCP2.6 which leads to small bias and variability, whereas scenario RCP4.5 is characterized by very small increases of greenhouse gas concentration. RCP8.5 is characterized by more than a double increment in greenhouse gas concentration, which implies higher bias and variability in precipitation.

Figures 2(d)–2(f) show the annual cycle of monthly biases in air temperature. In this case, the biases and their variability are generally strongly reduced. In some months, however, considerable errors remain after the correction (e.g., months belonging to Belg season), which is caused by different model error characteristics in the calibration and validation periods (i.e., by non-stationarity). It can be concluded that the bias correction for temperature leads to satisfactory results. Corrections are largest during the Bega months and smallest during Kiremt. This is mainly caused by the difference in mean temperature, as shown later in Figures 2(d)–2(f). This result indicates that using the RCM output without doing bias correction may lead to enormous uncertainty of hydrological analysis.

Table 4 presents the bias between corrected and uncorrected precipitation and temperature over the ensemble of climate models. The most pronounced biases occur for simulations of precipitation, in particular in the north and centre, a region with steep topographic and precipitation gradients (Table 4). The climate model tends to underestimate precipitation on the central and northern mountains by 15% or more, while it tends to greatly overestimate precipitation (>15%) in the low precipitation of the lower part of the river basin. The temperature behaves in the opposite way to precipitation, which is overestimated in the lowland part of the basin and underestimated over the mountainous region. Thus, to conclude, the temperature and precipitation biases in the uncorrected simulations are high in the northeast and north largely because of orographic rainfall effects (in this case, a rain shadow caused by a steep topographic gradient from the mountainous north to the lowlands in the north-east part of the UBNRB, Ethiopia).

Figure 3 illustrates the temporal variability and distribution of selected indices derived from projected precipitation and temperature variables over the selected six sites from the Upper Blue Nile River basin (UBNRB). In this study, indices, namely, maximum length of dry spell (MLDD), maximum length of wet day (MLWD), annual daily maximum precipitation (ADMP), and annual daily maximum temperature (ADMT) were selected from projected climate time series. As a result, Figure 3 shows that the station (right column: site 1) is dominated by increasing ADMT and MLDD, while the ADMP shows neither increasing nor decreasing but the variability is increasing with time. Most stations are identified by increasing dry spell length (MLDD), nevertheless, only two sites located in the north and central part of the basin are characterized by significant increases in MLDD. The MLDD and ADMT in the station have increased with time, as well as a slight decrease in daily precipitation extremes. This could be attributed to a combined effect of global warming, anthropogenic influences, natural climate variability (such as North Atlantic Oscillation, El Niño-Southern Oscillation), etc. MLWD is slightly decreased with time and this index is highly diverse spatially (among the selected sites). Mean MLWD in site 1 (Figure 3: right column) is around 6 days and 8 days in site 2 (Figure 3: left column). Mean MLDD of precipitation increases from 80 days to 110 days in site 1, while in site 2 it increases from 65 days to 100 days. The mean value of the indices in the far future (2071–2100) is different for different climate scenarios and models. Among the scenarios, RCP8.5 shows larger increase in MLDD and larger decrease in MLWT, and a similar condition also occurs for ADMT and ADMP in the far future, while RCP4.5 produces reasonable change in mean indices in the far future. Such differences appear due to the difference in radiation force level. The higher radiation force produces a drier future.

Inter-model spread is large, particularly over the south and west part of the river basin, which leads to higher bias in the ensemble mean estimation for projected impact analysis. Over the north and central parts, inter-model agreement is found to be higher, with about 90% of models agreeing on projected indices.

The number of seasonal Kiremt, Belg, and Bega wet days, dry days, and maximum temperature at the six selected sites in the UBNRB region are depicted in Table 5. The observed number of wet days in Kiremt season varied from 21 days at site 2 to 61 days at site 6. On average, there were more MLWD in a year at site 5. Temporally, there was more variability in the MLWD at site 2 and, conversely, less at site 6 than the other sites studied during 2071–2100.

On the other hand, in Belg season, MLDD varied from 41 days at site 3 to 80 days at site 5 during 2071–2100 (Table 5). There were, on average, more Belg season MLDD in a year at site 4 than the rest of the sites studied. The Belg season showed that there was moderate to high inter-annual variability in the MLDD at all the studied sites. This indicates that the MLDD in the Belg season was less dependable in the study area. The seasonal Belg MLDD varied from 41 days at site 3 to 80 days at site 5. Moreover, the present study has shown that the number of rainy days was more variable during the Belg season than during the Kiremt season, and the number of dry days was less variable than the number of rainy days in both seasons. From a hydrology, hydraulic structure, agricultural point of view, high inter-annual variability in the number of rainy days shows less dependability of the rains for planning activities, which may lead to crop failures, longer dam/reservoir filling time, and challenging hydrological operation. In particular, the high variability of rainy days for the Belg season could be a great problem for farmers who lack instruments to quantify rainfall amount but rather depend on the number of rainy days to plan their cropping calendar.

Sensitivity analysis is conducted before the calibration process to identify the most important/sensitive parameters for each site, and model components. Insensitive parameters can be fixed to suitable values to decrease the dimensionality of the calibration problem. For the six selected sites/catchments, the first year of input data measurements were used for the ‘warming-up’ of the models to estimate the initial state variables. The rest of the data were divided into two time periods, two-thirds of the data length for calibration (1971 to 2000) and one-third for validation (2001 to 2010) for the three hydrological models. We considered the Nash–Sutcliffe and log Nash–Sutcliffe criteria as objective functions for calibration; the statistical criteria evaluates how good the fit is between observed and simulated values. The parameters of the hydrological models are optimized with the SCE-UA method. We carefully calibrated and validated the three hydrological models. Figure 4 compares the modeled hydrograph from three hydrological models and observed streamflow during the calibration period from July 1981 to July 1997 for illustrative purposes for two sites (site 1 and site 5, respectively). This reveals that the three models work well in the study basin in reproducing the historical flow and in simulation of flood peaks and low flows. All peak flows are captured with high accuracy except for the peak at lowland catchments, where the modeled peak occurred earlier and was lower than the observed.

For the observed peak at site 2, which is located in the highlands of the basin, the model captured the first peak well. The less pronounced overprediction of low flows by HBV compared to GR4 J may indicate that the slow responding groundwater storage in HBV is less sensitive to different forecasted ensemble precipitation and temperature inputs; while the HMETS hydrological model behaves reasonably acceptably for high and medium flow. In evaluation using the second objective function (LogNSE), the recession limbs of the hydrographs are generally better modeled than the rising parts; this can be attributed to the limited ability of the model to simulate longer wet-weather periods, which are simulated better using the first objective function (NSE). Therefore, in this study we considered LogNSE objective function for low flow indices simulation and NSE for high flow indices simulation.

Table 6 shows the result of statistical performance in calibration and validation periods of the selected sites/catchments in the UBNRB, Ethiopia using three hydrological models and two objective functions. The results of calibration and validation range from 0.58 to 0.83, which shows very good agreement between the observed and estimated flows. The three models confirm that they can reproduce historical streamflow series with an acceptable accuracy. However, the low flow was well produced using LogNSE objective function and the high flow using NSE. Hence, in this study, we used both objective functions for further analysis. When we compare the hydrological models' performance, the HBV model has the highest NSE and GR4 J has the highest LogNSE values and HMETS has the lowest NSE and LogNSE values. The values in Table 6 demonstrate the basic capability of each model to reproduce daily observed streamflow in the selected sites in the Blue Nile River basin, which shows that both GR4 J and HBV models have high performance in reproducing historical high and low flow data for the study basin. This leads to a notion of model complexity. In general, the performances of GR4 J and HBV are similar in the calibration period, whereas HBV and HMETS perform better in the validation period. This is not surprising, since HBV has a more sophisticated model structure than GR4 J and HMETS.

In order to compare the performance of the three models, the daily discharge output of HBV, GR4 J, and HMTES is averaged to monthly max and min runoff, and the corresponding Nash–Sutcliffe coefficient and log Nash–Sutcliffe coefficient are computed. As shown in Table 6 and Figure 5, the Nash–Sutcliffe coefficients of HBV, GR4 J, and HMTES at daily time scale for the period 1971–2000 range from 0.65 to 0.89 and the log Nash–Sutcliffe coefficient from 0.58 to 0.83, respectively. Figure 5 shows a comparison of the three hydrological models for high flow (left column) and low flow (right column) simulation. It is found that the bias in monthly maximum flow was smaller using HBV and higher by the HMETS model during the reference period (1971–2000), while a smaller bias in monthly minimum flow was observed by GR4 J. Overall, HBV performs well in simulation of maximum flow whereas GR4 J performs well for low flow over other models due to the structure of the models.

To demonstrate the impact of bias correction on model results, for illustration, Table 7 shows the discharge of six hydrological sites simulated by using the HBV model driven by the raw and the bias-corrected output from one climate scenario (RCP4.5), for the control period 1971–2000. Without the bias correction, the model overestimates both the peak flow and the low flow in the control period. The bias of minimum/maximum/mean runoff during the reference period simulated by using the HBV model driven by the raw outputs of the climate models was 59.5, 37.4, and −16.9%, respectively. These biases were reduced to −9.5, −11.3, and −4.7% after the bias-corrected climate was used for site 1. When the same bias correction method was applied to drive future hydrologic simulations using the HBV hydrological model, a similar effect is established, with reductions in the future simulation (Table 7). The importance of bias correction to simulate future hydrological extremes is clearly visible.

In this section the changes in extremes of precipitation, temperature, and river flow are presented, respectively.

The projected changes of maximum wet spell (MLWD), maximum dry spell (MLDD), maximum precipitation (ADMP), and maximum temperature (ADMT) in the period 2071–2100 relative to the reference period are shown in Figure 6. The MLWD indicator demonstrates that the future wet spell will be increased in most sites over the river basin except for site 3 and site 4. However, the degree/range of change is not the same at all sites. Large range change is observed in the lowland part of the basin (e.g., site 2). The MLWD change has higher variability than the MLDD and ADMP. There is a pronounced change except in the sites in the central and northern part of the river basin, where small decreases in MLDD are expected with the maximum decrease reaching −23%. In the sites over the remaining areas of the river basin, the MLDD change reaches up to 35%. ADMP behaves with decreasing dominancy in the river basin. In the case of temperature, all sites and climate scenarios in the river basin showed positive change for maximum temperature. This may be due to the emission radian forcing direct relation with temperature and greenhouse gases. The main factor attributing to these changes is the impact of global warming on the temperature and precipitation patterns. The presence of short-length extreme wet spells and prolonged dry days, which will adversely affect agricultural practices and water management in the country, is discernible from the analysis (Figure 6).

The three climate scenarios show a consistency, decreasing in wet spell and increasing in wet spell and maximum temperature. However, the magnitude is different among the scenarios, with larger variability and magnitude by RCP8.5 and smaller by RCP2.6 climate scenarios. In particular, RCP8.5 reveals higher magnitude and large change range, but the RCP2.6 and RCP4.5 climate scenarios show more reasonable results and relatively small differences in their future extreme predictions. The reason is related to the level of radiation emission forcing, the higher greenhouse gas emission has a large effect on the wet/dry days in the region. The highest decrease in wet spell occurs in site 1 while the smallest is in site 6. The remaining sites are between those maximum and minimum site values. The slight decrease in MLWD can be seen at the sites over the upper northern and central part of the river basin by RCP4.5 and RCP2.6 climate scenarios (Figure 6).

As shown in Figure 7, in the period 2071–2100 (2080s) there may be a decrease in mean monthly precipitation for all months except June, July, August, and September for three scenarios (RCP2.6, RCP4.5, and RCP8.5). There is a decrease in annual mean precipitation for the 2080s and an increase for the Kiremt season and decrease for the Belg season, which means the annual decrease in change is more dominated by Belg than Kiremt. The RCP2.6 scenario displayed a mean monthly precipitation decrease up to −20% in January and December and an increase up to 18% in August, and RCP4.5 also shows a similar range of variability. RCP8.5 reflects quite large temporal variability with higher variability in its incremental (20%) and decline (−29.5%) range. Most likely, it is observed that precipitation during the Kiremt (wet season (June–September)) for the long-term future will increase, while precipitation for the Belg season (less rainy season (February–May)) will decrease in the far future (2071–2100). The Kiremit and Belg are the cropping seasons in Ethiopia. This gives an insight into the possible impact of climate change on agriculture and irrigation development in the study area.

The characteristics of changes in hydrological extremes are calculated from the model outputs obtained from three scenarios. In this study, the projected (2071–2100 period) flow indices were compared with the reference (1971–2000) period, and are presented in Figures 8 and 9. The results of seasonal change and extremes per site change are presented in Figures 8 and 9, respectively.

The relative change and variability of the climate models in the monthly maximum flows for the projected time period (2071–2100) versus the reference monthly maximum values were calculated and are shown in Figure 8. Generally, the scenario ensembles show lower monthly river flow at all considered stations, except for a large increase in June, July, August, and September. This means relative river flow increases are more substantial in the wet season compared to those in the dry season. The ensemble's projection ranges become markedly larger in the wet season, implying higher uncertainty in the hydrological change indicators. Among the scenarios, RCP4.5 shows the largest increase in November, while RCP8.5 shows the largest increases in January, February, and December. Although absolute increases are more substantial during the wet season months, relative increases are higher during the wet season. For instance, discharge in July and August could increase up to 65% at site 1, 41% at site 3, 150% at site 5, and 25% at site 4. Discharge in December, January, and February is projected to reduce slightly at all six sites, ranging between −8% at site 2, followed by −10% at site 6, and −5% at site 4. On the seasonal timescale, discharges increase at all stations during the wet seasons and the variability of GCMs/RCMs monthly discharge changes during the wet season are more variable compared to the dry season. This means that favorable water surplus conditions are more likely in the future during Kiremt season.

Figure 9 presets the percentage changes of hydrological extreme statistics (MAX, CMAX, NHF, Q10, MIN, CMIX, MIN7, and Q90) of the six sites under three scenarios for the 2080s period. CMAX and CMIX show a similar behavior in all sites; however, the range of percentage change is varied in space and time. The minimum flow shows lower variability than the high flows, which is maximum at sites 1 and 2 and minimum at site 6 (17%), and the CMIX varies from 10% at site 6 to 28% at site 2. Q90 and Q10 do not show a similar range of percentage changes, for Q10 changes are −20% at site 2 to 17% at site 5 while the range of changes in Q90 is much wider (−18% at site 2 to 18% at site 5). This result can be explained by low-flow seasons, which are usually more sensitive to changes in potential evapotranspiration than high-flow seasons. It is also partially because a relative percentage change value is magnitude-sensitive, e.g., the same increment or decrement gives a higher percentage change for low flows than for high flows. The increased low flows in this region suggest increased low flow during the mild Belg season (February to May). The RCP4.5 scenario projects an almost −23% decrease in the magnitude of the MAX at site 2 and increase of 12% at site 5. Under the RCP8.5 scenario, the increased temperature is in good accordance with the decrease in the magnitude of the MAX in site 2 and site 3. Also, more Kiremt flood events increase the number of high flows (NHF) by 6.5% at sites 1, 5, and 6.

In terms of extreme minimum flows, three of the scenarios projected an increase in the MIN except at site 2 which decreased by 10%. The increase in the MIN values is around 22%. Also, the 7-day minimum flow (MIN7) reflects a decrease in site 2 while in the other sites it shows increases by −18 to 22%. The increase in MIN7 flows can perhaps be explained by the shift in the period of low flows from the beginning of June toward the month of August, where the soil moisture storage is still affected by the wetter Kiremt season months. In summary, low flows (both MIN and MIN7) may become less extreme, less variable in terms of magnitude, and more variable in terms of temporal occurrence.

The presented results correspond well with the results published in other studies focused on the UBNRB. For example, Kim et al. (2008) estimated a large increase in low flow and wider change range (−25% to 60%) and slight increase in high flow and narrower change range (−15% to 20%), which implies reduced severe drought events in the region. Aich et al. (2014) also discussed an increase in high flows from 10% to 50% and an increase in low flows from 40% to 60% in the region.

This study assessed the expected climate change impact on hydrology extremes of six selected sites/catchments from the UBNRB and its implication for water resource management and planning and disaster prevention and preparedness. Seven RCM future projection runs, namely, canESM2, CCM4, CNRM-CM5, CSIRO-Mk3.6.0, HadGEM2-ES, MIROC5, and MPI-ESM-LR were acquired and used for this purpose. The climate models' runs are based on the CMIP5 RCP2.6, RCP4.5, and RCP8.5 scenarios. The run results for daily precipitation, and minimum and maximum temperatures were bias corrected against observed data using the DQM method for precipitation and temperature. The evaluation includes both biases and measures for temporal and inter-variable consistency. This is based on a split in the precipitation and temperature data with strictly independent calibration and validation periods. The hydrological extreme biases are reduced by DQM; larger reduction was seen in the lowland sites than mountainous parts of the region for all variables in most cases. Some exceptions were found, but did not restrict the method for improving the raw RCMs. Even in the case of high temporal variation, DQM still clearly improves the biases of the raw RCM. It is most important to use independent calibration and validation periods, which are affected by climate variability and change. Thus, these results give some indication for the performance of DQM applied to future scenarios.

Projected changes of extremes were derived from precipitation, temperature, and streamflow for the future period over selected sites in the UBNRB. The changes are calculated for the future period (2071–2100) relative to the 1971–2000 reference period under the RCP2.6, RCP4.5, and RCP8.5 emission scenarios. An ensemble of seven climate model outputs were used in calculations of the extreme precipitation and temperature indices (after bias correction was applied): AMDP (the number of heavy rainfall days), MLDD (the annual maximum of consecutive dry days), MLWD (the annual maximum of consecutive wet days), and ADMT. From the projection of extreme indices' changes, we conclude that most areas of northern, western, and the northeastern part of the river basin will likely become wetter in the wet season and drier in the dry season. The MLWD indicator demonstrates that the future wet spell will increase at most sites over the river basin except for site 3 and site 4. However, the degree/range of change is not the same at all sites. Large range change was observed in the lowland part of the basin (e.g., site 2). The MLWD change has higher variability than the MLDD and ADMP. There is a pronounced change except in the sites over the central and north part of the river basin, where small decreases in the MLDD are expected with the maximum decrease reaching −23%. In these sites over the remaining areas of the river basin, the MLDD change reaches up to 35%. ADMP behaves with decreasing dominancy in the river basin. In the case of temperature, all sites and climate scenarios in the river basin showed positive change for maximum temperature. This may be due to the emission radian forcing direct relation with temperature and greenhouse gases. The main factor attributed to these changes is the impact of global warming on temperature and precipitation patterns. The presence of short-length extreme wet spells and prolonged dry days will adversely affect the agricultural practices and water management in the country. This study concludes that climate change is affecting the precipitation characteristics significantly and it brings variability in precipitation behavior.

In this study we compare three hydrological models in the reference period. The result confirms the conclusion that HBV is more promising for future climate change impact study in the region. Climate change impact on the hydrological extremes in the 2080s has been carried out for six selected river basins in Ethiopia. The lumped conceptual hydrological model, HBV, was used to estimate the hydrological extremes under climate change conditions. The extreme indices are calculated from projected flow, namely, MAX, CMAX, NHF, Q10, MIN, CMIN, MIN7, Q90, and changes in the 2080s have been generated with respect to the reference period. The assessment of climate change impact was based on seven GCMs/RCMs simulation for RCP2.6, RCP4.5, and RCP8.5 emission scenarios. Overall, indices that related to low flows are projected to increase most prominently for lowland sites, due to the combined effects of projected decreases in Bega and Belg precipitation, and projected increases in evapotranspiration that reduce residual soil moisture in Bega and Belg. The indices related to the high flows project a slight increase in the central and upper part of the basin.

According to the three evaluated scenarios, climate change may have negative impacts on the hydrological extremes in the selected study areas. In this study, the RCP4.5 scenario represents an average change in all selected sites and hydrological indicators compared with the other two climate scenarios (RCP2.6 and RCP8.5). The future regime of maximum flows in the UBNRB can be characterized as less extreme in the far future and more variable in terms of both seasonal and annual occurrence and magnitude. Low flows may become moderately extreme and highly variable in spatial and temporal occurrences for most of the selected sites, due to the combined effects of reduced groundwater storage and projected increases in evapotranspiration that reduce residual soil moisture in Belg and Bega. In the future projections, low flows do not respond as strongly to increasing hydrological drought stress in these areas, because Bega soil moisture is already at very low levels for the reference period. Relatively high Bega and Belg soil moisture in the south and central part of the UBNRB, by comparison, can be reduced substantially by increased evapotranspiration and decreased Belg precipitation resulting in substantial reductions in base flows of the region. This robust evidence of changes in hydrological extremes (both low and high flow) shifts research and water resource planning and management focus to these low probability but potentially highly damaging events, which is important to reduce climate change impacts and associated risks in the far future period.

This study makes it clear that climate change impact assessment is an extremely complex issue, hence, encouraging more research on this topic in order to increase awareness and to help improve further investigations and predictions is very important.

We acknowledge National Meteorological Agency and Ministry of Water Resource, Ethiopia, for providing meteorological and hydrological data. Hadush K. Meresa performed the modeling strategy, analyzed the observed and future climate data, interpretation of the result, and wrote the paper. Mulusew T. Gatachew coordinated the data collection, supported the data analysis and writing, and proofread the paper. The authors declare no conflict of interest.