Flood hazard attribution and uncertainty analysis with climate changes in data scarce watersheds of southeastern Ethiopia

Flood hazard is the probability occurrence of potentially damaging flood phenomena within a specified period in a given area (Assefa 2018). Climate change refers to a change in the state of the climate that can be identified by changes in the mean and the variability of its properties that persists for extended periods, decades, or longer (IPCC 2007). The changing climate would bring changes to hydrological parameters like temperature, precipitation, evapotranspiration, and discharge in magnitude and frequency (Jain & Kumar 2012; Bhatt & Mall 2015). Rainfall can produce widespread surface flooding where water encounters dry ground and infiltrates, raising groundwater volumes (Adnan 2010). Temperature increases can lead to increases in evapotranspiration in rivers, dams, and other water reservoirs. Consequently, water availability for agricultural irrigation, domestic and non-domestic usage, and hydropower generation decreases (Adnan 2010). In tropical regions, annual flooding is associated with high intensity of rainfall and heat in the atmosphere (Dettinger 2009; Williams et al. 2012; Hall et al. 2014).

Several studies used different approaches to understand and quantify the effects of climate change on flooding. Studies like Taye & Willems (2012), Chen et al. (2012), and Li et al. (2020) analyzed time-series trends exhibited in the historical hydrological data. Other studies analyzed historical or current meteorological data coupled with hydrological models (Adnan 2010; Dile et al. 2013; Gebrechorkos et al. 2018). Others used a combination of climate models (i.e., a general circulation model, GCM and regional climate models, RCMs) with hydrological data for future projections (e.g., Adnan 2010; Endris et al. 2013; Ruiz-Villanueva et al. 2016; Musie et al. 2020). Gebrechorkos et al. (2018) evaluated multiple climate data sources to overcome data scarcity in East Africa based on different statistical measures used on daily, decadal, and monthly timescales. The result revealed that Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) and Africa Rainfall Climatology version 2.0 (ARC2) are the best-performing rainfall products in environmental resources management. Most of these studies were conducted in humid and semi-humid areas, where ground-based observed data are available relative to the southeastern lowland part of Ethiopia. In a hydrological forecast system, uncertainties are assessed based on quantifying uncertainty values from the inputs, like precipitation and temperatures (Lumbroso & Gaume 2012; Gaur et al. 2020). The study of uncertainty in modeled discharge from the hydrological model still needs further study.

The main objective of this paper is to quantify the impacts of climate change on streamflow and flood hazard in southeast Ethiopia, a case of the Wabi Shebele River Basin using a model-based approach. Wabi Shebele Basin is known as a data scarce watershed in Ethiopia (Abebe & Förch 2006; Wudineh et al. 2021). The specific objectives of the study are listed as follows: to test the performance of SWAT and identify sensitive parameters in flood prediction; to test climate sensitivity in flood hazard prediction; to analyze uncertainties of a climate model in streamflow prediction; and to estimate the share of impact between climate change and others on flood value in the region.

The climate of the basin is dependent on the altitude and strong latitudinal movement of the intertropical convergence zone (ITCZ) (Awass 2009). This study area is categorized as a semi-desert zone having a mean annual rainfall ranging between 271 mm at the lower arid portion (Gode) and 1,320 mm in the upstream highlands of the basin (Abebe & Förch 2006). The air temperature of the area varies with altitude and ranges between 20 and 25 °C (MoWR 2003). The mean annual evaporation reaches 2,800–3,000 mm in the southeast (Wudineh et al. 2021).

Flood is one of the frequently occurring natural disasters in the Wabi Shebele River Basin (MoWR 2003; Tadesse et al. 2016). In the basin, flood events have occurred regularly as flash floods in the lowland sections, as seen from the state of river beds and evidence of sheet erosion (IWMI 2015).

The spatial resolution of GCMs (currently 100–250 km) is too coarse for the direct outputs used in hydrological impact assessments on the catchment scale (Veijalainen et al. 2010; Musie et al. 2020). Although climate change is a worldwide concern, its impact on the hydrologic system is at the regional level that needs to be downscaled to appropriate scales. There are two groups of downscaling methods: dynamical and statistical downscaling (Fowler & Kilsby 2007). In dynamical downscaling, a regional climate model (RCM) or limited area model (LAM) of the higher spatial resolution set for a region (Hay et al. 2002; Veijalainen et al. 2010). RCMs use boundary conditions from the GCMs, but capture geographical details more precisely than GCMs (Hay et al. 2002). Few studies in the East African region used RCMs from dynamic downscaling (Endris et al. 2013; Musie et al. 2020). An ensemble of historical and future climate projections generated by the Coordinated Regional Downscaling Experiment (CORDEX) is available from the World Climate Research Program (Giorgi et al. 2008). CORDEX-Africa (http://cordex.org/domains/region-5-africa/) provides regionally downscaled climate data for the continent at a spatial resolution of 0.44° × 0.44°. Future projections of precipitation and temperature were investigated using CORDEX-Africa datasets. There are two methods commonly used to transfer the climate signal to the hydrological model: the delta change approach (Fowler & Kilsby 2007) and the direct RCM data approach. In using direct RCM data, the daily bias-corrected results from RCM were used as input to the hydrological model (Veijalainen et al. 2010). The delta change approach is classified as the simplest statistical downscaling method. However, the direct RCM data relies on dynamically downscaled by the RCM with an additional bias-corrected step.

The CORDEX-Africa downscaled from five GCMs of Climate Model Inter-comparison Project Phase 5 (CMIP5) to quantify the influence of future changes in regional climate on the hydrology of the Wabi Shebele River Basin. The latest version of the RCAs, developed by the Swedish Ross Centre Regional Atmospheric model (RCA4), is selected for this study (Samuelsson et al. 2011). In addition, two different scenarios of the Representative Concentration Pathways (RCPs) were considered for all models. For the model, climate data belonging to two RCPs emission scenarios (i.e., RCP4.5 and RCP8.5) extracted for 20th century climate (historical runs; 1981–2005) and future climate (2006–2100). These RCMs selected in studies (Diro et al. 2011; Gebrechorkos et al. 2018; Näschen et al. 2019; Musie et al. 2020) have shown that each model can reasonably reproduce the regional climate over the East Africa region. The simulations cover the period from 1951 to 2100, divided into a historical period from 1951 to 2005 and future projections from 2006 to 2100. The details of the RCA4 simulations used in this study are presented in Table 1.

In this study, the CMhyd tool (Rathjens 2016) was used to execute the bias correction techniques. The tool compares the raw RCM output with observed data, calculates the variation between observed and RCM-simulated data, and applies different bias correction methods to correct historical and future climate model output. The bias correction algorithms derived from historical RCM simulation and observed data are applied for future RCM bias correction processes. Temperature and rainfall data at 14 stations in the Wabi Shebele basin are used for the bias correction of the RCA4 climate data.

An ensemble of potential climate data is generated from the historical station record (1981–2000) and analysis of climate sensitivity. In the climate change impact assessment using climate sensitivity tests (Ficklin et al. 2009; Cheng et al. 2013), a combination of two weather variables is examined in this study: mean temperature (0, +1, +2, and +3 °C) and mean precipitation (0, +10, and +20%). Eleven total climate change conditions, including the current condition (0,0,0), are applied to the calibrated SWAT model. Flood indices were calculated on 18 years of data because the first 2 years are considered to spin up the SWAT model. The historical record is used as the baseline for comparison within the climate sensitivity datasets. Several studies (i.e., Yang et al. 2016; Xu et al. 2017) used climate sensitivity analyses for uncertainty analysis in climate change impact. The set of possible future temperature and precipitation conditions allows us to determine which factors and range of values affect environmental outcomes (Pianosi et al. 2016).

Based on simulated streamflow, using available observed data, 95%PU analysis is performed in Sequential Uncertainty Fitting Version 2 (SUFI-2) of SWAT-Calibration and Uncertainty Programs (SWAT-CUP) interface and taken as baseline uncertainty bound. For the climate model approach, the ensemble of five RCMs is selected for optimistic case emission scenarios (RCP4.5) and business-as-usual emission scenarios (RCP8.5) based on their general performance in the study area and region. For each model, I selected historical (1981–2000) and futures at mid and last 21st century (i.e., at 2041–2060 and 2081–2100) conditions for comparison. In both future climate approaches, only temperature and precipitation inputs changed in the SWAT model to automatically generate the other three weather inputs: solar radiation, relative humidity, and wind speed, as during the historical calibration and verification periods. As observed data, streamflow was simulated using historical time series of RCMs in Arc SWAT. If simulated streamflow values lie within the range of baseline uncertainty bound (95PPU) value obtained using observed climate data, the candidate RCMs used in the attribution process of flood change in the study area. If the simulated discharge falls outside the uncertainty band, the climate model is rejected from candidate RCMs for future climate impact analysis.

Climate models meeting the above criteria were assumed to be suitable as the ensemble in the climate change impacts on hydrology. An ensemble means of discharges and other water balance components simulated using the selected climate models can be used to assess the projected climate change impacts on the flood events of the Wabi Shebele River Basin.

The SWAT model is a watershed-scale and physically based distributed hydrological model (Neitsch et al. 2005; Abbaspour et al. 2007) developed to simulate the impact of land management practices on hydrology and water quality under complex watersheds with heterogeneous soil and land use conditions. In recent decades, it has been widely used for water cycle simulation and water resources management, especially for the analysis of streamflow variation under climate change and LULC change (Schulze 2000; Adamu 2014; Camici et al. 2014; Guo et al. 2016; Näschen et al. 2019; Gaur et al. 2020). In this study, the hydrologic model is used to see the impact of climate variables change on streamflow in the Wabi Shebele basin. The optimum parameters of the SWAT model is determined by sensitivity analysis, which assesses the sensitivity between a parameter and other parameters in different areas. Based on parameters available for water production identified by Arnold et al. (2012) and preliminary identification in the SWAT model, SWAT-CUP global sensitivity analysis is conducted to select the most sensitive parameters for watersheds. The p-value and t-statistic were used to eliminate non-sensitive parameters from the calibration process. The higher the absolute value of t-stat and the smaller value of p, the more sensitive is the parameter (Abbaspour et al. 2007; Moreira et al. 2018).

Six flood variables were extracted from the simulated discharges under different climate change and LULC change conditions. The impact levels of climate change and LULC change are analyzed in each index. These flood indices are: Annual maximum discharge (AMAX), Peak over threshold (3rd quartile) frequency (POTF), Peak over threshold (3rd quartile) magnitude, Seasonal peak discharge for winter (SMW), spring (SMSp), and summer (SMSu) used to define the extreme river flow.

To estimate flood hazard, two flood indices extracted from simulated daily flow rate at each sub-basin outlet: flood exceedance probability index (FEPI) and flood frequency index (FFI). These indices represent the duration and frequency of flood hazards. Flood characteristics can be derived from time series of observed or simulated hydro-meteorological variables using a user-defined threshold level. A peak over the threshold of a 2-year return period for wet season streamflow was used in all indices because it was used as a proxy to bank full discharge and threshold for flood events in past studies (BCEOM 1973; IWMI 2015).

The results indicate that most of the observations with different parameters are within a bracket of the 95PPU (i.e., p-factor ≥0.5 and r-factor <1.5), signifying SUFI-2 captures the model behavior. In terms of R² and NSE, the simulations give the value related to 0.6, which indicates that the SWAT model looks better for the prediction of discharge in the Wabi Shebele River Basin (as per Abbaspour et al. 2007), and the final parameter ranges were the best solution obtained for the basin. Most of the observed values obtained during the calibration and validation were within the boundaries of 95PPU, which indicates that SWAT model uncertainties were falling within the permissible limits. The study also approves the result obtained by Shawul et al. (2013) that the SWAT model performed well for simulation of monthly streamflow in Shaya mountainous watershed, southeastern Ethiopia.

The daily precipitation and temperature simulations of the climate models from the CORDEX-Africa, RCMs datasets are averaged over the watershed area, and their performance is evaluated using statistical parameters. Table 2 shows the absolute values and differences of RCMs from the meteorological observations of these parameters.

For all of the CORDEX-Africa RCMs, bias-corrected datasets using the meteorological observation overestimated the number of rainy days (Table 2). Although the differences in values to the gauge observation are considerably less, almost all the CORDEX-Africa RCMs underestimated average daily precipitation and standard deviation. However, most RCMs, except GFDL-ESM2M/RCA4 and MIROC-MIROC5/RCA4, overestimated maximum daily rainfall. The regional climate models such as GFDL-ESM2M/RCA4 and MIROC-MIROC5/RCA4 showed underestimation of extreme values of the average maximum daily precipitation. Similarly, the study conducted by Teshome & Zhang (2019) indicated that seasonal rainfall downscaled from RCMs over Ethiopian watersheds showed a decreasing situation and significant variability.

All bias-corrected RCMs on temperature showed similar to average maximum temperature with ground-based observed maximum temperature in the study area (Table 2). However, extremely high temperatures decrease from the observed temperature in CanESM2/RCA4, CM5A-MR/RCA4, and CNRM-CM5/RCA4 while indicating the increment in GFDL-ESM2M/RCA4 and MIROC-MIROC5. Furthermore, the standard deviation of maximum temperatures is estimated underestimation in all RCMs.

Uncertainties of climate change impact on hydrology arise from different sources: data quality, climate models and emission scenarios selected, the downscaling method used, and the hydrological model applied. Data scarcity and reliability to calibrate the hydrological model is the first source of uncertainty in the analysis. To minimize this uncertainty, checking and filling missed data using weather generators were carried out before the calibration, and validation of the hydrological model in this study.

Climate model uncertainty arises due to the response of different climate models to produce dissimilar changes in climate in the presence of the same radiative forcing. Relative to GCMs outputs, RCM output reduces uncertainties since it gives high-resolution climate information and provides a better description of orographic, land surface contrast, and land surface characteristics. Scenario uncertainty arises due to imperfect knowledge of the external factors affecting the climate system, e.g., future emission of greenhouse gasses. There are four groups of individual scenarios developed by the IPCC in 2014 to supersede SRES (IPCC 2014). Each scenario consists of a specific radiative forcing projection and makes assumptions about future population, GDP, and energy use, based on the radiative forcing. Hence, choosing among the scenarios also adds to the uncertainty.

The assumptions involved in the hydrologic model simulations are also a portion of the uncertainty. As described in the Method section, the determination of the impacted streamflow is only based on the precipitation and temperature changes in the future. The other climatic variables such as wind speed, solar radiation, and relative humidity were assumed to be constant throughout the future simulation periods. Even though it is definite that land use changes in the future, it is also assumed constant. But these assumptions can lead to a certain level of additional uncertainty. Therefore, all types of uncertainties discussed above propagated on the future predicted discharge volume. Thus, the uncertainty presented in the model and model outputs kept on cumulating while progressing toward the final result of the study.

The spatial distribution of all flood indices for the baseline case (T + 0 °C, P + 0%) shows greater flood hazard in the downstream part of the Wabi Shebele river (Table 3). Relatively higher flood indices were observed in the upstream gauging station at Dodola and the downstream gauging station at Gode. High flood indices in the upper gauging station may be due to consecutive days of rainfall in the northwestern highland part of the study area (MoWR 2003). However, high flood indices at the downstream gauging station may be explained by drainage size, where most of the main tributaries contribute flood discharges to this sub-basin. Drainage size is one of the major driving factors for flood discharge increment confirmed in different studies (e.g., Al-Rawas & Valeo 2010; Huang 2020). The spatial distribution of flood hazard analysis under baseline conditions approves that the SWAT model captured the impact of drainage size on stream discharge. Therefore, the downstream area of the Wabi Shebele River Basin could be a focus area for flood mitigation.

When precipitation increased by 10 or 20%, almost all sub-basins saw flood index increases from the baseline. Warmer temperatures caused a decrease in water yield, which counteracted the increase in precipitation. Therefore, in the sub-basins more sensitive to temperature change, flood indices were lower compared to the baseline scenario despite precipitation increases. Exceptionally in the upper sub-basin at Wabi Dodola station, flood exceedance probability exhibited an increasing tendency when temperature increased, indicating that headwaters were more sensitive to higher temperatures in terms of these two components of flood hazard. However, for both downstream gauging stations, Wabi at Legehida and Wabi at Gode, the level of change seems to be consistent across the watershed, suggesting that precipitation change has more impact on flood indices than temperature. This led us to the conclusion that more precipitation tended to perpetuate flood hazard in the study area, while warmer temperatures reduced flood hazard in middle and downstream areas.

It is evident from Figure 8 that under the RCP8.5 emission scenario, the magnitude of streamflow increases in 2081–2100 when all others estimated below the observed scenarios. FDCs at upper gauging stations at Dodola and Legehida become steep slopes, while FDCs at Wabi at Gode station show relatively flatter slopes during Q30–Q90 (medium flows). A flat curve indicates that groundwater contributions to the stream reach are significant that sustain the flow throughout the year (Chambers et al. 2017; Gaur et al. 2020). Additionally, FDCs at Gode station show a considerable increase in high flows (10% exceedance) during 2081–2100 relative to observed discharges (1981–2000).

A 2-year return period flood value is used as a threshold level to estimate flood characteristics during the reference period 1981–2000 and future climate scenarios. Two flood indices were calculated from extracted flood values and compared between different scenarios. As presented in Table 5, most of the flood simulated under future climate scenarios estimated flood indices below the reference period, except flood in the upper basin at Wabi at Dodola station, in which FEPI increases in most of the future climate scenarios. Furthermore, the FFI at Wabi Gode station shows a considerable increment between 2081 and 2100 in both scenarios.

The influence level of LULC and climate change on streamflow is estimated using the separation method. As presented in Table 6, the response of the discharges to climate change is higher than that of LULC change in the Wabi Shebele basin. However, LULC change also has a significant impact in middle and upper watersheds like Wabi at Legehida, Wabi at Dodola, Maribo, and Robe. Annual maximum discharge (AMAX) decreases in watersheds where forest and shrubland coverage increase in the study period. For instance, in Wabi at Legehida and Erer watersheds, the magnitude of floods decreases while the coverage of forest increases in condition one. In watersheds like Wabi at Dodola, Maribo, Robe, and Wabi at Legehida, flood discharges estimated using LULC of 2016 are more than flood estimates using LULC of 1986 by 3.91, 2.33, 1.92, and 128.66 m³/s, respectively. As a result, flood magnitude increases by 0.18, 1.83, 0.57, and 0.44% in watersheds. In Wabi at Gode watershed, flood discharge under condition one is more than flood magnitude in condition two by a value of 1,285.18 m³/s, which is contributed by climate change and LUCC, accounting for 105.12 and 5.12%, respectively. The results indicated that climate change as the main influential factor on the streamflow and flood values in the Wabi Shebele River Basin between 1980 and 2010, which is similar to the conclusion drawn by Akola et al. (2018).

In general, this study discussed the impact level of climate factor on flood formation in the Wabi Shebele River Basin, one of the largest flow basins in Ethiopia. The result revealed that climate change is the main driving force for flood hazards occurred in the study area. The impact of climate change on flood hazard is indicated to be higher in north western upstream and downstream lowland gauging stations at baseline scenario (i.e., with no change in temperature and precipitation, T + 0 °C, P + 0%). The Wabi Shebele River Basin is likely to experience an increase in flood hazard with an increase in precipitation in the future as temperatures increase less than 2 °C. This result strengthens the study result by Shiferaw et al. (2015) that flow in tropical river basins exhibits typical characteristics of tropical rainfall regimes. In such a case, where climate change is the leading cause of high hydrologic variability, development-based climate change adaptation mechanisms and flood risk management strategies are needed.

This paper addressed two main works: the first is to analyze the potential impacts of climate change on streamflow simulation along with quantification of projected flood characteristics, and the second is related to the quantification of uncertainties in streamflow projections. A semi-distributed hydrological model (i.e., SWAT) is used to simulate, analyse, and account for the spatial variability of streamflow. Model calibration/validation and parameter sensitivity analysis are performed through the SUFI-2 algorithm in SWAT-CUP. Model uncertainty analysis is done to establish the uncertainty bounds of the model using observed meteorological data, which is taken as a boundary limit to evaluate the significance of the climate change impact. The overall calibration and validation of the SWAT model were good in the basin except for Wabi at Gode station. At Gode station, the percentage of the simulated data within the uncertainty bound is only 28%. But for the other two sub-basins, the percentage of simulated flow within the uncertainty limit is more than 48%.

The model produced similar patterns of change in flooding due to temperature, and precipitation driven either by RCMs or climate sensitivity tests. All simulated discharges using RCMs fall within 95% probable uncertainty bands. Two flood indices are used in analysis, showed higher risk in lower basin areas (i.e., at Gode stations) and lower risk in the middle part of the basin areas at baseline scenario (T + 0 °C, P + 0%). The Wabi Shebele River Basin is likely to experience an increase in flood hazard with an increase in precipitation in the future as temperature increased less than 2 °C. When the precipitation increased by above 20%, flood hazard was most likely escalated in sub-basins. The influence level of LULC change and climate change on streamflow analyzed using the separation method indicates that climate change is the main factor influencing the streamflow and flood values in the Wabi Shebele River Basin. However, LULC change has a significant impact in middle and upper watersheds like Wabi at Legehida, Wabi at Dodola, Maribo, and Robe. These findings could provide information on extreme weather events and early warning alarms.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.