Evaluation of four bias correction methods and random forest model for climate change projection in the Mara River Basin, East Africa

The high-resolution global climate models (GCMs) and regional climate models (RCMs) are essential for understanding the global climate change variability and mitigation (Rauscher et al. 2010; Oo et al. 2020). As a major vulnerable region of climate variability, Africa is the first continent selected by the World Climate Research Program (WCRP) CORDEX (coordinated regional climate downscaling experiment) to generate an ensemble of high-resolution climate projections. The output of CORDEX evaluates the GCMs to understand the local climate variability and changes through the downscaling method. In general, it establishes coordination among regional climate downscaling for producing the GCMs (Dosio & Panitz 2016). Thus, downscaling the various GCMs output simulation is important for local climate change impact studies (Worku et al. 2020). Also, it is vital to select a better downscaling approach for GCMs simulation at the local scale to reduce the uncertainty.

Simulating the African climate models using the GCM is challenging due to complex and diverse processes, and the models have difficulties simulating the current climate (Laprise et al. 2013). Another factor is scale issues in the simulation process due to low spatial resolution 250 km × 250 km, respectively (Alemseged & Tom 2015), although GCM grid resolution (2°) is not a suitable approach for representing climate variables such as precipitation and temperature at regional climate (Dobler & Ahrens 2008; Urrutia & Vuille 2009; Lafon et al. 2013; Fiseha et al. 2014; Teng et al. 2015; Roy et al. 2018). Through the better representation of regional climate, regional climate models (RCMs) provide high-resolution (50 km × 50 km) climate simulation and variability (Urrutia & Vuille 2009). Most of the previous studies proved that the output of CORDEX domain RCMs could reproduce the feature of African climate (Kim et al. 2014; Dosio & Panitz 2016; Favre et al. 2016; Pinto et al. 2016). In general, very few studies simulated the CORDEX domain RCMs at a local scale in the African continent (Hernández-Díaz et al. 2013; Laprise et al. 2013; Alemseged & Tom 2015; Dosio & Panitz 2016). In addition, the evaluation of CORDEX shows a reasonable standard and is used in various climate change impact studies of Africa (Worku et al. 2020). However, the downscaled RCMs cannot be used directly without bias correction for climate change impact studies due to systematic error and the reliability of the simulation.

Bias correction is a scaling method of climate model variables with observation values (Gudmundsson et al. 2012; Lafon et al. 2013). Although various bias correction methods have been developed and used in numerous studies to reduce the RCMs simulation biases, in general, most of the existing studies only focus on projecting the future climate variable but not for bias correction methods (Kim et al. 2019). Tschöke et al. (2017) evaluate two bias correction methods – gamma distribution and power transformation (PT) for precipitation simulation; they found that gamma distribution methods are more efficient in reducing the error from RCMs. Worku et al. (2020) evaluate four bias correction methods such as distribution mapping (DM), linear scaling (LS), variance scaling and PT for adjusting the RCMs temperature and precipitation with observation and developed future climate scenarios. The study proved that the DM correction method is more effective in adjusting the frequency of temperature and rainfall. Moreover, very few studies utilize the statistical metrics (e.g. Räty et al. 2014; Worku et al. 2020) to find the suitable bias correction method that has not yet been tested in the Mara River Basin (MRB). In addition, there are some limitations on various bias correction methods due to its strong dependency on observation data (Teng et al. 2015). However, bias correction methods factor in various climatic studies for reducing the error from climate models. Therefore, it is important to investigate the strongest bias correction method from RCMs output before developing the future climate change scenario.

The bias correction approach cannot correct the non-stationary RCMs biases and uncertainties (Wang et al. 2018). The machine learning (ML) approach for an ensemble of various GCMs/RCMs is selected in numerous studies to reduce these uncertainties (Tripathi Srinivas & Nanjundiah 2006). Campozano et al. (2016) compared two ML approaches – artificial neural network (ANN) and least square support vector machines (LS-SVM) – and both methods performed equally for downscaling the monthly precipitation. Crawford et al. (2019) used six ML approaches, including random forest (RF) and SVM and linear regression (LR), to evaluate the performance of ensemble coupled model intercomparison project phase 5 (CMIP5); they found that the RF method performed better than other approaches to simulate temperature. In a similar context, Wang et al. (2018) compared four ML methods (RF, Bayesian model averaging, the arithmetic ensemble means (EM) and SVM) for reproducing the observed monthly temperature and rainfall. The output of this study indicates that the RF ensemble method performs well and can reproduce monthly climatic datasets. However, no published work is found for the application of ML to reproduce the future rainfall and temperature datasets from the bias-corrected multiple RCMs ensemble. This study considers an RF ML approach to reproduce the future rainfall and temperature.

The MRB hydrologic system has been selected to evaluate the various bias corrections and an ML method using RCMs from CORDEX. Dessu & Melesse (2012) evaluate the five GCMs for climate change assessment at the MRB using a single bias correction method. Similarly, Roy et al. (2018) used the CIMP5 GCM to develop the future climate change scenario on the hydrological application of the MRB. However, most studies were based on the GCM instead of the RCM, and very few used RCMs at the basin scale, although dissimilar topography and various extra-terrestrial climatic phenomena impact the MRB rainfall and temperature pattern. For example, the rainfall of MRB is strongly influenced by the intertropical convergence zone (ITCZ) and Indian ocean surface temperature. In addition, these rainfall variabilities are also controlled by Eastern Pacific Ocean surface temperature, ENSO (El Nino Southern Oscillation) and La-Nina conditions (USAID 2019). Furthermore, it is important to understand the complex interaction of the climate component of the region (Laprise et al. 2013). Therefore, high-resolution simulation from RCMs is needed to improve climate change assessment representation on water resources and agriculture at the MRB.

The main objective of this study is to evaluate the performance of four bias correction methods in adjusting the CORDEX domain RCMs at the regional scale of Africa. This evaluation mainly finds the most effective bias correction method using statistical metrics and cumulative distribution function (CDF) derived from the historical RCMs. It assumes that the most effective bias correction method from historical RCMs performed better and developed the future climate change assessment. Therefore, the best bias-corrected ensemble of RCMs was used to generate future climatic scenarios at the MRB. In addition, this study tested an ML (RF) approach based on the bias-corrected RCMs output to reproduce the future rainfall and temperature at a basin scale. Furthermore, the RF model determines the most important variable from each RCM. Moreover, to evaluate the performance of the bias correction method, the RF model was validated using various statistical metrics.

The MRB shares the transboundary between two African countries, Kenya and Tanzania, with latitudes −0.331573°S to 1.975056°S and longitudes 33.88372°W–35.907682°W (Figure 1). 65% of the upper MRB portion covers Kenya, and the remaining 35%, the lower part, is located in Tanzania (Mati et al. 2008). The Mara river (395 km) originates from the Mau Forest in Kenya at 3,000 m altitude and has fed into the Victoria Lake in Tanzania (Mango et al. 2011). The main tributaries of the MR are Nyangores and Amala, and other sub-tributaries are Engare Egito, Talek, Kenyo and Tambura. The river basin is affected by agriculture activity, deforestation, mining and other human activity.

The seasons of MRB are influenced by the movement of ITCZ and dry north-eastern wind, which comes from the Sahara Desert. The annual rainfall pattern of the MRB varies from 1,000 to 1,750 mm at the upper catchment and 700–850 mm at the lower catchment (around Musoma), which falls between April–September and November–December (Mati et al. 2008). The rainfall pattern in March to June is influenced by southeast trade winds from the Indian Ocean (Dutton et al. 2018). The average temperature of MRB is 25°C, where the maximum temperature recorded is 28 °C in the daytime and a minimum 10 °C at night. The climate is warm in the daytime and cool at night, which also depends on elevation. The temperature increased from October to February and slightly decreased from June to August.

The historical observation of daily rainfall and temperature data was obtained from four meteorological stations in the MRB, and it is available at https://geographic.org/global_weather/. Thus, the website creates a database to store historical weather station datasets for the world based on the National Centres for Environmental Information (NCEI) (https://www.ncdc.noaa.gov/). One major problem in simulating RCMs in the African region is the lack of observation data availability, especially for the regional level, such as the river basin. It is tough to obtain African historical data from 1900; however, the historical datasets are available in the studied region from 1980. Therefore, this study considers historical observation datasets during 1986–2005 (20 years) to evaluate the various bias correction methods and develop future climate projections.

This study obtained six historical and future RCMs from CORDEX projected to evaluate the bias correction method followed by Favre et al. (2016). The RCMs datasets are obtained from the Lawrence Livermore National Laboratory (LLNL) web portal powered by the Earth System Grid Federation (ESGF) (https://esgf-node.llnl.gov/search/esgf-llnl/). In this study, historical (1986–2005) and future (2026–2095) simulation of daily rainfall, maximum temperature (TMAX) and minimum temperature (TMIN) of ICHEC-EC-EARTH, CCCma-CanESM2 and MOHC-HadGEM2-ES models driven by six CORDEX RCMs (Table 1) were considered for bias correction and the RF model. Moreover, simulated and observation data vary in elevation and need bias adjustment in various locations (Worku et al. 2020). After retrieved the observation and RCMs rainfall and temperature, the data proceeded with a grid resolution of 0.44°×0.44° using the bilinear interpolation method to match the grid resolution of each RCMs. In addition, RCP 4.5, RCP 8.5 and RCP 2.6 emission scenarios were used to represent the radiative forcing level (Javadinejad et al. 2020; Oo et al. 2020). RCP 4.5 shows the intermediate emission levels of 4.5 W/m², while RCP 8.5 W/m² demonstrates the high radiative forcing with the emission level of 8.5 W/m², which could be stabilized after 2100 (Worku et al. 2020). Furthermore, RCP 2.6 is selected for considering the Paris agreement and sustainability for future climate change.

The DM technique is based on different approaches such as probability mapping (PMAP), empirical CDF mapping, quantile mapping (QMAP) and Kernel density distribution (KDDM) (McGinnis Nychka & Mearns 2015) and is used in various climate simulation studies (Lafon et al. 2013; Olsson et al. 2015; Switanek et al. 2017). The main principle of this method is to fit the transfer function of RCMs simulation with an observational mean value (Zhang et al. 2018). Mainly there are two functions suitable for precipitation and temperature – gamma and Gaussian distribution (Worku et al. 2020) – although the histogram of bias-corrected RCMs and observation values has properties of positive skewness (Tschöke et al. 2017) to adjust the mean value, extreme, standard deviation (SD) and distribution of precipitation.

RF is a classification approach used in various predictive climate change assessment studies. This method mainly consists of a decision tree approach where each classifier is developed from the bootstrap sample (Gaál 2012). The bootstrap sampling makes RF less susceptible to overfitting than a decision tree (Wang et al. 2018), although the decision tree methods are developed from partitioning the training data (y_i, x_i), i=1, ………, n, which is based on the independent variable (Crawford et al. 2019). However, the RF model contains a regression tree, which consists of many regression trees like a forest and obtained the prediction model based on the two parameters: (1) the n number of regression tree (n_tree) and (2) the selected evidential feature at each node (m_try) (Wang et al. 2018). Breiman (2001) and Segal (2003) have more details of the RF computation process.

In this study, we set the number of n_tree ranges from 500 to 100 and selected evidential feature m_try ranges from 5 to 20 for reproducing the future rainfall and temperature followed by Wang et al. (2018). The six bias-corrected RCMs output are considered a predictor variable in RF and used to determine the rank of relative impotence of each predictor, which may suggest the most important RCMs for contributing to reproducing the future climate scenario.

The CMhyd tool was used to process the four bias corrections methods (LS, DC, LI and DM) and retrieved the rainfall and temperature datasets from the RCMs (Zhang et al. 2018). However, the CDF is used to adjust the simulated and non-simulated rainfall and temperature datasets, then compared to the observational values (Figure 2) (Maurer & Pierce 2014). In addition, four statistical matrices (mean, SD, coefficient of variation (CV) and 90th percentile) are selected for evaluating the bias correction methods from historical RCMs and observation data. This evaluation is used to obtain the suitable bias correction method to obtain the future climate change scenario. The strongest bias correction method was applied for developing the near-term (2026–2045) and long-term (2076–2095) climate scenario at the MRB. Furthermore, the performance of the various bias correction methods and the RF model was validated using four evaluation metrics such as (a) the coefficient of determination (R²), (b) mean square error (MSE), (c) mean absolute error (MAE) and (d) root mean square error (RMSE).

The African continent is not suitable for a large global and regional model of precipitation change with different aspects (Déqué et al. 2017). In this study, results suggest that the four bias correction methods, LS, DC, LI and DM, are effective for adjusting the mean annual RCM simulation of rainfall. However, a significant difference was found between bias-corrected RCM and raw RCM for the spatial distribution of rainfall in the MRB (Figure 3). Also, the CDF results indicate the overestimation and underestimation of the annual rainfall were adjusted sufficiently in the entire sub-basin using LS, DC, LI and DM methods (Figure 2) (Worku et al. 2020). These four methods have comparable performance in adjusting the mean annual observed rainfall and RCMs output of the MRB (Figure 3). But the significant distribution of the total amount of rainfall was overestimated and underestimated depending on the changes in elevation (Gudmundsson et al. 2012).

The seasonal variation of precipitation derived from the bias-corrected RCMs shows comparable performance at a regional scale (Urrutia & Vuille 2009). In general, the mean monthly rainfall of simulated RCMs is adjusted through observed precipitation. The seasonal distribution of rainfall from bias-corrected RCMs and observation output provides the maximum peak of rainfall in the summer month of MJJ (May, June and July) and winter months of NDJ (November, December and January) (Figure 4(a)). Similarly, the LS and DC methods were found inadequate to adjust the CV of rainfall with an average value of 1.5, which is overestimated compared to the observation (Figure 4(d)). In the same contrast, LI and DM show good strength for CV with a value of 0.8 to adjust the rainfall from individual simulated RCMs. The rainfall pattern concerns the SD from various computed RCMs simulation (Figure 4(b)) at individual grid (Meehl Arblaster & Tebaldi 2005), although simulated CCCma-CanESM2 RCMs shows the maximum peak in SD compared to other simulated RCMs for rainfall and shows the poor performance with a MSE value of 204.33 mm (Table 2). In addition, the 90th percentile of precipitation indicates a significant result between the DM-corrected RCM and the observation value (Figure 4(c)). However, the LI and DM methods show similar results in adjusting the annual mean rainfall corresponding to observation values.

Other studies also suggest a comparable performance among various bias correction methods in adjusting the rainfall and temperature values (Liang 2004; Räty et al. 2014; Tschöke et al. 2017; Wang et al. 2018). In addition, there is a limitation in various bias correction processes at individual RCMs, which is mainly based on resolution (Rauscher et al. 2010). However, LS, DC and LI methods shows overestimated or underestimated rainfall variation in the statistical framework. Therefore, the DM bias correction method was selected to compare the performance of individual RCMs output against observation output, followed by Worku et al. (2020).

Four evaluation matrices (R², RMSE, MAE and MSE) were selected to evaluate the performance of bias-corrected individual RCM and the EM of six RCMs. The bias-corrected DMI-HIRHAM (R²=0.35) and HadGEM2-ES (R²=0.26) RCMs showed excellent performance to reproduce the rainfall and temperature compared to observation values (Table 2). In addition, RMSE error values of these two RCMs were 12.94 mm (DMI-HIRHAM) and 11.84 mm (HadGEM2- ES), respectively, although the bias-corrected KNMI (ICHEC-EC-Earth) RCMs output showed poor performance compared to observation values with the RMSE of 10.54 mm and the MSE of 111.2 mm. However, the ensemble of all bias-corrected RCMs showed excellent performance with a correlation coefficient (R²) of 0.4 and a minimum value for RMSE, MAE and MSE (3.43, 2.32 and 11.78). Therefore, this study selected DM-corrected ensemble of six RCMs output to develop the short-term (2026–2045) and long-term (2076–2095) future climate under the RCP 4.5, RCP 8.5 and RCP 2.6 scenario.

The future rainfall and temperature pattern in the MRB were projected using the ensemble RCMs output (Figure 5). The individual RCMs of rainfall results were compared to historical simulated RCMs (Table 3). However, the future mean annual rainfall and temperature were categorized into short-term (2026–2045) and long-term (2076–2095) based on the emission scenarios of RCP 2.6, RCP 4.5 and RCP 8.5. The spatial variation of rainfall for the near-term (2026–2045) and long-term (2076–2095) is shown in Figure 6. The results from the ensemble RCMs demonstrate the increasing trend of future rainfall variation under various RCP scenarios such as RCP 4.5 (+110.45%), RCP 8.5 (+ 108.87%) and RCP 2.6 (+134.28%) (Table 3) in the near term (2026–2045). Similarly, the long-term (2076–2095) climate of rainfall trend increased +114.32, +123.305 and +121.57% under the emission scenario of RCP 4.5 and RCP 8.5. But the highest increased rainfall recorded based on the ensemble RCM output in the near-term future (2026–2045) with the amount of +123, +117 and +115% under the emission scenarios of RCP 4.5, RCP 8.5 and RCP 2.6. In addition, a slight increasing rainfall trend was found (13–23%) for REMO (MPI-ESM-LR) RCMs in the near-term and long-term scenario. Moreover, all the RCMs output and EM suggest the increasing rainfall trends for short-term (2026–2045) and long-term (2076–2095) climate in the MRB.

Most of the individual RCMs showed an increasing trend in annual rainfall (Table 3), except simulated REMO output (Worku et al. 2020). Other studies show similar rainfall trends for the future scenario in the MRB. For example, Dessu & Melesse (2012) found a 30–50% increase in the rainfall trend during the wet season (MAM) using five GCMs output for the year 2050. Also, this study's results suggest a significant increase in the water flow of MRB from 2046 to 2065 and 2081 to 2100. Roy et al. (2018) projected 31 CMIP5 RCMs output to evaluate the daily precipitation and found an increased rainfall trend in the MRB for the short-term future scenario (2020–2050). Similar studies carried out by Mellander et al. (2013) showed the +6% increased rainfall pattern in the upper Blue Nile basin based on the downscale of ECHAM5/MP1-OM during 2050–2100. In addition, Fenta Mekonnen & Disse (2018) reported the downscaled annual mean precipitation increased from 2.1 to 43.8% under A1B, A2, and B1 scenarios for 2006–2100 using the LARS-WG model.

Similarly, maximum temperature (TMAX) and minimum temperature (TMIN) showed an increasing trend in the near-term (2026–2045) and long-term (2076–2095) climate under the emission scenarios of 2.6, 4.5 and 8.5 (Figures 7 and 8). However, the ensemble of temperature output have shown the rise of TMAX 1.31 °C and TMIN 2.66 °C under the RCP 4.5 scenario and an increasing trend of TMAX 1.85 °C and TMIN 3.6 °C under the RCP 8.5 scenario during the near-term period (2026–2045). The maximum increase trend from the ensemble of TMAX (11.81 °C) and TMIN (15.12 °C) has been recorded in the long-term (2076–2095) scenario based on the RCP 8.5. The lower emission of increased TMAX and TMIN was projected under the RCP 2.6 scenario (Table 4). The spatial pattern of future TMAX and TMIN is shown in Figures 7 and 8 based on ensemble RCMs and observed values. IPCC projected a similar condition of an increased trend in East Africa (IPCC 5th assessment), which is estimated to change from 0 to 4 °C warmer by 2050 under the RCP 4.5 scenario (Solomon et al. 2007). The other studies also found a similar increasing trend in the MRB. For example, USAID (2019) reported the average temperature of MRB would increase by 1.97 °C (2030) and 2.71 °C (2050), which is a similar output of the ensemble near-term (2026–2045) and long-term future (2076–2095) under the RCP 4.5 scenario (Table 4). Therefore, the emission scenario of RCP 4.5 and RCP 8.5 was selected for the RF model to reproduce the rainfall and temperature.

The study used EM of six observed RCMs output in the RF model to reproduce future climatic datasets. For developing the RF-based future climate scenario (precipitation and temperature), RCP 4.5 and RCP 8.5 were selected during the period (2006–2100). Most previous studies used 70% of datasets as training and the remaining 30% for testing in various ML models (Tripathi et al. 2006; Anandhi et al. 2012; Gaál 2012). Therefore, 70% of datasets (2006–2069) are considered training data to establish the relationship between observed EM and predicted EM and the rest of the 30% datasets (2070–2100) are used for testing. Four evaluation parameters (R², RMSE, MSE and MAE) were used to compare the performance of RF-based predicted EM against observed EM (Table 5). The results indicate that the RF model can predict and reproduce the future precipitation and temperature from each RCM. All the RF model-based EMs provide best agreement with observed EM, and each RCP shows an R² value of more than 0.90, which indicates excellent performance. The results also demonstrate the small error with the RMSE of 0.26 °C on average temperature and 0.45 mm on average rainfall.

Figure 8(a) and 8(b) shows the functional relationship between observed EM and predicted EM using the RF model during the testing period (2070–2100). RF-based predicted EM RCP 4.5 is similar to the observed EM for precipitation (Figure 8(a)), Tmax and Tmin (Figure 8(b)). The maximum difference between predicted EM and observed EM datasets is larger during the RCP 8.5 scenario for temperature and precipitation. Also, a precise degree of scattering was observed between estimated EM and predicted EM (Figure 8(c) and 8(d)) during the testing period, which indicates that the RF model can reproduce future rainfall and temperature datasets. The performance of each RCM was evaluated based on their relative importance shown in Figure 8(e) and 8(f). The relative performance of each RCM is automatically assessed in the RF model based on their variance of importance (Wang et al. 2018). The results suggest that the HadGEM2_ES RCM had more important variables in the RF model for production precipitation and CLMcom_CCLM for temperature, including RCP 4.5 and RCP 8.5 scenarios.

The study evaluates various bias correction methods from six RCMs (KNMI, CLMcom – CCLM, CLMcom – CCLM, REMO, CanESM2 and HadGEM2- ES), obtained from three CORDEX (EC – Earth, CCCma and MOHC). The uncertainty of the observation and individual bias-corrected RCMs pattern was captured from monthly mean rainfall and temperature (Figure 9). This study did not trigger the climate change of future rainfall (Räty et al. 2014). The bias-corrected monthly rainfall from RCMs showed a comparable increase trend in MJJ and NDJ. The EMs of RCMs and individual RCMs showed uniformity in future climate change scenarios with different magnitude. But strong similarities were found in observation and EMs rainfall. Furthermore, HadGEM2 RCMs projected the highest increasing rainfall trend under every emission scenario. The individual RCMs showed various rainfall patterns in the sub-grid process, which depends upon the grid resolution.

In this study, the results demonstrate the rising rainfall pattern in the near-term and long-term climate under the emission scenario of RCP 4.5 and RCP 8.5. However, this emission scenario described the changes in rainfall that may trigger the convective activity and moisture of the atmosphere over the MRB, although various regional climatic phenomena and local forcing are also the major driving factors for rainfall in the MRB, East Africa. Also, the changes in ITCZ could impact the rainfall variation in the near-term and long-term future. For considering the Paris agreement, the future temperature and rainfall are also investigated under the RCP 2.6 scenario. The RCP 2.6 scenario showed a lower percentage of rainfall increase in the MRB. KNMI (EC- EARTH) and CCLM (EC-EARTH) showed the lower increased rainfall rate under the RCP 4.5 and RCP 8.5 scenario. But the REMO (EC-EARTH) showed a slight increase in the rainfall pattern under the RCP 4.5, RCP 8.5 and RCP 2.6 scenario. The importance of large uncertainty was found in various RCMs, which is comparable to a different scenario. However, it found that the global emission will not be reduced by 2030 (10%), which could not be achieved according to the Paris climate agreement (Déqué et al. 2017). ̀

In a similar contrast, the TMAX and TMIN projected high emissions in the near-term and long-term future. The long-term future (2076–2095) scenario showed the high emission of greenhouse gases, increasing the temperature under the RCP 4.5 and RCP 8.5 scenario. But the lower increase rate is observed under RCP 2.6. The IPCC report also contributed that the emission scenario is a major driver of global temperature attributed to global climate change.

In addition, the RF method reduced the uncertainty of the bias-corrected ensemble RCM output and reproduced the future climate variable. The relative importance of each RCM for developing the future rainfall and temperature is beneficial for understanding the contribution of each model adding the value to the ensemble (Wang et al. 2018).

The bias correction method is a major pre-processing step to investigate the climate variable at a regional scale from RCMs results. This study tested four bias correction methods (LS, DC, LI and DM) to adjust the rainfall and temperature retrieved from six historical RCMs. These methods were evaluated in a statistical metrics framework to project future rainfall and temperature at the basin scale. Our results suggest that the DM method showed the best performance in adjusting the rainfall and temperature based on the historical period. Therefore, the DM method was applied to generate future climate scenarios at the MRB. The future climate scenario from the individual RCMs and all EM of RCMs revealed that wetter conditions and high temperature characterize the future climate of the MRB. In a similar context, the DM method can analyze the climatic extremes such as flood and drought and design new policies to reduce the impact of extreme events. In addition, to reduce the uncertainties of the EM of RCMs, we used an ML approach, namely RF. Thus, the method was applied to develop the most robust ensemble-based future climate scenario. The RF ensemble provides comparable results based on the evaluation matrices and perfectly reproduces future rainfall and temperature, although the relative importance was assigned based on the performance of each RCM in RF ensemble simulation. However, this study used DC, LS and LI methods to adjust the RCMs simulation and develop future climate scenarios. Moreover, it will assist in future water resource management and planning in the MRB.

This research was funded by ‘Integrated management for sustainable utilization of water resources in East Africa great Lakes basin (No. 2018YFE0105900)’ and the project commissioned by National Key R&D program of China (No. 2018YFE0105900).

All the work is in compliance with ethical standards.

None declared.

Publicly available datasets were analyzed in this study. This data can be found here: https://esgf-node.llnl.gov/search/esgf-llnl/ and https://www.ncdc.noaa.gov/.